text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:cs/0610053v1 [cs.CE] 10 Oct 2006
Towards a Bayesian framework for option
pricing
Henryk Gzyl∗
Enrique ter Horst†
Samuel W. Malone‡
February 1, 2008
Abstract
In this paper, we describe a general method for constructing the
posterior distribution of an option price. Our framework takes as inputs the prior distributions of the parameters of the stochastic process
followed by the underlying, as well as the likelihood function implied
by the observed price history for the underlying. Our work extends
that of Karolyi (1993) and Darsinos and Satchell (2001), but with the
crucial difference that the likelihood function we use for inference is
that which is directly implied by the underlying, rather than imposed
in an ad hoc manner via the introduction of a function representing
“measurement error.” As such, an important problem still relevant for
our method is that of model risk, and we address this issue by describing how to perform a Bayesian averaging of parameter inferences
based on the different models considered using our framework.
JEL CLassification: C11, C13, G12
Keywords: Bayesian analysis; Lévy processes; Option pricing; Risk-Neutral
measure; Model selection; Bayesian nonparametrics.
∗
Henryk Gzyl, Instituto de Estudios Superiores de Administración IESA, Caracas, DF,
Venezuela. [email protected]
†
Enrique ter Horst, Instituto de Estudios Superiores de Administración IESA, Caracas,
DF, Venezuela. [email protected]
‡
Samuel W. Malone,
Oxford University,
Oxford,
United Kingdom.
[email protected]
1
Acknowledgments
The authors would like to thank German Molina and Robert L. Wolpert for
helpful conversations.
1
Introduction
In this article, we present a novel method for “integrating out” parameters
from the risk-neutral pricing formula of an option. We are not the first
to propose a method that uses this well-known Bayesian technique for the
problem of option pricing. Previous innovations in the area of Bayesian
econometrics on techniques for integrating out parameters after the riskneutral pricing formula for the option has been derived in closed form include
Eraker et al (2000). However, no work has been done, to our knowledge, on
integrating out the parameters during the transformation of the physical
measure P to the risk neutral measure Q, except when testing sequentially
a precise hypothesis concerning the drift of a Brownian motion as in Rui
(2002), or computing Bayes factors between different models as in Polson
and Roberts (1994).
Karolyi (1993), Darsinos and Satchell (2001) are perhaps the previous
work most closely related to our own. In their articles, the authors use
the Bayesian technique of integrating out parameters to derive the posterior
distribution in closed-form for a European call option when the underlying
follows a geometric Brownian motion. To do this, however, they use the
sufficiency property of the unbiased estimator of σ 2 from discreetly sampled
observations. The method we propose differs crucially in this respect: to
obtain the posterior of the option price, we avoid what may be considered
an inconsistency in previous methods, by using the likelihood implied by the
stochastic process of the underlying. This is important on practical grounds,
because the different likelihoods, except in a special case, will lead to different
posterior distributions for the parameters, and this affects inference.
In general, the methodology developed in this paper is able to extend and
yield the posterior distribution, in closed-form, as in the work of Karolyi
(1993) and Darsinos and Satchell (2001). We construct these posterior distributions by combining the likelihood function that is implied by the underlying stochastic process with the prior distributions that are specified as the
views of the market participant.
2
The method described in the remainder of our paper serves as an illustration of the linkages to be made between the broad areas of mathematical
finance, on one hand, and Bayesian probability, on the other. We take a continuous time point of view for the purposes of exposition, but the technique
can be formulated in discrete time and implemented for practical purposes.
This is a subject of future research.
The outline of this paper is as follows. Section 2 reviews the literature on
Bayesian option pricing and motivates our framework. Section 3 presents the
methodology and describes how to find a likelihood function (Q the RadonNikodym derivative of the risk neutral measure with respect to the physical
measure P) by the use of the Esscher transform. It also addresses the question
as how to choose acceptable prior distributions for continuous-time finance
model parameters in order to perform a Bayesian analysis. Section 4 illustrates the methodology with two examples pertaining to the classical Black
and Scholes model, as well as in a diffusion case. Section 4 concludes.
2
Previous work on Bayesian option pricing
The mainstream Bayesian literature has concerned itself with using statespace models as a way to get posterior distributions for derivatives perturbed
around a Black & Scholes price of the following sort:
St
log
= µ + σ(Wt − Wt−1 )
(1)
St−1
Ct = BS(σ, St ) + ǫt
(2)
where Wt is a Brownian motion, the error term ǫt ∼ N(0, ν 2 ), and BS(σ, St )
is the option price from the classical Black and Scholes model. A more indepth discussion of this approach can be found in pp. 35-36 of Johannes and
Polson (2002). Both Johannes and Polson (2002), as well as Darsinos and
Satchell (2001) obtain the posterior distribution of the volatility parameter σ
from the discrete time version of the continuous time process. Although this
posterior distribution exists in discrete time, however, in continuous time it
is a degenerate point mass, as Polson and Roberts (1994) explain, under a
particular reference probability measure.
In order to get the posterior distribution of the theoretical Black and Scholes price, Johannes and Polson (2002) use a perturbation ǫt around the
3
theoretical Black and Scholes price to construct a likelihood. Our framework consists of retrieving the likelihood function directly from the Radon⋆
Nikodym process (Ztθ ≡ dQ
| ) used when performing a change of measure
dP Ft
from to the physical measure P to the risk-neutral measure Q, and where
θ⋆ is a function of the vector parameter θ governing the probability distribution of the logarithm of the exponential Lévy process St . Combining our
⋆
dP
likelihood (Ztθ )−1 ≡ dQ
|Ft with a prior π(θ) enables us to derive a posterior distribution for π(θ|{Ss : 0 ≤ s ≤ t}) given an observed price history
{Ss : 0 ≤ s ≤ t}. Darsinos and Satchell (2001), working within a Black and
Scholes model and using the density of the distribution of the underlying
with respect to the Lebesgue measure, are able to find a closed-form solution
for the posterior of the call option. Our method can be carried out with the
use of numerical simulations from the parameter’s posterior distribution by
standard Bayesian numerical methods as we illustrate throughout the paper.
Besides allowing us to integrate prior beliefs about the price process parameters together with the likelihood inherent in the price process to generate
a marginal distribution for the price of the option, our method has another
theoretical motivation as well. In the method summarized by equation 1 and
equation 2, the parameters σ and ν are estimated jointly using the call option
price data and the price data for the underlying (see Johannes and Polson
(2002) for the details). As a consequence, both of these parameters capture
the joint effects of risk aversion and price volatility on option pricing, and it
is somewhat unclear how to sort out these effects in that model. This is not
surprising, since in that case the error term ǫt is interpreted as some sort of
observation error.
The classical framework of option pricing supposes a call option C(t, St , θ)
whose payoff H(ST ) depends on our underlying St , and can be computed via
the following integration 1 :
C(t, St , θ) = exp (−r(T − t)) EQ {H(ST ) | Ft }
where integration is performed under the risk-neutral measure Q, such that
the discounted stock price exp (−rt) St is a Q-martingale. General integration
theory states that the following change of measure is also possible by invoking
1
For a European call option, the payoff function is equal to max(ST − K, 0) where K
is the strike price at termination date T .
4
the Radon-Nikodym theorem:
⋆
EP {ZTθ H(ST ) | Ft }
C(t, St , θ) = exp (−r(T − t))
⋆
EP {ZTθ | Ft }
and if considering a prior π(dθ), we can perform the following integration
with respect to the prior π(dθ):
Z
C(t, St ) = exp (−r(T − t)) C(t, St , θ)π(dθ)
ZΘ
⋆
EP {ZTθ H(ST ) | Ft }
= exp (−r(T − t)) π(dθ)
⋆
Ztθ
Θ
Z
dP
⋆
= exp (−r(T − t)) π(dθ)(
|Ft )EP {ZTθ H(ST ) | Ft }
dQ
ZΘ
⋆
= exp (−r(T − t)) g(Su : u ≤ t)π(dθ | Ft )EP {ZTθ H(ST ) | Ft }
Θ
⋆
dP
|Ft , and
where θ ∈ Θ, Θ is the parameter space, (Ztθ )−1 = dQ
Z
dP
|F
g(Su : u ≤ t) =
π(dθ)
dQ t
Θ
is the marginal probability distribution of the underlying St , which is thus a
constant given the information Ft .
The posterior distribution π(dθ | Ft ) comes from a direct application of Bayes
rule:
π(dθ | Ft ) =
dP
π(dθ)( dQ
|Ft )
g(Su : u ≤ t)
dP
|F )
π(dθ | Ft )g(Su : u ≤ t) = π(dθ)(
dQ t
This interesting result shows that in order to integrate out the uncertainty
related to the governing parameters from the probability distribution of the
underlying, one needs to use the likelihood that comes automatically by the
dP
specification of underlying through dQ
|Ft .
It is often the case that competing models could have generated the underlying process St . Let us suppose that we have k of those competing models2
2
Example of such models could be the case where the underlying process St is described
either by a jump-diffusion, diffusion, or by a pure jump Lévy process, or by the same
process with different parameters.
5
each of which has a prior probability of P(Mi ), a Radon-Nikodym derivative
dPi
| , and a model parameter vector θi , for i = 1, · · · , k. These vectors
dQi Ft
θi might have different meanings depending on the model they refer to. If
we wish to see which of the models models the data best, then one could
compute their posterior probabilities. The marginal likelihood P(St | Mi ) of
model Mi gives a measure of observing the data given that Mi is true. Each
different competing model will give rise to a different vector parameter θi ,
having similar or completely different interpretations. This last quantity can
be computed as follows:
Z
dPi
P(Su : u ≤ t | Mi ) =
( i |Ft )π(dθi ).
Θ dQ
The marginal likelihood together with the prior probability on model i enables to use Bayes theorem which updates our prior beliefs that the data
comes from model Mi as follows:
P(Su : u ≤ t | Mi )P(Mi )
P(Mi | Su : u ≤ t) = Pk
i=1 P(Su : u ≤ t | Mi )P(Mi )
As Cont (2006) points out, model uncertainty regarding option pricing
models leads to “model risk” which can be seen as an extension of model
misspecification. Bayesian model averaging is a way to incorporate model
uncertainty for option prices, where one could compute an option price integrating model uncertainty in the following way:
C̃(t, St ) =
k
X
C(t, St | Mi )P(Mi | Su : u ≤ t)
i=1
where the C(t, St | Mi ) are the option prices C(t, St ) computed under model
Mi . C̃(t, St ) would thus yield a model weighted “average” option price.
we do not always have a closed-form solution for the integral
R Although
dPi
(
| )π(dθi ), we can approximate it through classical Markov Chain
Θ dQi Ft
Monte Carlo methods to get the marginal likelihood. In order to do this, we
need a likelihood and a prior distribution on the model parameters in order
to perform a Bayesian analysis. In the next section, we show how to find this
likelihood.
6
3
3.1
The Framework
Preview of coming attractions
As motivation for the material in the following sections, we present a version
of the Black-Scholes model in discrete time, which stands between the binomial model and the continuous time Black-Scholes model, in which the posteriorization procedure does not involve stochastic calculus. Assume that our
price process is described by a sequence Sn = eµ+Rn Sn−1 , and S0 ∈ R is given,
and where the sequence {Rn | n ≥ 1} is an i.i.d sequence of N(0, 1) random
variables. We take (Ω = RN , Fn , F , P ) as sample space, where Rn will denote
the coordinate mappings and Fn = σ{Rk | k ≤ n}, F = σ{Rk | k < ∞} and
P is the obvious infinite product of gaussian measures on (Ω, F ).
It is a simple calculation to verify that the value of a such that
2 σ 2 /2
E[eaR−a
eµ+Rn ] = er
is a = r−ν
where ν = µ − σ 2 /2. Also, it is a standard exercise to verify
2
σ
Q
2 2
that the limit Z of the Wald martingale Zn ≡ ni=1 eaRn −a σ /2 provides us
with a measure Q ∼ P such that Zn = dQ/dP |Fn and that the present value
price process Sn∗ = Sn e−r is a Q-martingale.
Pn
2 2
It is also simple to verify that the likelihood Zn−1 = e−a 1 Rk +na σ /2 =
dP/dQ|Fn is a Q-martingale. The interest in Zn−1 is that it can be thought
of as the conditional density dP
(S0 , ..., Sn | µ). To see how this comes about,
dQ
Pn
Pn
note that k=1 Rk = k=1 ln Sk /Sk−1 − nµ, and after completing squares
and re-arranging the exponent we obtain
n
2
σ2
−
ln
S
/S
r
+
/n ×
n
0
2σ 2
2
2
σ2
n
− ln Sn /S0 /n .
× exp − 2 ν +
2σ
2
Zn−1 (S0 , ..., Sn | µ) = exp
From this point on, the marginalization to obtain the posterior of µ given
S0 , ..., Sn is a routine matter in the gaussian set up. Let us now proceed to
the continuous time case.
7
3.2
Change of measures and the likelihood function in
option pricing
It is common to use the following form when modelling the underlying St of
a derivative:
St = exp (Xt )
where Xt can either be a Lévy process or a diffusion, and thus St is the
exponential3 of this process discounted by exp(−rt) . In order to price options, one needs to find an equivalent probability measure Q ∼ P such that
St ≡ exp (−rt + Xt ) is a martingale under Q.
When performing a change of measure for a given stochastic process St
under P to Q, one can regard S as a random variable on the space Ω =
D[0, +∞) of càdlàg (Continue à droite limites à gauche) paths together with
its associated filtration (Ft )t≥0 . This measure P is therefore defined on the
dP
space of sample paths of X, and so the Radon-Nikodym derivative dQ
|Ft
4
with respect to the reference measure Q is the likelihood function after the
process has been observed up to time t.
It is easy to verify that the independence of the increments of X imply
that Ztθ⋆ = exp (θ⋆ Xt − tk(θ⋆ )). Furthermore, Ztθ⋆ is not only a positive Ft martingale under Pθ , but happens to be the
likelihood function given timeθ t.
∗
What is interesting here is that St ≡ St exp (rt) = exp (Xt − rt) is a Q ⋆ martingale for every θ as is easy to verify. This observation will enable us to
use Bayes theorem in the following sections to compute posterior distributions
of the parameters that govern the dynamics of the underlying St . Since we
will mostly work with continuous stochastic processes, we set Ω = C[0, ∞),
and5 (Ω, F , P) is the standard Wiener space.
3
See Applebaum (2004) and Oksendal (2003) for the case when Xt is a Lévy process
and a diffusion respectively.
4
See chapter X, section 2 of Jacod and Shiryaev (1987) on the equivalence between
Radon-Nikodym derivatives and likelihood function.
5
C[0, ∞) is the space of continuous functions defined from [0, ∞) into R.
8
4
Parameter posteriors for some popular continuous time models of the underlying
4.1
Geometric Brownian motion
In the classic paper by Black and Scholes (1973), the stock price St is solution
to the following SDE:
dSt
= µdt + σdWt
St
(3)
whose solution is equal to:
St
σ2
= S0 exp (µ − )t + σWt
2
Working with the discounted stock price, we obtain
σ2
exp (−rt) St = S0 exp (µ − r − )t + σWt
2
(4)
(5)
Invoking the Cameron-Martin-Girsanov (CMG) theorem enables us to determine the deterministic risk-neutral condition µ = r which is the one that
defines the unique martingale measure. We then get that St is a martingale6
under Q and is equal to:
σ2
St = S0 exp (r − )t + σWt
(6)
2
As it turns out, the martingale condition µ = r holds asymptotically in
our Bayesian framework, as t tends to infinity, but for all finite t the value
of µ consistent with no-arbitrage has a posterior distribution that is nondegenerate.
Theorem 1. Under the model given by equation 3, from the CameronMartin-Girsanov theorem the density process Ztθ is given by
"
2 #
t r−µ
r−µ
−
(7)
Ztθ = exp Wt
σ
2
σ
6
St is a Q-martingale, which is equivalent to showing that Ztθ St is a P-martingale.
9
and the posterior distribution for the drift parameter µ is given by
σWt σ 2
[µ|{Ss : 0 ≤ s ≤ t}] ∼ N r −
,
t
t
(8)
Proof. See Appendix.
Theorem (1) illustrates that, even in the setting where markets are complete and we can sample continuously, we obtain a posterior distribution for
the drift parameter µ that differs from the usual no-arbitrage condition that
µ = r with probability 1. We recover this no-arbitrage condition in the limit
as time goes to infinity, and when time is finite, we have that if Wt > 0, then
t
< r, and vice versa.
E[µ|{Ss : 0 ≤ s ≤ t}] = r − σW
t
2
One important observation is that as t → +∞ we get σt → 0 and thus
t
µ → θ0 = r since σW
→ 0 as t → +∞. Therefore, the posterior distribut
tion [µ|{Ss : 0 ≤ s ≤ t}] is consistent at r = θ0 . This is a property that a
posterior distribution should have in general if we wish to learn more about
a parameter, and is the Bayesian analogue of consistency for an estimator in
classical statistics. We now do a variation on the definition of consistency
from Ghosh and Ramamoorthi (2003) and propose the following one:
Definition 1. For each t, let π(θ|Ss : 0 ≤ s ≤ t) be a posterior probability
distribution given {Ss : 0 ≤ s ≤ t}. The collection {π(θ|Ss : 0 ≤ s ≤ t)} is
said to be consistent at θ0 if there is a Ω0 ⊂ Ω = C[0, ∞) with Pθ0 (Ω0 ) = 1
such that if ω is in Ω0 , then for every neighborhood U of θ0 ,
π(U|Ss : 0 ≤ s ≤ t) → 1 a.s. Pθ0
(9)
where Pθ0 is a probability measure defined on the space of right continuous
functions with left limits, and St is the underlying process. When performing a Bayesian analysis, it is important for the posterior to converge to a
degenerate point mass at the unknown parameter θ0 , which means in our
setup, the posterior variance tends to 0. In the previous example, we obtain
2
θ
consistency as the posterior variance σt → 0. It is worth to notice
that Zt
depends on the initial and final values S0 and S0 through log SS0t . Working
with continuously compounded returns enables us to use the same definition
of consistentcy as in Ghosh and Ramamoorthi (2003), as well as to prove
that two agents having different prior distributions π(θ) reach in the end
10
the same conclusions regarding θ. Throughout our examples we shall show
that our posterior distributions are consistent for both µ and σ 2 . We now
cite from Ghosh and Ramamoorthi (2003) the following second definition of
consistency as well as a theorem:
Definition 2. For each n, let π(θ|Rt1 , . . . , Rtn ) be a posterior given the iid
sequence Rt1 , . . . , Rtn . The sequence {π(θ|Rt1 , . . . , Rtn )} is said to be consistent
at θ0 if there is a W0 ⊂ W with7 P∞
θ0 (Ω0 ) = 1 such that if ω is in Ω0 , then
for every neighborhood U of θ0 ,
π(U|Rt1 , . . . , Rtn ) → 1 a.s. P∞
θ0
(10)
Si
where Rti ≡ log( Sti ) is the return during the i-th interval of length t. Sti and
0
S0i are the prices at the beginning and the end of the i-th interval of length t.
We have n of those independent returns since for a given time interval [0, T ]
such that T = nt, we use underlying processes for St with Stationary and
Independent Increments.
What if two different agents used the same probability model that generated the underlying price process St , but had different prior distributions
regarding the parameters governing the latter. Would both posterior distributions be consistent at the same value θ0 ? In other words, under what
conditions would two different priors lead to the same inference? In order to
answer the latter, we cite from Ghosh and Ramamoorthi (2003) the general
definition of consistency for a posterior distribution, together with a theorem.
Theorem 2. Assume that the family {Pθt : θ ∈ A} is dominated by a σ-finite
measure µ and let pθ denote the density of Pθt . Let θ0 be an interior point
of Θ, and π1 , π2 be two priors densities with respect to a measure ν, which
are positive and continuous at θ0 . Let π(θ|Rt1 , . . . , Rtn )i , i = 1, 2 denote the
posterior densities of θ given {Rt1 , . . . , Rtn }. If π(θ|Rt1 , . . . , Rtn )i , i = 1, 2 are
both consistent at θ0 then:
R
limn |π1 (θ|Rt1 , . . . , Rtn ) − π2 (θ|Rt1 , . . . , Rtn )|dν(θ) = 0 a.s. Pθ0
(11)
where Pθt is the probability distribution of the return Rt = log( SS0t ).
7
Here W = R∞ is the space of infinite sequences, P∞
θ0 = ⊗Pθ0 is the product probability
measure defined on the sigma algebra B(R∞ ) of W. Here Pθ0 is the probability distribution
of the returns Rti with respect to Qθ0 .
11
This last theorem shows that as n increases (and therefore our information set increases), the importance of the prior distribution fades away (and
thus the importance of the prior’s hyperparameters as well), since we get to
observe more and more data. In our framework, the dominating measure µ
of Ghosh and Ramamoorthi (2003) is our Q and their family Pθt corresponds
to our physical probability measure regarding the returns Rt .
4.2
A Bayesian Inference of the Black & Scholes model
As we just saw in the previous section, the log returns of the stock price
h
i
S
log( S t )
σ2 σ2
0
∼
N
µ
−
,
are normally distributed8 , thus the conditional liket
2
t
lihoods for µ and σ 2 in the Black & Scholes model are proportional to (see
appendix):
S
log S t
0 2
2
1
St
tσ
1 t(µ − t )
L(σ 2 | log( ), µ) ∝ (σ 2 )( 2 −1) exp −
(12)
+
2
S0
σ2
4
"
#2
St
2
log S0
St
t
σ
L(µ | log( ), σ 2 ) ∝ exp − 2
− (µ − )
(13)
2σ
S0
t
2
The first conditional likelihood given µ is nothing but proportional to a Generalized Inverse Gaussian with parameters λ, δ, and γ (in short GIG(λ, δ, γ))
whose density is equal to:
1
1 δ2
γ λ
(λ−1)
2
x
exp −
+γ x
(14)
f (x | λ, δ, γ) = ( )
δ 2Kλ (γδ)
2 x
where we use the same parametrization as in Silva et al (2006). We note
that equation (12) is proportional to a GIG(λ, δ, γ) with parameters equal
log
St
S0
)2 , and γ 2 = 4t , and that the conditional posterior
S
log S t
σ2 σ2
2
0
distribution of µ given σ is normally distributed as N
+ 2 , t . We
t
to λ = 21 , δ 2 = t(µ −
t
prove these last statements and more in the following two lemmas whose
proofs are in the appendix.
dP
|Ft is with respect to Lebesgue measure λ.
Here our Radon-Nikodym derivative dλ
One can compute this ratio with respect to the probability measure P using the following
dQ
dλ
chain rule: dQ
dP |Ft = dλ |Ft dP |Ft .
8
12
Lemma 1. When choosing flat uniform prior distributions on R and R+ for
µ and σ 2 in the Black and Scholes model,
then their posterior conditional dis
tributions are π(µ | σ 2 , log SS0t ) = N
log
St
S0
t
+
σ2 σ2
, t
2
and π(σ 2 | µ, log SS0t ) =
log
GIG(λ, δ, γ) respectively, where λ = 12 , δ 2 = t(µ −
uniform priors on defined on R and R+ respectively.
t
St
S0
)2 , and γ 2 = 4t , for
Proof. See appendix.
Lemma 2. When chosing normal prior distributions (π(µ) = N [m, s]) on
R for µ, and a π(σ 2 ) = GIG(λ, δ, γ) on R+ for σ 2 in the Black & Scholes model, their posterior conditional distributions are π(µ | σ 2 , log SS0t ) =
St σ2 t mσ2
log S + 2 + 2
σ2
s
0
N
, σ2 and π(σ 2 | µ, log SS0t ) = GIG(λ′ , δ ′ , γ ′ ), where
σ2
t+
λ′ = λ − 21 ,
δ ′2 = t(µ −
t+
s2
log
t
St
S0
s2
)2 + δ 2 , and γ ′2 =
t
4
+ γ2.
Proof. See appendix.
When modelling the underlying with a given process in continuous time
(in this example a Geometric Brownian Motion), there is already a likelihood
function implied by this latter. This methodology proposes an extension of
Karolyi (1993), Darsinos and Satchell (2001), where their posterior for σ 2 is
an Inverse Gamma distribution, which appears to be a specific case of the
GIG(λ, δ, γ) as explained in Silva et al (2006). Furthermore, both Karolyi
(1993), Darsinos and Satchell (2001) construct the likelihood for σ 2 relying
on the mathematical result that when log returns are normally distributed,
2
the statistic9 νs
∼ χ2 (ν) and independent of µ since the latter is sufficient.
σ2
Our extension relies on a more general mathematical framework that consists
of working with the Radon-Nikodym derivative Zt instead10 .
With this example we illustrate one of the core points of this paper that
one is not free to elicit a likelihood when imposing a mathematical model
for the underlying stochastic process, which already has one. These two
likelihoods can lead to different inferences for the parameter σ 2 unless they
9 2
1
n−1
10
χP(ν) is a chi-square random variable with ν = n − 1 degrees of freedom and s2 ≡
n
i
2
i=1 (Rt − R̄t ) .
Here the reference measure is just Lebesgue measure λ.
13
are by coincidence proportional one to another as Berger and Wolpert (1984)
point out.
From the Cameron-Martin-Girsanov theorem, the likelihood function for
a general diffusion process dXt = f (θ, t, Xt )dt + σ(t, Xt )dWt is given by11 :
Z t
(r − f (θ, s, Xs))
St
Zt = exp
×
d log
σ(s, Xs )
S0
0
Z
1 t (r − f (θ, s, Xs ))(r + f (θ, s, Xs ) − σ 2 )
ds
× exp −
2 0
σ2
n
o
Rt
where St = S0 exp Xt − 21 0 σ(s, Xs )2 ds is the stock price process. Polson and Roberts (1994) compute posterior probability distributions as well
as Bayes factors for the drift of the diffusion when f (θ, t, Xt ) = θf (t, Xt ).
Our posterior distribution for the drift reduces to example 4 in Polson and
Roberts (1994) when r = 0. We not only develop a Bayesian methodology
for consistently estimating parameters in the context of option pricing, but
also compute the posterior probability distributions for both σ 2 and µ, as
well as in continuous time, extending the work by Karolyi (1994).
5
Discussion and Conclusions
When traders and market participants use pricing formulas for derivatives,
the price they obtain is a function of the parameter values that they assume. However, it is reasonable to expect that these parameter values are
not known with certainty. As a result, it is worthwhile to take this uncertainty about parameter values into account in option pricing. The natural
way of doing this is through the use of Bayesian methods. In this paper,
we present a framework for Bayesian option pricing that can be applied to
a very general set of stochastic processes for the underlying. Using directly
the probability model for the stochastic process of the stock price process,
Girsanov’s theorem helps us derive the likelihood function which one of the
key ingredients that enables us to yield posterior probability distributions for
model parameters.
We show that in our framework, in which the posterior distribution for the
parameter estimates is computed using the likelihood implied by the underlying stochastic process and prior distribution that represents agents´ beliefs
11
See Oksendal (2003).
14
about the parameter values, the likelihood function is constructed from a
martingale. Asymptotically, as time goes to infinity and we collect more
data, our prior beliefs can be updated. Moreover, we discuss our framework with the example of geometric Brownian motion, where our findings
are consistent with those of Polson and Roberts (1994) in the context of the
posterior distribution of the drift µ. We extend the methodology by Karolyi
(1993), by noticing how to retrieve the implied likelihood imposed by the
stochastic process of the underlying. In this sense we show that when working with derivatives, it is important to have coherence between the likelihood
implied by the stochastic process of the underlying and the likelihood of the
econometric model used for inference.
In terms of future research, the most worthwhile application of these techniques would probably be to the case of a tractable version of a general Lévy
process, which in continuous time yields nontrivial likelihoods for the case of
jump-diffusion processes for all of the parameters, as we see in the appendix.
In addition, a potentially fruitful application of this technique in discrete
time, to binomial tree pricing methods, is the subject of current research.
Appendix
Posterior for µ
It is a standard
computation
to show that the value of a ∈ R such that
a2 t
a
Zt ≡ exp aWt − 2 is a density for the risk neutral measure Q making
Zta St exp (−rt) into a P-martingale is a = r−µ
. If θ = (µ, σ 2 ) are the model
σ
r−µ
parameters, then θ⋆ = a(θ) = σ2 .
We see how θ⋆ is a function of θ ≡ (µ,σ) and makes St = exp(Xt ) into a
| is equal to:
Q-martingale, where the Radon-Nikodym derivative dQ
dP Ft
dQ
σ2
σ 2 θ⋆2
t .
|F = exp θ⋆ Xt − θ⋆ (µ − ) +
dP t
2
2
Substituting θ⋆ in the Radon-Nikodym derivative and using a flat uniform
15
prior for µ, the posterior is proportional to:
dQ
t
σWt
2
|F ∝ exp − 2 [µ − 2µ(r −
)]
dP t
2σ
t
t
σWt 2
∝ exp − 2 [µ − (r −
)] .
2σ
t
Weh conclude that
i the posterior distribution for µ is:
σWt σ2
N r− t , t
A
Posterior probability distributions for JumpDiffusion processes
Consider first the case where the price process is St = eνt+σWt , with W
as above and ν = µ − σ 2 /2. In this case the getting (Ztθ⋆ )−1 ready for
posteriorization will be pretty much like in section 3.1 Note to begin with
ˆ
that from dS
t = µSt dt + σSt dWt and Ito’s formula that on the one hand
σdWt = dSt St − µdt and on the other hand, therefore
and therefore that
σ2t
d(ln St S0 +
) = dSt St
2
σWt = ln St S0 − tν.
With all this, just some arithmetics allows us to rearrange the exponent in
2 2
(Ztθ⋆ )−1 = exp − θσ⋆ σWt + (θ⋆ )2 σ t , where recall, θ⋆ = (r − µ)/σ, so that
t
σ 2 ln St S0 2
t
σ 2 ln St S0 2
θ⋆ −1
(r −
+
) − 2 (µ −
+
) .
(Zt ) = exp
2σ 2
2
t
2σ
2
t
When the asset prices are driven by discontinuous factors, part of the
routine is easy, but part is not. Assume for example that dSt = µSt dt +
σSt dWt + dJ(t), where the first two terms in the right hand side are as in the
P (t)
previous example, whereas J(t) = N
n=1 ξn is a compound Poisson process
such that N(t) is a Poisson Process with intensity λ and the i.i.d. sequence is
independent of both W and N, and for the sake of simplicity, let us assume
the ξn to be bounded.
16
The price equation has solution
N (t)
t(µ−σ2 /2)+σWt
St = e
Y
(1 + ξn ) = et(µ−σ
2 /2)+σW
t +Y
(t)
n=1
PN (t)
PN (t)
where Yt =
n=1 ln(1 + ξn ) ≡
n=1 ηn . Clearly we must assume that
(1+ξn) > 0 for the price process to beRpositive. The boundedness assumption
on the jumps yields the existence of (eθx − 1)n(dx), with n(dx) = dP (η1 ≤
x). With this, clearly E[eθ(σWt +Yt ) ] = etk(θ) , with
Z
σ2
k(θ) =
+ λ (eθx − 1)n(dx)
2
This time when we ask for the existence of a value θ⋆ such that
E[eθ(σWt +Yt )−tkθ) St ] = etr
we are led to the equation k(θ⋆ + 1) − k(θ⋆ ) = r − ν, with ν as above. This
amounts to solving
Z
r−ν
θ⋆ + λ (ex − 1)exθ⋆ n(dx) =
.
σ2
When θ∗ ranges from −∞ to +∞, the left hand side of the identity above
, ranges in an increasing way over the same range, thus the equation has a
solution. So there is at least one risk neutral measure for the given asset
ˆ
price. Furthermore, it is not hard to see using Ito’s
formula that this time
we also have
σWt + Yt = ln St S0 − tν
but the likelihood
Ztθ⋆
= exp
−θ⋆ (σWt +Yt )+tk(θ⋆ )
−θ⋆ ln(St S0 )−tν +tk(θ⋆ )
=e
does not lead to a posterior distribution for θ easy to sample from.
B
Proof of Lemmas 1 and 2
In this section we derive the full conditional posterior distributions for both
µ and σ 2 in the Black & Scholes model. In their model, the log returns are
17
S
normally distributed as
log( S t )
0
t
h
∼N µ−
σ2 σ2
, t
2
i
. Once we observe the whole
sample path St , we get the following function for µ and σ 2 :
"
#2
St
2
log
St
t
1
σ
S0
2
L(µ, σ | log( )) = p
exp − 2
− (µ − )
(15)
2
2σ
S0
t
2
(2π σt )
Proof for Lemma (1):
Proof. Since the priors for both µ and σ 2 are both flat priors, the joint
posterior distribution is proportional to the likelihood function, which is equal
to equation (15). Keeping only the terms in σ 2 in equation (15), we conclude
easily that the conditional posterior distribution for σ 2 is a GIG(λ′ , δ ′ , γ ′ )
log
St
S0
= t(µ − t )2 , andγ ′2 = 4t . The conditional posterior
S
log S t
2
2
0
distribution for µ is a N
+ σ2 , σt .
t
′
where λ =
1
,
2
δ
′2
Proof for Lemma (2):
Proof. By Bayes rule, the posterior π(σ 2 | log( SS0t ), µ) for σ 2 given µ and the
data is proportional to:
St
), µ)π(σ 2 )
S0
S
log S t
0 2
2
t(µ − t )
1
1 tσ
π(σ 2 )
∝ (σ 2 )( 2 −1) exp −
+
2 4
σ2
∝ L(σ 2 | log(
and the likelihood π(σ 2 | µ, log( SS0t )) times the GIG(λ, δ, γ) prior distribution
π(σ 2 ) yields:
S
log S t
2
0 2
(t(µ − t ) + δ )
1 2 t
2 (λ− 21 −1)
2
exp −
(σ )
σ ( +γ )+
2
4
σ2
which is proportional to a GIG(λ′ , δ ′ , γ ′ ), where λ′ = λ − 12 , δ ′2 = t(µ −
log
t
St
S0
)2 + δ 2 , and γ ′2 =
t
4
+ γ 2.
18
The full conditional posterior distribution for µ given σ 2 is given by π(µ |
σ , log( SS0t )) and is proportional to:
2
∝
∝
∝
∝
∝
"
#2
St
t
2
log S0
σ
π(µ) exp − 2
− (µ − )
2σ
t
2
"
#2
St
t
2
log S0
σ
1
− (µ − ) − 2 (µ − m)2
exp − 2
2σ
t
2
2s
St
log S
2
σ
2
2
0
mσ + ts
+ 2
(t + σ2 )
t
2
2
s
exp −
µ
−
2µ
2σ 2
ts2 + σ 2
"
!#)
(
2
2
2
m σs2 + log SS0t + tσ2
(t + σs2 ) 2
µ − 2µ
exp −
2
2σ 2
t + σs2
"
!#2
St
tσ2
σ2
(t + σ2 )
m s2 + log S0 + 2
s2
exp −
µ−
2
2σ 2
t + σs2
where π(µ) is a N(m, s2 ). We conclude therefore
conditional
poste σ2that Sthe
2
m 2 +log S t + tσ2
2
s
0
, σσ2 , which
rior distribution for µ is distributed as a N
σ2
t+
t+
s2
s2
is the same conditional posterior distribution as in Polson and Roberts (1994).
B.1
Consistency of posterior distributions for µ and σ 2
We now prove consistency of the posterior distibutions for both µ and σ 2 in
2
the sense of definition 1. The posterior variance of µ is σσ2 and goes to 0
t+
s2
as t → ∞, which together with Chebychev’s inequality shows consistency.
The posterior variance of σ 2 given the data and µ is given by Corollary
1.1 from Silva et al (2006):
2 "
2 #
δ
K
(γλ)
K
(γλ)
λ+1
λ+2
V ar(σ 2 | λ, δ, γ) =
−
γ
Kλ (γλ)
Kλ (γλ)
19
where Kν (ω)
function of the third kind and equal to
R is the modified
1 Bessel −1
1 ∞ ν−1
Kν (ω) = 2 0 x
exp − 2 ω (x + x ) dx. By noticing that
lim Kν (ω) = 0
ω→∞
p
1
, Kν (ω) ≈ ω − 2 exp (−ω) π2 (for ω very big), and that Kν (ω) ≥ 0, we get
the following inequalities for the posterior variance as ω → ∞:
2 "
2 #
δ
K
(γλ)
K
(γλ)
λ+2
λ+1
V ar(σ 2 | λ, δ, γ) =
−
γ
Kλ (γλ)
Kλ (γλ)
pπ
p π !2
2
− 21
− 21
ω exp (−ω) 2
δ ω exp (−ω) 2
≈
pπ −
p
1
1
−
γ
ω 2 exp (−ω) 2
ω − 2 exp (−ω) π2
= 0 as ω → ∞.
This last result shows that the posterior variance of σ 2 goes to zero as
ω → ∞. Combining this last fact together with Chebychev’s inequality
shows consistency of the posterior probability distribution of both σ 2 and µ.
B.2
Application of lemmas 1 and 2 for Gibbs sampling
Given the full conditional distributions for both µ and σ 2 from lemma 1 and
2, we can use them to construct a Gibbs sampler in order to get the posterior
probability distributions for both µ and σ 2 in the Black & Scholes model.
We now present the Gibbs algorithm:
• Initialize both µ0 and σ 2,0 with starting values
σ2,i−1
2,i−1
S
m
+log S t + tσ 2
σ2,i−1
s2
i
0
• Draw µ ∼ N
, σ2,i−1
σ 2,i−1
t+
t+
s2
s2
• Draw σ 2,i ∼ GIG(λ′ , δ ′ , γ ′ ), where the parameters λ′ , δ ′ , and γ ′ are
given by the above lemmas
Repeat for i = 1, · · · , I where I is big enough.
The posterior for µ given σ 2 converges to its posterior mean since the posterior variance for both µ and σ 2 goes to 0 as t → ∞ from subsection B.1.
20
References
[1] Applebaum, D. (2004)“Lévy Processes and Stochastic Calculus.”,
Cambridge University Press.
[2] Berger, J.O., and Wolpert, R.L. (1984)“The Likelihood Principle.”,
Hayward, CA: INstitute of Mathematical Statistics.
[3] Black, F. and Scholes, M. (1973)“The Pricing of Options and Corporate Liabilities” Journal of Political Economy, 81 637-654.
[4] Bliss, R. and Panigirtzoglou,N.(2004)“Option-implied risk aversion estimates.” The Journal of Finance LIX, No.1.
[5] Bohman, H. and Esscher, F. (1963)“Studies in risk theory with numerical illustrations concerning distribution functions and stop loss
premiums i,ii.” Skand. Aktuarietidskr. 174-225.
[6] Choudhuri, N., Ghosal, S. and Roy, A. (2004)“Bayesian Methods for
Function Estimation.” Tech. rep., North Carlina State University.
[7] Cont, R. (2006)“Model Uncertainty and its impact on the Pricing of
Derivative Instruments.” Mathematical Finance, 13 519-547. .
[8] Cont, R. and Tankov, P. (2003)“Financial Modelling with Jump Processes.” Chapman & Hall.
[9] Darsinos, T. and Satchell, S.E. (2001)“Bayesian Analysis of the Black
& Scholes Option Price.” Tech. rep., Cambridge Working Papers in
Economics.
[10] Eraker, B., Johannes, M. and Polson, N. (2000)“The impact of jumps
in volatility and returns.” Journal of Business 75 305-332.
[11] Gerber, H. and Shiu, E. (1994)“Option pricing by Esscher transforms.”
Trans. Soc. Actuar. XLVI 98-140.
[12] Ghosh, J. and Ramamoorthi, R. (2003)“Bayesian Nonparametrics.”
Springer Verlag.
[13] Hubalek, F. and Sgarra, C. (2006)“Esscher transforms and the minimal entropy martinagle measure for exponential Lévy models.” Journal
of Quantitative Finance (Forthcoming).
21
[14] Jacod, J. and Shiryaev, A.N. (1987)“Limit Theorems for Staochastic
Processes.” Springer-Verlag.
[15] Johannes, M. and Polson, N. (2002)“MCMC methods for financial
econometrics.”
[16] Kallsen, J. and Shiryaev, A. (2002)“The cumulant process and the
esscher’s change of measure” Finance Stoch. 6 397-428.
[17] Karolyi, A. (1993)“A Bayesian Approach to Modeling Stock Return
Volatility for Option Valuation.” Journal of Financial and Quantitative Analysis 28, December 1993.
[18] Lancelot, F. J. (2005)“Bayesian Poisson Process Partition Calculus
with an application to Bayesian Lévy Moving Averages.” Annals of
Statistics 33 1771-1799.
[19] Oksendal, B. (2003)“Stochastic Differential Equations: An Introduction with Applications.” Springer-Verlag.
[20] Paulo, R. (2002)“Problems on the Bayesian/Frequentist Interface.”
Ph.D. thesis, Duke University, ISDS.
[21] Polson, N. and Roberts, G. (1994)“Bayes factors for discrete observations from diffusion processes.” Biometrika 81 11-26.
[22] Silva, R.S. and Lopes, H.S. and Migon, H.S. (2006)“The Extended
Generalized Inverse Gaussian for Log-Linear Stochastic Volatility
Models.” Technical Report.
[23] Terence, C. (1999)“Pricing Contingent Claims on Stocks Driven by
Lévy Processes.” The Annals of Applied Probability 9 504-528.
[24] Wolpert, R.L. and Ickstadt, K. (1998)“Simulation of Lévy Random Fields.” Practical Nonparametric and Semiparametric Bayesian
Statistics 227-242.
22
| 5cs.CE
|
Work Analysis with
Resource-Aware Session Types
Ankush Das, Jan Hoffmann and Frank Pfenning
arXiv:1712.08310v1 [cs.PL] 22 Dec 2017
Carnegie Mellon University, Pittsburgh, PA, USA
Abstract. While there exist several successful techniques for supporting programmers in deriving static resource bounds for sequential code,
analyzing the resource usage of message-passing concurrent processes
poses additional challenges. To meet these challenges, this article presents
an analysis for statically deriving worst-case bounds on the total work
performed by message-passing processes. To decompose interacting processes into components that can be analyzed in isolation, the analysis is
based on novel resource-aware session types, which describe protocols and
resource contracts for inter-process communication. A key innovation is
that both messages and processes carry potential to share and amortize
cost while communicating. To symbolically express resource usage in a setting without static data structures and intrinsic sizes, resource contracts
describe bounds that are functions of interactions between processes.
Resource-aware session types combine standard binary session types and
type-based amortized resource analysis in a linear type system. This type
system is formulated for a core session-type calculus of the language
SILL and proved sound with respect to a multiset-based operational cost
semantics that tracks the total number of messages that are exchanged in
a system. The effectiveness of the analysis is demonstrated by analyzing
standard examples from amortized analysis and the literature on session
types and by a comparative performance analysis of different concurrent
programs implementing the same interface.
1
Introduction
In the past years, there has been increasing interest in supporting developers
to statically reason about the resource usage of their code. There are different approaches to the problem that are based on type systems [30,37,18,15,27],
abstract interpretation [25,5,16], recurrence relations [20,4,36], termination analysis [46,11,8,31], and other techniques [14,19]. Among the applications of this research we find the prevention of side channels that leak secret information [38,6,35],
identification of complexity bugs [39], support of scheduling decisions [1], and
help in profiling [26].
While there has been great progress in analyzing sequential code, relatively
little research has been done on analyzing the resource consumption of concurrent
and distributed programs [22,3,2]. The lack of analysis tools is in sharp contrast
to the need of programming language support in this area: concurrent and
distributed programs programs are both increasingly pervasive and particularly
difficult to analyze. For shared memory concurrency, we need to precisely predict
scheduling to account for synchronization cost. Similarly, the interactive nature
of message-passing systems makes it difficult to decompose the system into
components that can be analyzed in isolation. After all, the resource usage of
each component crucially depends on its interactions with the world.
In this paper, we study the foundations of worst-case resource analysis for
message-passing processes. A key idea of our approach is to rely on resourceaware session types to describe structure, protocols, and resource bounds for
inter-process communication that we can use to perform a compositional and
precise amortized analysis. Session types [32,33,12,13,45] prescribe bidirectional
communication protocols for message-passing processes. Binary session types
govern the interaction of two processes along a single channel, prescribing complementary send and receive actions for the processes at the two endpoints of
a channel. We use such protocols as the basis of resource usage contracts that
not only specify the type but also the potential of a message that is sent along a
channel. The potential (in the sense of classic amortized analysis [43]) may be
spent sending other messages as part of the network of interacting processes, or
maintained locally for future interactions. Resource analysis is static, using the
type system, and the runtime behavior of programs is not affected.
We focus on bounds on the total work that is performed by a system, counting
the number of messages that are exchanged. While this alone does not account yet
for the concurrent nature of message-passing programs it constitutes a significant
and necessary first step. The bounds we derive are also useful in their own right.
For example, the information can be used in scheduling decisions, to bound the
number of messages that are sent along a specific channel, or to statically decide
whether we should spawn a new thread of control or execute sequentially when
possible. Additionally, bounds on the work of a process can also serve as input
to a Brent-style theorem [10] that relates the complexity of the execution of a
program on a k-processor machine to the program’s work (the focus of this paper)
and span (the resource usage if we assume an unlimited number of processors).
We are working on a companion paper for deriving bounds on the span, which is
both conceptually and technically quite different.
Our analysis is based on a linear type system that combines standard binary
session types as available in the SILL language [44,40], and type-based amortized
resource analysis [30,27]. Both techniques are based on linear or affine type
systems, making their combination natural. Each session type constructor is
decorated with a natural number that declares a potential that must be transferred (conceptually!) along with the corresponding message. Since the interface
to a process is characterized entirely by the resource-aware session types of the
channels it can interact with, this design provides for a compositional resource
specification. For closed programs (which evolve into a closed network of interacting processes) the bound then becomes a single constant. In addition to the
natural compositionality of type systems we also preserve the good support for
deriving resource annotations via LP solving which is a key feature of type-based
amortized analysis. While we have not yet implemented a type inference algorithm, we designed the system with support for type inference in mind. Moreover,
resource-aware session types integrate well with type-based amortized analysis
for functional programs because they are based on compatible logical foundations
2
(intuitionistic linear logic and intuitionistic logic, respectively), as exemplified in
the design of SILL [44,40] that combines them monadically.
A conceptual challenge of the work is to express symbolic bounds in a setting
without static data structures and intrinsic sizes. Our innovation is that resourceaware session types describe bounds that are functions of interactions (i.e.,
messages sent) along a channel. A major technical challenge is to account for
the global number of messages sent with local derivation rules: Operationally,
local message counts are forwarded to a parent process when a sub-process
terminates. As a result, local message counts are incremented by sub-processes in
a rather non-local fashion. Our solution is that both messages and processes carry
potential to share and amortize the cost of a terminating sub-process proactively
as a side-effect of the communication.
Our main contributions are as follows. We present the first session type system
for deriving parametric bounds on the resource usage of message-passing processes.
We prove the nontrivial soundness of type system with respect to an operational
cost semantics that tracks the total number of messages exchanged in a network
of communicating process. We demonstrate the effectiveness of the technique by
deriving tight bounds for some standard examples from amortized analysis and
the literature on session types. We also show how resource-aware session types
can be used to specify and compare the performance characteristics of different
implementations of the same protocol. The analysis is currently manual, with
automation left for future work, as is a companion type system for deriving the
span of concurrent computations in the same language.
2
Overview
In this section, we motivate and informally introduce our resource-aware session
types and show how they can be used to analyze the resource usage of messagepassing processes. We start with building some intuition about session types.
Session Types. Session types have been introduced by Honda [32] to describe the
structure of communication just like standard data types describe the structure
of data. We follow the approach and syntax of SILL [44,40] which is based
on a Curry-Howard isomorphism between intuitionistic linear logic and session
types, extended by recursively defined types and processes. In the intuitionistic
approach, every channel has a provider and a client. We view the session type as
describing the communication from the provider’s point of view, with the client
having to perform matching actions.
As a first simple example, we consider natural numbers in binary form. A
process providing a natural number sends a stream of bits starting with the least
significant bit. These bits are represented by messages zero and one, eventually
terminated by $. Because the provider chooses which messages to send, we call
this an internal choice, which is written as
bits = ⊕{zero : bits, one : bits, $ : 1} .
3
Here, ⊕{l1 : A1 , . . . , ln : An } is an n-ary, labelled generalization of A ⊕ B of linear
logic, and 1 is the multiplicative unit of linear logic. Operationally, 1 means the
provider has to send an end message, closing the channel and terminating the
communication session. For example, the number 6 = (110)2 would be represented
by the sequence of messages zero, one, one, $, end.
The session type does not prescribe any particular implementation only the
interface to a process. In this example, a client of a channel c : bits has to branch
on whether it receives zero, one, or $. Note that as we proceed in a session, the
type of a channel must change according to the protocol. For example, if a client
receives the message $ along c : bits then c must afterwards have type 1. The
next message along c must be end and we have to wait for that after receiving $
so the session is properly closed.
As a second example we describe the interface to a counter. As a client, we
can repeatedly send inc messages to a counter, until we want to read its value
and send val. At that point the counter will send a stream of bits representing its
value as prescribed by the type bits. From the provider’s point of view, a counter
presents an external choice, since the client chooses between inc or val.
ctr = N{inc : ctr, val : bits}
The type former N{l1 : A1 , . . . , ln : An } is an n-ary labelled generalization of ANB
of linear logic. Operationally, the provider must branch based on which of the
labels li it receives. After receiving lk along a channel c : N{l1 : A1 , . . . , ln : An },
communication along c proceeds at type Ak .
Such type formers can be arbitrarily nested to allow more complex bidirectional
protocols. Consider for example the store protocol, which is defined by the
following type.
storeA = N{ ins : A ( storeA ,
del : ⊕{none : 1, some : A ⊗ storeA }}
A provider of a channel c of type storeA either accepts an ins or del message. If
it receives the ins message, the type of c is now A ( storeA , the implication
of linear logic. It means the provider now receives a channel of type A along
c and then behaves again like a store. If it receives a del message, it responds
with either the message none or the message some (an internal choice, ⊕). If it
sends none it next must send an end message and terminate. If it sends some, the
type of c is now A ⊗ storeA . This corresponds to the multiplicative conjunction
of linear logic and, operationally, requires the provider to now send a channel of
type A and then behave again as storeA . Linearity of the type system guarantees
that the channels retrieved from a store of this type are some permutation of the
channels inserted. It may, for example, behave as a stack or a queue (as explained
in Section 7).
Modeling a binary counter. We will now describe an implementation of a counter
and use our resource-aware session types to analyze its resource usage. Like in
the rest of the paper, the resource we are interested in is the total number of
messages sent along all channels in the system.
4
e
e
e
s1
s1
s1
b1
b1
b1
s2
b0
s2
b0
s2
b1
s3
s3
inc
s3
b1
b0
b0
s4
inc
s4
s4
Fig. 1. A binary counter system representing 5 = (101)2 with the messages triggered
when an inc message is received along s4 .
A well-known example of amortized analysis counts the number of bits that
must be flipped to increment a counter. It turns out the amortized cost per
increment is 2, so that n increments require at most 2n bits to be flipped. We
can see this by introducing a potential of 1 for every bit that is 1 and using the
potential to pay for the expensive case in which and increment triggers many
flips. When the lowest bit is zero, we flip it to one (costing 1) and also store a
remaining potential of 1 with this bit. When the lowest bit is one we use the
stored potential to flip the bit back to zero (with no stored potential) and use
the remaining potential of 2 for incrementing the higher bits.
We model a binary counter as a chain of processes where each process
represents a single bit (process b0 or b1) with a final process e at the end. Each of
the processes in the chain provides a channel of the ctr type, and each (except the
last) also uses a channel of this type representing the higher bits. For example,
in the first chain in Figure 1, the process b0 offers along channel s3 and uses
channel s2 . In our notation, we would write this as
· ` e :: (s1 : ctr)
s1 : ctr ` b1 :: (s2 : ctr)
s2 : ctr ` b0 :: (s3 : ctr)
s3 : ctr ` b1 :: (s4 : ctr)
We see that, logically, parallel composition with a private shared channel corresponds to an application of the cut rule. We do not show here the definitions
of e, b0, and b1, which can be found in Figure 3. The only channel visible to
an outside client (not shown) is s4 . Figure 1 shows the messages triggered if an
increment message is received along s4 .
Expressing resource bounds. Our basic approach is that messages carry potential
and processes store potential. This means the sender has to pay not just 1 unit
for sending the message, but whatever additional units to amortize future costs.
In the amortized analysis of the counter each bit flip corresponds exactly to an
inc message, because that is what triggers a bit to be flipped. Our cost model
5
focuses on messages as prescribed by the session type and does not count other
operations, such as spawning a new process or terminating a process. This choice
is not essential to our approach, but convenient here.
To capture the informal analysis we need to express in the type that we have
to send 1 unit of potential with the label inc. We do this using a superscript
indicating the required potential with the label, postponing for now the discussion
of val.
ctr = N{inc1 : ctr, val? : bits} .
When we assign types to the processes, we now use these more expressive types.
We also indicate the potential stored in a particular process as a superscript on
the turnstile.
t : ctr `0 b0 :: (s : ctr)
1
t : ctr ` b1 :: (s : ctr)
· `0 e :: (s : ctr)
(1)
(2)
(3)
With our formal typing rules (see Section 5) we can verify these typing constraints,
using the definitions of b0, b1, and e. Informally, we can reason as follows:
b0: When b0 receives inc it receives 1 unit of potential. It continues as b1 (which
requires no communication) which stores this 1 unit (as prescribed from the
type of b1 in Equation 2).
b1: When b1 receives inc it receives 1 unit of potential which, when combined
with the stored one, makes 2 units. It needs to send an inc messages which
consumes these 2 units (1 to send the message, and 1 to send along a potential
of 1 as prescribed in the type). It has no remaining potential, which is sufficient
because it transitions to b0 which stores no potential (inferred from the type
of b0 in Equation 1).
e: When e receives inc it receives 1 unit of potential. It spawns a new process e
and continues as b1. Spawning a process is free, and e requires no potential,
so it can store the potential it received with b1 as required.
How do we handle the type annotation val? : bits of the label val? Recall that
bits = ⊕{zero : bits, one : bits, $ : 1}. In our implementation, upon receiving a
val message, a b0 or b1 process will first respond with zero or one. It then sends
val along the channel it uses (representing the higher bits of the number) and
terminates by forwarding further communication to the higher bits in the chain.
The e process will just send $ and end, indicating the empty stream of bits.
We know we will have enough potential to carry out the required send
operations if each process (b0, b1, and e) carries an additional 2 units of potential.
We could impart these with the inc and val messages by sending 2 more units
with inc and 2 units with val. That is, the following type is correct:
bits = ⊕{zero0 : bits, one0 : bits, $0 : 10 }
ctr = N{inc3 : ctr, val2 : bits}
6
e
e
e
s1
s1
s1
b1
b1
b1
s2
b0
s2
b0
s2
b0
s3
s3
s3
b1
b1
=
s4
val
s4
one
s4
val
Fig. 2. A binary counter system representing 5 = (101)2 with the messages triggered
when a val message is received along s4 .
Here, the superscript 0 in the type of bits indicates that the corresponding
messages carry no potential.
However, this type is a gross over-approximation. The processes of a counter of
value n, would carry 2n additional units of potential while only 2 dlog(n + 1)e + 2
are needed. To obtain this more precise bound, we need families of session types.
A more precise analysis. A more precise session type for this example requires
that in the type we can refer either to the number of bits in the representation
of a number or its value. This form of internal measure is needed only for typechecking purposes, not at runtime. It is also not intrinsically tied to a property of
a representation, the way the length of a list in a functional language is tied to its
memory requirements. We indicate these measures using square brackets, so that
ctr[n] is a family of types, and ctr[0], for example, is a counter with value 0. Such
type refinements have been considered in the literature on session types (see, for
example, [24]) with respect to type-checking and inference. Here, we treat it as a
meta-level notation to denote families of types. Following the reasoning above,
we obtain the following type:
bits = ⊕{zero0 : bits, one0 : bits, $0 : 10 }
ctr[n] = N{inc1 : ctr[n + 1], val2dlog(n+1)e+2 : bits}
To check the types of our implementation, we need to revisit and refine the typing
of the b0, b1 and e processes.
t : ctr[n] `0 b0 :: (s : ctr[2n])
t : ctr[n] `1 b1 :: (s : ctr[2n + 1])
`0 e :: (s : ctr[0])
Our type system verifies these types against the implementation of b0, b1, and
e (see Section 3). The typing rules reduce the well-typedness of these processes
to arithmetic inequalities which we can solve by hand, for example, using that
log(2n) = log(n) + 1. The intrinsic measure n and the particular potential
7
annotations are not automatically derived but come from our insight about the
nature of the algorithms and programs.
Before introducing the formalism in which the programs are expressed, together with the typing rules that let us perform rigorous amortized analysis of the
code (as expressed in the soundness theorem in Section 6), we again emphasize
the compositional nature of the way resource bounds are expressed in the types
themselves and in the typing judgments for process expressions. Of course, they
reveal some intensional property of the implementations, namely a bound on
its cost, so different implementations of the same plain session type may have
different resource annotations.
The typing derivation provides a proof certificate on the resource bound for a
process. For closed processes, usually typed as
· `p Q :: (c : 1)
the number p provides a worst case bound for the number of messages sent during
the computation of Q, which always ends with the process sending end along c.
3
Resource-Aware SILL
We briefly introduce the linear, process-only fragment of SILL [44,40], which
integrates functional and concurrent computation. A program in SILL is a
collection of processes exchanging messages through channels. A new process is
spawned by invoking a process definition, which also creates a fresh channel that
is provided by the new process. The process that invokes a process definition
becomes the client of the new process, communicating with it according to the
session types of the channel. The exacting nature of linear typing provides strong
guarantees, including session fidelity (a form of preservation) and absence of
deadlocks (a form of progress).
We present an overview of the session types in SILL with a brief description
of their communication protocol. They follow the type grammar below.
S, T ::= V | ⊕{li : S} | N{li : S} | S ( T | S ⊗ T | 1
V denotes a type variable here. Types may be defined mutually recursively in a
global signature, where type definitions are constrained to be contractive [21].
This allows us to treat them equi-recursively, which means we can silently replace
a type variable by its definition for the purpose of type-checking.
Internal choice S ⊕ T and external choice S N T have been generalized to n-ary
labeled sums ⊕{li : Si }i∈I and N{li : Si }i∈I (for some index set I) respectively.
As a provider of internal choice ⊕{li : Si }i∈I , a process can send one of the labels
li to its client. As a dual, a provider of external choice N{li : Si }i∈I receives one of
the labels li which is sent by its client. We require external and internal choice to
comprise at least one label, otherwise there would exist a linear channel without
observable communication along it, which is computationally uninteresting and
would complicate our proofs. The connectives ⊗ and ( are used to send channels
8
Type
Contin- Process Term
(current)
uation
(current)
c.lk ; P
c : ⊕{liqi : Si } c : Sk
case c (li ⇒ Qi )i∈I
c : N{liqi : Si } c : Sk
case c (li ⇒ Pi )i∈I
c.lk ; Q
ContinDescription
uation
P
provider sends label lk along
c with potential qk
Qk
client receives label lk along
c with potential qk
Pk
provider receives label lk
along c with potential qk
Q
client sends label lk along c
with potential qk
q
c:S⊗T
q
c:S(T
c : 1q
c:T
c:T
send c w ; P
P
y ← recv c ; Qy
[w/y]Qy
y ← recv c ; Py
[w/y]Py
provider sends channel w : S
along c with potential q
client receives channel w : S
along c with potential q
provider receives channel w :
S along c with potential q
send c w ; Q
Q
client sends channel w : S
along c with potential q
−
close c
−
provider sends end along c
with potential q
wait c ; Q
Q
client receives end along c
with potential q
Table 1. Linear resource-aware session types
via other channels. A provider of S ⊗ T sends a channel of type S to its client
and then behaves as a provider of T . A provider of S ( T receives a channel of
type S from its client. The type of the provider and client changes consistently,
and the process terms of a provider and client come in matching pairs.
Formally, the syntax of process expressions of Resource-Aware SILL is same
as in SILL.
P, Q ::= x ← X ← y ; Q | x ← y | x.lk ; P | case x (li ⇒ P )
| send x w | y ← recv x ; P | close x | wait x ; P
The first term x ← X ← y ; Q invokes a process definition X to spawn a new
process P , which uses the channels in y as a client and provides service along
a fresh channel substituted for x in Q. A forwarding process x ← y (which
provides channel x) identifies channels x and y and terminates. The effect is that
clients of x will afterwards communicate along y. We saw an example of its use in
Figure 2. The rest of the program constructs concern communication between two
processes and are guided by their corresponding session type. Table 1 provides
an overview of session types, associated process terms, and their operational
description (ignore the portions in red). For each connective in Table 1, the
first line provides the perspective of the provider, while the second line provides
that of the client. The first two columns present the type of the channel before
(current) and after (continuation) the interaction. Similarly, the next two
columns present the process terms before and after the interaction. Finally, the
last column presents the operational description.
9
1: (t : ctr[n]) `0 b0 :: (s : ctr[2n])
2: s ← b0 ← t =
3:
case s (inc ⇒ s ← b1 ← t
4:
| val ⇒ s.zero ;
5:
t.val ;
6:
s ← t)
7: (t : ctr[n]) `1 b1 :: (s : ctr[2n + 1])
8: s ← b1 ← t =
9:
case s (inc ⇒ t.inc ;
10:
s ← b0 ← t
11:
| val ⇒ s.one ;
12:
t.val ;
13:
s ← t)
(t : ctr[n]) `1 s : ctr[2n + 1]
(t : ctr[n]) `2dlog(2n+1)e+2−1 s : bits
(t : bits) `2dlog(2n+1)e+1−2dlog(n+1)e−3 s : bits
(t : bits) `0 s : bits
%
%
%
%
%
%
%
%
%
14: · `0 e :: (s : ctr[0])
15: s ← e =
16:
case s (inc ⇒ t ← e ;
17:
s ← b1 ← t
18:
| val ⇒ s.e ;
19:
close s)
(t : ctr[n + 1]) `0 s : ctr[2n + 2]
(t : ctr[n]) `2dlog(2n+2)e+2−1 s : bits
(t : bits) `2dlog(2n+2)e+1−2dlog(n+1)e−3 s : bits
(t : bits) `0 s : bits
%
(t : ctr[0]) `1 (s : ctr[1])
%
· `2dlog(0+1)e+2−1 s : 10
Fig. 3. Implementations for the b0, b1 and e processes with their type derivations
demonstrating the binary counter.
We conclude by illustrating the syntax, types and semantics of SILL using a
simple example. Recall the counter protocol (ignoring the resource annotations
in red):
bits = ⊕{zero0 : bits, one0 : bits, $0 : 10 }
ctr[n] = N{inc1 : ctr[n + 1], val2dlog(n+1)e+2 : bits}
The type prescribes that a process providing service along a channel of type ctr
will either receive an inc or a val label. If it receives an inc label, the channel will
recurse back to the ctr type. If it receives a val label, it will continue by providing
bits, sending a sequence of labels zero and one closed out with $ and end.
We present implementations of the b0, b1 and e processes respectively that
were analyzed in Section 2 in Figure 3. In the comments we show the types of
the channels after the interaction on each line (again ignoring the annotations
in red). Since the b0 process provides and external choice along s, b0 needs to
branch based on the label received (line 3). If it receives the label inc, the type
of the channel updates to ctr, as indicated on the typing in the comment. At
this point, we can spawn the b1 process since the type on line 3 matches with
the type of the b1 process (line 7). If instead b0 receives the val label along s, it
continues at type bits. It sends zero (since the lowest bit is indeed zero). It then
requests the value of the higher bits by sending val along channel t. Now both s
and t have type bits (indicated in the typing on line 4) and b0 can terminate by
forwarding further communication along s to t.
10
The b1 process operates similarly, taking care to handle the carry upon
increment by sending an inc label along t. The e process spawns a new e process
and continues as b1 upon receiving the label inc and closes the channel after
sending $ when receiving val.
4
Cost Semantics
We present an operational cost semantics for Resource-Aware SILL that tracks
the total work performed by a system. Like previous work, our semantics is a
substructural operational semantics [41] based on multiset rewriting [17] and
asynchronous communication [40]. It can be seen as a combination of an asynchronous version of a recently introduced synchronous session-type semantics [9]
with the cost tracking semantics of Concurrent C0 [42]. The technical advantage
of our semantics is that it avoids the complex operational artifacts of Silva et
al. [42] such as message buffers: processes and messages can be typed with exactly
the same typing rules, changing only the cost metric.
We will only count communication costs, ignoring internal. To this end, we
introduce 3 costs, M label , M channel and M close , for labels, channels, and close
messages, respectively. A concrete semantics can be obtained by setting appropriate values for each of those metrics. For instance, setting M label = M channel =
M close = 1 will lead to counting the total number of messages exchanged.
Our cost semantics is asynchronous, that is, processes can continue their
evaluation without wait after sending a message. In order to guarantee session
fidelity the semantics must ensure that messages are received in the order they
are sent. Intuitively, we can think of channels as FIFO message buffers, although
we will formally define them differently. Synchronous communication can be
implemented in our language in a type-safe, logically motivated manner exploiting
adjoint shift operators (see [40]).
A collection of communicating process is called a configuration. A configuration
is formally modelled as a multiset of propositions proc(c, w, P ) and msg(c, w, M ).
The predicate proc(c, w, P ) describes a process executing process expression P
and providing channel c. The predicate msg(c, w, M ) describes the message M
on channel c. In order to guarantee that messages are received in they order
they are sent, only a single message can be on a given channel c. In order for
computation to remain truly asynchronous, every send operation (except for close)
on a channel c creates not only a fresh message, but also a fresh continuation
channel c0 for the next message. This continuation channel is encoded within the
message via a forwarding operation. Remarkably, this simple device allows us
to assign session types to messages just as if they were processes! Since M need
only encode a message, it has a restricted grammar.
M ::= c ← c0 | c.lk ; c ← c0 | c.lk ; c0 ← c
| send c e ; c ← c0 | send c e ; c0 ← c | close c
The work is tracked by the local counter w in the proc(c, w, P ) and msg(c, w, M )
propositions. For a process P , w maintains the total work performed by P so far.
11
Σ(X ) = x ← X ← y = Px,y
proc(d, w, x ← X ← e ; Qx )
proc(c, 0, [c/x, e/y]Px,y )
(c fresh)
proc(d, w, [c/x]Qx )
spawnc
proc(c, w, c ← d)
fwds
msg(c, w, c ← d)
msg(c, w0 , c ← d)
proc(d, w, P )
proc(c, w + w0 , [c/d]P )
msg(c, w0 , c ← d)
proc(e, w, Pc )
proc(e, w + w0 , [d/c]Pc )
proc(c , w + M
label
0
msg(c, 0, c.lk ; c ← c0 )
, [c /c]P )
msg(c, w, c.lk ; c ← c0 )
fwd−
r
(c0 fresh)
proc(c, w, c.lk ; P )
0
fwd+
r
⊕Cs
proc(d, w0 , case c (li ⇒ Qi )i∈I )
proc(d, w + w0 , [c0 /c]Qk )
(c0 fresh)
proc(c, w, send c e ; P )
proc(c0 , w + M channel , [c0 /c]P )
msg(c, w, send c e ; c ← c0 )
msg(c, 0, send c e ; c ← c0 )
proc(d, w0 , x ← recv c ; Qx )
proc(d, w + w0 , [c0 /c]Qe )
proc(d, w, send c e ; P )
(c0 fresh)
proc(d, w + M channel , [c0 /c]P )
msg(c0 , 0, send c e ; c0 ← c)
msg(c0 , w, send c e ; c0 ← c)
proc(c, w0 , x ← recv c ; Qx )
proc(c, w + w0 , [c0 /c]Qe )
proc(c, w, close c)
msg(c, w + M close , close c)
msg(c, w, close c)
⊗Cs
⊗Cr
( Cs
( Cr
1Cs
proc(d, w0 , wait c ; Q)
proc(d, w + w0 , Q)
⊕Cr
1Cr
Fig. 4. Cost semantics tracking total work for programs in SILL
When a process sends a message (i.e. creates a new msg predicate), we increment
its counter w by the cost for sending. When a processes terminates we remove
the respective predicate from the configuration but need to preserve the work
done by the process. A process can terminate either by sending a close message,
or by forwarding. In either case, we can conveniently preserve the process’ work
in the msg predicate to pass it on to the client process.
The semantics is defined by a set of rules rewriting the configuration that
consume the proposition in the premise of the rule and produce the propositions in the conclusion (rules should be read top-down!). A step consists of
non-deterministic application of a rule whose premises matches a part of the
configuration. Consider for instance the rule Cs that describes a client that sends
12
label lk along channel c.
proc(d, w, c.lk ; P )
proc(d, w + M label , [c0 /c]P )
(c0 fresh)
msg(c0 , 0, c.lk ; c0 ← c)
NCs
The rule can be applied to every proposition of the form proc(d, w, c.lk ; P ). When
applying the rule, we generate a fresh channel continuation channel c0 and replace
the premise by propositions proc(d, w+M label , [c0 /c]P ) and msg(c0 , 0, c.lk ; c0 ← c).
The message predicate contains the process c.lk ; c0 ← c which will eventually
deliver the message to the provider along c and will continue communication
along c0 (which is achieved by c0 ← c). The work of the process is incremented
by M label to account for the sent message, while the work of the message is 0.
Conversely, the rule NCr defines how a provider receives a label lk along c.
msg(c0 , w, c.lk ; c0 ← c)
proc(c, w0 , case c (li ⇒ Qi )i∈I )
NCr
proc(c, w + w0 , [c0 /c]Qk )
The rule replaces the msg and proc propositions in the configuration that match
the premises, with the single proc proposition in the conclusion. Since the provider
receives the label lk , it continues as Qk . However, we replace c with c0 in Qk
since the forwarding c0 ← c in the message process tells us that the next message
will arrive on channel c0 . If there is any work w encoded in the message, it is
transferred to the recipient. This is somewhat more general than necessary for
this particular rule, since in the current system the work w in a label-sending
message c.lk will always be 0.
The rest of the rules of cost semantics are given in Figure 4. The rule spawnc
describes the creation of a new channel c along with a spawning new process
X implemented by Pc . This implementation is looked up in a signature for
the semantics Σ (maps process names to the implementation code). The new
process is spawned with 0 work (as it has not sent any messages so far), while Qc
continues with the same amount of work. In the rule fwds a forwarding process
creates a forwarding message and terminates. The work carried by this special
message is the same as the work done by the process, now defunct. A forwarding
message form does not carry any real information (except for the work w!); it
just serves to identify the two channels c and d. In an implementation this could
be as simple as concatenating two message buffers. We therefore do not count
forwarding messages when computing the work. Another reason forward messages
are special is that unlike all other forms of messages, they are neither prescribed
by nor manifest in a channel’s type. In our formal rules, the forwarding message
−
can be absorbed either into the client (fwd+
r ) or provider (fwdr ), in both cases
preserving the total amount of work.
The rules of the cost semantics are successively applied to a configuration
until the configuration becomes empty or the configuration is stuck and none
of the rules can be applied. At any point in this local stepping, the total work
performed by the system can be obtained by summing the local counters w for
each predicate in the configuration. We will prove in Section 6 that this total
work can be upper bounded by the initial potential of the configuration that is
typed in our resource-aware type system.
13
5
Type System
We now present the resource-aware type system of our language which extends
the linear-only fragment of SILL [44,40] with resource annotations. It is in turn
based on intuitionistic linear logic [23] with sequents of the form
A1 , A2 , . . . , An ` A
where A1 , . . . An are the linear antecedents and A is the succedent. Under the
Curry-Howard isomorphism for intuitionistic linear logic, propositions are related
to session types, proofs to processes, and cut reduction in proofs to communication.
Appealing to this correspondence, we assign a process term P to the above
judgment and label each hypothesis as well as the conclusion with a channel.
(x1 : A1 ), (x2 : A2 ), . . . , (xn : An ) ` P :: (x : A)
The resulting judgment states that process P provides a service of session type
A along channel x, using the services of session types A1 , . . . , An provided along
channels x1 , . . . , xn respectively. The assignment of a channel to the conclusion
is convenient because, unlike functions, processes do not evaluate to a value but
continue to communicate along their providing channel once they have been
created. For the judgment to be well-formed, all the channel names have to
be distinct. Whether a session type is used or provided is determined by its
positioning to the left or right, respectively, of the turnstile.
Resource-aware session types are given by the following grammar.
q
q
S, T ::= V | ⊕{liqi : S} | N{liqi : S} | S ( T | S ⊗ T | 1q
Here, V is a type variable. The meaning of the types and the process terms
associated with it are defined in Table 1 (annotations and descriptions pertaining
to potentials are marked in red).
The typing judgment of Resource-Aware SILL has the form
Σ; Ω `q P :: (x : S) .
Intuitively, the judgment describes a process in state P using the context Ω and
signature Σ and providing service along channel x of type S. In other words,
P is the provider for channel x : S, and a client for all the channels in Ω. The
resource annotation q is a natural number and defines the potential stored in the
process P . Σ defines the signature containing type and process definitions. These
definitions are needed to typecheck processes which refer to a type definition or
spawn a new process.
The signature Σ is defined as a possibly infinite set of type definitions
V = SV and process definitions x : S ← X @ q ← y : W = Px,y . The equation
V = SV is used to define the type variable V as SV . We treat such definitions
equirecursively. For instance, ctr[n] = N{inc1 : ctr[n + 1], val2dlog(n+1)e+2 : bits}
exists in the signature for all n ∈ N for the binary counter system. Type families
14
q ≥ p + rk + M label
Σ ; Ω `p P :: (x : Sk )
(k ∈ I)
q
Σ ; Ω ` (x.lk ; P ) :: (x : ⊕{liri : Si }i∈I )
q + ri ≥ q i
Σ ; Ω (x :
Σ ; Ω (x : Si ) `qi Qi :: (z : U )
⊕{liri
(∀i ∈ I)
: Si }i∈I ) `q case x (li ⇒ Qi )i∈I :: (z : U )
q+r ≥p
Σ ; Ω (y : S) `p Py :: (x : T )
r
Σ ; Ω `q (y ← recv x ; Py ) :: (x : S ( T )
r
Σ ; Ω (w : S) (x : S ( T ) `q (send x w ; Q) :: (z : U )
q ≥ p + r + M channel
Σ ; Ω `p P :: (x : T )
r
Σ ; (w : S) Ω `q send x w ; P :: (x : S ⊗ T )
r
Σ ; Ω (x : S ⊗ T ) `q y ← recv x ; Qy :: (z : U )
⊗L
Σ ; Ω `p Q :: (z : U )
q+r ≥p
Σ ; Ω (x : 1 ) `q wait x ; Q :: (z : U )
r
(V = SV ) ∈ Σ
Σ ; Ω `q P :: (x : SV )
q
Σ ; Ω ` P :: (x : V )
(V = SV ) ∈ Σ
(L
⊗R
Σ ; Ω (y : S) (x : T ) `p Qy :: (z : U )
q ≥ r + M close
1R
Σ ; · `q close x :: (x : 1r )
⊕L
(R
Σ ; Ω (x : T ) `p Q :: (z : U )
q ≥ p + r + M channel
q+r ≥p
⊕Rk
µR
Σ ; Ω (x : SV ) `q P :: (z : U )
Σ ; Ω (x : V ) `q P :: (z : U )
1L
µL
Fig. 5. Typing rules for session-typed programs (remaining rules are given in the text)
exist only at the meta-level and ctr[n] is treated as a regular type variable. The
process definition x : S ← X @ q ← y : W = Px,y defines a (possibly recursive)
process named X that is implemented by Px,y provides along channel x : S,
and uses the channels y : W as a client. The process also stores a potential q,
shown as X @ q in the signature. For instance, for the binary counter system,
s : ctr[2n] ← b0 @ 0 ← t : ctr[n] = Ps,t (Ps,t defines the implementation of b0)
exists in the signature for all n ∈ N.
Messages are typed differently from processes as their work counters w
(introduced in the predicate msg(c, w, M )) are not incremented when they actually
deliver the message to the receiver. Hence, to type the messages, we define an
auxiliary cost-free typing judgment, Σ; Ω `q P :: (x : S), which follows the same
cf
typing rules as Figure 5, but with M label = M channel = M close = 0. This avoids
paying the cost for sending a message twice. A fresh signature Σ is used in the
derivation of the cost-free judgment.
15
The idea of the type system is that each message carries potential and the
sending process pays the potential along with the cost of sending a message from
its local potential. The receiving process receives the potential when it receives
the message and adds it to its local potential. For example, consider the rule
NLk for a client sending a label lk along channel x.
q ≥ p + rk + M label Σ ; Ω (x : Sk ) `p Q :: (z : U )
NLk
Σ ; Ω (x : N{liri : Si }) `q x.lk ; Q :: (z : U )
Since the continuation Q needs potential p to typecheck, and the potential to be
sent with the label is rk , we need a total potential of at least p+rk +M label , where
M label is the cost of sending a label. Hence, we get the constraint q ≥ p+rk +M label .
The rule NR describes a provider that is awaiting a message on channel x
and has local potential q available.
q + ri ≥ qi Σ ; Ω `qi Pi :: (x : Si ) (∀i ∈ I)
NR
Σ ; Ω `q case x (li ⇒ Pi )i∈I :: (x : N{liri : Si })
The second premise tells us that the branch Pi needs potential qi to typecheck.
But the branch Pi is reached after receiving the label li with potential ri . Hence,
the initial potential q must be able to cover the difference qi − ri . Since potential
q can typecheck all the branches, we get the constraint q ≥ qi − ri for all i.
To spawn a new process defined by X , we split the context Ω into Ω1 Ω2 , and
we use Ω1 to type the newly spawned process and Ω2 for the continuation Qx .
r ≥ p + q x0 : S ← X @ p ← y 0 : W = Px0 ,y0 ∈ Σ
Ω1 = y : W Σ ; Ω2 (x : S) `q Qx :: (z : U )
spawn
Σ ; Ω1 Ω2 `r (x ← X ← y ; Qx ) :: (z : U )
If the spawned process needs potential p (indicated by the signature) and the
continuation needs potential q then the whole process needs potential r ≥ p + q.
A forwarding process x ← y terminates and its potential q is lost. Since we do
not count forwarding messages in our cost semantics, we don’t need any potential
to type the forward.
q≥0
id
Σ ; y : S `q x ← y :: (x : S)
The rest of the rules are given in Figure 5. They are similar to the discussed rules
and we omit their explanation.
As an illustration, the resource-aware type for the binary counter was presented
in Section 3 (marked in red). Also, Figure 3 provides the type derivation of the
b0, b1 and e processes (again marked in red). The annotations, along with the
type derivation, prove that an increment has an amortized resource cost of 1
(potential annotation of inc in bits type) and reading a value has a resource cost
of 2 dlog(n + 1)e + 2.
16
S0
S
Σ ; Ω C :: Ω 0
emp
0
Σ ; Ω 0 C 0 :: Ω 00
S+S 0
Σ ; (·) (·) :: (·)
Σ; Ω
compose
(C C 0 ) :: Ω 00
Σ ; Ω1 `p P :: (x : A)
p+w
Cproc
Σ ; Ω Ω1 (proc(x, w, P )) :: (Ω (x : A))
Σ ; Ω1 `p P :: (x : A)
cf
Cmsg
p+w
Σ ; Ω Ω1 (msg(x, w, P )) :: (Ω (x : A))
Fig. 6. Typing rules for a configuration
6
Soundness
This section concludes the discussion of Resource-Aware SILL by proving the
soundness of the resource-aware type system with respect to the cost semantics.
So far, we have analyzed and type-checked processes in isolation. However, as
our cost semantics indicates, processes always exist in a configuration, where
they interact with other processes. Hence, we need to extend the typing rules to
configurations.
Configuration Typing At runtime, a program state in Resource-Aware SILL
is a set of processes interacting via messages. Such a set is represented as as
a multi-set of proc and msg predicates as described in Section 4. To type the
resulting configuration C, we first need to define a well-formed signature.
A signature Σ is said to be well formed if every process definition x : S ←
X @ p ← y : W = Px,y in Σ is well typed according to the process typing
judgment, i.e. Σ ; Ω `p Px,y :: (x : S).
We use the following judgment to type a configuration.
S
Σ; Ω1 C :: Ω2
It states that Σ is well-formed and that the configuration C uses the channels in
the context Ω1 and provides the channels in the context Ω2 . The natural number
S denotes the sum of the total potential and work done by the system. We call
S the weight of the configuration.
The configuration typing judgment is defined using the rules presented in
Figure 6. The rule emp defines that an empty configuration is well-typed with
weight 0. The rule compose composes two configurations C and C 0 : C provides
service on the channels in Ω 0 while C 0 uses the channels in Ω 0 . The weight of the
composed configuration C C 0 is obtained by summing up their individual weights.
The rule Cproc creates a configuration out of a single process. The weight of this
singleton configuration is obtained by adding the potential of the process and
the work performed by it. Similarly, the rule Cmsg creates a configuration out
of a single message. Since we account for the cost while typing the processes
(see Figure 5), using the same judgment to type the processes would lead to
17
paying for the same message twice. Hence, the messages are typed in a cost-free
judgment where M label = M channel = M close = 0.
Soundness Theorem 1 is the main theorem of the paper. It is a stronger version
of a classical type preservation theorem and the usual type preservation is a
direct consequence. Intuitively, it states that the weight of a configuration never
increases during an evaluation step. This also implies that the weight of the
initial configuration is an upper bound on the weight of any configuration it can
ever step to. The soundness connects the potential with the work (i.e. the type
system with the cost semantics).
Theorem 1 (Soundness). Consider a well-typed configuration C w.r.t. a wellS
formed signature Σ such that Σ; Ω1 C :: Ω2 . If C 7→ C 0 , then there exist Ω10 , Ω20
S0
and S 0 such that Σ; Ω10 C 0 :: Ω20 and S 0 ≤ S.
The proof of the soundness theorem is achieved by a case analysis on the cost
semantics, followed by an inversion on the typing of a configuration. The complete
proof is presented in Appendix A. The preservation theorem is a corollary of the
soundness, since we prove that the configuration C 0 is well-typed.
The soundness implies that the weight of a configuration is an upper bound
on the total work performed by an evaluation starting in that configuration. We
are particularly interested in the special case of a configuration that starts with 0
work. In this case, the weight corresponds to the initial potential of the system.
S
Corollary 1 (Upper Bound). If Σ; Ω1 C :: Ω2 , and C 7→∗ C 0 , then S ≥ W 0 ,
where W 0 is the total work performed by the configuration C 0 , i.e. the sum of the
work performed by each process and message in C 0 . In particular, if the work done
by the initial configuration C is 0, then the potential P of the initial configuration
satisfies P ≥ W 0 .
Proof. Applying the Soundness theorem successively, we get that if C 7→∗ C 0 and
S
S0
Σ; Ω1 C :: Ω2 and Σ; Ω10 C 0 :: Ω20 , then S 0 ≤ S. Also, S 0 = P 0 + W 0 , where
P 0 is the total potential of C 0 , while W 0 is the total work performed so far in C 0 .
Since P 0 ≥ 0, we get that W 0 ≤ P 0 + W 0 = S 0 ≤ S. In particular, if W = 0, we
get that P = P + W = S ≥ W 0 , where P and W are the potential and work of
the initial configuration respectively.
The progress theorem of Resource-Aware SILL is a direct consequence of
progress in SILL [44]. Our cost semantics are a cost observing semantics, i.e. it is
just annotated with counters observing the work. Hence, any runtime step that
can be taken by a program in SILL can be taken in Resource-Aware SILL.
7
Case Study: Stacks and Queues
As an illustration of our type system, we present a case study on stacks and
queues. Stacks and queues have the same interaction protocol: they store elements
18
of a variable type A and support inserting and deleting elements. They only differ
in their implementation and resource usage. We express their common interface
type as the simple session type storeA .
storeA = N{ ins : A ( storeA ,
del : ⊕{none : 1, some : A ⊗ storeA }}
The session type dictates that a process providing a service of type storeA , gives
a client the choice to either insert (ins) or delete (del) an element of type A. Upon
receipt of the label ins, the providing process expects to receive a channel of type
A to be enqueued and recurses. Upon receipt of the label del, the providing process
either indicates that the queue is empty (none), in which case it terminates, or
that there is a channel stored in the queue (some), in which case it deletes this
channel, sends it to the client, and recurses.
To account for the resource cost, we need to add potential annotations
leading to two different resource-aware types for stacks and queues. Since we
are interested in counting the total number of messages exchanged, we set
M label = M channel = M close = 1 in our type system to obtain a concrete bound.
Stacks The type for stacks is defined below.
0
stackA = N{ ins0 : A ( stackA ,
0
del2 : ⊕{none0 : 10 , some0 : A ⊗ stackA }}
A stack is implemented as a sequence of elem processes terminated by an empty
process. The implementation and type derivation of elem is presented below.
1: (x:A) (t:stackA ) `0 elem :: (s : stackA )
2: s ← elem ← x t =
3:
case s
4:
(ins ⇒ y ← recv s ;
% (y:A) (x:A) (t:stackA ) `0 s : stackA
0
5:
s ← elem ← x t ;
% (y:A) (s0 : stackA ) `0 s : stackA
0
6:
s ← elem ← y s
7:
8:
9:
| del ⇒ s.some ;
send s x ;
s ← t)
0
% (x:A) (t:stackA ) `1 s : A ⊗ stackA
% t:stackA `0 s : stackA
The recursive elem process stores an element of the stack. It uses channel x : A
(element being stored) and channel t : stackA (tail of the stack) and provides
service along s : stackA . The implementation demonstrates that if the elem
process receives an ins message along s, it receives the element y (line 4), spawns
a new elem process using its original element x (line 5), and continues with
another instance of the elem process with the received element y (line 6). In this
way, it adds the element y to the head of the sequence. Otherwise, elem receives
a del message along s and responds with the some label (line 7), followed by the
channel x it stores (line 8). It then forwards all communication along s to t.
19
Inserting an element has no resource cost, since no messages are sent by
the elem process. Similarly, deleting an element has a cost of 2, which is used
to send two messages: the some label and the element x. This is reflected by
the type stackA , which needs 0 and 2 potential units for insertion and deletion,
respectively, as indicated by the resource annotations.
We now implement and type the empty process.
10: · `0 empty :: (s : stackA )
11: s ← empty =
12:
case s (ins ⇒ y ← recv s ;
13:
e ← empty ;
14:
s ← elem ← y e
15:
| del ⇒ s.none ;
16:
close s)
% (y:A) `0 s : stackA
% (y:A) (e : stackA ) `0 s : stackA
% · `1 s : 1
The sequence of elem processes ends with an empty process, providing service
along channel s where it can receive the label ins or del. If it receives the label ins,
it receives the element y to be inserted (line 12), spawns a new empty process
(line 13), and continues execution as an elem process with the received element
(line 14). On receiving the label del, it just sends the none label (line 15) followed
by the close message (line 16), indicating that the stack is empty.
Inserting an element sends no messages and thus has cost 0. Deleting an
element sends two messages and has cost 2, which is reflected in the resource
annotations of the labels in the type storeA . Note that deleting an element requires
the system to send back two messages, either the none label followed by the close
message, or the some label followed by the element. Therefore, an implementation
of stacks will have a resource cost of at least 2 for deletion. This shows that
the above implementation is the most efficient w.r.t. our cost semantics because
insertion has no resource cost, and deletion has the least possible cost.
Queues Next, we consider the queue interface which is achieved by using the
same storeA interface and annotating it with a different potential. The tight
potential bound depends on the number of elements stored in the queue. Hence,
a precise resource-aware type needs access to this internal measure in the type. A
type queueA [n] intuitively defines a queue of size n, i.e. a process offering along
a channel of type queueA [n] connects a sequence of n elements.
0
queueA [n] = N{ ins2n : A ( queueA [n + 1],
0
del2 : ⊕{none0 : 10 , some0 : A ⊗ queueA [n − 1]}}
Similar to a stack, a queue is also implemented by a sequence of elem processes,
connected via channels, and terminated by the empty process. We show the
implementation of elem below.
1: (x:A) (t:queueA [n − 1]) `0 elem :: (s : queueA [n])
2: s ← elem ← x t =
20
3:
4:
5:
6:
7:
8:
9:
10:
case s (
ins ⇒ y ← recv s ; % (y:A) (x:A) (t:queueA [n − 1]) `2n s:queueA [n + 1]
0
t.ins ;
% (y:A)(x:A)(t:A ( queueA [n]) `1 s:queueA [n + 1]
send t y ;
% (x : A) (t : queueA [n]) `0 s : queueA [n + 1]
s ← elem ← x t
| del ⇒ s.some ;
send s x ;
s ← t)
0
% (x:A)(t:queueA [n − 1]) `1 s:A ⊗ queueA [n − 1]
% t:queueA [n − 1] `0 s : queueA [n − 1]
Similar to the implementation of a stack, the elem process provides along s :
queueA , stores the element x : A, and uses the tail of the queue t : queueA . When
the elem process receives the ins message along s, it receives the element y (line 4),
and passes the ins message (line 5) along with y (line 5) to t. Since the process at
the other end of t is also implemented using elem, it passes along the element to
its tail too. Thus the element travels to the end of the queue where it is finally
inserted. The deletion is similar to that for stack.
For each insertion, the ins label along with the element travels to the end of
the queue. Hence, the resource cost of each insertion is 2n where n is the size of
the queue and this is reflected in the type queueA in the potential annotation of
ins as 2n. Similar to the stack, deletion has a resource cost of 2 to get back the
some label and the element. We now consider the empty process.
· `0 empty :: (s : queueA [0])
s ← empty =
case s (
ins ⇒ y ← recv s ;
e ← empty ;
s ← elem ← y e
| del ⇒ s.none ;
close s)
% (y:A) `0 s : queueA [1]
% (y:A) (e : queueA [0]) `0 s : queueA [1]
% · `1 s : 1
The implementation of empty process is identical to that of stacks. Since insertion
does not cause the process to send any messages, its resource cost is 0. On the
other hand, deletion costs 2 units because the process sends back the none label
followed by the close message. This is correctly reflected in the queue type. Since
s : queueA [0], the annotation for ins is 2n = 2 · 0 = 0. Similarly, del is annotated
with a potential of 2.
The resource-aware types show that the implementation for stacks is more
efficient than that of queues. This follows from the potential annotation. The
label ins is annotated by 2n for stackA and with 0 for queueA . The label del has
the same annotation in both types.
Functional queues In a functional language, a queue is often implemented with
two lists. The idea is to enqueue into the first list and to dequeue from the second
list. If the second list is empty then we copy the first list over, thereby reversing
21
its order. Since the cost of the dequeue operation varies drastically between the
dequeue operations, amortized analysis is again instrumental in the analysis of
the worst-case behavior and shows that the worst-case amortized cost for deletion
is actually a constant.
Appendix B.3 contains an implementation of a functional queue in ResourceAware SILL. The type of the queue is
0
queue0A = N{ ins6 : A ( queue0A ,
0
del2 : ⊕{ some0 : A ⊗ queue0A , none0 : 10 }}
Resource-aware session types enable us to translate the amortized analysis to
the distributed setting. The type prescribes that an insertion has an amortized
cost of 6 while the deletion has an amortized cost of 2. The main idea here is
that the elements are inserted with a constant potential in the first list. While
deleting, if the second list is empty, then this stored potential in the first list is
used to pay for copying the elements over to the second list. The exact potential
annotations for the two lists can be found in Appendix B.3. As demonstrated
from the resource-aware type, this implementation is more efficient than the
previous queue implementation, which has a linear resource cost for insertion.
Generic clients The notion of efficiency of a store can be generalized and quantified
by considering clients for the stack and queue interface. A client interacts with
a generic store via a sequence of insertions and deletions. A provider can then
implement the store as a stack, queue, priority queue, etc. (same interface) and
just expose the resource-aware type for storeA . Our type system can then use
just the interface type and the generic client implementation to derive resource
bounds on the client. For simplicity, the clients are typed in an affine type system
which allows us to throw away dummy channels (see below).
We provide a general mechanism for implementing clients for a generic store.
We define a generic storeA type at which the potential annotations are arbitrary
natural numbers.
a
storeA [n] = N{ insi : A ( storeA [n + 1],
t
deld : ⊕{nonep : 1e , somes : A ⊗ storeA [n − 1]}}
A client is defined by a list ` of ins and del messages that it sends to the store.
We index the client C`,n using `, and the internal measure n of the storeA type.
The channel along which the client provides is irrelevant for our analysis and is
represented using a dummy channel d : D. For ease of notation, we define the
potential needed for a client C`,n as a function φ(`, n).
We implement the client C`,n as follows. First, consider the case when ` = [],
i.e. an empty list.
· `0 C[],n :: (d : D)
d ← C[],n = close d
22
The client for an empty list just closes the channel d. We assume that all clients
are typed with the cost-free metric to only count for the messages sent inside
the stores. So C[],0 needs 0 potential. For the potential function, this means
φ([], n) = 0.
Next, we implement the client when the head of the list ` is ins.
Ω (x : A) (s : storeA [n]) `q Cins::`,n :: (d : D)
d ← Cins::`,n ← Ω x s =
s.ins ;
send s x ;
d ← C`,n+1 ← Ω s
%
%
a
Ω (x : A) (s : A ( storeA [n + 1]) `q−i d : D
Ω (s : storeA [n + 1]) `q−i−a d : D
The client sends an ins label followed by the element x. If Cins::`,n needs a potential
q, then the type derivation informs us that C`,n+1 needs a potential q − i − a.
Thus, φ(ins :: `, n) = φ(`, n+1)+i+a. Finally, we show the client implementation
if the head of the list ` is del.
Ω (s : storeA [n]) `q Cdel::` :: (d : D)
d ← Cdel::`,n ← Ω s =
t
s.del ;
% Ω (s : ⊕{nonen : 1e , somes : A ⊗ storeA [n − 1]}) `q−d d : D
case s (
some ⇒ x ← recv s ;
% Ω (x : A) (s : storeA [n − 1]) `q−d+s+t d : D
d ← C`,n−1 ← Ω x s
| none ⇒ wait s)
% Ω `q−d+p+e d : D
The client sends the del label and then case analyzes on the label it receives. If it
receives the some label, it receives the element and then continues with C`,n−1 ,
else it receives the none label and waits for the channel s to close. In terms of
the potential function, this means
(
φ(`, n − 1) + d − s − t
if n > 0
φ(del :: `, n) =
max(0, d − p − e)
otherwise
Walking through the list ` and chaining the potential equations together, we can
achieve a resource bound on the client C`,n by computing φ(`, n).
The stackA and queueA interface types are specific instantiations of the storeA
type. For the stack interface, plugging in appropriate potential annotations
i = a = p = e = s = t = 0, and d = 2, we get (ignoring the case where the stack
becomes empty)
φ([], n) = 0
φ(ins :: `, n) = φ(`, n + 1)
φ(del :: `, n) = φ(`, n − 1) + 2
Similarly, considering the queue type as another instantiation of the storeA type,
and plugging a = p = e = s = t = 0, i = 2n and d = 2, we get
φ([], n) = 0
φ(ins :: `, n) = φ(`, n + 1) + 2n
23
φ(del :: `, n) = φ(`, n − 1) + 2
Finally, looking at queue0A as another instantiation of the storeA type and plugging
a = p = e = s = t = 0, i = 6 and d = 2, we get
φ([], n) = 0
φ(ins :: `, n) = φ(`, n + 1) + 6
φ(del :: `, n) = φ(`, n − 1) + 2
This allows us to compare arbitrary clients of two (same or different) interfaces
and compare their resource cost. The resource-aware types are expressive enough
to obtain these resource bounds without referring the implementation of the store
interface. For instance, an important property of queues is that every insertion is
more costly than the previous one. The cost of insertion depends on the size of
the queue, which, in turn, increases with every insertion. Hence, the complexity
of the queue system depends on the sequence in which inserts and deletes are
performed. In particular, we can consider the efficiency of two different clients
for the queue system, by solving the above system of equations.
For instance, consider two clients Q`1 ,n and Q`2 ,n , with two different message
lists `1 = [ins, . . . , ins, del, . . . , del], i.e. m insertions followed by m deletions, and
`2 = [ins, del, ins, del, . . . , ins, del], i.e. m instances of alternate insertions and
deletions. In both cases, we have the same number of insertions and deletions.
However, the resource cost of the two systems are completely different. Solving
the system of equations, we get that φ(`1 , n) = 2mn + m(m − 1) + 2m, while
φ(`2 , n) = 2m(n + 1), which shows that the second client is an order of magnitude
more efficient than the first one. More examples are presented in Appendix B.
8
Related Work
Session types have been introduced by Honda [32,33]. The technical development
in this article is based on previous work on [44,40]. By removing the potential
annotation from the type rules in Section 5 we arrive at the type system of loc.
cit. The internal measures and type families that we use are inspired by [24].
Other recent innovations in session types include sharing of resources [9] and
dynamic monitors [34]. In contrast to our work, all the aforementioned articles
do not discuss static resource analysis.
Static resource bound analysis for sequential programs has been extensively
studied. Successful approaches are based on refinement types [18,15], linear
dependent types [37], abstract interpretation [46,25,5,16], deriving and solving recurrence relations [19,20,4,36], term rewriting [11,8]. These works do not consider
message-passing programs nor concurrent or parallel evaluation.
Our work is based on type-based amortized resource analysis. Automatic
amortized resource analysis (AARA) has been introduced as a type system
to automatically derive linear [30] and polynomial bounds [27] for sequential
functional programs. It can also be integrated with program logics to derive
bounds for imperative programs [7,14]. Moreover, it has been used to derive
bounds for term-rewrite systems [31] and object-oriented programs [29]. A recent
work also considers bounds on the parallel evaluation cost (also called span) of
functional programs [28]. The innovation of our work is the integration of AARA
and session types and the analysis of message-passing programs that communicate
24
with the outside world. Instead of function arguments, our bounds depend on the
messages that are send along channels. As a result, the formulation and proof of
the soundness theorem is quite different from the soundness of sequential AARA.
We are only aware of a couple of other works that study resource bounds for
concurrent programs. Gimenez et al. [22] introduced a technique for analyzing the
parallel and sequential space and time cost of evaluating interaction nets. While
it also based on linear logic and potential annotations, the flavor of the analysis
is quite different. Interaction nets are mainly used to model parallel evaluation
while session types focus on the interaction of processes. A main innovation of our
work is that processes can exchange potential via messages. It is not clear how we
can represent the examples we consider in this article as interaction nets. Albert
et al. [3,2] have studied techniques for deriving bounds on the cost of concurrent
programs that are based on the actor model. While the goals of the work are
similar to ours, the used technique and considered examples are dissimilar. A
major difference is that our method is type-based and compositional. A unique
feature of our work is that types describe bounds as functions of the messages
that are sent along a channel.
9
Conclusion
We have introduced resource-aware session types, a linear type system that
combines session types [32,40] and type-based amortized resource analysis [30,27]
to reason about the resource usage of message-passing processes. The soundness of
the type system has been proved for a core session-typed language with respect to
a cost semantics that tracks the total communication cost in a system of processes.
We have demonstrated that our technique can be used to prove tight resource
bounds and supports amortized reasoning by analyzing standard session-type
data structures such as distributed binary counters, stacks, and queues.
Our approach addresses some of the main challenges of analyzing messagepassing programs such as compositionality and description of symbolic bounds.
However, there are several open problems that we plan to tackle as part of future
work. The technique we have developed in this paper does not yet account for
the concurrent execution cost of processes, or the span. We are working on a
companion paper that describes a type-based analysis to derive bounds on the
span; the earliest time a concurrent computation terminates assuming an infinite
number of processors. Due to data dependencies in a concurrent program, a
process needs to wait for messages from other processes, and computing these
waiting times statically makes span analysis challenging.
Similarly, we have focused on the foundations and meta theory of resourceaware session types in this paper. The next step is to implement our analysis. An
advantage of our method is that it is based upon type-based amortized resource
analysis for sequential programs. We will integrate the type system with SILL
for functional programs [27]. We designed the type system with automation in
mind and we are confident that we can support automatic type inference using
templates and LP solving similar to AARA [30,27]. To this end, we are working
on an algorithmic version of the declarative type system presented here.
25
References
1. Acar, U.A., Charguéraud, A., Rainey, M.: Oracle-guided scheduling for controlling granularity in implicitly parallel languages. J. Funct. Program. 26 (2016)
2. Albert, E., Arenas, P., Correas, J., Genaim, S., Gómez-Zamalloa, M., Martin-Martin,
E., Puebla, G., Román-Díez, G.: Resource analysis: From sequential to concurrent and
distributed programs. In: Formal Methods - 20th International Symposium (FM’15) (2015)
3. Albert, E., Flores-Montoya, A., Genaim, S., Martin-Martin, E.: May-Happen-in-Parallel
Analysis for Actor-Based Concurrency. ACM Trans. Comput. Log. 17(2) (2016)
4. Albert, E., Genaim, S., Masud, A.N.: On the Inference of Resource Usage Upper and Lower
Bounds. ACM Transactions on Computational Logic 14(3) (2013)
5. Alias, C., Darte, A., Feautrier, P., Gonnord, L.: Multi-dimensional Rankings, Program
Termination, and Complexity Bounds of Flowchart Programs. In: (SAS’10) (2010)
6. Antonopoulos, T., Gazzillo, P., Hicks, M., Koskinen, E., Terauchi, T., Wei, S.: Decomposition
Instead of Self-composition for Proving the Absence of Timing Channels. In: (PLDI’17)
7. Atkey, R.: Amortised Resource Analysis with Separation Logic. In: (ESOP’10)
8. Avanzini, M., Lago, U.D., Moser, G.: Analysing the Complexity of Functional Programs:
Higher-Order Meets First-Order. In: 29th Int. Conf. on Func. Prog. (ICFP’15)
9. Balzer, S., Pfenning, F.: Manifest sharing with session types (ICFP) (Aug 2017)
10. Brent, R.P.: Algorithms for minimization without derivatives. Courier Corporation (2013)
11. Brockschmidt, M., Emmes, F., Falke, S., Fuhs, C., Giesl, J.: Alternating Runtime and Size
Complexity Analysis of Integer Programs. In: (TACAS’14) (2014)
12. Caires, L., Pfenning, F.: Session types as intuitionistic linear propositions. In: Proc. of the
21st Int. Conf. on Concurrency Theory (CONCUR 2010). pp. 222–236 (Aug 2010)
13. Caires, L., Pfenning, F., Toninho, B.: Linear logic propositions as session types. Mathematical Structures in Computer Science 26(3), 367–423 (2016)
14. Carbonneaux, Q., Hoffmann, J., Reps, T., Shao, Z.: Automated Resource Analysis with
Coq Proof Objects. In: 29th Int. Conf. on Computer-Aided Verification (CAV’17) (2017)
15. Çiçek, E., Barthe, G., Gaboardi, M., Garg, D., Hoffmann, J.: Relational Cost Analysis. In:
44th Symposium on Principles of Programming Languages (POPL’17) (2017)
16. Cerný, P., Henzinger, T.A., Kovács, L., Radhakrishna, A., Zwirchmayr, J.: Segment
Abstraction for Worst-Case Execution Time Analysis. In: (ESOP’15) (2015)
17. Cervesato, I., Scedrov, A.: Relating state-based and process-based concurrency through
linear logic. Electron. Notes Theor. Comput. Sci. 165, 145–176 (Nov 2006)
18. Çiçek, E., Garg, D., Acar, U.A.: Refinement Types for Incremental Computational Complexity. In: 24th European Symposium on Programming (ESOP’15) (2015)
19. Danner, N., Licata, D.R., Ramyaa, R.: Denotational Cost Semantics for Functional Languages with Inductive Types. In: 29th Int. Conf. on Functional Programming (ICFP’15)
20. Flores-Montoya, A., Hähnle, R.: Resource Analysis of Complex Programs with Cost
Equations. In: Programming Languages and Systems - 12th Asian Symposiu (APLAS’14)
21. Gay, S.J., Hole, M.: Subtyping for session types in the π-calculus. Acta Informatica (2005)
22. Gimenez, S., Moser, G.: The complexity of interaction. In: (POPL’16) (2016)
23. Girard, J.Y.: Linear logic. Theoretical Computer Science 50(1), 1 – 101 (1987)
24. Griffith, D., Gunter, E.L.: Liquid pi: Inferrable dependent session types. In: Proceedings of
the NASA Formal Methods Symposium. pp. 186–197. Springer LNCS 7871 (2013)
25. Gulwani, S., Mehra, K.K., Chilimbi, T.M.: SPEED: Precise and Efficient Static Estimation
of Program Computational Complexity. In: 36th Principles of Prog. Langs. (POPL’09)
26
26. Haemmerlé, R., López-García, P., Liqat, U., Klemen, M., Gallagher, J.P., Hermenegildo,
M.V.: A Transformational Approach to Parametric Accumulated-Cost Static Profiling. In:
Functional and Logic Programming - 13th International Symposium (FLOPS’16) (2016)
27. Hoffmann, J., Das, A., Weng, S.C.: Towards Automatic Resource Bound Analysis for
OCaml. In: 44th Symposium on Principles of Programming Languages (POPL’17) (2017)
28. Hoffmann, J., Shao, Z.: Automatic Static Cost Analysis for Parallel Programs. In: 24th
European Symposium on Programming (ESOP’15) (2015)
29. Hofmann, M., Jost, S.: Type-Based Amortised Heap-Space Analysis. In: (ESOP’06)
30. Hofmann, M., Jost, S.: Static Prediction of Heap Space Usage for First-Order Functional
Programs. In: 30th ACM Symp. on Principles of Prog. Langs. (POPL’03) (2003)
31. Hofmann, M., Moser, G.: Multivariate Amortised Resource Analysis for Term Rewrite
Systems. In: 13th Int. Conf. on Typed Lambda Calculi and Appl. (TLCA’15) (2015)
32. Honda, K.: Types for dyadic interaction, pp. 509–523. Springer Berlin Heidelberg (1993)
33. Honda, K., Vasconcelos, V.T., Kubo, M.: Language primitives and type discipline for
structured communication-based programming. ESOP’98, Springer LNCS 1381
34. Jia, L., Gommerstadt, H., Pfenning, F.: Monitors and blame assignment for higher-order
session types. In: Proc. of the 43rd ACM Symp. on Prin. of Prog. Lang. POPL ’16
35. Jia Chen, Y.F., Dillig, I.: Precise Detection of Side-Channel Vulnerabilities using Quantitative Cartesian Hoare Logic. In: Conf. on Comp. and Comm. Security (CCS’17)
36. Kincaid, Z., Breck, J., Boroujeni, A.F., Reps, T.: Compositional recurrence analysis revisited.
In: Conference on Programming Language Design and Implementation (PLDI’17) (2017)
37. Lago, U.D., Petit, B.: The Geometry of Types. In: (POPL’13) (2013)
38. Ngo, V.C., Dehesa-Azuara, M., Fredrikson, M., Hoffmann, J.: Verifying and Synthesizing
Constant-Resource Implementations with Types. In: 38th IEEE Symp. on S&P (2017)
39. Olivo, O., Dillig, I., Lin, C.: Static Detection of Asymptotic Performance Bugs in Collection
Traversals. In: Conf. on Prog. Lang. Design and Impl. (PLDI’15) (2015)
40. Pfenning, F., Griffith, D.: Polarized Substructural Session Types. Springer Berlin (2015)
41. Pfenning, F., Simmons, R.J.: Substructural operational semantics as ordered logic programming. In: Proc. of 24th IEEE Symp. on Logic In Comp. Sc. LICS ’09 (2009)
42. Silva, M., Florido, M., Pfenning, F.: Non-blocking concurrent imperative programming
with session types. In: Proc. 4th Int. Workshop on Linearity, LINEARITY 2016. (2016)
43. Tarjan, R.E.: Amortized Computational Complexity. SIAM J. Alg. Disc. Methods (1985)
44. Toninho, B., Caires, L., Pfenning, F.: Higher-order processes, functions, and sessions: A
monadic integration. In: (ESOP’13). Springer LNCS 7792, Rome, Italy (Mar 2013)
45. Wadler, P.: Propositions as sessions. In: (ICFP’12). ACM Press, Copenhagen, Denmark
46. Zuleger, F., Sinn, M., Gulwani, S., Veith, H.: Bound Analysis of Imperative Programs with
the Size-change Abstraction. In: 18th Int. Static Analysis Symp. (SAS’11) (2011)
A
Proof of Soundness Theorem
We present the complete proof of the soundness theorem. We reiterate the rules
of cost semantics and typing. Figure 4 presents the cost semantics for our sessiontyped language, while Figure 5 presents the full set of typing rules. Finally,
Theorem 1 defines the soundness theorem establishing that the typing rules for a
configuration presented in Figure 6 are sound w.r.t. the rules for cost semantics
presented in Figure 4. We follow the complete proof of the soundness theorem.
27
Proof. The proof proceeds by case analysis on the cost semantics of our language,
i.e. on the judgment C 7→ C 0 . By the compose rule, we can split the configuration
0
0
such that C = (CM D) and C 0 = (CM
D) and CM 7→ CM
. Using the compose rule,
SM
Σ ; Ω CM :: Ω 0
SM +SD
Σ; Ω
SM
0
Σ ; Ω CM
:: Ω 0
0
SM
+SD
Σ; Ω
SD
Σ ; Ω 0 D :: Ω 00
(CM D) :: Ω
compose
00
SD
Σ ; Ω 0 D :: Ω 00
compose0
0
(CM
D) :: Ω 00
0
Hence, to show that S 0 ≤ S, it suffices to show that SM
≤ SM . We proceed by
0
case analysis on the CM 7→ CM judgment.
– Case (spawnc ) : CM = proc(d, w, x ← Px ← y ; Qx ). Inverting the typing
rule spawn on configuration CM , we get
r ≥p+q
Ω1 Ω2 `r x ← Px ← y ; Qx :: (d : U )
(4)
0
and in CM
= proc(c, 0, Pc ) proc(d, w, Qc ), we get (the premise due to inversion)
Ω1 `p Pc :: (c : S)
Ω2 (c : S) `q Qc :: (d : U )
From the cost semantics rule spawnc ,
proc(d, w, x ← Px ← y ; Qx )
spawnc
proc(c, 0, Pc )
proc(d, w, Qc )
Since S is the sum of work and potential of each process in the configuration,
0
we get SM = r + w, while SM
= (pP + wP ) + (pQ + wQ ) = (p + 0) + (q + w) =
p + q + w ≤ r + w = SM since p + q ≤ r (by Equation 4).
– Case (fwds ) : CM = proc(c, w, c ← d). Inverting the typing rule fwd on CM ,
q≥0
d : S `q c ← d :: (c : S)
0
Inverting the same rule in CM
,
q0 ≥ 0
0
d : S `q c ← d :: (c : S)
cf
Using the semantics rule fwds ,
proc(c, w, c ← d)
fwds
msg(c, w, c ← d)
0
Since we are free to choose q 0 , we set q 0 ≤ q. Computing SM and SM
, we get
0
SM
= q0 + w
≤q+w
= SM
28
0
– Case (fwd+
r ) : CM = proc(d, w, P ) msg(c, w , c ← d). Inverting the fwd rule
for CM ,
q1 ≥ 0
Ω `q1 P :: (d : S)
(5)
q2 ≥ 0
d : S `q2 c ← d :: (c : S)
Using Equation 5 and noting that c and d have the same type S, we get for
0
CM
,
0
q10 ≥ 0
Ω `q1 [c/d]P :: (c : S)
From the cost semantics rule fwd+
r , we get
proc(d, w, P )
msg(c, w0 , c ← d)
fwd+
r
proc(c, w + w0 , [c/d]P )
0
Since we are free to choose q10 , we set q10 ≤ q1 . Computing SM and SM
, we
get
0
SM
= q10 + w + w0
≤ q1 + q2 + w + w0
= (q1 + w) + (q2 + w0 )
= SM
0
– Case (fwd−
r ) : CM = proc(e, w, P ) msg(c, w , c ← d). Inverting the fwd rule
for CM ,
q1 ≥ 0
Ω (c : S) `q1 P :: (e : U )
(6)
q2 ≥ 0
d : S `q2 c ← d :: (c : S)
Using Equation 6 and noting that c and d have the same type S, we get for
0
CM
,
0
q10 ≥ 0
Ω (d : S) `q1 [d/c]P :: (e : U )
From the cost semantics rule fwd−
r , we get
proc(e, w, P )
msg(c, w0 , c ← d)
fwd−
r
proc(e, w + w0 , [d/c]P )
Since we are free to choose q10 , we set q10 ≤ q1 .
0
SM
= q10 + w + w0
≤ q1 + q2 + w + w0
= (q1 + w) + (q2 + w0 )
= SM
– Case (⊕Cs ) : CM = proc(c, w, c.lk ; P ). Inverting the typing rule ⊕Rk on
CM , we get
q1 ≥ p + rk + M label
Ω `q1 (c.lk ; P ) :: (c : ⊕{liri : Si }i∈I )
29
(7)
0
and in CM
, we get (the premise due to inversion)
Ω `p [c0 /c]P :: (c0 : Sk )
0
c0 : Sk `q1 (c.lk ; c ← c0 ) :: (c : ⊕{liri : Si }i∈I )
q10 ≥ q2 + rk
q2
0
q2 ≥ 0
0
c : Sk ` c ← c :: (c : Sk )
(8)
(9)
Using the cost semantics rule ⊕Cs , we get
proc(c, w, c.lk ; P )
0
proc(c , w + M
label
, [c0 /c]P )
msg(c, 0, c.lk ; c ← c0 )
⊕Cs
0
Again, SM
= (p + w + M label ) + (q10 + 0). Since, we need to prove that
0
there exists such an SM
, we can choose q10 and q2 arbitrarily such that
0
they satisfy Equations 8 and 9. We set q2 = 0 and q10 = rk . Hence, SM
=
label
(p + w + M
) + rk ≤ q1 + w ≤ SM (by Equation 7)
– Case (⊕Cr ) : CM = msg(c, w, c.lk ; c ← c0 ) proc(d, w0 , case c (li ⇒ Qi )i∈I ).
Inverting the typing rule ⊕L on CM , we get
q1 ≥ q2 + rk
c0 : Sk `q1 (c.lk ; c ← c0 ) :: (c : ⊕{liri : Si }i∈I )
q2 ≥ 0
q3 + rk ≥ qk
c0 : Sk `q2 c ← c0 :: (c : Sk )
(10)
(11)
Ω (c : ⊕{liri : Si }i∈I ) `q3 case c (li ⇒ Qi )i∈I :: (d : U ) (12)
0
and in CM
(premise of the typing rule due to inversion), we get
Ω (c0 : Sk ) `qk [c0 /c]Qk :: (d : U )
Using the cost semantics rule ⊕Cr , we get
msg(c, w, c.lk ; c ← c0 )
proc(d, w0 , case c (li ⇒ Qi )i∈I )
⊕Cr
proc(d, w + w0 , [c0 /c]Qk )
–
–
–
–
–
0
Again, SM
= qk + w + w0 ≤ q3 + rk + w + w0 ≤ q3 + q2 + rk + w + w0 ≤
q3 + q1 + w + w0 ≤ (q1 + w) + (q3 + w0 ) = SM (by Equations 10, 11 and 12).
Case (NCs ) : Analogous to ( Cs .
Case (NCr ) : Analogous to ( Cr .
Case (⊗Cs ) : Analogous to ⊕Cs .
Case (⊗Cr ) : Analogous to ⊕Cr .
Case (( Cs ) : CM = proc(d, w, send c e ; P ). Applying the rule ( L on CM ,
we get
q1 ≥ p + r + M channel
r
Ω (e : S) (c : S ( T ) `q1 (send c e ; P ) :: (d : U )
0
Inverting the same rule on CM
, we get
Ω (c0 : T ) `p [c0 /c]P :: (d : U )
30
q10 ≥ q2 + r
r
0
(e : S) (c : S ( T ) `q1 send c e ; c0 ← c :: (c0 : T )
q2 ≥ 0
c : T `q2 c0 ← c :: (c0 : T )
From the cost semantics rule ( Cs , we get
proc(d, w, send c e ; P )
proc(d, w + M
channel
, [c0 /c]P )
msg(c0 , 0, send c e ; c0 ← c)
( Cs
Since we can choose arbitrary values for q10 and q2 satisfying the above
0
inequalities, we set q2 = 0 and q10 = r. Computing SM and SM
, we get
0
SM
= (p + w + M channel ) + (q10 + w0 )
= (p + r + M channel + w) + (q10 − r + w0 )
≤ q1 + w + w0 )
= SM
– Case (( Cr ) : CM = msg(c0 , w, send c e ; c0 ← c) proc(c, w0 , x ← recv c ; Qx ).
Applying the rule ( L on the message in CM , we get
q1 ≥ q2 + r
r
(e : S) (c : S ( T ) `q1 send c e ; c0 ← c :: (c0 : T )
q2 ≥ 0
c : T `q2 c0 ← c :: (c0 : T )
Applying the rule ( R on the process in CM , we get
q3 + r ≥ p
r
Ω `q3 (x ← recv c ; Qx ) :: (x : S ( T )
0
Inverting the ( R rule, we get for CM
,
Ω (e : S) `p [c0 /c]Qe :: (c0 : T )
From the cost semantics rule ( Cr , we get
msg(c0 , w, send c e ; c0 ← c)
proc(c, w0 , x ← recv c ; Qx )
( Cr
0
proc(c, w + w , [c0 /c]Qe )
0
Computing SM and SM
, we get
0
SM
= p + w + w0
≤ q3 + r + w + w0
≤ q3 + q2 + r + w + w0
≤ q3 + q1 + w + w0
= (q1 + w) + (q3 + w0 )
= SM
31
– Case (1Cs ) : CM = proc(c, w, close c). Applying the 1R rule on CM , we get
· `q close c :: (c : 1r )
q ≥ r + M close
0
Inverting the same rule for CM
, we get
0
· `q close c :: (c : 1r )
q0 ≥ r
From the cost semantics rule 1Cs , we get
proc(c, w, close c)
msg(c, w + M close , close c)
1Cs
0
Setting q 0 = r and computing SM and SM
, we get
0
SM
= q 0 + w + M close
= r + w + M close
≤q+w
= SM
– Case (1Cr ) : CM = msg(c, w, close c) proc(d, w0 , wait c ; Q). Applying the
rule 1R on the message in CM , we get
· `q1 close c :: (c : 1r )
q1 ≥ r
Applying the 1L rule on the process in CM , we get
q2 + r ≥ p
Ω (c : 1r ) `q2 wait c ; Q :: (d : U )
0
Inverting the 1L rule for CM
, we get
Ω `p Q :: (d : U )
From the cost semantics rule 1Cr , we get
msg(c, w, close c)
proc(d, w0 , wait c ; Q)
1Cr
proc(d, w + w0 , Q)
0
Computing SM and SM
, we get
0
SM
= p + w + w0 + M close
≤ q2 + r + w + w0
≤ q2 + q1 + w + w0
≤ (q1 + w) + (q2 + w0 )
= SM
0
Hence, in all of the above cases, SM
≤ SM establishing that S 0 ≤ S, thus showing
that the potential type system is sound w.r.t. the cost semantics.
32
B
More Examples
Our type system is quite expressive and can be used to derive bounds on many
more examples. In this section, we will derive bounds on several list processes.
We will start with simple examples, such as the nil, cons and append processes.
We will then derive bounds on stacks and queues being implemented using lists.
Finally, we will conclude with some higher order functions such as map and f old.
For each of the following examples, we assume the standard cost metric, where
we count the number of messages exchanged, i.e. M label = M channel = M close = 1.
First, we consider the list protocol as a simple session type.
listA = ⊕{ cons : A ⊗ listA ,
nil : 1}
The type prescribes that a process providing service of type listA will either send
a label cons followed by an element of type A and recurse, or will send a nil
label followed by a close message and then terminate. On the client side (i.e. a
process that uses a channel of type listA in its context), the opposite behavior is
observed, i.e. a client receives the messages that the provider sends (sequence of
cons labels and elements terminated by a nil label and the close messsage). The
append process is a SILL implementation of the standard append function which
appends two lists.
B.1
Basic Processes
We present the implementations of nil, cons and append processes.
nil
The nil process is used to create an empty list. Formally, a nil process uses an
empty context, and provides an empty list along a channel l : listA . Concretely,
this means it sends a nil label followed by a close message along l. First, we
introduce the resource-aware session type for listA .
listA = ⊕{ nil0 : 10 ,
0
cons0 : A ⊗ listA }
This resource-aware type decorates each label and type operator with 0 potential, implying that none of the messages carry any potential, and the process
potential needs to pay only for the cost of sending the messages. We present the
implementation followed by the type derivation for the nil process.
· `2 nil :: (l : listA )
l ← nil =
l.nil ;
%
close l
%
· `1 l : 10
· `0 ·
33
The type of nil process shows that the process potential needed is 2, which
intuitively agrees with our cost model. The process sends two messages, each of
them costing unit potential. We explain the type derivation briefly. The initial
type of the process is · `2 l : listA . Now, for l to behave as an empty list, the
nil process needs to send the nil label first. As the listA type prescribes, the nil
label carries no potential, this send only costs 1. Updating the type of l and the
process potential, we get · `1 l : 10 . Finally, the nil process needs to send the
close message, which again costs 1 as the type 1 carries no potential in the type
definition of listA . Thus, our type system successully verifies that the nil process
needs a potential of 2, hence its resource usage is 2.
cons
Now, let’s look at the cons process.
(x : A) (t : list0A ) `2 cons :: (l : list0A )
l ← cons ← x t =
l.cons ;
send l x ;
l←t
0
% (x : A) (t : list0A ) `1 l : A ⊗ list0A
% (t : list0A ) `0 l : list0A
We can do another annotated type for cons.
list1A = ⊕{ nil0 : 1,
0
cons1 : A ⊗ list1A }
(x : A) (t : list1A ) `3 cons :: (l : list1A )
l ← cons ← x t =
l.cons ;
send l x ;
l←t
0
% (x : A) (t : list1A ) `1 l : A ⊗ list1A
% (t : list1A ) `0 l : list1A
append
Finally, let’s try the append process.
list2A = ⊕{ nil0 : 1,
0
cons2 : A ⊗ list2A }
(l1 : list2A ) (l2 : list0A ) `0 append :: (l : list0A )
l ← append ← l1 l2 =
case l1 (cons ⇒ x ← recv l1 ;
l.cons ;
send l x ;
l ← append ← l1 l2
| nil ⇒ wait l1 ;
l ← l2 )
34
% (x : A) (l1 : list2A ) (l2 : list0A ) `2 l : list0A
% (x : A) (l1 : list2A ) (l2 : list0A ) `1 l : list0A
% (l1 : list2A ) (l2 : list0A ) `0 l : list0A
% (l1 : 1) (l2 : list0A ) `0 l : list0A
% (l2 : list0A ) `0 l : list0A
B.2
Stacks as Lists
The stack interface introduced in Section 7 can be implemented using lists. First,
we define the resource-aware types.
list2A = ⊕{ nil2 : 10 ,
0
cons2 : A ⊗ list2A }
0
stack4A = N{ ins4 : A ( stack4A ,
0
del0 : ⊕{ some0 : A ⊗ stack4A ,
none0 : 10 }}
We need several sub-processes for the implementation of the stack using a
list. We will implement and type them first.
· `4 nil :: (l : list2A )
l ← nil =
l.nil ;
%
close l
%
· `1 l : 10
· `0 ·
(x : A) (t : list2A ) `4 cons :: (l : list2A )
l ← cons ← x t =
l.cons ;
send l x ;
l←t
0
% (x : A) (t : list2A ) `1 l : A ⊗ list2A
% (t : list2A ) `0 l : list2A
Finally, we can implement the stack interface using two processes, the first is
stack_new, which creates an empty list and uses it as an empty stack.
· `4 stack_new :: s : (stack4A )
s ← stack_new =
e ← nil ;
(e : list2A ) `0 (s : stack4A )
s ← stack ← e
The main process is called stack. It uses a list in its context and provides
service along s which behaves as a stack.
l : list2A `0 stack :: s : (stack4A )
s ← stack ← l
case s
(ins ⇒ x ← recv s ;
l0 ← cons ← x t
s ← stack ← l
del ⇒ case l
% (x : A)(l : list2A ) `4 s : stack2A
% l0 : list2A `0 s : stack4A
(cons ⇒ x ← recv l ;
s.some ;
35
0
%
(x : A)(l : list2A ) `2 s : ⊕{some0 : A ⊗ stack4A , none0 : 10 }
%
(x : A)(l : list2A ) `1 s : A ⊗ stack4A
0
send s x ;
s ← stack ← l
s.none
wait l ;
close s))
nil ⇒
B.3
%
(l : list2A ) `0 s : stack4A
% (l : 10 ) `1 s : 10
% · `1 s : 10
Queues as 2 Lists
A queue can be implemented using 2 lists. Insertion in such a queue has a constant
amortized cost. Since our type system supports amortized analysis, we can derive
a constant resource bound for such an implementation. The resource-aware types
we will be using are as follows.
(2,2)
listA
= ⊕{ nil2 : 10 ,
0
(2,2)
cons2 : A ⊗ listA
}
list4A = ⊕{ nil0 : 10 ,
0
cons4 : A ⊗ list4A }
(6,2)
queueA
0
(6,2)
= N{ enq6 : A ( queueA
2
0
,
0
(6,2)
deq : ⊕{ some : A ⊗ queueA
none0 : 10 }}
,
Again, we use the nil and cons sub-processes with a different type.
· `2 nil :: (l : list4A )
l ← nil =
l.nil ;
%
close l
%
· `1 l : 10
· `0 ·
(x : A) (t : list4A ) `6 cons :: (l : list4A )
l ← cons ← x t =
0
% (x : A) (t : list4A ) `1 l : A ⊗ list4A
% (t : list4A ) `0 l : list4A
l.cons ;
send l x ;
l←t
The main process queue2 acts as a client for 2 lists, and provides service along
a queue interface. We provide the implementation and its type derivation below.
(2,2)
(6,2)
0
(in : list4A ) (out : listA ) `
queue2 :: s : (queueA )
s ← queue2 ← l
case s
(2,2) 6
(6,2)
(enq ⇒ x ← recv s ;
% (x : A)(in : list4A ) (out : listA ) `
s : queueA
(2,2) 0
(6,2)
in0 ← cons ← x in
% (in0 : list4A ) (out : listA ) `
s : queueA
s ← queue2 ← in out
deq ⇒ case out
36
(cons ⇒ x ← recv out ;
%
(2,2)
(x : A) (in : list4A ) (out : listA
0
2
)`
(6,2)
s : ⊕{some0 : A ⊗ queueA
%
, none0 : 10 }
0
1
list0A ) `
list2A )
s.some ;
% (x : A) (in :
(out :
s : A ⊗ queue4A
0
send s x ;
% (in : list2A ) (out : list0A ) `
s : queue4A
s ← queue2 ← in out
4
wait out ;
% (in : list2A ) `
s : queue4A
0 2
0
0
out ← rev ← in%
; (out : listA ) ` s : queue4A
case out0
2
(cons ⇒ x ← recv out0 ; % (x : A) (out0 : list0A ) `
nil ⇒
0
%
s : ⊕{some0 : A ⊗ queue4A , none0 : 10 }
0
1
s.some ;
% (x : A) (out0 : list0A ) `
s : A ⊗ queue4A
0 0
0
send s x ;
% (out : listA ) ` s : queue4A
0
in0 ← nil ;
% (in0 : list2A ) (out0 : list0A ) `
s : queue4A
0
0
s ← queue2 ← in out
nil ⇒
B.4
0
wait out0 ;
s.none ;
close s)))
%
%
0
·`
s : ⊕{some0 : A ⊗ queue4A , none0 : 10 }
0
· ` s : 10
Higher Order Functions
Let’s consider some higher order functions. First, let’s consider the map function.
0
2
mapper2AB = N{ next0 : A ( B ⊗ mapper2AB ,
done0 : 12 }
list2A = ⊕{ nil1 : 10 ,
0
cons2 : A ⊗ list2A }
list0A = ⊕{ nil0 : 10 ,
0
cons0 : A ⊗ list0A }
Now, let’s consider the implementation of the map function.
(l : list2A ) (m : mapper2AB ) `0 map :: (k : list0A )
k ← map ← l m =
case l
(cons ⇒ x ← recv l ;
% (x : A) (l : list2A ) (m : mapper2AB ) `2 (k : list0A )
nil ⇒
0
2
(x : A) (l : list2A ) (m : A ( B ⊗ mapper2AB ) `1 (k : list0A )
m.next ;
%
send m x ;
y ← recv m ;
% (l : list2A ) (m : B ⊗ mapper2AB ) `0 (k : list0A )
% (l : list2A ) (y : B) (m : mapper2AB ) `2 (k : list0A )
k.cons ;
send k y ;
k ← map ← l m
wait l ;
% (l : list2A ) (y : B) (m : mapper2AB ) `1 (k : A ⊗ list0A )
% (l : list2A ) (m : mapper2AB ) `0 (k : list0A )
2
0
%
(m : mapper2AB ) `1 (k : list0A )
37
% (m : 12 ) `0 (k : list0A )
% · `2 (k : list0A )
% · `1 (k : 10 )
m.done ;
wait m ;
k.nil ;
close k)
Now, let’s consider the f old function.
0
0
0
folder0AB = N{ next0 : A ( B ( B ⊗ folder0AB ,
done0 : 10 }
list3A = ⊕{ nil1 : 10 ,
0
cons3 : A ⊗ list2A }
Now, let’s consider the implementation of the f old function.
(l : list3A ) (m : folder0AB ) (b : B) `0 f old :: (r : B)
r ← f old ← l m b =
case l
(cons ⇒ x ← recv l ;
% (x : A) (l : list2A ) (m : folder0AB ) (b : B) `3 (r : B)
nil ⇒
0
0
2
m.next ;
%
(x : A) (l : list3A ) (m : A ( B ( B ⊗ folder0AB ) (b : B) `2 (r : B)
send m x ;
%
(l : list3A ) (m : B ( B ⊗ folder0AB ) (b : B) `1 (r : B)
send m b ;
y ← recv m ;
k ← map ← l m
wait l ;
m.done ;
wait m ;
r ← b)
% (l : list3A ) (m : B ⊗ folder0AB ) `0 (r : B)
% (l : list2A ) (y : B) (m : folder0AB ) `2 (r : B)
0
2
2
% (m : folder0AB ) (b : B) `1 (r : B)
% (m : 10 ) (b : B) `0 (r : B)
% (b : B) `0 (r : B)
38
| 6cs.PL
|
arXiv:1707.06429v2 [math.AG] 28 Mar 2018
ULRICH BUNDLES ON NON–SPECIAL SURFACES
WITH pg = 0 AND q = 1
GIANFRANCO CASNATI
Abstract. Let S be a surface with pg (S) = 0, q(S)
= 1 and endowed with a
very ample line bundle OS (h) such that h1 S, OS (h) = 0. We show that such an
S supports families of dimension p of pairwise non–isomorphic, indecomposable,
Ulrich bundles for arbitrary large p. Moreover, we show that S supports stable
Ulrich bundles of rank 2 if the genus of the general element in |h| is at least 2.
1. Introduction and Notation
Throughout the whole paper we will work on an uncountable algebraically closed
field k of characteristic 0 and PN will denote the projective space over k of dimension
N. The word surface will always denote a projective smooth connected surface.
If X is a smooth variety, then the study of vector bundles supported on X is an
important tool for understanding its geometric properties. If X ⊆ PN , then X is
naturally polarised by the very ample line bundle OX (h) := OPN (1) ⊗ OX : in this
case, at least from a cohomological point of view, the simplest bundles F on X are
the ones which are Ulrich with respect to OX (h), i.e. such that
hi X, F (−ih) = hj X, F (−(j + 1)h) = 0
for each i > 0 and j < dim(X).
The existence of Ulrich bundles on each variety is a problem raised by D. Eisenbud
and F.O. Schreyer in [19] (see [10] for a survey on Ulrich bundles). There are many
partial results (e.g. see [2], [3], [7], [8], [9], [11], [12], [13], [15], [17], [18], [26], [27],
[28], [31]). Nevertheless, all such results and those ones proved in [20] seem to suggest
that Ulrich bundles exist at least when X satisfies an extra technical condition,
namely that X is arithmetically Cohen–Macaulay, i.e. projectively normal and such
that
hi X, OS (th) = 0
for each i = 1, . . . , dim(X) − 1 and t ∈ Z. When X is not arithmetically Cohen–
Macaulay, the literature is very limited (e.g. see [9] and [14]).
Now let S ⊆ PN be a surface and set pg (S) := h2 S, OS , q(S) := h1 S, OS ,
whence χ(OS ) := 1 − q(S) + pg (S) = 0. Thanks to the Enriques–Kodaira classification of surfaces, we know that κ(S) ≤ 1 and KS2 ≤ 0 (see [6], Theorem X.4 and
Lemma VI.1). In what follows we will denote by Pic(S) the Picard group of S: it
2010 Mathematics Subject Classification. Primary 14J60; Secondary 14J26, 14J27, 14J28.
Key words and phrases. Vector bundle, Ulrich bundle.
The author is a member of GNSAGA group of INdAM and is supported by the framework of
PRIN 2015 ‘Geometry of Algebraic Varieties’, cofinanced by MIUR.
1
2
GIANFRANCO CASNATI
is a group scheme and the connected component Pic0 (S) ⊆ Pic(S) of the identity is
an abelian variety of dimension q(S) parameterizing the line bundles algebraically
equivalent to OS .
In this paper we first rewrite the proof of Proposition 6 of [10], in order to be able
to extend its statement to a slightly wider class of surfaces.
Our modified statement (which holds also without the hypothesis that kis uncountable) is as follows: recall that OS (h) is called special if h1 S, OS (h) 6= 0,
non–special otherwise.
Theorem 1.1. Let S be a surface with pg (S) = 0, q(S) = 1 and endowed with a
very ample non–special line bundle OS (h).
If OS (η) ∈ Pic0 (S)\{ OS } is such that h0 S, OS (KS ±η) = h1 S, OS (h±η) = 0,
then for each general C ∈ |OS (h)| and each general set Z ⊆ C of h0 S, OS (h)
points, there is a rank 2 Ulrich bundle E with respect to OS (h) fitting into the exact
sequence
(1)
0 −→ OS (h + KS + η) −→ E −→ IZ|S (2h + η) −→ 0.
As pointed out in [10], Proposition 6, when S is a bielliptic surface then each
very ample line bundle OS (h) is automatically non–special and there always exists
a non–trivial OS (η) ∈ Pic0 (S) of order 2 satisfying the above vanishings: thus the
bundle E defined in Theorem 1.1 is actually special, i.e. c1 (E) = 3h + KS . We can
argue similarly if S is either anticanonical, i.e. | − KS | =
6 ∅, or geometrically ruled.
A condition forcing the indecomposability of a coherent sheaf F on an n–dimensional
variety X is its stability. Recall that the slope µ(F ) and the reduced Hilbert polynomial pF (t) of F with respect to the very ample polarisation OX (h) are
µ(F ) = c1 (F )hn−1 /rk(F ),
pF (t) = χ(F (th))/rk(F ).
The coherent sheaf F is called µ–semistable (resp. µ–stable) if for all subsheaves G
with 0 < rk(G) < rk(F ) we have µ(G) ≤ µ(F ) (resp. µ(G) < µ(F )).
The coherent sheaf F is called semistable (resp. stable) if for all G as above
pG (t) ≤ pF (t) (resp. pG (t) < pF (t)) for t ≫ 0.
On an arbitrary variety we have the following chain of implications
F is µ–stable ⇒ F is stable ⇒ F is semistable ⇒ F is µ–semistable.
Nevertheless, when we restrict our attention to Ulrich bundles, the two notions of
(semi)stability and µ–(semi)stability actually coincide.
A priori, it is not clear whether the bundles constructed in Theorem 1.1 are stable.
In Section 4 we deal with their stability as follows. The sectional genus of S with
respect to OS (h) is defined as the genus of a general element of |h|. By the adjunction
formula
h2 + hKS
π(OS (h)) :=
+ 1.
2
Notice that the equality π(OS (h)) = 0 would imply the rationality of S (e.g. see
[1] and the references therein), contradicting q(S) = 1. Thus π(OS (h)) ≥ 1 in our
setup.
ULRICH BUNDLES ON NON–SPECIAL SURFACES WITH pg = 0 AND q = 1
3
Theorem 1.2. Let S be a surface with pg (S) = 0, q(S) = 1 and endowed with a
very ample non–special line bundle OS (h).
If π(OS (h)) ≥ 2, then the bundle E constructed in Theorem 1.1 from a very
general set Z ⊆ C ⊆ S of h0 S, OS (h) points is stable.
Once that the existence of Ulrich bundles of low rank is proved, one could be
interested in understanding how large a family of Ulrich bundles supported on S
can actually be. In particular we say that a smooth variety X ⊆ PN is Ulrich–wild
if it supports families of dimension p of pairwise non–isomorphic, indecomposable,
Ulrich bundles for arbitrary large p.
The last result proved in this paper concerns the Ulrich–wildness of the surfaces
we are dealing with.
Theorem 1.3. Let S be a surface with pg (S) = 0, q(S) = 1 and endowed with a
very ample non–special line bundle OS (h). Then S is Ulrich–wild.
In Section 2 we list some general results on Ulrich bundles on polarised surfaces.
In Section 3 we prove Theorem 1.1. In Section 4 we first recall some easy facts about
the stability of Ulrich bundles, giving finally the proof of Theorem 1.2. In Section 5
we prove Theorem 1.3.
Finally, the author would like to thank the referee for her/his comments which
have allowed us to improve the whole exposition.
2. General results
In general, an Ulrich bundle F on X ⊆ PN collects many interesting properties
(see Section 2 of [19]). The following ones are particularly important.
• F is globally generated and its direct
summands are Ulrich as well.
• F is initialized, i.e. h0 X, F (−h) = 0 and h0 X, F 6= 0.
• F is aCM, i.e. hi X, F (th) = 0 for each i = 1, . . . , dim(X) − 1 and t ∈ Z.
Let S be a surface. The Serre duality for F is
hi S, F = h2−i S, F ∨ (KS ) ,
i = 0, 1, 2,
and the Riemann–Roch theorem is
h0 S, F + h2 S, F = h1 S, F +
(2)
c1 (F )(c1 (F ) − KS )
+ rk(F )χ(OS ) +
− c2 (F ).
2
Proposition 2.1. Let S be a surface endowed with a very ample line bundle OS (h).
If E is a vector bundle on S, then the following assertions are equivalent:
(1) E is an Ulrich bundle with respect to OS (h);
(2) E ∨(3h + KS ) is an Ulrich bundle with respect to OS (h);
(3) E is an aCM bundle and
c1 (E)h = rk(E)
(3)
c2 (E) =
3h2 + hKS
,
2
c1 (E)2 − c1 (E)KS
− rk(E)(h2 − χ(OS ));
2
4
GIANFRANCO CASNATI
(4) h0 S, E(−h) = h0 S, E ∨ (2h + KS ) = 0 and Equalities (3) hold.
Proof. See [14], Proposition 2.1.
The following corollaries are immediate consequences of the above characterization.
Corollary 2.2. Let S be a surface endowed with a very ample line bundle OS (h).
If OS (D) is a line bundle on S, then the following assertions are equivalent:
(1) OS (D) is an Ulrich bundle with respect to OS (h);
(2) OS (3h + KS − D) is an Ulrich bundle with respect to OS (h);
(3) OS (D) is an aCM bundle and
(4)
1
D 2 = 2(h2 − χ(OS )) + DKS ,
Dh = (3h2 + hKS );
2
(4) h0 S, OS (D − h) = h0 S, OS (2h + KS − D) = 0 and Equalities (4) hold.
Proof. See [14], Corollary 2.2.
3. Existence of rank 2 Ulrich bundles
We start this section by recalling that if S is any surface, then the connected
component Pic0 (S) of the identity inside Pic(S) is an abelian variety of dimension
q(S) called Picard variety of S. The quotient is a finitely generated abelian group
called Néron–Severi group of S.
Now, let S be a surface with pg (S) = 0 and q(S) = 1. Then Pic0 (S) is an elliptic
curve: in particular Pic0 (S) contains three pairwise distinct non–trivial divisors of
order 2.
In order to prove Theorem 1.1 we will make use of the Hartshorne–Serre correspondence on surfaces. We recall that a locally complete intersection subscheme Z
of dimension zero on a surface S is Cayley–Bacharach (CB for short) with respect to
′
a line bundle O
S (A) if, for each Z ⊆ Z of degree deg(Z) − 1, the natural morphism
H 0 S, IZ|S (A) → H 0 S, IZ ′ |S (A) is an isomorphism.
Theorem 3.1. Let S be a surface and Z ⊆ S a locally complete intersection subscheme of dimension 0.
Then there exists a vector bundle F of rank 2 on S fitting into an exact sequence
of the form
(5)
0 −→ OS −→ F −→ IZ|S (A) −→ 0,
if and only if Z is CB with respect to OS (A + KS ).
Proof. See Theorem 5.1.1 in [23].
We now prove Theorem 1.1 stated in the introduction. As we already noticed
therein, its proof for hKS = 0 coincides with
the one of Proposition 6 in [10] because
in this case the vanishing h1 S, OS (h±η) = 0 follows immediately from the Kodaira
vanishing theorem as we will show below in Corollary 3.3.
ULRICH BUNDLES ON NON–SPECIAL SURFACES WITH pg = 0 AND q = 1
5
Proof of Theorem 1.1. Recall that by hypothesis pg (S) = h1 S, OS (h) = 0 and
q(S) = 1. It follows that χ(OS ) = 0 and
h2 S, OS (h) = h0 S, OS (KS − h) ≤ h0 S, OS (KS ) = 0,
thus S ⊆ PN , where
h2 − hKS
− 1 ≥ 4,
N := h0 S, OS (h) − 1 =
2
because q(S) = 0 for each surface S ⊆ P3 .
Let C := S ∩ H ∈ |h| be a general hyperplane section and let i : C → S be the
inclusion morphism. The curve C is non–degenerate in PN −1 ∼
= H ⊆ PN . Indeed
the exact sequence
(6)
0 −→ IS|PN (1) −→ OPN (1) −→ OS (h) −→ 0
implies h0 PN , IS|PN (1) = h1 PN , IS|PN (1) = 0. Thus, the exact sequence
0 −→ IS|PN (1) −→ IC|PN (1) −→ IC|S (h) −→ 0
implies h0 PN , IC|PN (1) = 1, because IC|S (h) ∼
= OS . Finally the exact sequence
0 −→ IH|PN (1) −→ IC|PN (1) −→ IC|H (1) −→ 0
and the isomorphism IH|PN (1) ∼
= OPN yields h0 C, IC|H (1) = 0.
It follows the existence of a reduced subscheme Z ⊆ C ⊆ S of degree N + 1
whose points are in general position inside H ∼
= PN −1 . Thus Z is CB with respect
to OS (h), hence there exists Sequence (5) with OS (A) ∼
= OS (h − KS ), thanks to
Theorem 3.1.
Let OS (η) ∈ Pic0 (S)\{ OS } be such that h0 S, OS (KS ±η) = h1 S, OS (h±η) =
0 and set E := F (h + KS + η). The bundle E fits into Sequence (1)
and satisfies
0
0
∨
Equalities (3). If we show that h S, E(−h) = h S, E (2h + KS ) = 0, then we
conclude that E is Ulrich thanks to Proposition
2.1 above. Notice that the second
0
vanishing is equivalent to h S, E(−h
− 2η) = 0 because c1 (E) = 3h + KS + 2η.
0
The vanishing h S, OS (KS ± η) = 0 implies
h0 S, E(−h) ≤ h0 S, IZ|S (h + η) ,
h0 S, E(−h − 2η) ≤ h0 S, IZ|S (h − η) .
The exact sequence
0 −→ IC|S −→ IZ|S −→ IZ|C −→ 0
and the isomorphisms IC|S ∼
= OC (−Z) imply
= OS (−h) and IZ|C ∼
h0 S, IZ|S (h ± η) ≤ h0 C, OC (−Z) ⊗ OS (h ± η)
because h0 S, OS (±η) = 0. Thanks to the general choice of the points in Z, the
Riemann–Roch theorem on C and the adjunction formula OC (KC ) ∼
= i∗ OS (h + KS )
on S give
h0 C, OC (−Z) ⊗ OS (h ± η) = h0 C, i∗OS (h ± η) − deg(Z) =
= h2 + 1 − π(OS (h)) − deg(Z) + h1 C, i∗ OS (h ± η) = h0 C, i∗ OS (KS ∓ η) .
(7)
6
GIANFRANCO CASNATI
The exact sequence
(8)
0 −→ OS (−h) −→ OS −→ OC −→ 0
implies the existence of the exact sequence
H 0 S, OS (KS ∓ η) −→ H 0 C, i∗ OS (KS ∓ η) −→
−→ H 1 S, OS (KS − h ∓ η) ∼
= H 1 S, OS (h ± η) .
Thus the hypothesis on OS (KS ± η) and OS (h ± η) forces h0 C, i∗ OS (KS ∓ η) =
0.
It is natural to ask when the vanishings h1 S, OS (KS ±η) = h1 S, OS (h±η) = 0
actually occur. We list below some related result.
Corollary 3.2. Let S be a surface with pg (S) = 0, q(S) = 1 and endowed with a
very ample non–special line bundle OS (h).
Then S supports Ulrich bundles of rank r ≤ 2.
Proof. Since each direct summand of an Ulrich bundle is Ulrich as well, it follows
from Theorem 1.1 that it suffices
to prove the existence
of OS (η) ∈ Pic0 (S) \ { OS }
0
1
such that h S, OS (KS ± η) = h S, OS (h ± η) = 0.
Let P be the Poincaré line bundle on S × Pic(S). Recall that (e.g. see [29],
Lecture 19), if p : S × Pic(S) → Pic(S) is the projection on the second factor and
L ∈ Pic(S), then the restriction of P to the fibre p−1 (L) ∼
= S is isomorphic to the
line bundle L. The line bundle P is thus flat on Pic(S).
Let P0 be the restriction of P to S × Pic0 (S), A ⊆ S a divisor, s : S × Pic(S) → S
the projection on the first factor. The line bundle P0 ⊗ s∗ OS (A) is flat over Pic0 (S)
and parameterizes the line bundles on S algebraically equivalent to OS (A). Thus
the semicontinuity theorem (e.g. see Theorem III.12.8 of [22]) applied to the sheaf
P0 ⊗ s∗ OS (A) and the map p0 : S × Pic0 (S) → Pic0 (S) imply that for each i = 0, 1, 2
and c ∈ Z the sets
VAi (c) := { η ∈ Pic0 (S) | hi S, OS (A ± η) > c },
0
are closed inside Pic0 (S). In particular V := Vh1 (0) ∪ VK
(0) is closed.
S
0
By definition OS ∈ Pic (S) \ V =
6 ∅. Thus for each general OS (η) ∈ Pic0 (S), the
hypothesis h0 S, OS (KS ± η) = h1 S, OS (h ± η) = 0 is satisfied and the statement
is then completely proved.
Notice that the above result guarantees the existence of an Ulrich bundle E with
c1 (E) = 3h + KS + 2η fitting into Sequence (1). Such bundle is special if and only if
OS (η) has order 2. It is not clear if such a choice can be done in general. Anyhow in
some particular cases we can easily prove an existence result also for special Ulrich
bundles: we start from Beauville’s result for bielliptic surfaces, i.e. minimal surfaces
S with pg (S) = 0, q(S) = 1 and κ(S) = 0 (see Proposition 6 of [10]).
Corollary 3.3. Let S be a bielliptic surface endowed with a very ample line bundle
OS (h).
Then OS (h) is non–special and S supports special Ulrich bundles of rank 2.
ULRICH BUNDLES ON NON–SPECIAL SURFACES WITH pg = 0 AND q = 1
7
Proof. If κ(S) = 0, then KS is numerically trivial, hence h − KS ± η is ample for
each choice of OS (η) ∈ Pic0 (S), thanks to the Nakai criterion. Thus the vanishing
h1 S, OS (h ± η) = 0 follows from the Kodaira vanishing theorem: in particular
OS (h) is non–special.
We can find OS (η) ∈ Pic0 (S) \ { OS , OS (±KS ) } of order 2, because there are
three non–trivial and
pairwise non–isomorphic elements of order 2 in Pic0 (S). Thus
h0 S, OS (KS ± η) = 0 because KS ± η is not trivial by construction, hence the
statement follows from Theorem 1.1.
The surface S is anticanonical if | − KS | =
6 ∅: in particular pg (S) = 0. The
ampleness of OS (h) implies hKS < 0 in this case.
Corollary 3.4. Let S be an anticanonical surface with q(S) = 1 and endowed with
a very ample line bundle OS (h).
Then OS (h) is non–special and S supports special Ulrich bundles of rank 2.
Proof. If A ∈ | − KS|, then ωA ∼
= OA by the adjunction formula. We have
1
0
h A, OS (h ± η) ⊗ OA = h A, OS (−h ∓ η) ⊗ OA , for each OS (η) ∈ Pic0 (S).
On the one hand, if h0 A, OS (−h ∓ η) ⊗ OA > 0, then −hC ≥ 0 for some
irreducible component C ⊆ A. On the other hand OS (h) is ample, hence hC > 0.
The contradiction implies h0 A, OS (−h ∓ η) ⊗ OA = 0, hence the cohomology of
the exact sequence
0 −→ OS (h + KS ∓ η) −→ OS (h ∓ η) −→ OS (h ∓ η) ⊗ OA −→ 0
and the Kodaira vanishing theorem yield h1 S, OS (h ∓ η) = 0. In particular OS (h)
is non–special. Finally hKS < 0, hence h0 S, OS (KS ± η) = 0.
The statement then follows from Theorem 1.1 by taking any non–trivial OS (η) ∈
Pic0 (S) of order 2.
Recall that a geometrically ruled surface is a surface S with a surjective morphism
p : S → E onto a smooth curve such that every fibre of p is isomorphic to P1 . If S
is geometrically ruled, then pg (S) = 0 and q(S) is the genus of E (see [22], Chapter
V.2 for further details).
Remark 3.5. Let S be a geometrically ruled surface on an elliptic curve E so that
pg (S) = 0 and q(S) = 1. Thanks to the results in [22], Chapter V.2, we know
the existence of a vector bundle H of rank 2 on E such that h0 E, H 6= 0 and
h0 E, H(−P ) = 0 for each P ∈ E and S ∼
= P(H). Then p can be identified with
the natural projection map P(H) → E. The group Pic(S) is generated by the class
ξ of OP(H) (1) and by p∗ Pic(E). If we set OE (h) := det(H) and e := − deg(h), then
e ≥ −1 (see [30]). Moreover, KS = −2ξ + p∗ h.
There exists an exact sequence
(9)
0 −→ OE −→ H −→ OE (h) −→ 0.
The symmetric product of Sequence (9) yields
(10)
0 −→ H(−h) −→ S 2 H(−h) −→ OE (h) −→ 0.
8
GIANFRANCO CASNATI
Sequence (9) splits if and only if H is decomposable. Thus, if this occurs, then
S H(−h) contains OE as direct summand, whence
(11)
h0 S, OS (−KS ) ≥ h0 E, OE = 1.
because h0 S, OS (−KS ) = h0 E, S 2 H(−h) , thanks to the projection formula.
Assume that H is indecomposable. Then either OE (h) = OE or OE (h) 6= OE .
In the first case the cohomology of Sequences (9) and (10) again implies Inequality
(11).
If OE (h) 6= OE , then Lemma 22 of [4] implies that S 2 H(−h) is the direct
sum of
the three non–trivial elements of order 2 of Pic(E), hence h0 S, OS (−KS ) = 0.
We conclude that a geometrically ruled surface on an elliptic curve is anticanonical
if and only if e ≥ 0.
2
Thanks to the above remark and Corollary 3.4, we know that each geometrically
ruled surface S with q(S) = 1 and e ≥ 0 supports special Ulrich bundles of rank 2
with respect to each very ample line bundle OS (h). We can extend the result also
to the case e = −1.
Corollary 3.6. Let S be a geometrically ruled surface with q(S) = 1 and endowed
with a very ample line bundle OS (h).
Then OS (h) is non–special and S supports special Ulrich bundles of rank 2.
Proof. We have to prove the statement only for e = −1. If OS (h) = OP(H) (aξ + p∗ b),
then deg(b) > −a/2 (see [22], Proposition
V.2.21). Thus the0 Table in Proposition
1
3.1 of [21] implies that h S, OS (h ± η) = 0 for each η ∈ Pic (S).
Again the statement follows from Theorem 1.1 by taking any non–trivial OS (η)
of order 2.
Remark 3.7. The corollary above extends Propositions 3.1, 3.3 and Theorem 3.4
of [2] to the range e ≤ 0, when g = 1.
Recall that an embedded surface S ⊆ PN is called non–degenerate if it is not
contained in any hyperplane.
Corollary 3.8. Let S ⊆ P4 be a non–degenerate non–special surface with pg (S) = 0.
Then S supports special Ulrich bundles of rank 2.
Proof. The cohomology of Sequence (8) tensored by OS (h) implies h1 C, i∗ OS (h) =
0. In particular such surfaces are sectionally non–special (see [24] for details). Non–
special and sectionally non–special surfaces are completely classified in [24] and [25].
They satisfy q(S) ≤ 1 and, if equality holds, then they are either quintic scrolls over
elliptic curves, or the Serrano surfaces (these are very special bielliptic surfaces of
degree 10: see [32]). The results above and Section 4 of [14] yields the statement.
Remark 3.9. Linearly normal non–special surface S ⊆ P4 with pg (S) = 0 satisfy
3 ≤ h2 ≤ 10 (see [24] and [25]). If h2 ≤ 6, such surfaces are known to support Ulrich
line bundles: see [27] for the case q(S) = 0 and [10], Assertion 2) of Proposition 5
for the case q(S) = 1.
ULRICH BUNDLES ON NON–SPECIAL SURFACES WITH pg = 0 AND q = 1
9
4. Stability of Ulrich bundles
We start this section by recalling the following result: see [13], Theorem 2.9 for
its proof.
Theorem 4.1. Let X be a smooth variety endowed with a very ample line bundle
OX (h).
If E is an Ulrich bundle on X with respect to OX (h), the following assertions hold:
(1) E is semistable and µ–semistable;
(2) E is stable if and only if it is µ–stable;
(3) if
0 −→ L −→ E −→ M −→ 0
is an exact sequence of coherent sheaves with M torsion free and µ(L) =
µ(E), then both L and M are Ulrich bundles.
We now prove Theorem 1.2 stated in the introduction.
Proof of Theorem 1.2. Recall that E is constructed as follows. First we choose
C := S ∩ H ∈ |h| where H ∼
= PN −1 is a general hyperplane: from now on we denote
by i : C → S the inclusion morphism. The Hilbert scheme HC of 0–dimensional
subschemes of degree N + 1 on C has dimension N + 1 and contains an open non–
empty subset R ⊆ HC corresponding to reduced schemes of N + 1 points in general
position in H. If we choose a general Z ∈ R, then we finally construct E from Z by
means of Theorem 3.1.
We now show that if Z is very general inside R, i.e. it is in the complement of a
countable union of suitable proper closed subsets, then E is stable.
To this purpose, let OS (D) be an Ulrich line bundle on S (if any). By hypothesis
π(OS (h)) ≥ 2, then
(h + η − D)h = −
h2 + hKS
= 1 − π(OS (h)) ≤ −1,
2
hence
(12)
h0 S, IC|S (2h + η − D) = h0 S, OS (h + η − D) = 0,
i.e. there are no divisors in |2h + η − D| containing C. Thus the cohomology of
Sequence (8) tensored by OS (2h + η − D) yields the injectivity of the restriction
map
h0 S, OS (2h + η − D) → H 0 C, i∗ OS (2h + η − D) .
Since (2h + η − D)h = N + 1, it follows that each Z ⊆ A ∈ |2h + η − D| representing
a point the Hilbert scheme HC of subschemes of degree N + 1 on C, is actually cut
out on C by A.
Thus, if ZD denotes
the closed subset of HC of points Z such that h0 C, IZ|C ⊗
i∗ OS (2h + η − D) ≥ 1, then
dim(ZD ) = h0 C, i∗ OS (2h + η − D) − 1.
10
GIANFRANCO CASNATI
On the one hand, if i∗ OS (2h + η − D) is special, then the Clifford theorem and
the second Equality (4) imply
(2h + η − D)h
N +3
h0 C, i∗ OS (2h + η − D) ≤
+1=
≤ N,
2
2
because N ≥ 4 (see Inequality (6)). On the other hand, if i∗ OS (2h + η − D) is
non–special, the Riemann–Roch theorem on C and the second Equality (4) return
h0 C, i∗ OS (2h + η − D) = N + 2 − π(OS (h)) ≤ N,
because π(OS (h)) ≥ 2. It follows from the above inequalities that dim(ZD ) ≤ N −1.
Since q(S) = 1 and the Néron–Severi group of S is a finitely generated abelian
group, it follows that the set D ⊆ Pic(S) of Ulrich line bundles is contained in a
countable disjoint union of a fixed elliptic curve. In particular there is
[
Z ∈ R\
ZD
OS (D)∈D
because dim(R) = N + 1. Let E be the corresponding bundle.
Assume that E is not stable: then it is not µ–stable, thanks to Theorem 4.1. In
particular there exists a line subbundle OS (D) ⊆ E such that µ(E) = µ(OS (D)).
Again Theorem 4.1 implies that OS (D) is Ulrich.
On the one hand, if OS (D) is contained in the kernel K ∼
= OS (h + KS +
η) of the
map E → IZ|S (2h + η) in Sequence (1), then h0 S, OS (h + KS + η − D) 6= 0. On
the other hand, Equality (4) and Inequality (6) imply
h2 − hKS
(h + KS + η − D)h = −
= 1 − N ≤ −3,
2
whence h0 S, OS (h + KS + η − D) = 0.
We deduce that OS (D) 6⊆ K, hence the composite
map OS (D) ⊆ E → IZ|S (2h+η)
0
should be non–zero, i.e. h S, IZ|S (2h + η − D) ≥ 1. The cohomology of Sequence
(7) tensored by OS (2h + η − D) and Equality (12) then would imply
h0 C, IZ|C ⊗ OS (2h + η − D) ≥ h0 S, IZ|S (2h + η − D) ≥ 1,
contradicting our choice of Z: thus the bundle E is necessarily stable.
Remark 4.2. If π(OS (h)) = 1, then S is a geometrically ruled surface embedded
as a scroll by OS (h) ∼
= OS (ξ + p∗ b) over an elliptic curve, thanks to [1], Theorem A
(here we are using the notation introduced in Remark 3.5).
Moreover (h + η − D)h = 0, hence the argument in the above proof does not
lead to any contradiction when OS (D) ∼
= OS (h + η). Such a line bundle is actually
Ulrich, because one easily checks that it satisfies all the conditions of Corollary 2.2.
In [16], via a slightly different but similar construction, we are able to show the
existence special stable Ulrich bundles of rank 2 on elliptic scrolls.
Let S be a surface with pg (S) = 0, q(S) = 1 and endowed with a very ample
non–special line bundle OS (h). Let
c1 := 3h + KS + 2η,
c2 :=
5h2 + 3hKS
,
2
ULRICH BUNDLES ON NON–SPECIAL SURFACES WITH pg = 0 AND q = 1
11
where OS (η) ∈ Pic0 (S) \ { OS } satisfies
h0 S, OS (KS ± η) = h1 S, OS (h ± η) = 0.
If π(OS (h)) ≥ 2, then the coarse moduli space MsS (2; c1 , c2 ) parameterizing isomorphism classes of stable rank 2 bundles on S with Chern classes c1 and c2 is non–empty
s
(see Theorem 1.2). The locus Ms,U
S (2; c1 , c2 ) ⊆ MS (2; c1 , c2 ) parameterizing stable
Ulrich bundles is open as pointed out in [13].
Proposition 4.3. Let S be a surface with pg (S) = 0, q(S) = 1 and endowed with a
very ample non–special line bundle OS (h).
If π(OS (h)) ≥ 2, then there is a component US (η) of dimension at least h2 −KS2 in
Ms,U
S (2; c1 , c2 ) containing all the points representing the stable bundles E constructed
in Theorem 1.1.
Proof. Let us denote by HS the Hilbert flag scheme of pairs (Z, C) where C ∈ |OS (h)|
and Z ⊆ C is a 0–dimensional subscheme of degree N + 1. The general C ∈ |OS (h)|
is smooth and its image via the map induced by OS (h) generate a hyperplane inside
PN . Thus the set HSU ⊆ HS of pairs (Z, C) corresponding to sets of points Z in a
smooth curve C ⊆ PN which are in general position in the linear space generated
by C is open and non–empty.
We have a well–defined forgetful dominant morphism HS → |OS (h)| whose fibre
over C is an open subset of the (N + 1)–symmetric product of C. In particular HSU
is irreducible of dimension 2N + 1. Let (Z, C) represent a point of HSU : the Ulrich
bundles associated to such a point via the construction described in Theorem 1.1
correspond to the sections of
∨
Ext1S IZ|S (h − KS ), O ∼
= H 1 S, IZ|S (h) .
By definition of HSU , we have h0 C, IZ|C (h) = 0, hence the cohomology of the exact
sequence
0 −→ IZ|C (h) −→ OC (h) −→ OZ (h) −→ 0
and the Riemann–Roch theorem for OC (h) yield h1 C, IZ|C (h) = deg(Z)−χ(OC (h)) =
∼
0. Sequence (7), the isomorphism IC|S
= OS (−h) and the hypothesis q(S) =
1
pg (S) = 0 finally return h S, IZ|S (h) = 1. Thus we have a family E of Ulrich
bundles of rank 2 with Chern classes c1 and c2 parameterized by HSU .
If π(OS (h)) ≥ 2, then the bundles in the family are also stable for a general
choice of Z. Since stability is an open property in a flat family (see [23], Proposition
2.3.1 and Corollary 1.5.11), it follows the existence of an irreducible open subset
HSs,U ⊆ HSU ⊆ HS of points corresponding to stable bundles.
Thus, we have a morphism HSs,U → Ms,U
S (2; c1 , c2 ) whose image parameterizes
the isomorphism classes of stable bundles constructed in Theorem 1.1. In particular
such bundles, correspond to the points of a single irreducible component US (η) ⊆
Ms,U
S (2; c1 , c2 ).
Theorems 4.5.4 and 4.5.8 of [23] imply that dim(US (η)) ≥ 4c2 − c21 − 3χ(OS ).
Taking into account the definitions of c1 and c2 , simple computations finally yield
dim(US (η)) ≥ h2 − KS2 .
12
GIANFRANCO CASNATI
If we have some extra informations on the surface S, then we can describe US (η)
as the following proposition shows.
Proposition 4.4. Let S be an anticanonical surface with pg (S) = 0, q(S) = 1 and
endowed with a very ample line bundle OS (h).
If π(OS (h)) ≥ 2, then US (η) is non–rational and generically smooth of dimension
2
h − KS2 .
Proof. Thanks to Corollary 3.4 we know that OS (h) is non–special. Let A ∈ | −KS |:
the cohomology of
0 −→ OS (KS ) −→ OS −→ OA −→ 0
∨
tensored with E ⊗ E yields the exact sequence
0 −→ H 0 S, E ⊗ E ∨ (KS ) −→ H 0 S, E ⊗ E ∨ −→ H 0 A, E ⊗ E ∨ ⊗ OA .
Since E is stable (see Theorem 1.2), then it is simple, i.e. h0 S, E ⊗ E ∨ = 1 (see
[23], Corollary 1.2.8), hence the map
H 0 S, E ⊗ E ∨ −→ H 0 A, E ⊗ E ∨ ⊗ OA
is injective. We deduce that h2 S, E ⊗ E ∨ = h0 S, E ⊗ E ∨ (KS ) = 0.
Thus E corresponds to a smooth point of US (η) and dim(US (η)) = h2 −KS2 , thanks
to Corollary 4.5.2 of [23]. Finally, being q(S) = 1, then US (η) is irregular thanks to
[5] as well.
Remark 3.5 and the above proposition yield the following corollary.
Corollary 4.5. Let S be a geometrically ruled surface with q(S) = 1, e ≥ 0 and
endowed with a very ample line bundle OS (h).
If π(OS (h)) ≥ 2, then US (η) is non–rational and generically smooth of dimension
2
h .
5. Ulrich–wildness
Let S be a surface with pg (S) = 0 and q(S) = 1. Moreover π(OS (h)) ≥ 1 because
S is not rational, as pointed out in the introduction.
We will make use of the following result.
Theorem 5.1. Let X be a smooth variety endowed with a very ample line bundle
OX (h).
If A and Bare simple Ulrich bundles on X such that h1 X, A ⊗ B∨ ≥ 3 and
h0 X, A ⊗ B∨ = h0 X, B ⊗ A∨ = 0, then X is Ulrich–wild.
Proof. See [20], Theorem 1 and Corollary 1.
An immediate consequence of the above Theorem is the proof of Theorem 1.3.
Proof of Theorem 1.3. Recall that S is a surface with pg (S) = 0, q(S) = 1 and
endowed with a very ample non–special line bundle OS (h). We have π(OS (h)) ≥ 1,
χ(OS ) = 0 and KS2 ≤ 0.
If π(OS (h)) ≥ 2, then Theorems 1.1 and 1.2 yield the existence of a stable special
Ulrich bundle E of rank 2 on S.
ULRICH BUNDLES ON NON–SPECIAL SURFACES WITH pg = 0 AND q = 1
13
The local dimension of MsS (2; c1 , c2 ) at the point corresponding to E is at least
4c2 − c21 = h2 − KS2 ≥ 1. Thus, there exists a second stable Ulrich bundle G ∼
6= E of
rank 2 with ci (G) = ci , for i = 1, 2. Both E and G, being stable, are simple (see [23],
Corollary 1.2.8).
Due to Proposition 1.2.7 of [23] we have h0 F, E ⊗ G ∨ = h0 F, G ⊗ E ∨ = 0, thus
h1 F, E ⊗ G ∨ = h2 F, E ⊗ G ∨ − χ(E ⊗ G ∨ ) ≥ −χ(E ⊗ G ∨ ).
Equality (2) with F := E ⊗ G ∨ and the equalities rk(E ⊗ G ∨ ) = 4, c1 (E ⊗ G ∨ ) = 0
and c2 (E ⊗ G ∨ ) = 4c2 − c21 imply
h1 F, E ⊗ G ∨ ≥ 4c2 − c21 = h2 − KS2 ≥ 3.
because surfaces of degree up to 2 are rational. We conclude that S is Ulrich–wild,
by Theorem 5.1.
Finally let π(OS (h)) = 1. In this case, S is a geometrically ruled surface on an
elliptic curve E thanks to Theorem A of [1] embedded as a scroll bay OS (h). Using
the notations of Remark 3.5 we can thus assume that OS (h) = OS (ξ + p∗ b), where
deg(b) ≥ e + 3.
Assertion 2) of Proposition 5 in [10] yields that for each ϑ ∈ Pic0 (E) \ { OE }
the line bundle L := OS (h + p∗ ϑ) ∼
= OS (ξ + p∗ b + p∗ ϑ) is Ulrich. It follows from
is Ulrich too.
Corollary 2.2 that M := OS (2h + KS − p∗ ϑ) ∼
= p∗ OE (2b +
h − ϑ)
0
∨
0
Trivially, such bundles are simple and h S, L ⊗ M = h S, M ⊗ L∨ = 0
because L ∼
6= M. Since L ⊗ M∨ ∼
= OS (ξ − p∗ b − p∗ h + 2ϑ) and e = − deg(h) ≥ −1,
it follows from Equality (2) that
h1 S, L ⊗ M∨ ≥ −χ(L ⊗ M∨ ) = 2 deg(b) − e ≥ e + 6 ≥ 5.
The statement thus again follows from Theorem 5.1.
The following consequence of the above theorem is immediate, thanks to Corollaries 3.3, 3.4, 3.6.
Corollary 5.2. Let S be a surface endowed with a very ample line bundle OS (h).
If S is either bielliptic, or anticanonical with q(S) = 1, or geometrically ruled with
q(S) = 1, then it is Ulrich–wild.
The following corollary strengthens the second part of the statements of Theorems
4.13 and 4.18 in [27].
Corollary 5.3. Let S ⊆ P4 be a non–degenerate linearly normal non–special surface
of degree at least 4 with pg (S) = 0. Then S is Ulrich–wild.
Proof. As pointed out in the proof of Corollary 3.8 the surface S satisfies q(S) ≤ 1
and if equality holds it is either an elliptic scroll or a bielliptic surface. Theorem 1.3
above and Section 5 of [14] yields that S is Ulrich–wild.
References
[1] M. Andreatta, A. Sommese: Classification of irreducible projective surfaces of smooth sectional
genus ≤ 3. Math. Scand. 67 (1990), 197–214.
[2] M. Aprodu, L. Costa, R.M. Miró–Roig: Ulrich bundles on ruled surfaces. J. Pure Appl.
Algebra, 222 (2018),131–138.
14
GIANFRANCO CASNATI
[3] M. Aprodu, G. Farkas, A. Ortega: Minimal resolutions, Chow forms and Ulrich bundles on
K3 surfaces. J. Reine Angew. Math. 730 (2017), 225–249.
[4] M.F. Atiyah: Vector bundles over an elliptic curve. Proc. London Math. Soc. 7 (1957), 414–
452.
[5] E. Ballico, L. Chiantini: Some properties of stable rank 2 bundles on algebraic surfaces. Forum
Math. 4 (1992), 417–424.
[6] A. Beauville: Complex Algebraic Surfaces. L.M.S. Student Texts 34, (1996).
[7] A. Beauville: Determinantal hypersurfaces. Michigan Math. J. 48 (2000), 39–64.
[8] A. Beauville: Ulrich bundles on abelian surfaces. Proc. Amer. Math. Soc. 144 (2016), 4609–
4611.
[9] A. Beauville: Ulrich bundles on surfaces with pg = q = 0. Available at arXiv:1607.00895
[math.AG].
[10] A. Beauville: An introduction to Ulrich bundles. Available at arXiv:1610.02771 [math.AG].
[11] L. Borisov, H. Nuer: Ulrich bundles on Enriques surfaces. Available at arXiv:1606.01459
[math.AG].
[12] M. Casanellas, R. Hartshorne: ACM bundles on cubic surfaces. J. Eur. Math. Soc. 13 (2011),
709–731.
[13] M. Casanellas, R. Hartshorne, F. Geiss, F.O. Schreyer: Stable Ulrich bundles. Int. J. of Math.
23 1250083 (2012).
[14] G. Casnati: Special Ulrich bundles on non–special surfaces with pg = q = 0. Int. J. Math. 28
(2017), 1750061.
[15] G. Casnati, F. Galluzzi: Stability of rank 2 Ulrich bundles on projective K3 surface. Available
at arXiv:1607.05469 [math.AG]. To appear in Math. Scand..
[16] G. Casnati: On the existence of Ulrich bundles on geometrically ruled surface. Preprint.
[17] E. Coskun, R.S. Kulkarni, Y. Mustopa: The geometry of Ulrich bundles on del Pezzo surfaces.
J. Algebra 375 (2013), 280–301.
[18] E. Coskun, R.S. Kulkarni, Y. Mustopa: Pfaffian quartic surfaces and representations of Clifford algebras. Doc. Math. 17 (2012), 1003–1028.
[19] D. Eisenbud, F.O. Schreyer, J. Weyman: Resultants and Chow forms via exterior syzigies. J.
Amer. Math. Soc. 16 (2003), 537–579.
[20] D. Faenzi, J. Pons-Llopis: The CM representation type of projective varieties. Available at
arXiv:1504.03819 [math.AG].
[21] F.J. Gallego, B. P. Purnaprajna: Normal presentation on elliptic ruled surfaces. J. Algebra
186 (1996), 597–625.
[22] R. Hartshorne: Algebraic geometry. G.T.M. 52, Springer (1977).
[23] D. Huybrechts, M. Lehn: The geometry of moduli spaces of sheaves. Second edition. Cambridge
Mathematical Library, Cambridge U.P. (2010).
[24] M. Idà, E. Mezzetti: Smooth non–special surfaces in P4 . Manuscripta Math. 68 (1990), 57–77.
[25] E. Mezzetti, K. Ranestad: The non–existence of a smooth sectionally non–special surface of
degree 11 and sectional genus 8 in the projective fourspace. Manuscripta Math. 70 (1991),
279–283.
[26] R.M. Miró–Roig: The representation type of rational normal scrolls. Rend. Circ. Mat. Palermo
62 (2013), 153–164.
[27] R.M. Miró–Roig, J. Pons–Llopis: Representation Type of Rational ACM Surfaces X ⊆ P4 .
Algebr. Represent. Theor. 16 (2013), 1135–1157.
[28] R.M. Miró–Roig, J. Pons–Llopis: N –dimensional Fano varieties of wild representation type.
J. Pure Appl. Algebra 218 (2014), 1867–1884.
[29] D. Mumford: Lectures on curves on an algebraic surface. With a section by G. M. Bergman.
Annals of Mathematics Studies, 59 Princeton U. P. (1966).
[30] M. Nagata: On self–intersection numbers of a section on a ruled surface. Nagoya Math. J.,
37 (1970), 191–196.
ULRICH BUNDLES ON NON–SPECIAL SURFACES WITH pg = 0 AND q = 1
15
[31] J. Pons–Llopis, F. Tonini: ACM bundles on del Pezzo surfaces. Matematiche (Catania) 64
(2009), 177–211.
[32] F. Serrano: Divisors of bielliptic surfaces and embeddings in P4 . Math. Z. 203 (1990), 527–533.
Gianfranco Casnati,
Dipartimento di Scienze Matematiche, Politecnico di Torino,
c.so Duca degli Abruzzi 24, 10129 Torino, Italy
e-mail: [email protected]
| 0math.AC
|
arXiv:1009.1159v4 [math.AC] 23 Apr 2014
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC
PARAMETERS∗
Benjamin Antieau, Alexey Ovchinnikov, and Dmitry Trushin
Department of Mathematics, University of California at Los Angeles,
Los Angeles, California, USA
Department of Mathematics, CUNY Queens College, Queens, New York, USA
Department of Mathematics, CUNY Graduate Center, New York, New York, USA
Einstein Institute of Mathematics, The Hebrew University of Jerusalem,
Jerusalem, Israel
We develop a Galois theory for systems of linear difference equations with periodic parameters, for which we also introduce linear difference algebraic groups. We apply this to
constructively test if solutions of linear q-difference equations, with q ∈ C∗ and q not a root of
unity, satisfy any polynomial ζ-difference equations with ζt = 1, t > 1.
Key Words: Difference Galois theory; q-Difference equations; Difference pseudofields.
2010 Mathematics Subject Classification: Primary 12H10; Secondary 13N99, 20H25, 39A13.
1 INTRODUCTION
In this paper, we give a new Galois theory of systems of linear difference equations
with periodic (of finite order) difference parameters. This appears to be the first time that
equations with difference parameters have been treated in the literature. In this theory, the
Galois groups are linear difference algebraic groups, and they measure difference algebraic
dependence of solutions of difference equations. Our Galois theory and Galois correspondence works over fields of any characteristic. In characteristic p, our Galois correspondence
is presented here for separable extensions of difference pseudofields1. Among numerous
potential applications of our approach, we show how this can be applied to studying properties of solutions of q-difference equations.
For the purposes of this introduction, we briefly describe the setup and motivation
in the following simple case. Let q ∈ C \{0}. A q-difference equation of order n is an
equation in f of the form
f (qn z) + an−1(z) · f qn−1 z + . . . + a0(z) · f (z) = 0,
(1)
∗ Address correspondence to Prof. Alexey Ovchinnikov, Department of Mathematics, CUNY Queens College,
65-30 Kissena Blvd., Queens, NY 11367, USA; E-mail: [email protected]
1 Our proofs apply to the non-separable case by using non-reduced Hopf algebras, as pointed out to us by
Michael Wibmer.
1
2
ANTIEAU ET AL.
where a0 (z), . . . , an−1 (z) ∈ C(z) are given. Over C(z), a solution to such an equation will
not exist in general. However, there is a ring extension of C(z), called the Picard–Vessiot
ring for the equation that is universal for the property of having a full set of solutions to the
equation. Let f (z) be the solution in the Picard–Vessiot ring. When, in addition, we fix a
primitive tth root of unity, ζ, and we let Z/(t) act on C(z) by f (z) 7→ f (ζt), we show in
this paper that we can construct a Picard–Vessiot extension and an action of Z/(t) on the
ring extending the action of ζ on C(z). The motivation for developing a Galois theory of
difference equations with periodic parameters is to study the algebraic relations satisfied
by
f (z), f (ζz), . . . , f ζt−1 z .
In particular, as an application of the method of our difference Galois groups with parameters, Theorem 4.9 gives an explicit, complete description of all first-order q-difference
equations
f (qz) = a(z) f (z)
(2)
with rational coefficients whose solutions are ζ-difference algebraically independent over
the rational functions in variable z with coefficients belonging to the field k of q-invariant
meromorphic functions on C \{0}. This description is easy to use: the inputs are simple
functions in the multiplicities of the zeros and poles of a(z). Our proof requires a similar
approach to that of [21, Section 3], but it is substantially modified to take into account difference algebraic independence and make the result as explicit as possible. As an example
of our methods, we include a deduction of some algebraic independence properties of theta
functions (see Theorem 4.5).
The approach of this paper resembles the Galois theory of difference equations with
differential parameters studied in [20, 21, 22, 14, 15, 16, 17, 13, 12], where algebraic methods have been developed to test whether solutions of difference equations satisfy polynomial differential equations (see also [23] for a general Tannakian approach). In particular,
these methods can be used to prove Hölder’s theorem which says that the Γ-function, which
satisfies the difference equation Γ(x+ 1) = x·Γ(x), satisfies no non-trivial differential equation over C(x). However, when treating difference equations with differential parameters,
one may use fields as the rings of constants. This is not available when using difference
parameters, as Example 2.3 and [29, Proposition 7.3] show. The constants in our theory
are rings that have zero-divisors, and this fact requires numerous additional subtleties into
our approach. The key idea is to find a suitable notion of a difference closed ring. We
use the difference-closed pseudofields of [30], which we review in Section 2. Another approach to the question of difference algebraic closure is in [24], where difference versions
of valuation rings are given. However, since we require zero-divisors, Lando’s approach is
insufficient.
Picard–Vessiot extensions with zero divisors for systems of linear difference equation
have been considered in [32, 10, 27] with a non-linear generalization considered in [19].
Also, Galois theories of linear difference equations, without parameters, when the ground
ring has zero divisors have been studied in [4, 3, 1, 2, 34], where including zero divisors into
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
3
the ground ring is needed and provides a much more transparent Galois correspondence. In
all the mentioned cases, the ground ring must be a finite product of fields (called Noetherian
difference pseudofields).
Our approach allows us not only to treat parameters, but also prepares a foundation for
studying the non-Noetherian case as we base our methods on a natural geometric approach
to difference varieties developed in [30], which has been further generalized to the nonNoetherian case in [31]. However, to extend the theory to infinite parameter groups, it
is necessary to treat the non-Noetherian case, as the same construction results in a nonNoetherian ring. This generalization has been carried out in [28] using the methods of [34,
35] and the present paper (see also [18] for the Galois theory of linear differential equations
with difference parameters).
Some of our results can be treated in another way, via the method of faithfully flat
descent from algebraic geometry [26]. However, our theory gives a more flexible theory
than that obtained via descent, as explained in Section 3.6.
The paper is organized as follows. We give basic definitions in Section 2.1. The
main properties of difference pseudofields are detailed in Sections 2.2 and 2.3. Section 3
contains the development of our main technique, difference Galois theory (also called difference Picard–Vessiot theory) with periodic parameters. Difference algebraic groups are
introduced and studied in Section 3.3. We finish by showing in Section 4 how to use our
theory to study periodic difference algebraic dependencies among solutions of difference
equations. In particular, we apply these results to study Jacobi’s theta-function in Section 4.3 and to give a complete characterization to all first-order q-difference equations
with ζ-difference algebraically independent solutions over rational functions in variable z
with coefficients belonging to the field of q-invariant meromorphic functions on C \{0} in
Section 4.4.
2 BASIC DEFINITIONS
2.1 Difference rings
Most of the basic notions on difference algebra can be found in [11, 25]. Below, we
will introduce those that we use here. Let
Σ0 = Z,
Σ1 = Z /t1 Z ⊕ . . . ⊕ Z /ts Z,
and Σ = Σ0 ⊕ Σ1 ,
where each ti > 2. Let σ be a generator of Σ0 and ρi , 1 6 i 6 s, generate each component
of Σ1 .
A ring R equipped with an action of a fixed subgroup Σ′ ⊂ Σ by automorphisms is
called a Σ′ -ring.
Example 2.1. Let R = C(x) and σ(x) = px, ρ(x) = qx with p, q ∈ C∗ , |p| 6=
1 and q a primi-
tive m-th root of unity for some m > 2. Then Σ0 = {σn | n ∈ Z} and Σ1 = id, ρ, . . . , ρm−1 .
4
ANTIEAU ET AL.
Let R be a Σ′ -ring and let R [Σ′ ] = ∑ rτ τ rτ ∈ R, τ ∈ Σ′ denote the ring of difference
operators on R. The multiplication on R[Σ′ ] is given by τ · r = τ(r)τ. For a set Y , let
R{Y }Σ′ = R . . . , τy, . . . | τ ∈ Σ′ , y ∈ Y
denote the ring of Σ′ -polynomials over R with Y as the set of Σ′ -indeterminates.
Example 2.2. For example, if Σ′ = Σ1 = Z /2 Z and ρ is a generator of Σ1 , then R{y}Σ′ =
R[y, ρy] with the action of ρ given by ρ(y) = ρy and ρ(ρy) = y.
An ideal a ⊂ R is called a Σ′ -ideal if Σ′ (a) ⊂ a, where
Σ′ (a) := σ(a) | σ ∈ Σ′ , a ∈ a .
The smallest Σ′ -ideal containing a set F ⊂ R is denoted by [F]Σ′ . If Σ′ = Σ, then it is also
denoted simply by [F]. Let R1 and R2 be Σ′ -rings. A ring homomorphism f : R1 → R2 is
called a Σ′ -homomorphism if f (τ(r)) = τ( f (r)), for all τ ∈ Σ′ , r ∈ R1 .
The following example shows that even if we start with a base field, the constants of
the solution space as constructed in Section 3 have zero divisors.
Example 2.3. Let Σ1 = Z /4 Z with a generator ρ. Consider the equation σx = −x. The
procedure of constructing a solution space (called Picard–Vessiot extension) of the above
equation
described in Section 3 first takes C{x, 1/x}ρ , with σx = −x, and thenquotients
by ρx − ix, x4 − 1 , which is a maximal Σ-ideal. Thus, we arrive at the ring C[x] x4 − 1 ,
σx = −x and ρx = ix, which is a Σ-pseudofield generated by the solution
of the equation.
The subring of constants is generated by x2 and is isomorphic to C[t] t 2 − 1 , which is
not a field.
Denote the ring of Σ′ -constants of R by RΣ . In other words,
′
RΣ = r ∈ R | τ(r) = r for all τ ∈ Σ′ .
′
The set of all Σ′ -ideals of R will be denoted by IdΣ (R).
′
Definition 2.4. A Σ′ -ideal p of R is called pseudoprime if there exists a multiplicatively
closed subset S ⊂ R such that p is a maximal Σ-ideal with p ∩S = ∅.
Lemma 2.5. Let A and B be Σ-rings and ϕ : A → B be a Σ-homomorphism. Then for any
pseudoprime ideal q in B the ideal ϕ−1 (q) is pseudoprime.
Proof. See [30, Section 2].
The set of all pseudoprime ideals of R will be denoted by PSpec R or PSpecΣ R. For
s ∈ R, (PSpec R)s denotes the set of pseudoprime ideals of R not containing s. Let R1 and
R2 be Σ′ -rings and f : R1 → R2 be a Σ′ -homomorphism. Then f ∗ (q) := f −1 (q) defines a
map
f ∗ : PSpec R2 → PSpec R1
′
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
5
by Lemma 2.5. For an ideal a ⊂ R denote by aΣ′ the largest Σ′ -ideal of R contained in a.
Note that if p is a prime ideal of R, then the ideal pΣ′ is pseudoprime.
Recall that an R-module M with an action of Σ′ is called a Σ′ -module if for all τ ∈ Σ′ ,
r ∈ R, and m ∈ M, we have τ(rm) = τ(r)τ(m). A Σ′ -ring is called simple if it contains no
proper Σ′ -ideals except for (0).
Definition 2.6. A ring R is called absolutely flat if every R-module is flat. An absolutely
flat simple Σ′ -ring k is called a Σ′ -pseudofield (see [30]).
For every subset E ⊂ R{y1 , . . . , yn }Σ′ , let V(E) ⊂ Rn be the set of common zeroes of
E in Rn . Conversely, for every subset X ⊂ Rn , let
I(X) ⊂ R{y1 , . . . , yn }Σ′
on X. One sees that, for any
be the Σ′ -ideal of all polynomials in R{y1 , . . . , yn }Σ′ vanishing
√
reduced R and Σ′ -ideal I ⊂ R{y1 , . . . , yn }Σ′ , we have I ⊂ I(V(I)).
Definition 2.7. [30, Section 4.3] A Σ′ √
-pseudofield R is called difference closed if, for every
Σ′ -ideal I ⊂ R{y1 , . . . , yn }Σ′ , we have
I = I(V(I)).
2.2 Properties of pseudofields
Proposition 2.8. Let L be Σ′ -simple ring and K ⊂ L be an absolutely flat Σ′ -subring. Then,
K is a Σ′ -pseudofield.
Proof. Let 0 6= a ∈ K. We will show that the Σ′ -ideal of K generated by a contains 1. Since
K is absolutely flat, we may assume that a2 = a, since every principal ideal is generated by
an idempotent [5, Exercise II.27]. Since the Σ′ -ideal generated by a in L contains 1, there
exist hi ∈ L, 0 6 i 6 r, such that
1 = h0 a + h1σ1 (a) + . . . + hr σr (a)
(3)
for some σk ∈ Σ′ . Set σ0 = id for notation. We will show by induction on k 6 r that the hi ’s
can be selected so that hi ∈ K, 0 6 i 6 k. The base k = 0 is done in the same way as the
inductive step. Assume the statement for k − 1 ≥ 0. We will show it for k. Multiplying (3)
by 1 − σk (a) and using a2 = a, we have:
1 − σk (a) = (1 − σk (a))(h0 · a + . . . + hk−1 · σk−1 (a) + hk+1 · σk+1 (a) + . . . + hr · σr (a)).
Hence,
1 =(1 − σk (a))h0 · a + . . . + (1 − σk (a))hk−1 · σk−1 (a)+
+ σk (a) + (1 − σk (a))hk+1 · σk+1 (a) + . . . + (1 − σk (a))hr · σr (a)
with (1 − σk (a))h0 , . . . , (1 − σk (a))hk−1 , 1 ∈ K, which finishes the proof.
6
ANTIEAU ET AL.
Proposition 2.9. Let L be an absolutely flat ring and H ⊂ Aut(L). Then the ring LH is
absolutely flat.
Proof. Let 0 6= a ∈ LH . Then by [5, Exercise II.27] there exist unique an idempotent e and
a′ in L such that
e = aa′ , a = ea, and a′ = ea′ .
(4)
To see uniqueness, note that if (ē, ā′ ) is another such pair, then eē = eaā′ = aā′ = ē and,
similarly, eē = e. So, the element e is unique. Now, a′ = ea′ = ēa′ = aā′ a′ and, in the same
manner, ā′ = ēā′ = eā′ = aa′ ā′ .
We will show now that e and a′ are H-invariant. For σ ∈ H we have
a = σ(a) = σ(ae) = aσ(e).
Multiplying by a′ , we obtain e = eσ(e). Similarly, we obtain e = eσ−1 (e), which implies
that σ(e) = eσ(e). Hence, σ(e) = e. We, therefore, have
e = aσ(a′ ), a = ea, and σ(a′ ) = eσ(a′ ).
(5)
Since the pair (e, a′ ) is unique, (4) and (5) imply that σ(a′ ) = a′ . Applying [5, Exercise II.27] again, we conclude that LH is absolutely flat.
Proposition 2.10. Let A be a Σ1 -closed pseudofield. Then the ring R = A[Σ1 ] is completely
reducible: R ∼
= A ⊕ . . . ⊕ A as Σ1 -modules over A. In other words, every Σ1 -module over A
has a basis of Σ1 -invariant elements. Moreover, A[Σ1 ] ∼
= Mn (C) as rings, where C = AΣ1 .
Proof. Follows [30, Proposition 26 and Remark 27].
Proposition 2.11. Let R be a Σ-simple ring and A := Rσ be a Σ1 -difference closed pseud-
ofield. Let B be any Σ-A-algebra with σ acting as the identity. Then the Σ-homomorphism
B → R ⊗A B, with b 7→ 1 ⊗ b, b ∈ B, induces a bijection
IdΣ1 (B) ←→ IdΣ (R ⊗A B)
via a ⊂ B −→ ae := R ⊗A a, bc := b ∩B ←− b ⊂ R ⊗A B.
Proof. Let I be a Σ-ideal of the ring R ⊗A B and let I c = J. We will show that I = J e .
In other words, by passing to R ⊗A (B/J), we will show that if I c = (0), then I = (0).
By Proposition 2.10, there exists a basis {bi }i∈I of B over A consisting of Σ1 -invariant
elements. Then, every element of R ⊗A B is of the form
a1 ⊗ bi1 + . . . + an ⊗ bin
for some ai ∈ R, 1 6 i 6 n. Let 0 6= u ∈ I have the shortest expression of the form u =
a1 ⊗ b j1 + . . . + ak ⊗ b jk ,
M = a ∈ R | ∃c2 , . . . , ck ∈ R, i1 , . . . , ik ∈ I : a ⊗ bi1 + c2 ⊗ bi2 + . . . + ck ⊗ bik ∈ I .
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
7
As 0 6= a1 ∈ M, and since Σ(bi ) = bi , 1 6 i 6 n, the set M is a non-zero Σ-ideal of R. Hence,
1 ∈ M. Therefore, there exists u with a1 = 1. Since
u − σ(u) = (a2 − σ(a2 )) ⊗ bi2 + . . . + (ak − σ(ak )) ⊗ bik ∈ I
(6)
and has a shorter expression than u, we have
u − σ(u) = 0.
(7)
Since {bi }i∈I is a basis of B over A, {1 ⊗ bi}i∈I is a basis of R ⊗A B over R. Therefore, (6)
and (7) imply that σ(a2 ) = a2 , . . . , σ(ak ) = ak , that is, a2 , . . . , ak ∈ A. Thus,
u = 1 ⊗ bi1 + a2 bi2 + . . . + ak bik .
Hence,
0 6= bi1 + a2bi2 + . . . + ak bik ∈ I c ,
contradicting I c = (0). Therefore, we have shown that (I c )e = I. On the other hand, since R
is a free A-module, the B-module R ⊗A B is also free and, therefore, faithfully flat. Thus, by
[5, Exercise III.16] for every ideal J ⊂ B we have (J e )c = J, which finishes the proof.
Corollary 2.12. Let B be a Σ-ring containing a Σ-pseudofield L with CL := Lσ being a
Σ1 -closed pseudofield. Let C ⊂ Bσ be a Σ1 -subring such that CL ⊂ C. Then L ·C = L ⊗CL C.
Proof. The kernel I of the Σ-homomorphism
L ⊗CL C → L ·C ⊂ B,
l ⊗ c 7→ l · c,
is a Σ-ideal with I c = (0) ⊂ C. By Proposition 2.11, we conclude that I = 0.
2.3 Noetherian pseudofields
Lemma 2.13. Let A ⊂ B be Σ-rings such that for some s ∈ A the map Spec Bs → Spec As is
surjective. Then the map ϕ : (PSpec B)s → (PSpec A)s is surjective as well.
Proof. Let q ⊂ A be a pseudoprime ideal with s ∈
/ q. Then, since the maximal ideal not
intersecting a multiplicative subset is prime, by definition, there exists a prime ideal p ⊃ q
such that
\
q=
τ(p)
τ∈Σ
with q being a maximal Σ-ideal contained in τ(p), τ ∈ Σ. Since s ∈
/ q, there exists τ ∈ Σ
such that s ∈
/ τ(p). By our assumption, there exists a prime ideal p′ ⊂ B with p′ ∩A = pτ .
Then the ideal p′Σ is the pseudoprime ideal in B that is mapped to p by ϕ.
Lemma 2.14. Let A ⊂ B be Σ-rings such that A is Noetherian and reduced and B is a
finitely generated A-algebra. Then there exists 0 6= s ∈ A such that the map (PSpec B)s →
(PSpec A)s is surjective.
8
ANTIEAU ET AL.
Proof. There exists s ∈ A such that As is an integral domain. For instance, suppose that
(0) = p1 ∩· · ·∩pt is the representation of (0) as the intersection of the finitely many minimal
prime ideals in the Noetherian ring A. Let s ∈ p2 ∩ · · · ∩ pt be such that t ∈
/ p1 . Then, As is
a reduced ring with a single minimal prime ideal. Thus, it is integral. By [30, Lemma 30],
there exists t ∈ A such that the map SpecBst → SpecAst is surjective. The statement now
follows from Lemma 2.13.
Theorem 2.15. Let L be a Noetherian Σ-pseudofield with C := Lσ being a Σ1 -closed pseud-
ofield. Let R be a Σ1 -finitely generated Σ-simple ring over L. Then Rσ = C.
Proof. Let b ∈ Rσ . Since |Σ1 | < ∞, the ring R is finitely generated over L. Since R is Σ-
simple, it is reduced. Therefore, the Σ-subring of R generated by L and b, denoted by L{b},
is reduced as well. Hence, by Lemma 2.14, there exists a non-nilpotent element s ∈ L{b}
such that the map
(PSpec R)s → (PSpec L{b})s
is surjective. Therefore, since PSpec R = {(0)}, every non-zero pseudoprime ideal in L{b}
contains s. By Corollary 2.12, we have L{b} = L ⊗C C{b}. By Proposition 2.10, L is a free
C-module. Let {li }i∈I be a Σ1 -invariant basis over C. Then there exist r1 , . . . , rk ∈ C{b}
such that
s = l1 ⊗ r1 + . . . + lk ⊗ rk .
Since the ring L{b} is reduced, r1 is not nilpotent. Therefore, by [30, Proposition 34], there
exists a maximal Σ-ideal m in C{b} such that C{b}/ m = C and r1 ∈
/ m. Let
ϕ : L{b} = L ⊗C C{b} → L ⊗C C{b}/ m = L ⊗C C = L.
Then,
ϕ(s) = l1 r̄1 + . . . + lk r̄k ,
where r̄i are the images of ri modulo m, 1 6 i 6 k. Since {l1 , . . . , lk } are linearly independent
over C and r̄1 6= 0, the ideal L ⊗C m does not contain s. Since ϕ is a Σ-homomorphism,
L ⊗C m = ϕ−1 ((0)), and (0) is a pseudoprime ideal in L, the ideal L ⊗C m is pseudoprime
by Lemma 2.5. Therefore, L ⊗C m = (0) by the above. Thus, we see that b ∈ C by taking
σ-invariants as ϕ is an injective Σ-homomorphism.
Recall that an idempotent that is not a sum of several distinct orthogonal idempotents
is called indecomposable.
Proposition 2.16. Let L be a Noetherian Σ-pseudofield and let F = L/ m, where m is a
maximal ideal in L. Then, L ∼
= F × . . . × F. Moreover, Σ acts transitively on the set of
indecomposable idempotents of L.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
9
Proof. Since the ring L is Noetherian and dim L = 0, by [5, Theorem 8.5], the ring L is
Artinian. Therefore, by [5, Theorem VII.7], it is a finite product of local Artinian rings.
Since L is reduced, by [5, Proposition VIII.1],
L = F1 × . . . × Fn,
(8)
where Fi is a field, 1 6 i 6 n. Since L is Σ-simple, the group Σ acts transitively on Spec L.
Therefore, Fi ∼
= F1 , 1 6 i 6 n, as residue fields. Let e be an indecomposable idempotent in
L. Let OrbΣ (e) = {e1 , . . . , ek }. Then, the idempotent
E := e1 + . . . + ek
is Σ-invariant. Since L is Σ-simple, we have E = 1. Decomposition (8) implies that L has n
indecomposable idempotents, each one is of the form
(0, . . . , 0, 1, 0, . . . , 0)
and, thus, k = n and Σ acts transitively on the set of indecomposable idempotents of L.
Let B be a Σ0 -ring and let
FΣ1 (B) =
∏ B = { f : Σ1 → B},
(9)
µ∈Σ1
which is a Σ0 -ring with the component-wise action of Σ0 . Define
(µ f )(τ) = f µ−1 τ , f ∈ FΣ1 (B), µ, τ ∈ Σ1 .
The above makes FΣ1 (B) a Σ-ring. For every µ ∈ Σ1 define a Σ0 -homomorphism
γµ : FΣ1 (B) → B,
f 7→ f (µ).
(10)
Moreover, we have
γτ (µ f ) = (µ f )(τ) = f µ−1 τ = γµ−1 τ ( f ).
Proposition 2.17. Let A be a Σ-ring, B be a Σ0 -ring, and ϕ : A → B be a Σ0 -homomorphism.
Then for every µ ∈ Σ there exists unique Σ-homomorphism Φµ : A → FΣ1 (B) such that the
following diagram
FΣ1 (B)
<
Φµ ②②
②
γµ
②
②②
②
② ϕ
/B
A
is commutative.
10
ANTIEAU ET AL.
Proof. Since
Φµ (a) τ−1 µ = (τΦµ (a))(µ) = ϕ(τa),
where a ∈ A and τ ∈ Σ, the homomorphism Φµ is unique if it exists. Define
For every α ∈ Σ1 we have
Φµ (a)(τ) = ϕ µτ−1 a .
−1
Φµ (αa)(τ) = ϕ µτ−1 αa = ϕ µ α−1 τ a = Φµ (a) α−1 τ = (αΦµ (a))(τ)
Φµ (νa)(τ) = ϕ µτ−1 νa = ν ϕ µτ−1 a = ν(Φµ (a)(τ)) = ν(Φµ (a))(τ)
for all α, τ ∈ Σ1 , ν ∈ Σ0 , and a ∈ A. Thus, Φµ is a Σ-homomorphism.
Proposition 2.18. Let L be a Noetherian Σ-pseudofield such that Lσ is a Σ1 -closed pseud-
ofield. Then, there exists a Noetherian Σ0 -pseudofield B such that L ∼
= FΣ1 (B).
Proof. By [30, Theorem 17(4)], there exists an algebraically closed field K such that
Lσ = FΣ1 (K).
Let δτ ∈ FΣ1 (K) be the indicator of the point τ ∈ Σ1 and e = δid . Let also B = eL, which is
a Noetherian absolutely flat ring as a quotient of a Noetherian Σ-pseudofield. By Proposition 2.17, the homomorphism L → B, with a 7→ e · a, lifts to a unique Σ-homomorphism
φ : L → FΣ1 (B).
Since L is Σ-simple, φ is injective. To show that φ is surjective, we will prove that φ(L)
contains all indecomposable idempotents of FΣ1 (B). Every indecomposable idempotent of
the ring FΣ1 (B) is of the form δτ · f , f is an indecomposable idempotent of B. Let f = eh,
where h ∈ L. Since
φ(τ(e)h)(ν) = (eτ(e)h)(ν) = (τ(e) f )(ν) = e τ−1 ν f = δτ (ν) f ,
we are done. Finally, B is Σ0 -simple. Indeed, let b ⊂ B be a Σ0 -ideal. Let I ⊂ FΣ1 (B)
consist of all functions f with image contained in b. Since I is an ideal and Σ1 is acting on
the domain, I is invariant under the Σ1 -action. Since b is a Σ0 ideal, then I is a Σ0 -ideal as
well. Therefore, I is a Σ-ideal, which contradicts to L being a pseudofield.
Proposition 2.19. Let L be a Noetherian Σ-pseudofield such that Lσ is a Σ1 -closed pseud-
ofield. Then,
n
L∼
= ∏ FΣ1 (F)
i=1
as Σ1 -rings, where F is a field.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
11
Proof. By Proposition 2.18, L = FΣ1 (B), where B is a Noetherian Σ0 -pseudofield. Let
f1 , . . . , fn be all indecomposable idempotents of B. Then
L = f1 L × . . . × fn L.
On the other hand,
fi FΣ1 (B) = FΣ1 ( fi B) = FΣ1 (Fi ),
where Fi = fi B and F1 ∼
= Fi , 1 6 i 6 n.
Proposition 2.20. Let L be a Noetherian Σ-pseudofield and K ⊂ L be a Σ-pseudofield as
well. Then K is Noetherian.
Proof. Note that a pseudofield is Noetherian if and only if it contains a finite set of indecomposable idempotents e1 , . . . , en with
e1 + . . . + en = 1.
(11)
Necessity has been discussed above. To show sufficiency, note that if e is an indecomposable idempotent of an absolutely flat ring R, then eR is a field. Indeed, eR is an absolutely
flat ring without nontrivial idempotents [5, Exercise II.27]. Moreover, for every element
x ∈ R we have x = ax2 . Therefore, ax is an idempotent. So, either ax = 0 and, thus,
x = ax2 = 0, or ax = 1. Hence, equality (11) implies that R is finite product of fields and,
therefore, is Noetherian.
Thus, since every idempotent of K is an idempotent of L, which is Noetherian, the
ring K has finitely many indecomposable idempotents f1 , . . . , fk . Since f1 + . . . + fk is left
fixed by Σ, we have f1 + . . . + fk = 1. Again, by the above, the ring K is Noetherian.
Proposition 2.21. Let L be a Σ-field such that the subfield C := Lσ is algebraically closed.
Then there exists a Σ-pseudofield A and a Σ-embedding ϕ : L → A such that Aσ is the
Σ1 -closure of the Σ1 -field ϕ(C).
Proof. Set A = FΣ1 (L) and and let ϕ be the Taylor homomorphism for id : L → L by Proposition 2.17. Then, Aσ = FΣ1 (C) is the Σ1 -closure of C [30, discussions preceding Proposition 19].
3 PICARD–VESSIOT THEORY
3.1 Picard–Vessiot ring
Let K be a Noetherian Σ-pseudofield and let C = K σ be a Σ1 -closed pseudofield. Let
A ∈ GLn (K). Consider the following difference equation
σY = AY.
Let R be a Σ-ring containing K.
(12)
12
ANTIEAU ET AL.
Definition 3.1. A matrix F ∈ GLn (R) is called a fundamental matrix of equation (12) if
σF = AF.
Let F1 and F2 be two fundamental matrices of (12). Then for M := F1−1 F2 we have
σ(M) = σ(F1 )−1 σ(F2 ) = F1−1 A−1 AF2 = F1−1 F2 = M,
that is, M ∈ GLn (Rσ ).
Definition 3.2. A Σ-ring R is called a Picard–Vessiot ring for equation (12) if
1. there exists a fundamental matrix F ∈ GLn (R) for (12),
2. R is a Σ-simple ring, and
3. R is Σ-generated over K be the matrix entries Fi j and 1/ detF.
Proposition 3.3. Let K be a Noetherian Σ-pseudofield, K σ be a Σ1 -closed pseudofield, and
R be a Picard–Vessiot ring for equation (12). Then, Rσ = K σ .
Proof. Since R is a Σ1 -finitely generated algebra over K and |Σ1 | < ∞, R is finitely generated
over K. Then the result follows from Theorem 2.15.
Proposition 3.4. Let K be a Noetherian Σ-pseudofield with K σ being a Σ1 -closed pseud-
ofield. Then there exists a unique Picard–Vessiot ring for equation (12).
Proof. For existence, define the action of σ on the Σ1 -ring
R := K{Fi j , 1/ detF}Σ1
by σF = AF. Let m be any maximal Σ-ideal in R. Then R/ m is the Picard–Vessiot ring for
equation (12). For uniqueness, let R1 and R2 be two Picard–Vessiot rings of equations (12).
Let
R = (R1 ⊗K R2 )/ m,
where m is a maximal Σ-ideal. Since R1 and R2 are Σ-simple, the Σ-homomorphisms
ϕ1 : R1 → R, r 7→ r ⊗ 1,
ϕ2 : R2 → R, r 7→ 1 ⊗ r,
are injective. Let F1 and F2 be fundamental matrices of R1 and R2 , respectively. Then there
exists M ∈ GLn (Rσ ) such that ϕ1 (F1 ) = ϕ2 (F2 )M. Proposition 3.3 implies that Rσ = K σ .
Therefore, ϕ1 (F1 ) ⊂ ϕ2 (R2 ). Similarly, ϕ2 (F2 ) ⊂ ϕ1 (R1 ). Hence, ϕ1 (R1 ) = ϕ2 (R2 ) and,
thus, R1 ∼
= R.
= R2 ∼
Proposition 3.5. Let K be a Noetherian Σ-pseudofield with K σ being a Σ1 -closed pseud-
ofield and R be a Picard–Vessiot ring of equation (12). Then the complete quotient ring
L := Qt(R) is a Noetherian Σ-pseudofield with Lσ = K σ .
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
13
Proof. We will first show that L is Σ-simple. Let a be a non-zero Σ-ideal of L. Then
a ∩R 6= (0) and, therefore, 1 ∈ a.
We will now show that L is a finite product of fields. Since the ring K is Noetherian and
R is finitely generated over K, the ring R is Noetherian as well by the Hilbert basis theorem.
Hence, there exists a smallest set of prime ideals p1 , . . . , pn in R such that (0) = p1 ∩ . . . ∩pn .
S
The set of non-zero divisors in R coincides with R \ ni=1 pi . In Qt(R), all prime ideals
correspond to the pi ’s, that is, they are all maximal and their intersection is (0). Therefore,
by [5, Proposition 1.10]
Qt(R) ∼
= Qt(R/ p1 ) × . . . × Qt(R/ pn ),
which is absolutely flat and Noetherian.
Let c = ba ∈ Lσ . Using Theorem 2.15, it suffices to show that R{c}Σ1 is a Σ-simple
Σ-ring, since this would imply that c ∈ K σ . For this, we will show that every Σ-subring
D ⊂ L containing K is Σ-simple. Indeed, for every 0 6= d ∈ L there exists a ∈ R such that
0 6= ad ∈ R, which is true because L is the localization with respect to the set of non-zero
divisors. Therefore, for every nonzero ideal a of D we have a ∩R 6= {0}. Since R is Σsimple, 1 ∈ a.
3.2 Picard–Vessiot pseudofield
Let K be a Noetherian Σ-pseudofield with K σ being Σ1 -closed.
Definition 3.6. A Noetherian Σ-pseudofield L is called a Picard–Vessiot pseudofield for
equation (12) if
1. there is a fundamental matrix F of equation (12) with coefficients in L,
2. Lσ = K σ ,
3. L is Σ1 -generated over K by the entries of F.
It follows from Proposition 3.5 that every equation (12) has a Picard–Vessiot pseudofield. We will show that all Picard–Vessiot pseudofields are of this form.
Proposition 3.7. Let K be a Noetherian Σ-pseudofield, with C := K σ being a Σ1 -closed
pseudofield, and L be a Picard–Vessiot pseudofield for equation (12). Then, L ∼
= Qt(R),
where R is the corresponding Picard–Vessiot ring.
Proof. Let σ act on the Σ1 -ring R := L{Xi j , 1/ detX}Σ1 by σX = AX. Let F be a fundamental matrix of (12) with coefficients in L. Define Y = F −1 X. Then R = L{Yi j , 1/ detY }Σ1
and σY = Y . Therefore,
Rσ = C{Yi j , 1/ detY }Σ1 .
Moreover, we have a Σ-isomorphism
L ⊗K K{Xi j , 1/ det X}Σ1 ∼
= L ⊗C C{Yi j , 1/ detY }Σ1 .
(13)
14
ANTIEAU ET AL.
Recall that the Picard–Vessiot ring is given by R = K{Xi j , 1/ detX}Σ1 /I, where I is a maximal Σ-ideal. By Proposition 2.11 and isomorphism (13), the ideal L ⊗K I corresponds to
a Σ-ideal of the form L ⊗C J, where J is a Σ1 -ideal of C{Yi j , 1/ detY }Σ1 . This induces a
Σ-isomorphism
φ : L ⊗K R → L ⊗C B,
where B = C{Yi j , 1/ detC}Σ1 /J consists of σ-constants. Let m be a maximal Σ-ideal in B.
By [30, Proposition 14], we have
γ : B → B/ m ∼
= C,
since C is a Σ1 -closed pseudofield. Let ϕ be the Σ-homomorphism defined by
r7→1⊗r
φ
idL ⊗γ
l⊗c7→l·c
R −−−−→ L ⊗K R −−−−→ L ⊗C B −−−−→ L ⊗C C −−−−→ L.
Since R is Σ-simple, the homomorphism ϕ is injective. By the universal property, ϕ extends
to a Σ-embedding ϕ of Qt(R) into L. Since L is generated by the entries of its fundamental
matrix F, we finally conclude that ϕ(Qt(R)) = L.
3.3 Difference algebraic groups
3.3.1 Definitions
In analogy with differential algebraic groups [8, 9], we make the following definitions.
Throughout, C will denote a Σ1 -closed pseudofield. Recall that for E ⊆ C{y1 , . . . , yn }Σ1 , the
set V(E) is the set of all common zeroes for elements of E in Cn . The sets of the form V(E),
for some E, are called C-Σ1 -algebraic varieties (also called pseudovarieties in [30]).
Also recall that, for an arbitrary C-Σ1 -algebraic variety X, the set of all difference
polynomials vanishing on X will be denoted by I(X). Every difference polynomial defines
a polynomial function on X. The ring of all polynomial functions, thus, coincides with
C{y1 , . . . , yn }Σ1 / I(X).
Definition 3.8. A regular map f : X → Y of C-Σ1 -algebraic varieties is a map given by
difference polynomials in coordinates (for a general definition see [30, Section 4.6], in
particular, Theorem 40 there).
Definition 3.9. A C-Σ1 -algebraic group is a group supplied with a structure of a C-Σ1 algebraic variety such that the multiplication and inverse maps are regular.
Definition 3.10. A C-Σ1 -Hopf algebra is a C-Σ1 -algebra H supplied with comultiplication,
counit, and antipode morphisms that are all C-Σ1 -algebra morphisms.
Note that the ring of polynomial functions of a C-Σ1 -algebraic variety is a reduced
Hopf algebra such that the comultiplication, antipod, and counit are homomorphisms of
C-Σ1 -algebras.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
15
An example of a C-Σ1 -algebraic group is the group GLm,Σ1 (C), that is the set of all
m× m matrices with coefficients in C and having invertible determinant. The corresponding
ring of regular functions is
Hm = C{x11 , . . . , xmm , 1/detX}Σ1 .
The C-Σ1 -algebra Hm has a Hopf algebra structure defined on the Σ1 -generators in the usual
way and is extended by commuting to the Σ1 -monomials in the generators.
Example 3.11. Let us describe the structure of GL1,Σ1 (C) explicitly. Let
Σ1 = id, ρ, ρ2 , . . . , ρt−1
and consider
H1 = C{x, 1/x}Σ1 = C x, 1/x, ρ(x), 1/ρ(x), . . . , ρt−1 (x), 1/ρt−1 (x) .
Then, the comultiplication is
and the antipode map is
ρl (x) 7→ ρl (x) ⊗ ρl (x),
ρl (x) 7→ 1/ρl (x).
Since C is Σ1 -closed, it is of the form FΣ1 (K) for some algebraically closed field K. Then
the group GL1,Σ1 (C) has a natural structure of a K-algebraic group such that
GL1,Σ1 (C) = GL1 (K)t .
Definition 3.12. A linear C-Σ1 -algebraic group is a closed subgroup in GLm,Σ1 (C), that is
a subgroup given by difference polynomials.
In particular, this means that the C-Σ1 -Hopf algebra H of a linear C-Σ1 -algebraic group
is a quotient of Hm by a radical Σ1 -Hopf-ideal. More explicitly, the above equivalence also
follows from the equivalence of the categories of affine pseudovarieties and the category of
reduced Σ1 -finitely generated algebras [30, Proposition 42].
3.3.2 Difference algebraic subgroups of Gm,Σ1
Example 3.13. In the usual case of varieties over a field k, the
subgroups of Gm
algebraic
are given by equations xl = 1. The corresponding ideal of k x, x−1 is (xl − 1). In the case
of C-Σ1 -groups, where
Σ1 = Z /t1 Z ⊕ . . . ⊕ Z/t p Z =: {id = α1 , . . . , αt },
t := t1 · . . . · t p ,
there are more Σ1 -algebraic subgroups of Gm,Σ1 . Let C be an arbitrary Noetherian Σ1 pseudofield. Let also {e0 , . . . , es−1 } be all indecomposable idempotents of C with αi (e0 ) =
ei−1 , 1 6 i 6 s. Then the Σ1 -Hopf algebra of Gm,Σ1 is
C{x, 1/x}Σ1 = (K × . . . × K)[xα , 1/xα | α ∈ Σ1 ],
16
ANTIEAU ET AL.
where K = C/m for a maximal ideal m of C. We have
C{x, 1/x}Σ1 = e0C{x, 1/x}Σ1 × . . . × es−1C{x, 1/x}Σ1 ,
Ri = eiC{x, 1/x}Σ1 = K[xα , 1/xα | α ∈ Σ1 ].
As we can see, each Ri is a Hopf algebra. Let I be the Σ1 -ideal defining our Σ1 -closed
subgroup of Gm ,Σ1 . Then,
I = e0 I × . . . × es−1I.
For each i, 0 6 i 6 s − 1, the ideal ei I ⊂ Ri is defined by equations
ki,1,α1
xα 1
ki,m,α1
xα 1
k
t
· . . . · xαi,1,α
= 1,
t
..
.
k
t
· . . . · xαi,m,α
= 1.
t
So, if we collect all equations of all ideals ei I, 0 6 i 6 s − 1, we obtain the equations
e0 xk0,1,1 α2 xk0,1,2 · . . . · αt xk0,1,t
= e0 ,
..
.
= es−1 .
es−1 xks−1,m,1 α2 xks−1,m,2 · . . . · αt xks−1,m,t
Applying α−1
i to the equations with ei , 0 6 i 6 s, we can rewrite the above system in the
form
= e0 ,
e0 xk1,1 α2 xk1,2 · . . . · αt xk1,t
..
(14)
.
km,2
km,1
k
m,t
= e0 ,
· . . . · αt x
e0 x α2 x
which generate I as a Σ1 -ideal. The latter equations also give generators of the ideal e0 I.
So, by [33, Section 2.2] we must have m 6 t.
Now we claim that there is an equation in I of the form ϕ(x) − 1 = 0, where ϕ(xy) =
ϕ(x)ϕ(y). Indeed, for this, denote the first equation in (14) by ψ(x)− e0 . Then, the equation
∑
16k6s
αk (ψ(x) − e0 ) =
∑
16k6s
αk (ψ(x)) − 1.
is of the desired form, where the sum ∑16k6s αk (ψ(x)) is multiplicative because the ei ’s
are orthogonal.
Now suppose that s = t (this is the case, for example, when C is Σ1 -closed). In this
case, we know that the number m of equations does not exceed the number s of our idem-
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
17
potents. Then the following system defines the ideal I.
e0 xk1,1 α2 (xk1,2 ) · . . . · αt (xk1,t ) = e0 ,
..
.
(1)
e0 xkm,1 α2 (xkm,2 ) · . . . · αt (xkm,t ) = e0 ,
(m)
e0 = e0 ,
..
.
e0 = e0 .
(m+1)
(t)
Applying αi to the ith equation, 1 6 i 6 t, we obtain
e0 xk1,1 α2 (xk1,2 ) · . . . · αt (xk1,t ) = e0 ,
..
.
(1)
em−1 αm (xkm,1 )(αm α2 )(xkm,2 ) · . . . · (αm αt )(xkm,t−1 ) = em−1 ,
em = em ,
..
.
(m)
(m+1)
et−1 = et−1 .
(t)
By taking the sum of the above equations, we arrive at an equation of the form
ϕ(x) = 1.
(15)
Since the ei ’s are orthogonal, the left-hand side is multiplicative. Moreover, this equation
defines the same subgroup. Vice versa, every multiplicative ϕ(x) ∈ C{x, 1/x}Σ1 defines a
Σ1 -subgroup of Gm,Σ1 via (15). Note that it might happen that the set of solutions is empty.
For example, this is the case for ϕ = e, where e is idempotent and not equal to 1.
Example 3.14. Let C = C × C × C with ρ(a0 , a1 , a2 ) = (a2 , a0 , a1 ), ai ∈ C. By [30, Proposition 15], (C, ρ) is a Σ1 -closed pseudofield. Let
G = {a ∈ C | a · ρ(a) = 1},
(16)
a Σ1 -subgroup of Gm considered in Example 3.13. A calculation shows that G =
{(1, 1, 1), (−1, −1, −1)}. This demonstrates a major difference between Σ1 -subgroups and
differential algebraic subgroups (see [8, Chapter IV]) of Gm . More precisely, in the differential case the order of the defining equation coincides with the algebraic dimension of the
subgroup.
In our case, the order of ρ in (16) is equal to 1, however, the group is finite. Therefore,
in order to compute the algebraic dimension of a Σ1 -group one needs to do more calculation
than just to look at the ρ-order of the equation.
18
ANTIEAU ET AL.
3.4 Galois group
As before, let K be a Noetherian Σ-pseudofield with C := K σ being Σ1 -closed.
Definition 3.15. Let L be a Picard–Vessiot pseudofield of equation (12). Then the group of
Σ-automorphisms of L over K is called the difference Galois group of (12) and denoted by
AutΣ (L/K).
Let L be a Picard–Vessiot pseudofield of equation (12) and F ∈ GLn (L) be a fundamental matrix. Then for any γ ∈ AutΣ (L/K) we have
γ(F) = FMγ ,
(17)
where Mγ ∈ GLn (C), which, as usual, defines an injective group homomorphism from
AutΣ (L/K) into GLn (C). Since L is generated by the entries of F, the action of γ on L is determined by its action on F. This induces an identification of AutΣ (L/K) with AutΣ (R/K),
where R is the Picard–Vessiot ring corresponding to F.
We will now construct a map AutΣ (R/K) → MaxΣ (R ⊗K R), the maximal Σ-ideals of
R ⊗K R. For this, let F be a fundamental matrix of equation (12) with entries in R and
γ ∈ AutΣ (R/K). As above, γF = FMγ , where Mγ ∈ GLn (C). We will then map
γ 7→ [F ⊗ 1 − 1 ⊗ FMγ ]Σ ,
the smallest Σ-ideal containing F ⊗ 1 − 1 ⊗ FMγ . Since R is Σ-simple, the kernel of the
surjective Σ-homomorphism
(γ, Id) : R ⊗K R → R,
which is [F ⊗ 1 − 1 ⊗ FMγ ]Σ , is a maximal Σ-ideal in R ⊗K R.
To construct a map in the reverse direction, let
φ1 , φ2 : R → R ⊗K R,
with r 7→ r ⊗ 1 and r 7→ 1 ⊗ r, respectively. Let m be a maximal Σ-ideal of R ⊗K R. Then,
(R ⊗K R)/ m is a Picard–Vessiot ring of equation (12). As in Proposition 3.4, the composition homomorphisms
φi : R → R ⊗K R → (R ⊗K R)/ m
are isomorphisms. This induces an automorphism of the ring R defined by
−1
φm := φ2 ◦ φ1 .
Proposition 3.16. The correspondence AutΣ (R/K) → MaxΣ (R ⊗K R) constructed above
is bijective. Moreover, these bijections are inverses of each other.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
19
Proof. Let γ ∈ AutΣ (R/K) and M ∈ GLn (C) be such that γ(F) = FM. Set m = [F ⊗ 1 −
1 ⊗ FM]Σ . Since
φ1 (F) = F ⊗ 1,
F ⊗ 1 = 1 ⊗ FM in (R ⊗K R)/ m, and φ2 (FM) = 1 ⊗ FM,
we have φm (F) = FM.
Conversely, let m ∈ MaxΣ (R⊗K R). Then φm (F) = FM for some M ∈ GLn (C). Hence,
φ1 (F) = φ2 (FM).
Thus,
[F ⊗ 1 − 1 ⊗ FM]Σ ⊂ m .
Since, as above, the former ideal is Σ-maximal, it coincides with m.
Proposition 3.17. The Galois group G of equation (12) is a closed subgroup of GLn (C).
Moreover, if the ring R ⊗K R is reduced, then
R ⊗K R ∼
= R ⊗C C{G},
where C{G} is the ring of regular functions on G and R is a Picard–Vessiot ring of (12).
Proof. As before, define σ on the Σ1 -ring R{Xi j , 1/ detX}Σ1 by σX = AX. Let F be a
fundamental matrix of (12) with coefficients in R and let, as above, Y = F −1 X, which
implies that σY = Y . We have a Σ-isomorphism
R ⊗K K{Xi j , 1/ det X}Σ1 ∼
= R ⊗C C{Yi j , 1/ detY }Σ1 .
As in the proof of Proposition 3.7, this induces a Σ-isomorphism
R ⊗K R ∼
= R ⊗C B,
(18)
where B = C{Yi j , 1/ detY }Σ1 /J and J is a Σ1 -ideal.
By Proposition 3.16, AutΣ (R/K) as a set can be identified with MaxΣ (R ⊗K R). The
latter set, by Proposition 2.11 and isomorphism (18), can be identified with MaxΣ1 B. Since
C is Σ1 -closed, by [30, Proposition 14], the set MaxΣ1 B can be identified with a closed
subset of GLn (C). The group structure of G is preserved under this identification due
to (17). If the ring R ⊗K R is reduced, then the ideal J is radical and, therefore, B is the
coordinate ring of G.
3.5 Galois correspondence
Proposition 3.18. Let L be a Picard–Vessiot pseudofield of equation (12), R be its Picard–
Vessiot ring, and G be its Galois group. If the ring R ⊗K R is reduced, then LG = K.
20
ANTIEAU ET AL.
Proof. Let
a/b ∈ L \ K,
(19)
where a, b ∈ R and b is not a zero divisor. Set
d = a ⊗ b − b ⊗ a ∈ R ⊗K R.
We will show that d 6= 0. For this, let {e1 , . . . , en } be all indecomposable idempotents of
the Noetherian Σ-pseudofield K. Since b is not a zero divisor,
ei · b 6= 0,
1 6 i 6 n.
(20)
Suppose that for each i, 1 6 i 6 n, ei · a and ei · b are linearly dependent over ei K, that is,
λi · ei · a = µi · ei · b for all i. Then (20) implies that λi 6= 0, 1 6 i 6 n. Since ei K is a field,
we have ei · a = ei · b · µi/λi . Hence,
!
n
n
n
a
µi
µi
= ∑ ei ∈ K,
a = ∑ ei a = ∑ ei b, that is,
b i=1 λi
i=1
i=1 λi
which is a contradiction to (19). Therefore, there exists i, 1 6 i 6 n, such that ei · a and ei · b
are linearly independent over ei K. Then,
ei · a ⊗ ei · b − ei · b ⊗ ei · a 6= 0
in ei R ⊗ei K ei R. Hence,
a ⊗ b − b ⊗ a 6= 0 in
R ⊗K R.
We will now show that there is a maximal Σ-ideal in R ⊗K R that does not contain d. Since
R ⊗K R is reduced, then by Proposition 3.17 we have R ⊗K R ∼
= R ⊗C C{G}. Let {li }i∈I be
a basis of R over K. Then there exist r1 , . . . , rm ∈ C{G} such that
d = l1 ⊗ r1 + . . . lm ⊗ rm .
Since r1 is not nilpotent, there exists a maximal Σ1 -ideal m ⊂ C{G} such that r 1 6= 0 in
C{G}/ m. Then image of d in R ⊗C C{G}/ m ∼
= R is
d = l1 r1 + . . . + lm r m .
Since r1 6= 0, we have d 6= 0. Thus, d ∈
/ R ⊗C m. Using the correspondence between
maximal Σ-ideals in R ⊗K R and Σ-automorphisms of R over K, let
−1
φm = φ2 ◦ φ1
correspond to m as in the proof of Proposition 3.16. Then our choice of m implies that
(R ⊗K R)/ m ∋ φ1 (a)φ2 (b) − φ1 (b)φ2 (a) 6= 0.
−1
Applying
φ2
φm ba 6= ba .
(21)
to both sides of (21), we obtain that φm (a)b − φm (b)a 6= 0. Therefore,
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
21
Lemma 3.19. Let K ⊂ L be Noetherian Σ-pseudofields. Let H ⊂ AutΣ (L) such that LH = K.
Suppose that K ∼
= ∏ni=1 FΣ1 (F) as Σ1 -rings, where F is a field. Let {ei } be the corresponding idempotents. Then for each i the abstract group generated by Σ1 and H acts transitively
on the set of indecomposable idempotents of the ring ei L.
Proof. Let e ∈ ei L be an idempotent and S be its Σ1 ∗ H-orbit. The set S coincides with the
set of indecomposable idempotents if and only if ∑ f ∈S f = 1. This sum is H-invariant and,
therefore, it belongs to FΣ1 (F). Since it is Σ1 -invariant as well, it is equal to 1, because a
Σ1 -invariant idempotent of FΣ1 (F) generates a Σ1 -ideal.
Proposition 3.20. Let L be a Picard–Vessiot pseudofield for equation (12) and H be a
closed subgroup of the Galois group G. Then LH = K implies H = G.
Proof. As before, let F be a fundamental matrix with entries in L and σX = AX define the
action of Σ on the Σ1 -ring D := L{Xi j , 1/ detX}Σ1 . Let also Y = F −1 X. Again, as before,
L ⊗K K{Xi j , 1/ det X}Σ1 ∼
= L ⊗C C{Yi j , 1/ detY }Σ1 .
Suppose that H ( G and let I ( J be the defining ideals of G and H, respectively. Denote
their extensions to L{Xi j , 1/ detX} by (I) and (J), respectively. By Proposition 2.11, we
have (I) ( (J). Explicitly, we have
(I) = f (X) ∈ L{Xi j , 1/ detX}Σ1 | f (FM) = 0 for all M ∈ G
and
(J) = f (X) ∈ L{Xi j , 1/ detX}Σ1 | f (FM) = 0 for all M ∈ H .
(22)
Let T = (J) \ (I) 6= ∅. Define the action of H on L ⊗K K{Xi j , 1/ detX}Σ1 by
h(a ⊗ b) = h(a) ⊗ b,
h ∈ H.
Then, equality (22) implies that (J) is stable under this action of H. By Proposition 2.19,
K∼
= FΣ1 (F) × . . . × FΣ1 (F)
as Σ1 -rings, where F is a field. Let e1 , . . . , en be the idempotents corresponding to the
components FΣ1 (F) in the above product. By Proposition 2.10, the ring K{Xi j , 1/ detX}Σ1
has a Σ1 -invariant basis {Qα }. Then every element of the ring D is of the form
Q = q 1 Qα 1 + . . . + q n Qα n ,
(23)
where qi ∈ L, 1 6 i 6 n. Let Q be an element in T with the shortest presentation of the
form (23). Since Q = ∑i ei Q, there exists i such that ei Q ∈ T . Denote the latter polynomial
by Q as well. Now, we have Q ∈ ei D. Let { f1 , . . . , fm } be all indecomposable idempotents
of the Noetherian ring ei L. Then,
Q = ∑ j=1 f j Q.
m
22
ANTIEAU ET AL.
Hence, there exists j such that f j Q ∈ T . By Lemma 3.19, there exist ht ∈ Σ1 ∗ H such that
the coefficients of
Q′ := ∑ ht (Q)
t
are invertible in ei L. Therefore,
Q′ = e i Q1 + g 2 Q2 + . . . + g m Qm .
Since the ideal (J) is stable under the action of Σ1 ∗ H, we have Q′ ∈ T . Since ei ∈ K,
for every h ∈ H the polynomial Q′′h := Q′ − h(Q′ ) has a shorter presentation than Q and,
therefore, Q′′h ∈
/ T . That is,
Q′′h ∈ (I) for all h ∈ H.
(24)
We will show now that Q′′h = 0 for all h ∈ H. Suppose that Q′′h 6= 0 for some h ∈ H. Then (24)
implies that there exists j such that 0 6= f j Q′′h ∈ (I). Since Σ1 ∗ H acts transitively on the
indecomposable idempotents of ei L, there exist φt ∈ Σ1 ∗ H such that
Qh := ∑ φt Q′′h = r2 Q2 + . . . + rm Qm ∈ (I),
t
where r2 is invertible in ei L. Therefore, there exists r ∈ ei L such that g2 = rr2 . Then,
the polynomial Q′ − rQ ∈ T has a shorter presentation than Q′ , which is a contradiction.
We have shown that h(Q′ ) = Q′ for all h ∈ H. Hence, all coefficients of Q′ are in K and,
therefore, are invariant under the action of G as well. Since 0 = Q′ (F · id) = Q′ (F), we
have
0 = g(Q′ (F)) = g(Q′ )(FMg ) = Q′ (FMg )
for all g ∈ G. Thus, Q′ ∈ (I), which contradicts to Q′ ∈ T .
Lemma 3.21. Let M be a field,
D := M × . . . × M,
(25)
F ⊂ D be a subfield and H ⊂ Aut(D) with DH = F. Let f := (1, 0, . . . , 0) ∈ D and H1 ⊂ H
be the stabilizer of f . Then, f F = M H1 , where M is from the first component in (25).
Proof. Since f F is H1 -invariant, we have f F ⊆ M H1 . We will show the reverse inclusion.
Let l ∈ ( f D)H1 = M H1 . We need to show that there is an element a ∈ F such that l = f a.
Let the H-orbit of l be {l1 , . . . , lk }, where l = l1 . For each i, 1 6 i 6 k, there exists ai ∈ D
such that li = ai fi , where fi is the idempotent corresponding to the ith factor in D (so we
have f = f1 ), since if l 6= 0, then H1 is the stabilizer of l. Hence, for d = ∑ki=1 li we have
k
f d = ∑ f1 li = l1 = l
i=1
and H permutes the li ’s. Thus, d ∈ DH = F as desired.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
23
Proposition 3.22. Let K be a Noetherian Σ-pseudofield, R be a Σ-simple Noetherian alge-
bra over K, and L = Qt(R). Then for the statements
1. the ring R ⊗K R is reduced,
2. the ring L ⊗K L is reduced,
3. there exists a subgroup H ⊂ AutΣ (L/K) such that LH = K,
we have: (1) is equivalent to (2) and (3) implies (2). Moreover, if R is a Picard–Vessiot
ring over K, then the above statements are equivalent.
Proof. The equivalence of (1) and (2) follows from the fact that R ⊗K R ⊂ L ⊗K L and that
the latter ring is a localization of the former one. We will show that (3) implies (2). Let
{e1 , . . . , en } be the indecomposable idempotents of K. Then,
n
L ⊗K L = ∏ ei L ⊗ei K ei L.
i=1
It is enough to show that the ring ei L ⊗ei K ei L is reduced. Note that ei K is a field. Since ei ∈
K, they are all invariant under H and, moreover, (ei L)H = ei K. Let now { f1 , . . . , fm } be the
indecomposable idempotents of the ring ei L and let H1 be the stabilizer of f1 . Lemma 3.21
with D = ei L and F = ei K implies that
(ei f1 L)H1 = f1 ei K.
Since
ei L ⊗ei K ei L = ∏ ei fs L ⊗ei K ei ft L,
s,t
it remains to show that the ring
D := ei fs L ⊗ei K ei ft L
is reduced. By [6, Corollary 1, §7, no. 2], with A = ei fs L, B = ei ft L, N = B, and K = ei K,
the Jacobson radical of the ring D is zero. In particular, the ring D is reduced.
The last statement follows from Proposition 3.18.
Definition 3.23. A Picard–Vessiot extension L/K is called separable if one of the three
equivalent conditions in Proposition 3.22 is satisfied.
Theorem 3.24. Let R be a Picard–Vessiot ring of equation (12) and L = Qt(R) be separable
over K. Let F denote the set of all intermediate Σ-pseudofields F such that L is separable
over K and G denote the set of all Σ1 -closed subgroups H in the Galois group G of L over
K. Then the correspondence
F ←→ G , F 7→ AutΣ (L/F), H 7→ LH
is bijective and the above maps are inverses of each other. Moreover, H is normal in G if
and only if the Σ-pseudofield F := LH is G-invariant.
24
ANTIEAU ET AL.
F −→ G is well-defined by Proposition 3.17. Propositions 2.8 and 2.9
imply that LH ⊂ L is a Σ-pseudofield. By Proposition 2.20, it is Noetherian and, by Proposition 3.22, it is separable.
Let F ∈ F . Then the extension L over F is separable and is a Picard–Vessiot pseudofield for equation (12) considered over F. Moreover, F = F AutΣ (L/F) by Proposition 3.18.
Conversely, let H be a Σ1 -closed subgroup of G. Set F = LH . Then L is a Picard–Vessiot
pseudofield for equation (12) over F. By Proposition 3.20, we have H = AutΣ (L/F). The
equality
g(F) = r ∈ L | ghg−1r = r for all h ∈ H
Proof. The map
implies the statement about normality.
Remark 3.25. The base pseudofield K is a product of the fields, say L × . . . × L. If the field
L is perfect, then for every pseudofields F and E containing K the ring F ⊗K E is reduced.
Indeed, let e0 , . . . , et−1 be all indecomposable idempotents of K, then
t−1
F ⊗K E = ∏ ei F ⊗L ei E.
i=0
Since L is perfect and L-algebras ei F and ei E are reduced, then ei F ⊗L ei E is reduced as well
(see [7, A.V. 119, No. 5, Théorèm 3(d)]). Therefore, if L is perfect, then any Picard–Vessiot
extension is separable. If the field L is finite, algebraically closed or of characteristic zero,
then L is perfect. In this case, the set F contains all intermediate Σ-pseudofields.
3.6 Extension to non-faithful action
In the introduction, we alluded to the fact that some of our results could instead be
obtained via faithfully flat descent [26]. However, this requires the additional assumption2
that Σ1 acts faithfully in the setup for faithfully flat descent. The theory we have developed
in this paper works more generally, as we now illustrate. Let L be a field, Σ1 = Z/4Z, and
2
ρ be a generator. Suppose that Σ1 acts on L faithfully and K = Lρ . Then
Autρ (L/K) ∼
= Z/2Z but Aut LΣ1 /K Σ1 = {1},
where Autρ is the set of ρ-automorphisms (so, we also have to store the order of ρ – not
only the field of invariants). Roughly speaking, replacing a difference object by a nondifference one, we cannot recover the group of difference automorphisms. This example
(where L is a Picard–Vessiot extension and K is a base field) appears naturally in the Picard–
Vessiot theory if, for instance, the initial field contains only ρ-constants. Also note that if
we replace L by LΣ1 , then the σ-constants of LΣ1 are not necessarily algebraically closed.
2 We also assume this in Section 3, but it is only needed to guarantee the uniqueness of a parameterized PVextension, which we do not use in the applications. It also implies the existence, but is only a sufficient condition,
and there are situations when one does not have to make this assumption. The Galois correspondence in its
Hopf-algebraic version does not need this as we show below.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
25
Therefore, using descent, extra preparatory steps are required before we are able to apply
the standard non-parametric difference Galois theory.
Here is another example in which it is preferable to consider not necessarily faithful
actions of Σ1 . Let
L = C(x), σ(x) = 2 · x, ρ(x) = −x, K = C x2 , Σ1 = {id, ρ}
and the difference equation be
σ(y) = 2 · y.
(26)
There are no solutions of (26) in K and L is a Picard–Vessiot extension of the equation,
on which Σ1 acts faithfully, while the action of Σ1 on K is trivial. Of course, the field L
considered with the trivial action of ρ is also a Picard–Vessiot extension, but restricting
to this would not allow us to consider more interesting and useful cases outlined in this
example. We will show how one can generalize our results to include non-faithful actions
of Σ1 .
Theorem 3.26 ((Instead of Theorem 3.24)). Let R be a Picard–Vessiot ring over a pseud-
ofield K with the corresponding Σ1 -Hopf-algebra H (which replaces the Galois group –
see [3, Section 2]) and L = Qt(R) be the pseudofield of fractions. Let F denote the set of
all intermediate Σ-pseudofields and G denote the set of all Σ1 -Hopf-ideals in H. Then the
correspondence
G → F , I 7→ LI := {x ∈ L | 1 ⊗ x − x ⊗ 1 ∈ I · (L ⊗K L)}
is bijective. Moreover, I is normal in H if and only if the Σ-pseudofield LI is H-invariant.
A Σ1 -Hopf-ideal I is radical if and only if L is separable over LI .
In order to prove this result, one extends the Hopf-algebraic approach given in [2].
The main technical results one uses are [2, Proposition 3.10] and
Proposition 3.27 ((Instead of Proposition 2.11)). Let R be a Picard-Vessiot ring over K and
R ⊗K R ∼
= R ⊗C H
be the torsor isomorphism for R, where H is a Σ1 -Hopf-algebra over Σ1 -pseudofield C =
Rσ . Then this isomorphism induces a 1 − 1 correspondence between Σ1 -Hopf-ideals of H
and Σ-coideals of R ⊗K R. That is, if I ⊆ R ⊗K R is a Σ-coideal, then I ∩ H is a Σ1 -Hopfideal; if a is a Σ1 -Hopf-ideal, then R ⊗C a is a Σ-coideal.
From the Hopf-algebraic point of view, the Galois correspondence can be derived from
Theorem 3.26 above using the following result:
Theorem 3.28. Let H be a reduced Σ1 -Hopf-algebra over a difference closed pseudofield
K. Then H induces a functor F from the category of Σ1 -K-algebras to the category of
groups by the rule
F(R) = homΣ1 −K (H, R).
Then H can be recovered from F(K).
26
ANTIEAU ET AL.
3.7 Torsors
Let C be a Σ1 -closed pseudofield and K ⊃ C be a Noetherian Σ-pseudofield. Let G
be a Σ1 -group over C be C{G} be its Σ1 -Hopf algebra with comultiplication ∆, antipode S,
and counit ε.
Definition 3.29. A Σ1 -finitely generated K-algebra R supplied with a Σ-K-algebra homo-
morphism
ν∗ : R → R ⊗C C{G}
is called a G-torsor over K if the following statements are true:
1. R is a C{G}-comodule with respect to ν∗ ,
2. the vertical arrow in the following diagram is an isomorphism:
R ⊗K Rf◆
◆1⊗id
◆◆◆ R
♣
♣♣
R
R ▼▼
▼▼▼
qqq
xqq1⊗id
ν∗ &
R
C{G} ⊗C R
idR ⊗1♣♣8
In the above notation, the rings R and C{G} are finitely generated algebras over Artinian rings. Then the Krull dimension is defined for them, which we will denote by dim R
and dimC{G}, respectively. The isomorphism in (2) implies that dim R = dimC{G}.
Moreover, let e be an indecomposable idempotent in C and F := eC be the corresponding residue field. Then F ⊗C C{G} is a finitely generate F-algebra of dimension equal to
dimC{G}. Hence, for any minimal prime ideal p of the ring F ⊗C C{G},
tr. deg.F k(p) = dimC{G} = dim R,
where k(p) is the residue field of p.
Proposition 3.30. Let K be a Noetherian Σ-pseudofield with K σ being a Σ1 -closed pseud-
ofield. Let R be a Picard–Vessiot ring for equation (12) with L = Qt(R). Let G be the
Galois group of L over K. If R is separable over K, then R is a G-torsor over K.
Proof. Follows from Proposition 3.17.
4 APPLICATIONS
We will start by giving a difference dependence statement in the spirit of [21], which
we prove using our own methods. We then show how this is related to Jacobi’s thetafunction in Section 4.3 and demonstrate our applications in Section 4.4, in particular, in
Theorem 4.9, which provides a very explicit difference dependence test.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
27
4.1 General approach
For any nonzero complex number a we define an automorphism σa : C(z) → C(z) by
σa ( f )(z) = f (az).
Let Σ1 ⊆ C∗ be a finite subgroup. Then Σ1 is a cyclic group generated by a root of unity
ζ of degree t. Let q ∈ C be a complex number such that |q| > 1. Now we have an action
of the group Σ = Z ⊕ Z /t Z on C(z), where the first summand is generated by σq and the
second one is generated by σζ . Throughout this section the ring C(z) is supplied with this
structure of a Σ-ring.
Theorem 4.1. Let R be a Σ-ring containing the field C(z) such that k := Rσq is a field.
Suppose additionally that R contains the field k(z). Let f ∈ R and a ∈ C(z) be such that f
is an invertible solution of
σq ( f ) = a f .
(27)
Then f is σζ -algebraically dependent over k(z) if and only if
ϕ(a) = σq (b)/b
(28)
(x)
for some 0 6= b ∈ C(z) and 1 6= ϕ(x) = xn0 σζ (x)n1 · . . . · σt−1
ζ
nt−1
.
Proof. If (28) holds, then
σq (ϕ( f )/b) = ϕ(σq ( f ))/σq (b) = ϕ(a f )/σq (b) = ϕ(a)ϕ( f )/σq (b) = ϕ( f )/b.
Therefore,
ϕ( f )/b = c ∈ Rσq = k .
Thus, ϕ( f ) = c · b ∈ k(z), which gives a Σ1 -algebraic dependence for f over k(z). First,
note that z is algebraically independent over k. Indeed, suppose that there is a relation
an · zn + an−1 · zn−1 + . . . + a0 = 0
for some ai ∈ k. Applying σq n times, we obtain the following system of linear equations
1
1
... 1 1
an · zn
n−1
qn
qn−1
. . . q 1
an−1 · z
..
=0
..
.
.
.
..
..
.
. .. ..
.
(qn )n
n
(qn−1 )
...
qn
1
a0
Since the matrix is invertible, our relation is of the form a · zk = 0 for some a ∈ k. Since k
is a field, we have zk = 0. However, z ∈ C(z), which is a contradiction.
Assume now that f is Σ1 -algebraically dependent over k(z). Let C be the Σ1 -closure
of k and K be the total ring of fraction of the polynomial ring C[z], where σq (z) = qz and
28
ANTIEAU ET AL.
σζ (z) = ζz. So, the field k(z) is naturally embedded into K. Let D be the smallest Σ-subring
in R generated by k(z), f , and 1/ f and let
m ⊆ K ⊗k(z) D
be a maximal Σ-ideal. Then,
L = K ⊗k(z) D / m
is a Picard–Vessiot ring over K for equation (27). The image of f in L will be denoted by
f¯. Since f is Σ1 -algebraically dependent over k(z), f¯ is Σ1 -algebraically dependent over
K.
It follows from Section 3.7 that f¯ is Σ1 -algebraically dependent over K if and only the
Σ-Galois group G of equation (27) is a proper subgroup of Gm ,Σ1 . Then, by Example 3.13,
there exists a multiplicative
ϕ ∈ (FΣ1 Q){x, 1/x}Σ1
(see also (9)) such that G is given by the equation ϕ(x) = 1. Therefore, for all φ in the
Galois group,
φ(ϕ( f¯)) = ϕ(φ( f¯)) = ϕ(cφ · f¯) = ϕ(cφ ) · ϕ( f¯) = 1 · ϕ( f¯) = ϕ( f¯).
Hence, by Proposition 3.18, we have
b := ϕ( f¯) ∈ K = C(z),
as in [22, Proposition 3.1]. Since f is invertible, f¯ is also invertible and, since ϕ is multiplicative, ϕ( f¯) is invertible as well. Therefore,
(29)
ϕ(a) = ϕ σq ( f¯)/ f¯ = σq ϕ( f¯) /ϕ( f¯) = σq (b)/b.
We will show now that b can be chosen from (FΣ1 C)(z) satisfying (28) as in [22, Corollary 3.2]. We have the equalities a = ā/c and b = b̄/d, where ā, c ∈ C[z] and b̄, d ∈ C[z].
Consider the coefficients of b̄ and d as difference indeterminates. Then, equation (29) can
be considered as a system of equations in the coefficients of b̄ and d. Indeed, equation (29)
is equivalent to
b̄/d .
ϕ(ā/c) = σq b̄/d
So, we have
ϕ(ā) · σq (d) · b̄ − ϕ(c) · σq(b̄) · d = 0
(30)
The left-hand side of equation (30) is a polynomial in z. The desired system of equations
is given by the equalities for all coefficients. Now note that the condition of y ∈ C[z] being
invertible in C(z) is given by the inequation
y · σζ (y) · . . . · σt−1
ζ (y) 6= 0.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
29
Therefore, the coefficients of the polynomials b̄ and d are given by the system of equations
and inequalities. Since the pseudofield FΣ1 C is Σ1 -closed, existence of invertible b̄ and
d with coefficients in C implies existence of invertible b̄ and d with coefficients in FΣ1 C
(see [30, Proposition 25 (3)]).
We will now show that b ∈ C(z) and ϕ can be found of the desired form. We have
proven that
ϕ(a) = σq (b)/b
(31)
for some b ∈ (FΣ1 C)(z). It follows from Example 3.13 that
(x)n0,t−1 + . . . +
ϕ(x) =e0 · xn0,0 · σζ (x)n0,1 · . . . · σt−1
ζ
(x)nt−1,t−1 .
+ et−1 · xnt−1,0 · σζ (x)nt−1,1 · . . . · σt−1
ζ
Note that if a ∈ (FΣ1 C)(z) belongs to C(z), then
γe (a) = a
and γe (σiζ (a)) = σiζ (γe (a)),
where the σq -homomorphism γe : (FΣ1 C)(z) → C(z) is defined in (10). Applying this homomorphism to (31), we obtain
(a)n0,t−1 = σq (γe (b))/γe (b),
an0,0 · σζ (a)n0,1 · . . . · σt−1
ζ
which concludes the proof.
4.2 Setup for meromorphic functions
The ring of all meromorphic functions on C∗ will be denoted by M . For any nonzero
complex number a we define an automorphism σa : M → M by
σa ( f )(z) = f (az).
Let Σ1 ⊆ C∗ be a finite subgroup. Then Σ1 is a cyclic group generated by a root of unity ζ of
degree t. Let q ∈ C be such that |q| > 1. Now, we have an action of the group Σ = Z⊕ Z/tZ,
where the first summand is generated by σq and the second one is generated by σζ .
The set of all σq -invariant meromorphic functions will be denoted by k. As we can
see k is a Σ1 -ring. Let C be the Σ1 -closure of the field k. Supply the polynomial ring C[z]
with the following structure of a Σ-ring:
σq (z) = qz and σζ (z) = ζz.
Let K be the total ring of fractions of C[z], so, K is a Noetherian Σ-pseudofield with Σ1 closed subpseudofield of σq -constants C. The meromorphic function z is algebraically
independent over k. Hence, the minimal Σ-subfield in M generated by k and z is the ring
of rational functions k(z). Thus, this field can be naturally embedded into K with z being
mapped to z.
30
ANTIEAU ET AL.
4.3 Jacobi’s theta-function
We will study Σ1 -relations for Jacobi’s theta-function (being a solution of θq (qz) =
−qz · θq(z))
θq (z) = − ∑ (−1)n q
−n(n−1)
2
zn ,
n∈Z
z ∈ C,
(32)
with coefficients in k(z).
4.3.1 Relations for θq with q-periodic coefficients
First, we will show that there are many relations of such form:
1. Suppose that t > 3. Then, the function
2
λ = θq (z) · θ−2
q (ζz) · θq (ζ z)
is σq -invariant. Therefore, θq vanishes the following nontrivial Σ1 -polynomial:
y · σζ2 (y) − λ · (σζ (y))2 ∈ k(z){y}.
2. Suppose that t = m · n, where m and n are coprime. Then, there exist two numbers
vn
u 6= v such that the automorphisms σuζ 6= σvζ but σun
ζ = σζ 6= id. Then, the function
v
λ = θnq (ζu z) · θ−n
q (ζ z)
is σq -invariant. Therefore, θq vanishes the following nontrivial Σ1 -polynomial:
(σζu (y))n − λ · (σζv (y))n ∈ k(z){y}.
3. For any given ζ, the function
λ = θtq (z) · θ−t
q (ζz)
is σq -constant. Therefore, θq vanishes the following nontrivial Σ1 -polynomial:
yt − λ · (σζ(y))t ∈ k(z){y}.
4.3.2 Periodic difference-algebraic independence for θq with q-periodic coefficients
We will show now that in some sense these relations are the only possible ones.
Lemma 4.2. Suppose that for some rational function b ∈ k(z) there is a relation
k
(−qz)k0 (−qζz)k1 · . . . · −qζt−1 z t−1 = σq (b)/b
t−1
for some ki ∈ Z. Then, ∑ ki = 0.
i=0
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
31
Proof. The function σq (b)/b is of the following form σq (b)/b = h/g, where h and g have
the same degree and the same leading coefficient. The equality follows from the condition
on the degree.
Lemma 4.3. Suppose that there exist λ ∈ k(z) and η, q ∈ C such that σq (λ) = η · λ, where
|η| = 1 and 0 6= q is not a root of unity. Then, λ ∈ k and η = 1.
Proof. Let
λ = a · zr ·
(z − a1) · . . . · (z − an)
(z − b1) · . . . · (z − bm)
be the irreducible representation of λ, where ai , bi ∈ k, the algebraic closure of k. By the
hypothesis, we have
qr+n−m ·
(z − a1) · . . . · (z − an)
(z − a1/q) · . . . · (z − an /q)
= η·
.
(z − b1/q) · . . . · (z − bm /q)
(z − b1) · . . . · (z − bm)
Therefore, qr+n−m = η. Thus, r + n − m = 0 and η = 1. Moreover, the sets {a1 , . . . , an } and
{a1 /q, . . . , an /q} must coincide. If λ ∈
/ k, then, from r + n − m = 0, it follows that either
n > 0 or m > 0. Suppose that the first inequality holds. There exists i such that a1 = aqi .
If i = 1, then we set i0 = 1. Otherwise, i > 1 and, rearranging the elements {a j } for j > 1
suppose that i = 2. Again, a2 = aqi . If i = 1, we set i0 = i. Otherwise, i > 2 and rearranging
the elements {a j } for j > 2, suppose that i = 3, and so on. Since there are only finitely
many elements, the process will stop and we obtain a number i0 with the following system
of equations:
a1 = a2 /q, a2 = a3 /q, . . . , ai0 = a1 /q
Therefore, qi0 = 1. Thus, |q| = 1, which is a contradiction.
Proposition 4.4. Let the pseudofield K be as above. Let R be a Picard–Vessiot ring over
K for the equation σq (y) = −qz · y and L be the corresponding Picard–Vessiot pseudofield.
Suppose f is an invertible solution in R. Then L ⊗K R is a graded ring such that f is of
degree 1 and σq and σζ preserve the grading.
Proof. It follows from Proposition 3.17 that R ⊗K R = R ⊗C C{G}, where G is the corre-
sponding Galois group. Multiplying by L ⊗R −, we obtain: L ⊗K R = L ⊗C C{G}. Since
group G is a subgroup of Gm ,
C{G} = C{x, 1/x}Σ1 /J,
where the ideal J is generated by difference polynomials of the form
e0 · xk0 · σζ x
k1
kt−1
· . . . · (σt−1
− e0
ζ x)
32
ANTIEAU ET AL.
(see Example 3.13 for details). The ring C{x, 1/x}Σ1 is a graded ring such that x is homogeneous of degree 1 and σζ preserves the grading. In the proof of Theorem 4.1, we have
obtained that
k
(−qz)k0 · (−qζz)k1 · . . . · −qζt−1 z t−1 = σq (b)/b
for some b ∈ C(z). Thus, it follows from Lemma 4.2 that
t−1
∑ ki = 0.
i=0
Therefore, the ideal J is homogeneous. Hence, C{G} is graded. Thus, L ⊗C C{G} is
graded. Since f = f¯ · y, where f¯ ∈ L is a solution of the equation in L, then f is a homogeneous element of degree 1. Since x is σq -constant, σq preserves the grading.
Theorem 4.5. For every prime number t, every relation of the form
λ0 +
t−1
∑
d=1
d
= 0,
λ0d · θq (z)d + λ1d · θq (ζz)d + . . . + λt−1d · θq ζt−1 z
(33)
with λ0 , λi j ∈ k(z), implies that λ0 = λi j = 0.
Proof. Let the pseudofield K be as above, R be a Picard–Vessiot ring over K for the equation
σq (y) = −qz · y, and L be the corresponding Picard–Vessiot pseudofield for R. It follows
from Proposition 4.4 that D = L ⊗K R is a graded ring such that the image of θq in D is
homogeneous of degree 1. Suppose now that θq satisfies an equation of the form (33).
Then, the same equation holds in R and, after embedding R into D, it holds in D. Since D
is graded, our equation is homogeneous. Thus, it is of the form
d
λ0 · θq (z)d + λ1 · θq (ζz)d + . . . + λt−1 · θq ζt−1 z = 0
for some d. Consider the shortest equation and rewrite it as follows
d
θq (z)d + λr · θq (ζr z)d + . . . + λt−1 · θq ζt−1 z = 0,
where λr · θq (ζr z)d is the first nonzero summand immediately following θq (z)d . Applying
σq and dividing by (−qz)d , we obtain
θq (z)d + σq (λr ) · (ζr )d · θq (ζr z)d + . . . + σq (λt−1 ) · ζt−1
Therefore,
σq (λr ) = ζ−rd · λr .
d
d
· θq ζt−1 z = 0
Now, it follows from Lemma 4.3 that ζ−rd = 1, contradiction. Thus, θq (z)d = 0 must hold,
but Picard–Vessiot pseudofield is reduced, which is a contradiction again.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
33
4.3.3 Difference-algebraic independence for θq over C(z)
We will now show difference-algebraic independence for θq over C(z).
Example 4.6. Consider an equation
F(θq ) = ∑(n
1 ,...,n p )∈Z
p
gn1 ,...,n p (z) · θq (α1 z)n1 · . . . · θq (α p z)n p = 0,
where gn1 ,...,n p ∈ C(z) and 1 6= αi 6= α j in C∗ /qZ . We will show that all gn1 ,...,n p are equal
to zero. Since the sum is finite, there exists a monomial
M(θq ) = θq (α1 z)d1 · . . . · θq (α p z)d p
such that M(θq ) · F(θq ) contains monomials with negative powers. Now, we will calculate
the poles of a given monomial with negative powers. The poles of the i-th factor of the
monomial
1
1
M(θq ) =
· ... ·
.
θq (α1 z)n1
θq (α p z)n p
r
are α−1
i q for all r ∈ Z and the multiplicity of each of the poles is ni . The poles of distinct
factors are distinct. Indeed, suppose that
r1
−1
r2
α−1
i · q = αj · q .
Then, α j = αi · qr2 −r1 . Therefore, αi = α j in C∗ /qZ , which is a contradiction. Thus, the
r
set of all poles of the monomial M(θq ) is α−1
i · q with multiplicity ni .
Every function g ∈ C(z) has only finitely many poles and zeros, so, all of them are
inside of a disk Ud = {z ∈ C | |z| < d}. So, the set of all poles for M(θq ) and g · M(θq )
coincides in C \Ud for some d. There exists a disk Ud such that this property holds for all
summands in F. We can rewrite F as follows:
F(θq ) = ∑n ∑n ,...,n gn1 ,...,n p · θq (α1 z)n1 · . . . · θq (α p z)n p =
1
2
p
= ∑n Fn1 (θq ) = 0.
1
r1
The point α−1
1 q (where r1 is large enough positive if q > 1 and large enough negative
if q < 1) is a pole for all summands Fni and the multiplicity of this pole is different for
different ni . To annihilate these poles, Fn1 = 0 must hold for all n1 . Repeating the same
argument for all ni , we arrive at
gn1 ,...,n p (z) · θq (α1 z)n1 · . . . · θq (α p z)n p = 0
for each n1 , . . . , n p . Therefore, gn1 ,...,n p = 0.
It follows from this result that for an arbitrary root of unity ζ the function θq is σζ algebraically independent over C(z) in the field of meromorphic functions on C∗ . However,
to generalize this result to finitely many roots of unity, we need to require the following:
for all i and j
ζki = ζmj
implies ζki = ζmj = 1.
34
ANTIEAU ET AL.
Otherwise, the result is not true. Indeed, if ζki = ζmj 6= 1, then the relation
σkζi (θq ) − σm
ζ j (θq ) = 0
x are distinct
is non-trivial. Indeed, note that the difference indeterminates σkζ x and σm
ζj
i
even in the difference polynomial ring Q{x}Σ1 in spite of the fact that they define the same
automorphisms of meromorphic functions.
4.4 General order one q-difference equations
We will start by discussing several examples of σζ -dependence and independence and
finish by providing a general criterion in Theorem 4.9.
z+1
Example 4.7. For a(z) = z−1
, t = 2, and ζ = −1, we have
σζ (a)(z) · σ0ζ (a)(z) =
−z + 1 z + 1
·
= 1 = σq (1)/1.
−z − 1 z − 1
z+1
Let g be a meromorphic function on C \{0} such that σq (g) = z−1
· g. Then, g(z) · g(−z) is
σq -invariant:
z+1
z+1
z + 1 −z + 1
σq (g · σζ (g)) =
· g · σζ
·g =
·
· g · σζ(g) = g · σζ (g).
z−1
z−1
z − 1 −z − 1
So, the function g is σζ -algebraically dependent over k.
Example 4.8. For a(z) = 2z and t = 4 with ζ = i, we have
σ2ζ (a)(z) · σ0ζ (a)(z) = 2−z · 2z = 1 = σq (1)/1.
As before, let g be a meromorphic function on C \{0} such that σq (g) = 2z · g. Then,
g(z) · g(−z) is σq -invariant. Indeed,
σq (g · σ2ζ (g)) = 2z · g · σ2ζ (2z · g) = 2z · 2−z · g · σ2ζ (g) = g · σ2ζ (g).
So, the function g is σζ -algebraically dependent over k.
4.4.1 General characterization of periodic difference-algebraic independence
Let a ∈ C(z) and q, ζ ∈ C∗ be such that |q| > 1 and ζ is a primitive root of unity of
order t. Then, a can be represented as follows
t−1
a = λ · zT · ∏
N
R
∏ ∏
k=0 d=−N−1 i=1
z − ζk · qd · ri
sk,d,i
,
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
where λ, ri ∈ C∗ and the ri ’s are distinct in C
∑
sk,d,i , dk,i =
∑ ζk· j · ai, j , and D =
j=0
d=−N−1
35
ζZ · qZ . Let
t−1
N
ai,k =
∗
d0,1
d1,1
..
.
d0,2
d1,2
..
.
...
...
..
.
d0,R
d1,R
..
.
dt−1,1
dt−1,2
...
dt−1,R
.
The following result combined with Theorem 4.1 provides a complete characterization of
all equations (27) whose solutions are σζ -algebraically independent.
Theorem 4.9. Let a ∈ C(z) and D be as above. Then,
1. If T = 0 and, either λZ ∩ qZ 6= 1 or λ is a root of unity, then there exist b ∈ C(z) and
a multiplicative function
ϕ(x) = xn0 · σζ x
n1
· . . . · (σt−1
x)nt−1
ζ
such that ϕ(a) = σq (b)/b if and only if the matrix D contains a zero row.
2. If either T 6= 0 or, λZ ∩ qZ = 1 and λ is not a root of unity, then there exist b ∈ C(z)
and a multiplicative function
ϕ(x) = xn0 · σζ x
n1
· . . . · (σt−1
x)nt−1
ζ
such that ϕ(a) = σq (b)/b if and only if D contains a zero row other than the first
one.
Proof. We will write ϕ and b with undetermined coefficients and exponents. Suppose that
t−1
b = µ · zM · ∏
N
R
∏∏
k=0 d=−N i=1
z − ζk · qd · ri
lk,d,i
and ϕ(x) = xn0 · σζ x
n1
· . . . · (σt−1
x)nt−1
ζ
are such that ϕ(a) = σq (b)/b. Let us calculate the right and left-hand sides of this equality.
We see that
t−1
σq (b) = µ · qM+∑k,d,i lk,d,i · zM · ∏
N
R
∏∏
k=0 d=−N i=1
z − ζk · qd−1 · ri
lk,d,i
.
36
ANTIEAU ET AL.
Hence,
lk,d+1,i
R
t−1 N−1
σq (b)
·
= qM+∑k,d,i lk,d,i · ∏ ∏ ∏ z − ζk · qd · ri
b
k=0 d=−N−1 i=1
−lk,d,i
R
t−1 N
=
· ∏ ∏ ∏ z − ζk · qd · ri
k=0 d=−N i=1
lk,−N,i
t−1 R h
·
= qM+∑k,d,i lk,d,i · ∏ ∏ z − ζk · q−N−1 · ri
k=0 i=1
N−1
·
∏
d=−N
−lk,N,i i
lk,d+1,i−lk,d,i
.
z − ζk · qN · ri
z − ζk · qd · ri
Now, we calculate the left-hand side. We see that
t−1
σrζ a = λ · ζrT +∑k,d,i r·sk,d,i · zT · ∏
N
∏
R
sk,d,i
k−r
d
=
z
−
ζ
·
q
·
r
i
∏
k=0 d=−N−1 i=1
t−1
= λ · ζrT +∑k,d,i r·sk,d,i · zT · ∏
N
∏
R
sk+r,d,i
k
d
z
−
ζ
·
q
·
r
.
i
∏
k=0 d=−N−1 i=1
Hence,
t−1
t−1
t−1
ϕ(a) = λ∑r=0 nr · ζ(T +∑k,d,i sk,d,i )·(∑r=0 r·nr ) · zT ·(∑k=0 nr ) ·
∑t−1 nr sk+r,d,i
R
N
t−1
r=0
.
· ∏ ∏ ∏ z − ζk · qd · ri
k=0 d=−N−1 i=1
Now, the equation ϕ(a) = σq (b)/b gives the following system of equations
t−1
∑ sk+r,−N−1,i · nr = lk,−N,i
r=0
t−1
∑ sk+r,d,i · nr = lk,d+1,i − lk,d,i , −N 6 d 6 N − 1
r=0
t−1
∑ sk+r,N,i · nr = −lk,N,i
r=0
t−1
t−1
λ∑k=0 nr · ζ(T +∑k,d,i sk,d,i )·(∑r=0 r·nr ) = qM+∑k,d,i lk,d,i
t−1
T · ∑ nr = 0
r=0
In this system, the unknown variables are lk,d,i , nr , and M. If lk,d,i , nr , M is a solution of the
system such that not all nr ’s are zeroes, then t · lk,d,i ,t · nr ,t · M is a solution with the same
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
37
property. Therefore, we can replace the second equation with
t−1
λ∑k=0 nr = qM+∑k,d,i lk,d,i .
The first subsystem can be rewritten as follows:
sk,−N−1,i sk+1,−N−1,i . . . sk−1,−N−1,i
sk,−N,i
sk+1,−N,i
. . . sk−1,−N,i
..
..
.
.
..
..
.
.
sk,N,i
...
sk+1,N,i
sk−1,N,i
n0
n1
..
.
nt−1
lk,−N,i
lk,−N+1,i − lk,−N,i
=
.
..
.
−lk,N,i
This system has a solution in lk,d,i if and only if the sum of all equations is zero. Thus, we
can replace this system with the following:
∑
d=−N−1
N
N
N
n0 ·
sk,d,i + n1 ·
∑
d=−N−1
sk+1,d,i + . . . + nt−1 ·
∑
sk−1,d,i = 0.
d=−N−1
Using the definition of the ai, j ’s, we obtain the following system:
ai,0 ai,1 . . . ai,t−1
n0
ai,1 ai,2 . . . ai,0 n1
..
..
.. .. = 0.
.
.
.
.
.
. .
ai,t−1
ai,0
...
ai,t−2
nt−1
Thus, for some integers γk,d,i, j , we have:
n0
ai,0 ai,1 . . . ai,t−1
ai,1 ai,2 . . . ai,0 n1
..
.. .. = 0
..
.
.
.
.
. .
.
nt−1
a
a
.
.
.
a
i,t−1
i,0
i,t−2
t−1
λ∑k=0 nr = qM+∑k,d,i lk,d,i
t−1
T
·
∑ nr = 0
r=0
t−1
lk,d,i = ∑ γk,d,i,r · nr
r=0
Consider the first case: T = 0 and λZ ∩ qZ 6= 1. Then, for some u, v ∈ Z \ {0} we have
λu = qv . Hence, the second equation is equivalent to
t−1
v · ∑ nr = u · M + ∑k,d,i lk,d,i
r=0
38
ANTIEAU ET AL.
Suppose that nr and lk,d,i form a solution of all equations except for the second one, where
not all nr ’s are zero. Then,
t−1
u · nr ,
u · lk,d,i ,
M=
∑ (v · nr ) − ∑k,d,i (u · lk,d,i )
r=0
form a solution of the whole system. Thus, in this case, we may exclude the second equation. Now we will check the case T = 0 and λw = 1 for some w ∈ Z \ {0}. In this situation,
if nr , lk,d,i is a solution of all equations except for the second one, then
w · nr , w · lk,d,i , M = − ∑k,d,i w · lk,d,i
is a solution of the whole system. Therefore, in this
equivalent to
ai,0 ai,1 . . . ai,t−1
ai,1 ai,2 . . . ai,0
..
..
..
..
.
.
.
.
ai,t−1
ai,0
...
ai,t−2
case, the existence of ϕ and b is
n0
n1
..
.
nt−1
=0
(34)
having a nontrivial common solution.
Consider the second case: T 6= 0 or λZ ∩ qZ = 1 and λ is not a root of unity . If T 6=
Z
Z
0, then the third equation gives ∑t−1
r=0 nr = 0, and if λ ∩ q = 1 and λ is not a root of unity,
t−1
then the second equation gives ∑r=0 nr = 0. Therefore, in both cases, the second equation
is of the form
M + ∑k,d,i lk,d,i = 0.
Again, if nr , lk,d,i form a solution of all equations except the second one, where not all nr ’s
are zeroes, then
nr , lk,d,i , M = − ∑k,d,i lk,d,i
form a solution of the whole system with the same property. Thus, in this case, we need to
show the existence of a nontrivial solution of the system
n0
ai,0 ai,1 . . . ai,t−1
ai,1 ai,2 . . . ai,0 n1
..
.. .. = 0,
..
..
.
.
. .
.
nt−1
ai,t−1 ai,0 . . . ai,t−2
n0 + n1 + . . . + nt−1 = 0.
(35)
Since all the coefficients in (34) and (35) are integers, there is a nontrivial solution with
integral coefficients if and only if there is a nontrivial solution with complex coefficients.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
39
Define
1
1
E+ = 1
..
.
1 ζt−1
1
1
E− = 1
.
..
1
ζ2
ζ2·2
..
.
...
...
...
..
.
ζ2·(t−1)
ζ(t−1)·2
...
ζ(t−1)·(t−1)
1
ζ
ζ2
..
.
1
ζ−1
ζ−2
..
.
ζ−2
ζ−2·2
..
.
1 ζ−(t−1)
Ai =
ai,0
ai,1
..
.
ai,1
ai,2
..
.
...
...
..
.
ai,t−1
ai,0
...
ζt−1
..
.
...
...
...
..
.
1
1
,
1
ζ−(t−1)
ζ−2·(t−1)
..
.
ζ−(t−1)·2
ai,t−1
ai,0
.. ,
.
ai,t−2
,
. . . ζ−(t−1)·(t−1)
d0,i 0 . . .
0 d1,i . . .
Di = .
..
..
..
.
.
0
0
...
0
0
..
.
dt−1,i
.
A straightforward calculation shows that E+ · Ai = Di · E− . Let n be the vector with the
coordinates n0 , n1 , . . . , nt−1 . Hence, in the first case, the systems E+ · Ai · n = Di · E− · n = 0
have a nontrivial solution. This is equivalent to the condition that the systems Di · m = 0
have a nontrivial solution, where m = E− · n. Since the Di ’s are diagonal, there is a common
solution of all systems Di ·m = 0 if and only if the matrices Di have a zero in the same place.
In other words, there is an integer i0 such that for all i we have di0 ,i = 0. The latter condition
is equivalent to the condition that there is a zero row in the matrix D.
Consider the second case. Let l = (1, 1, . . . , 1) with t coordinates. We must to show
that the systems
Ai · n = 0, l · n = 0
have a nontrivial solution. Multiplying by E+ , we have
Di · E− · n = 0,
l · n = 0.
Let p1 , . . . , pu be the positions of all zero rows in the matrix D. And let E1 , . . . , Eu be the
−1
columns in E−
with the pi ’s as indices. Since the matrices Di are diagonal, every common
solution of the systems Di · E− · n = 0 is of the form:
n = W · µ,
W := (E1 , . . . , Eu ),
µ := (µ1 , . . . , µn )T .
Then, the equation l · n = 0 gives l ·W · µ = 0 Now, we find a condition when l · Ei is zero.
For this, note that (1, 0, . . . , 0) · E− = (1, 1, . . . , 1) and, therefore,
−1
(1, 1, . . . , 1) · E−
= (1, 0, . . . , 0).
40
ANTIEAU ET AL.
−1
Hence, only the first column of the matrix E−
gives nonzero elements in the vector l ·W .
The system l ·W · µ = 0 has only the zero solution if and only if W is just one column and
−1
l ·W 6= 0. Thus, this system has a nontrivial solution if and only if W contains a row of E−
other than the first one. In other words, the elements dk,i are zeroes for some k 6= 0 and all
i, 1 6 i 6 R. This is equivalent to the condition that D contains a zero row other than the
first one.
Corollary 4.10. In the situation of Theorem 4.1, if the zeros and poles of a ∈ C(z) are pairwise distinct modulo the group generated by ζ and q, then any non-zero solution f of the
equation σq ( f ) = a f is σζ -independent over k(z).
Example 4.11. If 0 6= c ∈ C, then any non-zero solution f of σq ( f ) = (z − c) · f is σζ independent over k(z).
ACKNOWLEDGMENTS
We are grateful to Yves André, Henri Gillet, Sergey Gorchinskiy, Charlotte Hardouin,
Manuel Kauers, Alexander Levin, Alice Medvedev, Eric Rosen, Jacques Sauloy, Michael
Singer, Lucia di Vizio, Michael Wibmer, and the referee for their helpful comments.
FUNDING
B. Antieau was supported by the NSF Grants DMS-0901373 and CCF-0901175.
A. Ovchinnikov was supported by the grants: NSF CCF-0901175, 0964875, and 0952591
and PSC-CUNY No. 60001-40 41.
REFERENCES
[1] K. Amano. Liouville extensions of Artinian simple module algebras. Communications in Algebra, 34(5):1811–1823, 2006. URL http://dx.doi.org/10.1080/00927870500542796.
[2] K. Amano and A. Masuoka.
Picard-Vessiot extensions of Artinian simple module algebras.
Journal of Algebra, 285(2):743–767, 2005.
URL
http://dx.doi.org/10.1016/j.jalgebra.2004.12.006.
[3] K. Amano, A. Masuoka, and M. Takeuchi.
Hopf algebraic approach to
Picard-Vessiot theory.
Handbook of Algebra, 6:127–171, 2009.
URL
http://dx.doi.org/10.1016/S1570-7954(08)00204-0.
[4] Y. André. Différentielles non commutatives et théorie de Galois différentielle ou aux
différences. Annales Scientifiques de l’École Normale Supérieure. Quatrième Série, 34(5):
685–739, 2001. URL http://dx.doi.org/10.1016/S0012-9593(01)01074-6.
[5] M. Atiyah and I. Macdonald. Introduction to Commutative Algebra. Addison-Wesley, Reading,
MA, 1969.
[6] N. Bourbaki. Algèbre. Chapitre 8: Modules et anneaux semi-simples. Hermann, Paris, 1958.
GALOIS THEORY OF DIFFERENCE EQUATIONS WITH PERIODIC PARAMETERS
41
[7] N. Bourbaki. Algèbre. Chapitres 4–7. Éléments de Mathématique. Springer, 2007. URL
http://dx.doi.org/10.1007/978-3-540-34499-5.
[8] P. Cassidy. Differential algebraic groups. American Journal of Mathematics, 94:891–954, 1972.
URL http://www.jstor.org/stable/2373764.
[9] P. Cassidy.
The differential rational representation algebra on a linear differential algebraic group.
Journal of Algebra, 37(2):223–238, 1975.
URL
http://dx.doi.org/10.1016/0021-8693(75)90075-7.
[10] Z. Chatzidakis, C. Hardouin, and M. F. Singer. On the definitions of difference Galois groups.
In Model theory with applications to algebra and analysis. Vol. 1, volume 349, pages 73–109.
2008. URL http://dx.doi.org/10.1017/CBO9780511735226.006.
[11] R. M. Cohn. Difference algebra. Interscience Publishers John Wiley & Sons, New YorkLondon-Sydeny, 1965.
[12] L.
Di
Vizio.
Approche
galoisienne
de
la
transcendance
différentielle.
Journée annuelle 2012 de la SMF, 2012.
URL
http://divizio.perso.math.cnrs.fr/PREPRINTS/16-JourneeAnnuelleSMF/difftransc.pdf.
[13] L. Di Vizio and C. Hardouin. Courbures, groupes de Galois génériques et D-groupoı̈des de
Galois d’un système aux D-différences. C. R. Math. Acad. Sci. Paris, 348(17–18):951–954,
2010. URL http://dx.doi.org/10.1016/j.crma.2010.08.001.
[14] L. Di Vizio and C. Hardouin. Galois theories for q-difference equations: comparison theorems.
2011. URL http://arxiv.org/abs/1002.4839.
[15] L. Di Vizio and C. Hardouin. Parameterized generic Galois groups for q-difference equations,
followed by the appendix “The Galois D-groupoid of a q-difference system” by Anne Granier.
2011. URL http://arxiv.org/abs/1002.4839.
[16] L. Di Vizio and C. Hardouin. On the Grothendieck conjecture on p-curvatures for q-difference
equations. 2011. URL http://arxiv.org/abs/1002.4839.
[17] L. Di Vizio and C. Hardouin. Descent for differential Galois theory of difference equations.
Confluence and q-dependency. Pacific Journal of Mathematics, 256(1):79–104, 2012. URL
http://dx.doi.org/10.2140/pjm.2012.256.79.
[18] L. Di Vizio, C. Hardouin, and M. Wibmer.
Difference Galois theory of linear differential equations.
Advances in Mathematics, 260:1–58, 2014.
URL
http://dx.doi.org/10.1016/j.aim.2014.04.005.
[19] A. Granier. A Galois D-groupoid for q-difference equations. Annales de l’Institut Fourier, 61
(4):1493–1516, 2011. URL http://dx.doi.org/10.5802/aif.2648.
[20] C. Hardouin. Hypertranscendance et groupes de Galois aux différences. 2006. URL
http://arxiv.org/abs/math/0609646.
[21] C. Hardouin. Hypertranscendance des systèmes aux différences diagonaux. Compositio Mathematica, 144(3):565–581, 2008. URL http://dx.doi.org/10.1112/S0010437X07003430.
[22] C. Hardouin and M. F. Singer.
Differential Galois theory of linear difference equations.
Mathematische Annalen, 342(2):333–377, 2008.
URL
http://dx.doi.org/10.1007/s00208-008-0238-z.
[23] M. Kamensky.
Tannakian formalism over fields with operators.
International Mathematics Research Notices, 2013(24):5571–5622, 2013.
URL
http://dx.doi.org/10.1093/imrn/rns190.
42
ANTIEAU ET AL.
[24] B. Lando.
Extensions of difference specializations.
Proceedings
of the American Mathematical Society,
79(2):197–202,
1980.
URL
http://dx.doi.org/10.1090/S0002-9939-1980-0565337-1.
[25] A. Levin. Difference algebra, volume 8 of Algebra and Applications. Springer, New York,
2008. URL http://dx.doi.org/10.1007/978-1-4020-6947-5.
[26] J. Milne. Étale Cohomology. Princeton University Press, Princeton, NJ, 1980.
[27] S. Morikawa. On a general difference Galois theory. I. Annales de l’Institut Fourier, 59(7):
2709–2732, 2009. URL http://dx.doi.org/10.5802/aif.2505.
[28] A. Ovchinnikov and M. Wibmer. σ-Galois theory of linear difference equations. International
Mathematics Research Notices, 2015. URL http://dx.doi.org/10.1093/imrn/rnu060.
[29] E. Rosen. A differential Chevalley theorem. 2008. URL http://arxiv.org/abs/0810.5486.
[30] D. Trushin.
Difference Nullstellensatz in the case of finite group.
2009.
URL
http://arxiv.org/abs/0908.3863.
[31] D. Trushin. Difference Nullstellensatz. 2009. URL http://arxiv.org/abs/0908.3865.
[32] M. van der Put and M. F. Singer.
Galois theory of difference equations, volume 1666 of Lecture Notes in Mathematics.
Springer-Verlag, Berlin, 1997.
URL
http://dx.doi.org/10.1007/BFb0096118.
[33] W. Waterhouse. Introduction to affine group schemes. Springer, New York, 1979. URL
http://dx.doi.org/10.1007/978-1-4612-6217-6.
[34] M. Wibmer. Geometric Difference Galois Theory. PhD thesis, University of Heidelberg, 2010.
[35] M. Wibmer. A Chevalley theorem for difference equations. Mathematische Annalen, 354(4):
1369–1396, 2012. URL http://dx.doi.org/10.1007/s00208-011-0770-0.
| 0math.AC
|
Linear identification of nonlinear systems:
A lifting technique based on the Koopman operator
arXiv:1605.04457v2 [cs.SY] 27 Aug 2016
Alexandre Mauroy and Jorge Goncalves
Abstract— We exploit the key idea that nonlinear system
identification is equivalent to linear identification of the socalled Koopman operator. Instead of considering nonlinear
system identification in the state space, we obtain a novel
linear identification technique by recasting the problem in the
infinite-dimensional space of observables. This technique can
be described in two main steps. In the first step, similar to
a component of the Extended Dynamic Mode Decomposition
algorithm, the data are lifted to the infinite-dimensional space
and used for linear identification of the Koopman operator. In
the second step, the obtained Koopman operator is “projected
back” to the finite-dimensional state space, and identified to the
nonlinear vector field through a linear least squares problem.
The proposed technique is efficient to recover (polynomial)
vector fields of different classes of systems, including unstable,
chaotic, and open systems. In addition, it is robust to noise,
well-suited to model low sampling rate datasets, and able to
infer network topology and dynamics.
I. I NTRODUCTION
Operator-theoretic techniques rely on the idea of lifting nonlinear dynamical systems to an infinite dimensional
space, where those systems have a linear representation.
For instance, the so-called Koopman operator is a linear
operator that describes the evolution of observable-functions
along the trajectories of the system. Its spectral properties
have been investigated over the past years and related to
geometric properties of the system (see e.g. [5]). On the
one hand, this body of work yielded novel methods for the
analysis of nonlinear systems described by a known vector
field (e.g. global stability analysis [4], global linearization
[2], analysis of monotone systems [10]). On the other hand,
the Koopman operator approach was shown to be conducive
to data analysis [8] and related to numerical methods such
as Dynamic Mode Decomposition (DMD) [9], yielding new
techniques for the analysis of nonlinear systems described by
trajectory data points (snapshots). In this context, the present
paper aims at bridging these two directions of analysis,
connecting data to vector field.
While Koopman operator techniques have been previously
considered in the framework of system analysis, we present a
first attempt –to the authors knowledge– to develop them for
system identification. Exploiting the lifting obtained through
the Koopman operator, our key idea is to identify a linear
operator in the space of observables, instead of identifying a
nonlinear system in the state space. Since Koopman operator
A. Mauroy and J. Goncalves are with the Luxembourg
Centre for Systems Biomedicine, University of Luxembourg,
4367 Belvaux, Luxembourg [email protected],
[email protected]
and state space representations are both equivalent, identifying the operator is indeed sufficient to recover the nonlinear
dynamical system. Because the Koopman operator is defined
in an infinite-dimensional space, we consider its projection
onto a finite basis of functions. The proposed method can
somehow be interpreted, in the space of observables, as an
analog of methods developed in [1], [6], which use dictionary
functions in the state space. Our method proceeds in two
main steps: from state space to space of observables, and
back to state space. In the first step, a finite-dimensional
projection of the Koopman operator is identified from data
through classical linear identification, a method that is related
to a component of the so-called Extended Dynamic Mode
Decomposition (EDMD) algorithm [12]. In the second step,
the projected operator is connected to the vector field of
the nonlinear system through an analytic derivation of the
infinitesimal generator (following similar steps as in [4]).
The expansion of the vector field in the finite basis is then
obtained by solving a linear least squares problem.
Thanks to the lifting technique, the proposed method has
the advantage of relying only on linear techniques (i.e. least
squares regression), in spite of the fact that the dynamical
system is nonlinear. In contrast to recent (direct) methods
(e.g. [1]), it is a (indirect) method that does not rely on the
estimation of time derivatives, which cannot be performed
when the sampling rate is too low. Therefore, the method is
well-suited to low sampling, and also robust to both measurement noise and process noise. In addition, it works efficiently
with different kinds of behaviors, including unstable and
chaotic systems (although limited so far to polynomial vector
fields) and can be applied to several small datasets obtained
from multiple trajectories. This flexibility is well-suited to
identification problems such as those encountered in the
context of system biology.
The rest of the paper is organized as follows. In Section II,
we introduce the system identification problem and present
the lifting technique obtained through the Koopman operator.
Our numerical method for nonlinear system identification
based on the lifting technique is described in Section III.
Extensions of the method to open systems (i.e. with outputs
or process noise) are discussed in Section IV. In Section V,
the method is illustrated with several numerical examples,
and applied to network reconstruction. Concluding remarks
and perspectives are given in Section VI.
II. KOOPMAN
OPERATOR TECHNIQUE
FOR SYSTEM IDENTIFICATION
A. Problem statement
Consider a (unknown) dynamical system
x ∈ Rn
ẋ = F(x) ,
(1)
of observables (at some points in the state space). Then classical linear identification is used to obtain an approximation
of the Koopman operator (in a finite-dimensional subspace
of F ). This corresponds to a component of the so-called
Extended Dynamic Mode Decomposition (EDMD) algorithm
[12]. In the second step, the vector field F is obtained through
the equalities (3) and (4).
with a vector field F(x) that is assumed to be polynomial.
We denote by ϕ : R+ × Rn → Rn the flow induced by
the system, i.e. t 7→ ϕt (x0 ) = ϕ(t, x0 ) is a solution of (1)
associated with the initial condition x0 .
We address the problem of identifying the vector field
F given K snapshot pairs (xk , yk ) ∈ Rn×2 . The data
points are such that yk = ϕTs (xk )(1 + vk ), where each
component of the measurement noise vk is a Gaussian
random variable with zero mean and standard deviation
σmeas . The sampling period Ts is assumed to be identical
for all pairs (xk , yk ). Note that the data points xk and yk
can belong to different trajectories. Stochastic systems with
process noise and systems with inputs will also be considered
in Section IV.
B. Lifting technique using the Koopman operator
The system (1) is described by its flow ϕt in the state space
R . Alternatively, the system can be lifted to an infinitedimensional space F of observable-functions f : Rn → R.
In this space, it is described by the semigroup of Koopman
operators U t : F → F , t ≥ 0, which governs the evolution
of the observables along the trajectories:
n
U t f = f ◦ ϕt
∀f ∈ F .
(2)
This lifting technique is of interest since it leads to a linear
(but infinite-dimensional) representation of the system. Even
if the system (1) is nonlinear, the operator is always linear
since we have
U t (a1 f1 +a2 f2 ) = a1 f1 ◦ϕt +a2 f2 ◦ϕt = a1 U t f1 +a2 U t f2 .
In addition, there is a bijective relationship between the
Koopman operator and its associated system. We denote by
L = limt→0 (U t − I)/t the infinitesimal generator of the
Koopman semigroup, which also satisfies
U t = eLt =
∞ k
X
t
k=0
k!
Lk .
(3)
Provided that the vector field F is continuously differentiable, it can be shown that
L = F· ∇,
(4)
where ∇ denotes the gradient (see e.g. [3]). This implies that
there is a one-to-one correspondence between the Koopman
operator and the vector field.
This paper exploits the equivalence between the two
descriptions of the system, proposing to identify the linear
Koopman operator in the space of observables. Figure 1 illustrates the two main steps of the method. In the first step, data
are lifted to the space of observables, providing the evolution
linear
identification
lifting
nonlinear
identification
Fig. 1. Classical nonlinear system identification is performed directly in
the state space. In contrast, the proposed Koopman operator identification
technique consists of two steps: (1) lifting of the data and linear identification of the Koopman operator; (2) identification of the vector field using
(3) and (4).
III. N UMERICAL
METHOD
This section describes the numerical scheme derived from
the Koopman operator identification method. Lifting of the
data and identification of the operator (step 1) are explained
in the first subsection. Identification of the vector field (step
2) is described in the second subsection.
Conceptually, the method works in the infinitedimensional space F of observables. But to be numerically
tractable, it is developed on a finite-dimensional subspace
FN which is spanned by a basis of N linearly independent
functions {pk }N
k=1 . In our case, we will choose a basis of
monomials with total degree less or equal to m:
pk (x) ∈ {xs11 · · · xsnn |(s1 , . . . , sn ) ∈ Nn , s1 +· · ·+sn ≤ m} .
This choice is motivated by the fact that the vector field is
assumed to be polynomial. Other bases of functions (e.g.
Fourier basis, radial basis functions) could be considered to
generalize the method. For practical purposes, the sequence
of monomials should be characterized by some order (e.g.
lexicographic order, weight order). The number of functions
in the basis is equal to N = (n+ m)!/(n! m!). We denote by
p(x) = (p1 (x), . . . , pN (x))T the vector of basis monomials.
The Koopman operator U t can be projected onto the
subspace FN and represented by a matrix U. Considering
f (x) = aT p(x) ∈ FN
(5)
PU t f (x) = bT p(x) ∈ FN
(6)
and
where P : F → FN is a projection operator, we have
Ua = b,
U∈R
N ×N
.
(7)
The matrix U provides a (approximate) finite-dimensional
linear representation of the nonlinear system. This representation is not obtained through local linearization techniques
and is valid globally in the state space. In the numerical
scheme described below, we compute U from data (step 1),
and then connect the obtained matrix to the vector field of
the nonlinear system (step 2).
A. Step 1: Lifting of the data and Koopman operator identification
We now describe the method used to compute the matrix
U given snapshot pairs {(xk , yk )}K
k=1 , which corresponds
to a component of the EDMD algorithm (see [12] for more
details).
For each snapshot pair (xk , yk ) ∈ Rn×2 , we construct the
new pair (p(xk ), p(yk )) ∈ RN ×2 . The data are somehow
lifted to (a subspace of) the space of observables and provide
snapshots of trajectories of particular observables (i.e. the
basis functions). Provided that the measurement noise vk is
small, we have yk ≈ ϕTs (xk ). It follows from (2) and (5)
that
PU Ts f (xk ) = aT PU Ts p(xk ) ≈ aT Pp(yk )
T
PU Ts f (xk ) = aT U p(xk ) ,
so that
p(xk )T U ≈ Pp(yk )T
(8)
where U denotes here the matrix representation of PU Ts .
The last equality implies that the jth column cj of U is so
that cTj p ≈ PU Ts pj . If the projection P is chosen so that it
yields the least squares fit at the points xk , k = 1, . . . , K
(i.e. finite-dimensional Galerkin projection), the matrix U is
given by the least squares solution
U = (Px )† Py
T
p(xK )
Ljk = pk
∂
,
∂xj
j = 1, . . . , n ,
k = 1, . . . , NF .
It is clear that the operators Ljk map polynomials of total
degree less or equal to m1 to polynomials of total degree
less or equal to m2 = m1 + mF − 1. Denote by pm (x) the
vector of monomials of total degree less or equal to m. The
projection of Ljk on the basis of monomials is represented by
j
the matrix Lk ∈ RN2 ×N1 , with N1 = (m1 + n)!/(m1 !n!) ≤
N2 = (m2 + n)!/(m2 !n!), which is so that
j
Ljk f (x) = (Lk a)T pm2 (x)
(12)
for all f (x) = aT pm1 (x). We have also Ljk pm1 (x) =
j
j
(Lk )T pm2 (x), which implies that the lth column of Lk corj
responds to the expansion of Lk pl in the basis of monomials
pm2 . Defining an index function Ψ = (ψ1 (k), . . . , ψn (k))
that encodes the order of the monomials in the vector p, i.e.
ψ (k)
ψ (k)
pk (x) = x1 1 · · · xn n , we have
Ljk pl = ψj (l)pΨ−1 (Ψ(k)+Ψ(l)−ej )
and from (6) and (7) that
with the K × N matrices
p(x1 )T
..
Px =
.
1) Derivation of the Koopman infinitesimal generator:
Our goal is to derive the matrix L which represents the
projection of the infinitesimal generator (4) on the basis
of monomials pk , and show that this is a function of the
unknown coefficients wkj of the vector field. Consider the
linear operators
(9)
where ej ∈ Rn is the jth unit vector. It follows that the
j
entries of Lk are given by
(
h ji
ψj (l) if Ψ(i) = Ψ(k) + Ψ(l) − ej ,
Lk =
(13)
il
0
otherwise .
j
Note that the matrices Lk can also be obtained by multiplying a multiplication matrix and a differentiation matrix (see
[4] for more details). Finally it follows from (4) and (11)
that
NF
n X
X
wkj Ljk
(14)
L=
j=1 k=1
Py =
T
p(y1 )
..
.
T
p(yK )
and where P† denotes the pseudoinverse of P.
and the matrix L ∈ R
N2 ×N1
(10)
B. Step 2: Identification of the vector field
We assume that the polynomial vector field F =
(F1 , . . . , Fn )T is of total degree less or equal to mF , so
that we can write
NF
X
Fj (x) =
wkj pk (x)
j = 1, . . . , n
(11)
k=1
where NF = (mF + n)!/(mF !n!) is the number of monomials pk of total degree less or equal to mF . The identification
problem is therefore reduced to identifying nNF coefficients
wkj .
L=
given by
NF
n X
X
j
wkj Lk
(15)
j=1 k=1
represents the projection of the Koopman infinitesimal generator on the basis of monomials pk .
2) Computation of the vector field: The coefficients wkj of
the vector field are finally computed by using the relationship
between U (obtained from data in (9)) and L (which depends
on wkj in (15)).
It can be shown easily (using (3)) that, if U and L
represent the projection of U t and L, respectively, then
U = eLt . This property and (9) imply that we can obtain an
estimation of L from data:
1
log((Px )† Py ) ∈ RN ×N .
(16)
Ldata =
Ts
The function log denotes the (principal) matrix logarithm.
Remark 1 (Matrix logarithm): The real eigenvalues of
U = (Px )† Py shall not be negative, so that the principal
logarithm exists and is real. However, numerical simulations
show that U can have negative real eigenvalues when the
number of data points is too small. In this case, there exists
no real matrix log(U) and the method might fail.
When the Koopman infinitesimal generator admits complex eigenvalues, the sampling period Ts must be sufficiently small so that the eigenvalues of Ts L lie in the strip
{z ∈ C : |ℑ{z}| < π}. If this condition is not satisfied,
log(exp(Ts L)) 6= Ts L and the accuracy of the method is
altered. This issue is discussed with more details in [13]. ⋄
Using (15), we obtain the equality
L=
NF
n X
X
j
ˆ
wkj Lk = L
data
(17)
j=1 k=1
IV. E XTENSIONS TO
OPEN SYSTEMS
In this section, we show that the proposed method is wellsuited to identify open systems that are driven by a known
input or by a white noise (i.e. process noise).
A. Systems with inputs
Consider an open dynamical system of the form
ẋ = F(x, u(t))
(19)
with x ∈ Rn and with the input u ∈ U : R+ → Rp .
We define the associated flow ϕ : R+ × Rn × U so that
t 7→ ϕ(t, x, u(·)) is a solution of (19) with the initial
condition x and the input u(·). Following the generalization
proposed in [7] for discrete-time dynamical systems, we
consider observables f : Rn × Rp → R and define the
semigroup of Koopman operators
U t f (x, u) = f (ϕt (x, u(·) = u), u)
ˆ
where L
data is the N2 × N1 matrix constructed with the N1
columns and N2 rows of Ldata associated with monomials
of total degree less or equal to m1 and m2 , respectively (note
that we will typically choose N = N2 ≥ N1 for the size of U
and Ldata ). By disregarding the N2 −N1 remaining columns
of Ldata , we only consider monomials that are mapped by
L onto the span of basis monomials (of total degree less or
equal to m2 ), for which the finite Galerkin projection (i.e.
the identity in this case) is exact.
The matrix equality (17) corresponds to a linear set of
equations (with nNF unknowns wkj ) that is overdetermined.
Its least squares solution yields an approximation of the
coefficients wkj :
where u(·) = u is a constant input. In this case, u can
be considered as additional state variables and the above
operator is the classical Koopman operator for the augmented
system ẋ = F(x, u), u̇ = 0. In particular, the infinitesimal
generator is still given by (4).
It follows that the method proposed in Sections III-A and
III-B can be used if the input can be considered as constant
between two snapshots (zero-order hold assumption), or
equivalently if the sampling rate is high enough. The matrix
U is obtained with snapshot pairs ([xk , uk ], [yk , uk ]) ∈
R(n+p)×2 and the rest of the procedure follows on similar
lines with the augmented state space Rn+p . The efficiency of
the identification method in the case of systems with inputs
is shown in Section V-B.
w11
|
..
1
. =
vec(L1 )
n
|
wN
F
B. Process noise
†
|
n
ˆ
. . . vec(LNF ) vec(Ldata )
|
(18)
where vec stands for the vectorization of the matrix. Note
that more advanced techniques could be used to solve (17),
for instance promoting sparsity of the vector of unknowns
wkj by adding a L1 -norm penalty term to (17), see e.g. the
LASSO method [11]. We leave this for future research.
Remark 2: Each entry of the matrix L yields an equality
j
in (18). However, if a given entry of Lk is zero for all j
and k, then the corresponding entry of L does not depend
on the coefficients wkj and the related equality in (18) can
be disregarded. In particular, when m1 = 1, the number of
effective equalities is equal to nNF and the problem is not
overdetermined. Moreover, (17) can be solved directly in this
case since each coefficient wkj is equal to an entry of L. ⋄
Remark 3: The identification problem could also be performed at the level of the Koopman semigroup. However
solving the equality U = eLTs (with a square matrix L)
amounts to solving a (nonconvex) nonlinear least squares
problem. Numerical simulations suggest that better results
are obtained by solving (17).
⋄
We have considered so far only measurement noise. We
will show that the proposed method is also robust to process noise. Consider a system described by the stochastic
differential equation
ẋk = Fk (x) + ηk (t)
(20)
where ηk (t) is a white noise that satisfies E[ηk (s)ηj (t)] =
2
σproc
δkj δ(t − s) (where E denotes the expectation). We
define the flow ϕ : R+ × Rn × Ω, where Ω is the probability
space, such that t 7→ ϕ(t, x, ω) is a solution of (20). In this
case, the semigroup of Koopman operators is defined by (see
e.g. [5])
U t f (x) = E[f (ϕ(t, x, ω))]
and its infinitesimal generator is given by
Lf = F · ∇f +
2
σproc
∆f
2
P
where ∆ = k ∂ 2 /∂x2k denotes the Laplacian operator that
accounts for diffusion. Note that the infinitesimal generator
is related to the so-called Kolmogorov backward equation.
Now we can show that the numerical scheme of the
proposed identification method does not need to be adapted
to take account of process noise. As explained in [12], the
first step of the method (Section III-A) is still valid for
identifying the matrix U. In the second step (Section III-B),
the procedure is the same, except that one has to consider
the Laplacian operator whose projection on the basis of
monomials is represented by a matrix D ∈ RN2 ×N1 . The
equality (17) is then replaced by
NF
n X
X
j
wkj Lk +
j=1 k=1
σ2
ˆ
D=L
data
2
where σ is an additional unknown. While the operators Ljk
map monomials of total degree m to monomials of total
degree greater or equal to m−1, the Laplacian operator maps
monomials of total degree m to monomials of total degree
m−2. Therefore all nonzero entries of D correspond to zero
j
entries of Lk , so that the addition of the diffusion term only
modifies entries of L which are zero—and do not depend
on wkj —when there is no process noise. In other words, the
diffusion term does not affect the equalities on wkj , whose
solution is still given by (18). In Section V-B, an example
illustrates the robustness of the method against process noise.
V. N UMERICAL
EXAMPLES
In this section, we illustrate our numerical method with
several examples. We show that the method is efficient
to recover the vector field of different kinds of systems,
including unstable, chaotic, and open systems. The method
is also applied to network reconstruction.
For all the simulations, we consider measurement noise
with standard deviation σmeas = 0.01. We also choose the
parameters m1 = 1 and mF = 3, so that m2 = m1 + mF −
1 = 3.
A. Periodic, unstable, and chaotic systems
We first consider three systems that exhibit different kinds
of behavior.
a) Van der Pol oscillator: The dynamics are given by
ẋ1
= x2
ẋ2
= (1 − x21 )x2 − x2
and possess a stable limit cycle. The data are obtained by
taking 3 snapshots at t ∈ {0, 0.5, 1} (sampling period Ts =
0.5) for 10 trajectories with initial conditions on [−1, 1]2 .
b) Unstable equilibrium: The dynamics are given by
ẋ1
ẋ2
= 3 x1 + 0.5 x2 − x1 x2 + x22 + 2 x31
= 0.5 x1 + 4 x2
and are characterized by an unstable equilibrium at the
origin. The data are obtained by taking 2 snapshots at
t ∈ {0, 0.1} (sampling period Ts = 0.1) for 20 trajectories
with initial conditions on [−1, 1]2.
c) Chaotic Lorenz system: The dynamics are given by
ẋ1
ẋ2
= 10(x2 − x1 )
= x1 (28 − x3 ) − x2
ẋ3
= x1 x2 − 8/3 x3
and exhibit a chaotic behavior. The data are obtained by
taking 31 snapshots over [0, 1] (sampling period Ts = 0.033)
for 10 trajectories with initial conditions on [−20, 20]3.
For each model, we identify the coefficient wkj of the
vector field and compute the root mean square error
v
u
NF
n X
2
u 1 X
(wkj )estim − (wkj )exact .
RMSE = t
nNF j=1
k=1
where (wkj )estim and (wkj )exact are the estimated and exact
values of the coefficients wkj . In order to compare the
results obtained with different systems, we also compute
the normalized RMSE: NRMSE = RMSE/w where w is
the average value of the nonzero coefficients |(wkj )exact |.
The RMSE and NRMSE averaged over 10 experiments are
given in Table I. The results show that the method performs
well, even in the case of unstable and chaotic systems (for
which the NRMSE is slightly larger). Note that the accuracy
increases with the number of data points (which is low here)
but a comprehensive study of this effect is beyond the scope
of this paper.
a) Van der Pol
b) Unstable equilibrium
c) Chaotic Lorenz
RMSE
0.035
0.157
0.569
NRMSE
0.035
0.092
0.074
TABLE I
(N ORMALIZED ) ROOT MEAN SQUARED ERROR OVER 10 SIMULATIONS .
B. Input and process noise
We consider the forced Duffing system
ẋ1
ẋ2
= x2
= x1 − x31 − 0.2 x2 + 0.2 x21 u
(21)
(22)
where u is the input. With u(t) = cos(t), we obtain 51
snapshots over [0, 10] (sampling period Ts = 0.2) for 5
trajectories with initial conditions on [−1, 1]2 . Our identification method provides a good estimation of the vector field
(including the forcing term 0.2 x21 u). The RMSE computed
over all coefficients (including those related to the forcing
term) and averaged over 10 simulations is equal to 0.025.
We now consider the effect of process noise and replace
the forcing term in (21) by a strong white noise with
σproc = 1 (note that we still add measurement noise with
σmeas = 0.01). Data points are obtained as previously,
but for 10 trajectories computed with the Euler-Maruyama
scheme. The method is robust against process noise, and
recovers the vector field with a RMSE (averaged over 10
simulations) equal to 0.078.
C. Network reconstruction
2
We consider a 12-dimensional system of the form
ẋj = −ξj xj +
3
X
k=1
σk
σk
ζj,k xν kj,1 xν kj,2
j,1
j = 1, . . . , 12 (23)
j,2
where the coefficients ξj are chosen according to a uniform
distribution on [0, 1] and ζj,k are distributed according to a
Gaussian distribution of zero mean and standard deviation
k
k
equal to one. The subscripts νj,1
and νj,2
are integers
k
randomly chosen on {1, . . . , 12} and the exponents σj,1
and
k
σj,2 are also integers randomly chosen on {0, 1, 2, 3} so that
k
k
σj,1
+ σj,2
∈ {2, 3}. The first term in (23) is a linear term
that ensures local stability of the origin. The other terms are
quadratic and cubic nonlinearities that define an interaction
network between the different states. We say that there is a
k
k
directed link from xl to xj if νj,r
= l and σj,r
6= 0 for some
k ∈ {1, 2, 3} and r ∈ {1, 2}.
For 500 different trajectories with initial conditions on
[−1, 1]12 , we consider three snapshots obtained at time t ∈
{0, 0.5, 1} (sampling period Ts = 0.5). Figure 2 shows
that, with this dataset, the method provides an accurate
estimation of all the coefficients of the polynomial vector
field (RMSE = 0.021). Note that we consider a total of 1500
data points to identify 12NF = 5460 coefficients wkj . As we
decrease the number of data points, results are less accurate
but still acceptable (RMSE = 0.106 with 400 trajectories
and RMSE = 0.151 with 300 trajectories). More advanced
methods promoting sparsity (e.g. LASSO method [11]) could
further improve the accuracy of the results.
Network reconstruction aims at predicting links in the
network. Very small coefficients are mainly due to measurement noise and have an exact value equal to zero. Also
we consider that there is a link from xl to xj if at least
one coefficient wkj related to a monomial containing xl is
above a given threshold value. With a threshold value equal
to 0.1, we obtain a true positive rate (i.e. number of correctly
identified links divided by the actual number of links) equal
to 0.875 and a false positive rate (i.e. number of incorrectly
identified links divided by the actual number of missing
links) equal to 0.023. It is noticeable that, in addition to this
good performance, the method also provides the nature of
the links (e.g. quadratic, cubic) and their weight (i.e. values
wkj ).
VI. C ONCLUSION
We have proposed a novel method for the identification
of nonlinear systems. This method is based on a lifting
technique which recasts nonlinear system identification as the
identification of the linear Koopman operator in an infinitedimensional space of observables. A key advantage of the
method is that it relies only on linear techniques and is an
indirect method that does not require the estimation of time
derivatives. It follows that the method is well-suited to model
low sampling rate data sets. It is also robust to noise and
efficient to recover the vector field of different classes of
systems.
1
wkj
0
-1
estimation
exact
1000
2000
3000
4000
5000
Index
Fig. 2.
The method is efficient to recover all the coefficients of the
polynomial vector field of a 12-dimensional system. The exact nonzero
j
values wk are depicted in blue. The estimated values are in red.
This paper presents preliminary results that open the
door to further developments of Koopman operator lifting
techniques for nonlinear system identification (e.g. use other
bases to identify non-polynomial vector fields). Since the
method is based on the (finite-dimensional) truncation of the
(infinite-dimensional) Koopman operator, a theoretical study
of its convergence should also be provided in future work.
VII. ACKNOWLEDGMENTS
The authors acknowledge support from the Luxembourg
National Research Fund (FNR 14/BM/8231540).
R EFERENCES
[1] S. L. Brunton, L. P. Proctor, and J. N. Kutz. Discovering governing
equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences,
113(15):3932–3937, 2016.
[2] Y. Lan and I. Mezić. Linearization in the large of nonlinear systems
and Koopman operator spectrum. Physica D, 242:42–53, 2013.
[3] A. Lasota and M. C. Mackey. Chaos, Fractals, and Noise: stochastic
aspects of dynamics. Springer-Verlag, 1994.
[4] A. Mauroy and I. Mezić. Global stability analysis using the eigenfunctions of the Koopman operator. IEEE Transactions On Automatic
Control, 2016. In press.
[5] I. Mezić. Spectral properties of dynamical systems, model reduction
and decompositions. Nonlinear Dynamics, 41(1-3):309–325, 2005.
[6] W. Pan, Y. Yuan, J. Goncalves, and G.-B. Stan. A sparse Bayesian
approach to the identification of nonlinear state-space systems. IEEE
Transactions On Automatic Control, 61(1):182–187, 2016.
[7] L. P. Proctor, S. L. Brunton, and J. N. Kutz.
Generalizing Koopman operator theory to allow for inputs and control.
http://arxiv.org/abs/1602.07647.
[8] C. W. Rowley, I. Mezić, S. Bagheri, P. Schlatter, and D. S. Henningson.
Spectral analysis of nonlinear flows. Journal of Fluid Mechanics,
641:115–127, 2009.
[9] P. J. Schmid. Dynamic mode decomposition of numerical and
experimental data. Journal of Fluid Mechanics, 656:5–28, 2010.
[10] A. Sootla and A. Mauroy. Operator-theoretic characterization of
eventually monotone systems. http://arxiv.org/abs/1510.01149.
[11] R. Tibshirani. Regression shrinkage and selection via the lasso.
Journal of the Royal Statistical Society. Series B (Methodological),
pages 267–288, 1996.
[12] M. O. Williams, I. G. Kevrekidis, and C. W. Rowley. A Data-Driven
Approximation of the Koopman Operator: Extending Dynamic Mode
Decomposition. Journal of Nonlinear Science, pages 1–40, June 2015.
[13] Z. Yue, J. Thunberg, L. Ljung, and J. Goncalves. Identification of
Sparse Continuous-Time Linear Systems with Low Sampling Rate:
Exploring Matrix Logarithms. https://arxiv.org/abs/1605.08590.
| 3cs.SY
|
Step-Indexed Normalization for a Language with General
Recursion
Chris Casinghino
[email protected]
Vilhelm Sjöberg
School of Engineering and Applied Science
University of Pennsylvania
[email protected]
Stephanie Weirich
[email protected]
The T RELLYS project has produced several designs for practical dependently typed languages. These
languages are broken into two fragments—a logical fragment where every term normalizes and
which is consistent when interpreted as a logic, and a programmatic fragment with general recursion and other convenient but unsound features. In this paper, we present a small example language
in this style. Our design allows the programmer to explicitly mention and pass information between
the two fragments. We show that this feature substantially complicates the metatheory and present
a new technique, combining the traditional Girard–Tait method with step-indexed logical relations,
which we use to show normalization for the logical fragment.
1
Introduction
The T RELLYS project is a collaborative initiative to design a dependently-typed language with simple
support for general recursion and other convenient but logically unsound features. To this end, the
present authors and their collaborators have proposed languages that are broken into two fragments: a
programmatic fragment with support for all the desired language features, and a logical fragment which
can reason about programs but is itself restricted for consistency [6, 31, 18, 29, 30].
As a simple example, consider the following natural number division function written in a Haskelllike syntax:
prog div : Nat -> Nat -> Nat
prog div n m = if n < m then 0 else 1 + (div (n - m) m)
This function computes the integer division of n by m unless m is 0, in which case it loops forever. We label
it “prog” to indicate it must be defined in the programmatic fragment described above. Disappointingly,
div can not be written directly in popular dependently-typed languages like Coq [28] or Agda [24]
because it is not total.
There are many sensible properties a programmer might wish to verify about div. For example,
that div 6 3 evaluates to 2, or that div n m <= n when m is not zero. Even though div itself is in
the programmatic fragment, we wish to state these properties in the consistent logical fragment. For
example:
log div63 : div 6 3 = 2
log div63 = refl
Above, the program (aka proof) div63 is tagged with “log” to indicate that it should be typechecked in
the logical fragment. The proof itself is just reflexivity, based on the operational behavior of div.
To encourage incremental verification, such a language should also include a way for programs
which are not known to be terminating to produce proofs. For example, programmers implementing a
J. Chapman and P. B. Levy (Eds.): Fourth Workshop on
Mathematically Structured Functional Programming (MSFP 2012).
EPTCS 76, 2012, pp. 25–39, doi:10.4204/EPTCS.76.4
Step-Indexed Normalization
26
complicated decision procedure might begin by writing in the programmatic fragment and come back
to prove termination at a later time. To support passing the proofs produced by such a procedure to the
logical fragment, the language may include an internalized logicality judgement—programs may assert
that other programs typecheck in a certain fragment. We use the new type form A@θ , where θ is L or P
for the logical or programmatic fragment, to claim that a term has type A in a particular fragment. For
example, a SAT solver which is not known to be terminating might be given the following type:
prog solver : (f : Formula) -> Maybe ((Satisfiable f) @ L)
Here, solver takes in some representation of a formula and optionally produces a proof that it is satisfiable. The type “(Satisfiable f) @ L” indicates that if a proof is produced, it will typecheck in the
logical fragment, even though the procedure itself is written in the programmatic fragment.
For these internalized judgements to be useful, the language must be able to produce them in one
fragment and use them in another. In general, any term which is produced in the logical fragment may be
safely used in the programmatic fragment. Additionally, values at certain “first-order” types (including
A@θ ) may be computed in the programmatic fragment and safely used by the logical fragment.
The metatheory of languages with this collection of features has proved challenging. This paper presents a new technique for demonstrating the normalization (and thus consistency) of the logical fragment in such a language. As we will show, direct adaptations of the Girard–Tait reducibility
method [15, 26] are insufficient. Since logical terms are permitted to make use of proofs produced
programmatically, it is necessary to simultaneously verify partial correctness properties of the programmatic fragment. To this end, our technique combines the traditional method with step-indexed logical
relations [2, 3].
Concretely, our contributions are:
• A small language with an internalized logicality judgement, sum types and recursive types (Section 2). While the language is insufficient for our examples, it retains enough features to exhibit
the difficulties we have encountered with traditional proofs (Section 3.2).
• A new, hybrid technique for proving normalization of the language’s logical fragment (Section 3.3).
This technique combines the Girard–Tait reducibility method with a step-indexed logical relation
for simultaneously verifying partial correctness properties of the programmatic fragment. This
combination seems to be essential to handle the internalized logicality judgement.
• A formalization of the language’s metatheory in Coq, including type safety and normalization
(Section 3.5).
• A comparison to related work on dependently-typed languages with general recursion and techniques for reasoning about them (Section 4).
The language we consider in this paper is simply-typed and thus insufficient to represent the examples
we have presented so far. However, this smaller language is still complex enough to exhibit the difficulties
we have encountered in proving normalization, and we are optimistic that our technique will scale up.
2
Language Definition
The language that we consider in this paper is a variant of the simply-typed call-by-value lambda calculus
with recursive types and general recursion. Its syntax is given in Figure 1. The chief novelty is the
presence of consistency classifiers θ . These classifiers are used by the typing judgement (written Γ ⊢θ
a : A) to divide the language into two fragments. The logical fragment, denoted by L, is a simply-typed
C. Casinghino, V. Sjöberg & S. Weirich
27
Types
A, B ::= Unit | A θ → B | A + B | A@θ | α | µ α .A
Terms
a, b ::= x | rec f x.a | a b | box a | unbox x = a in b
| () | inl a | inr a | case a of {inl x ⇒ a1 ; inr x ⇒ a2 } | roll a | unroll a
Language Classifiers
Environments
Values
θ ::= L | P
Γ ::= · | Γ, x :θ A
v ::= x | () | inl v | inr v | rec f x.a | box v | roll v
Syntactic Abbreviation:
λ x.a , rec f x.a
when f ∈
/ FV(a)
Figure 1: Syntax
lambda calculus with unit and sums. As we will show, all terms in this fragment are normalizing. The
programmatic fragment, denoted by P, adds general recursion and recursive types. The programmatic
fragment is a strict superset of the logical fragment: if Γ ⊢L a : A, then Γ ⊢P a : A as well.
Terms in the language may include subexpressions from both fragments. The box a term form and
corresponding A@θ type form mark such transitions. Intuitively, the judgement Γ ⊢θ box a : A@θ ′ holds
when fragment θ can safely observe that a has type A in the fragment θ ′ .
2.1 The typing judgement
We now describe the typing rules, given in Figure 2. As shown in rule TVAR, variables in the typing
context are tagged with a fragment. When a value is substituted for a variable, the value must check in
the corresponding fragment.
The fragments of the language may interact in several ways. Functions have arguments that are
tagged with consistency classifiers, as in A θ → B. The θ here specifies whether the function must be
applied to a logical or programmatic term. This classifier does not indicate in which fragment the function itself typechecks, and functions in each fragment are permitted to take arguments from the other.
Intuitively, the type may be read as “A@θ → B”, except that users need not explicitly box up arguments to functions. The rules for application (which involve the box form) ensure this does not cause
non-termination in the logical fragment, as we will discuss shortly.
There are two rules for type-checking functions. The first, TL AM, checks non-recursive functions in
the logical fragment. Here, λ x.b is syntax sugar for rec f x.b when f does not occur free in b. The second
rule, TR EC, checks (potentially) recursive functions in the programmatic fragment. Observe that, in both
cases, the consistency classifier for the argument is carried into the context when checking the body, but
does not directly influence the classifier of the function itself. The rules are otherwise standard.
The box a form effectively internalizes the typing judgement. It is checked by the three rules, describing the circumstances in which the fragments may safely talk about each other. The first rule, TB OX P,
says that the programmatic fragment may internalize any typing judgement—if a has type A in fragment
θ , then the programmatic fragment can observe that box a has type A@θ .
Rules TB OX L and TB OX LV check box in the logical fragment and are restricted to ensure termination. The former says that if a itself has type A in the logical fragment, then box a may also be formed in
the logical fragment (and checks at type A@θ for any θ , since logical terms are also programmatic). The
latter permits the logical fragment to observe that a term checks programmatically. In that case, the term
Step-Indexed Normalization
28
Γ ⊢θ a : A
x :θ A ∈ Γ
Γ ⊢θ x : A
Γ, x :θ A ⊢L b : B
Γ ⊢L λ x.b : A θ → B
TVAR
Γ ⊢θ a : A
Γ ⊢P box a : A@θ
Γ ⊢L a : A
Γ ⊢P a : A
Γ ⊢L a : A
Γ ⊢L box a : A@θ
TB OX P
Γ ⊢θ a : A@θ ′
′
Γ, x :θ A ⊢θ b : B
Γ ⊢θ unbox x = a in b : B
Γ, y :θ A, f :P A θ → B ⊢P a : B
Γ ⊢P rec f y.a : A θ → B
Γ ⊢P v : A
Γ ⊢L box v : A@P
TB OX L
Γ ⊢θ a : A θ → B
Γ ⊢θ box b : A@θ ′
Γ ⊢θ a b : B
TR EC
TB OX LV
′
TU NBOX
Γ ⊢P v : A FO (A)
Γ ⊢L v : A
TS UB
θ
Γ⊢ b:B
Γ ⊢θ inr b : A + B
TL AM
TI NR
Γ ⊢P a : [µ α .A/α ]A
Γ ⊢P roll a : µ α .A
TA PP
TFOVAL
Γ ⊢θ a : A
Γ ⊢θ inl a : A + B
Γ ⊢θ box a : (A1 + A2 )@θ ′
′
′
Γ, x :θ A1 ⊢θ b1 : B Γ, x :θ A2 ⊢θ b2 : B
Γ ⊢θ case a of {inl x ⇒ b1 ; inr x ⇒ b2 } : B
TROLL
TU NIT
Γ ⊢θ () : Unit
Γ ⊢P a : µ α .A
Γ ⊢P unroll a : [µ α .A/α ]A
TI NL
TC ASE
TU NROLL
FO (A)
FO (Unit)
FOU NIT
FO (A) FO (B)
FO (A + B)
FOS UM
Figure 2: Typing Rules
FO (A@θ )
FOAT
C. Casinghino, V. Sjöberg & S. Weirich
29
must be a value to ensure normalization. This restriction still permits the logical fragment to consider
programmatic terms (for example, recursive functions are values).
Rule TU NBOX checks the elimination form for boxed terms, which resembles a “let”. The term
unbox x = a in b typechecks when a has type A@θ ′ and b is parameterized by a value of type A in
fragment θ ′ . Intuitively, a will be evaluated first, eventually yielding a value box v, and v is substituted
into b. The operational semantics are discussed in more detail below. Note that no additional safety
restrictions occur in this rule—the box introduction rules handle everything required to ensure that the
logical fragment terminates.
The rule for function application, TA PP, makes use of the infrastructure for internalizing the typing
judgement. Recall that function types A θ → B demand arguments from a particular fragment. The box
and A@θ constructs already give us a way to safely check a term in different fragments, so we reuse
them here. To check the application a b in the fragment θ , we check that a has some function type
′
A θ → B, then check that box b can be given the type A@θ ′ in the current fragment. This has the effect of
restricting some applications to programmatic terms in the logical fragment—in general, programmatic
arguments to logical functions must be values, ensuring termination.
Rules TU NIT, TI NL and TI NR deal with the introduction forms for the unit and sum base types.
These terms may be used in either fragment and the typing rules are standard. Rule TC ASE checks
the pattern matching elimination form for sums. Notably, sums that typecheck in one fragment may
be eliminated in another—again we use the box infrastructure to ensure that this does not introduce
non-termination into the logical fragment.
Two rules describe the relationship between the fragments. As already discussed, any logical term
can be used programmatically—this is the content of rule TS UB. Rule TFOVAL is more surprising.
It allows potentially dangerous programmatic terms to be used in the logical fragment under certain
circumstances. In particular, the term must be a value (to ensure termination) and its type must be “first
order”. The first-order restriction, formalized by the FO (A) judgement in the same figure, intuitively
means that we move can move data but not computations from the programmatic fragment to the logical
one. For example, moving a natural number computed in P to L is safe, but moving a function from P to
L could cause non-termination when the function is applied.
Importantly, A@θ is a first-order type for any A. The programmatic fragment is permitted to compute logical values, including logical function values, and pass them back to the logical fragment. In a
language extended with dependent types, we believe this would be useful for working with proofs. For
example, a partial decision procedure could be written in the programmatic fragment and the resulting
proofs could be used in the logical fragment if the procedure terminates.
Finally, the language includes iso-recursive types [25]. These are checked by the two rules TROLL
and TU NROLL. Recursive types are restricted to the programmatic fragment because they can introduce
non-termination.
2.2 Operational Semantics
The language’s operational semantics are given in Figure 3. We use standard call-by-value evaluation
contexts and a small-step reduction relation. Note that reduction occurs inside box a terms, motivating
some of the restrictions from the previous section. The multi-step reduction relation is indexed by a
natural number—this will be useful in the step-indexed logical relation defined in Section 3.
Step-Indexed Normalization
30
Evaluation contexts E ::= [·] | [·]b | v[·] | inl [·] | inr [·] | case [·] of {inl x ⇒ a1 ; inr x ⇒ a2 }
| box [·] | unbox x = [·] in a | roll [·] | unroll [·]
a;b
(rec f x.a) v ; [v/x][rec f x.a/f ]a
SB ETA
case inl v of {inl x ⇒ a1 ; inr x ⇒ a2 } ; [v/x]a1
unbox x = box v in b ; [v/x]b
SC ASE L
case inr v of {inl x ⇒ a1 ; inr x ⇒ a2 } ; [v/x]a2
SC ASE R
a;b
E [a] ; E [b]
SU NROLL
SC TX
a ;∗ b
a ;n b
a ;0 a
unroll (roll v) ; v
SU NBOX
MSR EFL
a ;k b b ; b′
a ;(k+1) b′
MSS TEP
a ;k b
a ;∗ b
ASA NY
Figure 3: Operational Semantics
3
Metatheory
We now consider the metatheory of the small language presented in Section 2. Of particular interest is
the normalization result for the logical fragment, for which we employ a novel combination of traditional
and step-indexed logical relations. We motivate and explain this technique in Section 3.3. The system
also enjoys standard type-safety properties, as we show in Section 3.1. All the results presented here have
been mechanized using the Coq theorem prover, and we briefly describe the formalization in Section 3.5.
For this reason, we focus on a high-level description of the techniques and elide most proofs.
3.1 Type Safety
We prove type safety via syntactic progress and preservation theorems [32]. The progress result is direct
by induction on typing derivations, using appropriate canonical forms lemmas.
Theorem 1 (Progress). If · ⊢θ a : A then either a is a value or a ; a′ for some a′ .
For preservation, a substitution lemma is required. Because variables are values and our language
includes a value restriction (in the TB OX LV rule), we prove the substitution lemma only for values.
Lemma 2 (Substitution). If Γ, x :θ B ⊢θ a : A and Γ ⊢θ v : B, then Γ ⊢θ [v/x]a : A.
′
′
Since we employ a call-by-value operational semantics, this substitution lemma is enough to prove
preservation.
Theorem 3 (Preservation). If Γ ⊢θ a : A and a ; a′ , then Γ ⊢θ a′ : A.
C. Casinghino, V. Sjöberg & S. Weirich
31
3.2 Adapting the Girard-Tait Reducibility Method
To motivate the use of step-indexed logical relations in our normalization proof, we will first revisit the
standard Girard–Tait reducibility method [15, 26] and examine why more direct adaptations of it fail.
Traditional techniques for proving strong normalization typically begin by defining the “interpretation”
of each type. That is, for each type A, a set of terms JAK is defined approximating the type A and
where each term in the set is known to be strongly normalizing. Then a “soundness” theorem is proved,
demonstrating that if a has type A then a ∈ JAK. This implies a is strongly normalizing.
3.2.1
First attempt: ignoring the programmatic fragment
We begin by modify this technique in two ways to fit our setting. First, since we have a deterministic
call-by-value operational semantics, the interpretation of each type will be a set of values (not arbitrary
terms). Second, since the terms at a given type differ in the programmatic and logical fragments, we
index the interpretation by θ , writing JAKθ .
It is tempting to think that, because we do not care about the normalization behavior of the programmatic fragment, the programmatic interpretation of types can be very simple. Perhaps, for example, just
the well-typed values of the appropriate type will do. Consider the following interpretation:
JAKP
= {v | · ⊢P v : A}
JUnitKL
JA + BKL
JA θ → BKL
JA@θ KL
Jµ α .AKL
Jα KL
=
=
=
=
=
=
{()}
{inl v | v ∈ JAKL} ∪ {inr v | v ∈ JBKL}
{λ x.a | · ⊢L λ x.a : A θ → B and for any v ∈ JAKθ , [v/x]a ;∗ v′ ∈ JBKL}
{box v | v ∈ JAKθ }
0/
0/
Here, the logical interpretation of Unit contains only (). The logical interpretation of a sum type A + B
contains inl v for every v in the interpretation A, and inr v for every v in the interpretation of B. The logical
interpretation of functions types is standard, except for the addition of the consistency classifier: A θ → B
contains the term λ x.a if, for any v in the interpretation of the domain, [v/x]a reduces to a value in the
interpretation of the range (“related functions take related arguments to related results”). The logical
interpretation of A@θ comprises the values box v where v is in JAKθ . Finally, the logical interpretations
of recursive types and type variables are empty, since these are used only in the programmatic fragment.
Before we can state a soundness theorem, we must account for contexts. We use ρ for mappings of
variables to terms, and write Γ |= ρ if x :θ A ∈ Γ implies ρ x ∈ JAKθ . We let ρ a stand for the simultaneous
replacement of the variables in a by the corresponding terms in ρ .
In this setting, we would hope to be able to prove the following soundness theorem:
Soundness (take 1): Suppose Γ ⊢L a : A and Γ |= ρ . Then ρ a ;∗ v ∈ JAKL .
In a proof by induction on the typing derivation, most of the cases offer little resistance (the interested
reader is encouraged to write out the case for the TL AM and TA PP rules). However, the proof gets stuck
at the case for the first order rule:
Γ ⊢P v : A FO (A)
Γ ⊢L v : A
TFOVAL
Step-Indexed Normalization
32
Here, we must show that ρ v ∈ JAKL (substituting values into a value produces a value, so ρ v does not
step). However, since the premise is in the programmatic fragment, we have no induction hypothesis for
v. If A = Unit, we can complete the case using a canonical forms lemma (since we know by a substitution
lemma that · ⊢L ρ v : Unit). However, if A is B@L we are stuck. We could use a canonical forms lemma
to observe that ρ v must have the shape box v′ , but no induction hypothesis for v′ is available.
3.2.2
Second attempt: partial correctness for the programmatic fragment
Our previous attempt failed because the language permits values of first-order types to move from the
programmatic fragment to the logical fragment, but the theorem we were trying to prove didn’t capture
any information about the programmatic fragment. To fix this, we might try making two changes. First,
the programmatic and logical interpretations should agree at first-order types. Second, the programmatic
interpretation and the soundness theorem should be modified to prove a partial correctness result for the
programmatic fragment—we’ll need to know that if a programmatic term normalizes, then it is in the
appropriate interpretation.
These changes should allow us to handle the previously problematic TFOVAL case. Consider the
following modified interpretation, ignoring recursive types for the moment:
JUnitKθ
JA + BKθ
JA θ → BKL
JA θ → BKP
JA@θ ′ Kθ
=
=
=
=
{()}
{inl v | v ∈ JAKθ } ∪ {inr v | v ∈ JBKθ }
{λ x.a | · ⊢L λ x.a : A θ → B and for any v ∈ JAKθ , [v/x]a ;∗ v′ ∈ JBKL}
{rec f x.a | · ⊢P rec f x.a : A θ → B
and for any v ∈ JAKθ , if [v/x][rec f x.a/f ]a ;∗ v′ then v′ ∈ JBKP }
′
= {box v | v ∈ JAKθ }
Here, the logical interpretation is unchanged. The programmatic interpretation of the first-order types is
now the same as the logical interpretation. Finally, we have modified the programmatic interpretation of
function types to state a partial correctness property: if a function terminates when passed a value in the
interpretation of its domain, then the result must be in the interpretation of its range. We now restate the
soundness theorem similarly.
Soundness (take 2): Suppose Γ ⊢θ a : A and Γ |= ρ .
• If θ is L, then ρ a ;∗ v ∈ JAKL.
• If θ is P and ρ a ;∗ v, then v ∈ JAKP .
With the modified interpretation and soundness theorem, the TFOVAL case now goes through. Because
the rule only applies to values, the theorem now yields a useful induction hypothesis for the premise.
Unfortunately, this style of definition introduces a new problem: the programmatic interpretation of
recursive types. The previous definition (from Section 3.2.1) is insufficient to handle the TU NROLL case
of the new soundness theorem. To extend our partial correctness property, we might demand that when
unrolling results in a value, that value is in the interpretation of the unrolled type:
Jµ α .AKP = {roll v | · ⊢P roll v : µ α .A and v ∈ J[µ α .A/α ]AKP }
However, this is not a valid definition. If the interpretation is a function defined by recursion on the
structure of types, the substitution in J[µ α .A/α ]AKP ruins its well-foundedness.
C. Casinghino, V. Sjöberg & S. Weirich
33
3.3 A step-indexed interpretation
Happily, a technique exists in the literature to cope with the circularity introduced by iso-recursive types.
Step-indexed logical relations [2, 3] add an index to the interpretation, indicating the number of available
future execution steps. Terms in the relation are guaranteed to respect the property in question only
for the number of steps indicated. The interpretation is defined recursively on this additional index,
circumventing the circularity problem we encountered above.
Step-indexed logical relations intuitively describe partial-correctness properties—terms are certified
to be well behaved for a finite number of steps. For this reason, they have typically been used to prove
safety and program equivalence properties, not normalization. We will adopt a hybrid approach, where
the indices track execution of subterms in the programmatic fragment (where we need a partial correctness result) but not in the logical fragment (for which we are proving normalization).
Following Ahmed [2], our interpretation is split into two parts. The value interpretation V [[A]]θk
resembles the interpretations shown in the previous sections. The k index here indicates that when
a value appears in a larger term, its programmatic components will be “well-behaved” for at least k
steps of computation. The computational interpretation C [[A]]θk contains closed terms, not just values.
Its definition resembles the statement of the soundness theorem from the previous section, with steps
counted explicitly. Terms in C [[A]]Lk are guaranteed to normalize to values in V [[A]]Lk . On the other hand,
we have a partial correctness property for terms in C [[A]]Pk —if they reach a value in j steps for some
j ≤ k, then the value is in V [[A]]Pk−j .
V [[Unit]]θk
V [[A + B]]θk
V [[A@θ ′ ]]θk
′
V [[A θ → B]]Lk
V [[µ α .A]]Lk
V [[µ α .A]]Pk
{()}
{inl v | v ∈ V [[A]]θk } ∪ {inr v | v ∈ V [[B]]θk }
′
{box v | v ∈ V [[A]]θk }
′
{rec f x.a | · ⊢L rec f x.a : A θ → B
′
and ∀ j ≤ k, if v ∈ V [[A]]θj then [v/x]a ∈ C [[B]]Lj }
′
= {rec f x.a | · ⊢P rec f x.a : A θ → B
′
and ∀ j < k, if v ∈ V [[A]]θj then [v/x][rec f x.a/f ]a ∈ C [[B]]Pj }
= 0/
= {roll v | · ⊢P roll v : µ α .A and ∀ j < k, v ∈ V [[[µ α .A/α ]A]]Pj }
C [[A]]Pk
C [[A]]Lk
= {a | · ⊢P a : A and ∀ j ≤ k, if a ;j v then v ∈ V [[A]]P(k−j) }
= {a | · ⊢L a : A and a ;∗ v ∈ V [[A]]Lk }
V [[A θ → B]]Pk
′
=
=
=
=
The value interpretation is similar to the proposed interpretation in the previous section, with two
changes. First, the function type cases now refer to the computation interpretation rather than explicitly mentioning the reduction behavior. Second, the step indices track reductions in the programmatic
fragment. In particular, note that the programmatic interpretation of function types demands that related
functions take related arguments to related results at all strictly smaller indices, effectively counting the
one beta reduction step that this definition unfolds. The beta step in the logical interpretation is not
counted, since we are tracking only the reduction of programmatic components.
Unlike the proposed definition from the previous section, this interpretation is well defined. We can
formalize its descending well-founded metric as a lexicographically ordered triple (k, A, I ): here, k is
the index, A is the type and I is one of C or V with V < C . The third element of the triple tracks which
interpretation is being called—the computational interpretation may call the value interpretation at the
same index and type, but not vice-versa.
Step-Indexed Normalization
34
3.4 Normalization
The step-indexed interpretation from the previous section repairs the problems encountered in the first
two proposed interpretations and can be used to prove normalization for the logical fragment. Since
our results are formalized in Coq, we give only a high-level overview of the proof here. To begin, we
must update the Γ |= ρ judgement to account for steps. We now write Γ |=k ρ when x :θ A ∈ Γ implies
ρ x ∈ V [[A]]θk .
Three key lemmas are needed in the main soundness theorem. The first is a standard “downward
closure” property that often accompanies step-indexed logical relations. This lemma captures the idea
that we build a more precise interpretation of a type by considering terms that must be valid for more
steps.
Lemma (Downward Closure): For any A and θ , if j ≤ k then V [[A]]θk ⊆ V [[A]]θj and C [[A]]θk ⊆
C [[A]]θj .
We have two lemmas relating the programmatic and logical interpretations, corresponding to the TFOVAL and TS UB typing rules. The first says that the two interpretations agree on first-order types:
Lemma: If FO (A), then V [[A]]Lk = V [[A]]Pk .
The second captures the idea that the logical fragment is a subsystem of the programmatic fragment:
Lemma: For any A and k, V [[A]]Lk ⊆ V [[A]]Pk and C [[A]]Lk ⊆ C [[A]]Pk .
The content of the soundness theorem is essentially the same as in our second failed attempt, but
we can now state it more directly, using the computational interpretation. The theorem is proved by
induction on the typing derivation, using the lemmas outlined above.
Theorem (Soundness): If Γ ⊢θ a : A and Γ |=k ρ , then ρ a ∈ C [[A]]θk .
The normalization of the logical fragment is a direct consequence of this theorem and the definition of
the interpretation.
Lemma (Normalization): If · ⊢L a : A then there exists a value v such that a ;∗ v.
3.5 Formalization
The proof outline above has been formalized with the Coq proof assistant [28]. The proof scripts are
written in a heavily automated style, inspired by Chlipala’s work on practical dependently typed programming [8, 7]. They are available for download at the first author’s website:
http://www.seas.upenn.edu/~ccasin/papers/step_normalization.tar.gz.
The language formalized differs in several minor ways from the one presented in this paper. Namely,
• de Bruijn indices are used for binding instead of explicit names.
• Rather than being syntactic sugar, λ x.a is a separate form in the grammar of expressions.
• The reduction relation is formalized with explicit congruence rules rather than evaluation contexts.
• The formalized language includes natural numbers, but not unit.
Additionally, to prove certain facts about the interpretation, we found it necessary to add a standard
axiom of functional extensionality to Coq. This axiom is known to be consistent with Coq’s logic [27].
C. Casinghino, V. Sjöberg & S. Weirich
4
35
Related Work
4.1 Step-indexed logical relations
Our proof technique draws heavily from previous work on step-indexed logical relations. The idea to
approximate models of programming languages up to a number of remaining execution steps originated
in the work of Appel and McAllester on foundational proof-carrying code [3]. They observed that the
step indices allowed a natural interpretation of recursive types. Subsequently, Ahmed extended this
technique to languages involving impredicative polymorphism, mutable state and other features [2, 1].
Hobor, Dockins and Appel have proposed a general theory of indirection which captures many of the
common use-cases for step-indexed models [16]. They provide a general framework for applying these
approximation techniques to resolve certain types of apparent circularity (similar to the problems with
recursive types described above). In a recent draft [14], Dockins and Hobor have used this framework
to provide a Hoare logic of total correctness for a small language with function pointers and semantic
assertions. This work is closely related to the present development, but with different goals: they prove
the soundness of a logic which can reason about termination, while we prove that every term in the
logical fragment of our language terminates. We have not yet investigated whether their framework can
be adapted to our setting, but this connection is a promising avenue for future work.
4.2 Other approaches to recursion and partiality
Many authors have considered language features to model partiality and recursion in a consistent dependent type theory. The language described in the present paper is much simpler, but our goal is to provide
a foundation from which we may scale up to full dependent types, so we compare with some of the most
closely related approaches.
Partiality monad Capretta proposed representing potentially non-terminating computations via a coinductive partiality monad [5]. This technique can be used in existing languages like Coq and Agda, which
already support coinduction [10]. For example, Agda’s partiality monad has been used to present subtyping for recursive types [13] and represent potentially infinite parsing trees [12].
There are several differences between these approaches and the one outlined in this paper. Coinduction is a very general method for representing infinite data, which we do not consider. Our approach
has the advantage that terminating and potentially partial functions are defined and reasoned about in the
same way. By contrast, the reasoning principles for coinductively defined functions in Coq and Agda
require the user to consider so-called guardedness conditions that are not present for terminating functions. More, we are optimistic that splitting the language into two fragments will allow us to include
various other potential sources of logical unsoundness uniformly, restricting them to the programmatic
fragment. Admittedly, it remains to be seen how well this will work in practice and whether our proof
technique will scale.
Partial Types Constable and Smith [9] proposed adding partiality to the Nuprl type theory through the
addition of a type A of potentially nonterminating computations of type A. The general fixpoint operator,
for defining recursive computations then has type
(A → A) → A.
36
Step-Indexed Normalization
However, to preserve the consistency of the logic in dependent type theories, the type A must be restricted
to admissible. types. Crary [11] provides an expressive axiomatization of admissible types, but the
resulting conditions lead to significant proof obligations, in particular when using Σ types. Although we
have not yet formally proven the soundness of the system with arbitrary dependent types (including Σ
types), we do not believe that there will be any restrictions on the programmatic language, similar to
admissibility.
The “later” modality Nakano [23] introduced the “later” modality to define a total language with
guarded recursive types. Intuitively, a term of type •A (pronounced “later A”) will be useable as a term
of type A in the future. The recursive type µ α .A then unfolds to [•µ α .A/α ]A rather than [µ α .A/α ]A.
Using this modality, he is able to give the type (•A → A) → A to the Y combinator. This type allows programmers to define a variety of recursive functions while still ensuring that the language is normalizing.
Nakano uses a step-indexed realizability interpretation to prove the normalization property for his langauge, suggesting deep connections with the present work. One substantial difference is that Nakano’s
calculus is not call-by-value.
The later modality has been used by subsequent authors to design langauges for a variety of purposes.
Krishnaswami and Benton use it define a total language for functional reactive programming [20, 19].
Birkedal et al. [4] study the topos of trees, which they observe can model an extension of Nakano’s calculus to a full dependent type theory with guarded recursion. While these authors do not consider languages
with partiality and their settings have substantial differences from our own, their success in extending
step-indexing and closely-related techniques to model recursion in larger languages is promising.
Other T RELLYS approaches The T RELLYS group has been working simultaneously on an alternative
design, where the logical and programmatic languages are completely separate at a syntactic level [18].
This considerably simplifies the metatheory for the logical language, which is no longer a general programming language but rather a collection of principles for reasoning about the programmatic language.
On the other hand, it can restrict the expressiveness of the logic and create duplication between the two
fragments. We are exploring these trade-offs in our ongoing research.
4.3 Modal type systems for distributed computation
Modal logics allow one to reason from multiple perspectives, called “possible worlds”. It is tempting
to view the language presented here as such a system, where the possible worlds are θ , the logical and
computational fragments of the language.
One way to define a modal logic is to make the world explicit, for example using a judgement Γ ⊢θ A,
stating that under the assumptions in Γ, the proposition A is true at the world θ . Each assumption in the
context is tagged with the world where it holds (θ , A).
(θ , A) ∈ Γ
Γ ⊢θ A
In such as system, the at modality [17], internalizes the typing judgment into a proposition, with
introduction form
′
Γ ⊢θ A
Γ ⊢θ A@θ ′
and elimination form:
Γ ⊢θ A@θ ′
Γ, (θ ′ , A) ⊢θ C
Γ ⊢θ C
C. Casinghino, V. Sjöberg & S. Weirich
37
Our first-order rule is similar to the perspective-shifting rule, called get, from ML5 [22, 21].
Γ ⊢θ A A mobile
Γ ⊢θ A
′
This rule, shown above, allows a class of propositions to be directly translated between worlds. The class
of mobile types is very similar to our class of first-order types. For example, base types (such as strings
and integers) and the at modality (A@θ ) are always mobile, sums (A + B) are mobile only when their
components (A and B) are mobile, but implications are never mobile.
One difference between our system and modal logics is our treatment of implication (i.e. function
types). The functions in our system are annotated with a domain fragment, but this is not typically
the case in modal logics, where the domain and range of implications are in the same world. Such an
approach is incompatible with our subsumption rule:
Γ ⊢L a : A
Γ ⊢P a : A
TS UB
Suppose A were a function type B1 → B2 with no tag on the domain. When we defined such a function
in the logical fragment, the function’s body could make use of the fact that its argument checks logically.
If the subsumption rule were used to transport the function to the programmatic fragment, it could be
applied to terms that check only programmatically, potentially violating assumptions of its body.
5
Conclusion
In this paper, we have presented a small language with two fragments. The programmatic fragment
supports general recursion and recursive types, while every term in the logical fragment is normalizing.
Despite these differences, each fragment may explicitly mention and manipulate terms from the other
using the internalized logicality type, A@θ . We showed that direct adaptations of the Girard–Tait reducibility method fail to yield a normalization proof for the logical fragment. Finally, we proposed a
new technique involving step-indexed logical relations and used it to complete the proof, which has been
formalized in Coq.
The language considered here is small and unsuitable for real programming tasks. However, it constitutes the core of one of our designs for the T RELLYS programming language, and the metatheoretic
difficulties we explained and solved in this paper also appear there. In future work, we plan to add
polymorphism, type-level computation and dependent types back to this system. If our proof technique
scales, this will provide the basis for a practical, dependently-typed programming language which can
naturally express and reason about non-terminating computations.
References
[1] Amal Ahmed (2004): Semantics of Types for Mutable State. Ph.D. thesis, Princeton University. Available at
http://www.cs.princeton.edu/research/techreps/TR-713-04.
[2] Amal Ahmed (2006): Step-Indexed Syntactic Logical Relations for Recursive and Quantified Types. In:
ESOP, 2006, doi:10.1007/11693024_6.
[3] Andrew W. Appel & David McAllester (2001): An indexed model of recursive types for foundational proofcarrying code. ACM Trans. Program. Lang. Syst. 23(5), pp. 657–683, doi:10.1145/504709.504712.
38
Step-Indexed Normalization
[4] Lars Birkedal, Rasmus Ejlers Mogelberg, Jan Schwinghammer & Kristian Stovring (2011): First Steps in
Synthetic Guarded Domain Theory: Step-Indexing in the Topos of Trees. In: LICS 2011, pp. 55–64, doi:10.
1109/LICS.2011.16.
[5] Venanzio Capretta (2005): General Recursion via Coinductive Types. Logical Methods in Computer Science
1(2), pp. 1–18, doi:10.2168/LMCS-1(2:1)2005.
[6] Chris Casinghino, Harley D. Eades III, Garrin Kimmell, Vilhelm Sjöberg, Tim Sheard, Aaron Stump &
Stephanie Weirich: The Preliminary Design of the Trellys Core Language. Available at http://www.seas.
upenn.edu/~ccasin/papers/plpv11_slides.pdf. Talk and discussion session at PLPV 2011.
[7] Adam Chlipala (2010): An Introduction to Programming and Proving with Dependent Types in Coq. Journal
of Formalized Reasoning 3(2), pp. 1–93. Available at http://adam.chlipala.net/papers/CpdtJFR/.
[8] Adam Chlipala (2011): Certified Programming with Dependent Types.
chlipala.net/cpdt/.
Available at http://adam.
[9] Robert L. Constable & Scott Fraser Smith (1987): Partial Objects in Constructive Type Theory. In: LICS,
1987, pp. 183–193.
[10] Thierry Coquand (1994): Infinite objects in type theory. In: Proceedings of the international workshop on
Types for proofs and programs, Springer-Verlag New York, Inc., Secaucus, NJ, USA, pp. 62–78. Available
at http://dl.acm.org/citation.cfm?id=189973.189976.
[11] Karl Crary (1998): Type Theoretic Methodology for Practical Programming Languages. Ph.D. thesis, Cornell
University.
[12] Nils Anders Danielsson (2010): Total parser combinators. In: ICFP, 2010, ACM, New York, NY, USA, pp.
285–296, doi:10.1145/1863543.1863585.
[13] Nils Anders Danielsson & Thorsten Altenkirch (2010): Subtyping, Declaratively. In Claude Bolduc, Jules
Desharnais & Bchir Ktari, editors: Mathematics of Program Construction, Lecture Notes in Computer Science 6120, Springer Berlin / Heidelberg, pp. 100–118, doi:10.1007/978-3-642-13321-3_8.
[14] Robert Dockins & Aquinas Hobor (2010): A Theory of Termination via Indirection. In Amal Ahmed,
Nick Benton, Lars Birkedal & Martin Hofmann, editors: Modelling, Controlling and Reasoning About
State, Dagstuhl Seminar Proceedings 10351, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, Germany,
Dagstuhl, Germany. Available at http://drops.dagstuhl.de/opus/volltexte/2010/2805.
[15] Jean-Yves Girard (1972): Interprétation fonctionelle et élimination des coupures de l’arithmétique d’ordre
supérieur. Ph.D. thesis, Université Paris VII.
[16] Aquinas Hobor, Robert Dockins & Andrew W. Appel (2010): A Theory of Indirection via Approximation. In:
POPL, 2010, pp. 171–185. Available at http://msl.cs.princeton.edu/indirection.pdf.
[17] Limin Jia & David Walker (2004): Modal proofs as distributed programs (Extended Abstract). In: ESOP,
2004, Springer, pp. 219–233, doi:10.1007/978-3-540-24725-8_16.
[18] Garrin Kimmell, Aaron Stump, Harley D. Eades III, Peng Fu, Tim Sheard, Stephanie Weirich, Chris Casinghino, Vilhelm Sjöberg, Nathan Collins & Ki Yung Ahn (2012): Equational Reasoning about Programs
with General Recursion and Call-by-value Semantics. In: PLPV, 2012, doi:10.1145/2103776.2103780.
[19] Neelakantan Krishnaswami & Nick Benton (2011): A semantic model for graphical user interfaces. In:
ICFP, 2011, ACM, New York, NY, USA, pp. 45–57, doi:10.1145/2034773.2034782.
[20] Neelakantan Krishnaswami & Nick Benton (2011): Ultrametric Semantics of Reactive Programs. In: LICS,
2011, pp. 257–266, doi:10.1109/LICS.2011.38.
[21] Tom Murphy, VII (2008): Modal Types for Mobile Code. Ph.D. thesis, Carnegie Mellon. Available at
http://tom7.org/papers/. Available as technical report CMU-CS-08-126.
[22] Tom Murphy, VII, Karl Crary & Robert Harper (2007): Type-safe Distributed Programming with ML5. In:
Trustworthy Global Computing 2007, doi:10.1007/978-3-540-78663-4_9.
[23] Hiroshi Nakano (2000): A Modality for Recursion. In: LICS, 2000, IEEE Computer Society, Washington,
DC, USA, pp. 255–, doi:10.1109/LICS.2000.855774.
C. Casinghino, V. Sjöberg & S. Weirich
39
[24] Ulf Norell (2007): Towards a practical programming language based on dependent type theory. Ph.D. thesis,
Department of Computer Science and Engineering, Chalmers University of Technology.
[25] Benjamin C. Pierce (2002): Types and Programming Languages. MIT Press.
[26] William W. Tait (1967): Intensional Interpretations of Functionals of Finite Type I. The Journal of Symbolic
Logic 32(2), pp. pp. 198–212, doi:10.2307/2271658.
[27] The Coq Development Team (2011): The Coq Proof Assistant, Frequently Asked Questions. INRIA. Available at http://coq.inria.fr/faq/.
[28] The Coq Development Team (2011): The Coq Proof Assistant Reference Manual, Version 8.3. INRIA.
Available at http://coq.inria.fr/V8.3/refman/.
[29] Stephanie Weirich (2011): Combining Proofs and Programs. Invited lecture for RTA 2011 and TLCA 2011,
Novi Sad, Serbia.
[30] Stephanie Weirich (2011): Combining Proofs and Programs. Presentation at DTP 2011, Shonan Meeting
Seminar 007, Japan.
[31] Stephanie Weirich (2011): Combining Proofs and Programs in Trellys. Plenary Address at MFPS 26, Pittburgh, PA.
[32] Andrew K. Wright & Matthias Felleisen (1992): A Syntactic Approach to Type Soundness. Information and
Computation 115, pp. 38–94, doi:10.1006/inco.1994.1093.
| 6cs.PL
|
1
Towards a Functional System Architecture for
Automated Vehicles
arXiv:1703.08557v2 [cs.SY] 30 Mar 2017
Simon Ulbrich, Andreas Reschka, Jens Rieken, Susanne Ernst, Gerrit Bagschik, Frank Dierkes, Marcus Nolte,
and Markus Maurer
Abstract—This paper presents a functional system architecture
for an automated vehicle. It provides an overall, generic structure
that is independent of a specific implementation of a particular
vehicle project. Yet, it has been inspired and cross-checked with
a real world automated driving implementation in the Stadtpilot project at the Technische Universität Braunschweig. The
architecture entails aspects like environment and self perception,
planning and control, localization, map provision, Vehicle-To-X
communication, and interaction with human operators.
Index Terms—Functional System Architecture, Automated Vehicles, Autonomous Driving
Ulbrich et al. [14], or functional safety considerations resulting
from the application of the architecture in the aFAS project
[15].
This paper is organized as follows. First different approaches
for structuring driving tasks and their processing levels are
introduced and compared with other approaches in the literature. Then the functional system architecture from this paper
is presented by outlining its main columns and clarifying
its interfaces and comprised activities. For each aspect the
modifications to the state of the art are presented and provided
with a root cause for these. In the end, open issues are
highlighted and a conclusion is drawn.
I. I NTRODUCTION
W
HEN software is developed for an automated vehicle,
it is a bottom-up process for many teams. If existing
building blocks are just hacked together in some way, this will
lead to complex system designs. Yet, having a well structured
functional system architecture is key. It has central impact on
the system design and technical software development for an
automated vehicle — often for several years.
This paper presents an overall functional system architecture
for an automated vehicle. Implementation independent modules are grouped such that there are clean interfaces among
these modules.
This functional system architecture differs from others by
strictly using hierarchy and functional separation. It underwent
several iterations. Earlier versions of this architecture have
been published by [1]–[9]. It has strongly been influenced by
the functional system architecture developed by Dickmanns’
group [10]–[13]. Matthaei & Maurer [8], and Matthaei [9, p.
37 ff.]. The following architecture discussions will be based
on the last architecture revision in English [8].
Some refinements have already been published in German
[9, p. 37 ff.]. A goal of this article is to make our research
insights accessible to the international scientific community.
Apart from that, several other enhancements have been made
to incorporate more recent research such as the definition of
interfaces, e.g., the definition of a scene and a situation in
S. Ulbrich was with the Institute of Control Engineering at TU Braunschweig, Germany and is now with Audi AG.
A. Reschka, J. Rieken, S. Ernst, G. Bagschik, F. Dierkes, and M. Nolte are
with the Institute of Control Engineering at TU Braunschweig, Germany. A.
Reschka is currently visiting student researcher at Stanford University.
M. Maurer is professor and head of the Institute of Control Engineering at
TU Braunschweig, Germany.
Manuscript received March 10th, 2017.
II. BACKGROUND
The ISO 26262 standard proposes a functional system architecture as a part of the system design and defines “modularity”, “adequate level of granularity”, and “simplicity” as
requirements to the architectural design [16, part 4, p. 13].
As a property of a modular system design the ISO 26262
proposes “hierarchical design”, “precisely defined interfaces”,
“maintainability”, “testability”, and “avoidance of unnecessary
complexity” (ibid.).
To provide a structure for human behavior, Rasmussen [17]
distinguishes “skill-based behavior”, “rule-based behavior”,
and “knowledge-based behavior” as three levels of performance of skilled human operators. On the lowest, skill-based
level, reactive, sensory-motor activities take place without
conscious control. On a rule-based level, decisions are taken
based on a previously stored set of rules. If a situation is not
familiar and there is no stored rule for it, knowledge-based
behavior may be applied. Here a new strategy for goal archival
is developed from existing knowledge.
Donges distinguishes “navigation”, “guidance”, and “stabilization” as three hierarchical levels of driving tasks in his
publication from 1982 cited in [18]. Similar to Riemersma
[19] and Michon [20, p. 489] citing his inaugural lecture from
1971, one elementary (operational) layer is used for course
keeping and speed control, a second (tactical) layer is for any
behavior planning, and a third one is for strategic planning.
Hale et al. [21, p. 1383] suggest that the three levels of driving
tasks and the three levels of Rasmussen are rather orthogonal
to each other. While Donges’ driving tasks address what task
is to be solved, Rasmussen addresses how it is solved. Table I
illustrates the relationship between both. Most of the time
2
TABLE I
E XAMPLES FOR DRIVING TASKS AND THEIR PROCESSING LEVELS BASED
ON H ALE et al. [21, P. 1383] DEPICTED AS IN M UIGG [22, P. 8]
Driving task
Skill-based
Navigation
Guidance
Stabilization
Daily
commute
Negotiating
familiar
junctions
Road
following
around corners
Processing level
Rule-based
Knowledge-based
Choice
Navigating in
between
foreign town
familiar routes
Controlling a
Passing other
skid on icy
car
roads
Driving an
unfamiliar car
Learner on
first lesson
stabilization tasks are handled on a skill-based level. However,
if for example a car is unknown to the driver, he or she may not
have these subconscious skills to address the task. Yet, there
are learned rules that may be used. If even these rules need to
be formed from knowledge, because it is the first time a fresh
learner drives a car, it would be knowledge-based behavior.
On a tactical level, passing another car typically involves
situation assessment and a certain amount of stored rules and
experience, e.g., how much of a gap is necessary to overtake.
Humans address this typically by rule-based behavior. Last of
all, navigation may display the widest variety of processing
levels in everyday driving. Blindly commuting the same street
every day without looking for, e.g., changed traffic signs,
may almost be skill-based behavior. It becomes rule-based,
if it involves active tactical decisions between several route
options, and turns into knowledge-based behavior, if a driver
navigates in an unknown city for the first time.
Transferring the concept of driving tasks and processing levels
from human drivers to a technical system provides a starting
point for a technical architecture. In fact, this provides a
hierarchical abstraction of driving tasks as in Maurer [13] and
Matthaei et al. [23]. Another distinction of tasks for automated
driving may be derived from different processing steps of
perceiving and acting.
Extending work by Zapp [24], who described a functional
control-cycle for automated vehicles, Hock et al. [10] showed
an inverted “U” shaped signal flow from sensors to actuators with a hierarchical separation of processing levels for
driving tasks. Further specification and an exhaustive system
description can be found in Dickmanns et al. [11]. For
instance in [25, p. 239] and more clearly in [12, p. 441],
Dickmanns presents the concept of separating “recognition”
from “behavior (execution)” as well as the idea of aggregating
features into abstract symbolic representations. Maurer [13]
and Dickmanns [12, p. 185] highlight multiple feedback loops
at different hierarchical levels constituting the signal flow
in our architecture nowadays. In [25], [26, p. 595], [12,
p. 387], Dickmanns also illustrates the usage of a dynamic
knowledge base and background knowledge that is now named
as context1 modeling in this architecture. Additionally, [13, p.
40 ff.], [27, p.73 ff.], [28, p. 64], [12, p. 442] identify the
1 Context is here understood as the part of discourse that surrounds and
represents an element.
central role of system capabilities and translate them into a
hierarchical structure of abstraction. [12, p. 442] separates
between “scene understanding”, “planning”, and “gaze and
locomotion control”. As in [13], the situation assessment is
rather part of the perception. In [11], situation assessment
stretches into both worlds: the planning column as well as
the perception column. In this article, the goal- and valueindependent scene/context modeling is considered to be part of
the perception column. The goal- and value-specific situation
extraction and situation assessment is considered to be part
of the planning and control column. The idea of “situation
aspects” as a result of a “situation assessment” is presented in
[28, p. 51 ff.] for automated driving.
Matthaei [9, p. 25 ff.] provides a more comprehensive literature
review on the various forms of functional system architectures
that have been used by different teams in automated driving
as well as in robotics. For a broad literature review the reader
is referred to his dissertation. Yet, we would like to relate
our article to some recent publications on functional system
architectures for automated vehicles.
Tas et al. [29] compare the functional system architectures of
several automated vehicles. They summarize the advantages
and disadvantages of a distributed, modular architecture such
as ours. They highlight the importance of fault detection,
diagnosis, and self monitoring for system robustness in automated driving. Further, they unify the visual representation
of three other vehicle architectures ([30]–[32]) into a common visualization scheme. Here, [29]–[31] use a hierarchical
structuring of driving tasks as in Donges [18] for “mission
planning”, “behavior and motion planning”, and “vehicle
control and actuation”. Perception, localization, and “vehicle
state estimation” (cf. our self monitoring) are not hierarchically
structured in [29], [31], or [32]. A “scene understanding”
or “environment model” seems to be understood similarly to
our context/scene modeling as a central point for information
aggregation. [29] suggests to consider Vehicle-To-X (V2X)
communication as an “array of redundant sensors”. If such an
approach is chosen, it is of particular importance to keep in
mind that V2X information can be uncertain, incorrect, and
even intentionally misleading. Further V2X communication
might provide information on different levels of abstraction.
Hence, we treat information from V2X differently than information from onboard sensors (cf. section III-C).
Behere & Törngren [33] identify core components in a functional system architecture and group them under “perception”,
“decision and control”, and “vehicle platform manipulation”.
Their components resemble mostly our components. Yet, not
all of our components are part of their architecture. Additional components they identified are “energy management”
for “battery management” and “regenerative braking” and
“reactive control” for reflex responses to unexpected stimuli as
in automated emergency braking. For us, energy management
is considered as part of the vehicle. Reactive control is indeed
considered. It is part of the stabilization module and its lowlatency data link explained in section III-B. Like [29], Behere
& Törngren [33] consider V2X communication similar to
3
Fig. 1. Functional system architecture of an automated vehicle. Blocks represent modules for activities; arrows show information flows (TS & TL state est.
= traffic sign & traffic light (recognition and) state estimation, feature extrac. = feature extraction, dyn. env. = dynamic environment, aug. = augmentation, la.
cros. = lane crossing handling, la. cha. = lane change handling, exec. mon.= execution monitoring, HMI = human machine interface, V2X = Vehicle-To-X)
a sensor/actor. Finally, Behere & Törngren [33] identified
the otherwise often neglected aspect of “diagnosis and fault
management”. We agree to its importance. In our architecture
it is part of self perception (cf. section III-A) and execution
monitoring (cf. section III-B).
The functional system architecture presented here has been
inspired by and applied to the Stadtpilot project for automated
driving in urban environments and the aFAS project for an
unmanned protective vehicle for highway hard shoulder road
works. Hence, it is not just a top-down concept from a sketch
board but has actually been proven to work in real world
automated driving. It underwent several iterations. The foundations have been laid in [1]–[3], [5], the concept for context
modeling has been developed in [4], [6], [8], [14], environment
perception has been refined in [7]–[9], self representation has
been addressed by [15], [34], [35], and localization and map
provision has been discussed by [8], [9], [36]. The remainder
of this article will show the status quo of our functional
system architecture and present the enhancements compared
to previous publications.
III. F UNCTIONAL S YSTEM A RCHITECTURE
Figure 1 illustrates a revised architecture based on [1]–[9],
[11]–[13].
The vertical abstraction layers of the functional system architecture are aligned to the levels of driving tasks from
Donges [18], Riemersma [19], and Michon [20, p. 498]. One
elementary (operational) stabilization layer is used for course
keeping and speed control, a second (tactical) guidance layer
is for any behavior planning and a third one is for strategic
planning (navigation). Albus [37, p. 281] suggested the use of
such a hierarchical structure not only for behavior planning
and control but also for perception. Nothdurft [38] transferred
the concept of Oberlander et al. [39], to differentiate context
information in particular for digital maps2 by “topological,”
“semantic,” and “metric” properties to the field of automated
driving. In Figure 1, the terms road-level relate to the road
network topology, lane-level to the semantic relationships
among lanes and feature-level to the metric properties used
for a localization within a lane.
Certain modifications have been made by the team at the
Institute of Control Engineering at TU Braunschweig after the
2 Based on [40] and [14], we understand a map as a diagrammatic representation of an area’s scenery.
4
publication of previous architecture versions in Matthaei &
Maurer [8] and Matthaei [9, p. 37 ff.]. The following sections
describe the current state of the functional system architecture
and discuss the recent modifications. It is explained along the
inverted-“U”-shaped main signal flow through the components
in the architecture.
A. Environment and Self Perception
1) Interfaces: The environment and self perception column
has interfaces to localization and map provision, behavior
planning and control, and with the communication column.
Within the column there is an interface towards sensors.
Perception has an interface towards the automated vehicle’s
sensor systems. They have been clustered into environment
sensors covering external aspects around the vehicle (exteroceptive) and vehicle sensors to obtain information about
the vehicle itself and its internal state (proprioceptive). Environment sensors are sensors like cameras, lidar, and radar
sensors but also conventional sensors like a thermometer or
a rain sensor. Vehicle sensors provide information about the
movement or pitch of the ego vehicle, but also information
about the charging/filling level of the battery/the fuel tank,
for example. In a hardware architecture, sensor data feature
extraction and even model-based filtering may be allocated to
a sensor itself. Yet, in a functional system architecture, the
interface between the sensor block and the subsequent feature
extraction is raw sensor data.
Although the perception column is primarily based on sensor
data from within the column, it may use map information
together with a pose within that map as input on different
hierarchical levels of abstraction. On a macroscale level, there
are topological road network maps used to augment perceived
information with a-priori map information. On a mesoscale
level, lane level map information may be used to augment
context modeling even beyond the limited field of view from
on-board sensor systems. On a microscale level within a
lane, feature information may be used to provide additional
landmarks or to stabilize lane tracking. Likewise an input
might be Vehicle-To-X information obtained from other traffic
participants or infrastructure.
The algorithms in the planning and control column are the
primary data user of the perception column. On a navigation
level, a road network together with a traffic flow may be used
to calculate an optimal route. At the tactical level (guidance),
a scene3 as defined in Ulbrich et al. [14] is provided. On an
operational level (stabilization), the perception may provide
simple features and state variables as a low latency shortcut
to low level control as in Maurer [13, p. 42].
3 A scene describes a snapshot of the environment including the scenery and
dynamic elements, as well as all actors’ and observers’ self-representations,
and the relationships among those entities. Only a scene representation in a
simulated world can be all-encompassing (objective scene, ground truth). In
the real world it is incomplete, incorrect, uncertain, and from one or several
observers’ points of view (subjective scene) (cf. Ulbrich et al. [14]).
Perceived information is provided on different levels of abstraction (road-level, lane-level, feature-level) for map updates
or mapping. Sensor data (gyroscopes, wheel tick sensors, ...)
from the perception column may directly be used for localization and map provision.
Last of all, perception data may directly or indirectly be used
for broadcasting information via Vehicle-To-X communication
or visualization. The authors assume that there will always
be a goal and value specific context selection. Thus, rather
a for others as relevant classified situation subset will be
communicated or visualized. Yet, also with this intermediate
step, communication will be at least based on information
from perception.
2) Comprised Activities: Figure 2 provides details on the
environment and self perception. The dashed line symbolizes the separation between the perspective to the outside
(environment perception) and the often neglected perspective
to the inside (self perception) as in Maurer [13, p. 58 ff.],
Bergmiller [41, p. 145 ff.], and Reschka et al. [34]. Similar as
in Matthaei [9, p. 51], a green color codes that only relatively
certain internal information has been used. The blue color
indicates that only internal sensors and/or environment sensor
information has been used. The violet color indicates that
additionally map data with all possible errors in map-relative
localization and incorrect, possibly outdated map information
has been used. The yellow color indicates perceived data used
for map updates and Vehicle-To-X information.
Sensor data is used for feature extraction and subsequent
model-based filtering. Feature extraction and model-based
filtering is performed regarding several aspects. This includes
lane detection and tracking, dynamic element tracking4 , occupancy grid modeling plus subsequent feature extraction and
data filtering, traffic sign and traffic light recognition and state
estimation, as well as self monitoring of the automated vehicle.
Input to this block are raw or processed sensor data and
possibly feature-level map data. Models are used to identify
entities, associate measurements to entity hypotheses and track
entities over time. In lane tracking, dynamic element tracking,
and traffic light and traffic sign recognition a temporal validation or tracking is typically performed after an extraction of
relevant features. In occupancy grid mapping, widely used for
the stationary environment, a similar temporal filtering results
from a probabilistic filtering performed in different cells of
the occupancy grid itself. Entities and properties of these are
generated by a subsequent feature extraction from that grid.5
Any of the sensors are mounted to the automated vehicle.
Thus, their sensor data will be ego-relative. To transform
sensor data into a stationary coordinate system, it is necessary
to estimate ego motion. This is part of the data filtering in self
perception. We further suggest to integrate self monitoring into
self perception. The threshold between a self monitoring and
4 “Dynamic objects” form the set of “dynamic elements” by extending them
with non-object-model-compliant elements (cf. Ulbrich et al. [14]).
5 The feature extraction and model-based filtering is not discussed in further
detail here. Some details are provided in Rieken et al. [7] and will be discussed
in a future publication specifically on this topic.
5
Fig. 2. Environment and self perception based on Rieken et al. [7] and Matthaei & Maurer [8]. Green = only subject to vehicle sensor errors; blue = subject to
environment sensor errors; violet = also subject to map and localization related errors; yellow = perceived environment features and Vehicle-To-X information
(dyn. env. = dynamic environment, TS & TL state est. = traffic sign & traffic light (recognition and) state estimation, extrac. = extraction, pas. = passive, IR =
infrared sensors, USS = ultrasonic sensors, thermo = thermometer, PMD = photonic mixing device, rain = rain sensors, odomet. = odometry sensors, gyros. =
gyroscope, consum. = consumption sensors, fill lev. = fill level sensors, steer. ag. = steering angle sensors, rotation = rotation sensors, acceler. = acceleration
sensors, environ. = environment, V2X = Vehicle-To-X)
self representation on a context modeling level seems vague
at first. The authors suggest to use the same differentiation
as for other entities. The self monitoring provides information
about entities of the ego vehicle and their attributes like health
states or errors. The self representation provides semantic links
between those entities to derive a full context not only about
the environment but also about the automated vehicle itself.
The information from the feature extraction and model-based
filtering is used for context/scene modeling (cf. Ulbrich et al.
[14]). This subsumes several aspects of information modeling,
aggregation, and association. Scenery modeling combines lane
information with a scenery model. This scenery model may
use a-priori map data and a position in this map from the
localization and map provision column in Figure 1. Dynamic
environment modeling may interact with the scenery model
to incorporate model-based information. Dynamic elements,
for example, are more likely to move along lanes or paths.6
Dynamic elements and the scenery are associated with each
6 For safety applications and to model non-rule compliant behavior, it is
essential that this is only an information augmentation. The initial tracking
results still need to be maintained to avoid crashing into non-rule compliant
dynamic elements.
other to obtain an environment model. This is combined
with the self representation of the ego vehicle to yield a
context/scene model. This scene representation is transmitted
to modules in the planning and control column. Matthaei [9,
p. 52] differentiates a “local” scenery and scene modeling
from an “extended” one. The first is solely based on perceived
information and incorporates no map-related information. Its
output can be used for updating a map with perceived information. The distinction avoids loops in the information flow
and self-confirming hypotheses of confirming map data with
map-supported perception data.
The perception column is completed by modeling a road-level
environment. This subsumes a possible road topology and
traffic flow identification to estimate which lanes constitute
roads and whether these roads are congested or blocked.7 So
far, this module has not been implemented in the Stadtpilot
or aFAS project. The road network is simply piped through
as it is from an a-priori map from the localization and map
provision column towards subsequent modules.
7 A lane level traffic flow identification may still be considered as part of
the context modeling.
6
3) Enhancements to the State of the Art: The modifications
are shown towards Matthaei & Maurer [8] as the last broadly
accessible publication of our functional system architecture
in English. The sensors’ block is identical; feature extraction
and model-based filtering has only been marginally modified
regarding the self perception. Here, Matthaei only mentioned
the aspect of “motion estimation” and a rather vague “data
filtering” (Matthaei & Maurer, 2015, p. 162; Matthaei, 2015,
p. 51). Yet, as in Maurer [13, p. 58 ff.], Bergmiller [41, p.
145 ff.], and Reschka et al. [34] this is only part of the self
perception. It may further include friction coefficient estimation, vehicle component wear-and-tear estimation, component
diagnosis, energy level estimation, etc.
The aspect of traffic sign and traffic lights has marginally
been modified. Matthaei & Maurer [8, p. 162] called it traffic
sign and traffic light “detection”, Matthaei [9, p. 51] called it
traffic sign and traffic light “state estimation”. Of course it is
necessary to detect, recognize the position, type and in case
of traffic lights the state of an element. Other than tracking
stationary lane markings/lanes, Matthaei assumes no need for
an ego motion compensation for traffic signs and lights. Traffic
sign and traffic light estimation has not been implemented in
the Stadtpilot project. If this is purely frame-based, it may
indeed not need an ego motion estimation. If it stabilizes traffic
sign/light hypothesis over time, it will need an ego motion.
Thus it has been linked by a dotted line.
Context modeling has been restructured. Matthaei’s differentiation between “local” scenery/scene modeling and “extended”
scenery/scene modeling8 have been both subsumed under only
one scenery modeling and scene modeling with corresponding submodules. A dynamic environment modeling has been
introduced as an analogon to scenery modeling for static
environment aspects. This may include steps of validating
different tracks of dynamic elements against each other. For
instance if the contours of different elements overlap it might
be a sign of actually tracking the same object twice rather than
in fact observing a collision. Further, Matthaei & Maurer [8,
p. 162] and Matthaei [9, p. 51] called the step of associating
semantic information about the automated vehicle “vehicle
state modeling”. Aligned with Bergmiller [41, p. 145 ff.], the
authors prefer self representation as a name for this block.
Last of all, the name of the overall module seems odd at first.
While it is named context modeling, its output is only a scene
from a scene modeling as in Matthaei & Maurer [8]. With the
definitions from Ulbrich et al. [14] it is indeed correct to have
a scene as an output. Yet, the process itself entails aspects
of context modeling, too. Thus the name of the module is
extended to context/scene modeling.
Similar to Matthaei [9, p. 51], road topology identification and
modeling as well as traffic flow identification are summarized
in a block above context modeling. The block has been
renamed from “road topology and traffic flow modeling” to a
more general road-level environment modeling. Linguistically,
this makes room to identifying and modeling one day aspects
8 This entails information from a-priori map data about static and movable
elements.
like ferryboats affecting the mission planning due to limited
operating hours as a part of this block. Moreover, an arrow between road-level environment modeling and the context/scene
model has been added to represent such an information flow
of high-level road information towards information in a scene.
B. Planning and Control
1) Interfaces: The planning and control column has interfaces
to the perception and communication column and towards the
actuators within the column. Inputs from perception are:
•
•
•
A road network together with a traffic flow information
for navigation.
The scene described as in Ulbrich et al. [14] for tactical
planning (guidance).
Features with state variables as a low latency shortcut to
control as in Maurer [13, p. 42].
Outputs exist within the column towards actuators. These
entail gas, gears, brake commands, and steering. Yet, it may
also include actuation of other vehicle components like the
horn, indicators, or headlights. It may even include opening
a door lock or the trunk for freight delivery or loading, or
activating the wipers for removing dirt from the windscreen.
Interfaces towards the communication column will be detailed
in its according section.
2) Comprised Activities: Figure 3 illustrates details on the
planning and control in a functional system architecture. The
color coding for information flows is the same as in section
III-A. Modules for planning and control use the previously
mentioned scene as a central interface on a tactical level. The
modules have been divided into three levels according to the
hierarchy of driving tasks in Donges [18].
On a strategic level (navigation), the road network, information
about traffic flows or blockages and an externally provided
mission are used for navigation purposes. A mission planning
as in Dickmanns [12, p. 405] or Gregor et al. [42, p. 81 ff.] entails planning certain events like a cargo or passenger pickup.
They result in waypoints between which a route needs to be
planned. A route planning yields a —with respect to some
optimization criteria— best route but also route alternatives.9
The calculation of route alternatives may be triggered by
events or upon request of the tactical level (guidance). The
navigation may consider skill restrictions of an underlying
guidance layer. If, for instance, the battery of an electric
vehicle is too low to take a shorter but more energy consuming
route through a mountain area. Route alternatives to reach the
mission goals are recalculated to reflect ego position changes.
The guidance modules use this information to render a mission
executable. They use the current scene to select relevant
aspects and to augment it with additional information to derive
one or several situation representations for the automated
9 The aspect of route alternatives has so far not been implemented in the
Stadtpilot or aFAS project.
7
Fig. 3. Submodules for planning and control. Blue = subject to environment sensor errors; violet = also subject to map and localization related errors; yellow
= Vehicle-To-X information for perception (V2X = Vehicle-To-X, dyn. = dynamic, traj. = trajectory, alt. = alternative, HMI = human-machine-interface)
vehicle. Such a situation is used for situation assessment
and behavior planning regarding several situation aspects.
Among those are regular driving within a lane, lane changes,
lane crossings (e.g., at intersections), free space navigation
for parking, etc. (cf. “driving maneuvers [...] for automated
vehicles”, [35, p. 122 ff.]). Situation assessment for these situation aspects entails application specific situation assessment
expert algorithms and also skill and ability monitoring for that
particular situation aspects. Behavior planning entails not only
maneuver selection but also planning about how a maneuver
should be executed. This how does not include detailed velocity profile planning but rather a sequence of tactical behavior
decisions like longitudinal and/or lateral adjustments to a gap
or stopping points in an intersection, indicator activations
or maybe even honking one’s horn. The guidance block
is completed with execution monitoring of all components,
which ensures reliability (continuity of correct service) and
availability (readiness for correct service) [43]. This execution
monitoring has ultimate control over deactivating the system
or its modules.
Output of the tactical guidance layer is a set of target poses
for maneuvers. A target pose commands the stabilization layer
what to plan for. This may entail a target position, orientation,
velocity (and further derivatives), constraints for trajectory
planning like a drivable area, a reference corridor, sampling
ranges or target deviation costs, and a symbolic maneuver type
information.
The maneuver information may be utilized by an underlying
stabilization level to switch between algorithms as in Maurer
[13, p. 74]. The target pose may be linked to a vehicle with a
certain id to perform longitudinal vehicle following. It may be
set to the center of a neighboring lane for lane changing or it
is set towards a gap in traffic for longitudinal adjustments to
prepare lane changing. For parking, this pose may contain a
goal position and orientation in a parking lot. Even at complex
intersections, this interface seems sufficient to cover, e.g., to
stop at a stop sign, proceed to a line of sight and finally turning
through a lane with oncoming traffic.
Depending on the actual implementation only one or several10
target poses may be handed over to the stabilization level. In
case of the latter, target pose selection is implicitly done by
the knowledge of selection rules in the stabilization level.
The stabilization subsumes trajectory planning and low level
control, and execution monitoring as three major aspects. Trajectory planning calculates trajectory candidates for all these
target poses. Low level control translates those trajectories
10 So far, only one target pose has been implemented in the Stadtpilot or
aFAS project.
8
into actuator control variables. Execution monitoring detects
deviations between what is planned and executed.
Trajectory planning as in Werling [44], or path planning with a
subsequent velocity profile planning as in Kammel et al. [45],
Hundelshausen et al. [46], Wille [5], Broggi et al. [47] can be
generalized into a three step procedure of trajectory alternative
generation, trajectory alternative assessment and selection, and
transforming the results into a representation to be used for
low level control.
A path or trajectory planning may entail a subsampling of
further target poses around the provided target poses as in
Werling [44, p. 42]. Based on a cost function, the best
trajectory, according to a cost criteria, is selected.11 Depending
on the implementation of the trajectory planning, it is necessary to transform the trajectory from a geo-stationary, local
coordinate system of a scene or situation towards an egovehicle bound coordinate system in which the actuators and
low level controllers operate. If trajectory planning is executed
in a Frenet frame, this transform is performed as a last step.
A future point on this trajectory is used as input to the low
level controllers to command a steering angle, brake pressure,
or acceleration rate to the actuators of the automated vehicle.
To reduce latency, it may be necessary to obtain direct feature
updates from the previously mentioned model-based filtering
algorithms directly on the stabilization level. These feature
updates may be incorporated into the low level controllers or
even the trajectory planning.
Once more, the stabilization level entails execution monitoring
to ensure the correct functioning of these algorithms and
possibly to inform the guidance module about issues on
the stabilization level. Examples of this driving task relevant
information are if no collision-free trajectory can be calculated
or if the execution of commanded behavior is not possible
due to physical limitations in the vehicle’s dynamics. This
feedback is either used for execution monitoring in the tactical
level or even to adopt the tactical behavior planning (guidance)
or strategic mission planning (navigation). For instance, if
changing lanes to a highway exit lane jam-packed with traffic
requires high relative velocity adjustments and thus high
discomfort in trajectory planning, it may even affect the route
planning by avoiding such a maneuver and simply taking
an alternate route by choosing a next exit further down the
highway. Likewise, even low level control may provide such
feedback by reporting control deviations. If a high slip angle
indicates issues in vehicle stability, it may even affect tactical
behavior planning by changing to a lane with better friction.
3) Enhancements to the State of the Art: On the strategic
level of navigation, the route planning has been renamed to a
more general mission planning. When the scope of automated
driving becomes wider, mission planning may not only contain
route planning but even mission elements [48, p. 43] like
cargo pickup, or refueling. Matthaei & Maurer [8] mention
11 Selecting the best point could once more be considered as tactical decision
making. Hence, one could argue the necessity of a trajectory selection arbiter
block within the guidance module. For simplicity, it is excluded in Figure 3.
a “selection of a next navigation point” as a submodule of the
navigation block. Only transferring the next navigation point
to a guidance module imposes a severe limitation because
several route alternatives may exist. This can be illustrated
in the earlier mentioned example of an automated vehicle
performing a lane change onto an off-ramp jam-packed with
traffic. If there is a high risk to exceed the skills of the vehicle,
it may be better to avoid such a risky lane change and accept
a marginal detour rather than to enforce exiting where it was
planned.
This is not only a thought experiment but rather a real world
issue and addressed by the lane advice in Ulbrich & Maurer
[49]. For that reason, the authors deviate from Matthaei &
Maurer [8] by assuming not only one but several routes as
an output of the route/mission planning and dropping the
“selection of a next navigation point” altogether. Only if the
alternatives are known, an informed tactical decision about
following or deviating from what was planned at the navigation level is possible. Likewise to incorporate such knowledge
about limited skills from a tactical level (either from the self
representation as part of the scene) or the situation assessment
and behavior planning itself into the mission planning, an
upward facing arrow from guidance to navigation is added.
Deep changes have been made to tactical planning compared
to Matthaei & Maurer [8]. As illustrated in Ulbrich et al. [14] a
goals- and value specific context selection and augmentation is
added as an intermediate step between a goals- and value independent scene and a goals- and value related situation. There
may be one or several situation data structures for different
aspects of behavior planning. They can be used as an input
or even be augmented by modules for situation assessment.12
For instance, the results of a gap quality assessment might
be fed back into a situation. That information could be used
in an adaptive cruise control target pose selection module to
temporarily reduce a time gap towards a front vehicle to avoid
restricting gap adjustments to a gap slightly in front.
Behavior planning is used as an additional block to reflect
not only a maneuver selection but likewise the earlier introduced planning about how a maneuver should be executed.
The earlier introduced execution monitoring is added as an
additional block to the planning and control column. No clear
opinion has yet been formed if it is necessary to include
execution monitoring as a separate block or if every block
is supposed to have a sub-aspect of execution monitoring.
Yet, as mentioned earlier, it is indeed important to include
the upward information flow from stabilization to guidance.13
It was missing in [8] and has now been added.
The stabilization block has been detailed compared to Matthaei
& Maurer [8]. A feature updating block has been added to
12 Other than Matthaei & Maurer [8] the authors prefer the less ambiguous
term situation assessment instead of situation analysis. Yet, a situation is rather
the input of a situation assessment than its output. Only some situation aspects
may be needed for other modules in situation assessment and thus fed back
into the situation data structure.
13 This extension is based on discussions with Professor Chris Gerdes,
Stanford University in 2014.
9
reflect the updating process of, e.g., vehicle distances and velocities for low latency stabilization (cf. [13, p. 42]). Trajectory
target poses from the guidance level may be associated to
dynamic elements. Their dynamic state variables may be updated based on more recent information directly from modelbased filtering while bypassing the latency induced by the
more comprehensive context modeling, situation assessment,
and behavior planning. This leads to faster reactions in time
critical scenarios.
The set of actuators has been extended by adding indicators,
the horn, door locks, wipers, lights, etc. Matthaei & Maurer [8,
p. 164] highlight that some actuators are used for the purpose
of tactical communication (cf. “implicit communication” in
Ulbrich et al. [50]). These actuators (or rather: devices) have
not been part of the functional system architecture so far,
neither as part of the communication column nor of the
actuator block. Due to their similar nature as activating a
brake light, they are all grouped under the actuator module.
A module from the tactical guidance level may actuate those
devices through the operational stabilization level.
At last, the “planning and control” column has been renamed
from the linguistically ambiguous term “mission accomplishment”.
and control. Predictive warnings to inform a passenger may
either be considered as part of the situation or as a separate
information interface from the guidance module towards the
communication column.
In the opposite direction (towards perception and map provisioning), the communication column provides Vehicle-To-X
information to be incorporated into the scene and possibly likewise on a feature or road level. Likewise, a desired maneuver
may be commanded from an operator to the guidance module
[9, p. 57]. This could be to command an operator-initiated lane
change but also to command an emergency stopping maneuver
or a driver takeover request.
At the operational level (stabilization), short term warnings
may be issued or desired setpoints commanded [9, p. 57].
Short term warnings could be the activation of an electronic
stability control system in case of a higher than intended
slipping angle on a low friction road. A desired setpoint could
be the timegap towards a leading vehicle for an adaptive cruise
control driver assistance system. For a future level-5 system
these interfaces may not be necessary anymore, because by
definition the system needs to handle all these aspects without
driver intervention. Yet, as long as there is a transition between humans used to drive a vehicle by themselves and full
automation these interfaces may still exist as a legacy for a
long time.
C. Communication
1) Interfaces: The interfaces of the communication are illustrated in Figure 3. At the strategic level for navigation tasks,
a mission may directly be commanded from an operator via a
human-machine interface or even remotely via Vehicle-To-X
communication. The mission may entail a route destination as
well as goal criteria like a route with most comfort in automated driving, shortest travel distance, or the most economic
route alternative. As a feedback, the system may communicate
a planned route, resulting from the commanded mission. Yet,
the system may even provide route alternatives to an operator
to enhance mission selection. The authors agree with Matthaei
[9, p. 56] that for a SAE-level-5 system (cf. [51]) of an
automated vehicle, the only necessary input is on a strategic
level (navigation). Yet, for the sake of informing an operator
or in case of not-level-5 systems, additional communication
interfaces are necessary.
At a tactical level for guidance tasks, a situation is used as an
interface for visualization and Vehicle-To-X communication.
While the situation for Vehicle-To-X communication may be
different from the situation for behavior planning of the ego
vehicle, it is still a situation because not every aspect that is
part of the scene will be relevant for the (assumed) goals and
values of any of the information recipients in Vehicle-To-X
communication, or legal to be transmitted (cf. “enhancements”
section). Likewise, a situation for visualization will probably
be simplified and temporarily smoothed to reduce distraction.
Yet, it is still a situation because it shows what is relevant
regarding the goals and values of an operator or interested
passenger. It may entail information about planned maneuvers as part of the situation aspects derived from planning
2) Comprised Activities: An automated vehicle may have a
communication interface for communicating with an operator
or passenger (human-machine interface, HMI), as well as
for technical communication with other traffic participants or
the infrastructure via a Vehicle-To-X (V2X) communication
interface.
The human-machine interface entails both directions of communication: On the one hand, to obtain input from an operator
or passenger and on the other hand to provide information.
A special case are automated vehicles being monitored by
a central tele-operation unit. Here the aspect of a humanmachine interface and the usage of communication networks
are combined. Matthaei [9, p. 56] envisions the idea of navigation or guidance inputs for traffic management or clearing
corridors for emergency vehicles. For the latter, the reliability
and guaranteed coverage of current communication networks
is an issue. Yet, at least the technically less demanding centrally controlled deactivation of an automated driving function
within a certain amount of hours could be useful to ensure the
absence of hazardous states caused by a bug, after such a bug
has been discovered in the fleet of automated vehicles.
The aspect of Vehicle-To-X communication entails communication with other traffic participants or infrastructure. Depending on what other vehicles are able to provide the range
of applications is wide. Current research initiatives like KoHAF14 address aspects like obtaining map updates from fleets,
collaborative perception, and coordinating cooperative driving
maneuvers among traffic participants. Algorithms to implement such behavior are spread among the modules in the other
14 http://www.ko-haf.de/,
visited on Nov. 29th, 2016.
10
three columns of the functional system architecture. Yet, the
actual communication interface for 802.11p wireless local area
network communication, cellular network communication, or
other communication channels is part of this column.
3) Enhancements to the State of the Art: Certain modifications
have been made to the communication column since it was
published in Matthaei & Maurer [8].
Regarding interfaces, changes have been made to some contents of existing arrows. The interface between navigation and
communication in Matthaei [9, p. 57] is extended by not only
exchanging a “route” but rather a “mission” as input to the
navigation and by adding the aspect of route alternatives for
the opposite information flow.
While Matthaei15 assumed collaboration happens over the
interface left of the perception column, the authors suggest
to use the existing communication interface in the communication column. To the authors, there is no need for a separate
interface in the functional architecture, because aspects from
the perception column can be exchanged with one interface
at the very left. To allow an information flow from the
communication column to the perception and localization and
map provision columns, additional links have been added.
For transmitting Vehicle-To-X information, it is assumed that
a full scene will probably never be sent but only a relevant
extract of the aspects assumed to be relevant for the information recipients and their archival of their anticipated goals and
values (situation for Vehicle-To-X communication). If little
information exists about the goals and values of the information recipients, only obviously irrelevant aspects (e.g., privacy,
what was seen inside of buildings by accidentally looking
through windows) may be excluded and thus the relevant
extract may almost converge against the full information from
a scene.16 If legislation and communication channel width will
ever allow to broadcast a full scene, the aspect of information
selection could be dropped and the link between the perception
and communication blocks becomes bidirectional.
The localization and map provision column can exchange
V2X information with the communication interface. Thus, the
blocks in localization and map provision can receive and send
updates of map data on all layers of the architecture.
D. Localization and Map Provision
1) Interfaces: The localization and map provision column has
interfaces with the perception column to exchange:
•
•
•
road-level map features and map updates,
lane-level map features and map updates,
feature-level map features and map updates, and
15 Internal report “Cooperation, Collaboration, and Communication” from
March, 2015.
16 In the distinction between a scene and situation in Ulbrich et al. [14]
the focus was rather on goals and values of a vehicle. Here the distinction
has similarly been extended towards goals and values to be considered for
communication as they are stipulated by authorities, e.g., privacy.
•
vehicle and environment sensor data.
Further, it has an interface with localization sensors. According
to [8], the localization sensors like those in a global navigation
satellite system (GNSS) are not part of the environment
sensors at the bottom but are rather noted on the left due
to providing information on higher abstraction levels.
Within the column, information is exchanged between the
different hierarchical levels. The upward information flow
represents the use of, e.g., low level map features to extract
higher level lane information. Likewise, there is an information
flow downwards: Information about the existence of a road
might be used to establish semantic relationships and support
lane hypotheses in a lane level map.
2) Comprised Activities: The automated vehicle needs to
localize itself relative to its maps to make use of information
in these maps. The aspect of map provision entails providing
map information to other modules as well as the process of
mapping and map updating in order to have such information
to share. All these aspects are depicted in Figure 4.
Localization and map provision is executed on different hierarchical levels. Nothdurft [38] transferred the concept of
Oberlander et al. [39] to distinguish map information by
topological, semantic, and metric properties to the field of
automated driving. Based on Du et al. [52], Matthaei & Maurer
[8] differentiated between macroscale (road-level), mesoscale
(lane-level), and microscale (within lane) map information and
localization in those maps.
On each level, localization sensors provide data input to obtain
an absolute, global pose from localization algorithms. This
information is often combined in Bayesian filtering approaches
with inertial movement data (cf. blue data flow in Figure 4)
to provide a position even between position fixes from, e.g.,
a satellite-based localization sensor. Current approaches are
differentiated by their depth of data fusion (loosely, tightly,
ultra-tightly-coupled) and summarized by Skog & Händel [53]
cited by Matthaei [9, p. 43].
An absolute global pose is used together with perceived
environment features to obtain a map-relative pose estimation.
This map-relative pose is used to retrieve map information and
to provide it to modules in the perception column in order to
augment perceived information by map information.
Depending on the implementation, a second data flow from
the perception column towards the map provision column may
exist. This is to use features and a concurrently obtained maprelative pose to update maps with perceived information. This
concurrent map-relative localization and map-updating process
may be repeated on the earlier introduced hierarchical levels.
Information may be exchanged between the levels to keep
maps consistent.
Different technologies exist to serve the different (vertical)
levels in the functional system architecture with different needs
for accuracy. On a macroscale level (roads), global navigation
satellite system solutions found in today’s vehicle entertain-
11
Fig. 4. Localization and map provision based on Matthaei [9, p. 45]. Blue = subject to environment sensor errors; violet = also subject to map and localization
related errors; yellow = features for map updates (estim. = estimation)
ment systems are largely sufficient. For mesoscale localization
on a correct lane as well as for microscale localization within
a lane, a higher accuracy is needed. Signal distortions in
ionospheric layers can be compensated by utilizing different
carrier frequencies and correction data from ground stations
may be used to increase accuracy. Yet, accuracy as well as
reliability are insufficient to serve as a single, non-redundant
source for localization in automated driving. This becomes
particularly obvious in urban environments or complex multilevel highway interchanges.
All information from the localization and map-provision column is subject to errors in the localization as well as errors in
the maps itself. At the time of writing there is no guarantee
on information integrity and timeliness of data. Thus, incorrect
localization or map data may possibly propagate to subsequent
modules and compromise decisions and behavior. To ensure
awareness of this, every module that uses map data is colored
in violet.
3) Enhancements to the State of the Art: The localization
and map provision has been restructured. The “external data”
and “absolute global localization” columns in Matthaei &
Maurer [8] have been summarized into one “localization and
map provision” column. “External data” was renamed to map
provision to ensure that modules are activities, as in the
Unified Modeling Language (UML) standard. Hence, “external
data” is —similar to, e.g., a “scene”— a data container and
thus an arrow rather than a module. We think that the level
of abstraction for “localization” and “external data” seemed
less aggregated than for example “perception” or “planning
and control” which form other columns. Further, the titles of
the map provision blocks have been changed towards what
they actually provide: Maps. The “world modeling” used by
Matthaei & Maurer [8] leaves room for confusing it with
the term’s connotation in the community to where it reflects
activities which are here summarized under context modeling.
Matthaei & Maurer [8] did not provide details within the
map provision block. The refinements in Matthaei [9, p. 45]
within those blocks have now been incorporated to make them
accessible to a non-German-speaking audience.
IV. O PEN I SSUES
Despite long and intense discussions, there are still several
open issues in the functional system architecture. Three aspects
will be highlighted here:
First of all, the name of the context modeling seems counterintuitive due to the fact that it only outputs a scene. Indeed, a
scene is part of the context and according to its wide definition
(cf. section II), a full context may never be represented. Yet,
certain context information may be used for better scene
12
modeling. Thus, the term “context modeling” for the overall
block seems more appropriate.
III-B, we see these decisions to be under the decision-making
authority of the guidance level, but not without controversy.
Secondly, in architecture discussions with other research
groups in the Uni-DAS society17 the idea was voiced for
feedback from stabilization modules towards model-based
filtering modules. That is, to adopt models if, e.g., the ego
vehicle is not following a planned trajectory when it is drifting.
Another issue is where predictions are to be found in the
architecture. To the authors, a prediction is rather a tool to be
used in several modules. For instance, model-based filtering
will use prediction models. Likewise, a situation prediction
might be necessary in the guidance module or a movement
prediction in the stabilization module. One could ask if there
is a prediction even in the context model to provide not
only the current but even future scenes. A possible way to
illustrate predictions in the architecture could be to extend the
two-dimensional architecture by a third dimension in which
prediction is an additional layer. This comes to the price of
visual distinctiveness and presentability. Another way could be
to introduce multiple views on the architecture for particular
aspects.
Thirdly, no clear answer has yet been provided where a driver
or passenger18 monitoring camera should be located in the
architecture. One could argue that it is irrelevant if an operator
provides a maneuver input by a button or the camera and that
it thus shall be part of the human machine interface. Likewise,
it may be considered as a sensor and part of the perception
column. A driver or passenger monitoring is so far not part of
the Stadtpilot project.
Furthermore, an open point is the clear differentiation between
“occupancy grid mapping” in the perception column and
“feature map provision” in the localization and map provision
column. Occupancy grid mapping is necessary in perception
for local dynamic maps, free space extraction, or dynamic
classification. If static elements are aggregated in a global
feature map, it is part of the map provision column. Hence,
the age of features to be typically still maintained in the grid
or map is a distinguishing factor, but there is still room for a
better distinction between both.
In current discussions about the potentials and demands of
automated vehicles, a server-based shared map is a key to
the availability for automated vehicles. It is not explicitly
modeled in the architecture, since we think it is part of the
V2X connectivity. A more sophisticated integration into the
architecture of the ego vehicle seems not helpful, as it would
change the focus from the aspired architecture for a single
automated vehicle towards an overall architecture for a whole
traffic system. That would require several additional aspects
like trusted authorities for information validation or traffic
management authorities, which are out of scope of this article.
Moreover, there are still discussions on the point whether
navigation or guidance has ultimate decision power if a
planned route is followed or a route alternative is selected.
If a traffic jam is detected, it is clearly a navigation task to
adapt the route. Vice versa, if enforcing to take a highway exit
would result in a collision, it is the tactical guidance layer that
decides to not take the exit and to request a replanned route to
reflect the reality of having missed that particular exit. There
is a gray area in between where following the route is still
within the specifications of what the automated vehicle can
do, but where in the given situation it is just now, tactically a
better choice to rather pick a route alternative with a minimal
detour to avoid risk or maintain comfort goals. As in section
17 Uni-DAS workshop on functional system architectures in October 2015
in Darmstadt, Germany. www.uni-das.de
18 Might be necessary in a SAE level 5 system to help minors or elderly
passengers for instance in case of a medical emergency situation or to ensure
that they remain seated while driving.
Furthermore, the allocation of self representation to a particular block in the architecture is not as clear as it seems. For
sure, it is mainly a bottom up process to aggregate information
from vehicle sensors. Yet, execution monitoring might detect
that a vehicle’s deviation from its intended trajectory is high
and thus the maneuver capabilities of that vehicle are limited.
In other terms, there is goal and value specific information
for self modeling in the planning and control column. Hence,
certain aspects of self modeling could be spread over several
hierarchical levels and columns in the architecture and thus
limit the conceptual rigorousness that structure diagrams of
the architecture suggest. Once more, a third dimension with a
separate layer for self representation could alleviate this issue.
In this layer not only the self representation, but also all forms
of self monitoring and execution monitoring could be placed.
The result could be aggregated in the scene/context model
and used for decision making and control in the planning and
control column.
Possibly not fully covered is the aspect of cooperation and
competition between multiple agents. So far, implicitly cooperative behavior [54] and explicit Vehicle-To-Infrastructure
communication [55] has been implemented in the Stadtpilot
project. Yet, it seems likely that future research on cooperation
and competition may not be fully covered in the architecture.
We assume an additional view on the architecture might be
required to cover these aspects with all its various facets.
Last of all, the role of Vehicle-To-X communication is still
subject to discussions. While the current communication column is eligible to broadcast information from the planning
and control or perception column, an opposite communication
flow for Vehicle-To-X data input is harder to incorporate.
Currently, this induces a right-to-left information flow that
contradicts the main signal flow direction otherwise going
from left-to-right. A workaround would be once more to
open a third dimension or additional view for Vehicle-ToX communication as it has interfaces with many blocks. A
possible implementation specific addition to the architecture
could be a data flow from the decision modules to other traffic
participants or the infrastructure via Vehicle-To-X and vice
versa. E.g., the selected route, the selected maneuver as part
REFERENCES
13
of the situation for Vehicle-To-X communication, or a planned
trajectory on the stabilization level.
[6]
V. C ONCLUSIONS
This article presented a refined functional system architecture
for an automated vehicle. The concept of hierarchy and
functional separation has been introduced and applied. The
interfaces between the modules have been detailed and the
modifications to the state of the art have been presented. To the
authors, this functional system architecture is still an organic
structure that will be modified and refined to address the open
issues.
[7]
[8]
[9]
ACKNOWLEDGMENT
The authors would like to thank the Stadtpilot team at TU
Braunschweig for their valuable inputs and reviews of this
article. We thank Volkswagen Group Research and the AutoPilot team for their financial and institutional support of the
underlying research of this article. Further we thank the Federal Ministery for Economic Affairs and Energy for funding
our research as part of the aFAS project, the German Research
Foundation for supporting us in the Controlling Concurrent
Change project, and the Daimler and Benz Foundation for
funding the Value Based Decision Making project. We thank
Prof. Mykel Kochenderfer for hosting Andreas Reschka at
Stanford Intelligent Systems Lab while this article was written.
[10]
[11]
[12]
R EFERENCES
[1]
[2]
[3]
[4]
[5]
Wille, J. M., Saust, F. & Maurer, M., “Comprehensive
treated sections in a trajectory planner for realizing
autonomous driving in Braunschweig’s urban traffic,”
International IEEE Conference on Intelligent Transportation Systems (ITSC), pp. 647–652, 2010.
Saust, F., Bley, O., Kutzner, R., Wille, J. M., Friedrich,
B. & Maurer, M., “Exploitability of vehicle related
sensor data in cooperative systems,” in International
IEEE Conference on Intelligent Transportation Systems
(ITSC), 2010, pp. 1724–1729.
Reschka, A., Böhmer, J. R., Gacnik, J., Köster, F.,
Wille, J. M. & Maurer, M., “Development of Software for Open Autonomous Automotive Systems in
the Stadtpilot-Project,” in International Workshop on
Intelligent Transportation (WIT), 2011, pp. 81–86.
Nothdurft, T., Hecker, P., Frankiewicz, T., Gacnik, J.
& Köster, F., “Reliable Information Aggregation and
Exchange for Autonomous Vehicles,” in IEEE Vehicular
Technology Conference (VTC), 2011, pp. 1–5.
Wille, J. M., “Manöverübergreifende autonome
Fahrzeugführung in innerstädtischen Szenarien
am Beispiel des Stadtpilotprojekts (English Title:
Maneuver-Comprehensive
Autonomous
Vehicle
Guidance in Urban Scenarios in the Example of the
Stadtpilot Project),” Dissertation, TU Braunschweig,
2012.
[13]
[14]
[15]
[16]
[17]
[18]
Ulbrich, S., Nothdurft, T., Maurer, M. & Hecker, P.,
“Graph-based context representation, environment modeling and information aggregation for automated driving,” in IEEE Intelligent Vehicles Symposium (IV), 2014,
pp. 541–547.
Rieken, J., Matthaei, R. & Maurer, M., “Toward
Perception-Driven Urban Environment Modeling for
Automated Road Vehicles,” in International IEEE Conference on Intelligent Transportation Systems (ITSC),
2015, pp. 731–738.
Matthaei, R. & Maurer, M., “Autonomous driving a top-down-approach,” At - Automatisierungstechnik,
vol. 63, no. 3, pp. 155–167, 2015.
Matthaei, R., “Wahrnehmungsgestützte Lokalisierung in
fahrstreifengenauen Karten für Fahrerassistenzsysteme
und automatisches Fahren in urbaner Umgebung (English Translation: Perception-Supported Localization in
Lane-Level Maps for Advanced Driver Assistance Systems and Automated Driving in Urban Environments),”
Dissertation, TU Braunschweig, 2015.
Hock, C. & Dickmanns, E. D., “Intelligent Navigation
for Autonomous Robots Using Dynamic Vision,” in
Congress of the International Society for Photogrammetry and Remote Sensing (ISPRS), 1992, pp. 900–915.
Dickmanns, E., Behringer, R., Dickmanns, D., Hildebrandt, T., Maurer, M., Thomanek, F. & Schiehlen,
J., “The seeing passenger car ’VaMoRs-P’,” in IEEE
Intelligent Vehicles Symposium (IV), 1994, pp. 68–73.
Dickmanns, E. D., Dynamic Vision for Perception and
Control of Motion. London, United Kingdom: Springer
International Publishing, 2007.
Maurer, M., “Flexible Automatisierung von Straßenfahrzeugen mit Rechnersehen (English Translation:
Flexible Automation of Road Vehicles with Machine
Vision),” Dissertation, Universität der Bundeswehr in
München, 2000.
Ulbrich, S., Menzel, T., Reschka, A., Schuldt, F. &
Maurer, M., “Defining and Substantiating the Terms
Scene, Situation, and Scenario for Automated Driving,”
in International IEEE Conference on Intelligent Transportation Systems (ITSC), 2015, pp. 982–988.
Stolte, T., Reschka, A., Bagschik, G. & Maurer, M.,
“Towards Automated Driving: Unmanned Protective
Vehicle for Highway Hard Shoulder Road Works,” in
IEEE International Conference on Intelligent Transportation Systems (ITSC), 2015, pp. 672–677.
ISO 26262:2011 Road vehicles - Functional safety.
Geneva, Switzerland: International Organization for
Standardization (ISO), 2011.
Rasmussen, J., “Skills, rules, and knowledge; signals,
signs, and symbols, and other distinctions in human
performance models,” IEEE Transactions on Systems,
Man, and Cybernetics, vol. SMC-13, no. 3, pp. 257–
266, 1983.
Donges, E., “A Conceptual Framework for Active
Safety in Road Traffic,” Vehicle System Dynamics: International Journal of Vehicle Mechanics and Mobility,
vol. 32, no. 2-3, pp. 113–128, 1999.
REFERENCES
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
Riemersma, J. B. J., “Perception in traffic,” Urban
Ecology, vol. 4, no. 2, pp. 139–149, 1979.
Michon, J. A., “A Critical View of Driver Behavior
Models: What Do We Know, What Should We Do?”
In Human Behavior and Traffic Safety, Evans, L. &
Schwing, R. C., Eds., New York, USA: Springer International Publishing, 1985, pp. 485–524.
Hale, A. R., Stoop, J. & Hommels, J., “Human error
models as predictors of accident scenarios for designers
in road transport systems,” Ergonomics, vol. 33, no. 1011, pp. 1377–1387, 1990.
Muigg, A., “Implizites Workloadmanagement - Konzept
einer zeitlich-situativen Informationsfilterung im Automobil (English Translation: Implicit Workload Management - Concept of a Temporal-Situative Information Filtering in an Automobile),” Dissertation, TU München,
2009.
Matthaei, R., Reschka, A., Rieken, J., Dierkes, F.,
Ulbrich, S., Winkle, T. & Maurer, M., “Autonomous
Driving,” in Handbook of Driver Assistance Systems Basic Information, Components and Systems for Active
Safety and Comfort, Winner, H., Hakuli, S., Lotz, F. &
Singer, C., Eds., Springer, 2016, pp. 1519–1556.
Zapp, A., “Automatische Straßenfahrzeugführung durch
Rechnersehen (English Translation: Automated Road
Vehicle Guidance by Computer Vision),” Dissertation,
Universität der Bundeswehr in München, 1988.
Dickmanns, E. D. & Graefe, V., “Dynamic monocular machine vision,” Machine Vision and Applications,
vol. 1, no. 4, pp. 223–240, 1988.
Dickmanns, E. & Müller, N., “Scene recognition and
navigation capabilities for lane changes and turns in
vision-based vehicle guidance,” Control Engineering
Practice, vol. 4, no. 5, pp. 589–599, 1996.
Siedersberger, K.-H., “Komponenten zur automatischen Fahrzeugführung in sehenden (semi-)autonomen
Fahrzeugen (English Translation: Components for
Automated Vehicle Guidance in Perceiving (Semi)Autonomous Vehicles),” Dissertation, Universität der
Bundeswehr in München, 2003.
Pellkofer, M., “Verhaltensentscheidung für autonome
Fahrzeuge mit Blickrichtungssteuerung (English Translation: Behavior Decision Making for Autonomous Vehicles with Gaze Control),” Dissertation, Universität der
Bundeswehr in München, 2003.
Tas, O. S., Kuhnt, F., Zollner, J. M. & Stiller, C.,
“Functional system architectures towards fully automated driving,” in IEEE Intelligent Vehicles Symposium
(IV), 2016, pp. 304–309.
Wei, J., Snider, J. M., Gu, T., Dolan, J. M. & Litkouhi,
B., “A behavioral planning framework for autonomous
driving,” in IEEE Intelligent Vehicles Symposium (IV),
2014, pp. 458–464.
Jo, K., Kim, J., Kim, D., Jang, C. & Sunwoo, M.,
“Development of Autonomous Car - Part II: A Case
Study on the Implementation of an Autonomous Driving System Based on Distributed Architecture,” IEEE
14
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
Transactions on Industrial Electronics, vol. 62, no. 8,
pp. 5119–5132, 2015.
Kunz, F., Nuss, D., Wiest, J., Deusch, H., Reuter,
S., Gritschneder, F., Scheel, A., Stubler, M., Bach,
M., Hatzelmann, P., Wild, C. & Dietmayer, K., “Autonomous driving at Ulm University: A modular, robust,
and sensor-independent fusion approach,” in IEEE Intelligent Vehicles Symposium (IV), 2015, pp. 666–673.
Behere, S. & Törngren, M., “A functional architecture
for autonomous driving,” in Proceedings of the First
International Workshop on Automotive Software Architecture, 2015, pp. 3–10.
Reschka, A., Bagschik, G., Ulbrich, S., Nolte, M.
& Maurer, M., “Ability and skill graphs for system
modeling, online monitoring, and decision support for
vehicle guidance systems,” in IEEE Intelligent Vehicles
Symposium (IV), 2015, pp. 933–939.
Reschka, A., “Fertigkeiten- und Fähigkeitengraphen
als Grundlage des sicheren Betriebs von automatisierten Fahrzeugen im öffentlichen Straßenverkehr in
städtischer Umgebung (English Translation: Skill and
Ability Graphs as Basis for Safe Operation of Automated Vehicles in Urban Environments),” Dissertation,
forthcoming, TU Braunschweig, 2017, forthcoming.
Matthaei, R., Bagschik, G. & Maurer, M., “Maprelative localization in lane-level maps for ADAS and
autonomous driving,” in IEEE Intelligent Vehicles Symposium Proceedings, 2014, pp. 49–55.
Albus, J. S., “Mechanisms of planning and problem
solving in the brain,” Mathematical Biosciences, vol. 45,
no. 3, pp. 247–293, 1979.
Nothdurft, T., “Ein Kontextmodell für sicherheitsrelevante Anwendungen in der autonomen Fahrzeugführung
(English Title: A Context Model for Safety-Relevant
Applications in Automated Vehicle Guidance),” Dissertation, TU Braunschweig, 2014.
Oberlander, J., Uhl, K., Zollner, J. M. & Dillmann,
R., “A region-based SLAM algorithm capturing metric,
topological, and semantic properties,” in International
IEEE Conference on Robotics and Automation (ICRA),
2008, pp. 1886–1891.
Oxford Dictionaries, Map. URL: https : / / en .
oxforddictionaries . com / definition / map (visited on
02/10/2017).
Bergmiller, P., Towards Functional Safety in Drive-byWire Vehicles. Berlin, Germany: Springer International
Publishing, 2015.
Gregor, R., Lutzeler, M., Pellkofer, M., Siedersberger,
K.-H. & Dickmanns, E., “EMS-Vision: a perceptual
system for autonomous vehicles,” IEEE Transactions
on Intelligent Transportation Systems, vol. 3, no. 1,
pp. 48–59, 2002.
ISO/IEC/IEEE 24765: Systems and software engineering – Vocabulary. International Organization for Standardization (ISO), 2010.
Werling, M., “Ein neues Konzept für die Trajektoriengenerierung und -stabilisierung in zeitkritischen
Verkehrsszenarien (English Translation: A Novel Con-
15
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
cept for Trajectory Generation and Stabilization in
Time-Critical Traffic Scenarios),” Dissertation, Karlsruher Institut für Technologie, 2010.
Kammel, S., Ziegler, J., Pitzer, B., Werling, M., Gindele,
T., Jagzent, D., Schröder, J., Thuy, M., Goebl, M.,
Hundelshausen, F. von, Pink, O., Frese, C. & Stiller, C.,
“Team AnnieWAY’s autonomous system for the 2007
DARPA Urban Challenge,” Journal of Field Robotics,
vol. 25, no. 9, pp. 615–639, 2008.
Hundelshausen, F. von, Himmelsbach, M., Hecker, F.,
Mueller, A. & Wuensche, H.-J., “Driving with tentacles:
Integral structures for sensing and motion,” Journal of
Field Robotics, vol. 25, no. 9, pp. 640–673, 2008.
Broggi, A., Buzzoni, M., Debattisti, S., Grisleri, P.,
Laghi, M. C., Medici, P. & Versari, P., “Extensive Tests
of Autonomous Driving Technologies,” IEEE Transactions on Intelligent Transportation Systems, vol. 14,
no. 3, pp. 1403–1415, 2013.
Gregor, R., “Fähigkeiten zur Missionsdurchführung und
Landmarkennavigation (English Translation: Abilities
for Mission Execution and Landmark Navigation),”
Dissertation, Universität der Bundeswehr in München,
2002.
Ulbrich, S. & Maurer, M., “Situation Assessment in
Tactical Lane Change Behavior Planning for Automated
Vehicles,” in International IEEE Conference on Intelligent Transportation Systems (ITSC), 2015, pp. 975–981.
Ulbrich, S., Großjohann, S., Appelt, C., Homeier, K.,
Rieken, J. & Maurer, M., “Structuring Cooperative Behavior Planning Implementations for Automated Driving,” in International IEEE Conference on Intelligent
Transportation Systems (ITSC), 2015, pp. 2159–2165.
Society of Automotive Engineers (SAE), Taxonomy and
Definitions for Terms Related to On-Road Motor Vehicle
Automated Driving Systems J3016 Sep. 2016 (revised),
Society of Automotive Engineers (SAE), 2016.
Du, J., Masters, J. & Barth, J., “Lane-level positioning
for in-vehicle navigation and automated vehicle location (AVL) systems,” in International IEEE Conference
on Intelligent Transportation Systems (ITSC), 2004,
pp. 35–40.
Skog, I. & Händel, P., “State-of-the-Art In-Car Navigation: An Overview,” in Handbook of Intelligent Vehicles, Eskandarian, A., Ed., London, United Kingdom:
Springer International Publishing, 2012, pp. 435–462.
Ulbrich, S. & Maurer, M., “Probabilistic online POMDP
decision making for lane changes in fully automated
driving,” in International IEEE Conference on Intelligent Transportation Systems (ITSC), 2013, pp. 2063–
2067.
Saust, F., Wille, J. M. & Maurer, M., “EnergyOptimized Driving with an Autonomous Vehicle in
Urban Environments,” in IEEE Vehicular Technology
Conference (VTC Spring), 2012, pp. 1–5.
Simon Ulbrich is currently working at Audi AG on
piloted driving systems and is pursuing his PhD at
TU Braunschweig at the Institute of Control Engineering. Before that, he finished a Diploma Degree
in Electrical Engineering/Industrial Engineering at
TU Braunschweig and a Master of Science Degree
in Industrial and Systems Engineering at the Georgia
Institute of Technology. His main research interests
are tactical behavior planning and context modeling
for automated driving.
Andreas Reschka is currently pursuing a PhD at
the Institute of Control Engineering at TU Braunschweig and visiting the Stanford Intelligent Systems Laboratory. He holds a Bachelor of Science
in Network Computing from Technische Universität
Bergakademie Freiberg and a Master of Science in
Information Management and Information Technology from Universität Hildesheim. His main research
topics are self-awareness, safe behavior, functional
safety, and development processes for autonomous
vehicles.
Jens Rieken works as a research assistant at the Institute of Control Engineering at TU Braunschweig
since 2012 and is currently pursuing his PhD. He
holds a Master of Science Degree in Electrical Engineering from TU Braunschweig. His main research
topics are environment perception and scene understanding for urban scenarios, as well as designing
algorithms for point cloud processing.
Susanne Ernst is with the Institute of Control Engineering at TU Braunschweig since 2015 and is currently pursuing her PhD. She holds a Master
of Science Degree in Mechanical Engineering from TU Braunschweig. Her
main research topics are behavior recognition of other traffic participants and
decision making for automated vehicles.
Gerrit Bagschik works as a research assistant at
the Institute of Control Engineering at TU Braunschweig since 2013 and is currently pursuing his
PhD. He holds a Master of Science Degree in
Computer and Communications Engineering from
TU Braunschweig. His main research topics are
functional safety and scenario based hazard analysis
for automated vehicles.
16
Frank Dierkes works as a research assistant at
the Institute of Control Engineering at TU Braunschweig since 2013 and is currently pursuing his
PhD. He holds a diploma degree in Computer
Engineering from RWTH Aachen University. His
main research topic is behavior generation under
uncertainty for automated vehicles.
Marcus Nolte works as a research assistant at the Institute of Control Engineering at TU Braunschweig
since 2014 and is currently pursuing his PhD. He
received his Master of Science in Electrical Engineering from TU Braunschweig. His main research
topics are self-awareness and motion planning for
over-actuated automated vehicles.
Markus Maurer has held the chair for Vehicle Electronics at TU Braunschweig since 2008. His main
research interests include autonomous road vehicles,
driver assistance systems and automotive systems
engineering. From 2000 to 2007 he was active in
the development of driver assistance systems at Audi
AG.
| 3cs.SY
|
Published as a conference paper at 2018 IEEE Winter Conf. on Applications of Computer Vision (WACV’2018)
arXiv:1801.09718v1 [cs.CV] 29 Jan 2018
Object-based reasoning in VQA
Mikyas T. Desta
IBM Research, Almaden
Larry Chen∗
University of Chicago
Tomasz Kornuta
IBM Research, Almaden
[email protected]
[email protected]
[email protected]
Abstract
Visual Question Answering (VQA) is a novel problem domain where multi-modal inputs must be processed in order
to solve the task given in the form of a natural language. As
the solutions inherently require to combine visual and natural language processing with abstract reasoning, the problem is considered as AI-complete. Recent advances indicate
that using high-level, abstract facts extracted from the inputs might facilitate reasoning. Following that direction we
decided to develop a solution combining state-of-the-art object detection and reasoning modules. The results, achieved
on the well-balanced CLEVR dataset, confirm the promises
and show significant, few percent improvements of accuracy
on the complex ”counting” task.
1. Introduction
Due to successful application of machine learning, and
deep learning [24] in particular, we have lately observed
great progress in traditional Computer Vision (CV) tasks
such as image classification [23], object detection and image segmentation [10]. These advances have pushed researchers towards more complex tasks such as image caption generation [21] and VQA [4, 28]. Both of those tasks
combine Computer Vision with Natural Language Processing (NLP) and high-level reasoning, thus require utilization of multi-modal knowledge beyond a single sub-domain
(such as CV). In the case of caption generation, however,
there may be several valid captions describing a given image. Thus, it is hard to define a quantitative evaluation metric to track the progress. In contrast, in VQA the image
and question pairs might come with the ground-truth answers, enabling to measure the system accuracy. Moreover,
as there might be many questions referring to the same image, the system is in fact forced to learn various reasoning
types depending on the type of question. For those reasons
VQA is considered to be AI complete [4] and renews the
∗ Work
done during an internship at IBM Research, Almaden.
hope of building machines that could pass the Turing test in
open domains [28].
During the last three years several complex VQA
datasets have been introduced, however, many of the solutions show only marginal improvements over the strong
baselines. As there are many influencing factors it is often
hard to tell which modules of the system are working properly and why. The CLEVR dataset [18] was created with
aim of pushing forward the progress on VQA in a more systematic manner, with the main focus put on the validation
of reasoning skills. The dataset delivers a well-balanced set
of images and questions, where the information associated
with each image is complete and exclusive. Along with the
dataset the authors provided several, sometimes quite sophisticated baselines, and concluded that those models have
not learned the semantics of spatial reasoning at all.
In this paper we decided to focus on that aspect of
the problem, i.e. spatial reasoning.The hypothesis we
would like to validate is whether the operation on highlevel and abstract facts extracted from the image might improve the accuracy of the system (as suggested e.g. in [2]).
For that reason, we developed a solution operating on
object-relation-object triplets [7], similar to relational network [35], however, instead of using features extracted from
the images, we rely on the information associated with the
objects detected in the image. The achieved results are comparable with current state-of-the-art solutions in terms of
overall accuracy and show significant improvement on the
counting task.
2. Related work
Research on VQA has resulted in many interesting solutions. Those, however, could not have been developed without the existence of proper datasets and metrics for evaluation and comparison of the results.
From the historical perspective, the first dataset designed
as a benchmark for the VQA task was DAQUAR (DAtaset
for QUestion Answering on Real-world images) [27]. The
dataset contains 1.5k RGB-D images of indoor scenes from
NYU-Depth v2 dataset [37], with annotated semantic segmentations and question/answer pairs of two types: synthetic questions/answers generated automatically on the basis of NYU annotations and human question/answers collected from 5 annotators. The limitations of DAQUAR (i.e.
restriction of answers to a predefined set and strong bias
in human annotations) resulted in releases of several other
datasets, most prominently: COCO-QA [33] (based on images from MS COCO dataset [25], where substantial effort
was made in order to increase the scale of training data),
VQA [4, 11] (consisting of two sets, VQA-real with natural images, and VQA-abstract with cartoon images, and providing 17 additional (incorrect) candidate answers for each
question) and Visual Genome [22] (currently the largest
VQA dataset, with 1.7 million question/answer pairs and
with images contents additionally described by structured
annotations in the form of scene graphs).
The paper which introduced DAQUAR [27] also laid
the foundations for evaluation of the system accuracy, and
in consequence, monitoring the overall progress in the
field. The authors proposed two basic evaluation metrics.
First, by simply measuring the accuracy with respect to the
ground truth answer using string matching. Second, by using the Wu-Palmer Similarity (WUPS), enabling to evaluate
the similarity between common objects by their distance in
a taxonomy tree. Other evaluation methods include the previously mentioned one in a multiple-choice settings [4] or
the ”fill in the blanks” approach proposed along with the
Visual Madlibs dataset [48]. As the field started to mature,
researchers discovered the importance of biasses in the images and questions, which triggered the release of more balanced datasets (such as VQA v2 [11] and CLEVR [18]). It
also led to the analysis of the importance of the commonsense knowledge required to solve a given task, which in
turn resulted in efforts towards delivering ground truth containing rich description of the scene, facilitating both training and evaluation (e.g. Visual Genome [22]).
However, the most interesting achievements in the VQA
field are the novel algorithms and neural architectures.
In [43] the authors proposed to distinguish four categories:
joint embedding approaches, attention mechanisms, compositional models and models using external knowledge
bases, which also summarizes the four main research directions in VQA.
The efforts in joint embedding focus on the methods
for combining multi-modal representations. As in VQA
there are two distinct input modalities (image and text),
which makes this problem similar to the problems found
in other multi-modal domains. For example the projection
of user and item embeddings into the common representation space in neural recommender systems [15]. Exemplary
approaches developed for the VQA problem domain include e.g. Multimodal Compact Bilinear pooling (MCB) [9]
method that performed joint embedding of visual and text
features, or Relational Networks (RN) [35] where embedded question was concatenated with features extracted from
pairs of image regions, enabling the system to reason about
the relation between objects being present in those regions.
As some questions might require more than one reasoning step, a lot of researchers focused on attention mechanisms, which were initially introduced for neural translation [5], and later on adapted to the Question-Answering
problem [42]. In [36] the authors introduced a simple attention model, where an embedded question was used for the
generation of an attention mask (called region-question relevance), which was subsequently used for the calculation of
a weighted vector of concatenated question and image features. Yang et al. [47] introduced Stacked Attention Networks, which are able to infer the answer iteratively. This
architecture includes two attention layers, enabling two reasoning steps (in [42] called ”hops”), each driven by a different attention mask over the image features. The influence
of number of hops was analysed in [46], where the authors
introduced a spatial attention-based model called Spatial
Memory Network. Here, the model uses each word embedding to capture a fine-grained alignment between the image
and the question. The results indicated the superiority of
the two-hop model over the model with a single hop. The
described methods above focused only on visual attention
driven by the question. In contrast to those ideas, a major
innovation was proposed in [26], where Hierarchical CoAttention (HieCoAtt) model processed image and question
symmetrically, with the image guiding the attention over the
question and vice versa. In a quest to assess the quality of
the attention models, in more recent works researchers have
tried to validate whether the attention models learned by
neural models focus on the same image regions as humans
do [8].
The third direction results from the observation that different tasks might require totally different reasoning processes. For example in the ”query attribute” task the system
should focus its attention on a single aspect of the scene,
whereas in ”counting” it should perform a sequence of actions, focusing on the objects of interest one by one. This
observation led to the development of Neural Module Networks [3], where semantic parsers (e.g. Stanford Parser)
were used for question-driven dynamic assembly of the internal structure of the system, which is different for every
type of question. Most recent advances include programinduction [19] consisting of a program generator and program executor, where the former is a sequence-to-sequence
model (a pair of LSTMs, i.e. Long Short Term Memory recurrent neural nets [16]) responsible for constructing of an
internal representation of the reasoning process to be performed, whereas the latter is responsible for execution of
the resulting program in order to produce the answer.
Object
Encoding
Image encoder
...
MLP
...
Question
large
Sentence
encoding
(LSTM)
What is the size of
the object left of the
brown cylinder?
Word
embedding
(GLOVE)
MLP
MLP
(Softmax)
MLP
Aggregation
Object
detection
(Faster R-CNN)
Image
Question encoder
Semantic embedding
Answer
decoder
Answer
Figure 1: General architecture of the proposed system
The last direction includes research on models using external knowledge bases (i.e. external sources of data) for
answering the question. One such solution is proposed in
[44], where the authors developed a framework that extracts
image-related information from the DBpedia knowledge
base, enabling it to answer a broad range of questions, often requiring knowledge that could not be inferred directly
from the image. The improvements achieved by results
e.g. in the paper proposing the FVQA dataset [40] strongly
indicate that utilization of supporting-facts extracted from
large-scale structured knowledge bases might be the key ingredient for solving the Visual Turing test. As there are
other findings that also suggest the need for operation on
high-level, abstract facts extracted from the image (e.g. [2]),
we decided to investigate that direction further.
3. Our solution
Architecture of our system is presented in Fig. 1. It extends the Encoder-Decoder [6] architecture, which originally consisted of two RNNs, the first one used for encoding
a sequence of input symbols into a fixed length representation, and the other for decoding that representation into
another sequence of output symbols.
In VQA there are two input channels with different
modalities, thus we need two types of encoders and an intermediate module responsible for projection of both the visual and language representations into a common semantic
embedding space. Following that scheme, Xiao et al. [45]
adopted the convolutional layers of the VGG network followed by an average pooling across the output features as
the ”visual” encoder and used a two-layer stacked LSTM
network as the ”language” encoder. The authors assumed
that the output of the language encoder is already in the
semantic space and they used a two-layer-perceptron to
project the output of the visual encoder into the same space.
Similarly, the Refined Ask Your Neurons architecture [29] consisted of a visual encoder, question encoder,
a multimodal embedding module combining both encodings into a joint space, and an answer decoder. The authors
compared results achieved for several combinations of different question (e.g. BoW, CNN, GRU and LSTM) and
visual (several classical architectures including AlexNet,
GoogLeNet, VGG-19, ResNet-152) encoders.
3.1. Image encoder
In our system the image encoder realizes two major operations: object detection (OD) and object encoding. The
diagram for object detection is shown in Fig. 2. In a given
t-th image I t we detect a set of N objects:
t
dO
= [d ot0 , d ot1 , . . . , d dtN −1 ],
(1)
each defined as d otn =<d itn , d btn >, where d itn denotes the
class identifier (1 to 96) of the object and d btn contain four
parameters defining the object bounding box. We decided to
use Faster R-CNN [34] with ResNet-101 [14] pre-trained on
COCO dataset, since it is one of the most reliable methods
for object detection currently available.
In the next step (presented in Fig. 3) we decode each
object d otn and retrieve set of its attributes:
t
a on
=< ctn , mtn , stn , fnt , xtn , ynt >,
(2)
It
Object
detection
dO
feature vectors by concatenating pairs of vectors representing two objects with the encoded question, i.e. the last outt
put of LSTM (s qW
−1 ). This approach is somehow similar
to object-object-relation triplets used in [7].
t
Figure 2: Dataflow diagram of object detection
where ctn is the object color, mtn represents material it is
made from, stn indicates size, fnt represents its shape (form)
and xtn and ynt describe its position in the image. Simple encodings of each of those attributes (we used one-hot in case
of the former four and bucketing in the case of positional arguments, as explained in sec. 4.4) and concatenation of the
resulting vectors form the encoded object description e otn .
dO
t
Attribute
extraction
aO
t
Attribute
embedding
eO
t
t
e o0
t
e o0
t
s qW −1
t
e o0
t
e o1
t
s qW −1
...
t
t
t
e oN −1 e oN −1 s qW −1
...
The diagram for question processing is shown in Fig. 4.
We start with the question consisting of several words:
t
Qt = [q0t , q1t , . . . , qW
−1 ],
(3)
where W denotes the number of words constituting a
given question. Next, we use the GloVe word embedding model [32] to encode question words: e Qt =
t
[e q0t , e q1t , . . . , e qW
−1 ]. Finally, we pass the encoded words
one by one as inputs to the LSTM [16] to produce a list of
encoded output:
t
= [s q0t , s q1t , . . . , s qW
−1 ].
Word
embedding
Relation
aggregation
ar
t
t
rN
−1,N −2
Then, individual triplets are passed through a four layer
MLP (Multi-Layer Perceptron), each with 512 units and
ReLU non-linearities, which results in a vector of relations:
(5)
3.4. Answer decoder
3.2. Question encoder
Qt
Object relation
MLP
...
Those relations are further aggregated into a rt . We investigated several aggregation methods, but simple summation appeared to give the best results – please refer to
sec. 4.3.
Figure 3: Dataflow diagram of object encoding
What
is
the
object
...
...
t
r0,2
t
t
t
Rt = [r0,1
, r0,1
, . . . , rN
−1,N −1 ].
< ct5 , mt5 , st5 , f5t , xt5 , y5t >
t
Object relation
MLP
t
r0,1
Figure 5: Dataflow diagram of semantic embedding
< ct0 , mt0 , st0 , f0t , xt0 , y0t >
< ct1 , mt1 , st1 , f1t , xt1 , y1t >
sQ
Object relation
MLP
wQ
t
Sequence
embedding
(4)
sQ
t
Figure 4: Dataflow diagram of question encoding
3.3. Semantic embedding
There are many ways for combining image and question
features [38], such as concatenation, element-wise product
or bilinear operation. In our solution, similar to the relational network module reported in [35], we form a set of
Finally, we cast the answering problem as a classification task and pass the aggregated relation a rt through three
MLP layers consisting of 512, 1024, and 29 units. A 2%
dropout layer was added before the last MLP layer. ReLU
non-linearities were added in each layer except the last one,
where we applied the softmax function. The result from the
softmax was finally decoded using the word dictionary used
for initial word embeddings in Question encoder (Fig. 6).
ar
t
MLP
(Softmax)
at
Word
decoding
yt
Figure 6: Dataflow diagram of answer decoding
4. Analysis of the results
4.1. Question encoding
We have pretrained the GloVe word embedding model
on the Google News corpus [30] (3 billion running words)
word vector model and then generated embeddings for the
dictionary formed on the basis of 93 words present in
CLEVR questions and answers.
4.2. Image encoding
Utilization of object detector required its prior training
on images with ground truth bounding boxes and object
Figure 7: Exemplary CLEVR scene with objects detected
with our trained object detector
classes. Since scene descriptions in the CLEVR dataset do
not provide bounding boxes, we wrote a program for their
generation. We calculated the bounding boxes for every
object present in the scene relying on the associated metadata from the scene description, i.e. using object position
in the image (2d pixel coordinates and rotation), its position in the Cartesian space (3d coordinates) along with its
size and shape. In addition, we have assigned each object
a class identifier taking into account the possible combinations of object attributes as defined in (2), which resulted in
96 unique classes.
An exemplary scene is presented in Fig. 7. We compared
the reconstructed scene from the OD and the ground truth
scene from the scene description to check how accurately
the OD network is predicting the scenes. We were able to
get a precision of 0.99 and a recall of 0.99, thus we find the
results satisfactory.
Figure 9: The comparison of accuracy of our solution with
the CLEVR baselines in Query Attribute task
Figure 10: The comparison of accuracy of our solution with
the CLEVR baselines in Compare Integer task
4.3. Comparison with CLEVR baselines
We have reproduced the majority of baselines from
the original CLEVR paper [18], achieving similar results.
In our implementations we used the TensorFlow framework [1] and the associated object detection API [17]. As
Figure 11: The comparison of accuracy of our solution with
the CLEVR baselines in Compare Attribute task
Figure 8: The comparison of accuracy of our solution in
Count and Exist tasks with the CLEVR baselines
CLEVR does not provide ground truth for the test set, we
are reporting the results achieved on the validation set. During training we have used a learning rate of 1.0e−4 . In the
last step of Semantic embedding we have tried several aggregation methods, including simple concatenation of activations of Object relation MLPs, using mean values of those
(a) Question: what number of objects are
small cyan objects that are in front of the gray
object or big things that are in front of the big
rubber block ?
Predicted Answer: 3
Ground Truth: 3
(b) Question: how many balls are either big (c) Question: what number of things are eiblue matte objects or small red things ?
ther small shiny objects that are behind the
purple metallic cylinder or tiny purple metallic balls ?
Predicted Answer: 2
Predicted Answer: 1
Ground Truth: 2
Ground Truth: 1
Figure 12: Exemplary result on counting with heavily occluded object
activations and their sum. The experiments have shown that
summation of those activations gives the best results (accuracy of 94.6% in comparison to e.g. 92.9% in the case
of using mean value). The simple explanation is that summation ensures the invariance with respect to the order of
objects [35].
We have analyzed the results using the same taskoriented criteria as the original paper did. The accuracy
on ”exist” and ”count” tasks is presented in Fig. 8, where
the existence questions ask whether a certain type of object is present, while the count questions ask for the number
of objects fulfilling some conditions. The Query Attribute
task contains questions asking about an attribute of a particular object (Fig. 9). The Compare Integer questions ask
which of two object sets fulfilling given conditions is larger
(Fig. 10). Results for Attribute Comparison task, where
questions ask whether two objects have the same value for
a given attribute, are presented in Fig. 11.
Finally, we present the comparison of the overall accuracy of our solution against all the CLEVR baselines in
Fig. 13. Our solution has shown major improvement over
all the baselines, including the human accuracy collected
using Mechanical Turk.
Figure 13: The comparison of overall accuracy of our solution with CLEVR baselines
Figure 14: Comparison of our approach with results reported in the paper on Relational Networks
4.4. Comparison with results from RN
In Fig. 14 we present the comparison of our results
(shown in orange) with the results reported in [35] (shown
in blue). As one might notice, our overall accuracy is
slightly, by around 1%, lower. However, interestingly our
solution seems to give more consistent results for different
tasks and achieves better results in 2 out of 5 tasks. In particular, we achieved much better accuracy (by almost 5%)
in the counting task, which is currently considered to be
one of the hardest tasks in VQA (e.g. [39]). In order to realize the counting task, the system is supposed to perform
several quite different operations, i.e. understanding what
type of objects to focus on, finding those objects in the image and finally counting the instances. Our results indicate
Scene encoder
MLP
...
...
Question
large
Sentence
encoding
(LSTM)
What is the size of
the object left of the
brown cylinder?
Word
embedding
(GLOVE)
MLP
MLP
(Softmax)
MLP
Aggregation
{"color": "blue", "size": "large", "rotation":
269.8517172617167, "shape": "cube",
"3d_coords": [-1.3705521821975708,
2.0794010162353516, 0.699999988079071],
"material": "rubber", "pixel_coords": [269, 88,
12.661545753479004]}, ...
Scene
description
encoding
Scene description
Question encoder
Semantic embedding
Answer
decoder
Answer
Figure 15: The architecture of the system with scene description as input used for comparison with the CLEVR with state
description baseline from [35]
that a prior object detection naturally facilitates counting,
which we treat as support for our hypothesis that operation
on more abstract facts facilitates reasoning in general. In
Fig. 12 we present a few hard cases, where our system was
able to provide the correct answer. In particular, it managed
to properly answer questions about the scenes presented in
Fig. 12b and Fig. 12c despite the heavy occlusions of the
objects of interest, which in [35] was suggested to be the
main reason for failures in CLEVR.
4.4.1
In this approach we simply converted the object xtn and ynt
pixel coordinates into two separate vectors using one-hot
encoding and concatenated them into one vector of length
800 (as we have 480 possible values for xtn and 320 for
ynt ). As can be seen from tab. 1 this method was not able to
generalize as well as the others.
4.4.2
We further investigated why our solution does not improve in the other three tasks. After validating the accuracy
of object detection module (as reported in Sec. 4.2) we decided to reproduce the experiments from [35] in which the
Image encoder was replaced by the information retrieved
straight from the scene description. The architecture of
the resulting system is presented in Fig. 15. Despite using
the same learning hyperparameters for relational MLPs we
achieved the overall accuracy of 94.5%, which is almost 2%
worse from the accuracy of 96.4% reported for that setting
in the paper [35]. However, as the authors did not report
how they encoded the scene, we decided to further investigate that research direction. As the object attributes form
small dictionaries, we assumed that simple one-hot encoding is sufficient. However, the object position expressed
in pixel coordinates in the images (xtn and ynt ) might be
encoded in several different ways, which in turn might influence the final accuracy. Therefore, we have performed
several experiments with different methods of encoding the
object positions, briefly explained below.
One-hot encoding
Bucketing
Here we grouped n consecutive pixels to buckets along the x
and y axes separately, which resulted in a number of buckets
for x and y. We did experiments on three different bucket
sizes, i.e. 15, 20, and 30, which resulted in 32, 24, and 16
buckets for the x component and 22, 16, and 11 buckets for
the y component respectively. When given the object position coordinate xtn belongs to a range of a given bucket we
convert the bucket number to vector using one hot encoding.
We perform the same for ynt and concatenate both results. It
can be observed in tab. 1 that the solution generalizes better
when the bucket size is chosen to be 20.
4.4.3
Object enumeration
In this method, instead of relying on the pixel coordinates
we move to a more abstract representation and encoded the
position of the object in relation to the positions of other
objects in the scene. For x and y axes we form two separate lists, representing the order of the objects according
that axis. For instance, for three objects in a scene, the very
Encoding
One hot
Bucketing
Bucketing
Bucketing
Enumeration
Results from [35]
Bucket
size
–
15
20
30
–
–
Overall
Count
Exist
89.8
93.9
94.5
93.6
93.3
96.4
90.5
93.9
93.6
93.4
93.2
–
89.7
93.7
94.7
93.1
93.7
–
Compare
numbers
89.5
93.6
93.3
93.2
92.9
–
Query
attribute
90.6
94.6
95.2
94.9
93.5
–
Compare
attribute
88.5
93.5
94.4
93.4
93.1
–
Table 1: Results on different position encodings for solution with Image encoder replaced by the information retrieved
straight from the scene description
top object out of the three will have a value of one and the
bottom object will have a value of three. Finally, we encoded such represented positions using one hot encoding
and concatenated both encoded coordinates for every object.
Results of all the above mentioned experiments are presented in Tab. 1. The main finding is that our model was
able to generalize well when we grouped twenty pixel values into one bucket. Unfortunately, these experiments did
not improve the accuracy to the level reported in [35].
5. Conclusions
In this paper we have focused on reasoning using highlevel abstract facts. For that purpose, instead of relying on
features extracted from the images, we used facts in the
form of encoded objects detected by the Faster R-CNN detector. Those abstract facts were next passed to a reasoning
module, developed with the aim of learning object-object
relations. The achieved overall accuracy is comparable with
the current state-of-the-art results. Analysis of the results
unveiled that the proposed solution gives more stable results for different tasks, and, moreover, shows improvement
in 2 out of 5 CLEVR tasks. In particular, it gives significant, few percent improvement in the Counting task, which
is currently considered as one of the most complex tasks in
VQA and is gaining more and more attention, e.g. in [39]
the authors proposed a new dataset called HowMany-QA,
devoted only to that specific problem.
The detailed analysis of the operation of components of
our system has proven that OD is working properly, whereas
difference in accuracy comes from the Relational Network
module. According to the results reported in [35] the RN
module working in separation is supposed to give around
2% better overall accuracy than the accuracy that we have
managed to achieve. We treat that as a general promise that
RN operating on more abstract features should be able to
achieve even better results.
As the used reasoning module is the bottle-neck of the
system, in our future works we want to focus on this part
of the system and experiment with other approaches. The
most important research directions, aiming at improvement
of our system and overcoming its limitations, are as follows.
The first limitation is the used aggregation of relations detected by MLPs – as the number of objects in scenes might
vary, utilization of recurrent neural net seems to be natural.
An interesting solution to that problem was recently proposed in [31], where the authors developed a novel neuralbased message passing mechanism called Recurrent Relational Networks and have shown its superiority on Sudoku
puzzles and the BaBi textual QA. Second limitation of our
system is that the reasoning performed in the Semantic Embedding module is strictly sequential, with hardcoded number of reasoning steps, whereas number of reasoning steps
should in fact vary, depending on the task given. One of the
possible solutions addressing that problem is to employ the
Adaptive Computation Time (ACT) mechanism proposed in
[12]. Yet another promising direction is utilization of external memory (i.e. memory augmented neural networks such
as Neural Turing Machine [13]) for learning more complex
reasoning schemes and memorization of the abstract, graph
like representation of the observed scene before generating
the final answer. The latter is a natural extension for reasoning over explicit high-level representations of the contents
of the image and shows very good results as reported e.g.
in [41]. Finally, we also want to improve our model by
training it jointly on several tasks (e.g. word-level modeling, object detection and VQA) – it was recently shown e.g.
in [20] that such a multi-task learning enables the model to
develop unified representation and results in improved overall accuracy.
Acknowledgements
We would like to thank Alexis Asseman for setting up
and managing our hardware, enabling us to run our experiments smoothly, Ahmet S. Ozcan for his insights and
proofread, T.S. Jayram, Vincent Albouy, Ben Lyo and other
members of our Machine Intelligence team in IBM Research, Almaden, for critical feedback and discussions.
References
[1] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen,
C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, et al.
Tensorflow: Large-scale machine learning on heterogeneous
distributed systems. arXiv preprint arXiv:1603.04467, 2016.
[2] P. Anderson, X. He, C. Buehler, D. Teney, M. Johnson,
S. Gould, and L. Zhang. Bottom-up and top-down attention for image captioning and vqa. arXiv preprint
arXiv:1707.07998, 2017.
[3] J. Andreas, M. Rohrbach, T. Darrell, and D. Klein. Deep
compositional question answering with neural module networks. arxiv preprint. arXiv preprint arXiv:1511.02799, 2,
2015.
[4] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra,
C. Lawrence Zitnick, and D. Parikh. VQA: Visual Question Answering. In Proceedings of the IEEE International
Conference on Computer Vision, pages 2425–2433, 2015.
[5] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine
translation by jointly learning to align and translate. arXiv
preprint arXiv:1409.0473, 2014.
[6] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau,
F. Bougares, H. Schwenk, and Y. Bengio. Learning phrase
representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.
[7] B. Dai, Y. Zhang, and D. Lin. Detecting visual relationships
with deep relational networks. In Proceedings of the International Conference on Computer Vision and Pattern Recogition, 2017.
[8] A. Das, H. Agrawal, C. L. Zitnick, D. Parikh, and D. Batra.
Human attention in visual question answering: Do humans
and deep networks look at the same regions? arXiv preprint
arXiv:1606.03556, 2016.
[9] A. Fukui, D. H. Park, D. Yang, A. Rohrbach, T. Darrell,
and M. Rohrbach. Multimodal compact bilinear pooling
for visual question answering and visual grounding. arXiv
preprint arXiv:1606.01847, 2016.
[10] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Regionbased convolutional networks for accurate object detection
and segmentation. IEEE transactions on pattern analysis
and machine intelligence, 38(1):142–158, 2016.
[11] Y. Goyal, T. Khot, D. Summers-Stay, D. Batra, and
D. Parikh. Making the v in vqa matter: Elevating the role
of image understanding in visual question answering. arXiv
preprint arXiv:1612.00837, 2016.
[12] A. Graves. Adaptive computation time for recurrent neural
networks. arXiv preprint arXiv:1603.08983, 2016.
[13] A. Graves, G. Wayne, and I. Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401, 2014.
[14] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
770–778, 2016.
[15] X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua.
Neural collaborative filtering. In Proceedings of the 26th International Conference on World Wide Web, pages 173–182.
International World Wide Web Conferences Steering Committee, 2017.
[16] S. Hochreiter and J. Schmidhuber. Long short-term memory.
Neural computation, 9(8):1735–1780, 1997.
[17] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara,
A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, et al.
Speed/accuracy trade-offs for modern convolutional object
detectors. arXiv preprint arXiv:1611.10012, 2016.
[18] J. Johnson, B. Hariharan, L. van der Maaten, L. Fei-Fei,
C. L. Zitnick, and R. Girshick. Clevr: A diagnostic dataset
for compositional language and elementary visual reasoning.
In 2017 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pages 1988–1997. IEEE, 2017.
[19] J. Johnson, B. Hariharan, L. van der Maaten, J. Hoffman,
L. Fei-Fei, C. L. Zitnick, and R. Girshick. Inferring and
executing programs for visual reasoning. arXiv preprint
arXiv:1705.03633, 2017.
[20] L. Kaiser, A. N. Gomez, N. Shazeer, A. Vaswani, N. Parmar,
L. Jones, and J. Uszkoreit. One model to learn them all.
arXiv preprint arXiv:1706.05137, 2017.
[21] A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 3128–3137, 2015.
[22] R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz,
S. Chen, Y. Kalantidis, L.-J. Li, D. A. Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of
Computer Vision, 123(1):32–73, 2017.
[23] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in neural information processing systems, pages
1097–1105, 2012.
[24] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature,
521(7553):436–444, 2015.
[25] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In European conference on computer
vision, pages 740–755. Springer, 2014.
[26] J. Lu, J. Yang, D. Batra, and D. Parikh. Hierarchical
question-image co-attention for visual question answering.
In Advances In Neural Information Processing Systems,
pages 289–297, 2016.
[27] M. Malinowski and M. Fritz. A multi-world approach to
question answering about real-world scenes based on uncertain input. In Advances in Neural Information Processing
Systems, pages 1682–1690, 2014.
[28] M. Malinowski and M. Fritz. Towards a visual turing challenge. arXiv preprint arXiv:1410.8027, NIPS Workshop on
Learning Semantics, 2014.
[29] M. Malinowski, M. Rohrbach, and M. Fritz. Ask your
neurons: A deep learning approach to visual question answering. International Journal of Computer Vision, 125(13):110–135, 2017.
[30] T. Mikolov, K. Chen, G. Corrado, and J. Dean. Efficient
estimation of word representations in vector space. arXiv
preprint arXiv:1301.3781, 2013.
[31] R. B. Palm, U. Paquet, and O. Winther. Recurrent relational
networks for complex relational reasoning. arXiv preprint
arXiv:1711.08028, 2017.
[32] J. Pennington, R. Socher, and C. Manning. GloVe: Global
vectors for word representation. In Proceedings of the 2014
conference on empirical methods in natural language processing (EMNLP), pages 1532–1543, 2014.
[33] M. Ren, R. Kiros, and R. Zemel. Image question answering:
A visual semantic embedding model and a new dataset. Proc.
Advances in Neural Inf. Process. Syst, 1(2):5, 2015.
[34] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards
real-time object detection with region proposal networks. In
Advances in neural information processing systems, pages
91–99, 2015.
[35] A. Santoro, D. Raposo, D. G. Barrett, M. Malinowski,
R. Pascanu, P. Battaglia, and T. Lillicrap. A simple neural network module for relational reasoning. arXiv preprint
arXiv:1706.01427, 2017.
[36] K. J. Shih, S. Singh, and D. Hoiem. Where to look: Focus
regions for visual question answering. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 4613–4621, 2016.
[37] N. Silberman, D. Hoiem, P. Kohli, and R. Fergus. Indoor
segmentation and support inference from rgbd images. Computer Vision–ECCV 2012, pages 746–760, 2012.
[38] D. Teney, Q. Wu, and A. van den Hengel. Visual question
answering: A tutorial. IEEE Signal Processing Magazine,
34(6):63–75, 2017.
[39] A. Trott, C. Xiong, and R. Socher. Interpretable counting
in visual question answering. In Proceedings of the 6th International Conference on Learning Representations (ICLR),
2018.
[40] P. Wang, Q. Wu, C. Shen, A. Dick, and A. van den Hengel.
FVQA: Fact-based visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
[41] P. Wang, Q. Wu, C. Shen, and A. v. d. Hengel. The vqamachine: Learning how to use existing vision algorithms
to answer new questions. arXiv preprint arXiv:1612.05386,
2016.
[42] J. Weston, S. Chopra, and A. Bordes. Memory networks.
arXiv preprint arXiv:1410.3916, 2014.
[43] Q. Wu, D. Teney, P. Wang, C. Shen, A. Dick, and A. van den
Hengel. Visual question answering: A survey of methods
and datasets. Computer Vision and Image Understanding,
2017.
[44] Q. Wu, P. Wang, C. Shen, A. Dick, and A. van den Hengel. Ask me anything: Free-form visual question answering based on knowledge from external sources. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 4622–4630, 2016.
[45] F. Xiao, L. Sigal, and Y. J. Lee. Weakly-supervised visual
grounding of phrases with linguistic structures. In Proceedings of the International Conference on Computer Vision and
Pattern Recogition, 2017.
[46] H. Xu and K. Saenko. Ask, attend and answer: Exploring
question-guided spatial attention for visual question answering. In European Conference on Computer Vision, pages
451–466. Springer, 2016.
[47] Z. Yang, X. He, J. Gao, L. Deng, and A. Smola. Stacked
attention networks for image question answering. In Pro-
ceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 21–29, 2016.
[48] L. Yu, E. Park, A. C. Berg, and T. L. Berg. Visual madlibs:
Fill in the blank description generation and question answering. In Proceedings of the IEEE International Conference on
Computer Vision, pages 2461–2469, 2015.
| 1cs.CV
|
A Combinatorial Problem from Group Theory
arXiv:1701.08480v1 [math.CO] 30 Jan 2017
Eugene Curtin and Suho Oh
January 31, 2017
Abstract
Keller proposed a combinatorial conjecture on construction of an n-by-infinite matrix, which
comes from showing the existence of many orbits of different sizes in certain linear group actions
[1]. He proved it for the case n = 4, and we show that conjecture is true in the general case.
We also propose a combinatorial game version of the conjecture which even further generalizes
the problem.
1
The original problem
Keller conjectured [1] that given any n×∞ matrix of n element sets (Si,j ), it is possible to construct
an n × ∞ matrix (xi,j ) satisfying the following conditions:
• For all i and j, xi,j ∈ Si,j .
• The first n − 2 elements in each row are distinct and never repeated later in the row.
• For all t, the n sets {xi,1 , xi,2 , . . . , xi,t } (1 set from each row i) are distinct.
This problem comes from showing the existence of many orbits of different sizes in certain linear
group actions. Keller showed that it is possible when n = 4, and conjectured that it was true in
general.
The problem is very easy to solve in certain special cases. When all Si,j are all pairwise disjoint,
any choice of xi,j ’s would lead to a valid construction. When all Si,j = {1, . . . , n},
1
2
3
4
..
.
2
3
4
5
..
.
···
···
···
···
..
.
n 1 ···
n−2 n−1 n−1
n−1
n
n
n
1
1
1
2
2
..
..
..
.
.
.
n−3 n−2 n−2
···
···
···
···
..
.
n−1
n
1
2
···
···
···
···
···
n − 2 ···
gives an easy solution. So the problem is trivial when all Si,j are pairwise disjoint, or when all
of them are the same.
In Section 2, we will show that the conjecture is true for any value of n. In Section 3, we propose
a combinatorial game which generalizes this original problem.
1
2
The proof of the problem
We will first show that it is enough to prove when one has finitely many number of columns.
Lemma 1 (Kőnig). [2] Let G be a connected graph with infinitely many vertices such that each
vertex is adjacent to only finitely many other vertices. Then G contains a path with no repeated
vertices that starts at one vertex and continues from it through infinitely many vertices.
Corollary 1. If one can show that the conjecture is true for all n-by-k matrices, it implies that
the conjecture is true for n-by-∞ matrices.
Proof. Fix n and Si,j ’s. Let G be a graph where the vertices are given by all possible solutions
to the n-by-k cases, where k goes from 1 to ∞. We connect two vertices a, b by an edge if a is a
solution to the n-by-(k − 1) case, b is a solution to the n-by-k case, and a is exactly b with last
column deleted. Then this graph has infinitely many vertices, and only has finitely many number
of components (since there are only finitely many possible solutions to the n-by-1 case). Now pick
a component that has infinitely many vertices, and apply Kőnig’s infinity lemma in order to get a
non-repeated path of infinite length. This path gives a solution to the n-by-∞ case.
Now we will introduce some notation to make the proof smoother.
Definition 1. We say that a sequence {ai } is a representative of a sequence of sets {Si } if for
each i, ai ∈ Si . Given two sequences {a1 , a2 , . . .} and {b1 , b2 , . . .}, we say that they are equivalent
if {a1 , . . . , ak } = {b1 , . . . , bk } for all k. And we will say that they are compatible if {a1 , . . . , ak } =
6
{b1 , . . . , bk } for all k. A representative is regular if xj 6∈ {x1 , . . . , xj−1 } when Sj \ {x1 , . . . , xj−1 }
has at least 2 elements.
For example, the sequences {1, 2, 3, 1, 1, . . .} and {1, 2, 3, 2, 2, . . .} are equivalent. The sequences
{1, 2, 3, 4, 4, . . .} and {2, 3, 4, 5, 5, . . .} are compatible. One can think of a regular sequence as a
sequence where an element that has already appeared before may be chosen only if there aren’t
at least two new elements to choose from. Regularity of a sequence implies the second condition
of Keller’s original problem in Section 1. The following is the reformulation of Keller’s problem,
except that we relax the condition on the cardinality of the Si,j ’s.
Definition 2. We say that an n-by-k matrix where the entries are given by sets Si,j of cardinality
at least n is an (n, k)-system. A solution of this system is an n-by-k matrix with entries xi,j
such that:
• Each row of the solution is a regular representative of the corresponding row in the system.
• All rows are pairwise compatible.
We now prove our main result.
Theorem 1. There exists a solution to any (n, k)-system. Moreover, consider the (n, 1)-system
obtained by just looking at the Si,1 ’s. Any solution to this (n, 1)-system can be extended to a solution
of the (n, k)-system.
2
Proof. Set S to be the union of all the Si,j ’s. We will first show that it is possible to assume that
S
j≥2 Si,j = S for all rows. If not, extend the (n, k)-system to an (n, k + 1)-system where Si,k+1 = S
for all i. If we can find a solution to this (n, k + 1)-system, we can just delete the last column to
obtain a solution to the original (n, k)-system.
The statement is obviously true when n = 1. For sake of induction, assume the statement is
true for n, and we will show that it is also true for n + 1. Pick an arbitrary solution p of the
(n + 1, 1)-system. For each i 6= j, we set qi,j as the least t ≥ 2 such that pi,1 ∈ Sj,t. These numbers
are well defined thanks to the assumption we made in the beginning of the proof. Properly reorder
the rows so that q1,2 is the minimum among the qi,j ’s. Set a to be p1,1 . Set d1 = 1, and for each
i ≥ 2, denote di as the least l ≥ 2 such that a ∈ Si,l . (So di = q1,i for i ≥ 2.) Now reorder the
rows 2, . . . , n + 1 such that d2 ≤ · · · ≤ dn+1 . The positions (i, di )′ s will form a diagonal of sorts,
encoding the earliest possible position where a can appear in each row (except the first column).
Let b be pn+1,1 , the entry of the last row of the solution.
Now in the (n + 1, k)-system we are looking at, do the following:
• delete the last row,
• for each row i, for all columns t past the diagonal di (t ≥ di ), remove a from Si,t ,
• for each row i, for all early columns t before the diagonal (t ≤ min(di , n − 1)), remove b from
Si,t .
• and erase the entries at (i, di ).
(The description of the third operation might seem odd. Avoiding early occurrences of b in the
first n rows of x we get later on helps achieve compatibility with the eventual (n + 1)-th row of x,
and we allow later occurrences to maintain regularity.) Now we have an (n, k − 1)-system. Set y
to be yi,1 = pi,1 for 2 ≤ i ≤ n, and set y1,1 to be any element of S1,2 \ {a, y2,1 . . . , yn,1 }. Then y
is a solution to the (n, 1)-system, and we can extend this to a solution of the (n, k − 1)-system via
the induction hypothesis. Name this solution again as y. Insert a into the di -th position of each
row of y to get x. We claim that x is a solution to the (n, k) system obtained from the original
(n + 1, k)-system by deleting the last row. The rows of x are pairwise compatible since the rows
of y are pairwise compatible and a appears exactly once per each row of x. Now assume for the
sake of contradiction that some row i of x is not regular. This means there is some entry j such
that xi,j ∈ {xi,1 , . . . , xi,j−1 } and |Si,j \ {xi,1 , . . . , xi,j−1 }| ≥ 2. Now j obviously can’t be ≥ di since
y is regular and a appears only once per row. It also can’t be between 1 and n − 1 since repeated
elements in each row can only start appearing from column n. And for j between n and di , Si,j
hasn’t been changed so regularity of y gives a contradiction. This finishes the proof of the claim
that x is a solution to the (n, k) system obtained from the original (n + 1, k)-system by deleting
the last row.
All that remains now is to fill out the (n + 1)-th row of x. If dn+1 ≤ n, then dj ≤ n for all j
and a must appear before b in each of the first n rows of x. In this case, any regular representative
not using a will suffice. When dn+1 > n, consider any regular representative of the last row up to
the n-th position, b = z1 , . . . , zn . Each row of x puts a partial ordering on {z2 , . . . , zn } : look at
the first position where zi appears in that row and order the zi ’s according to that position so that
the ones appearing earlier are smaller. Extend that ordering into a total ordering by making the
elements that didn’t appear bigger than the ones that appeared. So we get n − 2 total orderings
3
from row 3 to row n in x. There is some symbol, say zt , that is not the last in any of the n − 2
orderings. Replace zt with any z ′ ∈ Sn+1,t \ {z2 , . . . , zn }, and extend the sequence to a regular
representative of the last row of the system by imposing xn+1,j 6= zt for n + 1 ≤ j < dn+1 and
xn+1,j 6= a for j ≥ dn+1 . (This is possible since we are always allowed to avoid one particular
element while constructing a regular representative.)
Now we want to show that this sequence, the new n+1-th row of x, is compatible with rows 1 to
n of x. For row 1 it is obvious since row 1 starts with a and (n + 1)-th row does not contain a at all.
Denote {xi,1 , . . . , xi,j } by C(i, j). Assume for the sake of contradiction that C(i, j) = C(n + 1, j)
for some 3 ≤ i ≤ n. Since row n + 1 starts with b and contains no a, we need b ∈ C(i, j) and
a∈
/ C(i, j), hence j ≥ n. But this implies {z2 , . . . , zn } \ {zt } ⊂ C(n + 1, j) = C(i, j). From the
definition of zt , it can’t be the last element to occur among {z2 , . . . , zn } in row 3 to row n, so it
has to be contained in C(i, j). We get zt ∈ C(i, j) = C(n + 1, j), which implies j ≥ dn+1 . From
di ≤ dn+1 , we have a ∈ C(i, j) = C(n + 1, j), but we get a contradiction since a never appears in
the last row. The only case remaining to check is the compatibility against row 2. Again assume
for the sake of contradiction that C(2, j) = C(n + 1, j) for some j. Since a never appears in the
(n + 1)-th row, j < d2 . But b being the first element in the n + 1-th row means b ∈ C(2, d2 − 1).
This contradicts the choice of a, which should be an entry of p that can be repeated the earliest
among different rows. Therefore this finishes the proof of the claim that the n + 1-th row of x is
compatible with all other rows of x. Hence x is a solution of the (n + 1, k)-system, and this proves
the theorem by induction.
As we have mentioned in Corollary 1, combining this result with Kőnig’s Lemma implies the
following result:
Theorem 2. Keller’s conjecture is true for any value of n.
3
Combinatorial game version
In this section we propose a combinatorial game. If a solution to such game always exists, it would
imply the solution to Keller’s problem. Recall that [n] denotes {1, · · · , n}.
Problem 1. Fix a natural number n. You start with n sets Si = {i}. Each round, for each i ∈ [n],
an adversary offers two elements from [n] \ Si . Your goal is to pick one of the two elements offered,
to add to Si , such that all Si ’s are still pairwise different. Can you always make the correct choice
for n − 2 rounds? (So that Si ’s become the n subsets of cardinality n − 1 of the set [n]).
It is easy to see that solving this problem would imply a solution to Keller’s problem. A slightly
weaker version of the problem can be described in terms of finding n vertex disjoint paths in n
subposets of a Boolean lattice Bn . We say that a subposet P of Bn is branching if every nonmaximal element A ∈ P is covered by at least 2 elements of P of cardinality |A| + 1. We call the
maximal elements of a branching poset to be the leaves of the poset. A subposet of Bn is rooted
if it contains the minimal element ∅. A subposet of Bn is wide if it contains all the singletons. A
full path is a simple path from some {i} to some [n] \ {j} in Bn .
Problem 2. Consider the Boolean lattice Bn . Let P1 , . . . , Pn be rooted branching subposets of Bn ,
such that {i} ∈ Pi for all i ∈ [n]. Can one find a simple path for each Pi so that they are all vertex
disjoint?
4
It is not hard to see that a solution to this problem would give a solution to Keller’s problem:
think of a sequence a1 , . . . , ak as the set {a1 , . . . , ak }. So considering all possible sets you get
from S
all possible sequences of a given row in Keller’s problem gives us a subposet of Bn , where
n = | i,j Si,j |. This poset is a branching poset thanks to the regularity condition.
We will show that this problem is easily solved when P1 = · · · = Pn . We denote this poset as
P , and P has to be wide. Assume for the sake of contradiction that we can’t find n vertex disjoint
full paths in P . Using the following Menger’s theorem tells us that there is a set C of vertices in P
that separates the singletons from the complement of the singletons, with |C| < n.
Theorem 3 ([3]). Let G be a graph with A and B two disjoint vertex subsets of G, with |A| =
|B| = n. Then either G contains n pairwise vertex disjoint paths from A to B, or there exists a set
of fewer than n vertices that separates A from B.
Take C to be an inclusion-wise minimal set that separates the singletons from the complement
of the singletons. Now consider all paths from ∅ to a vertex of C, that does not pass through other
vertices of C. Then the union of these paths is again a wide rooted branching poset, and C is
exactly the set of maximal elements of this poset. Now we will show the following claim:
Proposition 1. Any wide rooted branching poset P of Bn has at least n leaves.
To do so we will be using the following lemma:
Lemma 2. Any rooted branching subposet P of Bn has at least h leaves, where h is the maximal
height of the subposet.
Proof of Lemma 2. Let P x be the induced subposet below the singleton x, so that P x = {y|y ≥P
{x}}. We use induction on the number of vertices to prove the lemma. When there is only one
vertex, the statement is trivial. Assume for the sake of induction that the statement is true for
subposets of size < k. Let v1 , . . . , vr be the vertices at height 2. Relabel them as {{1}, . . . , {r}},
such that P 1 has maximal height. Note that P 1 is isomorphic to a rooted branching subposet of
Bn−1 of height h − 1, so we may apply the induction hypothesis to obtain that P 1 has at least
h − 1 leaves. Now consider P 2 . Starting from {2}, go down P 2 by avoiding the element 1. Note
that if A ∈ P and 1 6∈ A, then at most one set containing 1 covers A, so branching guarantees we
can avoid 1. Thus we reach a leaf of P 2 which does not contain 1 and thus cannot be in P 1 . This
proves that P has at least h leaves.
Now we will prove the Proposition:
Proof of Proposition 1. Assume for the sake of contradiction that we have at most n − 1 leaves.
We will show a contradiction by a fractional counting argument
For each singleton
P on the leaves.
P
1
1
s and leaf l, define f (s, l) = |l| if s ≤P l, 0 otherwise. Then s,l f (s, l) ≤ l |l| ∗ |l|
≤ n − 1. Since
P
there are n singletons, there has to exist some singleton q such that l f (q, l) < 1. Now consider
P q , the subposet
q. By Lemma 2, P q has at least h leaves where h is the height of P q . So
P
P below
1
1
l f (q, l) =
l∈P q |l| ≥ h ∗ h = 1, which leads to a contradiction.
The above proposition tells us that |C| < n cannot happen, and therefore has to be n vertex
disjoint full paths in P .
Corollary 2. It is possible to find a set of vertex disjoint paths for Problem 2 in the case where
P1 = · · · = Pn .
5
References
[1] Thomas Keller. A combinatorial problem arising in group theory. Australasian Journal of
Combinatorics, 61(2):175–181, 2014.
[2] Dénes Kőnig. Sur les correspondances multivoques des ensembles. Fundamenta Mathematicae,
8:114–134, 1926.
[3] Karl Menger. Zur allgemeinen kurventheorie. Fundamenta Mathematicae, 10:95–115, 1927.
6
| 4math.GR
|
arXiv:1803.03311v1 [math.CT] 8 Mar 2018
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
HANNO BECKER
Abstract. We study liftings of abelian model structures to categories of chain complexes and construct a realization functor D(A ) → Ho(M) for any cofibrantly generated, hereditary abelian model structure M on a Grothendieck category A .
1. Introduction
Given a Grothendieck abelian category A , we may form its derived category D(A ),
which – very vaguely – one might think of as underlying a generic homotopy theory built
upon A . While we don’t know how to elaborate this into a precise statement, this article
deals with the following approximation:
Question (Realization Problem). If M is a ’reasonable’ hereditary abelian model structure on A with triangulated homotopy category Ho(M), does there exist a triangulated
functor D(A ) → Ho(M) such that the following diagram commutes up to isomorphism?
Ho(M)
real
D(A )
A
This question will be answered affirmatively for cofibrantly generated, hereditary
abelian model structures on Grothendieck categories in this article.
Outline. In this introduction, we will begin by recollecting some by now classical results
of Gorenstein homological algebra leading to a solution of the realization problem for
the Gorenstein-projective model structure over an Iwanaga-Gorenstein ring. The rest of
the introduction will then indicate how these results can be generalized, and the details
and proofs of our three main theorems will be supplemented in the remaining sections.
Acknowledgements. The results of this article are mostly part of my PhD thesis written under the supervision of Prof. Catharina Stroppel at the University of Bonn. I am
very grateful to Catharina for her help and support in all phases of the project. I am
also grateful to Jan Stovicek for a very stimulating discussion at the beginning of my
work on the topic of this article, and to Pieter Belmans for numerous helpful comments.
Date: March 12, 2018.
1
2
HANNO BECKER
1.1. Gorenstein homological algebra via triangulated categories. Let R be an
Iwanaga-Gorenstein ring [EJ11, Definition 9.1.1], i.e. R is Noetherian and of finite
injective dimension as a left and as a right module over itself. Further, denote
G-Proj(R) := {X ∈ R -Mod | ∀ i > 0 : ExtiR (X, R) = 0}
its Frobenius category of Gorenstein-projective R-modules, and
G-Proj(R) := G-Proj(R)/ Proj(R)
its stable category, a triangulated category. We have the following well-known theorem:
Theorem 1.1. There is an equivalence of triangulated categories
H : G-Proj(R) ∼
= Kac (Proj(R)),
where the rhs. is the homotopy category of acyclic complexes of projective R-modules.
This is a consequence of G-Proj(R) being a Frobenius category with projectiveinjective objects given by Proj(R) together with Theorem 1.7 below.
Next, the chain-complex description of the stable category relates to the ordinary
derived category through a recollement:
Theorem 1.2 ([Mur07, Theorem 5.15], [Kra05, Corollary 4.3]). There are recollements
of triangulated categories
Kac (Proj(R))
Kac (Inj(R))
K(Proj(R))
D(R)
K(Inj(R))
D(R)
In fact, the injective recollement exists in a considerably more general situation, see
[Kra05]. Finally, we have the following description of the stabilization functor [Kra05,
§5] of the recollement from Theorem 1.2. In the remainder of this paper, left-directed
left resp. right adjoints in recollements we be denoted λ resp. ρ.
Theorem 1.3 ([Kra05, §7]). The composition
ρ
λ
H
R -Mod ֒→ D(R) ֒−−−→ K(Proj(R)) −−−→
→ Kac (Proj(R)) −−−→ G-Proj(R)
is equivalent to the Gorenstein-projective approximation
Here, by Gorenstein-projective approximation the right adjoint R -Mod → G-Proj(R)
to the inclusion functor is meant.
As the next section will show, Theorem 1.3 is essentially the solution to a realization
problem in the present situation. Also, note that Theorems 1.1, 1.2, 1.3 have analogues
in the Gorenstein-injective case.
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
3
1.2. Gorenstein homological algebra via model categories. The results of the
previous section can be lifted to the world of abelian model categories and thereby be
seen as intimately related to a suitable realization problem in this context.
For brevity, we refer to [Bec14] for the basics on abelian model categories and also
keep the notation of op. cit.; in particular, abelian model structures will be denoted
M, and their classes of cofibrant, weakly trivial and fibrant objects in an abelian model
category will be denoted C, W and F, respectively.
The first step is the introduction of an abelian model structure for G-Proj(R):
Theorem 1.4 ([Hov02, Theorem 8.6]). Let R be a Gorenstein ring.
(1) There exists an abelian model structure on R -Mod, called the Gorenstein projective model structure and denoted MG-Proj (R), with C = G-Proj(R), W =
P<∞ (R) (the modules of finite projective dimension) and F = R -Mod.
(2) There exists an abelian model structure on R -Mod, called the Gorenstein injective model structure and denoted MG-inj (R), with C = R -Mod, W = P<∞ (R)
and F = G-inj(R).
Moreover, MG-Proj (R) and MG-inj (R) are cofibrantly generated, and their homotopy categories are canonically equivalent to the stable categories of Gorenstein-projective resp.
Gorenstein-injective R-modules.
The previous section solves the realization problem for MG-Proj (R): The Gorensteinprojective approximation of an R-module is its cofibrant replacement in the Gorensteinprojective model structure MG-Proj (R), so Theorem 1.3 shows that the diagram
Ho(MG-Proj (R)) ∼
= G-Proj(R)
H ◦λ ◦ ρ
D(R -Mod)
R -Mod
commutes up to isomorphism, as wished.
The following theorem lifts Theorem 1.2 to the level of abelian model categories. Its
generalizability to suitable abelian model structures will be the key to constructing a
realization functor in the general setting.
Theorem 1.5 ([Bec14, Proposition 2.2.4]). The injective recollement associated with a
Gorenstein ring R admits a lifting to a butterfly of abelian model structures (see below):
Mco
sing (R)
Minj (R)
R
R
L
R
L
Mco (R)
L
L
L
R
i
Mco
sing (R)
R
L
R
m
Minj (R)
Similarly, the projective recollement from Theorem 1.2 arises from a butterfly.
For the definition of the model structures involved in the previous theorem, see loc.cit.
4
HANNO BECKER
Definition 1.6. A diagram of abelian model structures and Quillen adjunctions of shape
Ml
Mr
R
(⊲⊳)
R
L
R
L
M
L
L
L
R
L
R
R
M′l
M′r
is called a butterfly if the following properties hold:
(⊲⊳.i) The left and right vertical adjunctions are Quillen equivalences.
(⊲⊳.ii) The two wings in the following following diagram commute:
Ho(Ml )
Ho (Mr )
L id
(Ho(⊲⊳))
∼
=L id
L id
∼
=
R id
Ho(M)
R id
Ho (M′l )
R id
Ho (M′r )
(⊲⊳.iii) The derived horizontal adjunctions in (⊲⊳) realize the homotopy categories of the
left and right sides as full subcategories of the homotopy category of the middle
term, and the horizontal functors in (Ho(⊲⊳)) form an exact sequence.
1.3. Generalization. We now describe our generalizations of Theorems 1.1, 1.2, 1.3.
Setup. In the following, let A be a Grothendieck abelian category and let M = (C, W, F)
be a cofibrantly generated and hereditary (cgh) abelian model structure on A . In particular, C ∩ F is a Frobenius category with respect to the short exact sequences inherited
from A , its class of projective-injective objects equals ω := C ∩ W ∩ F, and the homotopy
category Ho(M) is canonically equivalent to the stable category C ∩ F/C ∩ W ∩ F.
Recall the following basic theorem on Frobenius categories generalizing Theorem 1.1:
Theorem 1.7. Let F be a Frobenius category and ω := Proj-Inj(F ) its class of projectiveinjective objects. Then there are equivalences of triangulated categories
Q0 , Z0 : Kac (ω) −−→ F = F /ω
which coincide up to shift, Q0 ∼
= Σ ◦ Z0 .
This suggests that a lifting of M to a Quillen equivalent model structure on Ch(A )
e on Ch(A )
could be obtained by providing a choice for an abelian model structure M
such that the class of acyclic complexes with values in ω := C ∩ W ∩ F is equal to the
e The base for implementing this idea will be the following
class of bifibrant objects in M.
Theorem of Gillespie extending the author’s previous work:
Theorem 1.8 ([Bec14, Proposition 1.4.2], [Gil14, Theorem 1.1]). Let A be an abelian
e and (C,
e D) be complete (small), hereditary cotorsion pairs over A
category and (C, D)
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
5
e ⊆ C and C ∩ D
e =C
e ∩ D. Then there exists a unique (cofibrantly generated) abelian
with C
model structure (C, W, D), and its class W of weakly trivial objects is given by
(1.1)
e →C
e→0
W = {X ∈ A | ∃ 0 → X → D
e →C
e → X → 0}.
0→D
and
The situation of Theorem 1.8 turns out to be abundant and deserves its own name:
α
Notation 1.9. We write (D′ , E′ ) −
→ (D, E), or (D′ , E′ ) −
→ (D, E), as an abbreviation for
′
′
(D , E ) and (D, E) being complete and hereditary cotorsion pairs having the same core
D′ ∩ E′ = D ∩ E and satisfying D′ ⊂ D. Such a situation will be called a localization
context. Its induced model structure from Theorem 1.8 is denoted Loc(α) := (D, ?, E′ )
on A and called its localization.
♦
Leaving the details for later, the technical heart of this article is the following zoology
of localization contexts induced by a single cgh abelian model structure. For the notation,
we refer to Definition 2.1.
Theorem A. Let M = (C, W, F) be a cofibrantly generated, hereditary abelian model
structure on the Grothendieck category A with core ω := C ∩ W ∩ F. Then Figure 1
shows a diagram of localization contexts on Ch(A ) with common core.
An arrow (D′ , E′ ) → (D, E) signifies (independent of its style) that D′ ⊆ D, and ⊥
indicates that the corresponding entry is the left/right orthogonal of the other entry.
e ∩ dw- W ∩ F)
(⊥, dg- F
e ∩ dw- W ∩ F)
(⊥, F
e
(C^
∩ W, dg- F)
e
(dg- C^
∩ W, F)
(dw- C ∩ W, ⊥)
(dw- C ∩ W, ⊥)
(⊥, dw- W ∩ F)
(⊥, dw- W ∩ F)
e dg- W
^
(C,
∩ F)
e W
^
(dg- C,
∩ F)
e ∩ dw- C ∩ W, ⊥)
(C
e ∩ dw- C ∩ W, ⊥)
(dg- C
Figure 1 Localization contexts on Ch(A ) induced by a cofibrantly
generated, hereditary abelian model structure on A
With the multitude of localization contexts from Theorem A and each of them inducing an abelian model structure on Ch(A ) by virtue of Theorem 1.8, we end up with
6
HANNO BECKER
a large number of abelian model structures on Ch(A ). The following helps gaining a
rough understanding on their interrelation and will be elaborated later:
(i) Each triangle
in Figure 1 gives rise to a localization sequence between the
three model structures induced by its edges.
in Figure 1 yields a butterfly-shaped diagram of adjunctions
(ii) Each square
between the two localization sequences associated via (i) to its triangle faces.
However, these are not necessarily butterflies.
Using this, we get the desired generalization of Theorem 1.2:
Theorem B. For a cofibrantly generated and hereditary abelian model structure M =
(C, W, F) over a Grothendieck category A , consider Figure 2. It shows a diagram of
identity Quillen adjunctions between cofibrantly generated and hereditary abelian model
structures on Ch(A ) with the following properties:
(1)
(2)
(3)
(4)
Both horizontal layers are butterflies in the sense of Definition 1.6.
All the vertical adjunctions are Quillen equivalences.
The four model structures on the left side are Quillen equivalent to M.
The four model structures on the right side hand have their homotopy category
canonically equivalent to D(A ).
Finally, the induced functor D(A ) → Ho(M) solves the realization problem:
Theorem C. In the situation of Theorem B, the composition
e −→ Ho(C
e ∩ dw- C ∩ W, ?, dg- F)
e ∼
D(A ) ∼
∩ W, Acyc(A ), dg- F)
= Ho(dg- C^
= Ho(M)
of the left stabilization functor associated to the lower butterfly in Figure 2 and the
equivalence from Theorem B.(3) makes the following diagram commutative,
Ho(M)
D(A )
A,
thereby solving the realization problem for M.
2. Proof of Theorem A
We begin by recalling the definitions of some crucial classes of complexes; see e.g.
[Gil04, Definition 3.3], [Šťo10, Notation 4.1]:
Definition 2.1. Let (C, D) be a cotorsion pair in the abelian category A .
(1)
(2)
(3)
(4)
dw- C := Ch(C) and dw- D := Ch(D).
e denotes the class of acyclic complexes in A with syzygies in C.
C
e denotes the class of acyclic complexes in A with syzygies in D.
D
e := ⊥ D
e and dg- D
e := C
e⊥ .
dg- C
Example 2.2. Considering the cotorsion pairs (P, A ) and (A , I) one recovers the classes
of dg projective and dg injective complexes as dg- P and dg- I, respectively.
♦
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
e ?, dg- F
e ∩ dw- W ∩ F)
(C,
R
R
L
L
R
L
L
e ?, dg- F
e ∩ dw- W ∩ F)
(dg- C,
L
L
e ?, F
e ∩ dw- W ∩ F)
(dg- C,
e ∩ dw- C ∩ W, ?, dg- F)
e
(C
L
R
R
R
R
L
L
e ∩ dw- C ∩ W, Acyc(A ), [C
e ∩ dw- C ∩ W]⊥ )
(dg- C
L
L
R
e ∩ dw- C ∩ W, ?, dg- F)
e
(dg- C
R
R
L
e ∩ dw- C ∩ W, ?, F)
e
(dg- C
L
e ∩ dw- W ∩ F], Acyc(A ), dg- F
e ∩ dw- W ∩ F)
(⊥ [F
R
L
e Acyc(A ), dg- W
^
(dg- C,
∩ F)
R
R
R
7
L
L
R
L
R
R
e
(dg- C^
∩ W, Acyc(A ), dg- F)
Figure 2 Abelian model structures on Ch(A ) induced by an abelian
model structure on a Grothendieck abelian category A
Proposition 2.3. Let (C, D) be a cotorsion pair over A and X ∈ Ch(A ).
e
(1) X ∈ dg- C if and only if X ∈ dw- C and [X, D] = 0 for all D ∈ D.
e
(2) X ∈ dg- D if and only if X ∈ dw- D and [C, X] = 0 for all C ∈ C.
Further, we have the following inclusions:
e and Ch+ (D) ⊂ dg- D.
e
(iii) Ch− (C) ⊂ dg- C
+
−
⊥
e
e ⊂ [dw- C]⊥ .
(iv) Ch (A ) ∩ C ⊂ [dw- D] and Ch (A ) ∩ D
e ⊂ dg- C
e and D
e ⊂ dg- D.
e
(v) C
Proof. This is mostly contained in [Gil04, §3], but for convenience we include an argument here. First, the (exact) adjoints G± : A Z → Ch(A ) to the (exact) forgetful
e resp. D.
e Hence, given X ∈ dg- C
e
functor (−)♯ : Ch(A ) → A Z map CZ resp. DZ to C
1
1
Z
−
♯
♯
⊥
Z
Z
∼
and D ∈ D , we have 0 = ExtCh(A ) (X, G (D)) = ExtA Z (X , D), so X ∈ (D ) = C ,
e ⊂ dw- D.
i.e. X ∈ dw- C. Similarly, we have dg- D
Next, if X ∈ dw- D and Z ∈ dw- C, then any short exact sequence 0 → X → Y →
Z → 0 in Ch(A ) is degree-wise split, so that Ext1Ch(A ) (Z, X) is canonically isomorphic
to the extension group Ext1dw- Ch(A ) (Z, X) ∼
= [Z, ΣX] with respect to the degree-wise
split exact structure on Ch(A ). This proves the first two claims (i) and (ii).
8
HANNO BECKER
The proof of the inclusions in (iii) and (iv) is analogous to the proof of the classical fact
that chain maps from bounded above complexes of projectives to acyclic complexes are
nullhomotopic, as are chain maps from acyclic complexes to bounded below complexes
e ⊂ dg- C
e and D
e ⊂ dg- D
e from part (v) are proved
of injectives. Finally, the inclusions C
in [Gil04, Lemma 3.9].
The following beautiful theorem is the result of long work by Gillespie [Gil04, Gil06]
in his studies of the flat model structures on Ch(R -Mod) and Ch(OX ) (for a survey, see [Hov07, §7]) and new results of Stovicek [Šťo10] on deconstructible classes in
Grothendieck categories. To be precise, [Gil06, Proposition 3.6, Corollary 3.7] essentially
e dg- D)
e crucial for part (ii) is guarprove parts (i) and (iii), while the completeness of (C,
e established in [Šťo10, Theorem 4.2]. We collect
anteed by the deconstructibility of C
these arguments and give a proof for convenience of the reader.
Theorem 2.4 ([Gil06, Šťo10]). Let A be a Grothendieck category and (C, D) be a small
and hereditary cotorsion pair in A . Then the following hold:
e D)
e is a small, hereditary cotorsion pairs in Ch(A ).
(1) (dg- C,
e
e
(2) (C, dg- D) is a small, hereditary cotorsion pairs in Ch(A ).
e = dg- C
e ∩ Acyc(A ) and D
e = dg- D
e ∩ Acyc(A ).
(3) C
e Acyc(A ), dg- D)
e is a cofibrantly
Corollary 2.5. In the situation of Theorem 2.4, (dg- C,
generated abelian model structure on Ch(A ) with homotopy category D(A ).
Example 2.6. Applying Corollary 2.5 to the cotorsion pair (A , I) gives rise to the injective model (Ch(A ), Acyc(A ), dg- I) for D(A ). Similarly, if A has enough projectives
the cotorsion pair (P, A ) yields the projective model (dg- P, Acyc(A ), Ch(A )) for D(A ).
A nontrivial example – and in fact the example that started the theory – is obtained
from the flat cotorsion pair (flat(R), flat(R)⊥ ) on R -Mod, with R any ring: in this case,
one obtains Gillespie’s flat model structure.
♦
Proof of Theorem 2.4. We first prove (i). To begin, [Gil04, Proposition 3.6] shows that
e D)
e and (C,
e dg- D)
e are cotorsion pairs; in addition to (C, D) being a cotorsion pair,
(dg- C,
this only needs the assumption that C is generating and D is cogenerating in A .
e D)
e and (C,
e dg- D),
e applying [Šťo10, Theorem 4.2]
Concerning the smallness of (dg- C,
(generalizing the ideas used by Gillespie [Gil04] in the case of the flat model structure
e and C
e are deconstructible, so it remains to check that
on Ch(R -Mod)) shows that dg- C
e
C is generating. For this, note that since C is generating in A , CZ is generating in A Z ;
the counit G+ (X ♯ ) → X being an epimorphism for X ∈ Ch(A ), it follows that G+ (CZ )
e so C
e is generating, too.
is generating in Ch(A ). We have G+ (CZ ) ⊂ C,
e
e
e
e
To check that (dg- C, D) and (C, dg- D) are hereditary, it suffices (by [Bec14, Corollary
e is resolving while dg- D
e is coresolving. We only check that
1.1.12]) to show that dg- C
e
e
dg- C is resolving, the proof of dg- D being analogous. For that, recall from Proposition
e consists of those X ∈ dw- C for which Hom∗ (X, D) ∈ Acyc(Z) for all
2.3 that dg- C
A
e and suppose 0 → X → Y → Z → 0 is a short exact sequence in Ch(A ) with
D ∈ D,
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
9
e Then, firstly, X ∈ dw- C since C is resolving and Y, Z ∈ dw- C. Further, for
Y, Z ∈ dg- C.
e
any D ∈ D (even any D ∈ dw- D), the sequence of complexes of abelian groups
0 → Hom∗A (Z, D) → Hom∗A (Y, D) → Hom∗A (X, D) → 0
is exact, and since Hom∗A (Z, D) and Hom∗A (Y, D) are exact by our assumption that
e it follows that Hom∗ (X, D) is exact, too.
Y, Z ∈ dg- C,
A
e ∩ Acyc = C,
e and in view of
Concerning (ii), [Gil04, Theorem 3.12] shows that dg- C
e
e
[Gil04, Lemma 3.14(1)] and the completeness of (dg- C, D) we already proved, this also
e ∩ Acyc = D.
e
shows dg- D
Suppose now that M = (C, W, F) is a cofibrantly generated and hereditary abelian
model structure on the Grothendieck category A , and put ω := C ∩ W ∩ F, the core of
M. We will be concerned with quite a number of induced cotorsion pairs all of which
e∩W
e =: ω
^
will have C
∩ F = C^
∩W ∩ F
e as their core, the class of acyclic complexes
with syzygies in ω. In view of the following lemma, these are precisely the contractible
complexes with entries in ω:
Lemma 2.7. For A abelian and X ∈ Acyc(A ), the following are equivalent:
(1) X is contractible.
(2) The exact sequences 0 → Zn X → X → Zn+1 X → 0 split.
Proof. Omitted.
Lemma 2.8. Let A be an abelian category and ω be a self-orthogonal class of objects
in A , i.e. ω ⊂ ⊥ ω. Then ω
e , the class of contractible complexes with entries in ω, is the
largest self-orthogonal, Σ-stable class in Ch(A ) contained in dw- ω.
Proof. By self-orthogonality of ω, any short exact sequence 0 → X → Y → Z → 0 with
X, Y ∈ dw- ω is degree-wise split, and hence determined by a homotopy class in [Z, ΣX].
This shows that ω
e is self-orthogonal in Ch(A ).
Conversely, suppose E ⊂ dw- ω is self-orthogonal and Σ-stable, i.e. ΣE ⊂ E. Then,
given any X ∈ E we have 0 = Ext1Ch(A ) (ΣX, X) = Ext1dw- Ch(A ) (ΣX, X) ∼
= [ΣX, ΣX],
so X is contractible.
Proof of Theorem A. We begin by showing that all cotorsion pairs are small.
For the top square, it suffices to check that the right hand sides of the cotorsion
pairs listed in it are of the form S⊥ for a generating set S ⊂ Ch(A ). This property
e dg- F
e and dw- W ∩ F
is preserved under intersection, so we need to check it for F,
h
i⊥
e = dg- C^
only. Concerning the first two, we know from Theorem 2.4 that F
∩W
i⊥
h
e = C^
∩ W and dg- C^
∩ W both deconstructible and generating.
∩ W , with C^
and dg- F
For dw- W ∩ F, note that X ∈ dw- W ∩ F if and only if for all C ∈ CZ we have 0 =
Ext1A Z (C, X ♯ ) ∼
= Ext1Ch(A ) (G+ (C), X), so that by cocontinuity and exactness of G+ we
conclude that dw- W ∩ F = G+ (S)⊥ for S ⊂ A some set chosen such that C = filt- S; as C
is generating, we may assume S generating, too, and then G+ (S) is generating in Ch(A )
10
HANNO BECKER
since the counit G+ X ♯ → X is an epimorphism for all X ∈ Ch(A ). This concludes the
proof that all cotorsion pairs in the upper square are small.
For the middle square, all cotorsion pairs contained in it are of the form studied
in Theorem 2.4, hence small. Finally, to prove that the cotorsion pairs in the lower
square are small it suffices to show that their left hand sides are generating and deconstructible. They are generating as they all contain the generating class G+ (C ∩ W), and
deconstructibility follows from the stability of deconstructible classes under intersection
e dg- C
e and dw- C^
[Šťo10, Proposition 2.9] as well as the deconstructibility of C,
∩ W [Šťo10,
Theorem 4.2].
Next we check that all cotorsion pairs are hereditary. For the ones in middle square,
this follows from Theorem 2.4 above. Concerning the ones in the upper square, their
e dg- F
e and dw- W ∩ F,
right hand sides are coresolving as intersections of the classes F,
each of which is coresolving: the first two are again treated as part of Theorem 2.4, while
dw- W ∩ F is coresolving since W ∩ F is. Applying [Bec14, Corollary 1.1.12] then shows
that all cotorsion pairs in the upper square are hereditary, and the reasoning for the
lower square is analogous.
Finally we check that all cotorsion pairs have core equal to ω
e , the class of contractible
complexes with values in C ∩ W ∩ F. First, using the fact Ext1dw- Ch A (−, −) ∼
= [Σ(−), −]
it is a quick check that ω
e is contained in all the cores. For the reverse inclusion, Lemma
2.8 and the stability under shift of all the classes involved show that it suffices to check
that all cores are contained in dw- ω. For the middle square, this is clear. For the upper
square, all the cotorsion pairs in it have their right hand sides contained in dw- W ∩ F,
and the fact that they are all connected to a cotorsion pair in the middle row by a chain
of arrows shows that their left hand sides are all contained in dw- C. Similarly, the left
hand sides of the cotorsion pairs in the lower square are all contained in dw- C ∩ W, while
all of them receiving an arrow from the middle square shows that their right hand sides
are all contained in dw- F.
3. Proof of Theorem B
We now elaborate on the following statements, leading to the proof of Theorem B:
in Figure 1 gives rise to a localization sequence between the
(i) Each triangle
three model structures induced by its edges (Proposition 3.5).
(ii) Each square
in Figure 1 yields a butterfly-shaped diagram of adjunctions
between the two localization sequences associated via (i) to its triangle faces.
However, these are not necessarily butterflies.
(iii) The dashed arrows in Figure 1 yield four models for D(A ) (Proposition 3.4).
(iv) The snaked arrows yield four models structures on Ch(A ) Quillen equivalent to
our given model structure (C, W, F) on A (Theorem 3.1).
(v) The zigzag arrows induce model structures analogous to the co- and contraderived
model structures, and the upper right and lower left triangles induce localization
sequences connecting them to the models for D(A ) from (iii) and the model
structures associated to the dashdotted arrows. See Example 3.6.
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
11
(vi) Each of the two tilted squares in Figure 1 composed out of two dashed, two snaked
and one dotted arrow, gives rise to a butterfly between the model structures
associated to the dotted arrows and the models from (iii) and (iv).
We begin with the model structures induced by the snaked arrows in Figure 1:
Theorem 3.1. Let M = (C, W, F) be a cofibrantly generated, hereditary abelian model
structure on the Grothendieck category A . Then there is a square of cofibrantly generated
abelian model structures on Ch(A ) and identity Quillen equivalences:
(3.1)
e ?, dg- F
e ∩ dw- W ∩ F)
(C,
R
L
R
L
e ∩ dw- C ∩ W, ?, dg- F)
e
(C
e ?, F
e ∩ dw- W ∩ F)
(dg- C,
R
L
R
L
e ∩ dw- C ∩ W, ?, F)
e
(dg- C
Their homotopy categories are equivalent to the homotopy category of acyclic complexes
with entries in C ∩ W ∩ F and syzygies in C ∩ F. Moreover, there are Quillen equivalences
(3.2)
(3.3)
e ∩ dw- C ∩ W, ?, dg- F)
e ⇄ (C, W, F) : ι0
Q0 : (C
e ?, F
e ∩ dw- W ∩ F) : Z0
ι0 : (C, W, F) ⇄ (dg- C,
which on the homotopy categories yield the classical equivalences from Theorem 1.7.
In particular, one has to beware that the derived adjoint equivalences of (3.2) and
(3.3) are not isomorphic, but are shifts of one another.
Example 3.2. Suppose R is a Gorenstein ring and consider Hovey’s Gorenstein projective model structure MG-Proj (R) = (G-Proj(R), P<∞ (R), R -Mod) on R -Mod. In
this case, we recover some constructions and results of [Bec14, Sections 2, 3.1]: Namely,
e ∩dw- C∩W, ?, dg- F)
e coincides with the projective singular contraderived model struc(C
p ctr
ture Msing (R) := (Ch(Proj(R)), ?, Ch(R)) (see [Bec14, Proposition 2.2.1]) since the
syzygies of any acyclic complex of projectives are automatically Gorenstein projective.
The equivalence (3.3) therefore agrees with the Quillen equivalence obtained in [Bec14,
e ∩ dw- C ∩ W, ?, F)
e coincides with the clasProposition 3.1.3]. Also, we claim that (dg- C
sical singular contraderived model structure Mctr
(R)
:=
(Ch(Proj(R)),
?, Acyc(R)) (see
sing
⊥g
[Bec14, Definition 2.1.2]), i.e. that Ch(Proj(R)) ⊆ I<∞ : For this, note first that for
P ∈ Ch(Proj(R)) the short exact sequence 0 → P → G+ (ΣP ♯ ) → ΣP → 0 exhibits
P as the syzygy of ΣP in the abelian category Ch(R), since G+ (ΣP ♯ ) is projective in
Ch(R); see [Bec14, Section 1.3] for the relevant results and notions. Iterating this procedure, we see that any P ∈ Ch(Proj(R)) is an arbitrarily high syzygy in Ch(R). On the
<∞ admits a finite resolution by contractible complexes of
other hand, any complex in Ig
injectives, i.e. has finite injective dimension in the abelian category Ch(R). The claim
<∞ follows.
♦
Ch(Proj(R)) ⊆ ⊥ Ig
Proof of Theorem 3.1. It only remains to prove that (3.2) and (3.3) are Quillen equivalences. We start by checking that (3.2) is a Quillen adjunction. Suppose ι : X → Y
e ∩ dw- C ∩ W, ?, dg- F),
e i.e. ι is a monomorphism in Ch(A ) with
is a cofibration in (C
12
HANNO BECKER
e ∩ dw- C ∩ W. Since C
e ⊂ Acyc(A ), it follows as in
cokernel Z := coker(ι) belonging to C
0
0
0
[Bec14, Proposition 3.1.3] that 0 → Q X → Q Y → Q Z → 0 is exact in A , and as
Q0 Z ∈ C by definition, it follows that Q0 ι is a cofibration in (C, W, F). If ι is a trivial
e = C^
cofibration, then Z ∈ ⊥ dg- F
∩ W, hence exact, and we deduce an exact sequence
0
0
0
0 → Q X → Q Y → Q Z → 0 with Q0 Z ∈ C ∩ W. This shows that (3.2) is a Quillen
adjunction, and for (3.3) the proof is analogous.
Next we prove that (3.2) is a Quillen equivalence. In the one direction, consider
e ∩ dw- C ∩ W, ?, dg- F),
e that is, X ∈ C
e ∩ dg- F
e ∩ dw- C ∩ W. Since
a bifibrant X in (C
e
e
e
e
dg- F∩Acyc(A ) = F, we have X ∈ C∩F∩dw- C∩W, and in particular Q0 X ∈ F is fibrant.
Hence, to show the derived unit is an equivalence, it suffices to show that such an X,
e ∩ dw- C ∩ W, ?, dg- F).
e
the underived unit ε : X → ι0 Q0 (X) is a weak equivalence in (C
We have that ε is an epimorphism and ker(ε) ∼
= τ≤0 X ⊕ σ>0 X, and we consider the
e which is
two summands separately. The first summand τ≤0 X belongs to Ch− (A ) ∩ F
⊥
e ∩ dw- C ∩
contained in [dw- C ∩ W] by Proposition 2.3(iv), hence trivially fibrant in (C
+
e
W, ?, dg- F). The second summand σ>0 (X) belongs to Ch (W ∩ F) which is contained in
e⊥ by Proposition 2.3(iii), hence trivially fibrant, too. It follows that ε : X → ι0 Q0 (X)
C
e ∩ dw- C ∩ W, ?, dg- F).
e
is indeed weak equivalence in (C
We have just proved that the derived unit id ⇒ R ι0 ◦ L Q0 is an equivalence, which
means that L Q0 is fully faithful. To prove that L Q0 ⊣ R ι0 is an equivalence, it is
therefore enough to show that L Q0 is also essentially surjective. For this, it suffices to
check that any bifibrant M ∈ C ∩ F occurs as the 0-th syzygy Q0 X of some bifibrant
e∩F
e ∩ dw- C ∩ W. Such a resolution can be built inductively
“complete resolution” X ∈ C
using the completeness of the cotorsion pairs (C ∩ W, F) and (C, W ∩ F).
The proof that (3.3) is a Quillen equivalence is analogous.
Corollary 3.3. Any hereditary and cofibrantly generated abelian model structure on a
Grothendieck category is Quillen equivalent to an abelian model structure on Ch(A ).
Next we study the model structures induced by the dashed arrows in Figure 1.
Proposition 3.4. Let M = (C, W, F) be a cofibrantly generated, hereditary abelian model
structure on the Grothendieck category A . Then there is a square of cofibrantly generated
abelian model structures on Ch(A ) and identity Quillen equivalences:
(3.4)
L
e ∩ dw- C ∩ W, Acyc(A ), (C
e ∩ dw- C ∩ W)⊥ )
e
(dg- C
(dg- C^
∩ W, Acyc(A ), dg- F)
R
L
R
R
e ∩ dw- W ∩ F), Acyc(A ), dg- F
e ∩ dw- W ∩ F)
(⊥ (F
L
R
L
e Acyc(A ), dg- W
^
(dg- C,
∩ F)
Their homotopy categories are equivalent to the ordinary derived category D(A ).
Proof. Applying Gillespie’s Theorem 1.8 to the dashed arrows in Figure 1 gives four
model structures matching the triples listed in (3.4) in the left and right hand parts; it
therefore suffices to check that their classes of weakly trivial objects all coincide with the
class Acyc(A ) of acyclic complexes.
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
13
e → (dg- C^
e
For the model structures associated to the arrows (C^
∩ W, dg- F)
∩ W, F)
e dg- W
e dg- W
^
^
and (C,
∩ F) → (dg- C,
∩ F) we already know this from Corollary 2.5 above.
e
e
Next, consider model structure associated to (C∩dwC∩W, ⊥) → (dg- C∩dwC∩W, ⊥):
By (1.1) the weakly trivial objects in the associated model structure are those X ∈
Ch(A ) which admit a short exact sequence of the form 0 → F → C → X → 0 with
h
i⊥
e ∩ dw- C ∩ W
e and C ∈ C
e ∩ dw- C ∩ W. Note that F, C ∈ Acyc(A ),
F ∈ dg- C
⊂ F
so the existence of such a sequence implies that X ∈ Acyc(A ). Conversely, suppose
X ∈ Acyc(A ) and pick an approximation sequence 0 → F → C → X → 0 for the
h
i⊥
e ∩ dw- C ∩ W, dg- C
e ∩ dw- C ∩ W ). Then again F ∈ Acyc(A ), and
cotorsion pair (dg- C
e ∩ Acyc(A ) ∩ dw- C ∩ W = C
e ∩ dw- C ∩ W.
also X ∈ Acyc(A ) by assumption, so C ∈ dg- C
The proof that the weak equivalences in the model structure associated to the arrow
e ∩ dw- W ∩ F) → (⊥, F
e ∩ dw- W ∩ F) in Figure 1 is analogous.
(⊥, dg- F
Finally, we study the relation between the model structures induced by the tilted
squares in Figure 1, beginning with some observations that hold in general:
β
α
Proposition 3.5. For localization contexts (D′′ , E′′ ) −
→ (D′ , E′ ) −
→ (D, E) their induced
model structures are related via a localization sequence of triangulated categories:
(3.5)
Loc(α)
(D′ , ?, E′′ )
L id
R id
Loc(β ◦ α)
L id
R id
(D, ?, E′′ )
Loc(β)
(D, ?, E′ )
∼ C′ ∩
Proof. For ω the common core of the given cotorsion pairs, we have Ho Loc(α) =
E′′ /ω and Ho Loc(β◦α) ∼
= C∩E′′ /ω, and L id : Ho Loc(α) → Ho Loc(β◦α) is the canonical
′
′′
functor C ∩ E /ω → C ∩ E′′ /ω, hence fully faithful. Similarly, R id : Ho Loc(β) →
Ho Loc(β ◦ α) is given by the embedding of D ∩ E′ /ω into D ∩ E′′ /ω, hence fully faithful.
It remains to prove the exactness of (3.5). Up to isomorphism in Ho Loc(β ◦ α),
L id
ker[Ho Loc(β ◦ α) −−→ Ho Loc(β)] consists of those D ∈ D admitting a short exact
sequence 0 → E → D′ → D → 0 with E ∈ E, D′ ∈ D′ (recall that these characterize
the weakly trivial objects in Loc(β)). As E is the class of trivially fibrant objects of
Loc(β ◦ α), such a sequence already implies D ∼
= D′ in Ho Loc(β ◦ α). Hence, up to
L id
isomorphism in Ho Loc(β ◦ α), ker[Ho Loc(β ◦ α) −−→ Ho Loc(β)] consists of the objects
L id
of D′ , and the same is true for im[Ho Loc(α) −−→ Ho Loc(β ◦ α)] by definition of L id.
Example 3.6. Proposition 3.5 applies to the chain of localization contexts
(3.6)
(3.7)
e dg- W
e W
^
^
(⊥, dw- W ∩ F) → (C,
∩ F) → (dg- C,
∩ F),
e → (dg- C^
e → (dw- C ∩ W, ⊥)
(C^
∩ W, dg- F)
∩ W, F)
in the upper right resp. lower left corner of Figure 1. The model structures
(3.8)
e ?, dw- W ∩ F) and (dw- C ∩ W, ?, dg- F)
e
(dg- C,
associated with the composed localization contexts are generalizations of both the injective/projective models for D(A ) and the contraderived/coderived model structures:
14
HANNO BECKER
e ?, dw- W ∩ F) = Mco (A )
– Choosing (C, W, F) as (A , A , I(A )), we have (dg- C,
inj
e
and (dw- C ∩ W, ?, dg- F) = M (A ).
– Choosing (C, W, F) as (P(A ), A , A ) in case A has enough projectives, we have
e ?, dw- W ∩ F) = Mproj (A ) and (dw- C ∩ W, dg- F)
e = Mctr (A ).
(dg- C,
Further, in the case of (A , A , I(A )) the localization sequences induced by (3.6) is the
known one D(A ) ⇄ K(I(A )) ⇄ Kac (I(A )), while the one induced by (3.7) is the trivial
localization sequence 0 ⇄ D(A ) ⇄ D(A ). Similarly, in the case of (P(A ), A , A ), the
localization sequence associated with (3.6) is trivial, while the one associated with (3.7)
is the classical one Kac (P(A )) ⇄ K(P(A )) ⇄ D(A ).
♦
Fact 3.7. Suppose given a square of localization contexts
(D′′′ , E′′′ )
β
(D′ , E′ )
α
(D′′ , E′′ )
ε
γ
δ
(D, E)
Then their localizations fit into a diagram of identity Quillen adjunctions
(D′ , ?, E′′′ )
(D, ?, E′ )
L
(3.9)
L
R
R
R
(D, ?, E′′′ )
L
L
R
R
R
L
L
(D, ?, E′′ )
(D′′ , ?, E′′′ )
Proof of Theorem B. According to Fact 3.7, each oriented square in Figure 1 gives rise to
a diagram of the form (3.9) in which the upper and lower rows are exact by Proposition
3.5. Applying this to the two tilted squares in Figure (1) gives rise to the diagram of
model structures depicted in Figure 2.
The left vertical Quillen adjunctions are Quillen equivalences by Theorem 3.1; in fact,
they share their classes of bifibrant objects, as do the two model structures connected by
the middle vertical adjunction, which is therefore a Quillen equivalence, too. Moreover,
this argument also proves one half of the wing-commutativity condition (⊲⊳.ii) from Definition 1.6 of butterflies. Similarly, the right vertical adjunctions are Quillen equivalences
by Proposition 3.4; in fact, they share their classes of weakly trivial objects, proving the
remaining part of condition (⊲⊳.ii).
The remaining properties are established by Proposition 3.5, so the proof is finished.
4. Proof of Theorem C
Denote T the common homotopy category of the two middle model structures in
Figure 2; explicitly, this is the homotopy category of complexes with components in
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
15
e and dg- F.
e Passing to homotopy categories in
ω = C ∩ W ∩ F which belong to both dg- C
Figure 2 now yields a recollement
Ho(M)
T
D(A )
and we claim that the induced stabilization functor real : D(A ) → Ho(M) makes the
following diagram commutative:
Ho(M)
real
D(A )
A
We begin with some generalities on stabilization functors associated to recollements:
Definition 4.1 (see [Kra05, §5]). Given a recollement T ′ T T ′′ of triangulated categories, the functors Iρ Qλ : T ′′ → T ′ resp. Iλ Qρ : T ′′ → T ′ are called the left resp. right
stabilization functors associated to the recollement.
Fact 4.2. For any X ′′ ∈ T ′′ there is a non-canonical isomorphism
Iλ Qρ X ′′ ∼
= ΣIρ Qλ X ′′ .
Proof. For X ∈ T the localization sequence T ′ ⇄ T ⇄ T ′′ induces a non-canonical
distinguished triangle Qλ QX → X → IIλ X → ΣQλ QX, which for X = Qρ X ′′ with
X ′′ ∈ T ′′ transforms into Qλ X ′′ → Qρ X ′′ → IIλ Qρ X ′′ → ΣQλ X ′′ . Applying Iρ from
the left annihilates Qρ X and hence yields an isomorphism Iλ Qρ X ′′ ∼
= ΣIρ Qλ X ′′ .
Generalizing the inclusions Acyc− ⊂ Wctr ⊂ Acyc ⊃ Wco ⊃ Acyc+ , we have:
Fact 4.3. The model structures (3.8) have the following properties:
e ?, dw- W ∩ F)
(1) The class W of weakly trivial objects in the model structure (dg- C,
+
satisfies Acyc (A ) ⊆ W ⊆ Acyc(A ).
e
(2) The class W of weakly trivial objects in the model structure (dw- C ∩ W, ?, dg- F)
−
satisfies Acyc (A ) ⊆ W ⊆ Acyc(A ).
e ?, dw- W ∩ F) arises as the localization of
Proof. Recall that the model structure (dg- C,
the composed localization context in (3.6). Therefore, by Gillespie’s Theorem 1.8, its
class W of weakly trivial complexes consists of those X ∈ Ch(A ) which admit a short
^
exact sequence 0 → X → F → C → 0 with F ∈ W
∩ F and C ∈ ⊥ [dw- W ∩ F]. Now
^
W
∩ F ⊆ Acyc(A ) by definition, and ⊥ [dw- W ∩ F] ⊆ Acyc(A ) as witnessed by the
upper right dashdotted arrow in Figure 1, so W ⊆ Acyc(A ) by the 2-out-of-3 property
of Acyc(A ). Conversely, suppose X ∈ Acyc+ (A ). Applying [Bec14, Resolution Lemma
1.4.4] to the cotorsion pair (C, W ∩ F) and all short exact sequences 0 → Zn X → X n →
Zn+1 X → 0 we can construct a short exact sequence 0 → X → F → C → 0 in Ch(A )
e ∩ Ch+ (A ). Since C
e ∩ Ch+ (A ) ⊆ ⊥ [dw- W ∩ F], it
^
with F ∈ W
∩ F ∩ Ch+ (A ) and C ∈ C
follows that X ∈ W. This finishes the proof of statement (i), and (ii) is analogous.
16
HANNO BECKER
Proof of Theorem C. Step 1: Firstly, we note that the derived functor
(4.1)
e → Ho(dg- C
e ∩ dw- C ∩ W, ?, dg- F)
e
Qλ : Ho(dg- C^
∩ W, Acyc(A ), dg- F)
may be computed naively on Ch− (A ): Namely, it can be computed through any resolution by a quasi-isomorphic bounded above complex with entries in C ∩ W, and by Fact
e hence a fortiori
4.3 such a resolution is still a weak equivalence in (dw- C ∩ W, ?, dg- F),
e
e
also in (dg- C ∩ dw- C ∩ W, ?, dg- F).
Step 2: We claim that the stabilization functor in question annihilates all X ∈ W.
For that, step 1 and the exactness (Proposition 3.5) of the sequence of functors
e ∩ dw- C ∩ W, Acyc(A ), [C
e ∩ dw- C ∩ W]⊥ )
Ho(dg- C
R id
e ∩ dw- C ∩ W, ?, dg- F)
e
Ho(dg- C
R id
e ∩ dw- C ∩ W, ?, F)
e
Ho(C
e ∩ dw- C ∩ W]⊥
show that it suffices to prove that any X ∈ W ⊂ Ch(A ) belongs to [C
e
e
up to weak equivalence in (dg- C ∩ dw- C ∩ W, ?, dg- F). Now, the presence of enough
injectives with respect to (C ∩ W, F) shows that X admits a resolution ι : X F with
F ∈ Ch≥0 (F) such that Zk F ∈ C ∩ W for k > 0. The thickness of W then implies
^
that even F ∈ Ch≥0 (W ∩ F) ⊂ dg- W
∩ F, and moreover Z := coker(ι) ∈ C^
∩ W. Since
e
e
^
^
C ∩ W are the trivially cofibrant objects in (dg- C ∩ dw- C ∩ W, ?, dg- F) and dg- W
∩F ⊆
⊥
e ∩ dw- C ∩ W] as witnessed by the right vertical arrows in Figure 2, it follows that
[C
e ∩ dw- C ∩ W]⊥ in Ho(dg- C
e ∩ dw- C ∩ W, ?, dg- F)
e as claimed.
X∼
= F ∈ [C
Step 3: Since any X ∈ A admits a functorial resolution of the form 0 → X → F →
C → 0 with F ∈ F and C ∈ C ∩ W, the second step shows that it suffices to prove the
commutativity of (C) when restricted to the fibrant objects F ⊆ A . In this case, by
definition as well as step 1, both the left stabilization
e −→ Ho(C
e ∩ dw- C ∩ W, ?, dg- F)
e
Ho(dg- C^
∩ W, Acyc(A ), dg- F)
and the functor
e ∩ dw- C ∩ W, ?, dg- F)
e
R ι0 : Ho(M) → Ho(C
can be computed naively, and the commutativity of (C) follows.
Dually, the composition of the right stabilization functor associated to the upper
butterfly in Figure 2 and the equivalence R Z0 from 3.3 in Theorem 3.1 gives another
functor D(A ) → Ho(M) making (C) commutative. Comparing the two functors D(A ) →
Ho(M) obtained this way, Fact 4.2 and L Q0 ∼
= Σ R Z0 show that they are pointwise noncanonically isomorphic; to show that they are even naturally isomorphic we’d need to
find an enhancement of Fact 4.2.
A REALIZATION FUNCTOR FOR ABELIAN MODEL CATEGORIES
17
References
[Bec14] Hanno Becker. Models for singularity categories. Adv. Math., 254:187–232, 2014.
[EJ11] Edgar E. Enochs and Overtoun M. G. Jenda. Relative homological algebra. Volume 1, volume 30
of De Gruyter Expositions in Mathematics. Walter de Gruyter GmbH & Co. KG, Berlin, 2011.
[Gil04] James Gillespie. The flat model structure on Ch(R). Trans. Amer. Math. Soc., 356(8):3369–
3390, 2004.
[Gil06] James Gillespie. The flat model structure on complexes of sheaves. Trans. Amer. Math. Soc.,
358(7):2855–2874, 2006.
[Gil14] James Gillespie. How to construct a hovey triple from two cotorsion pairs, 2014, 1406.2619.
[Hov02] Mark Hovey. Cotorsion pairs, model category structures, and representation theory. Math. Z.,
241(3), 2002.
[Hov07] Mark Hovey. Cotorsion pairs and model categories. In Interactions between homotopy theory
and algebra, volume 436 of Contemp. Math. Amer. Math. Soc., Providence, RI, 2007, 0701161.
[Kra05] Henning Krause. The stable derived category of a Noetherian scheme. Compos. Math., 141(5),
2005.
[Mur07] Daniel Murfet. The mock homotopy category of projectives and Grothendieck duality. PhD
Thesis, available online at http://www.therisingsea.org/thesis.pdf, 2007.
[Šťo10] Jan Šťovı́ček. Deconstructibility and the Hill lemma in Grothendieck categories, 2010,
1005.3251.
E-mail address: [email protected]
| 0math.AC
|
Logical Methods in Computer Science
Vol. 5 (4:1) 2009, pp. 1–25
www.lmcs-online.org
Submitted
Published
Jan. 9, 2009
Nov. 2, 2009
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
SYLVAIN LEBRESNE
Purdue University
e-mail address: [email protected]
Abstract. We present an extension of System F with call-by-name exceptions. The
type system is enriched with two syntactic constructs: a union type for programs whose
execution may raise an exception at top level, and a corruption type for programs that
may raise an exception in any evaluation context (not necessarily at top level). We present
the syntax and reduction rules of the system, as well as its typing and subtyping rules.
We then study its properties, such as confluence. Finally, we construct a realizability
model using orthogonality techniques, from which we deduce that well-typed programs
are weakly normalizing and that the ones who have the type of natural numbers really
compute a natural number, without raising exceptions.
1. Introduction
Exceptions are a convenient mechanism for handling errors in programming languages.
Most modern languages use them: Java, ML, C++, . . . . The main computational features
of exceptions are:
(1) One can raise an exception instead of any other expression (or instruction);
(2) It propagates automatically by default;
(3) One can catch it only when they need to.
Exceptions have long been confined to call-by-value languages and are usually presented
as a mechanism which “cuts through” the normal control flow of a program when raised.
This is viewing the raising of an exception as an effect of the calculus. Unfortunately, this
view makes exceptions hard to transpose to call-by-name calculi since those do not cope
well with effects.
This is a well-known problem. While in call-by-value the effect of a term t u can be
simply predicted based solely on the effects of t and u, in call-by-name it also depends on
the actual term t. Indeed, in call-by-name u may well not be evaluated thus not producing
its effect (or evaluated many times, producing the effect many times). For exceptions, this
means that in call-by-name, the fact that u raises an exception does not necessarily imply
that t u will. Hence in call-by-name, as summarized by S. Peyton Jones et al. [14], “(. . . )
the only productive way to think about an expression is to consider the value it computes,
not the way in which the value is computed”. Based on this observation, they proposed the
1998 ACM Subject Classification: D3.1, F4.1.
Key words and phrases: Exceptions, Call-By-Name, System F, Type System, Realizability.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
c
DOI:10.2168/LMCS-5 (4:1) 2009
CC
S. Lebresne
Creative Commons
2
S. LEBRESNE
idea of exceptions-as-values: a value is either a “normal” value, or an “exceptional” one. In
their framework, exceptions are not effects anymore. And while they present this idea in
the context of the Haskell programming language, this is a very general idea for exceptions
in call-by-name calculi.
From a typing perspective, exceptions are no simple beasts. Indeed, the type system
should allow the use of exceptions in any part of a program. What should then be the type
of the operation for raising an exception1? A solution, used in ML for instance, is to allow
the operation of raising an exception to have any type. In [14], S. Peyton Jones et al. chose
a similar solution, making exceptional values inhabitants of all types. While simple, this
solution comes at a price, the loss of type safety with respect to exceptions. The type of an
expression never ensures that no exceptions can be raised during evaluation.
If we want the type of an expression to reflect which exception it may raise, a more precise typing is in order. For call-by-value languages, since exceptions are effects, a convenient
and efficient solution is to add to the type system an effect system [13, 7]. Unfortunately
and unsurprisingly, this solution is unadapted to the typing of call-by-name exceptional
values. Indeed, tracking values with types is much more difficult than tracking effects.
The call-by-name evaluation is well represented amongst type theoretical calculi which
are at the core of many proof assistants (coq [4], lego [15], . . . ). We believe this reinforce
the case for studying exceptions in call-by-name and their precise typing. Indeed, the
solution of having exceptional values inhabiting all types would be inconsistent in these
settings.
This paper proposes a type system for exceptions in call-by-name calculi. By introducing the new notion of corruption, this type system is able to track which exceptions may
escape from a term during evaluation. By using subtyping, this notion is able to cope with
the automatic propagation of exceptions and to respect the modularity of typing. This type
system is presented in the context of an extension of System F with exceptions. System F
is used here as a first step towards more elaborate type theoretical frameworks.
Meta-theoretical properties of the resulting calculus are proved in particular by exhibiting a realizability model. While parts of the proofs are given in this paper, more detailed
proofs of the results presented2 are available in the Ph.D thesis of the author [8].
The remaining of the paper is organized as follows. We explain our design in Section 2:
we justify the kind of exception-as-values we use and describe the three levels of corruption
our type system distinguishes. We formally present our calculus in Section 3 and state the
properties it enjoys and Section 4 provides some examples. Then in Section 5 we design
a realizability model of our calculus that gives some insight on the meaning of corruption
and we prove its soundness. Finally, we present in Section 6 some related works before
concluding in Section 7 with future works.
2. Design of the system
2.1. Which exceptions-as-values? As stressed above, exceptions in call-by-name calculi
should be values. But there are essentially two designs for exceptions as values: either we
1Remark that imperative language alleviate this problem by making the operation of raising an exception
an instruction and not an expression.
2A notable difference is the presence of lists in the language described in this document, while in [8] the
data type of lists is only presented for a first-order typed version of the language.
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
3
encode them explicitly in the language, or we make them primitives. The first option is a
well-known one and let us first present its drawbacks in order to justify the need for the
primitive solution.
Encoding explicitly exceptions is an old idea [21, 16]: to each type A is associated a type
Maybe A which is either values of A tagged as correct values or exceptional values (this idea
is nicely explained, for the Haskell programming language, in [14]). It has later been realized
that the Maybe type constructor forms a monad [11, 22]. And P. Wadler and P. Thiemann
proposed in [23] to add effects to monads, allowing for the detection of uncaught exceptions
in such monadic encoding. However this approach has some drawbacks, namely:
• Terms using exceptions are crippled by extra clutter. For example, in Haskell, to apply a
function f :: Int -> Int to a value x :: Maybe Int we are forced to write:
do a <- x
return (f a)
Using exceptions is not as transparent for the programmer as it is in call-by-value languages;
• As remarked in [14], modularity and code re-use are compromised, especially for higher
order functions. Consider the following sorting function:
sort :: (a -> a -> Bool) -> [a] -> [a]
This function cannot be applied to a comparison function that may raise exceptions such
as:
cmp :: a -> a -> Maybe Bool
Indeed, with monads, we need to know where the sort function uses the comparison in
order to add the monad’s operations;
• Monads force the evaluation of arguments (in the example above, the evaluation of x
is forced before the application to f). One could not see that as an inconvenience, and
this is indeed desirable for most uses of monads (IO, states, . . . ). Nonetheless, this is
a constraint and it makes exceptions not usable in non monadic call-by-name code. We
think that this can be avoided for exceptions.
This leads us to the second design choice: making exceptions primitives. This has been
first proposed by S. Peyton Jones et al. [14] with imprecise exceptions. The idea is that
a value of any type is either a “normal” value, or an “exceptional” one. The resulting
mechanism allows exceptions to be used in place of any other term (as for more traditional
“call-by-value” exceptions and contrarily to monadic ones). Note that since values may
be exceptional, we can have for instance, a list, which is fully defined but for which some
elements are exceptional values (see Section 4). These exceptions are raised only when (and
if) the list is evaluated. A main difference with the call-by-value mechanism of exceptions
is for example that a term like (λx. 0) (raise ε) (where 0 is simply the constant zero and ε
some exception) will reduce to 0 and not to raise ε.
Our system, named Fx, adapts this idea to System F, adding it two new term constructions: raise and try. But while the exceptions of [14] are not precisely typed (the
raising operation is in all types), we propose a type system where the type of an expression
indicates which exceptions the expression may raise.
4
S. LEBRESNE
2.2. Expected properties. The type system we will present enjoys the following properties:
• If a term can raise an exception, its type indicates it. In particular, programs of type N
are not able to raise exceptions;
• Programmers can use a term raise ε in place of any other term. In particular, raise ε
type as a function;
• Exceptions and their typing discipline do not jeopardize modularity and code re-use.
A function defined without exceptions in mind still accepts exceptional arguments and
behave in a sensible way. Moreover, this is done without knowing the actual code of the
function.
2.3. Three levels of corruption. We call corrupted, a term that may mention exceptions. Given a type A (say the type N of natural numbers), we distinguish three levels of
corruptions for the terms related with this type:
• Terms of A. They are not corrupted, either they do not mention exceptions or the ones
they mention are caught or erased during reduction;
⋆ {ε}. They are terms of A or terms that reduce to the exception ε, i.e. reduce
• Terms of A ∪
to raise ε (we then say that they raise ε).
• Terms of A{ε} . They are terms of A that may mention the exception ε but do not
necessarily reduce to it (for instance, if S is the successor function, S (raise ε) has type
⋆ {ε} since it has not type N nor does it reduce to raise ε).
N{ε} , but not type N ∪
Moreover, to handle the properties of corruption, we use a subtyping relation. And in
⋆ {ε} ≤ A{ε} .
particular we have the subtyping: A ≤ A ∪
The following section explains why the need to distinguish at least those three levels.
But one might wonder why we do not distinguish more levels. Like the terms containing
exceptions but not at top level. Or terms having an exception at a depth of at most 2 (like
raise ε or λx. raise ε but not λx. λy. raise ε), etc. As of now, while such more precise
notion may well be sound, we have not study them. The main reason is that they would
complicate and clutter the type system while we are not convinced they would prove useful
in practice.
⋆ {ε} is really
2.4. Why we need to distinguish these three levels. The construction A ∪
needed because of the typing of the try operation, since for a try to catch an exception in
its body, this body has to reduce to the exception.
But because we do not want to change the typing rule of application, the construction
⋆ {ε} clearly does not fulfill all our needs. Firstly, we cannot use it to type S (raise ε).
A∪
⋆ {ε}.
Secondly, given a function M of type A → B, we cannot apply it to a term N of type A ∪
⋆
Indeed, M N is generally not of type B ∪ {ε} (note however that it would be the case in a
call-by-value calculus). Consider for instance M = λx. λy. x (of type A → (C → A)) and
⋆ {ε} (since
N = raise ε, then M N reduces to λy. raise ε which is not of type (C → A) ∪
it is neither a function of type C → A nor the exception ε)3.
3Note that this is typically this example, of the typing of a term like (λx. λy. x) N when N may raises
an exception, that makes effect system [13] unsuited to call-by-name exceptions.
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
5
To solve these problems, we use a second type construction, the corruption of a type
A by an exception of name ε, denoted A{ε} . The main property the corruption enjoys is a
good behavior with respect to arrow types:
.
(A → B){ε} = A{ε} → B {ε}
This subtyping equality4 may seem paradoxical with the usual subtyping rule of arrow
(contra-variance to the left, co-variance to the right). This is however justified by the
realizability model of Section 5.
Intuitively, terms of type A{ε} should be seen as terms of type A where some sub-terms
may have been replaced by raise ε (hence, programmers can use raise ε wherever they
⋆ {ε} are
want, which, in turns, corrupts the resulting type). Equivalently, while terms of A ∪
{ε}
terms that may reduce to raise ε at top-level, terms of A
are the ones that may reduce
to raise ε in any evaluation context.
Now, with corruption, we can apply a function f : A → B to a potentially exceptional
term. Indeed, we have that
.
⋆ {ε} ≤ (A → B){ε} = A{ε} → B {ε} .
A → B ≤ (A → B) ∪
Remark that since we use subtyping, there is no need to actually know the term f . This
allows for modularity: to type the application of some (external) function f to a term u,
it is enough to know the type of f , and this even when u may raise exceptions but the
exported type of f does not mention exceptions. This is in particular convenient for primitive functions like the successor function S, allowing to type-check S (raise ε) with the
type N{ε} without the need to give S a complicated type (the type of S is simply N → N).
2.5. Exceptions by the millions. While we have only used one exception names ε in the
above section, it is useful to be able to handle more than one exception at a time. To that
⋆ ∆ and A∆ where ∆ is a set of exception names.
end, the general type constructions are A ∪
Using sets of exceptions requires some type identification using the following subtyping
rules:
.
⋆ ∆′ ) ∪
⋆ ∆ = A∪
⋆ (∆∪∆′ )
(A ∪
′
′
.
∆
∆
=
A(∆∪∆ )
(A )
.
⋆ ∅
A∪
=
A
.
∅
=
A
A
.
4The subtyping equality A =
B is simply defined as shorthand for A ≤ B and B ≤ A.
6
S. LEBRESNE
3. Formal presentation
We present the Fx calculus, an extension of System F with typed exceptions, natural numbers and lists.
3.1. Syntax, reductions and associated properties.
3.1.1. Syntax of terms. We consider a countable set E of names of exceptions and a distinguished set of variables V.
Definition 3.1 (Terms). A term of Fx is a term generated by the following grammar:
M, N
::= x | λx. M | M N
| raise ε | try M with ε 7→ N
| 0 | S | rec | [ ] | cons | list rec
In this definition, variables are ranged over by x, y, . . . while exception names are ranged
over by ε, ε′ , . . .. Notions of free and bound variables are defined as usual, as well as
the external operation of substitution (written M {x := N }). The set of all closed terms
is denoted T and terms are considered up to α-equivalence. Note that the construction
try M with ε 7→ N does not bind the occurrences of ε. The term raise ε is called an
exception, ε being its name, but, as an abuse of terminology, we also call ε an exception. In
the term try M with ε 7→ N we will sometimes call M the body and N the handler of the
try construction.
To the terms of the lambda calculus, we add the constructions to raise and catch
exceptions as well as two usual structured data types: the natural numbers and the lists.
3.1.2. Computation in Fx.
Definition 3.2 (Regular values). A regular value is a (closed) term of Fx having one of
the following form:
RV
::= λx. M | 0 | S | S N | rec | rec M | rec M N
| [ ] | cons | cons M | cons M N | list rec | list rec M | list rec M N
Note that S N is a regular value for any term N and hence S (raise ε) is a regular
value as well.
Definition 3.3 (Values). A values is a (closed) term of Fx having one of the following
form:
V ::= RV | raise ε
where RV is a regular value and ε any exception name.
For well-typed term, a value corresponds to a weak head normal form.
Definition 3.4 (Computation). The notion of reduction > for the calculus is defined by the
rules of Figure 1. Computation in Fx is defined from the notion of reduction by the relation
of reduction ≻ whose rules are given in Figure 2. We note ≻∗ the reflexive and transitive
closure of ≻ and we note = its reflexive, transitive and symmetric closure. Moreover, if
M = N , we will say that M is equivalent to N .
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
(λx. M ) N
(raise ε) M
try (raise ε) with ε 7→ N
try (raise ε′ ) with ε 7→ N
try V with ε 7→ N
rec X Y 0
rec X Y (S N )
rec X Y (raise ε)
list rec X Y [ ]
list rec X Y (cons E L)
list rec X Y (raise ε)
7
> M {x := N }
> raise ε
> N
> raise ε′
> V
(if ε 6= ε′ )
(if V is a regular value)
> X
> Y N (rec X Y N )
> raise ε
> X
> Y E L (list rec X Y L)
> raise ε
Figure 1: Notion of reduction for Fx
M
M
M
λx. M
> M′
≻ M′
≻ M′
≻ λx. M ′
M
M N
N
M N
≻ M′
≻ M′ N
≻ N′
≻ M N′
M
try M with ε 7→ N
N
try M with ε 7→ N
≻ M′
≻ try M ′ with ε 7→ N
≻ N′
≻ try M with ε 7→ N ′
Figure 2: Relation of reduction for Fx
Note that, as usual, the scope of capture of the try construction is dynamic: in the
term (λx. try x with ε 7→ 0) (raise ε), the exception is caught during reduction and the
whole term reduces to 0. We say that a term M raises the exception ε if M ≻∗ raise ε
(that is, if M reduces to the exception named ε).
Definition 3.5 (to have a value). We will say that a term M has a value if and only if it
reduces to a value, that is if there exists a value V such that M ≻∗ V .
It can be proved [8] that this notion is equivalent to the one of having a normal form
for the weak head reduction of the calculus.
We now show that adding raise and try does not break the confluence of the calculus:
Theorem 3.6 (Confluence). If M , N and N ′ are terms such that M ≻∗ N and M ≻∗ N ′ ,
then there exists a term P such that N ≻∗ P and N ′ ≻∗ P.
Proof. We adapt the proof originated by Tait and Martin-Löf for the confluence of pure
lambda-calculus that can be found in [1] for example. We define the notion of parallel
reduction ≫ for Fx, we show that it satisfies the diamond property and conclude since
≻∗ = ≫∗ . Proofs of these properties are easy to tackle inductions we leave to the interested
reader. We however give in appendix A the definition of the parallel reduction for Fx.
8
S. LEBRESNE
3.2. The type system. As stressed in Section 2.3, Fx uses a subtyping relation ≤. Thus,
Fx is in fact an extension of the second-order lambda calculus with subtyping introduced
by Mitchell [10, 24] (and we will call this calculus System Fη in the following). Note that
we will however use a presentation of this calculus that differs from the original one and
that can be found for example in [19].
Definition 3.7 (Types). The syntax of types for Fx is built upon the one of System F.
Type of Fx are generated by the following grammar:
⋆ ∆ | A∆
A, B ::= α | N | A list | A → B | ∀α. A | A ∪
⋆ ∆ and A∆ , ∆ is a finite set of exceptions names (∆ ⊆ E). Moreover, α stands
In A ∪
for a type variable taken from the set of type variables A. Notions of free and bound
type variable are defined as usual, as well as the external operation of substitution (written
A{α := B}). We denote by F V (A) the set of all the free type variables of the type A.
Types are considered up to α-equivalence. Precedences for the arrow construction and the
⋆ ∆ and A∆ being higher.
universal quantifier are the usual ones; the precedences of A ∪
Moreover, we will often write A list∆ for (A list)∆ .
3.2.1. Typing.
Definition 3.8 (Typing context). A typing context Γ is a finite set of declarations having
the form Γ ≡ x1 : A1 , . . . , xn : An where x1 , . . . , xn are pairwise distinct term variables and
where A1 , . . . , An are arbitrary types.
The set F V (Γ) of free variables of Γ denotes the union of the sets of free type variables
for the types used in Γ, that is to say:
[
F V (x1 : A1 , . . . , xn : An ) =
F V (Ai )
i ∈ {1...n}
Definition 3.9 (Typing). The type system of Fx is defined from the typing judgment
Γ⊢M :A
that reads ‘in the typing context Γ, the term M has type A’. This judgment is inductively
defined by the rules of Figure 3.
Remark that the typing rules from System Fη are unchanged, we simply add rules.
Also note that the usual typing rules for the recursion operators can be retrieved from (rec)
and (fold) by taking ∆ = ∅ (theses rules are in fact typing schemes).
3.2.2. Subtyping.
Definition 3.10 (Subtyping). The subtyping relation between two types A and B, written
A ≤ B, is inductively defined by the rules of Figure 4.
.
The equality A = B is defined as short for “A ≤ B and A ≥ B”. In the inference rules,
.
when the equality A = B appears as a premise, it figures for the two premises A ≤ B and
A ≥ B. And when it appears as a conclusion, it figures for two inference rules, one having
A ≤ B as a conclusion, the other one having A ≥ B.
The subtyping rules from Fη are unchanged. The rules (ex-noexc), (eq-uu) and (eqcc) dealt with sets of exceptions. The hierarchy of corruption (see 2.3) is implemented by
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
9
System Fη typing rules:
(x : A) ∈ Γ
Γ, x : A ⊢ M : B
(ax)
(abs)
Γ⊢x:A
Γ ⊢ λx. M : A → B
Γ ⊢ M : A α∈
/ F V (Γ)
(gen)
Γ ⊢ M : ∀α. A
Γ⊢M :A→B Γ⊢N :A
(app)
Γ⊢M N :B
Γ⊢M :A A ≤ B
(subs)
Γ⊢M :B
Natural numbers typing rules:
Γ⊢0:N
(zero)
Γ⊢S:N→N
(succ)
⋆ ∆ → (N∆ → α ∪
⋆ ∆ → α∪
⋆ ∆) → N∆ ∪
⋆ ∆′ → α ∪
⋆ (∆∪∆′ )
Γ ⊢ rec : ∀α. α ∪
(rec)
List typing rules:
(nil)
(cons)
Γ ⊢ cons : ∀α. α → α list → α list
(fold)
⋆ ∆ → (β ∆ → β list∆ → α ∪
⋆ ∆ → α∪
⋆ ∆)
Γ ⊢ list rec : ∀α. ∀β. α ∪
⋆ ∆′ → α ∪
⋆ (∆∪∆′ )
→ β list∆ ∪
Γ ⊢ [ ] : ∀α. α list
Exceptions handling typing rules:
⋆ {ε}
Γ ⊢ raise ε : ∀α. α ∪
(raise)
⋆ {ε} Γ ⊢ N : A
Γ ⊢ M : A∪
(try)
Γ ⊢ try M with ε 7→ N : A
Figure 3: Typing judgments
(ex-uni) and (ex-corrupt). The rules (ex-fallc) and (ex-fallu) are justified by the absence
of computational content of the universal quantification. Moreover, corruption and union
commutes (eq-uc).
The subtyping is stable by union (ex-ctx), but also by corruption (this is proved by
⋆ ∆ is
Theorem 3.11). Rule (ex-arru) simply says that, since a term M of type (A → B) ∪
either a term of type A → B or an exception of ∆, it can always be applied to a term of
type A, resulting in a term of type B (if M is a true function) or an exception of ∆ (if so
is M ).
As discussed in Section 2.4, the rule (eq-arrc) is the main rule of corruption and allows
exceptions to be used anywhere. Note that we really need an equality here on pain of losing
the subject-reduction property.
Finally, the list construction is monotonic (rule (ex-lcor)) and a list of corrupted elements is in particular a corrupted list (rule ex-lctx ).
The subtyping associated to the notion of corruption is a quite flexible one, especially
with respect to arrows. As noted in Section 2.4, it allows to derive
∆
∆
A1 → A2 → . . . → An ≤ A∆
1 → A2 → . . . → An
10
S. LEBRESNE
System Fη rules :
A ≤ A
(st-id)
A ≤ B
A ≤ B B ≤ C
(st-trans)
A ≤ C
α∈
/ F V (A)
A ≤ ∀α. B
(f-gen)
A′ ≤ A B ≤ B ′
(st-arrow)
A → B ≤ A′ → B ′
∀α. A ≤ A{α := B}
α∈
/ F V (A)
∀α. (A → B) ≤ A → ∀α. B
(f-inst)
(f-distr)
Exception related rules :
⋆ ∆
A ≤ A∪
(ex-uni)
⋆ ∆ ≤ A∆
A∪
A∅ ≤ A
A ≤ B
(ex-ctx)
⋆ ∆ ≤ B∪
⋆ ∆
A∪
∀α. A∆ ≤ (∀α. A)∆
A∆
list ≤ A
(ex-noexc)
⋆ ∆ ≤ A→B∪
⋆ ∆
(A → B) ∪
(ex-fallc)
list∆
(ex-corrupt)
⋆ ∆) ≤ (∀α. A) ∪
⋆ ∆
∀α. (A ∪
(ex-arru)
(ex-fallu)
A ≤ B
(ex-lctx)
A list ≤ B list
(ex-lcor)
Exception related equality rules :
(eq-uu)
.
⋆ (∆∪∆′ )
⋆ ∆) ∪
⋆ ∆′ = A ∪
(A ∪
(eq-uc)
′
′
.
⋆ ∆
⋆ ∆)∆ = A∆ ∪
(A ∪
(eq-cc)
′
′
.
(A∆ )∆ = A(∆∪∆ )
(eq-arrc)
.
(A → B)∆ = A∆ → B ∆
Figure 4: The subtyping relation
but also that A∆ → B ≤ A∆ → B ∆ or that A → B ∆ ≤ A∆ → B ∆ for instance5. However, what the subtyping of corruption forbids is the removable of corruption in covariant
position. That is, corruption allows the use of functions with exceptions they do not handle
themselves, but it then always ensure that the return type mentions those exceptions.
3.2.3. Typing the recursion operations. Fx uses natural numbers and lists. To work with
these data types, we have equipped the calculus with recursion operators (rec and list rec)
of Gdel’s System T[6]. However, in the presence of exceptions, the usual typing of these
5Proofs for all those relations follow the same pattern. Corruption is introduced on the right with (ex-
uni) and (ex-corrupt) and is then distributed over the operands of the arrow with (eq-arrc). Lastly, double
corruption ((A∆ )∆ ) is eliminated with (eq-cc) if needed.
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
11
operators is not precise enough. Indeed, consider the case of the natural numbers. The
usual typing rule of the recursion operator rec is
∀α. α → (N → α → α) → α → α,
and hence, using the corruption type and its associated subtyping rules, it also has the type
∀α. α∆ → (N∆ → α∆ → α∆ ) → α∆ → α∆
for any set of exception names ∆. However, this last type is not precise enough and for
instance, it does not precisely account for the reduction rule
rec X Y (raise ε) ≻ raise ε.
Dealing with this imprecision is the reason of the addition of the set ∆′ in the typing rule
of rec (rule (rec) of Figure 3). Moreover, the function eval which will be introduced in
section 4 reveals another imprecision. Given a corrupted natural number, this function
returns either a well formed natural number or an exception at top-level. But to give this
⋆ ∆, we need the addition of
function the type we want, that is to say the type N∆ → N ∪
the set ∆ in the typing rule of the recursion operator (rule (rec) of Figure 3). The typing
rule of the recursion operator list rec follows the same modifications.
3.3. Properties of typing. The subtyping relation is stable by corruption:
Theorem 3.11. If A and B are two types such that A ≤ B, then for any set of exception
names ∆, A∆ ≤ B ∆ .
Proof. We proceed by induction on the derivation of A ≤ B. All the cases are easily
resolved since corruption commutes with all type constructions. For example, taking the
⋆ ∆′ )∆ ≤ (A → B ∪
⋆ ∆′ )∆ . But
case of rule (ex-arru), we have to show that ((A → B) ∪
∆
∆
′
∆
∆
′
⋆ ∆ ≤ A → B ∪
⋆ ∆ and we conclude using the fact
using rule (ex-arru), (A → B ) ∪
.
.
⋆ ∆′ )∆ .
⋆ ∆′ )∆ and A∆ → B ∆ ∪
⋆ ∆′ = (A → B ∪
⋆ ∆′ = ((A → B) ∪
that (A∆ → B ∆ ) ∪
A few remarkable subtyping rules are also easily derivable from the ones of Figure 4:
Theorem 3.12. The following subtyping relations hold:
⋆ ∅ ≤ A
A∪
⋆
⋆ ∆)
(∀α. A) ∪ ∆ ≤ ∀α. (A ∪
∆
∆
(∀α. A)
≤ ∀α. (A )
⋆ ∅ ≤ A comes from rules (ex-corrupt) and (ex-noexc). The proofs for
Proof. Proof of A ∪
⋆ ∆ ≤ ∀α. (A ∪
⋆ ∆) and (∀α. A)∆ ≤ ∀α. (A∆ ) are similar. For instance, for the
(∀α. A) ∪
⋆ ∆. Then, using (ex-ctx)
former one, we use (f-inst) and (ex-uni) to show that ∀α. A ≤ A ∪
⋆
⋆
and (eq-uu), we show that (∀α. A) ∪ ∆ ≤ A ∪ ∆. And we conclude with (f-gen).
12
S. LEBRESNE
M ⊑∆ M
(c-id)
M ⊑∆ M ′
(c-lam)
λx. M ⊑∆ λx. t′
ε∈∆
(c-rai)
M ⊑∆ raise ε
M ⊑∆ M ′ N ⊑∆ N ′
(c-app)
M N ⊑∆ M ′ N ′
M ⊑∆ M ′ N ⊑∆ N ′
(c-try)
try M with ε 7→ N ⊑∆ try M ′ with ε 7→ N ′
Figure 5: Corruption relation
Definition 3.13. The corruption relation ⊑∆ between terms is inductively defined Figure 5.
To have M ⊑∆ N means that N is obtained from M by replacing some sub-terms in any
position by raise ε, ε belonging to ∆.
Thus, Theorem 3.14 formally states that, in term of programming, exceptions can be
used in any place, but with the added cost of corrupting the type.
Theorem 3.14 (corruption). If M and N are two terms, A a type and ∆ a set of exceptions
such that Γ ⊢ M : A and M ⊑∆ N , then Γ ⊢ N : A∆ .
Proof. This theorem is proved by induction on the statement M ⊑∆ N . The proof presents
no major difficulty as long as we first prove the three following “inversion” results :
(1) If M is a term, A a type and Γ a typing context such that
Γ ⊢ λx. M : A,
→
then there exists a set of type variable −
α∈
/ F V (Γ) and two terms B and C such that
−
→
∀ α . (B → C) ≤ A and Γ, x : B ⊢ M : C.
(2) If M and N are two terms, A a type and Γ a typing context such that
Γ⊢M N :A,
then there exists a term C such that Γ ⊢ M : C → A and Γ ⊢ N : C.
(3) If M and N are two terms, A is a type and Γ is a typing context such that
Γ ⊢ try M with ε 7→ N : A,
⋆ {ε} and Γ ⊢ N : A.
then Γ ⊢ M : A ∪
Proofs of these three results are straightforward inductions on the derivation of the initial
typing judgment.
4. Examples
A simple yet classical function on natural numbers which can raise an exception is the
predecessor function. In Fx, we can define:
pred ≡
rec (raise ε) (λx. λy. x)
:
⋆ {pred err }
N → N∪
≻∗
It has the expected reductions, i.e. pred 0
raise ε and pred (S N ) ≻∗ N . We can then
define a “safe” predecessor pred′ from pred which returns 0 when applied to 0:
pred′
≡
λn. try (pred n) with pred err 7→ 0
:
N→N
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
13
Having exceptions, it is possible to define the functions that return the head and the
tail of a list:
⋆ {hd fail }
hd : A list → A ∪
≡ list rec (raise hd fail) (λe. λl. λ . e)
⋆ {tl fail }
: A list → A list ∪
≡ list rec (raise tl fail ) (λe. λl. λ . l)
⋆ {div by 0 }) and the
We can also define the Euclidean division (div : N → N → N ∪
mapping of a function to a list (map : ∀α. ∀β. (α → β) → α list → β list). The type
system being modular, we can (and will) use these two functions without having to exhibit
a particular implementation. But using them allows us to define the following function that
2
maps the function n 7→ n−1
to a list:
tl
⋆ {div by 0 }) list
: N list → (N{pred err } ∪
≡ map (λn. div 10 (pred n))
f
Remark that the result is always a list but that can contain exceptional values. For instance,
f [2; 1; 5] computes the list [5; raise div by 0 ; 2] which does not reduces to raise div by 0 .
Again, exceptions are values that propagate only when used. Now we can get the first
element of the result of this function with:
⋆ {hd fail }
g : N list → N{pred err ,div by 0 } ∪
≡ λl. hd (f l)
We can apply g to some argument and catch the exception hd fail with a try, but we cannot
catch the two other exceptions since these ones are not necessarily at top-level. If we want
to catch them, we need a function that evaluates a natural number potentially corrupted
and returns either a well formed natural number or an exception. It is the purpose of the
following function:
eval
⋆ ∆
: N∆ → N ∪
≡ λn. (rec (λa. a) (λm. λr. λa. r (S a)) n) 0
With this function, we can now capture the exceptions that can appear in the result of the
function g above. That is what the following function does (where we use a straightforward
shortcut allowing the try to catch all the exceptions):
h
: N list → N
≡ λl. try (eval (g l))
with pred err , div by 0 , hd fail 7→ 0
Note that for instance h [2; 1; 5] will return 5 in our system since the part of the list that
would yield an exception (the second element after the mapping) is never used (we only use
the head of the list). In contrast, a similar function in say Caml would have yielded 0.
5. Realizability model
We will define a realizability model for Fx using techniques of orthogonality (see [12, 20] for
examples of use of such techniques). The choice of those orthogonality techniques is mainly
motivated by two reasons: we believe that it offers a nice way to handle second order and
it will come in handy for the definition of the interpretation of corruption, allowing a much
more simple definition than a “direct” model would allow. We start by introducing a few
definitions necessary to the construction of the model.
14
S. LEBRESNE
5.1. Daimon and contexts. We add a new and distinguished term, the daimon (denoted z) similar to the one of [5]. This term computationally behaves like an uncatchable
exception. We also introduce the new term construction M ; N . This construction tests
if M is the daimon and if so, return N . Otherwise, it does not reduce. The reduction
rules for these two additions are given Figure 6. Moreover, z is added to the definition of
value. Note that none of these constructs have typing rules and as such, they cannot be
used in well-typed terms. It can also easily be proved that they do not break the confluence
property of the language. In those respects, they are only convenient technical addition for
the model and should not be considered as inherent part of the language.
zN
try z with ε 7→ N
rec X Y z
list rec X Y z
z; N
>
>
>
>
>
z
z
z
z
N
Figure 6: Reduction rules for z and ;.
The daimon has two purposes in the model. First, it will inhabit all type interpretation,
property that will be used to show that all the terms of the interpretation are weakly
normalizing (see Lemma 5.13). Secondly, our model is a realizability one, types will be
interpreted by sets of terms. But the principle of our orthogonality model is to not define
those sets directly, but instead to first define the interpretation of types as sets of evaluation
contexts. Then, to each such set S of evaluation contexts is associated the set of all the
terms that “behave correctly” for all the contexts of S. This notion of a term M “behaving
correctly” in a context C is the orthogonality relation. For our model we chose it to be that
M put in the context C reduces to this distinguished term z. But to define formally this
orthogonality relation, let us first define formally the evaluation contexts we will use:
Definition 5.1 (Context). A context is a term with a hole (denoted by [ ]) defined by:
C ::= [ ] | C N | try C with ε 7→ z | rec M N C | list rec X Y C
The set of all contexts is noted C and the term obtained by filling the hole of a context C
with the term M is noted C[M ].
Note that our definition of context is more restrictive than the usual one (where a
context is any term with a hole). Actually, save the restriction in the handler of try to z
(which will allow for a simpler interpretation of corruption), our contexts are the evaluation
contexts of call-by-name evaluation.
Moreover, we will not care about the order of two adjacent try in a context. Since the
set of all exception names E is countable, we can fix a priori a bijection φ : E → N and
we define the following notation:
Notation 5.2. If ∆ is a (finite) set of exception names, then try [ ] with ∆ is a notation
for
• the context try (. . . (try [ ] with ε1 7→ z) . . .) with εn 7→ z if ∆ 6= ∅ and if ε1 , . . . εn are
the elements of ∆ arranged according to φ (that is, φ(ε1 ) < . . . < φ(εn )).
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
15
• the empty context otherwise (if ∆ = ∅).
Contexts have the following property :
Lemma 5.3. If C is a context and M is a term such that C[M ] has a value, then M has
a value.
Proof. By case on the form of the context C and by induction on the length of the reduction
of C[M ] to the value. No case raises specific difficulties.
5.2. Orthogonality relation.
Definition 5.4 (Orthogonality relation). If M is a term and C a context, then M ⊥ C
(and we say that M and C are orthogonal) if and only if C[M ] ≻∗ z.
Moreover, if S is a set of contexts, we define the set of terms S ⊥ by
S⊥
{ M | ∀ C ∈ S, M ⊥ C }
=
Note that as with any orthogonality relation, we can easily check that ⊥ verifies the following
properties:
Lemma 5.5. If S and T are two context sets such that S ⊆ T , then T ⊥ ⊆ S ⊥ .
Lemma 5.6. If I is any set and (Si )i ∈ I is a family of set of contexts indexed by i, then
S
T
( i ∈ I Si )⊥ = i ∈ I Si ⊥ .
5.3. Operations on sets. We recall the two standard definition of concatenation · (of a
set of terms and a set of contexts) and composition ◦ (of two sets of contexts):
A · S = { C [ ] N | C ∈ S, N ∈ A }
S ◦ T = { C[D[ ]] | C ∈ S, D ∈ T }
For instance,
{0, 1} · {try [ ] with ε 7→ z} = {try ([ ] 0) with ε 7→ z, try ([ ] 1) with ε 7→ z}
{try [ ] with ε 7→ z, [ ] (λx. x)} ◦ {[ ] 1} = {try ([ ] 1) with ε 7→ z, ([ ] 1) (λx. x)}
We then define two operations on sets of contexts:
↓∆ S
↑∆ S
=
=
S ◦ { try [ ] with ∆ }
{ try [ ] with ∆ } ◦ S
and thus, for instance,
C1 ≡ ↓∆ (rec 0 (λx. λy. y) [ ]) = { rec 0 (λx. λy. y) (try [ ] with ∆) }
C2 ≡ ↑∆ (rec 0 (λx. λy. y) [ ]) = { try (rec 0 (λx. λy. y) [ ]) with ∆ }
and if ε ∈ ∆,
C1 [raise ε]
C1 [S (raise ε)]
C2 [raise ε]
C2 [S (raise ε)]
=
=
=
=
rec 0 (λx. λy. y) (try raise ε with ∆)
rec 0 (λx. λy. y) (try S (raise ε) with ∆)
try (rec 0 (λx. λy. y) (raise ε)) with ∆
try (rec 0 (λx. λy. y) (S (raise ε))) with ∆
≻∗
≻∗
≻∗
≻∗
z
raise ε
z
z
16
S. LEBRESNE
It can be checked that by definition we have the following equalities:
↑∆ (A · S) =
↑∆ (↓∆′ S) =
A · ↑∆ S
↓∆′ (↑∆ S)
Moreover, we have the following lemma:
Lemma 5.7. If S is a set of contexts and ∆ and ∆′ are sets of exception names, then
(↑∆ (↑∆′ S))⊥
(↓∆ (↓∆′ S))⊥
=
=
(↑∆∪∆′ S)⊥
(↓∆∪∆′ S)⊥
Proof. We only give the proof for (↑∆ (↑∆′ S))⊥ = (↑∆∪∆′ S)⊥ since the other one is similar.
We show two inclusions:
• We show that (↑∆ (↑∆′ S))⊥ ⊆ (↑∆∪∆′ S)⊥ :
Let t ∈ (↑∆ (↑∆′ S))⊥ and C ∈ ↑∆∪∆′ S, by definition C = try D with (∆∪∆′ ) where
D ∈ S. Hence try (try D[t] with ∆′ ) with ∆ ≻∗ z and
− either D[t] ≻∗ z, but then C[t] ≻∗ z.
− or D[t] ≻∗ raise ε when ε ∈ ∆′ or ε ∈ ∆. But then again, C[t] ≻∗ z.
• We show (↑∆∪∆′ S)⊥ ⊆ (↑∆ (↑∆′ S))⊥ :
Let t ∈ (↑∆∪∆′ S)⊥ and C ∈ ↑∆ (↑∆′ S), by definition, C = try (try D with ∆′ ) with ∆
where D ∈ S. Hence try D[t] with (∆∪∆′ ) ≻∗ z and
− either D[t] ≻∗ z, but then C[t] ≻∗ z.
− or D[t] ≻∗ raise ε when ε ∈ ∆∪∆′ . But then again, C[t] ≻∗ z.
Along with the definition of ↑∆ , this lemma implies (A · ↑∆ (↑∆′ S))⊥ = (A · ↑∆∪∆′ S)⊥ .
5.4. Model definition. We call valuation function any function ρ from type variables to
the power set of C minus the empty set (ρ : A → (P(C))+ ). To each type A we associate
two sets:
A set of contexts
| A |ρ ⊆ C
A set of terms
J A Kρ ⊆ T
The set J A Kρ is uniformly defined from | A |ρ by
J A Kρ = | A |ρ ⊥ = { M | ∀C ∈ | A |ρ , M ⊥ C }.
The set | A |ρ is defined by induction on A. Its definition is given Figure 7.
⋆ ∆ and A∆ follows, to
Note that the interpretation in the model of the construction A ∪
⋆ ∆ are terms that may raise an exception only
some extends, the idea that terms of type A ∪
at top level, where terms of A∆ are those that may raise an exception in any evaluation
context. This is emphasized by the “opposition” of the operations ↓∆ and ↑∆ . Remark
that it is only the restriction to z in the handlers of try contexts that allows for such a
simple definition of the interpretation of corruption. Indeed, thanks to this restriction we
ensure that for any context C, C[raise ε] will always reduce to raise ε or z.
The other interesting point of the model is the interpretation of arrow types. In Fx, a
function f which has type A → B has also all the types A∆ → B ∆ for any ∆. Our arrow
type is thus smaller than the usual realizability one and so, functions of Fx are in particular
realizability functions. More formally,
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
| α |ρ
| N |ρ
| A list |ρ
⋆ ∆ |ρ
|A∪
| A∆ |ρ
| A → B |ρ
17
=
=
=
=
=
ρ(α)
{ rec z (λy. λx. x) [ ] }
{ list rec z (λe. λl. λr. (| A |ρ )[e]; r) [ ] }
↓∆ | A |ρ
↑∆ | A |ρ
[
⊥
=
| A∆ |ρ · | B ∆ |ρ
∆⊆E
| ∀α. A |ρ
=
[
| A |ρ; α ← S
S ⊆ C+
Figure 7: Definition of the | A |ρ set of contexts
Lemma 5.8. If A and B are two types and ρ a valuation function, then
\
J A → B Kρ =
{ M | ∀N ∈ J A∆ Kρ , M N ∈ J B ∆ Kρ }.
∆⊆E
Proof. We prove the two inclusions forming the equality separately, both being simple consequences of definitions.
⋆ ∆ is a union and that the interWe can moreover show that the interpretation of A ∪
pretations of the natural numbers and the lists are standards:
Lemma 5.9. If A is a type, ∆ a set of exception names and ρ a valuation function, then
⋆ ∆ Kρ = J A Kρ ∪ { M | M ≻∗ raise ε, ε ∈ ∆ }.
JA∪
Proof. By definition,
⋆ ∆ Kρ = (↓∆ | A |ρ )⊥ = { M | ∀ C ∈ | A |ρ , C[try M with ∆] ≻∗ z } .
JA∪
We show each side of the inclusion separately:
• For M ∈ { t | ∀ C ∈ | A |ρ , C[try M with ∆] ≻∗ z } we have try M with ∆ ∈ J A Kρ .
Hence try M with ∆ has a value (Lemma 5.13) and M as well (Lemma 5.3) :
− Either M ≻∗ raise ε for ε ∈ ∆ and we conclude directly.
− Or M ≻∗ V where V is a regular value and then try M with ∆ ≻∗ V . But the
interpretation being closed by equivalence (Lemma 5.12), V ∈ J A Kρ and M ∈ J A Kρ .
• Let M ∈ J A Kρ ∪ { M | M ≻∗ raise ε, ε ∈ ∆ } and let C ∈ | A |ρ . We have to show
C[try M with ∆] ≻∗ z:
− If M ≻∗ raise ε for ε ∈ ∆, then try M with ∆ ≻∗ z and C[try M with ∆] ≻∗ z.
− Otherwise, M ∈ J A Kρ and then try M with ∆ ∈ J A Kρ (because, M having a value
(Lemma 5.13), there exists V such that M ≻∗ V . But then try M with ∆ ≻∗ V and
we use the closure by equivalence of the interpretation (Lemma 5.12)). Finally, using
the definition of orthogonality, C[try M with ∆] ≻∗ z.
18
S. LEBRESNE
Lemma 5.10. If ρ is a valuation function, ∆ a set of exception names and if Φ represents
one of 0, z or raise ε for some ε ∈ ∆, then
J N∆ Kρ = { M | M ≻∗ S n Φ, n ∈ N }.
Proof. In the following, Φ will always represent one of 0, z or raise ε for some ε ∈ ∆.
• If for some integer n, M ≻∗ S n Φ, then by induction on n it is easy to show that
try (rec z (λy. λx. x) M ) with ∆ ≻∗ z.
• If M ∈ J N∆ Kρ , then there exists k such that try (rec z (λy. λx. x) M ) with ∆ ≻k z.
Hence, we show by induction on k that for any k′ ≤ k and for any term M :
if
′
try (rec z (λy. λx. x) M ) with ∆ ≻k z then
M ≻∗ S n Φ
− We cannot have k = 0 since try (rec z (λy. λx. x) M ) with ∆ is not z.
− We have try (rec z (λy. λx. x) M ) with ∆ ≻ N ≻k z. But the first reduction can
only occurs either if M is 0, z or raise ε (and in the last case we have ε ∈ ∆), or if
M is S M ′ . In this last case, we easily conclude using the reduction of rec and the
induction hypothesis.
Lemma 5.11. If ρ is a valuation function, A a type, ∆ a set of exception names and if Φ
represents one of 0, z or raise ε for some ε ∈ ∆, then
J (A list)∆ Kρ
= { M | M ≻∗ cons a0 (. . . (cons an Φ) . . .), n ∈ N, ∀ 0≤i≤n, ai ∈ J A∆ Kρ } .
Proof. The proof follows the same structure as the one of Lemma 5.10.
5.5. Model properties.
Lemma 5.12 (closure by equivalence). If M and N are two terms, A is a type and ρ is a
valuation function such that M ∈ J A Kρ and M = N , then
N ∈ J A Kρ .
Proof. Let M ∈ J A Kρ and M = N . Let C ∈ | A |ρ , by definition C[M ] ≻∗ z. But since
M = N , C[M ] = C[N ]. Thus, by confluence of the reduction (theorem 3.6), C[N ] ≻∗ z
and N ∈ J A Kρ .
In particular, the interpretation is closed by reduction and anti-reduction.
Lemma 5.13. If M is a term, A a type and ρ a valuation function such that M ∈ J A Kρ ,
then M has a value.
Proof. By definition M ∈ J A Kρ yields M ∈ { M | ∀C ∈ | A |ρ , C[M ] ≻∗ z }. Thus, if
C ∈ | A |ρ , C[M ] has a value and, using Lemma 5.3, we have that M has a value too. We
simply have to make sure that there always exists such a context C, that is | A |ρ is never
empty. But for any type A, it can be easily proved by induction on A that | A |ρ 6= ∅
and z ∈ J A Kρ (remark that both properties have to be proved simultaneously since the
non-emptiness of | A → B |ρ depends upon the non-emptiness of J A Kρ which comes (by
induction hypothesis) of the non-emptiness of | A |ρ ).
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
19
Lemma 5.14. If A is a type, ρ a valuation function and ∆ a set of exception names, then
for all ε ∈ ∆, raise ε ∈ J A∆ Kρ .
Proof. Let C ∈ | A∆ |ρ , by definition C = {try [ ] with∆}◦D where D ∈ | A |ρ . We can easily
show that either D[raise ε] ≻∗ z or D[raise ε] ≻∗ raise ε. In any case, C[raise ε] ≻∗ z.
An important and essential property of the model is that it validates the subtyping rule
(eq-arrc):
Lemma 5.15. If A and B are two types, ρ is valuation function and ∆ is a set of exception
names, then
J (A → B)∆ Kρ = J A∆ → B ∆ Kρ
Proof. Using Lemmas 5.6 and 5.7, we have that
\
⊥
J (A → B)∆ Kρ =
((↑∆′ | A |ρ )⊥ · ↑∆∪∆′ | B |ρ )
′
∆ ⊆E
\
⊥
∆
∆
J A → B Kρ =
((↑∆∪∆′ | A |ρ )⊥ · ↑∆∪∆′ | B |ρ )
∆′ ⊆ E
It directly follows that J (A → B)∆ Kρ ⊆ J A∆ → B ∆ Kρ .
For the other inclusion, let M ∈ J A∆ → B ∆ Kρ . If ∆′ is a set of exception names and
if C ∈ (↑∆′ | A |ρ )⊥ · ↑∆∪∆′ | B |ρ , since we can show that (↑∆′ | A |ρ )⊥ ⊆ (↑∆∪∆′ | A |ρ )⊥ we
have C ∈ (↑∆∪∆′ | A |ρ )⊥ · ↑∆∪∆′ | B |ρ and we conclude.
Lemma 5.16. The interpretations validate the following equalities:
| (∀α. A)∆ |ρ
⋆ ∆ |ρ
| (∀α. A) ∪
′
⋆ ∆)∆ |ρ
| (A ∪
′
J (A∆ )∆ Kρ
⋆ ∆) ∪
⋆ ∆′ K ρ
J (A ∪
=
=
=
=
=
| ∀α. A∆ |ρ
⋆ ∆ |ρ
| ∀α. A ∪
′
∆
⋆ ∆ |ρ
| (A ) ∪
′
∆∪∆
Kρ
JA
⋆ (∆∪∆′ ) Kρ
JA∪
Proof. The three first equalities are direct consequence of the definitions, the two last are
direct consequence of Lemma 5.7.
5.6. Model soundness. We first show that subtyping is sound with respect to the interpretation we have defined.
Lemma 5.17 (Subtyping soudness). If A and B are two types and ρ a valuation function
such that A ≤ B, then for any set of exception names ∆,
J A∆ Kρ ⊆ J B ∆ Kρ
Proof. We reason by induction on the derivation of A ≤ B. Many cases are either trivial
((st-id) and (st-trans)) or direct consequences of the lemmas we have defined so far. We
only give in the following the cases that do not belong to one of these categories:
(st-arrow): Consider ∆ ⊆ E and M ∈ J (A → B)∆ Kρ = J A∆ → B ∆ Kρ . We now have
to show that M ∈ J (A′ → B ′ )∆ Kρ = J A′∆ → B ′∆ Kρ . For C ∈ | A′∆ → B ′∆ |ρ we will es′
tablish C[M ] ≻∗ z. By definition of | A′∆ → B ′∆ |ρ , there exist ∆′ , N ∈ J A′∆∪∆ Kρ and
′
′
′
D ∈ | B ′∆∪∆ |ρ such that C = D[[ ] N ]. By induction hypothesis, J A′∆∪∆ Kρ ⊆ J A∆∪∆ Kρ .
20
S. LEBRESNE
′
But since M ∈ J A∆ → B ∆ Kρ , using Lemma 5.8, we have M N ∈ J B ∆∪∆ Kρ . Then by
′
′
induction hypothesis, J B ∆∪∆ Kρ ⊆ J B ′∆∪∆ Kρ and finally, C[M ] = D[M N ] ≻∗ z.
(f-gen): For ∆ ⊆ E and M ∈ J A∆ Kρ we will show that M ∈ J (∀α. B)∆ Kρ = J ∀α. B ∆ Kρ .
Let C ∈ | ∀α. B ∆ |ρ , there exists S such that C ∈ | B ∆ |ρ; α ← S . Moreover, since α ∈
/ FV(A),
∆
∆
∆
∆
we have J A Kρ = J A Kρ; α ← S . And since J A Kρ; α ← S ⊆ J B Kρ; α ← S by induction hypothesis, M ∈ J B ∆ Kρ; α ← S and finally, C[M ] ≻∗ z.
(f-inst): Given ∆ ⊆ E, we will show that | (A{α := B})∆ |ρ ⊆ | (∀α. A)∆ |ρ and then conclude by orthogonality (Lemma 5.5). Let C ∈ | (A{α := B})∆ |ρ , we show by a straightforward induction on A that | (A{α := B})∆ |ρ = | A∆ |ρ; α ← | B |ρ . Moreover, by definition of | ∀α. A∆ |ρ , we have | A∆ |ρ; α ← | B |ρ ⊆ | ∀α. A∆ |ρ , from which if follows that
C ∈ | ∀α. A∆ |ρ = | (∀α. A)∆ |ρ .
(f-distr): Consider ∆ ⊆ E and t ∈ J (∀α. (A → B))∆ Kρ = J ∀α. (A∆ → B ∆ ) Kρ , we will
show that t ∈ J (A → ∀α. B)∆ Kρ = J A∆ → ∀α. B ∆ Kρ . Let C ∈ | A∆ → ∀α. B ∆ |ρ , by def′
′
′
inition there exists ∆′ , u ∈ J A∆∪∆ Kρ and D ∈ | (∀α. B ∆ )∆ |ρ = | ∀α. B ∆∪∆ |ρ such that
′
/ FV(A),
C = D[[ ] u]. Then there exists S such that D ∈ | B ∆∪∆ |ρ; α ← S and since α ∈
′
u ∈ J A∆∪∆ Kρ; α ← S . Thus by definition of J ∀α. (A∆ → B ∆ ) Kρ and using Lemma 5.8, we
′
have t u ∈ J B ∆∪∆ Kρ; α ← S and finally C[t] = D[t u] ≻∗ z.
⋆ ∆′ )∆ Kρ = J (A∆ → B ∆ ) ∪
⋆ ∆′ Kρ , we
(ex-arru): Consider ∆ ⊆ E and t ∈ J ((A → B) ∪
⋆ ∆′ )∆ Kρ = J A∆ → B ∆ ∪
⋆ ∆′ Kρ . Consider ∆′′ ⊆ E and
have to show that t ∈ J (A → B ∪
′′
′′
∆∪∆
⋆ ∆′ Kρ . But using
Kρ , using Lemma 5.8, we must show that t u ∈ J B ∆∪∆ ∪
u∈ JA
Lemma 5.9 we have:
• Either t ≻∗ raise ε for ε ∈ ∆′ , but then t u ≻∗ raise ε and hence using Lemma 5.9,
′′
⋆ ∆′ K ρ .
t u ∈ J B ∆∪∆ ∪
′′
• Or t ∈ J A∆ → B ∆ Kρ and using Lemma 5.8, t u ∈ J B ∆∪∆ Kρ which in turn gives
′′
⋆ ∆′ Kρ (Lemma 5.9).
t u ∈ J B ∆∪∆ ∪
(ex-ctx): This case is trivial with the use of Lemma 5.9.
(ex-uni): This case is trivial with the use of Lemma 5.9.
′
′
⋆ ∆′ )∆ Kρ = J A∆ ∪
⋆ ∆′ Kρ ⊆ J (A∆ )∆ Kρ = J A∆∪∆ Kρ .
(ex-corrupt): We need to show J (A ∪
⋆ ∆′ Kρ = J A∆ Kρ ∪ { t | t ≻∗ raise ε, ε ∈ ∆′ } and it can be easily shown that
But J A∆ ∪
′
if ∆ ⊆ ∆′ , then J A∆ Kρ ⊆ J A∆ Kρ and that if ε ∈ ∆, then raise ε ∈ J A∆ Kρ (Lemma 5.14).
We define the interpretation corrupted by some set of exception names ∆ (eventually empty)
of a typing context Γ by:
J Γ∆ K ρ
=
{ σ | ∀ (x : A) ∈ Γ, σ(x) ∈ J A∆ Kρ }
Moreover, if σ is a substitution of term variables and M is a term, we use the notation M [σ]
for the parallel substitution of M by σ, which consists in applying σ to all free variables of
M in parallel. We can now show that our interpretation is sound with respect to typing:
Theorem 5.18 (Model soundness). If M is a term, A a type and Γ a typing context such
that Γ ⊢ M : A, then for all valuation function ρ, for all set of exception names ∆ and for
all substitution σ ∈ J Γ∆ Kρ , we have M [σ] ∈ J A∆ Kρ .
Proof. We use induction on the derivation of Γ ⊢ t : A. Note that since J A Kρ ⊆ J A∆ Kρ
(Lemma 5.17), we will only show that t[σ] ∈ J A Kρ when possible. We give here only the
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
21
interesting cases. The other cases are either simple ((ax), (subs), (zero), (succ), (nil),
(cons)) or, for (fold), follows closely the structure of the proof for (rec).
(abs): Let C ∈ | (A → B)∆ |ρ = | A∆ → B ∆ |ρ , we need to show that C[λx. t[σ]] ≻∗ z.
′
′
By definition, there exists ∆′ , u ∈ J A∆ Kρ and D ∈ | B (∆∪∆ ) |ρ such that C = D[[ ] u].
′
′
′
Then, if δ = σ + {x → u}, we have δ ∈ J (Γ, x : A)(∆∪∆ ) Kρ (σ ∈ J Γ∆ Kρ ⊆ J Γ(∆∪∆ ) Kρ and
′
′
′
u ∈ J A∆ Kρ ⊆J A∆∪∆ Kρ ), and also by induction hypothesis, t[δ] ∈ J B (∆∪∆ ) Kρ . However,
′
(λx. t[σ]) u ≻ t[δ] and J B (∆∪∆ ) Kρ is closed by anti-reduction (Lemma 5.12), and thus
′
(λx. M [σ]) u ∈ J B (∆∪∆ ) Kρ and finally C[λx. M [σ]] = D[(λx. M [σ]) u] ≻∗ z.
(app): We easily conclude using Lemma 5.8.
(gen): Let C ∈ | (∀α. A)∆ |ρ = | ∀α. A∆ |ρ , by definition there exists S non empty such that
C ∈ | A∆ |ρ; α ← S . Moreover, since α ∈
/ FV(Γ), J Γ∆ Kρ = J Γ∆ Kρ; α ← S . It follows that by
∆
induction hypothesis, t[σ] ∈ J A Kρ; α ← S . Finally, C[t[σ]] ≻∗ z and t[σ] ∈ J (∀α. A)∆ Kρ .
(rec): We have to show that
⋆ ∆ → (N∆ → α ∪
⋆ ∆ → α∪
⋆ ∆) → N∆ ∪
⋆ ∆′ → α ∪
⋆ (∆∪∆′ ) Kρ .
rec ∈ J ∀α. α ∪
Using Lemma 5.8, we have to show that for any non empty set of contexts S, for any
∆1 , ∆2 and ∆3 and for
⋆ ∆ Kρ; α ← S
z ∈ J α∆1 ∪
∆
∪∆
1
2 ∪∆ → α∆1 ∪∆2 ∪
⋆ ∆ → α∆1 ∪∆2 ∪
⋆ ∆ Kρ; α ← S
f ∈ JN
∆
∪∆
∪∆
′
1
2
3
⋆
n ∈ JN
∪ ∆ Kρ; α ← S
⋆ (∆∪∆′ ) Kρ; α ← S . With Lemma 5.9, we have either
we have rec z f n ∈ J α∆1 ∪∆2 ∪∆3 ∪
n ≻∗ raise ε for ε ∈ ∆′ (and we easily conclude), or n ∈ J N∆1∪∆2 ∪∆3 ∪∆ Kρ; α ← S . In
this last case, there exists some k such that n ≻∗ S k Φ where Φ is one of 0, z or
raise ε for ε ∈ ∆1 ∪∆2 ∪∆3 ∪∆ (Lemma 5.10). We then proceed by induction on k.
If k = 0 (n ≻∗ Φ), we easily conclude in each case of Φ. Otherwise, we must show
⋆ (∆∪∆′ ) Kρ; α ← S . But rec z f (S (S k Φ)) ≻
that rec z f (S (S k Φ)) ∈ J α∆1 ∪∆2 ∪∆3 ∪
k
k
f (S Φ) (rec z f (S Φ)). We then conclude using Lemma 5.8 with f , the fact that
S k Φ ∈ J N∆1∪∆2 ∪∆3 ∪∆ Kρ (Lemma 5.10) and the induction hypothesis.
(raise): It is a direct using Lemma 5.9.
⋆ ∆)∆ Kρ = J A∆ ∪
⋆ ∆ Kρ and u[σ] ∈ J A∆ Kρ . Us(try): By induction hypothesis, t[σ] ∈ J (A ∪
∆
ing Lemma 5.9, we have either that t[σ] ∈ J A Kρ or that t[σ] ≻∗ raise ε. Since terms
inhabiting the interpretation have values (Lemma 5.13), in both case we can show that
either try t[σ] with ε 7→ u[σ] reduces to u[σ] or, if t[σ] ∈ J A∆ Kρ , it reduces to some t′
such that t[σ] ≻∗ t′ . In both case, we can conclude.
Note that in this model, we only consider closed terms by construction. For this very reason,
we cannot establish a strong normalization theorem using this model. But, from the model,
we obtain a form a weak head normalization theorem (let us recall that values corresponds
to weak head normal form):
Theorem 5.19 (Weak head normalization). If M is a closed term, A a type and Γ a typing
context such that Γ ⊢ M : A, then M has a value.
Proof. This comes directly from the model soundness theorem and Lemma 5.13.
22
S. LEBRESNE
The model allows us to prove for instance that our typing of exceptions is safe for the
primitive data types of the natural numbers:
Lemma 5.20 (type safety for natural numbers). If M is a term such that ⊢ M : N, then
M ≻∗ S n 0 for some n≥0.
Proof. If ⊢ M : N, then with Theorem 5.18, M ∈ J N Kρ . We conclude using Lemma 5.10
and the fact that M is well typed and there is no typing rule for z.
Hence, if a program is of the type of the natural numbers, we assure that it will compute
a true natural number without producing errors.
6. Related Works
The static detection of uncaught exceptions has been studied in many works, based on
typing or not. For instance, for the OCaml languages, J.C. Guzmn and A. Surez [7] have
proposed an extension of the type system where arrows are annotated by the exceptions a
function can raise. Later, X. Leroy and F. Pessaux [13] have proposed a similar system but
have added polymorphism over these annotations. Their solution is efficient and covers all
the Ocaml language, including modularity. However, all these works consider exceptions in
call-by-value languages and rely heavily on the exceptions-as-control-flow paradigm.
In call-by-name, it is standard to use monads to encode exceptions [21, 16]. We have
however already explained in section 2.1 the drawbacks of such approach. As already
stressed, from a computational point of view, the exception mechanism described in this
paper is very similar to the imprecise exceptions of S. Peyton Jones et al. [14] who are
implemented in the ghc Haskell compiler [17]. The novelty of this paper is to provide a
precise type system for this exception mechanism while in [14] exceptional values inhabit
all types. The “imprecision” of imprecise exceptions comes from the willingness to not force
a particular reduction strategy for primitive binary operators. For instance, with imprecise
exception the term (raise ε) + (raise ε′ ) evaluates to the set {raise ε, raise ε′ } (hence
exceptional values are sets). Since in Fx we do not have binary primitive operators, we
have no need for such so-called imprecision. However, in Fx, the addition should be coded
using the rec operator, such coding being bound to be non commutative for exceptions (the
coding have to choose on which operand of the addition the recursion should be performed).
We however believe that if needed, the typing of exceptions presented in this paper could be
adapted with almost no changes to the case of imprecise exceptions since our type notions
already deal with sets of exceptions.
In the literature, exceptions are often considered as control operators. Note however
that exceptions have a dynamic semantic, and as such, cannot be compared to static control
operators like first-class continuations [18]. In particular, the typing of exceptions does not
necessarily lift the logic to a classical one. Besides, in this paper, we address the problem of
the static detection of uncaught exceptions. We do not know of previous works on control
operators dealing with this particular problem.
Exceptions in type theoretical settings have been less studied. However, R. David
and G. Mounier [3] have designed a typed mechanism of exceptions for the language AF2.
However, as with monads, the propagation of exceptions in their system has to be forced by
means of Krivine’s storage operators. Besides, their exceptions are restricted in the sense
that only data types can carry exceptions and for example, exceptions cannot be used as
functions.
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
23
7. Conclusion and future works
We have presented the Fx calculus, an extension of System F with typed exceptions. We
have presented a mechanism of exceptions that does not force a particular β-reduction
strategy for the calculus. We have also provided a type system for this mechanism that
performs static detection of uncaught exceptions. This type system is modular and allows
the use and propagation of exceptions to be transparent for the programmer. Finally, we
have justified the semantic of our calculus by exhibiting a realizability model.
This calculus can be improved in a certain number of ways. First, by proving more metatheoretical properties. Our realizability model only allows to prove weak head normalization
but it could probably be modified in order to prove strong normalization. In fact, we believe
that the simple change of the definition of the orthogonality relation (definition 5.4) to
“M ⊥ C if and only if C[M ] ≻∗ z and C[M ] is strongly normalizing”, would yield a strong
normalization model (but with this new notion the interpretation will not be closed by antireduction anymore and proofs will have to be adapted). Moreover, we have not completed
yet the proof of subject-reduction for Fx. However, a detailed proof of subject-reduction
for the restriction of the calculus to first-order can be found in [8] (showing that corruption
does not break intrinsically the subject-reduction property). Adapting this proof to second
order (and thus to Fx) is however not trivial, not because of corruption, but because of the
subtyping rules of quantification. Besides, the realizability model already proves a form a
type safety for the calculus.
Type inference for Fx is obviously undecidable [25]. But type inference for restrictions
of Fx, to first-order for instance, remains to be studied, and we have good hopes since we
know that in such a restriction, the subtyping relation is decidable.
Exceptions in Fx are simple names. We would like to extend the calculus so that they
carry arguments. However, we will then need to account in the type system for the types
of these arguments, which will complicates notably the type system.
As mentioned in the introduction, we think that corruption is a promising notion for the
addition of exceptions to proof assistants based on type theoretical calculi. To that end, we
think that a natural extension would be to add dependent product to our calculus. As our
type system is heavily based on subtyping, we would build on previous works on subtyping
in dependent calculus [2, 9]. Moreover, we already know how to extend our realizability
model to handle the dependent product: if T is a type and Ux a type family indexed by x,
we can take
[
∆
| Πx : T. U |ρ =
{ M · C | M ∈ J T ∆ Kρ ∧ C ∈ | UM
|ρ }
∆⊆E
References
[1] H.P. Barendregt. The lambda calculus. North-Holland, 1984.
[2] G. Chen. Subtyping calculus of construction, extended abstract. In The 22nd International Symposium
on Mathematical Foundation of Computer Science, volume 1295. Springer.
[3] R. David and G. Mounier. An intuitionistic λ-calculus with exceptions. Journal of Functional Programming, 15(01):33–52, 2004.
[4] The Coq development team. The Coq Proof Assistant Reference Manual v8.1, 2006.
[5] J.Y. Girard. Locus Solum: From the rules of logic to the logic of rules. Mathematical Structures in
Computer Science, 11(03):301–506, 2001.
[6] J.Y. Girard et al. Proofs and types. Cambridge University Press New York, 1989.
24
S. LEBRESNE
[7] J. Guzman and A. Suarez. An extended type system for exceptions. Proceedings of the ACM SIGPLAN
Workshop on ML and its Applications, pages 127–135, 1994.
[8] Sylvain Lebresne. Une approche de la détection statique d’exceptions non rattrapées en appel par nom.
PhD thesis, Université Paris Diderot – Paris 7, 2008.
[9] A. Miquel. The implicit calculus of constructions. In Proceedings of the 5th International Conference
on Typed Lambda Calculi and Applications (TLCA 2001), volume 2044, pages 344–359, 2001.
[10] J.C. Mitchell. Polymorphic type inference and containment. Information and Computation, 76(2-3):211–
249, 1988.
[11] E. Moggi. Notions of computation and monads. INF. COMPUT., 93(1):55–92, 1991.
[12] M. Parigot. Strong normalization for second order classical natural deduction. Logic in Computer Science, 1993. LICS’93., Proceedings of Eighth Annual IEEE Symposium on, pages 39–46, 1993.
[13] F. Pessaux and X. Leroy. Type-based analysis of uncaught exceptions. Proceedings of the 26th ACM
SIGPLAN-SIGACT symposium on Principles of programming languages, pages 276–290, 1999.
[14] S. Peyton Jones, A. Reid, F. Henderson, T. Hoare, and S. Marlow. A semantics for imprecise exceptions.
ACM SIGPLAN Notices, 34(5):25–36, 1999.
[15] Randy Pollack. The LEGO Proof Assistant, 1998.
[16] M. Spivey. A functional theory of exceptions. Science of Computer Programming, 14(1):25–42, 1990.
[17] The
GHC
Team.
GHC,
the
control.exception
module,
2009.
http://www.haskell.org/ghc/docs/latest/html/libraries/base/Control-Exception.html.
[18] H. Thielecke. Comparing Control Constructs by Double-Barrelled CPS. Higher-Order and Symbolic
Computation, 15(2):141–160, 2002.
[19] J. Tiuryn and P. Urzyczyn. The subtyping problem for second-order types is undecidable. Logic in
Computer Science, 1996. LICS’96. Proceedings., Eleventh Annual IEEE Symposium on, pages 74–85,
1996.
[20] J. Vouillon and P.A. Melliès. Semantic types: a fresh look at the ideal model for types. Proceedings of
the 31st ACM SIGPLAN-SIGACT symposium on Principles of programming languages, pages 52–63,
2004.
[21] P. Wadler. How to Replace Failure by a List of Successes A method for exception handling, backtracking,
and pattern matching. Functional Programming Languages and Computer Architecture, 1985.
[22] P. Wadler. Comprehending monads. Proceedings of the 1990 ACM conference on LISP and functional
programming, pages 61–78, 1990.
[23] P. Wadler and P. Thiemann. The marriage of effects and monads. ACM Transactions on Computational
Logic (TOCL), 4(1):1–32, 2003.
[24] J.B. Wells. The undecidability of Mitchells subtyping relation. Technical Report 95-019, Boston University, Boston, Massachusetts, 1995.
[25] J.B. Wells. Typability and type checking in System F are equivalent and undecidable. Annals of Pure
and Applied Logic, 98(1-3):111–156, 1999.
Appendix A. Parallel reduction for Fx
M ≫M
M ≫ M′
λx. M ≫ λx. M ′
M ≫ M′ N ≫ N′
(λx. M ) N ≫ M ′ {x := N ′ }
N ≫ N′
try (raise ε) with ε 7→ N ≫ N ′
M ≫ M′ N ≫ N′
M N ≫ M′ N′
(raise ε) M ≫ raise ε
try (raise ε′ ) with ε 7→ N ≫ raise ε′
M ≫ M′ N ≫ N′
try M with ε 7→ N ≫ try M ′ with ε 7→ N ′
A TYPE SYSTEM FOR CALL-BY-NAME EXCEPTIONS
25
V ≫ V ′ V is a regular value
try V with ε 7→ N ≫ V ′
X ≫ X′
rec X Y 0 ≫ X ′
X ≫ X′ Y ≫ Y ′ N ≫ N ′
rec X Y (S N ) ≫ Y ′ N ′ (rec X ′ Y ′ N ′ )
rec X Y (raise ε) ≫ raise ε
X ≫ X′
list rec X Y [ ] ≫ X ′
X ≫ X ′ Y ≫ Y ′ E ≫ E ′ L ≫ L′
list rec X Y (cons E L) ≫ Y ′ E ′ L′ (list rec X ′ Y ′ L′ )
list rec X Y (raise ε) ≫ raise ε
This work is licensed under the Creative Commons Attribution-NoDerivs License. To view
a copy of this license, visit http:// reative ommons.org/li enses/by-nd/2.0/ or send a
letter to Creative Commons, 171 Second St, Suite 300, San Francisco, CA 94105, USA, or
Eisenacher Strasse 2, 10777 Berlin, Germany
| 6cs.PL
|
ML 2005 Preliminary Version
Practical Datatype Specializations with
Phantom Types and Recursion Schemes
arXiv:cs/0510074v1 [cs.PL] 24 Oct 2005
Matthew Fluet 1 Riccardo Pucella 2
Department of Computer Science
Cornell University
Ithaca, NY 14853 USA
Abstract
Datatype specialization is a form of subtyping that captures program invariants
on data structures that are expressed using the convenient and intuitive datatype
notation. Of particular interest are structural invariants such as well-formedness.
We investigate the use of phantom types for describing datatype specializations. We
show that it is possible to express statically-checked specializations within the type
system of Standard ML. We also show that this can be done in a way that does not
lose useful programming facilities such as pattern matching in case expressions.
Key words: data types and structures, invariants
1
Introduction
Data structures that are used pervasively in an application are often implemented with a degree of genericity that makes them suited, though not always
well suited, to their various uses. A benefit of this genericity is that it ensures
that core functionality of the data structure is available to all clients. A
limitation is that a client that is required to produce, consume, or maintain
an instance of the data structure subject to a particular invariant has difficulty enforcing the invariant. While many languages boast a type system
that statically enforces basic safety properties on data structures, few allow
programmers to directly capture these additional invariants within the type
system. This is unfortunate: when invariants are reflected into a type system,
compile-time type errors will indicate code that could violate these invariants.
To make our discussion concrete, consider Boolean formulas, which we use
as a running example throughout this paper. A straightfoward representation
of formulas is the following:
1
2
Email: [email protected]
Email: [email protected]
This is a preliminary version. The final version will be published in
Electronic Notes in Theoretical Computer Science
URL: www.elsevier.nl/locate/entcs
Fluet and Pucella
datatype fmla = Var of string | Not of fmla
| True | And of fmla * fmla
| False | Or of fmla * fmla
We can easily define a function eval that takes a formula and an environment
associating every variable in the formula with a truth value, and returns the
truth value of the formula. Similarly, we can define a toString function that
takes a formula and returns a string representation of the formula. As is well
known, a propositional formula can always be represented in a special form
called Disjunctive Normal Form (or DNF), as a disjunction of conjunctions of
variables and negations of variables. A formula in DNF is still a formula, but it
has a restricted structure. Some algorithms that operate on formulas require
their input to be presented in DNF; therefore, it makes sense to statically
check that a formula is in DNF.
One way to perform this static checking is to simply introduce one datatype
for (generic) formulas and another datatype for DNF formulas, and provide
functions to explicitly convert between them. This is, of course, inefficient. For
example, converting a formula in DNF to a string via toString would require
two complete traversals of the DNF formula (one to convert it to a generic
formula, and one to build the string representation), as well as the allocation
of an intermediate structure (of the same size as the original formula). An
alternative is to define a DNF formula as a specialization of formulas. For
the sake of presentation, we assume a special syntax for specializations. This
syntax should be self-explanatory:
datatype fmla
=
|
|
withspec atom =
and lit =
and conj =
and dnf =
Var of string
True | And of
False | Or of
Var of string
Var of string
True | And of
False | Or of
| Not of fmla
fmla * fmla
fmla * fmla
| Not of atom
lit * conj
conj * dnf
Roughly speaking, the specialization dnf of the datatype fmla is restricted so
that the Or constructor creates list of conjunctions terminated with the False
constructor. A conjunction is defined by another specialization conj of the
datatype fmla that restricts the And constructor to forming lists of literals.
A literal is essentially a variable or a negated variable. This can be captured
using two specializations, atom for atomic literals and lit for literals. Notice
that to define the dnf specialization, we need all the specializations dnf, conj,
lit, and atom. These specializations induce a simple subtyping hierarchy:
fmlaH
lit
HHH
ss
sss
conj
dnf
atom
For uniformity, we consider the datatype as a degenerate specialization.
Abstracting from this example, we define a specialization of a datatype to
2
Fluet and Pucella
be a version of the datatype that is generated by a subset of the datatype
constructors, which may themselves be required to be applied to specializations of the datatype. If we view elements of a datatype as data structures,
then we can view elements of the specialization as data structures obeying
certain restrictions. A set of specializations of a datatype may be specified in
a mutually recursive fashion.
In this paper, we show that we can implement, in Standard ML [15], much
of what one would expect from a language with a type system that directly
supports the kind of specializations described above. For example, we can
write a toDnf function that statically guarantees not only that its result is
a formula, but also that it is a DNF formula. The advantage of implementing specialization invariants in SML is that type systems directly supporting
specializations are complex and not widely available. (A type system that
enforces similar, but strictly more powerful, invariants is the refinement types
system [3,8]; see Section 4.)
What are the key features of specializations that we would like available?
For one, we would like the representation of values of specialized types to be
the same as the representation of the original datatype. This is important to
avoid expensive run-time conversions and to allow code reuse. For example,
we should be able to implement a single function to evaluate not only an
unspecialized formula, but also any specialization of formulas, such as the dnf
specialization. Moreover, we would like to write case expressions that do not
include branches for constructors that do not occur in the specialization of
the value being examined. For example, if we perform a case analysis on a
value with specialization dnf, we should only need to supply branches for the
False and Or constructors.
Drawing on previous work relating phantom types and subtyping [7] and
the intuition that the specializations of a datatype induce a subtyping hierarchy, we present in Section 2 an informal translation based on phantom
types from a set of specializations of a datatype to an interface providing
constructors, destructors, and coercions corresponding to the specializations.
This interface forms a minimal set of primitive operations that provide the
functionality (and static guarantees) of the specializations. It uses the SML
type system to enforce the invariants embodied by the specializations.
Unfortunately, the interface leaves much to be desired, particularly if we
wish to promote datatype specializations as a practical programming technique that integrates naturally with standard, idiomatic SML usage. In order
to overcome these deficiencies, notably the lack of pattern matching, we draw
on another programming techique: recursion schemes [28]. We present in Section 3 an informal translation from the set of specializations of the datatype,
using recursion schemes and the interface described above, to an improved
interface that provides the functionality of the specializations through bona
fide SML datatypes. As before, the SML type system enforces the invariants
of the specializations.
3
Fluet and Pucella
The reader may well wonder why we choose to encode specializations in
SML, rather than designing a language extension. First, there are numerous
SML compilers [9,16,17,18,20,24,25]; while many of these compilers implement
some language extensions, the SML language specified by the Definition [15]
behaves the same in all compilers. Hence, the technique described here is
available now to all SML programmers. Furthermore, encoding specializations
in SML is more expedient than writing our own compiler or modifying an
existing compiler.
Our notion of datatype specializations is quite general, and we believe that
it can capture a good number of useful invariants. For example, we can define
a specialization that ensures that the formula contains no variables:
datatype fmla
=
|
|
withspec grnd =
|
|
Var of string
True | And of
False | Or of
Not of grnd
True | And of
False | Or of
| Not of fmla
fmla * fmla
fmla * fmla
grnd * grnd
grnd * grnd
The following specializations distinguish between zero and non-zero:
datatype nat
= Zero | Succ of nat
withspec zero
= Zero
and nonzero = Succ of nat
Lists give rise to interesting specializations. Consider the following specializations, distinguishing between empty, singleton, and nonempty lists:
datatype ’a list
withspec ’a empty
and ’a singleton
and ’a nonempty
=
=
=
=
Nil | Cons of ’a * ’a list
Nil
Cons of ’a * ’a empty
Cons of ’a * ’a list
The following specializations distinguish between lists of even and odd length:
datatype ’a list
= Nil | Cons of ’a * ’a list
withspec ’a even = Nil | Cons of ’a * ’a odd
and ’a odd = Cons of ’a * ’a even
Finally, we can use specializations to define abstract syntax trees that distinguish between arbitrary expressions and well-formed expressions (e.g., welltyped expressions [4,13], expressions in normal forms, etc.). A simple example
of this is the following:
datatype exp
=
|
|
withspec boolexp =
|
and intexp =
|
Bool of bool | And of exp * exp
Int of int | Plus of exp * exp
If of exp * exp * exp
Bool of bool | And of boolexp * boolexp
If of boolexp * boolexp * boolexp
Int of int | Plus of intexp * intexp
If of boolexp * intexp * intexp
More involved examples that can be expressed using specializations include
red-black trees that check the critical invariant that no red node has a red child
after inserting a new element, and constructors for expressions in the simply4
Fluet and Pucella
typed λ-calculus that permit only the building of type-correct expressions,
essentially using an encoding manipulating de Brujin indices.
2
Specializations with Phantom Types
How should we write (in SML) an implementation of the formula specializations so that the type system enforces the appropriate structural invariants?
In this section, we give a highly-stylized implementation that achieves this
particular goal. We hope that the reader will grasp the straightforward generalization of this implementation to arbitrary specializations. 3 We feel that a
fully elaborated example is more instructive than a formal translation, where
definitions and notation become burdensome and obfuscating.
We first review the essence of the phantom-types technique and its application to subtyping [7]. The phantom-types technique uses the definition
of type equivalence to encode information in a superfluous type variable of a
type. (Because instantiations of this type variable do not contribute to the
run-time representation of values of the type, it is called a phantom type.) Unification can then be used to enforce a particular structure on the information
carried by two such types.
When applied to subtyping, the information we wish to encode is a position
within a subtyping hierarchy. We require an encoding hσi of each specialization σ in the hierarchy; this encoding should yield a type in the SML type
system, with the property that hσ1 i unifies with hσ2 i if and only if σ1 is a
subtype of σ2 in the hierarchy (written σ1 ≤ σ2 ). An obvious issue is that
we want to use unification (a symmetric relation) to capture subtyping (an
asymmetric relation). The simplest solution is to use two encodings h·iC and
h·iA defined over all the specializations in the hierarchy. A value of specialization σ will be given a type using hσiC . We call hσiC the concrete encoding of
σ, and we assume that it uses only ground types (i.e., no type variables). In
order to restrict the domain of an operation or constructor to the set of values
that are subtypes of a specialization σ, we use hσiA , the abstract encoding of
σ. In order for the underlying type system to enforce the subtyping hierarchy,
we require the encodings h·iC and h·iA to respect the subtyping hierarchy by
satisfying the following property:
For all specializations σ1 and σ2 , hσ1 iC unifies with hσ2 iA iff σ1 ≤ σ2 .
To allow for unification, the abstract encoding introduces free type variables.
Since, in the SML type system, a top-level type cannot contain free type
variables, the abstract encoding is always a part of some polymorphic type
3
Our running example uses a first-order, monomorphic datatype, of which abstract syntax
trees are a typical example. Extending the implementation to handle first-order, polymorphic datatypes is straightforward. It is also possible to handle higher-order datatypes; we
briefly consider this in Section 4.
5
Fluet and Pucella
signature FMLA = sig
(* specialization type *)
type ’a t
(* abstract types *)
type ’a AFmla = {fmla: ’a} t
type ’a ALit = {lit: ’a} AFmla
type ’a AAtom = {atom: ’a} ALit
type ’a AConj = {conj: ’a} AFmla
type ’a ADnf = {dnf: ’a} AFmla
(* concrete types *)
type CFmla = unit AFmla
type CLit = unit ALit
type CAtom = unit AAtom
type CConj = unit AConj
type CDnf = unit ADnf
structure Fmla : sig
(* constructors *)
val Var
: string -> CFmla
val Not : ’a AFmla -> CFmla
val True
: CFmla
val And : ’a AFmla * ’b AFmla -> CFmla
val False : CFmla
val Or : ’a AFmla * ’b AFmla -> CFmla
(* destructor *)
val dest
: ’a AFmla -> {Var
: string -> ’b, Not : CFmla -> ’b,
True : unit -> ’b,
And : CFmla * CFmla -> ’b,
False : unit -> ’b,
Or : CFmla * CFmla -> ’b} -> ’b
(* coercion *)
val coerce : ’a AFmla -> CFmla
end
structure Lit : sig
val Var
: string -> CLit
val Not : ’a AAtom -> CLit
val dest
: ’a ALit -> {Var : string -> ’b, Not : CAtom -> ’b} -> ’b
val coerce : ’a ALit -> CLit
end
structure Atom : sig
val Var
: string -> CAtom
val dest
: ’a AAtom -> {Var : string -> ’b} -> ’b
val coerce : ’a AAtom -> CAtom
end
structure Conj : sig
val True
: CConj
val And : ’a ALit * ’b AConj -> CConj
val dest
: ’a AConj -> {True : unit -> ’b, And : CLit * CConj -> ’b} -> ’b
val coerce : ’a AConj -> CConj
end
structure Dnf : sig
val False : CDnf
val Or : ’a AConj * ’b ADnf -> CDnf
val dest
: ’a ADnf -> {False : unit -> ’b, Or : CConj * CDnf -> ’b} -> ’b
val coerce : ’a ADnf -> CDnf
end
end
Fig. 1. The FMLA signature
scheme. This leads to some restrictions on the uses of abstract encodings,
the details of which are beyond the scope of this paper (but, see our previous
work [7] for a thorough discussion). We assert that abstract encodings are
used appropriately in the following presentation.
Figures 1 and 2 give a signature and corresponding implementation of the
specializations above. The amount of code may seem staggering for such a
small example, but we shall see that most of it is boilerplate code. (In fact,
it is straightforward to mechanically generate this code from a declarative
6
Fluet and Pucella
structure Fmla :> FMLA = struct
structure Rep = struct
(* representation type *)
datatype t = Var of string | Not of t
| True | And of t * t
| False | Or of t * t
(* destructor *)
fun fail _ = raise Match
fun dest c (v,n,t,a,f,o) = case c of Var(s) => v (s) | Not(f) => n (f)
| True => t () | And(f,g) => a (f,g)
| False => f () | Or(f,g) => o (f,g)
(* coercions *)
fun coerce (v) = v
end
(* specialization type *)
type ’a t = Rep.t
(* abstract types *)
type ’a AFmla = {fmla: ’a} t
type ’a ALit = {lit: ’a} AFmla
type ’a AAtom = {atom: ’a} ALit
type ’a AConj = {conj: ’a} AFmla
type ’a ADnf = {dnf: ’a} AFmla
(* concrete types *)
type CFmla = unit AFmla
type CLit = unit ALit
type CAtom = unit AAtom
type CConj = unit AConj
type CDnf = unit ADnf
structure Fmla = struct
open Rep
fun dest c {Var=v,Not=n,True=t,And=a,False=f,Or=o} = Rep.dest c (v,n,t,a,f,o)
end
structure Lit = struct
open Rep
fun dest c {Var=v,Not=n} = Rep.dest c (v,n,fail,fail,fail,fail)
end
structure Atom = struct
open Rep
fun dest c {Var=v} = Rep.dest c (v,fail,fail,fail,fail,fail)
end
structure Conj = struct
open Rep
fun dest c {True=t,And=a} = Rep.dest c (fail,fail,t,a,fail,fail)
end
structure Dnf = struct
open Rep
fun dest c {False=f,Or=o} = Rep.dest c (fail,fail,fail,fail,f,o)
end
end
Fig. 2. The Fmla structure
description of the specializations; see Section 5.) Moreover, all the action is
in the signature! The implementation is trivial. Part of the reason for this
explosion of code is that we implement specializations notionally as abstract
types with explicit constructors and destructors. (Section 3 shows how to improve upon this seemingly draconian implementation.) With this in mind, let
us examine the different elements of the signature and their implementation.
Types. The first part of the signature defines the types for the specializations
7
Fluet and Pucella
of formulas. We introduce a polymorphic type ’a t, representing the values
of the specializations.
The first series of type abbreviations combines the abstract encodings of
the specializations with the specialization type, yielding the abstract types.
Consider the definition of the type ALit:
type ’a ALit = {lit: ’a} AFmla (* = {fmla: {lit: ’a}} t *)
Here, {fmla: {lit: ’a}} is the abstract encoding of the specialization lit in
the subtyping hierarchy given above. Note that the sequence of record labels
describes the path through the subtyping hierarchy from the root to the lit
specialization. 4
The second series of type abbreviations instantiates the abstract types
with unit, yielding the concrete types. We can verify that CLit and CAtom
unify with ALit. We can also verify that all concrete types unify with AFmla,
capturing the fact that the fmla specialization is the top element of the subtyping hierarchy. Yet CLit does not unify with AAtom, AConj, or ADnf. In
other words, the encodings respect the subtyping hierarchy.
In the implementation of Figure 2, we see that the type Rep.t is implemented as a bona fide SML datatype, while the type ’a t is implemented as
a type abbreviation whose polymorphic type variable is ignored, but serves as
a placeholder for a position in the subtyping hierarchy. Hence, all the specializations share the same representation. We use opaque signature matching in
Figure 2. This is crucial to get the required behavior for the phantom types.
Constructors. For every specialization, the interface provides a function for
each constructor of the specialization. For instance, the atom specialization
has a single constructor (Var), so we provide a function 5
val Atom.Var : string -> CAtom
that returns an element of the specialization atom (and hence, of type CAtom).
We allow subtyping on the constructor arguments, where appropriate. Hence,
the And constructor for conj is available as:
val Conj.And : ’a ALit * ’b AConj -> CConj
The implementation of these constructors is trivial. They are simply aliases
for the actual constructors of the representation type (brought into scope by
open Rep). Placing them in different structures allows us to constrain their
particular type, depending on the specialization we want them to yield.
Destructors. For every specialization, the interface provides a destructor
4
This is essentially the encoding of tree hierarchies given in our previous work [7]. All of
the other encodings given in that work are also applicable in this setting.
5
In the following and in Section 3, we occasionally replicate declarations in the signature
or structure using long identifiers where the SML syntax requires an (unqualified) identifer.
We do so in order to unambiguously denote the appropriate portion of the code; in all cases,
the meaning should be clear.
8
Fluet and Pucella
function that can be used to simultaneously discriminate and deconstruct elements of the specialization, similar to the manner in which the case expression
operates in SML. Each destructor function takes an element of the specialization as well as functions to be applied to the arguments of the matched
constructor.
As an example, consider Conj.dest, the destructor function for the specialization conj. Because the elements of conj are built using only the True
and And constructors, deconstructing elements of such a specialization can
only yield the True constructor or the And constructor applied to a lit value
and to a conj value. Therefore, we give Conj.dest the type:
val Conj.dest : ’a AConj -> {True : unit -> ’b,
And : CLit * CConj -> ’b} -> ’b
Similar reasoning allows us to drop or refine the types of various branches in
the destructor functions for the other specializations.
Informally, the invariants of the specializations are preserved when we
use concrete encodings in covariant type positions and abstract encodings
in contravariant type positions. This explains the appearance of concrete
encodings in the argument types of the branch functions.
Destructor functions also have a trivial implementation. They are simply
implemented as an SML case expression. On the branches for which no function is provided, we raise an exception. If the invariants of the specializations
are enforced, we know that those exceptions will never be raised! By virtue of
our encoding of subtyping, static typing ensures that this exception will never
be raised by programs that use the interface [7].
Coercions. Finally, the interface provides coercion functions that convert
subtypes to supertypes. Such coercion functions are necessary because, intuitively, phantom types provide only a restricted form of subtyping. To
illustrate the problem, consider the following function; it does not typecheck
because the type of the true branch is CLit and the type of the false branch
is CAtom – two types that cannot be unified:
fun bad b = if b then Lit.Var ("p") else Atom.Var ("q")
Instead, we must write the following:
fun good b = if b then Lit.Var ("p") else Lit.coerce (Atom.Var ("q"))
Technically, this behavior is due to the fact that type subsumption occurs
only at type application (implicit in SML), which most often coincides with
function application. Thus, when two expressions of different specializations
occur in contexts that must have equal types, such as the branches of an if
expression, subsumption does not occur, and the expressions must be coerced
to a common supertype. Coercions are also useful to work around a restriction
in SML that precludes the use of polymorphic recursion [10,11]. We shall
shortly see an example where this use of a coercion is necessary.
The implementation of coercion functions is trivial. They are simply iden9
Fluet and Pucella
tity functions that change the (phantom) type of a value.
2.1 Examples
Let us give a few examples of functions that can be written against the interface of formula specializations given above. First, consider a simple function
to identify the top-level operator of a formula:
fun identify (f:
Fmla.dest f {
Var
= fn _
True = fn _
False = fn _
’a AFmla): string =
=> "variable",
Not = fn _ => "negation",
=> "constant true", And = fn _ => "conjunction",
=> "constant false", Or = fn _ => "disjunction"}
Note that the type of identify, ’a AFmla -> string, asserts that the function may be safely applied to any formula specialization.
A more interesting example is a recursive toString function that returns a
string representation of a formula. A simple implementation is the following:
fun toString (f: ’a AFmla): string
fun toString’ (f: CFmla): string
Fmla.dest f {
Var
= fn s => s,
Not
= fn f => concat ["-",
True = fn () => "T",
And
= fn (f1,f2) => concat
False = fn () => "F",
Or
= fn (f1,f2) => concat
in toString’ (Fmla.coerce f) end
= let
=
toString’ f],
["(", toString’ f1, " & ", toString’ f2, ")"],
["(", toString’ f1, " | ", toString’ f2, ")"]}
Note that the inferred type of the toString’ function is CFmla -> string,
because it is recursively applied to variables of type CFmla in the Fmla.dest
branches and SML does not support polymorphic recursion. However, we
can recover a function that allows subtyping on its argument by composing
toString’ with an explicit coercion. Now the SML type system infers the
desired type for the toString function.
If we had polymorphic recursion, we could directly assign the type ’a
AFmla -> string to toString’. One may ask whether the lack of polymorphic recursion in SML poses a significant problem for the use of specializations
as we have described them. Fortunately, under reasonable assumptions, 6 it
can be shown that use of specializations never needs polymorphic recursion
in an essential way. This is a consequence of the fact that the use of phantom types in specializations only influences the type of an expression or value,
never its representation. The argument proceeds as follows.
Consider a recursive function f with type σ -> τ , for a specialization σ.
Any recursive call in the body of f must be applied to an argument x of
specialization σ ′ , where σ ′ is a subtype of σ. Since σ ′ is a subtype of σ, there
exists a coercion from σ ′ to σ. Hence, we can always implement the function
6
The main assumption is that the desired recursive function on the unspecialized datatype
can itself be written without polymorphic recursion. If this is not the case, then the function
cannot be written in SML with or without phantom types and specializations. (Thus, we
exclude specializations of non-regular datatypes.) On the other hand, if this is the case,
then one can write the function using only the interface provided by specializations.
10
Fluet and Pucella
fun andConjs (f: CConj, g: CConj): CConj =
Conj.dest f {True = fn () => g,
And = fn (f1,f2) => Conj.And (f1, andConjs (f2, g))}
fun orDnfs (f: CDnf, g: CDnf): CDnf =
Dnf.dest f {False = fn () => g,
Or = fn (f1,f2) => Dnf.Or (f1, orDnfs (f2, g))}
fun andConjDnf (f: CConj, g: CDnf): CDnf =
Dnf.dest g {False = fn () => Dnf.False,
Or = fn (g1,g2) => Dnf.Or (andConjs (f, g1), andConjDnf (f, g2))}
fun andDnfs (f: CDnf, g: CDnf): CDnf =
Dnf.dest f {
False = fn () => Dnf.False,
Or = fn (f1,f2) => Dnf.dest g {
False = fn () => Dnf.False,
Or = fn (g1,g2) => Dnf.Or (andConjs (f1, g1),
orDnfs (andConjDnf (f1, g2),
orDnfs (andConjDnf (g1, f2),
andDnfs (f2, g2))))}}
fun litToDnf (f: ’a ALit): CDnf = Dnf.Or (Conj.And (f, Conj.True), Dnf.False)
fun toDnf (f: ’a AFmla): CDnf = let
fun toDnf’ (f: CFmla): CDnf =
Fmla.dest f {
Var = fn s => litToDnf (Atom.Var s),
Not = fn f => Fmla.dest f {
Var = fn s => litToDnf (Lit.Not (Atom.Var s)),
Not = fn f => toDnf’ f,
True = fn () => toDnf’ Fmla.False,
And = fn (f,g) => toDnf’ (Fmla.Or (Fmla.Not f, Fmla.Not g)),
False = fn () => toDnf’ Fmla.True,
Or = fn (f,g) => toDnf’ (Fmla.And (Fmla.Not f, Fmla.Not g))},
True = fn () => Dnf.Or (Conj.True, Dnf.False),
And = fn (f,g) => andDnfs (toDnf’ f, toDnf’ g),
False = fn () => Dnf.False,
Or = fn (f,g) => orDnfs (toDnf’ f, toDnf’ g)}
in toDnf’ (Fmla.coerce f) end
Fig. 3. The toDnf function
as we did for the toString function: set the domain of an auxilary function f’
to the concrete encoding of σ and write the recursive calls as f’ (σ.coerce
x). (In the toString example, all of the recursive calls are on values of type
CFmla, so the coercion can be elided.) Finally, we recover the appropriate
subtyping in the argument of f by composing f’ with σ.coerce.
Figure 3 gives an extended example culminating with a toDnf function that
converts any formula into an equivalent DNF formula. The type of this function, ’a AFmla -> CDnf, ensures that the result formula is a DNF formula.
We further note that the use of type annotations in Figure 3 is completely
superfluous. Type inference will deduce precisely these types.
There is a difference between the guarantee made by the SML type system
and the guarantee made by a specialization library employing the phantomtypes technique. While the former ensures that “a well-typed program won’t
go wrong,” the latter ensures that “a well-typed client won’t go wrong, provided the specialization library is correctly implemented.” We hope that this
section has demonstrated that the implementation of a specialization library
is straightforward. However, a more subtle point is that one must choose specializations that capture properties and invariants of interest in order to gain
the benefits of additional static guarantees.
11
Fluet and Pucella
3
Specializations with Recursion Schemes
While Section 2 describes an interesting theoretical result—the ability to define a toDnf function whose type statically enforces a structural invariant without recourse to separate datatypes—it is not yet clear whether the methodology presented is usable in practice. In particular, pattern matching on specialization types must be performed via application of the destructor functions,
rather than via SML’s built-in syntactic support for pattern matching. The
result is the “truly unreadable code” 7 in Figure 3. While any use of pattern matching can be desugared to applications of the destructor functions,
important aspects of the pattern-matching programming idiom are seriously
inhibited by the encoding in the previous section.
Examining Figure 3 reveals two particularly glaring assaults on readability
that could be improved with pattern matching. First, consider the andDnfs
function. Informally, one can describe the intended behavior of the function in
the following way: consider both arguments in the dnf specialization: if either
argument is a False element, return False; if both arguments are Or elements,
return the disjunction of the pairwise conjunction of the elements’ arguments.
Unfortunately, the written code obscures this behavior. This highlights two
missing aspects of pattern matching: the ability to match simultaneously via
the nesting of datatype patterns within tuple patterns, and the ability to give
a wild-card match.
Second, consider the toDnf’ function, more specifically, the nested application of the caseFmla destructor function. Here, one misses the ability to
write nested patterns, which would combine the two applications of destructor
functions into a single pattern match.
For comparison, Figure 4 implements the toDnf function for the unspecialized Rep.t datatype, addressing the concerns above. However, this apparent
improvement has come at a cost. The compiler now issues multiple nonexhaustive match warnings. More importantly, the type system provides no
assurance that the result is in Disjunctive Normal Form.
We seek a solution that brings the expressiveness and convienence of pattern matching to specializations. Recall the observation that we made in the
introduction: the same static invariants can be obtained by using distinct
datatypes for each specialization and providing functions to convert between
them. As we pointed out, this is inefficient. However, there is a middle-ground
solution, one that uses distinct datatypes for their induced patterns and finegrained coercions to localize coercions to and from the specialization types
and the specialization datatypes.
We take as inspiration Wang and Murphy’s recursion schemes [28]. Exploiting two-level types, which split an inductively-defined type into a component that represents the structure of the type and a component that ties the
7
courtesy of an anonymous reviewer
12
Fluet and Pucella
fun andConjs (f: Rep.t, g: Rep.t): Rep.t =
case f of Rep.True => g
| Rep.And (f1,f2) => Rep.And (f1, andConjs (f2, g))
fun orDnfs (f: Rep.t, g: Rep.t): Rep.t =
case f of Rep.False => g
| Rep.Or (f1,f2) => Rep.Or (f1, orDnfs (f2, g))
fun andConjDnf (f: Rep.t, g: Rep.t): Rep.t =
case g of Rep.False => Rep.False
| Rep.Or (g1,g2) => Rep.Or (andConjs (f, g1), andConjDnf (f, g2))
fun andDnfs (f: Rep.t, g: Rep.t): Rep.t =
case (f, g) of
(Rep.False, _) => Rep.False
| (_, Rep.False) => Rep.False
| (Rep.Or (f1,f2), Rep.Or (g1, g2)) => Rep.Or (andConjs (f1, g1),
orDnfs (andConjDnf (f1, g2),
orDnfs (andConjDnf (g1, f2),
andDnfs (f2, g2))))
fun litToDnf (f: Rep.t): Rep.t = Rep.Or (Rep.And (f, Rep.True), Rep.False)
fun toDnf (f: Rep.t): Rep.t = let
fun toDnf’ (f: Rep.t): Rep.t =
case f of
Rep.Var s => litToDnf (Rep.Var s)
| Rep.Not (Rep.Var s) => litToDnf (Rep.Not (Rep.Var s))
| Rep.Not (Rep.Not f) => toDnf’ f
| Rep.Not Rep.True => Rep.False
| Rep.Not (Rep.And (f,g)) => toDnf’ (Rep.Or (Rep.Not f, Rep.Not g))
| Rep.Not Rep.False => toDnf’ Rep.True
| Rep.Not (Rep.Or (f,g)) =>toDnf’ (Rep.And (Rep.Not f, Rep.Not g))
| Rep.True => Rep.Or (Rep.True, Rep.False)
| Rep.And (f,g) => andDnfs (toDnf’ f, toDnf’ g)
| Rep.False => Rep.False
| Rep.Or (f,g) => orDnfs (toDnf’ f, toDnf’ g)
in toDnf’ f end
Fig. 4. The toDnf function (via the unspecialized Rep.t datatype)
recursive knot, recursion schemes provide a programming idiom that can hide
the representation of an abstract type while still supporting pattern matching.
Roughly speaking, the technique suggests defining a datatype for each
specialization, which represents the top-level structure of the specialization.
Pattern matching on a specialization is performed by first converting part of
the specialization into the appropriate datatype and then matching on the
result. The important point is that we do not need to convert the whole
specialization into the specialization datatype, but rather only as much as is
needed to perform the pattern matching.
Figures 5 and 6 give a signature. and corresponding implementation of
datatypes for the specializations of the formulas. (Again, while the quantity of
code is large, it is largely boilerplate code that can be mechanically generated.)
Note that the signature and implementation are written against the FMLA
signature and Fmla structure; in particular, the FMLA DT signature and FmlaDT
structure do not require access to the representation type Rep.t. Therefore, we
need not make any additional arguments about the safety of using the datatype
interface to the specializations. That is, we cannot violate the invariants
imposed by the specializations by using the datatype interface. As we did
in Section 2, let us examine the different elements of the signature and their
13
Fluet and Pucella
signature FMLA_DT = sig
include FMLA
structure Fmla : sig
(* specialization datatype *)
datatype (’not,’andl,’andr,’orl,’orr) t’ = Var’ of string | Not’ of ’not
| True’ | And’ of ’andl * ’andr
| False’ | Or’ of ’orl * ’orr
(* injection *)
val inj : (’a AFmla, ’b AFmla, ’c AFmla, ’d AFmla, ’e AFmla) t’ -> CFmla
(* projection *)
val prj : ’a AFmla -> (CFmla, CFmla, CFmla, CFmla, CFmla) t’
(* map *)
val map : ((’not1 -> ’not2) *
(’andl1 -> ’andl2) * (’andr1 -> ’andr2) *
(’orll1 -> ’orll2) * (’orr1 -> ’orr2)) ->
(’not1,’andl1,’andr1,’orll1,’orr1) t’ ->
(’not2,’andl2,’andr2,’orll2,’orr2) t’
end
structure Lit : sig
datatype ’not t’ = Var’ of string | Not’ of ’not
val inj : ’a AAtom t’ -> CLit
val prj : ’a ALit -> CAtom t’
val map : (’not1 -> ’not2) -> ’not1 t’ -> ’not2 t’
end
structure Atom : sig
datatype t’ = Var’ of string
val inj : t’ -> CAtom
val prj : ’a AAtom -> t’
val map : t’ -> t’
end
structure Conj : sig
datatype (’andl, ’andr) t’ = True’ | And’ of ’andl * ’andr
val inj : (’a ALit, ’b AConj) t’ -> CConj
val prj : ’a AConj -> (CLit, CConj) t’
val map : ((’andl1 -> ’andl2) * (’andr1 -> ’andr2)) ->
(’andl1,’andr1) t’ -> (’andl2,’andr2) t’
end
structure Dnf : sig
datatype (’orll, ’orr) t’ = False’ | Or’ of ’orll * ’orr
val inj : (’a AConj, ’b ADnf) t’ -> CDnf
val prj : ’a ADnf -> (CConj, CDnf) t’
val map : ((’orll1 -> ’orll2) * (’orr1 -> ’orr2)) ->
(’orll1,’orr1) t’ -> (’orll2,’orr2) t’
end
(* specialization datatype types *)
type (’not,’andl,’andr,’orl,’orr) DFmla = (’not,’andl,’andr,’orl,’orr) Fmla.t’
type ’not DLit = ’not Lit.t’
type DAtom = Atom.t’
type (’andl,’andr) DConj = (’andl,’andr) Conj.t’
type (’orl,’orr) DDnf = (’orl,’orr) Dnf.t’
end
Fig. 5. The FMLA DT signature 8
implementation.
Datatypes. The first part of each substructure defines the datatype used to
represent a specialization. The datatype for a specialization has a datatype
8
The signature given for FMLA DT is not valid SML, in that the substructures (Fmla, Lit,
etc.) are extended. While this abuse of notation seems acceptable in an exposition, an
implementation must textually duplicate and extend the FMLA signature. This argues that
SML could benefit from a more expressive language of signatures [21].
14
Fluet and Pucella
structure FmlaDT : FMLA_DT = struct
open Fmla
structure Fmla = struct
open Fmla
(* specialization datatype *)
datatype (’not,’andl,’andr,’orl,’orr) t’ = Var’ of string | Not’ of ’not
| True’ | And’ of ’andl * ’andr
| False’ | Or’ of ’orl * ’orr
(* injection *)
fun inj f =
case f of Var’ s => Var s | Not’ f => Not f
| True’ => True | And’ (f1, f2) => And (f1, f2)
| False’ => False | Or’ (f1, f2) => Or (f1, f2)
(* projection *)
fun prj f =
dest f {Var = Var’, Not = Not’,
True = fn () => True’, And = And’,
False = fn () => False’, Or = Or’}
(* map *)
fun map (F1,F2,F3,F4,F5) f =
case f of Var’ s => Var’ s | Not’ f => Not’ (F1 f)
| True’ => True’ | And’ (f1, f2) => And’ (F2 f1, F3 f2)
| False’ => False’ | Or’ (f1, f2) => Or’ (F4 f1, F5 f2)
end
structure Lit = struct
open Lit
datatype ’not t’ = Var’ of string | Not’ of ’not
fun inj f = case f of Var’ s => Var s | Not’ f => Not f
fun prj f = dest f {Var = Var’, Not = Not’}
fun map F f = case f of Var’ s => Var’ s | Not’ f => Not’ (F f)
end
structure Atom = struct
open Atom
datatype t’ = Var’ of string
fun inj f = case f of Var’ s => Var s
fun prj f = dest f {Var = Var’}
fun map f = case f of Var’ s => Var’ s
end
structure Conj = struct
open Conj
datatype (’andl, ’andr) t’ = True’ | And’ of ’andl * ’andr
fun inj f = case f of True’ => True | And’ (f1, f2) => And (f1, f2)
fun prj f = dest f {True = fn () => True’, And = And’}
fun map (F1,F2) f = case f of True’ => True’ | And’ (f1, f2) => And’ (F1 f1, F2 f2)
end
structure Dnf = struct
open Dnf
datatype (’orll, ’orr) t’ = False’ | Or’ of ’orll * ’orr
fun inj f = case f of False’ => False | Or’ (f1, f2) => Or (f1, f2)
fun prj f = dest f {False = fn () => False’, Or = Or’}
fun map (F1,F2) f = case f of False’ => False’ | Or’ (f1, f2) => Or’ (F1 f1, F2 f2)
end
(* specialization datatype types *)
type (’not,’andl,’andr,’orl,’orr) DFmla = (’not,’andl,’andr,’orl,’orr) Fmla.t’
type ’not DLit = ’not Lit.t’
type DAtom = Atom.t’
type (’andl,’andr) DConj = (’andl,’andr) Conj.t’
type (’orl,’orr) DDnf = (’orl,’orr) Dnf.t’
end
Fig. 6. The FmlaDT structure
15
Fluet and Pucella
constructor for each constructor in the specialization. Wherever a constructor
has a specialization as an argument, we introduce a type variable, creating a
polymorphic datatype. The polymorphic type allows the datatype to represent unfoldings of the specialization with encoded specialization types at the
“leaves” of the structure. For instance, the dnf specialization
withspec dnf = False | Or of conj * dnf
becomes
datatype (’orl, ’orr) Dnf.t’ = Dnf.False’ | Dnf.Or’ of ’orl * ’orr
replacing the references to the specializations conj and dnf with the polymorphic type variables ’orl and ’orr. Thus, the structure of a dnf specialization
is given without reference to specific types for conj or dnf.
The polymorphic type variables allow the specialization datatype to represent arbitrary, finite unrollings of the specialization type. For example, the
type (CConj, CDnf DDnf) DDnf corresponds to unrolling the dnf specialization into the Dnf.t’ datatype once at the top-level and once again at the
second argument to the Or constructor. Hence, it has as elements Dnf.False’,
Dnf.Or’ (e1, Dnf.False’), and Dnf.Or’ (e1, Dnf.Or’ (e2, e3)) for
any e1 and e2 of type CConj and e3 of type CDnf.
Injections. For every specialization, the interface provides a function for
coercing from the specialization datatype to the specialization type. In particular, the coercion is from one top-level unrolling of the specialization type (into
the specialization datatype) back to the specialization type. The implementation is straightforward: map each datatype constructor to the corresponding
specialization constructor. For the Dnf.t’ datatype, this yields
fun Dnf.inj f =
case f of Dnf.False’ => Dnf.False
| Dnf.Or’ (f1, f2) => Dnf.Or (f1, f2)
with the type (’a AConj, ’b ADnf) DDnf -> CDnf. We find injections to
be used infrequently in practice. One usually wishes to build values at the
specialization type, and the specialization constructors are better suited for
this purpose than injecting from the specialization datatype. Hence, we do
not use injections in our examples.
Projections. Much more practical are the projection functions, which convert from a specialization type to the specialization datatype. Again, the
conversion is from the specialization type to a single top-level unrolling of the
specialization type (into the specialization datatype). The implementation
builds upon the destructor functions, by mapping each dispatch function to
the corresponding datatype constructor. For the DDnf datatype, this yields
fun Dnf.prj f =
Dnf.dest f {False=fn () => Dnf.False’,
Or = fn (f1,f2) => Dnf.Or’ (f1,f2)}
16
Fluet and Pucella
fun andConjs (f: CConj, g: CConj): CConj =
case Conj.prj f of Conj.True’ => g
| Conj.And’ (f1,f2) => Conj.And (f1, andConjs (f2, g))
fun orDnfs (f: CDnf, g: CDnf): CDnf =
case Dnf.prj f of Dnf.False’ => g
| Dnf.Or’ (f1,f2) => Dnf.Or (f1, orDnfs (f2, g))
fun andConjDnf (f: CConj, g: CDnf): CDnf =
case Dnf.prj g of Dnf.False’ => Dnf.False
| Dnf.Or’ (g1, g2) => Dnf.Or (andConjs (f, g1), andConjDnf (f, g2))
fun andDnfs (f: CDnf, g: CDnf): CDnf =
case (Dnf.prj f, Dnf.prj g) of
(Dnf.False’, _) => Dnf.False
| (_, Dnf.False’) => Dnf.False
| (Dnf.Or’ (f1,f2), Dnf.Or’ (g1,g2)) => Dnf.Or (andConjs (f1, g1),
orDnfs (andConjDnf (f1, g2),
orDnfs (andConjDnf (g1, f2),
andDnfs (f2, g2))))
fun litToDnf (f: ’a ALit): CDnf = Dnf.Or (Conj.And (f, Conj.True), Dnf.False)
fun toDnf (f: ’a AFmla): CDnf = let
fun toDnf’ (f: CFmla): CDnf =
case Fmla.map (Fmla.prj, id, id, id, id) (Fmla.prj f) of
Fmla.Var’ s => litToDnf (Atom.Var s)
| Fmla.Not’ (Fmla.Var’ s) => litToDnf (Lit.Not (Atom.Var s))
| Fmla.Not’ (Fmla.Not’ f) => toDnf’ f
| Fmla.Not’ Fmla.True’ => toDnf’ Fmla.False
| Fmla.Not’ (Fmla.And’ (f,g)) => toDnf’ (Fmla.Or (Fmla.Not f, Fmla.Not g))
| Fmla.Not’ Fmla.False’ => toDnf’ Fmla.True
| Fmla.Not’ (Fmla.Or’ (f,g)) => toDnf’ (Fmla.And (Fmla.Not f, Fmla.Not g))
| Fmla.True’ => Dnf.Or (Conj.True, Dnf.False)
| Fmla.And’ (f,g) => andDnfs (toDnf’ f, toDnf’ g)
| Fmla.False’ => Dnf.False
| Fmla.Or’ (f,g) => orDnfs (toDnf’ f, toDnf’ g)
in toDnf’ (Fmla.coerce f) end
Fig. 7. The toDnf function (via datatype interface)
with the type ’a ADnf -> (CConj, CDnf) DDnf.
Maps. One final useful family of functions are the structure-preserving maps.
These functions, similar in flavor to the familiar map on polymorphic lists,
apply a function to each polymorphic element of a structure, but otherwise
leave the structure’s “shape” intact. For instance, the map
fun Lit.map F f =
case f of Lit.Var’ s => Lit.Var’ s
| Lit.Not’ f => Lit.Not’ (F f)
transforms an ’not1 DLit to an ’not2 DLit via a function F of type ’not1
-> ’not2. Since the polymorphic elements of a specialization datatype correspond to the nested specializations used by that datatype, maps are useful
for localizing unfoldings of a specialization for nested pattern matching.
3.1 Example
Figure 7 reproduces the code from Figure 3 using the datatype interface to the
formula specializations. While some may argue that the syntactic differences
are minor, the differences are important ones. All of the functions are written
in a familiar, readable pattern-matching style. In particular, the andDnfs
function uses simultaneous pattern matching and wild-card matches. Also, the
17
Fluet and Pucella
toDnf’ function uses nested patterns to fold all the branches into a single case
expression. Note the use of the Fmla.map function to unfold only the CFmla
under the Fmla.Not’ constructor, while leaving all other fmla specializations
folded. The expression discriminated by the case has the following type:
((CFmla, CFmla, CFmla, CFmla, CFmla) DFmla,
CFmla, CFmla, CFmla, CFmla) DFmla
Finally, a word about the efficiency of the compiled code. In the presence of cross module inlining, smart representation decisions, and some local
constant folding, the overhead of projections from a specialization type to a
specialization datatype can be almost entirely eliminated. Specifically, when
projections appear directly as the expression discriminated by a case, then a
compiler can easily fold the case expression “buried” in the projection function into the outer case expression.
4
Discussion
We discuss the relationship between our implementation of specializations
and related approaches and techniques in the literature. We also discuss some
limitations, and how they could be lifted.
Refinement Types. As we have already mentioned, refinement types [8]
enforce invariants similar in spirit to (and strictly more expressive than) specializations. In fact, many of the examples considered here were inspired
by similar examples expressed using refinement types. However, there are a
number of critical differences between refinement types and specializations as
described and implemented in this paper. These differences permit us to encode specializations directly in the SML type system. The most significant
difference concerns the “number” of types assigned to a value. In short, refinement types use a limited form of type intersection to assign a value multiple
types, each corresponding to the evaluation of the value at specific refinements,
while our technique assigns every value exactly one type. 9 For example, in
Section 2.1, we assigned the litToDnf function the (conceptual) type lit ->
dnf. With refinement types, it would be assigned the type (fmla -> fmla)
∧ (lit -> dnf), indicating that in addition to mapping literals to DNF formulae, the function can also be applied to an arbitrary formula, although the
resulting formula will not satisfy any of the declared refinements.
While this demonstrates the expressiveness of refinement types, it does not
address the utility of this expressiveness. In particular, one rarely works in
a context where all possible typings of an expression are necessary. In fact,
the common case, particularly with data-structure invariants, is a context
where exactly one type is of interest: a “good” structure is either produced
or preserved. This is exactly the situation that motivates our examples of
9
Both systems in fact employ a form of subtyping to further increase the “number” of
types assigned to a value.
18
Fluet and Pucella
specializations. In this sense, our technique is closer in spirit to refinementtype checking [3], which verifies that an expression satisfies the user-specified
refinement types.
Deforestation. We pointed out in the introduction and again in Section 3
that the same static invariants that we obtain through the use of specializations can be obtained by using distinct datatypes for each specialization and
explicit conversion functions between the datatypes. We also pointed out that
this approach is inefficient. Deforestation is a compiler optimization that aims
to reduce the allocation of intermediate structures that are created when using
functions that operate on data structures such as lists and trees [26]. It is quite
possible that a judicious use of such techniques can alleviate the inefficiency
inherent in using multiple datatypes for specialization. While one would gain
more readable case expressions, avoiding the need for projections, one would
lose the subsumption of function arguments available through the phantomtypes technique. Deforestation suffers from the same criticism as refinement
types: it is not widely implemented in compilers, particularly in languages like
SML, where strict evaluation and effects often invalidate the optimization. In
constrast, our approach is applicable independent of the underlying compiler
technology.
Phantom Types. The phantom-types technique underlies many interesting uses of type systems. It has been used to derive early implementations
of extensible records [27,22,2], to provide a safe and flexible interface to the
Network Socket API [23], to interface with COM components [6], to type
embedded compiler expressions [13,4], to record sets of effects in type-andeffect type systems [19], to embed a representation of the C type system in
SML [1], to guarantee conforming XHTML documents are produced by SML
scriplets [5], and to provide a safe interface to GUI widgets [12]. The application to type-embedded compiler expressions and to guarantee the production
of conforming XHTML documents are most closely related to our specialization technique. Many of the other applications focus on enforcing a required
external structure on an essentially unstructured internal representation (e.g.,
the SML representation of a socket is just a 32-bit integer). Our specialization technique extends the idea to enforce internal invariants on a common
internal representation.
Recursion Schemes. While the implementation described in Section 3 is
inspired by recursion schemes [28], there are significant differences. Recursion schemes are designed to support monomorphic recursive types, with a
straightforward extension to polymorphic recursive types. However, it is less
clear how to extend the idiom to mutually recursive types, which may arise in
specializations. Hence, we have chosen to allow each reference to a specialization to be typed independently. Clearly, we do not wish to identify conj and
dnf in the dnf specialization, because the two specializations are distinct. It
19
Fluet and Pucella
is debatable whether all occurences of the fmla specialization within the fmla
specialization should be identified; that is, whether we should introduce
datatype (’not,’andl,’andr,’orl,’orr) Fmla.t’ =
Var’ of string | Not’ of ’not
| True’ | And’ of ’andl * ’andr
| False’ | Or’ of ’orll * ’orr
or
datatype ’fmla Fmla.t’ =
Var’ of string | Not’ of ’fmla
| True’ | And’ of ’fmla * ’fmla
| False’ | Or’ of ’fmla * ’fmla
The latter is closer to recursion schemes (and, hence, could be given a simple
categorical interface [28]). For pragmatic reasons, we have instead adopted
the former, because it gives finer grained control over coercions to and from
the datatype. However, as our entire implementation does not require access
to the representation type, either or both definitions could be used without
difficulty.
Extensions. We have made a number of implicit and explicit restrictions
to simplify the treatment of specializations. There are a number of ways of
relaxing these restrictions that result in more expressive systems. For example,
we can allow a specialization to use the same constructor at multiple argument
types. Consider defining a DNF formula to be a list of Or-ed conjuctions that
grows to either the left or the right:
withspec dnf = False | Or of conj * dnf | Or of dnf * conj
One can easily define two constructor functions for Or that inject into the dnf
specialization. However, the “best” destructor function that one can write is:
val Dnf.dest : ’a ADnf -> {False : unit -> ’b,
Or : CFmla * CFmla -> ’b} -> ’b
not, as might be expected,
val Dnf.dest : ’a ADnf -> {False : unit -> ’b,
Or : (CConj * CDnf -> ’b) *
(CDnf * CConj -> ’b)} -> ’b
While the second function can be written with the expected semantics, it
requires a run-time inspection of the arguments to the Or constructor to distinguish between a conj and a dnf. We do not consider this implementation
to be in the spirit of a primitive case expression, because sufficiently complicated specializations could require non-constant time to execute a destructor
function. Instead, the first function corresponds to the least upper bound of
conj * dnf and dnf * conj in the upper semi-lattice (i.e., the subtyping hierarchy) induced by the specializations. The loss of precision in the resulting
type corresponds to the fact that no specialization exactly corresponds to the
union of the conj and dnf specializations. We could gain some precision by
20
Fluet and Pucella
introducing such a specialization, at the cost of complicating the interface.
In the examples discussed previously, we have restricted ourselves to monomorphic, first-order datatypes. This restriction can be relaxed to allow
polymorphic, first-order datatypes in the obvious manner: the specialization
type makes use of both non-phantom and phantom type variables, the nonphantom type variables being applied to the representation type. For example,
the specializations given in Section 1 that distinguish between lists of even and
odd length would induce the following signature:
signature EVENODDLIST = sig
type (’a, ’b) t
type (’a, ’b) AList = (’a, {list: ’b}) t
type (’a, ’b) AEven = (’a, {even: ’b}) AList
type (’a, ’b) AOdd = (’a, {odd: ’b}) AList
...
end
and imlementation:
structure EvenOddList :> EVENODDLIST = struct
structure Rep = struct
datatype ’a t = Nil | Cons of ’a * ’a t
...
end
type (’a, ’b) t = ’a Rep.t
type (’a, ’b) AList = (’a, {list: ’b}) t
type (’a, ’b) AEven = (’a, {even: ’b}) AList
type (’a, ’b) AOdd = (’a, {odd: ’b}) AList
...
end
We may also extend the technique to higher-order datatypes, although
this extension requires the specialization subtyping hierarchy to induce a full
lattice of types (rather than an upper semi-lattice), due to the contravariance
of function arguments. There are additional restrictions on the subtyping
relation at function types; see our previous work for more details [7].
One final limitation of the procedure described here is that it applies to a
single datatype. At times, it may be desirable to consider specializations of one
datatype in terms of the specializations of another datatype. The technique
described in this paper can be extended to handle this situation by processing
all of the specialized datatypes simultaneously. Although this decreases the
modularity of a project, this is required in order to define each representation
type in terms of other representation types, which otherwise would be hidden
by the opaque signatures.
It is worth pointing out that some limitations of our implementation technique are due to the encodings of the subtyping hierarchy induced by the specializations. So long as the encodings respect the hierarchy [7], the techniques
described in this paper are completely agnostic to the specifics of the encoding. However, if the encodings of the subtyping hierarchy possess properties
beyond respecting the hierarchy, these properties may be used to provide a
21
Fluet and Pucella
signature BITS = sig
(* specialization type *)
type ’a t
signature BITS = sig
(* specialization type *)
type ’a t
(* abstact and concrete types *)
type ’a ABits = {bits: ’a} t
type ’a AEven = {even: ’a} ABits
type ’a AOdd = {odd: ’a} ABits
type CBits = unit ABits
type CEven = unit AEven
type COdd = unit AOdd
structure Bits : sig
val Nil : CBits
val Zero : ’a ABits -> CBits
val One : ’a ABits -> CBits
...
end
structure Even : sig
val Nil : CEven
val Zero : ’a AEven -> CEven
val One : ’a AOdd -> CEven
...
end
structure Odd : sig
val Zero : ’a AOdd -> COdd
val One : ’a AEven -> COdd
...
end
end
(a) The BITS signature
(* abstact and concrete types *)
type (’a, ’b) ABits = {bits: (’a * ’b)} t
type (’a, ’b) AEven = ({z: ’a}, ’b) ABits
type (’a, ’b) AOdd = (’a, {z: ’b}) ABits
type CBits = (unit, unit) ABits
type CEven = (unit, unit) AEven
type COdd = (unit, unit) AOdd
(* general constructors *)
val Zero : (’a, ’b) t -> (’a, ’b) t
val One : (’a, ’b) t -> (’b, ’a) t
structure Bits : sig
val Nil : CBits
val Zero : ’a ABits -> CBits
val One : ’a ABits -> CBits
...
end
structure Even : sig
val Nil : CEven
val Zero : ’a AEven -> CEven
val One : ’a AOdd -> CEven
...
end
structure Odd : sig
val Zero : ’a AOdd -> COdd
val One : ’a AEven -> COdd
...
end
end
(b) An alternative BITS signature
Fig. 8.
more flexible implementation of specializations. The following example should
give a flavor of the kind of flexibility we have in mind.
Consider a datatype of bit strings, and specializations that capture the
parity of bit strings:
datatype bits
= Nil | Zero of bits | One of bits
withspec even = Nil | Zero of even | One of odd
and odd = Zero of odd | One of even
Following the approach described in this paper, it is straightforward to derive
an interface BITS to the specializations; see Figure 8(a). An analysis of the
specializations and implementation reveals that we could safely define a single
Zero constructor that applies to all specializations, with type ’a t -> ’a t.
Unfortunately, we cannot similarly define a single One constructor.
However, if we choose the encoding of the subtyping hierarchy carefully, we
can in fact come up with an alternative interface BITS to bit strings and their
specializations that allows for the definition of a single one constructor; see
Figure 8(b). One can verify that the concrete and abstract encodings respect
the induced subtyping hierachy. However, they also satisfy a symmetry that
22
Fluet and Pucella
makes it possible to write single instances of the Zero and One constructors
that apply to all specializations. In particular, note that the type of the One
function makes it explicit that the parity of the bit string is flipped.
Formalizing Results. In Section 2, we noted that a fully elaborated example of our specialization technique is more instructive than a formal translation. However, it is worth considering how we could formalize the technique,
and what results we would obtain. First, we would define a core calculus that
captures the essence of specializations; such a calculus would include a primitive specialization type with its specializations, constructors into and destructors from the specializations, and a type system with the induced subtying
relation. We would prove the type safety of the core calculus, a consequence
of which is that a well-typed program may never attempt to apply a destructor to an inappropriate specialization. We would next define a type-directed
translation of this core calculus into SML. Finally, we would show that this
translation preserves both the type and the meaning of expressions, the latter by a simulation argument whereby the runtime behavior of a well-typed
core calculus program is simulated by the translated SML program. Since
the translation preserves meaning, the translated program will never raise the
Match exception—that runtime behavior has no analogue in the core calculus.
In fact, the type-preservation result is a special case of the more general
result stating that phantom types may be used to capture a particular notion
of subtyping [7], where we take the specialization constructors and destructors
as primitive operations. Under the assumption that the operational behavior
of the constructors and destructors is sound with respect to their assigned
specialization types, we obtain a result that ensures the absence of runtime
exceptions arising from the use of the primitive operations. What the current
work adds is the fact that, for datatype specializations, we need not take
the constructors and destructors as primitive operations in the translation;
instead, we may directly encode them in SML.
5
Conclusion
We applied two programming techniques, the phantom-types technique [7] and
the recursion-schemes technique [28], to the problem of capturing structural
invariants in user-defined datatypes. By modeling an abstract datatype as a
collection of constructor, destructor, and coercion functions, we can define an
implementation of datatype specializations using the techniques developed in
this paper. We further described methods by which the clumsy destructor
functions can be replaced by familiar pattern matching by injecting into and
projecting from datatypes inspired by recursion schemes.
We have collected a set of interesting examples of datatype specializations at http://www.cs.cornell.edu/People/fluet/specializations, all
using the techniques of this paper. We have also begun work on a tool to me23
Fluet and Pucella
chanically generate an implementation from the concise datatype/withspec
declaration. A proof-of-concept version of the tool may be obtained at the
same website.
Acknowledgments. Thanks to Simon Peyton Jones who provided feedback
on the application of phantom types to datatype specializations. The third
anonymous reviewer suggested using substructures as an organizing principle,
rather than ad-hoc name mangling.
References
[1] Blume, M., No-Longer-Foreign: Teaching an ML compiler to speak C “natively”,
in: Proc. BABEL’01, Electronic Notes in Theoretical Computer Science 59.1
(2001).
[2] Burton, F., Type extension through polymorphism, ACM Transactions on
Programming Languages and Systems 12 (1990), pp. 135–138.
[3] Davies, R., “Practical Refinement-Type Checking,” Ph.D. thesis, School of
Computer Science, Carnegie Mellon University (2005), available as Technical
Report CMU-CS-05-110.
[4] Elliott, C., S. Finne and O. de Moor, Compiling embedded languages,
in: Workshop on Semantics, Applications, and Implementation of Program
Generation, 2000.
[5] Elsman, M. and K. Larsen, Typing XHTML Web applications in ML,
in: Proc. 7th International Symposium on Practical Aspects of Declarative
Languages (PADL’04), Lecture Notes in Computer Science 3350 (2004), pp.
224–238.
[6] Finne, S., D. Leijen, E. Meijer and S. Peyton Jones, Calling hell from heaven
and heaven from hell, in: Proc. 4th ACM SIGPLAN International Conference
on Functional Programming (ICFP’99) (1999), pp. 114–125.
[7] Fluet, M. and R. Pucella, Phantom types and subtyping, in: Proc. 2nd IFIP
International Conference on Theoretical Computer Science (TCS’02) (2002),
pp. 448–460.
[8] Freeman, T., “Refinement types for ML,” Ph.D. thesis, School of Computer
Science, Carnegie Mellon University (1994), available as Technical Report
CMU-CS-94-110.
[9] HaMLet, http://www.ps.uni-sb.de/hamlet.
[10] Henglein, F., Type inference with polymorphic recursion, ACM Transactions on
Programming Languages and Systems 15 (1993), pp. 253–289.
[11] Kfoury, A., J. Tiuryn and P. Urzyczyn, Type reconstruction in the presence
of polymorphic recursion, ACM Transactions on Programming Languages and
Systems 15 (1993), pp. 290–311.
24
Fluet and Pucella
[12] Larsen, K. and H. Niss, mGTK: An SML binding of Gtk+, in: Proc. USENIX
Annual Technical Conference, 2004, pp. 127–134.
[13] Leijen, D. and E. Meijer, Domain specific embedded compilers, in: Proc. 2nd
Conference on Domain-Specific Languages (DSL’99), 1999, pp. 109–122.
[14] Milner, R., A theory of type polymorphism in programming, Journal of
Computer and Systems Sciences 17 (1978), pp. 348–375.
[15] Milner, R., M. Tofte, R. Harper and D. MacQueen, “The Definition of Standard
ML (Revised),” MIT Press, 1997.
[16] ML Kit, http://www.itu.dk/research/mlkit.
[17] MLton Standard ML compiler, http://www.mlton.org.
[18] Moscow ML, http://www.dina.dk/~sestoft/mosml.html.
[19] Pessaux, F. and X. Leroy, Type-based analysis of uncaught exceptions, in:
Proc. 26th Annual ACM Symposium on Principles of Programming Languages
(POPL’99) (1999), pp. 276–290.
[20] Poly/ML, http://www.polyml.org.
[21] Ramsey, N., K. Fisher and P. Govereau, An expressive language of signatures,
in: Proc. 10th ACM SIGPLAN International Conference on Functional
Programming (ICFP’05) (2005).
[22] Rémy, D., Records and variants as a natural extension of ML, in: Proc. 16th
Annual ACM Symposium on Principles of Programming Languages (POPL’89)
(1989), pp. 77–88.
[23] Reppy, J. H., A safe interface to sockets, Technical memorandum, AT&T Bell
Laboratories (1996).
[24] SML.NET compiler, http://www.cl.cam.ac.uk/Research/TSG/SMLNET.
[25] Standard ML of New Jersey, http://www.smlnj.org.
[26] Wadler, P., Deforestation: Transforming programs
Theoretical Computer Science 73 (1990), pp. 231–248.
to eliminate trees,
[27] Wand, M., Complete type inference for simple objects, in: Proc. 2nd Annual
IEEE Symposium on Logic in Computer Science (LICS’87), 1987.
[28] Wang, D. C. and T. Murphy VII, Programming with recursion schemes (2003),
available from http://www-2.cs.cmu.edu/~tom7/papers.
25
| 6cs.PL
|
MassChroQ : A versatile tool for mass spectrometry quantification
Benoit Valot1, Olivier Langella2, Edlira Nano1, Michel Zivy2.
1
INRA, Plateforme d'Analyse Protéomique de Paris Sud-Ouest, UMR 0320/UMR 8120 de
Génétique Végétale, Gif sur Yvette, France.
2
CNRS, PAPPSO, Plateforme d'Analyse Protéomique de Paris Sud-Ouest, UMR 0320/UMR 8120 de
Génétique Végétale, Gif sur Yvette, France.
Correspondence :
Dr Benoît Valot, PAPPSO, UMR de Génétique Végétale, INRA/Université Paris-Sud/
CNRS/AgroParisTech, Ferme du Moulon, F-91190 Gif-sur-Yvette, France.
Email : [email protected]
Fax : +33 1 69 33 23 40
Abbreviations :
MassChroQ, Mass Chromatogram Quantification; HR, High Resolution; LR, Low Resolution; RT,
Retention Time; XIC, eXtracted Ion Chromatogram
Keywords :
Bioinformatics, Extracted Ion Chromatogram, Mass spectrometry, Proteomics , Quantification,
Software
Number of words : 2773
Abstract
Recently, many software tools have been developed to perform quantification in LC-MS analyses.
However, most of them are specific to either a quantification strategy (e.g. label-free or isotopic
labelling) or a mass-spectrometry system (e.g. high or low resolution).
In this context, we have developed MassChroQ, a versatile software that performs LC-MS data
alignment and peptide quantification by peak area integration on extracted ion chromatograms.
MassChroQ is suitable for quantification with or without labelling and is not limited to high
resolution systems. Peptides of interest (for example all the identified peptides) can be determined
automatically or manually by providing targeted m/z and retention time values. It can handle large
experiments that include protein or peptide fractionation (as SDS-PAGE, 2D-LC). It is fully
configurable. Every processing step is traceable, the produced data are in open standard format and
its modularity allows easy integration into proteomic pipelines. The output results are ready for use
in statistical analyses.
Evaluation of MassChroQ on complex label-free data obtained from low and high resolution mass
spectrometers showed low CVs for technical reproducibility (1.4%) and high coefficients of
correlation to protein quantity (0.98).
MassChroQ
is
freely
available
under
http://pappso.inra.fr/bioinfo/masschroq/.
the
GNU
General
Public
Licence
v3.0
at
In the last years, the continuous improvement of quantitative mass spectrometry methods has
opened new perspectives in proteomics. The amount and complexity of the data to be processed has
grown, evidencing the need for new automatic computational tools. While spectral counting has
been used for semi-quantitative analysis, quantitative experiments are mostly based on
quantification of the MS signal. Several tools using various algorithms have been developed (e.g [13]). These tools handle the different steps of the analysis : signal denoising, peak detection, peak
area measurement, de-isotoping, LC-MS runs alignment, etc. However, most of them deal only with
a specific problem or type of data, and are for example restricted to high-resolution (HR)
spectrometers, to isotope labelling or to label-free quantification. In addition, they often present
platform specificities, proprietary data format dependencies and do not allow integration in
proteomics pipelines like TPP [4] or TOPP [5].
We have developed MassChroQ (which stands for Mass Chromatogram Quantification) with the
aim of being as experiment-independent as possible, while being able to take into account complex
experimental designs. MassChroQ processes quantification data from their rough state to a form
ready to be used by statistical software. It is fully configurable and every step of the analysis is
traceable. MassChroQ allows the user to: i) process data obtained from spectrometers with various
levels of resolution; ii) analyse label-free as well as isotopic labelling experiments; iii) analyse
experiments in which samples were fractionated prior to LC-MS analysis (as in SDS-PAGE, SCX,
etc.); iv) time-efficiently process a large number of samples.
Low-resolution (LR) instruments (e.g. LTQ ion traps) can provide valuable quantitative data from
samples of low or medium complexity. In order to be able to quantify data obtained from LR as
well as from HR instruments (e.g. Orbitrap), we chose a quantification method based on eXtracted
Ion Chromatograms (XIC) rather than on feature detection on the virtual 2D image (e.g algorithms
derived from 2D gels analysis or “peak picking” [6-10]). Indeed, the latter needs high resolution in
MS mode to identify isotopic profiles. By contrast, quantification based on XICs is obtained by
extracting the intensity corresponding to the m/z of the selected peptides along the LC-MS run, and
by integrating the peak area at their retention time (RT). This strategy can be used with LR as well
as HR mass spectrometers by adapting the window size of XIC extraction. It can be used with labelfree (e.g. [11,12] as well as isotopic methods (e.g. SILAC, ICAT, N15, [13-15]).
The main features of MassChroQ are :
i) Determination of peptides to be quantified. If an experiment includes MS/MS acquisition for
identification, the identified peptides and protein descriptions can be provided to MassChroQ. He
will then automatically quantify them in all samples, including those where the peptides were not
identified. Peptides can also be specified by providing a list of m/z or m/z-RT values. If isotopic
labelling was performed, the different labels can be described by specifying the modified sites (e.g.
amino acids, peptide N- or C-terminal) and the mass shifts.
ii) XIC extraction, peak detection and quantification. XICs of peptides of interest are extracted from
the original data file. Filters are used to correct baselines or to remove artefactual spikes (Fig S1).
XICs are then smoothed with an average filter before performing a closing and an opening
mathematical morphology operation with a small flat structuring element [16]. The closing
operation eliminates thin valleys and conserves the intensity of local maxima, while the opening
operation eliminates thin peaks (i.e. remaining spikes) and conserves the intensity of local minima.
Hence, detection of peak positions is performed on the closed profile, and the opened profile is used
to eliminate remaining spikes (Fig. S2). The peak boundaries are searched on the closed profile, and
the peak area (i.e. the quantification value) is computed on the unaltered XIC, by integrating the
intensity between these boundaries.
iii) Alignment and peak matching. As distortions can occur between LC-MS runs, MS RTs must be
aligned before matching. Two alignment methods are proposed in MassChroQ: the OBI-Warp
(Ordered Bijective Interpolated Warping) alignment method [17] which is based on MS data only,
and an in-house MS/MS alignment method. The latter uses common MS/MS identifications as
landmarks to evaluate time deviation (ΔMS) along the chromatography. More precisely, the ΔMS/MS
difference between the MS/MS RT in the run to align and the MS/MS RT in the run chosen as a
reference is computed for each common peptide. Then, for each MS RT of the run to align, ΔMS is
evaluated by linear interpolation between the ΔMS/MS values of its two closest surrounding MS/MS
RTs. Both ΔMS/MS and ΔMS data are smoothed before use (with average and median filters) to
eliminate low-scale RT heterogeneities. After alignment, peak matching is performed as follows: the
quantitative value of a peak is assigned to an identified peptide if and only if the MS/MS RT of this
peptide is within the boundaries of this peak. Of course, only similar LC-MS runs should be
aligned. For example, if samples were fractionated by SCX, only LC-MS runs from the same SCX
fraction should be aligned. To take this into account the user defines groups of LC-MS runs that can
be compared to each other. Alignment and peak matching will be performed only within these
groups. If necessary, alignment and quantification methods can be specifically defined for each
group.
To evaluate MassChroQ performances, we prepared 6 samples made each of 700 ng of the same
total protein digest of Saccharomyces cerevisiae, spiked with 6 different amounts of BSA digest
(4.5, 15, 45, 105, 450 and 1500 fmol). These samples were analysed with an LR and an HR system
(respectively a Thermo-Fisher LTQ XL coupled to an Eksigent 2D-ultra-nanoLC, and a ThermoFisher Orbitrap Discovery coupled to a Dionex U3000 nanoLC; see supplement materials). All runs
included MS and MS/MS acquisition. Two groups of runs were defined in MassChroQ to separate
the six LR runs from the six HR runs. The alignment was performed by using the MS/MS alignment
method (Fig. 1A). Since spectrometers do not trigger MS/MS at the exact RT of peptide peaks,
MS/MS RTs showed a non negligible dispersion. However, data points were numerous enough to
allow the computation of the tendency deviation curve along the reference LC, that was used by the
alignment algorithm to correct RTs . The standard deviation of peptide RT was clearly reduced by
the alignment in both LR and HR systems (Fig. 1B and 1C), showing its efficiency. Although LCMS runs showed small deviations before alignment, the alignment significantly impacted data
quality affecting the matching of 5% of the peaks (data not shown).
XIC extraction was performed with an m/z window of 0.3 Th for LR data and 10 ppm for HR data.
All identified peptides were selected for quantification. Combining all LR and HR LC-MS/MS
runs, 5831 different peptide sequences allowed the identification of 556 proteins (with a false
discovery rate of 0.3%), distributed in 492 groups of proteins sharing at least one peptide. A total of
5936 and 2467 XICs were extracted from respectively LR and HR LC-MS runs. Almost all detected
peptides were found reproducible (i.e. detected in at least five of the six replicates) in the HR
system (97%), against 67% in the LR system (Fig. 2A). Peptide reproducibility was clearly
correlated to peptide intensity in LR data (Fig. S3), most probably due to noisy XICs. Altogether,
418 of the 492 identified proteins were represented by at least one reproducible peptide.
After normalization and log10-transformation (see supplement materials), the mean coefficient of
variation of peptide quantitative values was 1.31% in HR and 1.40% in LR data (Fig. 2B, 2C). This
small technical variation is similar to other reported data (see [1]) and attests the accuracy of the
detection/quantification process. Moreover, a correlation of 0.89 between the mean intensity of
peptides common to LR and HR data (1179 peptides, Fig. 2D) showed that the quantification
process extracted similar results from both systems, despite a high sample complexity not
favourable to LR analysis. The few high coefficients observed for abundant peptides in the HR data
were mostly due to a poor determination of the ends of smearing peaks.
Twenty-five and fourteen BSA peptides were quantified in at least three samples in respectively LR
and HR systems. All HR peptide intensities except one were highly correlated and linearly related to
injected BSA quantities with a mean coefficient of correlation greater than 0.98 on three orders of
magnitude. This exception was due to a single datapoint (Fig. 3A). Nineteen of the twenty-five LR
peptides responded linearly to BSA quantities with a mean coefficient of correlation higher than
0.98 on two orders of magnitude (Fig. 3B). The lower correlation observed for the six remaining
peptides was mainly due to miss-assignments at low BSA quantities (<45 fmol): the BSA peptide
peak was contaminated by a peak of the yeast digest of similar m/z and RT values (Fig. S4). Thus,
quantification performances were lower with the LR than with the HR system, mainly because of
mismatches caused by the high complexity of the yeast lysate. This confirms that accurate
measurements can be expected with LR systems only when analysing peptide samples of lower
complexity. Nevertheless, the observed correlations between peptide intensity and protein quantity
were globally similar to those obtained by other software [4,5,12].
MassChroQ is written in C++ with Qt and runs both on Linux and Windows platforms. It is a
command-line standalone program and it comes with a library for integration in proteomics
pipelines.
MassChroQ is fully configurable via an XML input file (in masschroqML format) where the user
indicates the chosen processing steps, parameters and data files to analyse (see example on Fig. S5).
This file can be automatically generated by any XML editor by using the provided schema, or
manually by using a text editor. Parameters of XIC creation, filtering and detection, which depend
on the type, precision and noise level of the spectrometer, can all be configured in the
masschroqML file. Templates for several experiment scenarios are provided in the documentation.
LC-MS data input files can be in mzXML [18] or mzML format [19]. If X!Tandem [20] is used for
protein identification, a complete masschroqML input file containing identified peptides and protein
descriptions
can
be
automatically
generated
via
our
X!Tandem
pipeline
tool
(http://pappso.inra.fr/bioinfo/xtandempipeline/). If another identification engine is used, identified
peptides to be quantified can be provided to MassChroQ via TSV or CSV text files (Tab or Comma
Separated Values). MassChroQ results can be exported in TSV, gnumeric spreadsheet or
masschroqML XML format. TSV and spreadsheet formats allow direct import of data to statistical
software and the XML format allows their upload in proteomics databases like PROTICdb [21].
XICs can also be exported for visualization.
Computation time depends on data size and on the number of extracted XICs. In the present study,
the processing of the twelve LC-MS runs (6GB) where more than 5000 different peptide XICs were
extracted took 1 hour with a 2.93 GHz CPU on a Linux platform. Most of that time was spent
analysing non-centroid data from the LR system.
In conclusion, we showed that MassChroQ efficiently aligns and quantifies LR and HR LC-MS
data. Low coefficients of variation and high coefficients of correlation to protein quantity attested
the quality of the quantification measurements. MassChroQ is currently being successfully used in
our laboratory on both isotopic and label-free large experiments (data not shown). Its very modular
structure facilitates implementation of new algorithms and integration in other pipelines. Future
developments will focus on handling SRM data, developing a graphical user interface for
parameter adjustment with XIC visualization and computation time optimisation. This program is
licensed under the GNU General Public License v3.0 (http://www.gnu.org/licenses/gpl.html). The
source code is available at https://sourcesup.cru.fr/projects/masschroq. Compiled binaries for Linux
and Windows platforms and documentation (including data and results of this test set) can be found
at http://pappso.inra.fr/bioinfo/masschroq/.
Acknowledgements
This work was partially supported by "Infrastructures en Biologie Santé et Agronomie (IbiSA)" (E
N salary). The authors want to thank Mélisande Blein for the yeast extracts and SourceSup for their
subversion repository support.
Conflict of interest statement
The authors have declared no conflicts of interest.
References
[1]
America, A.H.P., Cordewener, J.H.G., Comparative LC-MS: a landscape of peaks and valleys.
Proteomics 2008, 8, 731-749.
[2]
Mueller, L.N., Brusniak, M.-Y., Mani, D.R., Aebersold, R., An assessment of software
solutions for the analysis of mass spectrometry based quantitative proteomics data. J.
Proteome Res 2008, 7, 51-61.
[3]
Neilson, K.A., Ali, N.A., Muralidharan, S., Mirzaei, M. et al., Less label, more free:
approaches in label-free quantitative mass spectrometry. Proteomics 2011, 11, 535-553.
[4]
Deutsch, E.W., Mendoza, L., Shteynberg, D., Farrah, T. et al., A guided tour of the TransProteomic Pipeline. Proteomics 2010, 10, 1150-1159.
[5]
Kohlbacher, O., Reinert, K., Gröpl, C., Lange, E. et al., TOPP--the OpenMS proteomics
pipeline. Bioinformatics 2007, 23, e191-197.
[6]
Li, X.-jun, Yi, E.C., Kemp, C.J., Zhang, H., Aebersold, R., A software suite for the generation
and comparison of peptide arrays from sets of data collected by liquid chromatography-mass
spectrometry. Mol. Cell Proteomics 2005, 4, 1328-1340.
[7]
Bellew, M., Coram, M., Fitzgibbon, M., Igra, M. et al., A suite of algorithms for the
comprehensive analysis of complex protein mixtures using high-resolution LC-MS.
Bioinformatics 2006, 22, 1902-1909.
[8]
Sturm, M., Bertsch, A., Gröpl, C., Hildebrandt, A. et al., OpenMS - an open-source software
framework for mass spectrometry. BMC Bioinformatics 2008, 9, 163.
[9]
Mueller, L.N., Rinner, O., Schmidt, A., Letarte, S. et al., SuperHirn - a novel tool for high
resolution LC-MS-based peptide/protein profiling. Proteomics 2007, 7, 3470-3480.
[10] Palagi, P.M., Walther, D., Quadroni, M., Catherinet, S. et al., MSight: an image analysis
software for liquid chromatography-mass spectrometry. Proteomics 2005, 5, 2381-2384.
[11] Mortensen, P., Gouw, J.W., Olsen, J.V., Ong, S.-E. et al., MSQuant, an open source platform
for mass spectrometry-based quantitative proteomics. J. Proteome Res 2010, 9, 393-403.
[12] Tsou, C.-C., Tsai, C.-F., Tsui, Y.-H., Sudhir, P.-R. et al., IDEAL-Q, an Automated Tool for
Label-free Quantitation Analysis Using an Efficient Peptide Alignment Approach and Spectral
Data Validation. Molecular & Cellular Proteomics 2010, 9, 131 -144.
[13] Li, X.-J., Zhang, H., Ranish, J.A., Aebersold, R., Automated statistical analysis of protein
abundance ratios from data generated by stable-isotope dilution and tandem mass
spectrometry. Anal. Chem 2003, 75, 6648-6657.
[14] Bouyssié, D., Gonzalez de Peredo, A., Mouton, E., Albigot, R. et al., Mascot file parsing and
quantification (MFPaQ), a new software to parse, validate, and quantify proteomics data
generated by ICAT and SILAC mass spectrometric analyses: application to the proteomics
study of membrane proteins from primary human endothelial cells. Mol. Cell Proteomics
2007, 6, 1621-1637.
[15] Cox, J., Matic, I., Hilger, M., Nagaraj, N. et al., A practical guide to the MaxQuant
computational platform for SILAC-based quantitative proteomics. Nat Protoc 2009, 4, 698705.
[16] Serra, J., Image analysis and mathematical morphology, Academic Press, London 1982.
[17] Prince, J.T., Marcotte, E.M., Chromatographic alignment of ESI-LC-MS proteomics data sets
by ordered bijective interpolated warping. Anal. Chem 2006, 78, 6140-6152.
[18] Pedrioli, P.G.A., Eng, J.K., Hubley, R., Vogelzang, M. et al., A common open representation
of mass spectrometry data and its application to proteomics research. Nat. Biotechnol 2004,
22, 1459-1466.
[19] Martens, L., Chambers, M., Sturm, M., Kessner, D. et al., mzML - a Community Standard for
Mass Spectrometry Data. Mol Cell Proteomics 2010, 10, R110.000133.
[20]
Craig, R., Beavis, R.C., TANDEM: matching proteins with tandem mass spectra.
Bioinformatics 2004, 20, 1466-1467.
[21] Ferry-Dumazet, H., Houel, G., Montalent, P., Moreau, L. et al., PROTICdb: a web-based
application to store, track, query, and compare plant proteome data. Proteomics 2005, 5, 2069-
2081.
Figure 1. Evaluation of alignment. (A) Example of RT correction along an HR LC-MS run.
Points correspond to peptides identified by MS/MS in both the reference LC-MS run and the
LC-MS run being aligned. The line is the computed time deviation used for alignment of MS
data. (B, C) Standard deviation of peptide RT before and after alignment respectively in HR
and LR experiments.
Figure 2. Evaluation of reproducibility. (A) Number of peptides detected in 1 to 6 replicates in LR
and HR LC-MS experiments. (B, C) Influence of mean peptide intensity on peptide coefficient of
variation (all reproducible peptides) in HR and LR systems. (D) Correlation between HR and LR
mean values of peptides identified in both systems (1179 peptides, r = 0.89).
Figure 3. Linear relation between BSA peptide intensity and BSA
quantity. (A) and (B) : respectively HR (25 peptides) and LR (14
peptides) experiments. Each line corresponds to a different peptide.
Only peptides detected in at least 3 samples are shown. Dashed line:
peptides showing a correlation to BSA quantity lower than 0.98.
Figure S1. XIC creation, filtering and peak detection: example for peptide LVNELTEFAK (45 fmol of BSA digest spiked in yeast total digest) in LCMS runs from LR (A) and HR (B) experiments. Filtering involves baseline correction for LR experiments and spike removing for HR ones. Red
circles: detected peaks; blue circles: detected peak matching the peptide's RT.
Figure S2. Signal treatment for peak detection. (A) The original XIC contains high frequency noise that is partly eliminated by smoothing (average and/
or median filters). (B) Morphological closing (blue profile) and opening (red profile) by a small flat structuring element are performed on the smoothed
XIC (gray profile). The closing operation eliminates many noisy peaks by filling small valleys and preserves the actual position of the remaining peaks.
Hence peaks are detected on this profile if they are greater than a threshold (blue line). Only peaks that are thick compared to the structuring element
stay high in the opened profile. Then, to avoid detection of thin artifactual spikes, peaks detected on the closed profile are filtered according to the
intensity at the same position in the opened profile : intensity in the opened profile must be greater than a second threshold (red line).
Figure S3. Relation between peptide intensity and detection
reproducibility in the LR experiment.
Figure S4. Peak assignment error in LR data analysis: example of the BSA peptide HLVDEPQNLIK.
The BSA digest quantity is indicated on the top-left corner of each graph. Red circles: detected peaks. Blue
circles: Peaks matching the peptide's RT. A yeast peak at RT and m/z values very close to those of
HLVDEPQNLIK, induced a mismatch at 15 fmol. In addition, peaks assigned to HLVDEPQNLIK in other
samples partly contained the intensity that should have been assigned to the yeast peptides.
SUPPLEMENTAL FIGURE S5
<!--Example of MassChroQ processing file-->
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<masschroq>
<!-- List of LC-MS run files in open
<rawdata>
<data_file id="samp0" format="mzxml"
<data_file id="samp1" format="mzxml"
<data_file id="samp2" format="mzxml"
<data_file id="samp3" format="mzxml"
type="profile"/>
<data_file id="samp4" format="mzxml"
type="profile"/>
</rawdata>
<groups>
format : mzXML or mzML -->
path="Labelling-light-heavy.mzXML" type="centroid"/>
path="Label-free-samp1.mzXML" type="centroid"/>
path="Label-free-samp2.mzXML" type="centroid"/>
path="MS-data-without_identification-samp1.mzXML"
path="MS-data-without_identification-samp2.mzXML"
<!-- Grouping of LC-MS runs. Within a group:
- all LC-MS runs are aligned (with the same alignment method);
- peptides observed in at least one LC-MS run are quantified in all LC-MS runs
of this group (using the same quantification method) -->
<group data_ids="samp0" id="G1"/>
<group data_ids="samp1 samp2" id="G2"/>
<group data_ids="samp3 samp4" id="G3"/>
</groups>
<!-- The peptide features list can be defined in two ways :
in separate spreadsheets containing identified peptides for each LC-MS run as follows :-->
<peptide_files_list>
<peptide_file data="samp0" path="labelling_peptide_list.txt"/>
</peptide_files_list>
<!-- directly in this file with proteins/peptides list as follows : -->
<protein_list>
<protein desc="conta|P02769|ALBU_BOVIN SERUM ALBUMIN PRECURSOR." id="P1.1"/>
</protein_list>
<peptide_list>
<peptide id="pep0" mh="1463.626" mods="114.08" prot_ids="P1.1" seq="TCVADESHAGCEK">
<observed_in data="samp2" scan="755" z="2"/>
<observed_in data="samp3" scan="798" z="2"/>
</peptide>
<peptide id="pep1" mh="1103.461" mods="57.04" prot_ids="P1.1" seq="ADESHAGCEK">
<observed_in data="samp2" scan="663" z="2"/>
</peptide>
</peptide_list>
<!-- Definition of different labels for isotopic experiments.
Example with a dimethylation of primary amine -->
<isotope_label_list>
<isotope_label id="light">
<mod at="Nter" value="28.0"/>
<mod at="K" value="28.0"/>
</isotope_label>
<isotope_label id="heavy">
<mod at="Nter" value="32.0"/>
<mod at="K" value="32.0"/>
</isotope_label>
</isotope_label_list>
<!-- Definition of different alignment methods. Two alignment algorithms
are implemented : MS/MS alignment and OBI-Warp alignment -->
<alignments>
<alignment_methods>
<alignment_method id="ms2_1">
<ms2>
<ms2_tendency_halfwindow>10</ms2_tendency_halfwindow>
<ms2_smoothing_halfwindow>5</ms2_smoothing_halfwindow>
<ms1_smoothing_halfwindow>0</ms1_smoothing_halfwindow>
</ms2>
</alignment_method>
<alignment_method id="obiwarp1">
<obiwarp>
<lmat_precision>1</lmat_precision>
<mz_start>500</mz_start>
<mz_stop>1200</mz_stop>
</obiwarp>
</alignment_method>
</alignment_methods>
<!-- Perform alignment on each group using the above defined methods.
A reference LC-MS run should be defined for each group. All other runs of
the group will be aligned towards this reference run -->
<align group_id="G2" method_id="ms2_1" reference_data_id="samp1"/>
<align group_id="G3" method_id="obiwarp1" reference_data_id="samp3"/>
</alignments>
<!-- Definition of different quantification methods and parameters for
XIC creation, XIC filtering and peak detection -->
<quantification_methods>
<quantification_method id="quanti1">
<!-- XIC creation on mz or ppm range using TIC (sum) or basepeak (max) -->
<xic_extraction xic_type="sum">
<ppm_range max="10" min="10"/>
</xic_extraction>
<!-- XIC filtering with spike removing, baseline correction or smoothing -->
<xic_filters>
<anti_spike half="5"/>
<background half_mediane="5" half_min_max="15"/>
<smoothing half="3"/>
</xic_filters>
<!-- XIC detection with threshold -->
<peak_detection>
<detection_zivy>
<mean_filter_half_edge>1</mean_filter_half_edge>
<minmax_half_edge>3</minmax_half_edge>
<maxmin_half_edge>2</maxmin_half_edge>
<detection_threshold_on_max>5000</detection_threshold_on_max>
<detection_threshold_on_min>3000</detection_threshold_on_min>
</detection_zivy>
</peak_detection>
</quantification_method>
</quantification_methods>
<!-- Quantification area -->
<quantification>
<!-- Definition of the export files and formats for the quantification results :
spreadsheet formats (tsv, gnumeric) or xml format -->
<quantification_results>
<quantification_result output_file="results" format ="tsv"/>
</quantification_results>
<!-- Definition of the export files for the XIC traces in spreadsheet (tsv) format :
one can trace all the XICs and/or a list of given peptides and/or a list of
given m/z values, and/or a list of m/z-rt values -->
<quantification_traces>
<all_xics_traces output_dir="all_xics_traces" format="tsv"/>
<peptide_traces peptide_ids="pep0 pep1" output_dir="peplist_xics_traces" format="tsv"/>
</quantification_traces>
<!-- For each group, start quantification on : -->
<!-- all the peptides -->
<quantify withingroup="G1" quantification_method_id="quanti1">
<peptides_in_peptide_list mode="real_or_mean" isotope_label_refs="light heavy"/>
</quantify>
<quantify withingroup="G2" quantification_method_id="quanti1" id="q1">
<peptides_in_peptide_list mode="real_or_mean"/>
</quantify>
<!-- and/or on a list of given m/z values -->
<quantify withingroup="G3" quantification_method_id="quanti1" id="q2">
<mz_list>732.317 449.754 552.234 464.251 381.577 569.771 575.256</mz_list>
</quantify>
<!-- Quantification can also be performed on a list of m/z-rt values -->
</quantification>
</masschroq>
Supplementary materials and methods
1 Protein extraction and digestion
The proteins were extracted from yeast cell pellets using a TCA/acetone method: proteins were
precipitated in 10% TCA, 0.07% 2-Mercaptoethanol in acetone and the pellet was rinsed in 0.07%
2-Mercaptoethanol in acetone. Precipitated proteins were resuspended in urea 8 M, thiourea 2 M,
CHAPS 2%. Protein concentration was determined by using the PlusOne 2-D Quant Kit (GE
Healthcare), and 10 µg of proteins were reduced and alkylated with respectively 10 mM of DTT
and 55 mM of iodoacetamide. Proteins were diluted 10 times with 50 mM ammonium bicarbonate
and digested with 1/50 (w/w) trypsin (Promega) at 37 °C overnight. Digestion was stopped by
acidification with TFA. Bovin Serum Albumin (initial fractionation by heat shock, Sigma) was
digested similarly.
2 LC-MS/MS analysis
2.1 Low resolution analysis
HPLC was performed on a NanoLC-Ultra system (Eksigent). Seven-hundred ng of protein digest
were loaded at 7.5 µL/min-1 on a precolumn cartridge (stationary phase: C18 PepMap 100, 5 µm;
column: 100 µm i.d., 1 cm; Dionex) and desalted with 0.1% HCOOH. After 3 min, the precolumn
cartridge was connected to the separating PepMap C18 column (stationary phase: C18 PepMap 100,
3 µm; column: 75 µm i.d., 150 mm; Dionex). Buffers were 0.1% HCOOH in water (A) and 0.1%
HCOOH in ACN (B). The peptide separation was achieved with a linear gradient from 5 to 30% B
for 60 min at 300 nL/min-1. Including the regeneration step at 95% B and the equilibration step at
95% A, one run took 90 min.
Eluted peptides were analysed on-line with a LTQ XL ion trap (Thermo Electron) using a
nanoelectrospray interface. Ionization (1.5 kV ionization potential) was performed with liquid
junction and a non-coated capillary probe (10 µm i.d.; New Objective). Peptide ions were analysed
using Xcalibur 2.0.7 with the following data-dependent acquisition steps: (1) full MS scan (m/z of
300 to 1300, enhanced profile mode) and (2) MS/MS (qz = 0.25, activation time = 30 ms, and
collision energy = 35%; centroid mode). Step 2 was repeated for the three major ions detected in
step 1. Dynamic exclusion was set to 45 s.
2.2 High resolution (HR) analysis
HPLC was performed on an Ultimate 3000 LC system (Dionex). Seven-hundred ng of protein
digest were loaded at 7.5 µL/min-1 on a precolumn cartridge (stationary phase: C18 PepMap 100, 5
µm; column: 300 µm i.d., 5 mm; Dionex) and desalted with 0.08% TFA and 2% ACN. After 3
minutes, the precolumn cartridge was connected to the separating PepMap C18 column (stationary
phase: C18 PepMap 100, 3 µm; column: 75 µm i.d., 150 mm; Dionex). Buffers were 0.1%
HCOOH, 3% ACN (A) and 0.1% HCOOH, 80% ACN (B). The peptide separation was achieved
with a linear gradient from 4 to 36% B for 60 min at 300 nL/min -1. Including the regeneration step
at 100% B and the equilibration step at 100% A, one run took 90 min.
Eluted peptides were analysed on-line with a LTQ-Orbitrap Discovery (Thermo Electron)
using a nanoelectrospray interface. Ionization (1.3 ionization potential) was performed with liquid
junction and a non-coated capillary probe (10 µm i.d.; New Objective). Peptide ions were analysed
using Xcalibur 2.0.7 with the following data-dependent acquisition steps: (1) FTMS scan on
Orbitrap (m/z of 300 to 1300, 15000 resolution, profile mode), (2) MS/MS on the LTQ (qz = 0.25,
activation time = 30 ms, and collision energy = 35%; centroid mode). Step 2 was repeated for the
two major ions detected in step 1. Dynamic exclusion was set to 90s.
3 Protein identification
Database search was performed with the X!Tandem software (version 2010.01.01.4)
(http://www.thegpm.org/TANDEM/). Enzymatic cleavage was declared as a trypsin digestion with
one possible misscleavage. Cys carboxyamidomethylation and Met oxidation were set to
respectively static and possible modifications. Precursor mass precision was set to 2.0 Da for LR
and 20 ppm for HR. Fragment mass tolerance was 0.5 Th for both LR and HR data. A refinement
search was added with the same parameters except that semi-trypsic peptide and protein N-ter
acetylation
were
also
searched.
The
Saccharomyces
Genome
Database
(http://downloads.yeastgenome.org, 5885 entries) was searched together with a contaminant
database (trypsin, keratins...). The decoy database comprised the reverse protein sequences of the
Saccharomyces database. Only peptides with an E-value smaller than 0.1 were reported.
Identified proteins were filtered and sorted by using the XTandem pipeline
(http://pappso.inra.fr/bioinfo/xtandempipeline/). Criteria used for protein identification were (1) at
least two different peptides identified with an E-value smaller than 0.05, (2) a protein E-value
(product of unique peptide E-values) smaller than 10-4 . These criteria led to a False Discovery Rate
of 0.3% for peptide and protein identification. To take redundancy into account (i.e. the fact that the
same peptide sequence can be found in several proteins), proteins with at least one peptide in
common were grouped. Grouped proteins had similar functions. Within each group, proteins with at
least one specific peptide relatively to other members of the group were reported as sub-groups.
4 Statistical analysis
In order to take into account possible global quantitative variations between LC-MS runs,
normalization was performed. For each LC-MS run, the ratio of all peptide values to their value in
the chosen reference LC-MS run was computed. Normalization was performed by dividing peptide
values by the median value of peptide ratios. Subsequent statistical analyses were performed on
log10-transformed normalized data.
BSA peptides showed linear relationships to protein quantity, but the slopes were peptide-
specific. These peptide-specific global effects were estimated by a two-way ANOVA on normalized
data, where the two factors were peptide and BSA quantity. This enabled the estimation of the
specific effect of each BSA peptide, which was removed from normalized data in Figure 3. This
allowed the representation of all peptides on the same scale, but had no effect on curve shapes.
| 5cs.CE
|
1
Approximate Random Matrix Models for
Generalized Fading MIMO Channels
Muralikrishnan Srinivasan
Sheetal Kalyani
arXiv:1707.09734v1 [cs.IT] 31 Jul 2017
Department of Electrical Engineering,
Indian Institute of Technology, Madras,
Chennai, India 600036.
{ee14d206,skalyani}@ee.iitm.ac.in
Abstract
Approximate random matrix models for κ − µ and η − µ faded multiple input multiple output
(MIMO) communication channels are derived in terms of a complex Wishart matrix by keeping the
degree of freedom constrained to the number of transmitting antennas and by matching matrix variate
moments. The utility of the result is demonstrated in a) computing the average capacity/rate of κ−µ/η−µ
MIMO systems b) computing Symbol Error Rate (SER) for optimum combining with Rayleigh faded
users and an arbitrary number of κ − µ and η − µ faded interferers. These approximate expressions are
compared with Monte-Carlo simulations and a close match is observed.
Index Terms
Random matrices, Wishart matrices, Generalized fading, κ − µ, η − µ, MIMO, capacity, optimum
combining
I. I NTRODUCTION
The need for high data rates has been one of the driving factors for the evolution of the wireless
systems from Single Input Single Output (SISO) systems to Multiple Input Multiple Output
(MIMO) systems. MIMO systems are being used increasingly in modern wireless standards and
it is imperative to study the channel capacity and other Quality of Service (QoS) metrics of such
systems.
The capacity of wireless channels depends on channel fading statistics and also on whether the
statistics are known at the receiver and the transmitter. To capture the fading statistics, channel
August 1, 2017
DRAFT
2
gain is characterized by a single random variable in SISO systems. But for MIMO systems,
the channel is in the form of a matrix, hence the characterization of random matrices plays an
indispensable role in studying MIMO channel metrics. The MIMO system is mathematically
modeled by an NR × NT channel gain matrix H, where NR is the number of receive antennas
and NT is the number of transmit antennas. Various performance metrics such as capacity, rate,
etc. require the eigenvalue statistics of the Gram matrix HHH (or HH H). When the elements
of H are i.i.d. circular symmetric complex Gaussian with zero or non-zero mean, i.e., when
Rayleigh or Rician faded MIMO channels are considered, the Gram HHH can be characterized
by Wishart matrices - central or non-central respectively [1]. These random matrix models have
been used widely for deriving capacity expressions in the case of Rayleigh faded MIMO channels
[2]–[4] and also Rician faded MIMO channels [5], [6].
Recently, there has been significant focus on generalized fading models namely κ − µ and
η −µ models introduced in [7]. These distributions model the small scale variations in the fading
channel in the line of sight and non-line of sight conditions respectively. Further, these generalized
fading distributions include Rayleigh, Rician, Nakagami, One-sided Gaussian distributions as
special cases. These generalized fading distributions have been widely used in capacity and
outage probability analysis for SISO and MISO systems. Average channel capacity of single
branch κ − µ and η − µ faded receivers is studied in [8]. The outage, coverage probability
and rate of these generalized fading channels are analyzed in [9]–[13] and references therein.
Secrecy capacity analysis is carried out in [14] and effective throughput in MISO systems is
determined in [15], [16]. Outage probability of MRC in κ − µ fading channels in the presence
of co-channel interference is studied in [17]. Analysis of decode and forward relay system for
generalized fading models is performed in [18]–[20].
The capacity of MIMO systems for these generalized fading channels has been less analyzed
for want of a random matrix model that characterizes the channel matrix. Nevertheless, some
random matrix models have been developed for Nakagami and Rician-shadowed fading channels.
A random matrix model has been developed for Nakagami-q fading in [21] and the pdf of eigen
values of the Gram HHH is obtained in terms of a Pfaffian. In [22], the ergodic capacity of
MIMO correlated Nakagami-m fading channel has been derived using the concept of a copula.
But the work presents an analysis only for 2 × 2 MIMO channel and determining the capacity
of MIMO channels with a larger number of receive and transmit antennas using this method
is cumbersome. Recently a MIMO capacity upper bound was derived for the κ − µ and η − µ
August 1, 2017
DRAFT
3
fading channels in [23]. In [24], a MIMO model has been developed for Rician-shadowed fading
as a unification model for MIMO-Rayleigh and MIMO-Rician fading models.
Given the complicated pdf structure of complex κ − µ and η − µ fading distributions [25],
[26], it is challenging to develop the matrix distribution and the eigenvalue statistics for HHH ,
even when the elements of H are assumed to be i.i.d. κ − µ or η − µ random variables. Hence,
in this paper, we develop an approximate matrix model for HHH (or HH H) in terms of a
Wishart distribution, which is a very well-studied matrix distribution [1]. Approximating any
matrix distributions by central Wishart by means of Taylor expansion is studied in [27], but
the approximation requires the knowledge of not only one or more cumulants and moments
of the random matrix that is to be approximated but also the derivatives of central Wishart
matrix. Also, the approximation of non-central Wishart matrix by a central Wishart by means
of Laguerre polynomial expansion is given in [28] and by means of the moment generating
functions in [29]. In this paper, we propose a Wishart distributed approximation of HHH , such
that the approximation has its first moment matched with the original matrix distribution of
HHH and the degree of freedom is constrained to be the number of columns of the matrix H.
This method requires only the knowledge of the expectation of HHH with respect to the original
distribution and this can be found out for both the κ − µ and η − µ case. We also show that our
method is equivalent to minimizing the K-L divergence between the actual MIMO matrix and
the Wishart distributed approximation1.
The proposed approximation is discussed in Section II. In Section III, the utility of the
approximation is shown in two applications. In one application, the proposed approximation
is used to determine the capacity of MIMO systems with i.i.d. κ − µ or η − µ channel gains.
Further, the approximation is also used to determine the asymptotic capacity of these MIMO
systems. To the best of our knowledge, ours is the first work to derive even an approximate
capacity expression in the presence of κ − µ/η − µ MIMO channels. Another application of the
approximation is in determining the Symbol Error Rate (SER) of an optimum combining (OC)
receiver [30] for Rayleigh faded user and κ−µ or η−µ faded interferers. The SINR expression for
OC involves a covariance matrix formed by interferer channel gains. Using the derived Wishart
approximation, closed form expressions for SER of OC systems for κ − µ or η − µ faded
1
Since the κ − µ/η − µ fading distributions are fairly complicated, computing and matching the higher moments is fairly
difficult.
August 1, 2017
DRAFT
4
interferers are obtained. To the best of our knowledge, no prior work has given SER expressions
for a receiver diversity system employing OC under the case of Rayleigh faded user and κ−µ or
η−µ faded interferers. In Section IV, the derived capacity and SER approximations are compared
with Monte-Carlo simulations and a close match is found between the theoretical results and
simulation results. While we have only shown the utility of the approximation in two applications
namely capacity computation and SER computation in OC system, the approximation can be
used in any application which deals with random κ − µ/η − µ matrix models.
Basic notation: Ex (.) denotes expectation with respect to distribution x. |X| and det(X) denote
determinant of a matrix X. etr(X) denotes an exponential raised to trace of the matrix X.
II. P ROPOSED A PPROXIMATION
Let H be an n1 × n2 random matrix with independent and identically distributed elements.
and X = HHH be an n1 × n1 random matrix. The exact matrix distribution of X denoted by
p(X) is not known2. Hence, we propose to approximate the density p(X) by an n1 × n1 Wishart
matrix whose distribution is q(X) = CW n1 (n2 , Σ) with n2 degrees of freedom and covariance
matrix Σ, such that Ep [X] = Eq [X], i.e., their first moments are matched.
Eq (X) = n2 Σ = Ep [X].
(1)
We now show that this is also equivalent to minimizing the the K-L divergence between p(X)
and a Wishart distribution. Let q(X) be that Wishart distribution which minimizes the K-L
divergence between p(X) and all the Complex Wishart distributions CW n1 (n, Σ), i.e.,
Z
q(X) = argminKL(p(X)||q(X)) = argmax p(X)[ln(q(X)) − ln(p(X))]dX
q(X)
q(X)
= argmax
q(X)
Z
p(X)ln(q(X))dX.
(2)
Note that we assume an unknown degrees of freedom as n in this case, while in the Wishart
approximation we had constrained the degree of freedom to n2 . The density of an n1 × n1
complex Wishart matrix X ∼ CW n1 (n, Σ) is given by [31],
q(X) =
2
1
etr(−Σ−1 X)(detX)n−n1 ,
CΓn1 (n)(detΣ)n
The pdf of entries of H are known. But finding the matrix variate pdf of X = HHH is fairly complicated, since we require
the joint pdf of all entries of X. On the other hand, the diagonal elements of X are well characterized.
August 1, 2017
DRAFT
5
where CΓ. (.) is the complex multivariate gamma function [31]. Substituting the density in (2),
we obtain
q(X) = argmax
q(X)
Z
p(X) − ln(CΓn1 (n)) − nln|Σ| + T r(−Σ X) + (n − n1 )ln|X| dX
−1
−1
= argmax − ln(CΓn1 (n)) − nln|Σ| + T r(−Σ Ep [X]) + (n − n1 )Ep [ln|X|] .
q(X)
Denoting Z = Ep [X] and Y = Ep [ln|X|], we get
−1
q(X) = argmax − ln(CΓn1 (n)) − nln|Σ| + T r(−Σ Z) + (n − n1 )Y .
(3)
q(X)
To obtain the minimizing distribution, we can differentiate the above equation with respect to
two variables namely, Σ and n. Differentiating equation (3) w.r.t. Σ, we obtain
dq(X)
= −nΣ−1 + Σ−1 ZT Σ−1 .
dΣ
When the above equation is equated to zero, we obtain
Σ=
1
1 T
Z = Ep [X].
n
n
(4)
Note that we obtain the same Σ when we equate the expectations of the matrix with respect to
distributions p(X) and q(X), i.e., Ep [X] = Eq [X] and by fixing the degrees of freedom to be
n2 . But the question arises as to whether the degree of freedom of the distribution obtained from
minimizing K-L divergence is indeed n2 , which is nothing but the number of columns of the
matrix H. In order to answer this question, we will continue the K-L divergence minimization
by differentiating (3) w.r.t. n. Now differentiating equation (3) w.r.t. n, we obtain
n
1
X
dq(X)
= −ln|Σ| + Y −
ψ(n − i + 1),
dn
i=1
where ψ(.) is the digamma function [32]. Equating the derivative to zero, we get
−ln|Σ| + Y −
By substituting Σ =
1 T
Z
n
n1
X
i=1
ψ(n − i + 1) = 0.
from (4), we obtain,
ln(n) − ln|Z| + Y −
n1
X
i=1
ψ(n − i + 1) = 0.
(5)
Matching the expectations Ep [ln|X|] = Eq [ln|X|] also leads to (5). Hence minimizing the K-L
divergence has reduced to a simple case of matching expectations Ep [X] and Ep [ln|X|] with
Eq [X] and Eq [ln|X|] respectively.
August 1, 2017
DRAFT
6
Solving (5) for n requires the knowledge of Y = Ep [ln|X|]. Since finding the random
matrix variate distribution of X is mathematically intractable, finding the expectation of the
log-determinant analytically is not possible. We now argue that approximating n by n2 is a
plausible solution to the problem.
a) While Y = Ep [ln|X|] cannot be theoretically computed, it can be empirically computed by
using simulated κ − µ/η − µ matrix elements and then solving for n in (5). We observe that in
all our simulations, this leads to n being a real number which is very close to n2 .
b) Note, n2 denotes the number of transmitter antennas or the number of interferers in the MIMO
channel matrix. Hence, it makes sense to retain the same number n2 , even in approximation,
given that there is no correlation in the transmitter side and all the elements of the matrix are
i.i.d.
c) For any Wishart distributed matrix A = BBH ∼ CW n1 (n2 , Σ), the degrees of freedom also
denote the number of columns of complex Gaussian B. In fact, when a non-central Wishart
matrix was approximated by a Wishart matrix in [29], the approach of keeping n = n2 was
followed.
Based on the above reasoning it can be argued that, n ≈ n2 . Hence, the minimizer complex
Wishart distribution, q(X) given by CW n1 (n2 , n12 ZT ), where Z = Ep [X], is the closest to the
actual unknown distribution among all central Wishart distributions in terms of K-L divergence.
We will now apply the approximation procedure to i.i.d. κ − µ and η − µ fading MIMO channels
and theoretically determine the covariance matrix Σ =
1
(Ep [X])T
n2
=
1 T
Z
n2
of the corresponding
approximate Wishart matrix.
Once the first moment is matched by the above procedure, we now quantify how close the
second moment of the two matrices are to one another. For this, we determine NMSE (Normalized
P
XXH −
Mean Square Error) between the second moments of the two matrices. First, let X̂ =
P
P ′ ′H
. denotes the empirical
X X , where X′ is the Wishart approximated matrix and
average of a matrix. Also, X = HHH where H is a matrix with i.i.d. κ − µ elements. In
other words, X̂ is used to measure the discrepancy between the second moments of the actual
matrix and Wishart distributed approximation. Second, we determine the absolute value of all
the elements of the matrix X̂, i.e., X̃(i, j) = |X̂(i, j)|. We now determine NMSE, i.e., the ratio
P
XXH as
of average of elements of X̃ to the average of elements of
NMSE =
August 1, 2017
1X̃1T
.
P
XXH 1T
1
DRAFT
7
TABLE I: NMSE for n1 = 2
n2
κ = 2 and µ = 3
κ = 6 and µ = 3
κ = 2 and µ = 6
κ = 6 and µ = 6
2
0.4247
0.4633
0.4631
0.4769
4
0.2182
0.2328
0.2362
0.2417
6
0.1511
0.1563
0.1574
0.1631
8
0.1115
0.1198
0.1195
0.1205
Table I shows the NMSE for various values of degree of freedom n2 , κ and µ. We can see that
as the degree of freedom n2 increases, NMSE decreases, i.e., the approximation becomes tighter,
where as, if κ or µ increases, NMSE increases.
A. κ − µ model
In κ − µ fading model, the signal is divided into different clusters of waves. The number of
clusters is µ and in each of the clusters, there is a deterministic LOS component with arbitrary
power and scattered waves with identical powers. Note, κ is the ratio between the total power
of the dominant components and the total power of the scattered waves. Suppose the elements
hi,j = xij + jyij of H are i.i.d. κ − µ random variables, where xij and yij are the real and
imaginary components respectively, then the joint distribution is given by [25],
(xij − p)2 + (yij − q)2
|xij yij |µ/2
exp(−
)
fxy (xij , yij ) = 4
4σ |pq|µ/2−1
2σ 2
qyij
|pxij |
|qyij |
pxij
(6)
sech( 2 )sech( 2 )I µ2 −1 ( 2 )I µ2 −1 ( 2 ).
σ
σ
σ
σ
P
P
Here p2 = µi=1 p2i and q 2 = µi=1 qi2 , where pi and qi are the LOS components of in-phase and
quadrature components respectively of multipath waves of each cluster. κ =
p2 +q 2
,
2µσ2
where σ 2 is
the power of the scattered waves. We are interested in the distribution of HHH . However, an
exact characterization of the matrix variate distribution is intractable, because it involves finding
the joint pdf of entries of HHH . Hence, we now approximate HHH by the Wishart matrix with
n2 degrees of freedom and Σ =
1
E [HHH],
n2 p
where the expectation is with respect to the κ − µ
pdf. The diagonal elements of Z = Ep [HHH ], i.e, zii are nothing but the mean of n2 κ − µ
August 1, 2017
DRAFT
8
envelope square variables. i.e., zii = E[
where κ =
p2 +q 2
.
2µσ2
zij = E[
n2
X
k=1
P n2
2
j=1 [xij
+ yij2 ]]. Hence, zii = 2σ 2 n2 (1 + κ)µ from [7],
The off diagonal elements of zij are given by
[(xik + jyik )(xkj − jykj )]] =
Since xik and yik are i.i.d., we obtain ∀i, j,
zij =
n2
X
n2
X
k=1
E[xik xkj + yik ykj − jxik ykj + jxkj yik ].
((E[xik ])2 + (E[ykj ])2 ).
(7)
k=1
E[xik ] and E[yik ] are given by,
Z ∞
(x − p)2
px
|x|µ/2
|px|
E[xik ] =
)sech( 2 )I µ2 −1 ( 2 )dx
x 2 µ/2−1 exp(−
2
2σ
σ
σ
−∞ 2σ |p|
∞
(y − q)2
qy
|y|µ/2
|qy|
µ
)sech(
)I
)dy.
E[yik ] =
y 2 µ/2−1 exp(−
(
−1
2σ 2
σ2 2
σ2
−∞ 2σ |q|
Z
(8)
(9)
∀i, j. The closed form expressions for the above integrals seem mathematically intractable.
However an approximation for the integral is derived in Appendix A and given by (39) and
(40).
3
Since all the κ − µ elements of the matrix H are i.i.d., the mean of all the off-diagonal
elements are equal. Substituting the results from (39) and (40) in (7), we obtain ∀i, j and i 6= j,
"
µ
i2
h
p2
µ
µ
2p2 π
2p2
4σ 2
µ
+1 Γ( 2 + 1)
2
)
+
1,
1,
3/2,
,
,
)
zij ≈ n2 2pe− 2σ2 ( 2
Ψ
(
1
4σ + 2p2 π
Γ( µ2 )
2
2 2p2 π + 4σ 2 4σ 2 + 2p2 π
#
µ
h
i2
2
2
2
q2
+
1)
Γ(
µ
µ
µ
2q
π
2q
4σ
+ 2qe− 2σ2 ( 2
) 2 +1 2 µ Ψ1 ( + 1, 1, 3/2, , 2
,
) .
2
4σ + 2q π
Γ( 2 )
2
2 2q π + 4σ 2 4σ 2 + 2q 2 π
(10)
Since Σ =
Z
,
n2
we have
Σii = 2σ 2 (1 + κ)µ and Σij = zij /n2 , i 6= j
(11)
Now we look at a special case of κ − µ distribution namely the Rician distribution with µ = 1.
3
These closed form approximations are also compared with both numerical integration evaluation of (8) and (9) and Monte-
Carlo simulation and an excellent match is observed with both.
August 1, 2017
DRAFT
9
Rician: Suppose the elements hi,j = xij + jyij ∼ CN (pij + jqij , 2σ 2 ) of H are independent
Rician distributed random variables, where xij and yij are the real and imaginary components
respectively, then the joint distribution is given by [25],
fxy (xij , yij ) =
(xij − pij )2 + (yij − qij )2
1
exp(−
)
2πσ 2
2σ 2
2
p2ij +qij
.
2
2σ
The diagonal elements zii of n2 Σ are nothing but mean
P 2 2
of n2 sum of Rician envelope square variables. i.e., zii = E[ nj=1
[xij + yij2 ]] = 2σ 2 n2 (1 + κ)
P 2
(pik + jqik )(pkj − jqkj ).
and the off-diagonal elements zij = nk=1
with identical Rice factor κ =
Since the elements of H have a Rician envelope, X is a non-central Wishart matrix. In
this case, we are simply approximating a non-central Wishart matrix by a central Wishart
matrix. This has been well studied in [29], which obtains the same approximation as us, but by
deriving the moment generating function (mgf) of the non-central Wishart matrix and retaining
the same degree of freedom. If the columns of the matrix H with dimension n1 × n2 are
distributed as complex Gaussian CN (mi , Σ′ ), then X is a non-central Wishart matrix distributed as CW n1 (n2 , Σ′ , MMH ), where M = [m1 , m2 , ..., ms ]. From [29], a non-central Wishart
matrix CW n1 (n2 , Σ′ , MMH ) can be approximated by a Wishart matrix CW n1 (n2 , Σ), where
Σ = Σ′ +
1
MMH .
n2
It is also known that the approximation becomes tighter as the degree of
freedom n2 increases. In our case also, it can be shown that Σ = Σ′ +
1
MMH .
n2
If the variables are identically distributed, i.e., pij s are equal to p and qij s are equal to q,
then the off-diagonal elements are n2 (p2 + q 2 ) = 2σ 2 n2 κ and it can be seen from a numerical
evaluation that it is approximately equal to (10) evaluated at µ = 1.
B. η − µ model
The η − µ is a fading distribution that represents small scale fading effects in non-line of
sight condition. The elements hij of H are independent and identical η − µ distributed random
variables with density [26],
x2
yij2
µ2µ |xij yij |2µ−1
ij
+
)
exp(−µ
fxy (xij , yij ) =
ΩµX ΩµY Γ2 (µ)
ΩX
ΩY
(12)
where Ω is the power parameter given by Ω = 2σ 2 µ, σ 2 is the power of the Gaussian variable
in each cluster, µ is the number of clusters. Note, ΩX = (1 − η)Ω/2, ΩY = (1 + η)Ω/2,
−1 ≤ η ≤ 1 and the diagonal elements zii are means of sums of η − µ envelope square
August 1, 2017
DRAFT
10
variables. i.e., zii = E[
elements are given by,
P n2
2
j=1 [xij
+ yij2 ]]. Hence by [7], zii = n2 (ΩX + ΩY ). The off-diagonal
n2
X
zij = E[ [(xik + jyik )(xkj − jykj )]]
k=1
Since xik and yik are i.i.d., we obtain ∀i, j and i 6= j,
n2
X
(E[xik ]2 + E[ykj ]2 ) = 0
zij =
k=1
The off-diagonal elements are zero, because distributions fx (xij ) and fy (yij ) are odd functions.
Therefore, Σ = (ΩX + ΩY )In1 . It is interesting to note that, the approximation doesn’t depend
on η.
C. Nakagami-m model
The elements hij of H are independent and identically distributed Nakagami-m variables with
density,
fxy (xij , yij ) =
m
mm |xij |m−1 |yij |m−1
exp(− (x2ij + yij2 ))
m
2
Ω Γ (m/2)
Ω
We can obtain a Nakagami random variable by substituting η = 0 i.e., ΩX = ΩY = Ω/2 and
m = 2µ in η − µ random variable given by (12). It can also be obtained by substituting κ = 0
i.e., p = q = 0 and Ω = 2µσ 2 in κ−µ random variable given by (6). Hence, both the approaches
yield the same result, i.e., Σ = ΩIn1 . It is interesting to note that, existing work [22] has analyzed
MIMO ergodic capacity of correlated Nakagami-m fading channels and derived a joint pdf of
eigen-values of HHH using copula. But the analysis is performed only for 2 × 2 channel matrix
and it becomes fairly difficult even for a 3 × 3 channel matrix.
III. A PPLICATIONS
OF THE APPROXIMATION
In this section, to demonstrate the utility of our work, we apply the above approximation in
two very different applications namely, finding MIMO channel capacity for κ − µ/η − µ faded
channel coefficients and finding SER expressions for optimum combining with κ − µ/η − µ
faded interferers. Finding ergodic MIMO channel capacity involves finding the expectation of
log determinant of Gram matrix HHH , where entries of H are κ − µ/η − µ faded. On the other
hand, finding SER expressions for optimum combining involves determining the SINR given by
η = cH R−1 c, where R = HHH is the Gram matrix formed by κ − µ/η − µ faded interferers
and c is the user signal.
August 1, 2017
DRAFT
11
A. MIMO channel Capacity
We consider an NR × NT MIMO channel matrix H, where NR denotes the number of receive
antennas and NT denotes the number of transmit antennas. Let x be the NT × 1 transmitted
vector and n be the NR ×1 zero mean i.i.d. complex Gaussian noise vector. The NR ×1 received
vector y is given by, y = Hx+n. Assuming that the transmitter has no channel state information
(CSI), the capacity of the MIMO channels when the transmitter has no CSI, is given by [3],
C ′ = log2 det(I +
ρ
HHH ),
NT
(13)
where ρ is the average signal to noise ratio (SNR) per receiving antenna. Since HHH and HH H
have the same non-zero eigenvalue statistics, from [3],
′
C =
n1
X
i=1
log2 (1 +
ρ
λi ),
NT
where n1 = min(NR , NT ) and λ1 , ...., λn1 are the non-zero eigenvalues of R, which is given by,
HHH if NR ≤ NT
R=
HH H if NR > NT .
Hence, the mean value of C ′ is given by [2],
n1
hX
i
ρ
C = EΛ
log2 (1 +
λi ) .
NT
i=1
We do not know the exact eigenvalue distribution of R, when H comprises i.i.d. κ − µ or η − µ
variables. Hence, we apply the Wishart approximation developed in the last section and then
determine C. We approximate R by a Wishart matrix with a degree of freedom NT and NR ×NR
covariance matrix Σ, given by (11) for κ − µ interferers and (ΩX + ΩY )INR for η − µ interferers.
1) κ − µ: The approximate expression for C is derived in Appendix. B. For NT ≥ NR , by
substituting n2 = NT and n1 = NR in (44), we can get the average capacity approximation as,
C ≈ (−1)
1
N (NR −1)
2 R
1
QNR
1
NT (w − w )NR −1
2
1
j=1 (NT − j)! (detΣ)
QNR −2
j=1
j!
NX
R −1
k=1
|N k |,
(14)
where N k is given in (45) and w1 = 2σ 2 (1 + κ)µ − y and w2 = 2σ 2 (1 + κ)µ + (NR − 1)y are
the eigenvalues of Σ−1 with multiplicity NR − 1 and 1 respectively.
In case NT < NR , we approximate HH H instead of HHH since both have the same non-zero
eigenvalues. We therefore approximate HH H by a central Wishart W ∼ CW NI (NR , Σ). For
August 1, 2017
DRAFT
12
NT ≤ NR , by substituting n2 = NR and n1 = NT in (44), we can get the capacity approximation
as,
1
C ≈ (−1) 2 NT (NT −1) QNT
1
1
j=1 (NR − j)!
k
|Σ|NR (w
2
− w1
)NT −1
2
QNT −2
j=1
2
j!
NX
T −1
k=1
|N k |,
(15)
where N is given in (45) and w1 = 2σ (1 + κ)µ − y and w2 = 2σ (1 + κ)µ + (NT − 1)y are
the eigenvalues of Σ with multiplicity NT − 1 and 1 respectively.
2) η−µ: If R = HHH with H having i.i.d. η−µ elements, we use the Wishart approximation
of R and follow a procedure similar to that used for κ−µ. For NT ≥ NR , by substituting n2 = NT
and n1 = NR in (46), we can obtain the capacity approximation as,
N
R
X
((ΩX + ΩY ))−NR NT
|N k |,
C ≈ QNR
QNR
i=1 (NR − i)! k=1
i=1 (NT − i)!
(16)
where N k is given by (47). Similarly for NT ≤ NR , by substituting n2 = NR and n1 = NT in
(46), we can obtain the capacity approximation as,
N
T
X
((ΩX + ΩY ))−NR NT
|N k |,
C ≈ QNT
QNT
i=1 (NT − i)! k=1
i=1 (NR − i)!
(17)
where N k is given by (47). Since, ΩX = (1 − η)Ω/2 and ΩY = (1 + η)Ω/2, the approximate
capacity expressions depends only on the power parameter Ω and not on the η parameter. In [5]
and [2], exact capacity expressions are derived for Rayleigh faded MIMO channels. The results
from these expressions match our η/µ capacity expressions for Rayleigh faded MIMO channels,
i.e., for η = 0 and µ = 1. Also, an upper bound for the ergodic capacity of κ − µ and η − µ
faded MIMO channels is derived in [23]. However, the upper bound requires computation of the
mean of each entry of H given by E[hij ], for which a numerical computation is done in [23].
Hence, we can apply our mean approximation in [23] to evaluate the upper bound. The upper
bound is plotted in Section IV and compared with our theoretical approximation.
3) Asymptotics: Since we have approximated η − µ faded MIMO channels by a complex
Wishart matrix, a lot of existing properties and results of complex Wishart matrix can be exploited
to get interesting results for these channels. One such application is in determining the asymptotic
capacity of generalized fading channels, especially η −µ faded channel. The asymptotic capacity
of Rayleigh faded channels is studied in detail in [33]. We now use their analysis to study η − µ
asymptotics. For NT = NR = N, and η − µ fading, the capacity is given by
C = EΛ
N
X
i=1
August 1, 2017
log2 (1 +
ρ(ΩX + ΩY )
λi ),
N
DRAFT
13
where λi are the eigen values of H ∼ CN N (0, IN ). Using [33], we obtain the asymptotic capacity
as
where
Z ∞
N
ρ(ΩX + ΩY )
1 X
log2 (1 +
λi ) =
log2 (1 + ρ(ΩX + ΩY )λ)g(λ)dλ
lim EΛ
N →∞
N i=1
N
0
g(λ) =
q
1 1 −
π
λ
0
1
4
0≤λ≤4
(18)
(19)
o.w.
Solving the above integral, we obtain the asymptotic capacity as
N
ρ
ρ(ΩX + ΩY ) 3 F2 (1, 1, 3/2; 2, 3; −4ρ(ΩX + ΩY ))
1 X
log2 (1 + λi ) =
.
lim EΛ
N →∞
N i=1
N
ln2
(20)
where 3 F2 (.) is a Hypergeometric function. With a first order approximation of the logarithm
at low SNR as in [33],
1
lim
C≈
N →∞ N
Z
∞
ρ(ΩX + ΩY )λg(λ)dλ = ρ(ΩX + ΩY ).
0
Similarly at high SNR, we obtain from [33],
1
C ≈ log2 (ρ(ΩX + ΩY )/e).
N →∞ N
lim
Since (ΩX + ΩY ) = Ω = 2µσ 2 , the capacity grows as a linear function of µ and SNR ρ, at low
SNR and capacity grows as a logarithmic function of µ and SNR ρ, at high SNR.
For κ − µ random variables, the capacity is given by
C = E[log2 det(I +
ρ
HHH )],
N
(21)
where H ∼ CN N (0, Σ) and w1 = 2σ 2 (1 + κ)µ − y and w2 = 2σ 2 (1 + κ)µ + (N − 1)y are the
eigenvalues of Σ with multiplicity N − 1 and 1. Unlike η − µ random variable, it is difficult to
obtain the asymptotic capacity like in (20) for all SNR values, due to the presence of correlation
matrix Σ. Hence, we will derive approximate asymptotic capacity only at high SNR. At high
SNR, the capacity is given by
ρ
ρ
HHH )] = E[log2 det( Σ1/2 H′ H′H Σ1/2 )]
N
N
ρ ′ ′H
= E[log2 det( H H )] + log2 det(Σ)
N
N
X
ρ
= EΛ
log2 ( λi ) + log2 (w1N −1 w2 ).
N
i=1
C ≈ E[log2 det(
August 1, 2017
DRAFT
14
where H′ ∼ CN N (0, I). Therefore, the asymptotic capacity is given by,
N
N −1
1
ρ
1 X
1
log2 ( λi ) + lim
C ≈ lim EΛ
log2 (w1 ) + lim
log2 (w2 )
lim
N →∞
N →∞
N →∞ N
N →∞ N
N i=1
N
N
Z 4
ρ
=
log2 (ρλ)g(λ) + log2 (w1 ) + 0 = log2 ( ) + log2 (w1 ).
e
0
(22)
From the above equation it is clear that, at high SNR, capacity grows as a logarithmic function
of SNR ρ.
B. SER Optimum combining
One other application where our approximation can be used is in determining SER expressions
for OC with κ − µ/η − µ interferers. Though there exist some results that compute bounds for
the capacity of κ − µ/η − µ faded MIMO channels, there exists no such prior literature for
OC, to the best of our knowledge, where the interferers are κ − µ/η − µ faded. Let c denote
the NR × 1 channel from the desired transmitter to the user, ci denote the NR × 1 channel
from the ith interferer to the user, x denotes the desired user symbol belonging to unit energy
QAM constellation and xi denote the ith interferer symbol also belonging to a unit energy QAM
constellation. The NR × 1 received vector is given by,
y = cx +
NI
p X
EI
ci xi + n,
(23)
i=1
where EI is the mean interferers power, n is the NR × 1 additive white complex Gaussian noise
vector with power σ 2 per dimension, i.e., n ∼ CN (0, σ 2 INR ) and NI denotes the number of
interferers. The user channel is modeled as i.i.d. Rayleigh i.e., c ∼ CN (0, INR ). The interferer
channels are modeled as equal power i.i.d. κ − µ or η − µ. The NR × NR covariance matrix of
the interference term is,
NI
H
I
R =E[(ΣN
i=1 ci xi )(Σi=1 ci xi ) ].
(24)
In order to derive the expression for SER, we first consider the expression for SINR of OC given
by [34],
η=
σ 2 −1
1 H
c (R +
I) c.
EI
EI
(25)
Let λ1 , λ2 , ..., λNR denote the eigenvalues of R. Then, R = UΛUH by eigen-value decomposition, where U is the matrix composed of orthonormal eigen vectors, corresponding to the
August 1, 2017
DRAFT
15
eigenvalues of R and Λ is the diagonal matrix of eigenvalues. The received SINR is now given
by,
η=
1 H
σ 2 −1
1 H
c (R + σ 2 I)−1 c =
I) c̃,
c̃ (Λ +
EI
EI
EI
(26)
where c̃ = UH c. Defining pk = |c̃k |2 ,
η=
NR
X
k=1
pk
EI
λk +
σ2
EI
.
(27)
Since c̃ is spherically invariant, it will have the same distribution as c. Since ck are i.i.d. complex
Gaussian with zero means, pk are i.i.d. exponential random variables with unit means. Using
the standard assumption that the contribution of the interference and the noise at the output of
optimal combiner, for a fixed η, can be well-approximated to be Gaussian, as in [35] and [36]
and references therein, the probability of symbol error for an M-ary square QAM constellation
is given by [37],
Pe = k1 Q(
p
k2 η) − k3 Q(
k2
p
k2 η)2 ,
(28)
where k1 = 4(1 − √ 1 ), k2 = M3−1 , k3 = 41 and the Q-function is given by Q(x) =
(M )
R ∞ −u2 /2
1
e
du. The assumption is valid even when the number of interferers NI is small [36]
2π x
and such a system model assumption is made in a number of papers [34], [38], [39] to derive
the SER expression. Using the popular approximation Q(x) ≈
can write Pe as,
Pe =
5
X
1 − 12 x2
e
12
2
2
+ 14 e− 3 x from [40], one
al e−bl η ,
(29)
l=1
3
3
3
where a1 = k121 , a2 = k41 , a3 = −k
, a4 = −k
, a5 = −k
, b1 = k22 , b2 = 2k32 , b3 =k2 , b4 = 4k32 and b5 = 7k62 . The
144
16
24
exponential approximation of the Q-function is shown to be tight in [40] and a similar approximation is used in [41], [42]. The average SER obtained by averaging Pe over all channel
realizations is derived as follows:
5
5
X
X
−bl η
al Eη [e−bl η ].
al e
]=
SER = Eη [Pe ] = Eη [
l=1
(30)
l=1
Substituting for η from (27) in the above equation and also rewriting the expectation over η
using the fact that Λ and p = [p1 p2 ... pNR ] are independent, we get,
pk
##
" "
PNR
EI
5
−b
X
l
2
k=1
σ
λk +
EI
.
SER =
al EΛ Ep e
(31)
l=1
August 1, 2017
DRAFT
16
Each of the pk is an independent exponential random variable and the m.g.f. of an exponential
random variable X with mean ω is given by E[etX ] =
ω
ω−t
for t < ω. Hence, we can write (31)
as,
SER =
5
X
al EΛ Ep
NR
Y
−bl
[e
pk
EI
2
λk + σ
EI
k=1
l=1
] =
5
X
al EΛ
l=1
The problem reduces to determining an expectation EΛ
NR
Y
λk +
λ +
k=1 k
Qn1
σ2
EI
+
bl
EI
2
σ
λk + E
I
k=1 λ + σ 2 + bl
k E
E
I
and λ1 , λ2 , ..., λn1 denote the eigenvalues of R. Let J(l, n1 ) = EΛ
σ2
EI
I
Qn1
.
(32)
, where l = 1, ..., 5
2
σ
λk + E
I
k=1 λ + σ 2 + bl
k E
E
I
I
. For this
case also, we use the Wishart approximation of R, i.e., we approximate R to a Wishart matrix
with degree of freedom NI and NR × NR covariance matrix Σ.
1) κ − µ: The approximate expression for J(l, n1 ) is given in (49) in Appendix C. For
NI ≥ NR , we can get the SER approximation directly by substituting n2 = NI and n1 = NR in
the approximation for J(l, n1 ) given in (49) in Appendix C as,
SER ≈
5
X
l=1
|M(NR , NI )|
QNR −2 ,
NI (w − w )NR −1
(N
−
j)!
|Σ|
I
2
1
j=1
j=1 j!
1
al (−1) 2 NR (NR −1) QNR
1
where M matrix is given in (50) and w1 = 2σ 2 and w2 = NR 2σ 2 (1 + κ)µ − (NR − 1)2σ 2 are
the eigenvalues of Σ−1 with multiplicity NR − 1 and 1 respectively.. For NI ≤ NR , the number
of non-zero eigenvalues of HHH is only NI . Hence,
5
X
al σ2
SER =
l=1
EI
σ2
EI
+
bl
EI
NR −NI
EΛ
NI
Y
λk +
λ +
k=1 k
σ2
EI
σ2
EI
+
bl
EI
.
(33)
Hence, for NI ≤ NR , we apply the same logic that was applied in the capacity calculations for
NT ≤ NR . We thus obtain the SER approximation from J(l, n1 ) in (49), but with n2 = NR and
n1 = NI ,
5
X
1
1
bl NI −NR
|M(NI , NR )|
(−1) 2 NI (NI −1) QNI
al 1 + 2
SER ≈
QNI −2 ,
NI −1
σ
(N
−
j)!
(w
−
w
)
R
2
1
j=1
j=1 j!
l=1
(34)
where M matrix is given in (50) and w1 ≈ 2σ 2 and w2 ≈ NT 2σ 2 (1 + κ)µ − (NT − 1)2σ 2 are
the eigenvalues of Σ−1 with multiplicity NT − 1 and 1 respectively..
August 1, 2017
DRAFT
17
2) η − µ: For NI ≥ NR , we can get the SER approximation directly by substituting n2 = NI
and n1 = NR in (51) in Appendix C as,
5
X
SER ≈
l=1
((ΩX + ΩY ))−NI NR
QNR
i=1 (NR − i)!
i=1 (NI − i)!
al QNR
× det {Γ(NI − NR + i + j − 1)(ΩX + ΩY )NI −NR +i+j−2
σ2
bl
σ2
b
!
+ ElI
bl ΩEI ++ΩEI
EI
× (ΩX + ΩY −
e X Y ENI −NR +i+j−1 (
))}
EI
ΩX + ΩY
1 ≤ i, j ≤ NR .
(35)
For NI ≤ NR , the number of non-zero eigenvalues of HHH is NI and NR −NI zero eigenvalues.
Hence,
5
X
SER =
al σ2
EI
l=1
σ2
EI
+
bl
EI
NR −NI
EΛ
NI
Y
k=1
λk +
λk +
σ2
EI
σ2
EI
+
bl
EI
.
For NR ≤ NI , we obtain the SER approximation by using (51) in Appendix C, but with n2 = NR
and n1 = NI . Hence,
5
X
bl NI −NR
((ΩX + ΩY ))−NI NR
al 1 + 2
SER ≈
QNI
QNI
σ
(N
−
i)!
R
i=1 (NI − i)!
i=1
l=1
× det {Γ(NR − NI + i + j − 1)(ΩX + ΩY )NR −NI +i+j−2
σ2
bl
σ2
b
+ ElI
bl ΩEI ++ΩEI
E
e X Y ENR −NI +i+j−1 ( I
))}
× (ΩX + ΩY −
EI
ΩX + ΩY
!
1 ≤ i, j ≤ NI .
(36)
Since ΩX = (1 − η)Ω/2 and ΩY = (1 + η)Ω/2, similar to the capacity case, the approximate
SER expressions depends only on the power parameter Ω and not on the η parameter.
IV. N UMERICAL R ESULTS
AND
S IMULATIONS
A. Capacity
The derived capacity expressions are verified using Monte-Carlo simulations for both NR ≥
NT and NR ≤ NT . For each Monte-Carlo simulation, the NR × NR random matrix HHH is
generated such that H has i.i.d. κ − µ or η − µ complex variables following the distribution
that is given in [25], [26]. For a given SNR ρ and NT , capacity is evaluated using (13). This
procedure is repeated over many realizations of HHH and the mean is taken to obtain the average
capacity. The approximate average capacity value is obtained by using the expressions (14) and
August 1, 2017
DRAFT
18
45
25
Simulation
Theoretical Approximation
Simulation
Theoretical Approximation
Upper Bound from [23]
40
N R=6, SNR=20 dB
Capacity in bps/Hz
Capacity in bps/Hz
35
30
N R=6, SNR=10 dB
25
N R=2, SNR=20 dB
20
20
µ=1, 3, 5, 7
N R=6, NT=6
N R=3, NT=6
15
N R=2, SNR=10 dB
15
µ=1, 3, 5, 7
10
5
2
3
4
5
6
7
8
9
10
10
1
2
3
4
NT
5
6
7
8
κ
(a) Capacity vs NT for κ = 4 and µ = 3
(b) Capacity vs κ for SNR= 10dB
Fig. 1: Capacity for varying κ ,µ, NR and NT
12
24
Simulation
Theoretical Approximation
Existing result [5]
22
10
20
µ=9
Capacity in bps/Hz
Capacity in bps/Hz
18
16
14
12
10
8
µ=5
6
4
µ=1
8
2
µ=1, 3, 5, 7
6
4
Simulation
Asymptotic capacity
High SNR approximation
5
10
15
20
25
0
5
10
SNR in dB
15
20
25
SNR in dB
(a) Capacity vs SNR ρ for NR = 4, NT = 2, η = 0
(b) Asymptotic capacity vs SNR ρ for
NR = 8, NT = 8, η = 0.3
Fig. 2: Capacity vs SNR ρ
(15) for NR ≥ NT and NR ≤ NT respectively for κ − µ case. Similarly, the approximate average
capacity value is obtained by using the expressions (16) and (17) for NR ≥ NT and NR ≤ NT
respectively for η − µ case. This procedure is repeated for various values of κ/η, µ, NR and NT .
A close match is found between the theoretical and simulation results for all the cases as can
be seen from the Fig. 1- Fig. 2.
It can be observed from Fig. 1 (a), that capacity increases with NT , for a fixed NR , but
August 1, 2017
DRAFT
19
10
0
10
0
Simulation
Theoretical Approximation
10
Simulation
Theoretical Approximation
-1
10
-1
64 QAM
10
10 -2
SER
SER
10 -2
-3
κ=1, µ=1 κ=1, µ=5
κ=1, µ=9
10 -4
10
16 QAM
-3
10 -4
κ=5, µ=9
κ=9, µ=9
10
-5
10 -6
10
5
10
15
20
10 -6
25
4 QAM
-5
5
10
SNR in dB
15
20
25
SNR in dB
(a) NR = 2, NI = 1, EI = −20dB
(b) NR = 10, NI = 6, κ = 5, µ = 7, EI = −1dB
Fig. 3: SER vs SNR for κ − µ
10 0
10 0
Simulation
Theoretical Approximation
Simulation
Theoretical Approximation
64 QAM
10 -1
µ=9
E I =-10 dB
16 QAM
SER
SER
µ=5
µ=1
10 -2
10 -1
µ=5
10 -3
µ=1
4 QAM
E I =-20 dB
10 -4
5
10
15
20
25
10 -2
5
SNR in dB
10
15
20
25
SNR in dB
(a) 4 − QAM, NR = 2, NI = 1, η = −0.3
(b) NR = 3, NI = 4, η = 0.3, µ = 5, EI = −17dB
Fig. 4: SER vs SNR for η − µ
saturates for large values of NT . For any further increase in capacity one has to increase either
NR or the SNR. From Fig. 1 (b) and Fig. 2 (a), it can be seen that the average capacity, increases
with increase in the number of clusters µ of κ − µ or η − µ distribution. But the increase is
diminished as µ increases. Similarly, the asymptotic capacity increases with increase in the
number of clusters µ of η − µ distribution, as seen in Fig. 2 (b). Also, the average capacity
increases with κ, as observed in Fig. 1 (b). The capacity upper bound from [23] is plotted in
Fig. 1 (a). Similarly, the existing results for Rayleigh faded MIMO channels from [5] are plotted
August 1, 2017
DRAFT
20
in Fig. 2 (a) and a close match with our η − µ results are observed for η = 0 and µ = 1.
B. Optimum combining
The derived SER expressions are verified using Monte-Carlo simulations for both NR ≥ NI
and NR ≤ NI . For each Monte-Carlo simulation, the random matrix R = HHH is generated,
where H has i.i.d. κ − µ or η − µ complex variables following the distribution that is given
in [25], [26]. R is decomposed into its eigen-values λ1 , λ2 , ..., λNR and exponential random
variables with unit mean, pk for k = 1, ..., NR , are generated for the user channel. For a given
noise value σ 2 , SINR η is evaluated using (27) and is substituted in (28), to obtain the exact
probability of error over one iteration. This procedure is repeated over many realizations of R and
the exponential random variables pk and the average of all these values is taken to get the final
SER. Instead of using (28) to compute the probability of error, one can use the approximation
given in (29) and average over many realizations of C and pk to get the final SER.
The approximate SER value is obtained by using the expressions (33) and (34) for NR ≥ NI
and NR ≤ NI respectively for the κ − µ case. Similarly, the approximate SER value is obtained
by using the expressions (35) and (36) for NR ≥ NI and NR ≤ NI respectively for the η − µ
case. This procedure is repeated for various values of κ or η, µ, NR , NI and EI . A close match
is found between the theoretical and simulation results for all the cases as can be seen from Fig.
3 and Fig. 4.
We can observe from Fig. 3(a) that SER increases with increase in κ or µ. As we keep κ
constant and increase µ, the increase in SER diminishes as µ becomes larger. The same can be
said for an increase in κ with µ kept constant. Even for the case of η − µ, we can observe from
Fig. 4(a) that, the SER increases as there is an increase in either EI or µ. As µ increases, the
increase in SER also diminishes, as seen from the plots for µ = 1, 5 and 9, for EI = −10dB.
V. C ONCLUSIONS
Approximate random matrix models have been derived for HHH when the elements of H are
i.i.d κ − µ or η − µ random variables. The approximation is terms of a complex Wishart matrix
having the same first moment as the original matrix distribution with the degree of freedom being
constrained to the number of columns of H. The utility of our result is shown by a) deriving
approximate capacity expressions for κ − µ or η − µ MIMO models b) deriving approximate
expressions for the SER of an optimum combining system with Rayleigh faded users and κ − µ
August 1, 2017
DRAFT
21
or η − µ faded interferers. For both these applications, extensive Monte-Carlo simulations have
been performed and an excellent match with the approximate expressions has been observed.
A PPENDIX A
A PPROXIMATE
MEAN OF COMPLEX
κ − µ RANDOM
VARIABLES
The expectations to be approximated are,
Z ∞
(x − p)2
px
|px|
|x|µ/2
)sech( 2 )I µ2 −1 ( 2 )dx
E[xik ] =
x 2 µ/2−1 exp(−
2
2σ
σ
σ
−∞ 2σ |p|
Z
E[yik ] =
∞
−∞
y
|qy|
|y|µ/2
(y − q)2
qy
exp(−
)sech( 2 )I µ2 −1 ( 2 )dy.
2
µ/2−1
2
2σ |q|
2σ
σ
σ
(37)
(38)
The expectation E[xik ] is rewritten, using the trigonometric identity tanh(z) = 1 − e−z sech(z),
as,
E[xik ] = 2
Z
0
∞
x2
p2
px
px
xµ/2+1
µ
exp(−
)exp(−
)tanh(
)I
)dx.
(
−1
2σ 2 |p|µ/2−1
2σ 2
2σ 2
σ2 2
σ2
The above integral cannot be solved to obtain a solution in closed form. Alternatively, we can
√
approximate like in [43], tanh( σpx2 ) by erf ( 2π σpx2 ) to obtain,
√
Z ∞
x2
p2
px
xµ/2+1
π px
exp(− 2 )exp(− 2 )erf (
)I µ2 −1 ( 2 )dx.
E[xik ] ≈ 2
2
µ/2−1
2
2σ |p|
2σ
2σ
2 σ
σ
0
2
1
( 2z )v 0 F1 (v + 1, z4 ) from [44], we get,
Using the identity Iv (z) = Γ(v+1)
√
Z ∞
x2
p2
xµ/2+1
µ p2 x2
px µ −1
π px 1
2
E[xik ] ≈ 2
exp(−
)exp(−
)erf
(
)
)
,
)dx.
F
(
(
0
1
µ
2σ 2 |p|µ/2−1
2σ 2
2σ 2
2 σ 2 Γ( 2 ) 2σ 2
2 4σ 4
0
Expanding the hypergeometric series and interchanging the integration and summation, we
obtain,
√
µ
x 2 +1
x2
p2
π px 1
px
E[xik ] ≈ 2
exp(− 2 )exp(− 2 )erf (
) µ ( 2 )µ/2−1
2
µ/2−1
2
2σ |p|
2σ
2σ
2 σ Γ( 2 ) 2σ
0
∞
X 1
p2 x2 n
(
×
) dx
µ
4
)
n!
(
4σ
n
2
n=0
√
Z ∞
∞
X
π px
p2n − p22 1
1
x2
µ+2n
=2
e 2σ
x
exp(−
)erf
(
)dx
µ
µ
µ
2
2
+2n
2) 2
)
n!
)
(
Γ(
2σ
2
σ
(2σ
n
0
2
2
n=0
R∞
2
2 2
Now using the integration identity 0 erf (ax)e−b x xp dx = √aπ b−p−2 Γ( p2 +1) 2F1 ( 21 , 2p +1, 23 , − ab2 )
Z
∞
for b2 > 0 and p > −2 from [45], we obtain,
∞
2 X
p2
1
3 2p2 π
1
− p2
( 2 )n µ
Γ(µ/2 + n + 1) 2 F1 ( , µ/2 + n + 1, , − 2 ).
E[xik ] = 2pe 2σ
2σ ( 2 )n n!Γ(µ/2)
2
2
σ 4
n=0
August 1, 2017
DRAFT
22
z
) for the Gauss HypergeUsing the transformation 2 F1 (a, b, c, z) = (1 − z)−b 2 F1 (c − a, b, c, z−1
ometric function from [46], we obtain,
−
p2
2σ 2
−
p2
2σ 2
E[xik ] ≈ 2pe
= 2pe
∞
X
3
2p2 π
p2 Γ( µ + n + 1)
µ
2p2 π − µ −n−1
2
)
+
n
+
1,
,
)
( 2 )n µ2
F
(1,
(1
+
2 1
µ
2 4
2 π + 4σ 2
2σ
(
σ
2
2
2p
)
n!Γ(
)
n
2
2
n=0
∞
X
2p2
Γ(µ/2 + n + 1)
4σ 2
µ/2+1
(
)
)n µ
( 2
2
2
2
4σ + 2p π
4σ + 2p π ( 2 )n n!Γ(µ/2)
n=0
3
2p2 π
× 2 F1 (1, µ/2 + n + 1, , 2
).
2 2p π + 4σ 2
Expanding the 2 F1 as series
2
− p2
2σ
E[xik ] = 2pe
2p2
2p2 π
Γ( µ2 + 1) X X ( µ2 + 1)n+k (1)k ( 4σ2 +2p2 π )n ( 2p2 π+4σ2 )k
µ
4σ 2
+1
( 2
)2
.
4σ + 2p2 π
Γ( µ2 ) n=0 k=0
n!
k!
( 32 )k ( µ2 )n
∞
∞
Rewriting the above using confluent Appell function Ψ1 [47],
Γ(µ/2 + 1)
4σ 2
)µ/2+1
2
2
4σ + 2p π
Γ(µ/2)
2p2
2p2 π
,
).
Ψ1 (µ/2 + 1, 1, 3/2, µ/2, 2
2p π + 4σ 2 4σ 2 + 2p2 π
p2
E[xik ] ≈ 2pe− 2σ2 (
(39)
Similarly,
Γ(µ/2 + 1)
4σ 2
)µ/2+1
2
2
4σ + 2q π
Γ(µ/2)
2q 2
2q 2 π
,
).
Ψ1 (µ/2 + 1, 1, 3/2, µ/2, 2
2q π + 4σ 2 4σ 2 + 2q 2 π
q2
E[yik ] ≈ 2qe− 2σ2 (
(40)
We have compared (39) and (40) with numerical evaluation of the expectation integrals and also
empirical average of simulated κ − µ variables for a wide range of parameters. In all cases, an
excellent match has been observed.
A PPENDIX B
κ − µ AND η − µ
P 1
We have to determine an approximation for C = EΛ [ ni=1
log2 (1 +
C APACITY
FOR
ρ
NT
λi )], where λk for
k = 1, .., n1 are eigenvalues of a n1 × n1 random matrix R = HHH , where H have i.i.d. κ − µ
or η − µ elements.
August 1, 2017
DRAFT
23
A. κ − µ
We approximate the matrix R by a n1 × n1 central Wishart matrix W ∼ CW n1 (n2 , Σ), such
that n1 ≤ n2 and Σ as in (11). The eigenvalue distribution of the unordered eigenvalues of W
is given by,
f (Λ) = (−1)
1
n (n −1)
2 1 1
n1
λnj 2 −n1
1 det({e−λi wj }) ∆(Λ) Y
n1 !
|Σ|n2
∆(Σ−1 ) j=1 (n2 − j)!
(41)
where w1 > w2 > .... > wn1 are the eigenvalues of Σ−1 and λ1 , ..., λn1 are the eigenvalues of
W. But if some eigenvalues of Σ−1 are not distinct, then the above distribution cannot be used
because det({e−λi wj }) = ∆(Σ−1 ) = 0 leading to an indeterminate form. Hence, we apply the
following theorem from [31], to modify the distribution and account for non-distinct eigenvalues.
Theorem 1. Let f1 , ..., fN be a family of infinitely differentiable functions and let x1 , ..., xN ∈ R.
Denote
det {fi (xj )}
R(x1 , .., xN ) , Q
.
i<j (xj − xi )
Then, for N1 , ..., Np such that N1 + ... + Np = N and for y1 , ..., yp ∈ R distinct,
lim
x1 ,...,xN1 →y1
....
xN−Np +1,..,xN →yp
R(x1 , ..., xN )
(N −1)
(N −1)
det fi (y1 ), fi′ (y1), ..., fi 1 (y1 ), ..., fi (yp ), fi′(yp ), ...., fi p (yp )
=
.
Q
Qp QNl −1
Ni Nj
j!
(y
−
y
)
j
i
j=1
1≤i<j≤p
l=1
In our case, Σ−1 has two eigenvalues w1 and w2 with multiplicity n1 − 1 and 1 respectively.
Hence, applying the above theorem to (41), we obtain, the eigenvalue distribution as,
1
n1
λnj 2 −n1
(−1) 2 n1 (n1 −1) det({e−λi w1 (−λi )e−λi w1 ... (−λi )n1 −2 e−λi w1 e−λi w2 }) ∆(Λ) Y
f (Λ) =
.
Q 1 −2
n1 !
|Σ|n2 j=1 (n2 − j)!
j!
(w2 − w1 )n1 −1 nj=1
(42)
Hence
C ≈ (−1)
1
n (n −1)
2 1 1
1
1
Qn1
Qn1 −2
n1 ! j=1 (n2 − j)! |Σ|n2 (w2 − w1 )n1 −1 j=1
j!
Z
0
n1
∞X
i=1
log2 (1 +
ρ
λi )
NT
× λnk 2 −n1 ∆(Λ)det({e−λi w1 (−λi )e−λi w1 ... (−λi )n1 −2 e−λi w1 e−λi w2 })dΛ.
August 1, 2017
DRAFT
24
From Theorem 3 in Appendix of [3], it can be observed that, for two arbitrary n1 × n1 matrices
Φ(y) and Ψ(y) with ij th elements φi (yj ) and Ψi (yj ), and two arbitrary functions ξ(.) and ξ ′ (.),
where y = [y1 y2 ... yn1 ]T , the following identity holds:
Q
PN ′
|Φ(y)||Ψ(y)| N
n=1 ξ(yn )
k=1 ξ (yk )dy1 ..dyn1
!
Z d
N
X
,
det
φi (y)Ψj (y)ξ(y)Uk,j (ξ ′ (y))dy
= N!
R
···
R
d≥yi ≥c
k=1
where, Uj,k (x) = x,
c
(43)
1≤i,j≤n1
if k = j and Uj,k (x) = 1,
if k 6= j. Applying the above identity, we
obtain,
n
C ≈ (−1)
1
n (n −1)
2 1 1
where
Nki,j (n1 , n2 ) =
1
X
1
1
Qn1
det(N k ].
Q
n1 −2
n
n
−1
2
1
(n
−
j)!
|Σ|
(w
−
w
)
j!
2
2
1
j=1
j=1
k=1
(44)
R ∞ n −n i−1
λ 2 1 λ (−λ)j−1 e−λw1 dλ; 1 ≤ i ≤ n1 , 1 ≤ j ≤ n1 − 1, j 6= k,
0
R∞
log2 (1 + NρT λ)λn2 −n1 λi−1 (−λ)j−1 e−λw1 dλ;
0
1 ≤ i ≤ n1 , 1 ≤ j ≤ n1 − 1, j = k,
R ∞ n −n i−1 −λw
λ 2 1 λ e 2 dλ; 1 ≤ i ≤ n1 , j = n1 , j 6= k,
0
R ∞ log (1 + ρ λ)λn2 −n1 λi−1 e−λw2 dλ; 1 ≤ i ≤ n , j = n , j = k.
1
1
2
NT
0
First writing the logarithm in terms of Meijer-G function using the identity ln(1 + x) =
R ∞ v−1 −µx
1,2 1,1
G2,2
e dx = Γ(v)µ−v and
1,0 x [48] and solving the integrals using identities 0 x
R ∞ −ρ −βx m,n a1 ,...,ap
m,n+1 ρ,a1 ,...,ap α
from [49], we obtain,
x e
Gp,q b1 ,...,bq αx dx = β ρ−1 Gp+1,q
b1 ,...,bq
β
0
(−1)j−1 Γ(n2 − n1 + i + j − 1)w1−n2 +n1 −i−j+1;
1 ≤ i ≤ n1 , 1 ≤ j ≤ n1 − 1, j 6= k,
1,3 1−n2 +n1 −i−j+1,1,1
ρ
(−1)j−1 1 G3,2
w1−n2+n1 −i−j+1;
1,0
ln2
NT w 1
k
Ni,j (n1 , n2 ) =
1 ≤ i ≤ n1 , 1 ≤ j ≤ n1 − 1, j = k,
Γ(n2 − n1 + i)w2−n2 +n1 −i ; 1 ≤ i ≤ n1 , j = n1 , j 6= k,
1,3 1−n2 +n1 −i,1,1
ρ
1 G3,2
w2−n2 +n1 −i ; 1 ≤ i ≤ n1 , j = n1 , j = k.
1,0
ln2
NT w 2
(45)
August 1, 2017
DRAFT
25
B. η − µ
We approximate the matrix R by a n1 × n1 central Wishart matrix W ∼ CW n1 (n2 , cIn1 ) such
that n1 ≤ n2 . The eigenvalue distribution of the unordered eigenvalues of W is given by,
n
n
1
1
Y
1X
(c)−n1 n2
Qn1
λi ),
λni 2 −n1 |V (Λ)|2 exp(−
c i=1
i=1 (n2 − i)!
i=1 (n1 − i)! i=1
f (Λ) = Qn1
where λ1 , ..., λn1 are the eigenvalues of W and the term V (Λ) denotes the Vandermonde matrix
formed by the eigenvalues. For our case, c = (ΩX + ΩY ). Hence,
Pn1
Z ∞X
n1
n1
λ
Y
i=1 i
ρ
((ΩX + ΩY ))−n1 n2
n2 −n1
2 − (ΩX +ΩY )
Qn1
C≈
[
log2 (1 +
λi )] Qn1
dΛ
λi
|V (Λ)| e
N
(n
−
i)!
(n
−
i)!
T
2
1
0
i=1
i=1
i=1
i=1
Z ∞X
n1
−n1 n2
λi
((ΩX + ΩY ))
ρ
−
Qn1
≈ Qn1
λi )]λni 2 −n1 |V (Λ)|2 e ΩX +ΩY dΛ.
[
log2 (1 +
NT
i=1 (n2 − i)!
i=1 (n1 − i)! 0
i=1
By applying Theorem 3 in Appendix of [3], we obtain,
n
where
1
X
((ΩX + ΩY ))−n1 n2
Qn1
|N k |,
C ≈ Qn1
i=1 (n2 − i)!
i=1 (n1 − i)! k=1
(46)
λ
R
∞ λn2 −n1 λi+j−2e− ΩX +ΩY dλ; 1 ≤ i, j ≤ n1 , j 6= k,
0
k
Ni,j (n1 , n2 ) = R
λ
∞ log2 (1 + ρ λ)λn2 −n1 λi+j−2 e− ΩX +ΩY dλ; 1 ≤ i, j ≤ n1 , j = k.
NT
0
Solving similar to the κ − µ case, we obtain,
Γ(n2 − n1 + i + j − 1)(ΩX + ΩY )n2 −n1 +i+j−1 ; 1 ≤ i, j ≤ n1 , j 6= k,
1,3 1−n2 +n1 −i−j+1,1,1 ρ(ΩX +ΩY
1
Nki,j (n1 , n2 ) =
G
(ΩX + ΩY )n2 −n1 +i+j−1;
1,0
ln2 3,2
NT
1 ≤ i, j ≤ n1 , j = k.
(47)
A PPENDIX C
κ−µ
Qn1
We have to determine an approximation for J(l) = EΛ
k=1
O PTIMUM
COMBINING FOR
2
σ
λk + E
σ2
λk + E
I
I
b
+ El
I
, where λk for
k = 1, .., n1 are eigenvalues of a n1 × n1 random matrix R = HHH , where H have i.i.d. κ − µ
or η − µ elements.
August 1, 2017
DRAFT
26
A. κ − µ
We approximate the matrix R by a n1 × n1 central Wishart matrix W ∼ CW n1 (n2 , Σ) such
that n1 ≤ n2 and Σ as in (11). Σ−1 has two eigenvalues w1 and w2 with multiplicity n1 − 1
and 1 respectively. Hence, using the eigenvalue distribution from (42),
2
Z ∞Y
n1
λk + Eσ I
1
1
(−1) 2 n1 (n1 −1)
J(l) ≈
bl
σ2
n1 !
λ + EI + EI
0
k=1 k
n1
λnj 2 −n1
det({e−λi w1 (−λi )e−λi w1 ... (−λi )n1 −2 e−λi w1 e−λi w2 }) ∆(Λ) Y
×
dΛ
Q 1 −2
|Σ|n2 j=1 (n2 − j)!
j!
(w2 − w1 )n1 −1 nj=1
2
Z ∞Y
n1
λk + Eσ I
1
1
1
n
(n
−1)
Q 1
≈ (−1) 2 1 1
Qn1 −2
2
n1 ! nj=1
(n2 − j)! |Σ|n2 (w2 − w1 )n1 −1 j=1
j! 0 k=1 λk + Eσ I +
bl
EI
× λnk 2 −n1 ∆(Λ)det({e−λi w1 (−λi )e−λi w1 ... (−λi )n1 −2 e−λi w1 e−λi w2 })dΛ.
From Theorem 2 in Appendix of [3], it can be observed that, for two arbitrary n1 × n1 matrices
ρ(y) and Ψ(y) with ij th elements ρi (yj ) and Ψi (yj ), and an arbitrary function ρ(.), where
y = [y1 y2 ... yn1 ]T , the following identity holds:
R
···
R
d≥yi ≥c
|Ξ(y)||Ψ(y)|
Qn1
k=1 ρ(yk )dy1 ..dyn1
= n1 !det
n
R
d
ξ (y)Ψj (y)ρ(y)dy
c i
o
1≤i,j≤n1
.
(48)
Using the above relation to simplify the expectation, we obtain,
1
1
Qn1 −2 det(M)
n2
n1 −1
j=1 (n2 − j)! |Σ| (w2 − w1 )
j=1 j!
1
J(l) ≈ (−1) 2 n1 (n1 −1) Qn1
where M is given by,
R∞
0
Mi,j (n1 , n2 ) = R
∞
0
(49)
2
σ
λ+ E
2
I
b
σ
λ+ E
+ El
I
2
I
2
σ
λ+ E
I
b
σ
λ+ E
+ El
I
λn2 −n1 λi−1 (−λ)j−1 e−λw1 dλ;
λn2 −n1 λi−1 e−λw2 dλ;
I
1 ≤ i ≤ n1 , 1 ≤ j ≤ n1 − 1,
1 ≤ i ≤ n1 , j = n1 ,
R∞
which can be simplified using identities 0 xv−1 e−µx dx = Γ(v)µ−v from [49] as,
b
σ2
2
bl ( EI + ElI )w1
j−1 1−j−i+n1 −n2
e
(−1)
w
(1
−
w1 En2 −n1 +i+j−1 [( Eσ I + EblI )w1 ])
1
E
I
×Γ(n2 − n1 + i + j − 1); 1 ≤ i ≤ n1 , 1 ≤ j ≤ n1 − 1,
Mi,j (n1 , n2 ) =
b
σ2
−i+n1 −n2
bl ( EI + ElI )w2
σ2
w
(1
−
e
+ EblI )w2 ])Γ(n2 − n1 + i);
w
E
[(
2
n
−n
+i
2
1
2
E
E
I
I
1 ≤ i ≤ n1 , j = n1 .
(50)
where E.(.) is the exponential integral function [44].
August 1, 2017
DRAFT
27
B. η − µ
We approximate the matrix R by a n1 × n1 central Wishart matrix W ∼ CW n1 (n2 , cIn1 ) such
that n1 ≤ n2 . The eigenvalue distribution of the unordered eigenvalues of W is given by,
n
n
1
1
Y
1X
(c)−n1 n2
Qn1
λi )
λni 2 −n1 |V (Λ)|2 exp(−
c i=1
i=1 (n2 − i)!
i=1 (n1 − i)! i=1
f (Λ) = Qn1
where λ1 , ..., λn1 are the eigenvalues of W and the term V (Λ) denotes the Vandermonde matrix
formed by the eigenvalues. For our case, c = (ΩX + ΩY ). Hence,
2
Z ∞Y
n1
λk + Eσ I
((ΩX + ΩY ))−n1 n2
Q
Qn1
J(l, n1 ) ≈
n
2
1
λ + Eσ I + EblI i=1 (n2 − i)! i=1 (n1 − i)!
0
k=1 k
n1
Y
n
1
X
1
λi )dΛ
(Λ)| exp(−
(ΩX + ΩY ) i=1
i=1
2
Z ∞Y
n1
λi
λi + Eσ I
((ΩX + ΩY ))−n1 n2
n2 −n1
2 − ΩX +ΩY
Qn1
≈ Qn1
dΛ.
λ
|V
(Λ)|
e
2
i
b
σ
l
λ
+
+
i
i=1 (n2 − i)!
i=1 (n1 − i)! 0
EI
EI
i=1
λni 2 −n1 |V
2
Using Theorem 2 in Appendix of [3], we obtain for 1 ≤ i, j ≤ n1
2
Z ∞
λ + Eσ I
λ
((ΩX + ΩY ))−n1 n2
n2 −n1 i+j−2 − ΩX +ΩY
Qn1
dλ}).
det({
λ
λ
e
J(l, n1 ) ≈ Qn1
2
b
σ
λ + EI + ElI
0
i=1 (n2 − i)!
i=1 (n1 − i)!
R∞
This can be further simplified using identities 0 xv−1 e−µx dx = Γ(v)µ−v from [49] to obtain,
((ΩX + ΩY ))−n1 n2
Qn1
i=1 (n2 − i)!
i=1 (n1 − i)!
J(l, n1 ) ≈ Qn1
det({Γ(n2 − n1 + i + j − 1)(ΩX + ΩY )n2 −n1 +i+j−2
σ 2 + bl
σ2
b
+ ElI
bl ΩEI +ΩEI
E
e X Y En2 −n1 +i+j−1 ( I
))}1 ≤ i, j ≤ n1 ).
(ΩX + ΩY −
EI
ΩX + ΩY
(51)
where E.(.) is the exponential integral function [44].
R EFERENCES
[1] A. T. James, “Distributions of matrix variates and latent roots derived from normal samples,” Ann. Math. Statist., vol. 35,
no. 2, pp. 475–501, 06 1964. [Online]. Available: http://dx.doi.org/10.1214/aoms/1177703550
[2] E. Telatar, “Capacity of Multi-antenna Gaussian Channels,” European Transactions on Telecommunications, vol. 10,
no. 6, pp. 585–595, 1999. [Online]. Available: http://dx.doi.org/10.1002/ett.4460100604
[3] M. Chiani, M. Win, and A. Zanella, “On the capacity of spatially correlated MIMO Rayleigh-fading channels,” IEEE
Transactions on Information Theory, vol. 49, no. 10, pp. 2363–2371, Oct 2003.
[4] A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, “Capacity limits of MIMO channels,” IEEE Journal on Selected
Areas in Communications, vol. 21, no. 5, pp. 684–702, June 2003.
August 1, 2017
DRAFT
28
[5] M. Kang and M. S. Alouini, “Capacity of MIMO Rician channels,” IEEE Transactions on Wireless Communications,
vol. 5, no. 1, pp. 112–122, Jan 2006.
[6] M. R. McKay and I. B. Collings, “General Capacity Bounds for Spatially Correlated Rician MIMO Channels,” IEEE
Transactions on Information Theory, vol. 51, no. 9, pp. 3121–3145, Sept 2005.
[7] M. Yacoub, “The κ-µ distribution and the η-µ distribution,” Antennas and Propagation Magazine, IEEE, vol. 49, no. 1,
pp. 68–81, Feb 2007.
[8] D. B. D. Costa and M. D. Yacoub, “Average channel capacity for generalized fading scenarios,” IEEE Communications
Letters, vol. 11, no. 12, pp. 949–951, December 2007.
[9] D. Morales-Jimenez and J. F. Paris, “Outage probability analysis for η-µ fading channels,” IEEE Communications Letters,
vol. 14, no. 6, pp. 521–523, June 2010.
[10] J. F. Paris, “Outage Probability in η-µ/ η-µ and κ-µ /η-µ Interference-Limited Scenarios,” IEEE Transactions on
Communications, vol. 61, no. 1, pp. 335–343, January 2013.
[11] N. Y. Ermolova and O. Tirkkonen, “Outage Probability Analysis in Generalized Fading Channels with Co-Channel
Interference and Background Noise: η-µ/ η-µ, η-µ / κ-µ and κ-µ / η-µ Scenarios,” IEEE Transactions on Wireless
Communications, vol. 13, no. 1, pp. 291–297, January 2014.
[12] S. Kumar, G. Chandrasekaran, and S. Kalyani, “Analysis of Outage Probability and Capacity for κ-µ /η-µ Faded Channel,”
IEEE Communications Letters, vol. 19, no. 2, pp. 211–214, Feb 2015.
[13] S. Kumar and S. Kalyani, “Coverage Probability and Rate for κ-µ /η-µ Fading Channels in Interference-Limited Scenarios,”
IEEE Transactions on Wireless Communications, vol. 14, no. 11, pp. 6082–6096, Nov 2015.
[14] N. Bhargav, S. L. Cotton, and D. E. Simmons, “Secrecy Capacity Analysis Over κ-µ Fading Channels: Theory and
Applications,” IEEE Transactions on Communications, vol. 64, no. 7, pp. 3011–3024, July 2016.
[15] J. Zhang, Z. Tan, H. Wang, Q. Huang, and L. Hanzo, “The Effective Throughput of MISO Systems Over κ-µ Fading
Channels,” IEEE Transactions on Vehicular Technology, vol. 63, no. 2, pp. 943–947, Feb 2014.
[16] J. Zhang, M. Matthaiou, Z. Tan, and H. Wang, “Effective rate analysis of MISO η-µ fading channels,” in 2013 IEEE
International Conference on Communications (ICC), June 2013, pp. 5840–5844.
[17] D. Morales-Jimenez, J. F. Paris, and A. Lozano, “Outage Probability Analysis for MRC in κ-µ Fading Channels with
Co-Channel Interference,” IEEE Communications Letters, vol. 16, no. 5, pp. 674–677, May 2012.
[18] W. G. Li and M. Chen, “Outage capacity of dual-hop decode-and-forward relaying system over generalized fading channels,”
in 2010 2nd International Conference on Future Computer and Communication, vol. 3, May 2010, pp. V3–827–V3–831.
[19] M. K. Fikadu, P. C. Sofotasios, S. Muhaidat, Q. Cui, and M. Valkama, “SER of M-QAM decode-and-forward multi-relay
systems under generalized fading conditions,” in 2016 23rd International Conference on Telecommunications (ICT), May
2016, pp. 1–5.
[20] P. Kumar and K. Dhaka, “Performance Analysis of a Decode-and-Forward Relay System in κ-µ and η-µ Fading Channels,”
IEEE Transactions on Vehicular Technology, vol. 65, no. 4, pp. 2768–2775, April 2016.
[21] S. Kumar and A. Pandey, “Random Matrix Model for Nakagami-Hoyt Fading,” IEEE Transactions on Information Theory,
vol. 56, no. 5, pp. 2360–2372, May 2010.
[22] M. H. Gholizadeh, H. Amindavar, and J. A. Ritcey, “On the capacity of MIMO correlated Nakagami-m fading channels
using copula,” EURASIP Journal on Wireless Communications and Networking, vol. 2015, no. 1, p. 1, 2015.
[23] V. M. Vergara and S. E. Barbin, “MIMO capacity upper bound for κ-µ and η-µ faded channels,” in 2012 IEEE Radio and
Wireless Symposium, Jan 2012, pp. 367–370.
[24] L. Moreno-Pozas and E. Martos-Naya, “On Some Unifications Arising from the MIMO Rician Shadowed Model,” in 2016
IEEE 83rd Vehicular Technology Conference (VTC Spring), May 2016, pp. 1–5.
August 1, 2017
DRAFT
29
[25] U. S. Dias, M. D. Yacoub, and D. B. da Costa, “The κ-µ; phase-envelope joint distribution,” in 2008 IEEE 19th International
Symposium on Personal, Indoor and Mobile Radio Communications, Sept 2008, pp. 1–5.
[26] D. B. da Costa and M. D. Yacoub, “The η-µ; Joint Phase-Envelope Distribution,” in 2007 IEEE Wireless Communications
and Networking Conference, March 2007, pp. 1906–1908.
[27] T. Kollo and D. von Rosen, “Approximating by the Wishart distribution,” Annals of the Institute of Statistical Mathematics,
vol. 47, no. 4, pp. 767–783, 1995.
[28] W. Tan and R. Gupta, “On approximating the non-central wishart distribution by central wishart distribution a monte carlo
study,” Communications in Statistics-Simulation and Computation, vol. 11, no. 1, pp. 47–64, 1982.
[29] S. H.S. and R. J.J.J., “ Approximations for the non-central Wishart distribution,” South African Statistical Journal, vol. 6,
no. 2, pp. 165–172, Jan 1962.
[30] J. Winters, “Optimum Combining in Digital Mobile Radio with Cochannel Interference,” IEEE Journal on Selected Areas
in Communications, vol. 2, no. 4, pp. 528–539, July 1984.
[31] R. Couillet and M. Debbah, Random Matrix Methods for Wireless Communications.
Cambridge University Press, 2011.
[Online]. Available: https://books.google.co.in/books?id= j7pT9HjKAUC
[32] K. Granstrom and U. Orguner, “Properties and approximations of some matrix variate probability density functions,” 2011.
[33] C.-N. Chuah, D. N. C. Tse, J. M. Kahn, and R. A. Valenzuela, “Capacity scaling in MIMO wireless systems under
correlated fading,” IEEE Transactions on Information Theory, vol. 48, no. 3, pp. 637–650, Mar 2002.
[34] M. Chiani, M. Win, A. Zanella, R. Mallik, and J. Winters, “Bounds and Approximations for Optimum Combining of
Signals in the Presence of Multiple Cochannel Interferers and Thermal Noise,” IEEE Transactions on Communications,
vol. 51, no. 2, pp. 296–307, Feb 2003.
[35] H. V. Poor and S. Verdu, “Probability of error in MMSE multiuser detection,” IEEE Transactions on Information Theory,
vol. 43, no. 3, pp. 858–871, May 1997.
[36] A. Giorgetti and M. Chiani, “Influence of Fading on the Gaussian Approximation for BPSK and QPSK with Asynchronous
Cochannel Interference,” IEEE Transactions on Wireless Communications, vol. 4, no. 2, pp. 384–389, March 2005.
[37] J. Proakis and M. Salehi, Fundamentals of Communication Systems.
Pearson Education, 2007.
[38] M. Chiani, M. Win, A. Zanella, and J. Winters, “Exact Symbol Error Probability for Optimum Combining in the Presence
of Multiple Co-channel Interferers and Thermal Noise,” in Global Telecommunications Conference, 2001. GLOBECOM
’01. IEEE, vol. 2, 2001, pp. 1182–1186 vol.2.
[39] ——, “A Simple and Asymptotically Tight Upper Bound on the Symbol Error Probability of Adaptive Antennas with
Optimum Combining,” in . IEEE International Conference on Communications, 2002. ICC 2002, vol. 3, 2002, pp. 1878–
1881 vol.3.
[40] M. Chiani, D. Dardari, and M. K. Simon, “New Exponential Bounds and Approximations for the Computation of Error
Probability in Fading Channels,” IEEE Transactions on Wireless Communications, vol. 2, no. 4, pp. 840–845, July 2003.
[41] M. McKay, A. Zanella, I. Collings, and M. Chiani, “Error probability and SINR analysis of optimum combining in rician
fading,” IEEE Transactions on Communications, vol. 57, no. 3, pp. 676–687, March 2009.
[42] S. Menon and S. Kalyani, “SER for Optimal Combining in the Presence of Multiple Correlated Co-channel Interferers,”
IEEE Communications Letters, vol. PP, no. 99, pp. 1–1, 2015.
[43] M. A. Reyes and R. Arcos-Olalla, “Supersymmetric features of the Error and Dawson’s functions,” Revista mexicana de
fı́sica, vol. 61, no. 6, pp. 475–480, 2015.
[44] M. Abramowitz and I. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables,
ser. Applied mathematics series. U.S. Government Printing Office, 1972.
August 1, 2017
DRAFT
30
[45] E. W. Ng and M. Geller, “A table of integrals of the error functions,” Journal of Research of the National Bureau of
Standards B, vol. 73, no. 1, pp. 1–20, 1969.
[46] I.
Wolfram
Research.
Generalized
laguerre
polynomials.
[Online].
Available:
http://functions.wolfram.com/07.23.16.0002.01
[47] Y. A. Brychkov and N. Saad, “On some formulas for the Appell function F2 (a, b, b’; c, c’; w; z),” Integral Transforms
and Special Functions, vol. 25, no. 2, pp. 111–123, 2014.
[48] I.
Wolfram
Research.
Generalized
laguerre
polynomials.
[Online].
Available:
http://functions.wolfram.com/01.04.26.0003.01
[49] A. Jeffrey and D. Zwillinger, Table of Integrals, Series, and Products, ser. Table of Integrals, Series, and Products Series.
Elsevier Science, 2007.
August 1, 2017
DRAFT
| 7cs.IT
|
Face Aging with Contextual Generative Adversarial Nets
Si Liu
SKLOIS, IIE, CAS
[email protected]
arXiv:1802.00237v1 [cs.CV] 1 Feb 2018
Renda Bao
SKLOIS, IIE, CAS
School of Cyber Security, UCAS
roger [email protected]
Yao Sun∗
Defa Zhu
SKLOIS, IIE, CAS
[email protected]
SKLOIS, IIE, CAS
School of Cyber Security, UCAS
[email protected]
Wei Wang
Xiangbo Shu
University of Trento, Italy
[email protected]
Nanjing University of Science and
Technology
[email protected]
Shuicheng Yan
Qihoo 360 AI Institute, Beijing, China
National University of singapore
[email protected]
Age
ABSTRACT
Face aging, which renders aging faces for an input face, has attracted
extensive attention in the multimedia research. Recently, several
conditional Generative Adversarial Nets (GANs) based methods
have achieved great success. They can generate images fitting the
real face distributions conditioned on each individual age group.
However, these methods fail to capture the transition patterns,
e.g., the gradual shape and texture changes between adjacent age
groups. In this paper, we propose a novel Contextual Generative
Adversarial Nets (C-GANs) to specifically take it into consideration.
The C-GANs consists of a conditional transformation network and
two discriminative networks. The conditional transformation network imitates the aging procedure with several specially designed
residual blocks. The age discriminative network guides the synthesized face to fit the real conditional distribution. The transition
pattern discriminative network is novel, aiming to distinguish the
real transition patterns with the fake ones. It serves as an extra
regularization term for the conditional transformation network,
ensuring the generated image pairs to fit the corresponding real
transition pattern distribution. Experimental results demonstrate
the proposed framework produces appealing results by comparing with the state-of-the-art and ground truth. We also observe
performance gain for cross-age face verification.
KEYWORDS
Face Aging, Generative Adversarial Nets, Contextual Modeling
∗ corresponding
0‐10
11‐
11‐18
19‐
19‐29
30‐
30‐39
40‐
40‐49
50‐
50‐59
60+
groundtruth
2
1
Input face
3
4
5
6
7
synthetic
Conditional Transformation Network
1: Age
g discriminative network
…
Age
g group
g p1
2: Transition Pattern discriminative network
…
Group
p 6 ‐> ggroup
p7
Figure 1: The proposed C-GANs algorithm for face aging.
The input image can be transformed to any specific age
group. The synthesized results of C-GANs are natural due
to the two discriminative networks: the age discriminative
network modeling the distribution of each individual age
group, while the transition pattern discriminative network
modeling the correlations between adjacent groups.
ACM Reference format:
Si Liu, Yao Sun, Defa Zhu, Renda Bao, Wei Wang, Xiangbo Shu, and Shuicheng
Yan. 2017. Face Aging with Contextual Generative Adversarial Nets. In
Proceedings of MM ’17, Mountain View, CA, USA, October 23–27, 2017, 9 pages.
DOI: 10.1145/3123266.3123431
author
1
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
MM ’17, Mountain View, CA, USA
© 2017 ACM. 978-1-4503-4906-2/17/10. . . $15.00
DOI: 10.1145/3123266.3123431
INTRODUCTION
Face aging, also known as age progression [28, 35], is attracting
more and more research interests. It has plenty of applications in
various domains including cross-age face recognition [23], finding
lost children, and entertainments [40]. In recent years, face aging
has witnessed various breakthroughs and a number of face aging
models have been proposed [9]. Face aging, however, is still a very
challenging task in practice for various reasons. First, faces may
have many different expressions and lighting conditions, which
MM ’17, October 23–27, 2017, Mountain View, CA, USA
Si Liu, Yao Sun, Defa Zhu, Renda Bao, Wei Wang, Xiangbo Shu, and Shuicheng Yan
Age Discriminative Network
Transition Pattern Discriminative Network
Image
Age
Real/Fake face?
Real/Fake face pair?
Real
Pair
Real
image
Image
Fake Image
Fake
k
Pair
Paired Image
Discriminator
TTransforme
er
Discriminator
Age
Conditional Transformation Network
Age
Figure 2: The structure of the proposed C-GANs.
pose great challenges to modeling the aging patterns. Besides, the
training data are usually very limited and the face images for the
same person only cover a narrow range of ages.
Traditional face aging approaches can be roughly split into to
two classes, i.e., the prototyping ones [15, 38], and the modeling
ones [36, 37]. However, these approaches often require face aging
sequences of the same person with wide range of ages which are
very costly to collect. Generative Adversarial Networks (GANs) [10]
better deal with age progressions. Many GANs based methods [2,
46] can generate the most plausible and realistic images which are
hard to distinguish from real data conditioned on the age. However,
all of these methods do not make full use of the sequential data.
Therefore, these methods cannot explicitly consider the transition
patterns which are defined as the facial feature correlations between
different age groups for one person. Therefore, their results are
usually unable to maintain face identity, or cannot satisfy the crossaging transition rules well.
In this paper, we mainly consider the cross-age transition pattern. Specifically, transition pattern contains two aspects. One is
the identity consistency, and the other is the appearance changes.
Identity preserving is critical in face aging based applications, e.g.,
cross-age face verification. Appearance changes include texture
and shape alterations. Transition pattern is age-aware. For example, when one grows from baby to teenagers, the main appearance
difference is the face becomes larger. When one grows from the
age of 50 to 60, the main facial changes lie on the texture alteration,
such as the gradually developed eye bag, senile plaques and wrinkle.
Different from traditional GANs which only model the real data
distribution of each individual age, we focus on the higher-order
cross-age correlations, which will make the face aging results more
appealing. To model the above-mentioned transition patterns, we
propose a Contextual Generative Adversarial Nets (C-GANs). Figure 1 illustrates C-GANs briefly. For an input face, C-GANs can
generate faces for any target age group. To ensure the generated images real, C-GANs uses two discriminative networks to model the
distribution of each individual age group as well as the transition
patterns of two adjacent groups respectively.
More specifically, C-GANs consists of three networks, which
is shown in Figure 2. The conditional transformation network
transforms the input face to the desired age; the age discriminative network assists generating images indistinguishable with the
real ones; the transition pattern discriminative network regularize the generated images satisfying the cross-age aging rules. The
proposed C-GANs can be trained end-to-end and very easy to reproduce. In order to facilitate the presentation, we only mention
face aging/progression in this paper. Actually, C-GANs can also
achieve face regression without any further modification.
The contributions of this paper are summarized as follows.
(1) We design an effective and efficient contextual GANs based
face aging system whose aging results are significantly
better than existing face aging methods. The source code
of our method will be released to the academic area for
further research.
(2) We introduce a novel transition pattern discriminative network in the C-GANs to regularize the synthesized faces
satisfying the cross-age face aging rules.
(3) The conditional face transformation network in C-GANs
is different with existing GANs generators in that it is
much deeper, with several specially designed skip layers
to preserve both the high-level semantics and low-level
information. It makes the generated images more natural
and real.
2 RELATED WORKS
2.1 Face Aging
Traditional face aging models can be roughly divided into physical model approaches and prototype approaches. Physical model
approaches explicitly or implicitly models shape and texture parameters for each age group. For example, Suo et al. [36] present
a hierarchical And-Or graph based dynamic model for face aging.
Other model-based age progression approaches include active appearance model [16], support vector regression [25] and implicit
function [4]. The prototype approache [15] aim at constructing a
k3n64s1
conv
conv
conv
k1n64s1 k3n64s2
concate
conv
conv
k1n64s1 k3n64s1 k3n64s1
conv
concate
conv
conv
conv
label
image
k3n3s1 k3n64s1 k3n64s1
MM ’17, October 23–27, 2017, Mountain View, CA, USA
concate
Face Aging with Contextual Generative Adversarial Nets
Skip connection
conv
Legend:
leakyRelu
co
onv
decconv
co
onv
k3n3s1 k3n64s2 k1n64s1
con
ncate
co
onv
co
onv
co
onv
co
onv
k3n3s1 k3n64s1 k1n3s1
con
ncate
k1n3s1
con
ncate
co
onv
taanh
k1n3s1
b
bn
Figure 3: Architecture of the conditional transformation network i-th corresponding kernel size (k), number of feature maps
(n) and stride (s) indicated for each convolutional layer.
relighted average face as prototypes for different age groups, and
transferring the texture difference between the prototypes to the
test image. However, the limitation of this model is they are based
on general rules, they totally discard the personalized information.
Recently, Shu et al. propose a coupled dictionary learning (CDL)
model [35]. It encodes the aging patterns by the dictionary bases.
Every two neighboring dictionaries are learned jointly. However,
this method still has ghost artifacts as the reconstruction residual
does not evolve over time. Wang et al. [40] introduce a recurrent
face aging (RFA) framework based on a recurrent neural network.
They employ a two-layer gated recurrent unit as the basic recurrent module whose bottom layer encodes a young face to a latent
representation and the top layer decodes the representation to a
corresponding older face. Generally, these techniques requires
sufficient age sequences as the training data, which limits these
methods’s practicality.
2.2
Generative Adversarial Networks
Recently, GANs [10] has achieved great success in many image
synthesis applications, including super resolution [17], image-toimage translation by pix2pix [13] and CycleGAN [48], in-paining
[24], visual manipulation on the images [47].
Antipo et al. [2] propose the GAN-based method for automatic
face aging. They particularly emphasize preserving the original person’s identity by introducing a “Identity-Preserving” optimization
of GAN’s latent vectors. Zhang [46] propose a conditional adversarial autoencoder (CAAE) that learns a face manifold, traversing on
which smooth age progression and regression can be realized simultaneously. Two adversarial networks are imposed on the encoder
and generator, respectively, forcing to generate more photo-realistic
faces. Li et al. [18] presents a Deep convolutional network model
for Identity-Aware Transfer (DIAT) of facial attributes. However,
these GANs based methods independently model the distribution of
each age group, without capturing the cross-age transition patterns.
3
APPROACH
The architecture of the proposed C-GANs is shown in Figure 2.
The input image x is first aligned and parsed (Section 3.1). Then
x is paired with an arbitrary age label ỹ to feed into the conditional transformation network G (Section 3.2). The synthesized
face G (x, ỹ) is judged by the age discriminative network D a to be
real/fake (Section 3.3). Moreover, the age pair composed of a real
image and its fake counterpart, is fed into the transition pattern discriminative network D t which predicts whether it is from the real
image pair distribution (Section 3.4). Finally, the objective function
and the training strategy is introduced (Section 3.5).
3.1
Image Preprocessing
The input image x is aligned via the face alignment techniques
which locates 68 points on the faces. The landmark is used to align
the faces. Then we use Deeplab v2 [6] to parse the human face into
facial and non-facial regions. The non-facial region containing the
background, the hair and clothes, are masked with gray color to
facilitate the GANs training.
3.2
Conditional Transformation Network
Given the input face x and the desired age ỹ, the Conditional Transformation Network generates the synthesized face xỹ = G (x, ỹ).
The architecture of the generator is shown in Figure 3. The input
and output face images are 128 × 128 RGB images. The output is in
range [−1, 1] through the hyperbolic tangent function. Normalizing
the input may make the training process converge faster. The conditions of C-GANs are 7-dim one-hot age vectors, and reshaped as
7-channel tensor with the same spatial dimensions with the input
face. The input faces and the labels are concatenated and fed into
MM ’17, October 23–27, 2017, Mountain View, CA, USA
Si Liu, Yao Sun, Defa Zhu, Renda Bao, Wei Wang, Xiangbo Shu, and Shuicheng Yan
con
nv
lab
bel
k4n64s2
sigmoid
conv
conv
conv
conv
conv
concate
conv
image
k4n128s2 k4n256s2 k4n512s2 k4n1024s2 k4n2s1
10‐>20
50‐>60
Figure 4: The structure of the age discriminative network.
Figure 5: Illustration of the transition patterns of differnt
age ranges.
the further processing. To make fair concatenation, the elements
of label is also confined to [−1, 1], where −1 corresponds to 0.
The conditional transformation network mainly contains several
residual blocks [11] with several skip layers. Following DCGAN
[27], the convolution of stride 2 is employed instead of pooling.
The first three residual blocks donwsize the feature maps to half
the resolution of the input image. The “deconv” layer upsamples
the feature map to the original resolution. Such layers have been
used previously [34, 44]. Note that to perform the refinement, we
adopt several skip layer. In this way we preserve both the highlevel information passed from coarser feature maps and fine local
information provided in lower layer feature maps.
the gradually developped wrinkles. Despite the big appeareance
changes between adjecent age groups, the aging face should keep
the identity of the input face. In other word, the learned transition
pattern should also has the nice property of identity preserving.
To this end, C-GANs contains a transition pattern discriminative
network as shown in Figure 2. The network D t xy , xy+1 , y, y + 1
indicates the transition pattern between xy aged at y to the image
xy+1 at the next age group y + 1. For notation simplity, the D t is
denoted as D t xy , xy+1 , y . The networks distinguishes the real
joint distribiton xy , xy+1 , y ∼ pdat a (xy , xy+1 , y) from the fake one.
The optimization objetive is:
3.3
Age Discriminative Network
min max E θG , θ D t =
The structure of the age discriminative network D a (x, y) is shown
in Figure 4, similar to the conditional GAN [22]. θ D a is the network
parameters. More formally, the training can be expressed as an
optimization of the function E(θG , θ D a ), where θG and θ D a are
parameters of G and D a , respectively:
min max E θG , θ D a = E xy ,y∼pd at a (xy ,y) log D a xy , y
G D
a
+E x ∼px , ỹ∼py [log (1 − D a (G (x, ỹ) , ỹ))].
(1)
Note that the age label is resized to a tensor similar with the
conditional transformation network. The image and the labels
go through one convolution layers individually and concatenated
to feed to D a to make it discriminative on both age and human
face. During training, the
positive samples are the real faces and
their corresponding age xy , y , while nagetive samples are {xỹ , ỹ},
where xỹ is the generated/fake face and ỹ is the corrsponding
label used during the generation process. Note that we specifically
randomly sample the label ỹ for the fake images to enhance the
generalization ability of the C-GANs model.
3.4
Transition Pattern Discriminative Network
For better face aging results, we specifically model the cross-age
transition patterns, which is defined as the facial feature correlations betweeen different age groups. In this paper, we only consider
adjacent age groups for simplicity. The long-range transition pattern can be represented as the combination of a series of transitions
patterns between adjacent age group. The transion pattern is ageaware. As shown in Figure 4, when a person grows from the age of
10 to 20, the facial shape alters as the skull grows. However, when
one grows from the age of 50 to 60, the most obvious change is
G Dt
= E xy ,xy+1,y∼pd at a (xy ,xy+1,y) log D t xy , xy+1 , y
+ 12 E xy ,y∼pd at a (xy ,y) log 1 − D t xy , G xy , y + 1 , y
+ 12 E xy ,y∼pd at a (xy ,y) log 1 − D t G xy , y − 1 , xy , y − 1 .
(2)
The second term of Equation
2
guides
the
transformation
network
to
generate the fake pair xy , G xy , y + 1 to obey the real transition
pattern distribution. Similarily, the third term are imposed on
the
network to generate real convincing fake pair
tranformaion
G xy , y − 1 , xy .
3.5
Objective Function & Training Strategy
Comprehensively considering the losses of the conditional transformation network G, the age discriminative network D a as well
as the transition pattern discriminative network D t , the overall
objective function is:
min max max E θG , θ D a , θ D t =Ea + Et + λTV
G Da Dt
= E xy ,y∼pd at a (xy ,y) log D a xy , y
+E x ∼px , ỹ∼py [log (1 − D a (G (x, ỹ) , ỹ))]
+E xy ,xy+1,y∼pd at a (xy ,xy+1,y) log D t xy , xy+1 , y
+ 12 E xy ,y∼pd at a (xy ,y) log 1 − D t xy , G xy , y + 1 , y
+ 12 E xy ,y∼pd at a (xy ,y) log 1 − D t G xy , y − 1 , xy , y − 1
+λ TV G xy , y − 1 +TV G xy , y + 1 + TV (G (x, ỹ)) .
(3)
where TV(·) denotes the total variation which is effective in removing the ghosting artifacts. The coefficient λ balances the smoothness
and high resolution. During training, D a , D t and G are alternatively optimized. More specifically, for one iteration, the D a and G
Face Aging with Contextual Generative Adversarial Nets
Original image
0-10
11-18
MM ’17, October 23–27, 2017, Mountain View, CA, USA
19-29
30-39
40-49
50-59
60+
Figure 6: The input face and the generated faces for 7 age groups.
are updated. For the next iteration, the parameters of D t and G are
refined.
Testing Phase The test image x t and the desized label y is fed
into the conditional transformation network in Section 3.2, and the
output face xyt is obtained.
In the future, we would like to explore the face aging in videos
by making use of the state-of-the-art video processing [20, 43],
face parsing [41], object tracking [21, 45] and video deblur [29, 30]
techniques.
4 EXPERIMENTS
4.1 Dataset Collection
Our C-GANs requires both sequential and non-sequential data. For
sequential data, we select 575, 649, 1, 962, 695 and 166 images from
CACD [5], FGNET [1], LFW [12], Morph [31] and SUP [40] dataset,
respectively. The whole dataset contains 4, 047 images with equal
female/male and age distribution. We generate 3, 992 positive pairs
from the sequence data for training. Note that we only use 575
images from CACD for model training, and reserves 4, 689 images
for the face verification experiment in Section 4.6.
For non-sequential data, we use the public IMDB-Wiki dataset
[32]. We manually delete the images with heavy occlusion or low
resolution. As the dataset contains very few senior people, we also
MM ’17, October 23–27, 2017, Mountain View, CA, USA
Si Liu, Yao Sun, Defa Zhu, Renda Bao, Wei Wang, Xiangbo Shu, and Shuicheng Yan
20+ ‐> 60+
FT demo
CDL
RFA
CAAE
Ours
10+ ‐> 20+
FT demo
CDL
RFA
CAAE
Ours
30+ ‐> 60+
FT demo
CDL
RFA
CAAE
Ours
10+ ‐> 40+
FT demo
CDL
RFA
CAAE
Ours
20+ ‐> 40+
FT demo
CDL
RFA
CAAE
Ours
4+ ‐> 10+
FT demo
CDL
RFA
CAAE
Ours
8+ ‐> 30+
FT demo
CDL
RFA
CAAE
Ours
20+ ‐> 40+
FT demo
CDL
CDM
CAAE
Ours
40+ ‐> 60+
FT demo
d
CDL
RFA
CAAE
Ours
20+ ‐> 30+
FT demo
CDL
CDM
CAAE
Ours
6+ ‐> 50+
FT demo
CDL
IAAP
CAAE
Ours
20+ ‐> 60+
FT demo
CDL
CDM
CAAE
Ours
Figure 7: Comparison with the State-of-the-arts.
selected some images from the webface dataset [19], adiencedb [8]
and CACD [5]. Like [2], we divide the age into six age categories:
0 − 10, 11 − 18, 19 − 29, 30 − 39, 40 − 49, 50 − 59 and 60+ years. The
non-sequential data consists of 15, 030 face images with a uniform
distribution on both gender and age.
4.2
Implementation Details
The C-GANs net is trained via torch based on the public available
codes DCGAN [27]. Similarly, the learning rate is set 0.0002 and
beta1 is 0.5 with a mini-batch size 28. Both faces and ages are
fed to the network. The generator and discriminator networks
are optimized alternatively by the Adam method. Generally, the
network needs 100 epochs to generate favorable images, which
may takes 12 hour by using NVIDIA Titan X. It takes about 0.028
sec. for testing one image.
4.3
Face Aging Results
We show the face aging results in Figure 6. We can see that C-GANs
can generate quite appealing results. We can draw the following
observations. First, the faces are quite real and natural. Second,
the generated faces can change gradually when getting older. For
example, in the output faces of the last 5 age groups (in red boxes)
of the first and second rows, the beards appear and become white.
Third, C-GANs can synthesize images with large age gaps. For
example, the input face of the third row is quite young, but the
synthesized faces in the 60+ group (in yellow box) is still quite real.
For another example, the face in the fourth row is a senior lady.
We can produce very child-looking young face (in yellow box) for
the 0 − 10 age group. Fourth, C-GANs can produce very detailed
texture changes. For example, in the fifth, sixth and seventh rows,
the synthesized faces in the red boxes contain frighteningly real
convincing enough crow’s feet, canthus wrinkles and eye bags.
Fifth, the shapes of face and facial features also change during the
aging/progression. For example, in the last two rows, when the
seniors are transformed to babies, their face become smaller, and
their eyes and ocular distance become larger.
4.4
Comparison with the State-of-the-arts
Baseline methods: Some prior works on age progression have
posted their best face aging results with inputs of different ages,
including [36, 39]. We mainly compare with 9 baselines, including
FT demo: an online fun demo Face Transformer, IAAP: state-of-theart illumination-aware age progression [15], RFA: recurrent face
aging [40], CDL: coupled dictionary learning [35], acGAN: face
Face Aging with Contextual Generative Adversarial Nets
MM ’17, October 23–27, 2017, Mountain View, CA, USA
40+
40->60
60+
10+
10->9
9
30+
30->20
20+
40+
40->30
30+
50+
50->30
30+
30+
30->50
50+
50+
50->30
30+
5
5->10
10+
7
7->10
10+
40+
40->50
50+
60+
60->50
50+
50+
50->40
40+
40+
40->30
30+
30+
30->60
60+
4
4->10
10+
60+
60->30
30+
40-49
9 50-59 60+
11-18 19-29 30-39
0.5
0.8
1
0.7
0.7
0.6
0.9
0.6
0.6
1
0.6
0.5
0.5
0.9
0.7
0.7
0.9
0.8
0.9
1
09
0.9
1-10
aging with conditional generative adversarial networks [2], CAAE:
conditional adversarial autoencoder [46], CDM: Compositional and
Dynamic Model [36] and [33, 39]. There are 246 aging results with
72 inputs in total. Our face aging for each input is implemented
to generate the aging results with the same ages (ranges) of the
posted results.
Qualitative Evaluation: Figure 7 plots the results of the comparison. Compared with other methods, the aged face generated by
our method has more realistic and noticeable transformations in
appearance. For instance, the shape of facial features has changed
obviously when children grow up(row 1 col 2, row 3 col 2). The
aged face in the row 2 col 1 gets more wrinkles, and her eyes become smaller during the process. We can also observe that eye bags
and more wrinkles appear in row 5 col 1. Meanwhile, our method
preserve the identity very well, which can be observed in most
images.
Quantitative Evaluation: To quantitatively evaluate the performance of the proposed method, we designed a user study from
43 volunteers. Each volunteer is shown with three generated images each time. The candidate images are generated from given age
groups and supposed to have the target ages. Among these images,
there is the one generated by C-GANs and two other results generated by FT demo, CAAE or other prior methods. Every volunteer
is asked to choose one of the following three options. We added
1 point if our result was chosen as the best one, 0.5 when “all are
likely” was chosen, and 0 if one result from prior work was chosen
as the best. The score was normalized by number of responses per
cell and shown in Figure 9. The x-axis is the input age group and
the y-axis is the target age group.
From Figure 9 we can see that the proposed method outperforms
prior work almost all the time. Particularly, our approach performs
Targe
et Age
Figure 8: Comparison between our results and ground truth.
1-10
1
1
0.9
0.6
0.9
0.8
0.5
1
0.5
0.6
0.8
0.9
0.7
0.6
0.8
0.6
06
0.6
08
0.8
07
0.7
08
0.8
09
0.9
11-18 19-29
30-39
40-49 50-59 60+
Input Age
Figure 9: Comprehensive comparison to prior works.
very good when the input and target age groups are contiguous. We
believe this manly credits to the newly proposed Transition Pattern
Discriminative network. We also notice that our approach is a bit
poor when generating faces of 60+ from children. We think this is
because of the significant changes of appearances from children to
olds. We will try to improve the performance in future works.
MM ’17, October 23–27, 2017, Mountain View, CA, USA
Si Liu, Yao Sun, Defa Zhu, Renda Bao, Wei Wang, Xiangbo Shu, and Shuicheng Yan
10+
20+
20+
50+
60+
50+
60+
40+
5
8
20+
40+
Figure 11: Face verification results.
Figure 10: Movie stars. In each group, a single image on the
far left is age progressed to different ages in the first row,
and the real images of that age are shown in the second row.
4.5
Comparison with ground truth
Here we qualitatively compare the generated faces with the ground
truth. The qualitative comparisons of both female and male are
shown in Figure 8, which shows appealing similarity. In each triplet,
the first and third images are the ground truths with Age Group 1
and Age Group 2, while the second image is our aging result.
In this experiment, we first crop the face from images of Age
Group 1. Then we proceed these images by C-GANs and get the
aging faces with the same ages as Age Group 2. At last, we run
graph cuts to find an optimal seam followed by poisson blending
to blend the aged face into the real head photo [7].
In Figure 8, we can observe that the generated aging faces almost
have no difference from real ones. This indicates the C-GANs could
correctly synthesis the real age progression.
To better demonstrate the age progressing capability of C-GANs,
we collect some images of two movie starts, including Brad Pitt and
Shirley Temple. For Temple, we find several real images through
her life. We cut out some frames from movies and take them as the
older appearance of Pitt. The results are shown in Figure 10. The
age groups from 20 − 50 are omitted due to the limitation of data.
Note that our method can successfully simulate the transformation
characteristics of all ages of the stars, especially for the ages of
children and olds.
than 20 years. The set of these 4, 000 pairs is called “Original Pairs”.
In each original pair, we proceed the younger face to the aging
face with the same age of the older face by C-GANs, and assume
the aging face has the same age as well. We replace the younger
face in original pairs by the newly generated aging face, and then
construct 4, 000 new pairs, called “Our Synthetic Pairs”. To evaluate
the performance of our C-GANs, we also generated the “CAAE
Synthetic Pairs” by the state-of-the-art age progression method
[46]. The state-of-the-art center Loss based face verification [42] is
used for testing on the above three sets of pairs.
The FAR-FRR (false acceptance rate-false rejection rate) curves
are illustrated in Figure 11. The EER (the equal error rates) on
C-GANs Synthetic Pairs, CAAE Synthetic Pairs, and Original Pairs
are 8.72%, 11.05%, and 17.41% respectively. We can see that the face
verification on C-GANs Synthetic pairs achieves better ERR than
on other pairs. This implies the aging faces generated by C-GANs
can effectively alleviate the face verification errors cost by age gaps.
5
CONCLUSION AND FUTURE WORKS
In this paper, we propose a contextual generative adversarial nets
to tackle the face aging problem. Different from existing generative
adversarial nets based methods, we explicitly model the transition
patterns between adjacent age groups during the training procedure. From baby to teenagers period, the transition patterns is
shown in the way that the face becomes bigger, while from the
ages of 30 to the age of 50, the transition patterns include the grad4.6 Cross Age Face Verification
ually developed wrinkle. To this end, the C-GANs consists of two
The proposed age progression method can also improve the perdiscriminative networks, i.e., an age discriminative network and a
formance of cross-age face verification significantly. We collected
transition pattern discriminative network. They are collaboratively
2, 000 intra-person pairs and 2, 000 inter-person pairs with cross
contribute to the appealing results.
ages on the FGNET database, using 2, 044/2, 645 images of males/females
Currently, our model is based on DCGAN [27]. In future, we
respectively. In both sets of pairs, the number of male pairs and
plan to employ other GANs to improve the performance, such as
female pairs are equal, and the age span of the pairs are all more
Wasserstein GAN [3], LS GAN [26], EB GANs [14].
Face Aging with Contextual Generative Adversarial Nets
6
ACKNOWLEDGMENT
This work was supported by National Natural Science Foundation
of China (No.61572493, Grant U1536203) and Natural Science Foundation of Jiangsu Province (Grant No. BK20170856)
REFERENCES
[1] 2000. Face and gesture recognition. In Network: FG-NET Aging Database. (2000).
[2] Grigory Antipov, Moez Baccouche, and Jean Luc Dugelay. 2017. Face Aging
With Conditional Generative Adversarial Networks. arXiv:1702.01983 (2017).
[3] Martin Arjovsky, Soumith Chintala, and Lon Bottou. 2017. Wasserstein GAN.
(2017).
[4] Alexandre Cruz Berg, Francisco Jos Perales Lopez, and Manuel Gonzlez. 2006. A
Facial Aging Simulation Method Using flaccidity deformation criteria. IV (2006).
[5] Bor Chun Chen, Chu Song Chen, and Winston H. Hsu. 2014. Cross-Age Reference
Coding for Age-Invariant Face Recognition and Retrieval.
[6] Liang Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and
Alan L. Yuille. 2016. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. In arXiv:1606.00915.
[7] Neeraj Kumar Samreen Dhillon Peter Belhumeur Shree K. Nayar Dmitri, Bitouk.
2008. In ACM Trans. on Graphics (also Proc. of ACM SIGGRAPH).
[8] E Eidinger, R Enbar, and T Hassner. 2014. Age and Gender Estimation of Unfiltered Faces. TIFS (2014).
[9] Y. Fu, G. Guo, and T. S. Huang. 2010. Age synthesis and estimation via faces: a
survey. TPAMI (2010).
[10] Ian J. Goodfellow, Jean Pougetabadie, Mehdi Mirza, Bing Xu, David Wardefarley,
Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative Adversarial
Networks. NIPS (2014).
[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. Deep Residual
Learning for Image Recognition. In CVPR.
[12] Gary B. Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. 2008.
Labeled Faces in the Wild: A Database forStudying Face Recognition in Unconstrained Environments. (2008).
[13] Phillip Isola, Jun Yan Zhu, Tinghui Zhou, and Alexei A. Efros. 2016. Image-toImage Translation with Conditional Adversarial Networks. In arXiv:1611.07004.
[14] Yann LeCun Junbo Zhao, Michael Mathieu. 2017. Energy-based Generative
Adversarial Network. arXiv:1609.03126 (2017).
[15] Ira Kemelmacher-Shlizerman, Supasorn Suwajanakorn, and Steven M Seitz. 2014.
Illumination-Aware Age Progression. In CVPR.
[16] A Lanitis, C. J Taylor, and T. F Cootes. 2002. Toward automatic simulation of
aging effects on face images. TPAMI (2002).
[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, and
Zehan Wang. 2016. Photo-Realistic Single Image Super-Resolution Using a
Generative Adversarial Network. In arXiv:1609.04802.
[18] Mu Li, Wangmeng Zuo, and David Zhang. 2016. Deep Identity-aware Transfer
of Facial Attributes. In arXiv:1610.05586.
[19] L. Liu, C. Xiong, H. Zhang, and Z. Niu. 2016. Deep Aging Face Verification With
Large Gaps. TMM (2016).
[20] Si Liu, Changhu Wang, Ruihe Qian, Han Yu, Renda Bao, and Yao Sun. 2017.
Surveillance Video Parsing with Single Frame Supervision. CVPR (2017).
[21] Si Liu, Tianzhu Zhang, Xiaochun Cao, and Changsheng Xu. 2016. Structural
correlation filter for robust visual tracking. In CVPR.
[22] Mehdi Mirza and Simon Osindero. 2014. Conditional Generative Adversarial
Nets. Computer Science (2014).
MM ’17, October 23–27, 2017, Mountain View, CA, USA
[23] U Park, Y. Tong, and A. K. Jain. 2010. Age-invariant face recognition. TPAMI
(2010).
[24] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A.
Efros. 2016. Context Encoders: Feature Learning by Inpainting. CVPR (2016).
[25] E Patterson, A Sethuram, M Albert, and K Ricanek. 2007. Aspects of Age Variation
in Facial Morphology Affecting Biometrics. In ICB.
[26] Guo-Jun Qi. 2017. Loss-Sensitive Generative Adversarial Networks on Lipschitz
Densities. arXiv:1701.06264 (2017).
[27] Alec Radford, Luke Metz, and Soumith Chintala. 2015. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks.
Computer Science (2015).
[28] Lejian Ren, Si Liu, Yao Sun, Jian Dong, Liu Luoqi, and Yan Shuicheng. 2017. Time
Traveler: a real-time face aging system. In ACM MM.
[29] Wenqi Ren, Xiaochun Cao, Jinshan Pan, Xiaojie Guo, Wangmeng Zuo, and
Ming-Hsuan Yang. 2016. Image deblurring via enhanced low-rank prior. IEEE
Transactions on Image Processing 25, 7 (2016), 3426–3437.
[30] Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, and Ming-Hsuan
Yang. 2016. Single image dehazing via multi-scale convolutional neural networks.
In European Conference on Computer Vision. Springer, 154–169.
[31] Karl Ricanek and Tamirat Tesafaye. 2006. MORPH: a longitudinal image database
of normal adult age-progression. FG (2006).
[32] Rasmus Rothe, Radu Timofte, and Luc Van Gool. 2015. DEX: Deep EXpectation
of Apparent Age from a Single Image. In CVPR workshop.
[33] Amrutha Sethuram, Karl Ricanek, and Eric Patterson. 2010. A hierarchical
approach to facial aging. (2010).
[34] Evan Shelhamer, Jonathon Long, and Trevor Darrell. 2017. Fully Convolutional
Networks for Semantic Segmentation. TPAMI (2017).
[35] Xiangbo Shu, Jinhui Tang, Hanjiang Lai, Luoqi Liu, and Shuicheng Yan. 2015.
Personalized Age Progression with Aging Dictionary. In ICCV.
[36] Jinli Suo, Song Chun Zhu, Shiguang Shan, and Xilin Chen. 2010. A Compositional
and Dynamic Model for Face Aging. TPAMI (2010).
[37] Yusuke Tazoe, Hiroaki Gohara, Akinobu Maejima, and Shigeo Morishima. 2012.
Facial aging simulator considering geometry and patch-tiled texture. In ACM
SIGGRAPH.
[38] Bernard Tiddeman, Michael Burt, and David Perrett. 2001. Prototyping and
transforming facial textures for perception research. CGA (2001).
[39] Junyan Wang, Yan Shang, Guangda Su, and Xinggang Lin. 2006. Age simulation
for face recognition. In ICPR.
[40] Wei Wang, Zhen Cui, Yan Yan, Jiashi Feng, Shuicheng Yan, Xiangbo Shu, and
Nicu Sebe. 2016. Recurrent Face Aging. In CVPR.
[41] Zhen Wei, Yao Sun, Jinqiao Wang, Hanjiang Lai, and Si Liu. 2017. Learning
Adaptive Receptive Fields for Deep Image Parsing Network. In CVPR.
[42] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. 2016. A Discriminative
Feature Learning Approach for Deep Face Recognition. In ECCV.
[43] Han Yu, Guanghui Ren, Ruihe Qian, Yao Sun, Changhu Wang, hanqing Lu, and
Si Liu. 2017. RSVP: A Real-Time Surveillance Video Parsing System with Single
Frame Supervision. In ACM MM.
[44] Matthew D. Zeiler and Rob Fergus. 2014. Visualizing and Understanding Convolutional Networks. In ECCV.
[45] Tianzhu Zhang, Si Liu, Changsheng Xu, Shuicheng Yan, Bernard Ghanem, Narendra Ahuja, and Ming-Hsuan Yang. 2015. Structural sparse tracking. In CVPR.
[46] Zhifei Zhang, Yang Song, and Hairong Qi. 2017. Age Progression/Regression by
Conditional Adversarial Autoencoder. In arXiv:1702.08423.
[47] Jun Yan Zhu, Philipp Krhenbhl, Eli Shechtman, and Alexei A. Efros. 2016. Generative Visual Manipulation on the Natural Image Manifold. In ECCV.
[48] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. 2017. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks.
arXiv:1703.10593 (2017).
| 1cs.CV
|
arXiv:1403.4862v1 [math.AC] 19 Mar 2014
Green’s Hyperplane Restriction Theorem:
an extension to modules
Ornella Greco
Abstract
In this paper, we prove a generalization of Green’s Hyperplane Restriction Theorem to the case of modules over the polynomial ring,
providing in particular an upper bound for the Hilbert function of the
general linear restriction of a module M in a degree d by the corresponding Hilbert function of a lexicographic module.
1
Introduction
The extremal properties of the Hilbert function constitute a relevant topic
in commutative algebra and, within this topic, an important role is played
by lexicographic ideals and binomial representation of integers.
Many results about extremal behavior of Hilbert functions have been proved.
In particular, Macaulay’s Theorem (see [3, 10, 13]), which characterizes the
possible Hilbert functions of homogeneous k-algebras; or, the analogous result for exterior algebra, proved by Kruskal and Katona (see [9, 12]), which
also provides a characterization of f -vectors of simplicial complexes.
More recently, in [8], Hulett proved an extension of Macaulay’s Theorem to
modules over the symmetric algebra.
Another classical, but slightly recent, result in this topic is the Hyperplane
Restriction Theorem (HRT), given by Green in [7], which gives a bound for
the codimension of the generic linear restriction of a vector space generated
in a certain degree by the codimension of a lex-segment space with same
degree and dimension. This result was also used by Green to give another,
less technical, proof of Macaulay’s Theorem.
It has been also been applied by several authors (see for instance [1, 2, 14])
to get some results about level and Gorenstein algebras, with focus on the
1
weak Lefschetz property.
In this paper, we will focus on Green’s theorem, trying to follow Hulett’s
path in generalizing these results to modules over the polynomial ring.
One generalization in this direction has already been done by Gasharov ([5]),
but we will give an extension of Green’s theorem that provides a bound which
is actually achieved by lexicographic modules.
We will first give the tools to state Green’s Hyperplane Restriction Theorem: namely we introduce the binomial expansion of a positive integer, also
called Macaulay representation, then, the concept of generic linear forms.
Afterwards, we will state Green’s theorem.
In the next section, we will extend some of these definitions to the submodule
of a finitely generated free module, in particular we will define a monomial
order induced by the deglex ordering, and will provide the concept of lexicographic module, which is a particular class of monomial modules.
We first try to extend Green’s theorem in the simplest case of a rank 2 module, by proving an inequality of Macaulay representations. Later we prove
by induction a new inequality which imply the main theorem, namely the
extension of the Hyperplane Restriction Theorem to the case of modules over
the polynomial ring.
As a consequence of our theorem, we will derive a result which constitute
another version of the theorem ”after scaling”, which means that we divide
by the respective Hilbert function of the polynomial ring.
Finally, we apply our main theorem to level algebras, by giving some
conditions in which the bound given by the extension to modules of the
Green’s HRT is given than the one obtained trough the Green’s HRT.
2
Green’s Theorem
Let S = k[x1 , . . . , xn ], where k is an infinite field, and let S be standard
graded. Let us fix the deglex monomial ordering on S with x1 > x2 > · · · >
xn , namely
xa >deglex xb ⇔ |a| > |b| or |a| = |b| and a >lex b,
(see [11]). Notice that in a homogeneous component Sd the monomial orderings, deglex and lex, coincide.
2
Definition 2.1. Let a, d ∈ N, then the d-th Macaulay representation of a
(also called the binomial representation of a in base d) is the unique way to
write
aδ
ad−1
ad
,
+···+
+
a=
δ
d−1
d
where ad > ad−1 > · · · > aδ ≥ δ and δ = min{i | ai ≥ i}.
Let us set dc = 0 whenever c < d, then, given the d-th Macaulay representation of a, let us define the integer
aδ − 1
ad−1 − 1
ad − 1
.
+···+
+
ahdi =
δ
d−1
d
Sometimes it will
convenient
to use
the extended d-th Macaulay represen bead−1
a1
ad
tation, i.e. d + d−1 + · · · + 1 , where ad > ad−1 > · · · > a1 ≥ 0.
Moreover, since ab = 0 if a < b, we will set the numerators of such binomial
coefficients in the Macaulay representation equal to zero.
d−1
d−1
+· · ·+ b11 ,
+· · ·+ a11 and b = bdd + bd−1
Observation 2.2. If a = add + ad−1
a ≥ b if and only if (ad , . . . , a1 ) ≥lex (bd , . . . , b1 ).
Definition 2.3. Let d ∈ N, a lex-segment is a set constituted by all monomials in Sd lexicographically larger or equal than f , for some f ∈ Mon(Sd ).
A k-vector subspace V of Sd is a lex-segment space if Mon(Sd )∩V is a k-basis
of V and a lex-segment.
Definition 2.4. A graded monomial ideal I in S is said to be lexicographic
if, for all degrees d, the homogeneous component Id is a lex-segment space.
Definition 2.5. We say that a property P holds for a generic linear form
ℓ if there is a non-empty Zariski open set U ⊆ S1 such that P holds for all
ℓ ∈ U.
If V ⊆ Sd a lex-segment space, we denote by Vℓ the image of V in S/(ℓ).
Proposition 2.6. [10, Proposition 5.5.23] Let k be an infinite field, d ∈ N,
V ⊆ Sd a lex-segment space. Then:
1. For a generic linear form ℓ ∈ S1 , there is a homogeneous linear change
of coordinates φ : S → S such that φ(V ) = V and φ(ℓ) = xn ;
2. For a generic linear form ℓ ∈ S1 , we have
codimk (Vℓ ) = codimk (Vxn ) = codimk (V )hdi .
3
So, after a general change of coordinates, we may assume that the generic
linear form is xn .
From now on, we will denote by H(M, d) the Hilbert function of the Smodule M in degree d, i.e. dimk (Md ). Moreover, if ℓ ∈ S1 is a linear form
and R = S/I, we denote by Rℓ the ring S/[I + (ℓ)].
Theorem 2.7 (Green’s Hyperplane Restriction Theorem). Let I be a homogeneous ideal in S, d ∈ N, then
H((S/I)ℓ , d) ≤ H(S/I, d)hdi ,
where ℓ is a generic linear form. Moreover, equality holds when Id is a lexsegment space.
3
Extension to modules
In the case of the restriction of an S-module M to a generic hyperplane ℓ
we expect not only an upper bound on the Hilbert function H(Mℓ , d), but
also that this upper bound is fully described by lexicographic modules, in
complete analogy with Green’s theorem.
In this section, we will be able to prove some numerical properties of the
Macaulay’s representations, that will lead to a version of Green’s HRT for
modules valid in any characteristic of the infinite field k.
Let F be a finitely generated graded free S-module, let us fix a homogeneous
basis {e1 , e2 , . . . , er } and let deg(ei ) = fi , where, without loss of generality,
we may assume that and f1 ≤ f2 ≤ · · · ≤ fr .
We now define the monomial modules and we induce a monomial ordering
on F in such a way that the concept of lexicographic module may be defined.
We also define another monomial ordering on F , a particular revlex, that
will let us use the tool of generic initial modules in our circumstances.
Definition 3.1. A monomial in F is an element of the form mei where
m ∈ Mon(S). A submodule M ⊆ F is monomial if it is generated by
monomials, in this case it can be written as I1 e1 ⊕ I2 e2 ⊕ · · · ⊕ Ir er , where Ii
is a monomial ideal for i = 1, 2, . . . , r.
4
One can extend a monomial ordering, defined in the polynomial ring S,
to a finitely generated free S-module in a really natural way, in particular
here we define the deglex ordering in F .
Definition 3.2. Given two monomials in F , mei and nej , we say that
mei >deglex nej if either i = j and m >deglex n in S or i < j. In particular, we have that e1 > e2 > · · · > er .
Again, as in the polynomial ring, when we restrict to a precise homogeneous component, the two induced monomial orderings, deglex and lex,
coincide.
Definition 3.3. A monomial graded submodule L is a lexicographic module
if, for every degree d, Ld is spanned by the largest, with respect to the
lexicographic order (or deglex), H(L, d) monomials.
Let us now recall the definition of reverse lexicographic order on F as
given in ([4, pg. 339]), some easy results about it.
Definition 3.4. The reverse lexicographic order on F is defined by choosing
an order on the basis of F , say e1 > · · · > er and by setting mei >revlex nej
iff either deg(mei ) > deg(nej ) or the degrees are the same and m >revlex n
in S or m = n and i < j.
Remark 3.5. Notice that the deglex is a POT (Position over Term) monomial
order, on the contrary the revlex is TOP (Term over Position) monomial
order.
Definition 3.6. The initial module of a submodule M, denoted by in(M),
is the submodule of F generated by the set {in(m)| m ∈ M}, i.e. by all the
leading terms (according to the lexicographic order or deglex or revlex) of
elements in M.
Proposition 3.7. [4, Proposition 15.12] Suppose that F is a free S-module
with basis {e1 , . . . , er } and reverse lexicographic order. Let M be a graded
submodule.
Then in(M + xn F ) = in(M) + xn F and (in(M) :F xn ) = in(M :F xn ).
Let us define the concept of generic initial module as done in ([15, Chapters 1 and 2]).
P
Let GL(n) be the general linear group. It acts on S by (aij )xj = ni=1 aij xi .
5
Let GL(F ) be the group of the S-module automorphisms of F . An element
φ in GL(F ) is a homogeneous automorphism and
P can be represented by a
matrix (tij ), with tij ∈ Sfi −fj , where φ(ei ) = rj=1 tij ej . We also have an
action of GL(n) on GL(F ) given by a · φ = aφa−1 , for all a ∈ GL(n) and
φ ∈ GL(F ). The two groups together act on F by mean of their semi-direct
product G = GL(n) ⋊ GL(F ).
Let us consider B(n) the subgroup of GL(n) of upper triangular invertible
matrices: from the action given above, a matrix in B(n) will send x1 to a
multiple of itself, and xn to the linear combination of all variables xi ’s with
coefficients the last column of tha matrix. Let B(F ) be the subgroup of
all automorphisms in GL(F ) represented by lower triangular matrices: they
send each el to an S-linear combination of e1 , . . . , el . Let B = B(n) ⋊ B(F ).
Pardue, in [15], proved the generalization to modules of the Galligo’s theorem, in a more general setting than ours, and in particular he proved the
following result:
Proposition 3.8. Let M ⊂ F be a graded module, and let < be a monomial
order on F , then there exists a Zariski open set U ⊆ G such that in< (φM) is
constant for every φ ∈ U. Moreover, in< (φM) is fixed by the action of group
B.
Definition 3.9. The monomial submodule in< (φM), φ ∈ U, is called the
generic initial module of M, and denoted by gin< (M).
The generic initial module has really nice properties in case we use, as
monomial order on F , the revlex order defined above.
Proposition 3.10. For a generic linear form ℓ and a graded submodule M
of F , we have that
gin(Mℓ ) = gin(M)xn .
The revlex gin(M) has the properties in 3.7, and moreover, since it is
B-fixed, we can work in generic coordinates, and choose the coordinates in
such a way that ℓ = xn .
If ℓ ∈ S1 is a linear form, and M a graded submodule of F , let us denote by (F/M)ℓ the restriction of the module F/M to ℓ, which is equal to
F/(M + ℓF ). Notice that Fℓ is a module over Sℓ and that Sℓ can be thought
as polynomial ring in n − 1 variables.
6
Before the statement of the main theorem of the section, we are going to
prove two propositions, regarding the behavior of the function ∗hdi : N → N:
in particular we need to provide an upper bound to the function ahd1 i + bhd2 i ,
where d1 ≥ d2 , generalizing [5, Lemma 4.4]; and then, we will extend the
latter inequality to the sum of r = rank(F ) integers.
These two proposition will be just the numerical translations of the extension of Green’s theorem, respectively in the rank(F ) = 2 case, and in the
rank(F ) = r case.
Lemma 3.11. [5, Lemma 4.4, Lemma 4.5] For any a, b, d ∈ N, the two
following inequalities hold: ahdi + bhdi ≤ (a + b)hdi ; ahd+1i ≤ ahdi .
Remark 3.12. If a, d ∈ N and a = add + · · · + aδδ (aδ ≥ δ) is the d-th
Macaulay representation of a, then
(a − 1)hdi = ahdi if and only if aδ = δ.
Proposition 3.13. Given a, b ∈ N, a ≤ N1 = n+dd11 −1 , b ≤ N2 =
and d1 ≥ d2 , d1 , d2 ∈ N, then
(
(a + b)hd2 i
if a + b ≤ N2 ,
ahd1 i + bhd2 i ≤
(a + b − N2 )hd1 i + (N2 )hd2 i if a + b ≥ N2 .
n+d2 −1
d2
Proof We will now prove the claim by induction on i = a + b.
The case i = 0 is trivial. Thus, let the claim be true for all a′ , b′ such that
0 < a′ + b′ <i.
Let a = add11 + · · · + add22 + · · · + a11 , b = bdd22 + · · · + b11 , where some
of the last top coefficients in the two Macaulay representations may be zero,
and let us denote these Macaulay representations as vectors a = (ad1 , . . . , a1 )
and b = (bd2 , . . . , b1 ).
Let us now distinguish between two cases.
First Case If the d1 -Macaulay representation of a ends after degree d2 ,
i.e. min{i| ai ≥ i} ≤ d2 , we write a = (ad1 , . . . , ad2 , . . . , a1 ) and
b = (bd2 , . . . , b1 ), where in both vectors the last entries, after degree
d2 , may be zero.
Then, we redefine a and b as follows, sorting them entry by entry:
a = (ad1 , . . . , ad2 +1 , min{ad2 , bd2 }, . . . , min{a1 , b1 }),
b = (max{ad2 , bd2 }, . . . , max{a1 , b1 }).
7
Observe that, in this process, we are decreasing a of the same amount
we increase b (we are not modifying a + b) and we are not changing
ahd1 i + bhd2 i , since the sum of the two binomial expansions will still
contain the same terms. Moreover, in this way, we obtain that the
Macaulay representation of a ends sooner than that of b.
The idea is to find a′ , b′ such that a′ + b′ = a + b and a ≤ a′ , b′ ≥ b and
such that the following are satisfied:
1. a′hd1 i + b′hd2 i ≥ ahd1 i + bhd2 i ;
2. if i < N2 , either a′ = 0 and b′ = i; if i ≥ N2 , either a′ = i − N2
and b′ = N2 ; or at least one between a′ and b′ has Macaulay
representation ending with a binomial coefficient of the form tt
(since, in this case, we can use Remark 3.12 and apply induction
hypothesis).
So, we can suppose that
a = (ad1 , . . . , ad2 , ad2 −1 , . . . , as , 0, . . . , 0)
b = (bd2 , bd2 −1 , . . . , bt , 0, . . . , 0),
where as > s and bt > t (otherwise we can use Remark 3.12 and apply
induction hypothesis), and s ≥ t.
We need to show how to find a′ and b′ , by decreasing a and increasing
b of the same amount in three different cases.
In the case in which s = t = 1, we have a1 > 1 and b1 > 1, we
can consider (ad1 , . . . , ad2 , . . . , a1 − 1) and (bd2 , . . . , b2 , b1 + 1), without
changing ahd1 i + bhd2 i . Applying this several times we will get to the
point in which a1 = 1, so we have reached the desired a′ and b′ .
The
is when
t = 1 and s > 1, here we have to write
second
Ps case
as −(j+1)
as
,
and
then consider (ad1 , . . . , ad2 , . . . , as+1 , as −
=
j=0
s−j
s
1, . . . , as −s) and (bd , . . . , b2 , b1 +1) (we are decreasing a and increasing
b by 1). After this step, we are again in the first case.
Notice that in these first two cases if b1 + 1 = b2 we have to consider
(bd2 , . . . , b2 + 1, 0) and then we sort again by taking the min{ai , bi } and
the max{ai , bi }.
Ps−t as −(j+1)
+
The last case is when s > t > 1: here we write ass = j=0
s−j
8
as −(s−t+1)
t−1
; so we consider:
(ad1 , . . . , ad2 , . . . , as+1 , as − 1, . . . , as − (s − t + 1) (position t),
as − (s − t + 1) (position t − 1), 0, . . . , 0)
(bd2 , bd2 −1 , . . . , bt , 0, . . . , 0),
and we resort by taking the min and the max in each modified degree.
Second Case When the Macaulay representation of a ends before the degree d2 , say in degree j, say k is the smallest degree for which bk ≥ k.
If d2 > 1, then let us consider the decomposition:
j
X
aj − (j − k)
aj − (j − i + 1)
aj
;
+
=
k
i
j
i=k+1
so we write
a = (ad1 , . . . , aj+1 , aj − 1, . . . , aj − (j − d2 + 1), . . . ,
aj − (j − k), aj − (j − k), 0, . . . , 0),
and we compare, entry by entry, with b moving the smaller entries in
the new vector a and the larger entries in b.
The only case in which this cannot work is in the trivial case b = (β),
a = (ad1 , . . . , α, α) (after having used the previous decomposition) and
α = β, in this situation we just consider (ad1 , . . . , α, α − 1) and (β + 1).
So, for all the possible cases we have found a way to subtract to a and add
to b the same number in such a way that the function f (x, y) = xhd1 i + yhd2 i
does not change: we will continue, recursively, to do this until we get the
extremal cases, or until one of the two numbers has Macaulay representation
ending with tt .
✷
Observation 3.14. Notice that the previous proposition is just the numerical
translation of the generalization of Green’s HRT to the case in which we
have F = Se1 ⊕ Se2 , with e1 >deglex e2 , deg(e1 ) = f1 , deg(e2) = f2 and
f1 ≤ f2 . In fact,
N, we set d1 = d − f1 , d2 = d − f2 , and
for any degree d ∈n+d
n+d1 −1
2 −1
= dimk (Sd2 ).
= dimk (Sd1 ), N2 =
N1 =
d2
d1
9
Remark 3.15. The previous proposition implies the first part of Lemma 3.11:
in fact, when we have f1 = f2 , and so N1 = N2 = N, we can repeat the same
argument of the proof of Proposition 3.13, with the only difference that we
continue decreasing a and increasing b of the same amount until a = 0 and
b = i, i.e. without distinguish between the cases in which i ≤ N and i ≥ N.
The following proposition just extends the previous inequality to the case
of more summands.
Proposition 3.16. Let d1 , . . . , dr ∈ N a non-increasing sequence,
Pr and set
n+di −1
N =
. Let ai ≤ Ni be an integer. Then, if, for some j, i=j+1 Ni ≤
di P
Pi r
r
a
≤
i=j Ni , the following inequality holds:
i=1 i
r
X
aihdi i ≤
i=1
r
X
ai −
r
X
Ni
i=j+1
i=1
!
+
hdj i
r
X
Nihdi i .
i=j+1
Proof By induction on r. The case r = 2 is given by the previous
result,
Pr
let us nowP
suppose that the inequality holds for s < r. Let i=j+1 Ni ≤
P
r
r
i=1 ai ≤
i=j Ni . We then distinguish between two cases.
In case aj + aj+1 ≤ Nj+1 , then aj hdj i + aj+1 hdj+1 i ≤ (aj + aj+1)hdj+1 i .
P
P
P
P
Since ri=j+1 Ni ≤ ri=1 ai ≤ ri=j Ni ≤ Nj−1 + ri=j+1 Ni , we can apply
the induction to have the following:
!
r
r
r
r
X
X
X
X
aihdi i + (aj + aj+1 )hdj+1 i ≤
ai −
Ni
+
aihdi i ≤
i=1
+
r
X
i=j+1
i=1
i=1,i6=j,j+1
Nihdi i ≤
r
X
i=1
ai −
r
X
i=j+1
Ni
!
+
hdj i
r
X
i=j+1
hdj−1 i
Nihdi i .
i=j+1
In the case in which aj + aj+1 ≥ Nj+1, then, by the base of induction,
aj hdj i + aj+1hdj+1 i ≤ (aj + aj+1 − Nj+1 )hdj i + Nj+1 hdj+1 i . We have that aj +
P
P
P
aj+1 − Nj+1 ≤ Nj and ri=j+2 Ni ≤ ri=1 ai − Nj+1 ≤ Nj + ri=j+2 Ni , thus
by induction:
!
r
r
r
r
X
X
X
X
aihdi i + (aj + aj+1 −Nj+1 )hdj i ≤
ai −
Ni
+
Nihdi i ,
i=1,i6=j,j+1
i=1
10
i=j+1
hdj i
i=j+2
By adding Nj+1hdj+1 i to both sides, we get
r
X
aihdi i ≤
i=1
r
X
i=1
ai −
r
X
Ni
i=j+1
!
+
hdj i
r
X
Nihdi i .
i=j+1
✷
The idea for the extension of the Green’s theorem to modules is to reduce to the monomial case: a monomial submodule is direct sum of monomial
components, and so we may just apply Theorem 2.7 to each of these components. Afterwards we need to bound the Hilbert function of the quotient by
the monomial submodule by the Hilbert function of the quotient by lexicographic module.
Let us now give a definition that will be used inside the statement of our
generalization of Green’s theorem, and then we are ready to prove our main
result.
Definition 3.17. Let M be a submodule in F , and m ∈ N. Set di = m − fi
({d1 , d2, . . . , dr } is a non-increasing sequence) and Ni = n+ddii −1 = dimk Sdi .
P
P
Then, if ri=j+1 Ni ≤ H(F/M, m) ≤ ri=j Ni , for some j, we define
H(F/M, m){m,r} =
H(F/M, m) −
r
X
i=j+1
Ni
!
+
hdj i
r
X
Nihdi i .
i=j+1
Theorem 3.18. Let F = Se1 ⊕ · · · ⊕ Ser where deg(ei ) = fi for all i. Let
M be a submodule in F , then
H((F/M)ℓ , m) ≤ H((F/L)ℓ , m)
where ℓ is generic linear form, m ∈ N, and L is a submodule that in degree
m is generated by a lex-segment of length H(M, m).
Moreover,
H((F/L)ℓ , m) = H(F/M, m){m,r} .
Proof Assume that M is a monomial submodule of F . Otherwise, let us take
the generic initial module of M, gin(M), with respect to the monomial order
revlex. We get a new submodule which has the same Hilbert function of
11
M and moreover is a monomial submodule of F . We have that gin(M)xn =
gin(Mℓ ). If we work in generic coordinates, we can choose ℓ = xn and so we
get that H((F/M)ℓ , d) = H((F/gin(M))ℓ , d). So we may substitute M by
gin(M) in order to get a monomial submodule. Thus, we can write M as
I1 e1 ⊕ · · · ⊕ Ir er , where each Ii is a monomial ideal, hence a homogeneous
ideal in R.
Let d ∈ N and ℓ be a generic linear form, then
(F/M)ℓ = F/(M + ℓF ) ∼
=
= F/[(I1 + (ℓ))e1 ⊕ · · · ⊕ (Ir + (ℓ))] ∼
(S/I1 )ℓ e1 ⊕ · · · ⊕ (S/Ir )ℓ er .
So we have that, by applying Green’s HRT to each (R/Ii )ℓ ,
H((F/M)ℓ , m) =
r
X
H((R/Ii )ℓ , m − fi ) ≤
i=1
r
X
H(R/Ii , m − fi )hm−fi i . (1)
i=1
Finally the last sum in (1) is bounded by H((F/L)ℓ , m) where L is a submodule such that Lm is generated by a lex-segment of length H(M, m): this
follows directly from Proposition 3.16.
✷
3.1
Green’s HRT after scaling for modules
We now derive from last theorem a result, that will give the possibility to
apply Green’s HRT in case it is difficult to have the Macaulay binomial representations. The idea is to consider the Hilbert function up to a multiplication
by an integer. In this way, we will obtain a linear inequality, that will hold
for all modules.
Corollary 3.19. Let M be a module over S = k[x1 , . . . , xn ], generated in
degree 0, and let ℓ be a generic linear form, d ∈ N, then
n−1
H(M, d).
n+d−1
δ
L
Mℓ , d = δH(Mℓ , d), so we
Proof For any integer δ, we have that H
H(Mℓ , d) ≤
i=1
can choose δ such that there is β ∈ N satisfying the following equality
n+d−1
,
δH(M, d) = βdimk (Sd ) = β
d
12
so by Theorem 3.18 we have
!
δ
M
n+d−2
n+d−1
=
=β
H
Mℓ , d ≤ β
d
d
hdi
i=1
!
δ
M
n−1
n+d−1
n−1
β
=
H
M, d ,
d
n+d−1
n+d−1
i=1
then, dividing by δ, we get the claim.
✷
Observation 3.20. This corollary can be described by the expression ”after
scaling” because the last fraction satisfies the following equality
n−1
dim(Rd )
=
,
n+d−1
dim(Sd )
where S = k[x1 , x2 , . . . , xn ] and R = k[x1 , x2 , . . . , xn−1 ] ∼
= Sℓ .
Therefore, the claim of the Corollary may be written as:
H(M, d)
H(Ml , d)
≤
,
H(Sl , d)
H(S, d)
so we have scaled by respective value of the Hilbert function of the rings over
which we have the two modules.
4
Application to level algebras
In this section we will apply the extension of the Green’s HRT to level algebras.
Definition 4.1. Let A = S/I = A0 ⊕ · · · ⊕ Ac be an artinian k-algebra. Let
us consider the socle of A, which is defined by soc(A) = annm (A), where
m = ⊕i>0 Ai . If soc(A) = U0 ⊕ · · · ⊕ Uc , we say that A is level of type b if
soc(A) = Uc = Ac and dim(Ac ) = b. We say that A is Gorenstein when it is
level of type 1.
Let A = A0 ⊕· · ·⊕Ac be a level algebra with Hilbert function (h0 , h1 , . . . , hc ).
By considering the dual of A, A∨ = Homk (A, k), we get a level module, since
13
is an artinian S-module with generators in a single degree, and with onedimensional socle, i.e. also concentrated in a single degree.
We want to consider the multiplication by a generic linear form ℓ ∈ A1 .
We first apply Theorem 2.7. In this way we get the following bound on the
Hilbert function of the ideal (ℓ).
dimk (ℓ)i ≥ hi − (hi )hii , 0 ≤ i ≤ c.
We can also apply Theorem 3.18 to the level module M = A∨ (−c). Notice that M has Hilbert function (hc , hc−1 , . . . , h0 ). We get in this case the
following bound for the module ℓM:
dim(ℓM)i ≥ hc−i − (hc−i ){i} , 0 ≤ i ≤ c.
In the right hand side of this
inequality, a{i} denotes q(si )hii + rhii , where
i+n−1
.
a = qsi + r and si =
i
∼
We observe that (ℓ) = (ℓM)∨ (−c + 1). This implies that we can give two
lower bounds for the Hilbert function of (ℓ), namely hGM
and hG
i
i , where we
GM
G
set hi = hi − (hi ){c−i} and hi = hi+1 − (hi+1 )hi+1i .
We would like to find conditions in which hGM
> hG
i
i , so when the bound
given by the HRT for modules is better than the one given by the Green’s
HRT.
In [6], Geramita, Harima, Migliore and Shin, in case the polynomial ring
S has three variables, provided a list of all the possible Hilbert functions for
level algebras of socle in degree 3, 4, 5 and in degree 6 and type 2. Among
these, the condition hGM
> hG
i
i is true only for the ones contained in Table 1
in page 15.
Proposition 4.2. For i such that the following conditions are all verified,
we have that hGM
≥ hG
i
i :
i − 1 ≤ c − i;
hi = hi+1 ;
sc−i > hi+1 .
Proof We have that hGM
= hi − (hi ){c−i} = hi+1 − (hi+1 ){c−i} . Since
i
sc−i > hi+1 , (hi+1 ){c−i} = (hi+1 )hc−ii . We may then compare (hi+1 )hc−ii and
(hi+1 )hi+1i , using Lemma 3.11, and the fact that i − 1 ≤ c − i, so we get that
14
hGM
≥ hG
i
i .
✷
The previous condition does not characterize all cases in which hGM
> hG
i
i ,
but only the ones in bold font in the table. Notice that if in the first column we have a certain number i this means, according to our notation, that
G
hGM
i−1 > hi−1 .
Question. What are the sufficient conditions for which the bound given by
the HRT for modules is better than the bound given by the classical HRT?
Position
2
2
2
4
4
4
2
4
4
2
3
3
4
4
5
5
5
5
5
5
5
h
1, 3, 3, 3, 2
1, 3, 3, 3, 3
1, 3, 3, 3, 3, 2
1, 3, 6, 8, 5, 2
1, 3, 6, 9, 5, 2
1, 3, 6, 10, 6, 2
1, 3, 3, 3, 3, 3
1, 3, 6, 8, 5, 3
1, 3, 6, 10, 6, 3
1, 3, 3, 3, 3, 3, 2
1, 3, 4, 4, 4, 3, 2
1, 3, 4, 4, 4, 4, 2
1, 3, 5, 6, 5, 4, 2
1, 3, 6, 8, 6, 4, 2
1, 3, 6, 8, 10, 6, 2
1, 3, 6, 9, 10, 6, 2
1, 3, 6, 9, 11, 6, 2
1, 3, 6, 9, 12, 6, 2
1, 3, 6, 10, 10, 6, 2
1, 3, 6, 10, 11, 6, 2
1, 3, 6, 10, 12, 6, 2
hGM
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
2,
2,
3,
5,
5,
5,
3,
5,
5,
3,
4,
4,
4,
5,
5,
5,
5,
5,
5,
5,
5,
1
1
2,
5,
5,
6,
2,
5,
6,
3,
3,
3,
5,
6,
6,
6,
6,
6,
6,
6,
6,
1
2
2
2
1
2
2
2,
3,
3,
3,
3,
6,
6,
6,
6,
6,
6,
6,
1
1
2
2
2
2
2
2
2
2
2
2
hG
1, 2,
1, 2,
1, 2,
1, 3,
1, 3,
1, 3,
1, 2,
1, 3,
1, 3,
1, 2,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
1, 3,
3,
3,
3,
6,
6,
6,
3,
6,
6,
3,
3,
3,
5,
6,
6,
6,
6,
6,
6,
6,
6,
2
3
3,
4,
4,
5,
3,
4,
5,
3,
4,
4,
4,
5,
8,
8,
9,
9,
8,
9,
9,
2
2
2
2
3
3
3
3,
3,
4,
4,
4,
5,
5,
5,
5,
5,
5,
5,
2
2
2
2
2
2
2
2
2
2
2
2
Table 1: Hilbert functions of Level artinian algebras
Acknowledgments. I would like to thank my supervisor Mats Boij for his
great support, and also Ralf Fröberg for the useful comments. Moreover, I
15
also thank Giulio Caviglia for his suggestions, that were crucial in the last
phase of preparation of the paper.
References
[1] J. Ahn and Y. S. Shin, Artinian level algebras of codimension 3, J.
Pure Appl. Algebra, 216 (2012), pp. 95–107.
[2] M. Boij and F. Zanello, Some algebraic consequences of Green’s
hyperplane restriction theorems, J. Pure Appl. Algebra, 214 (2010),
pp. 1263–1270.
[3] W. Bruns and J. Herzog, Cohen-Macaulay rings, vol. 39 of Cambridge Studies in Advanced Mathematics, Cambridge University Press,
Cambridge, 1993.
[4] D. Eisenbud, Commutative algebra, vol. 150 of Graduate Texts in
Mathematics, Springer-Verlag, New York, 1995. With a view toward
algebraic geometry.
[5] V. Gasharov, Extremal properties of Hilbert functions, Illinois J.
Math., 41 (1997), pp. 612–629.
[6] A. V. Geramita, T. Harima, J. C. Migliore, and Y. S. Shin,
The Hilbert function of a level algebra, Mem. Amer. Math. Soc., 186
(2007), pp. vi+139.
[7] M. Green, Restrictions of linear series to hyperplanes, and some results
of Macaulay and Gotzmann, in Algebraic curves and projective geometry
(Trento, 1988), vol. 1389 of Lecture Notes in Math., Springer, Berlin,
1989, pp. 76–86.
[8] H. A. Hulett, A generalization of Macaulay’s theorem, Comm. Algebra, 23 (1995), pp. 1249–1263.
[9] G. Katona, A theorem of finite sets, in Theory of graphs (Proc. Colloq.,
Tihany, 1966), Academic Press, New York, 1968, pp. 187–207.
[10] M. Kreuzer and L. Robbiano, Computational commutative algebra.
2, Springer-Verlag, Berlin, 2005.
16
[11]
, Computational commutative algebra 1, Springer-Verlag, Berlin,
2008. Corrected reprint of the 2000 original.
[12] J. B. Kruskal, The number of simplices in a complex, in Mathematical
optimization techniques, Univ. of California Press, Berkeley, Calif., 1963,
pp. 251–278.
[13] F. S. Macaulay, The algebraic theory of modular systems, Cambridge
Mathematical Library, Cambridge University Press, Cambridge, 1994.
Revised reprint of the 1916 original, With an introduction by Paul
Roberts.
[14] J. Migliore and F. Zanello, The strength of the weak Lefschetz
property, Illinois J. Math., 52 (2008), pp. 1417–1433.
[15] K. Pardue, Nonstandard borel-fixed ideals, ProQuest LLC, Ann Arbor,
MI, 1994. Thesis (Ph.D.)–Brandeis University.
Royal Institute of Technology, Department of Mathematics,
S-10044 Stockholm, Sweden.
E-mail address: [email protected]
17
| 0math.AC
|
Distributed Stochastic Model Predictive Control
Synthesis for Large-Scale Uncertain Linear Systems
arXiv:1703.06273v2 [math.OC] 28 Mar 2017
VAHAB ROSTAMPOUR AND TAMÁS KEVICZKY
Abstract. This paper presents an approach to distributed stochastic model predictive control
(SMPC) of large-scale linear systems with additive disturbances and multiplicative uncertainties in
a plug-and-play (PnP) framework. Typical SMPC approaches for such problems involve formulating
a large-scale finite-horizon chance-constrained optimization problem at each sampling time, which
is in general non-convex and difficult to solve. Using an approximation, the so-called scenario
approach, we formulate a large-scale scenario program and provide a theoretical guarantee to quantify
the robustness of the obtained solution. However, such a reformulation leads to a computational
tractability issue, due to the large number of required scenarios. To this end, we present two
novel ideas in this paper to address this issue. We first provide a technique to decompose the
large-scale scenario program into distributed scenario programs that exchange a certain number
of scenarios with each other in order to compute local decisions. We show the exactness of the
decomposition with a-priori probabilistic guarantees for the desired level of constraint fulfillment.
As our second contribution, we develop an inter-agent soft communication scheme based on a set
parametrization technique together with the notion of probabilistically reliable set to reduce the
required communication between each subproblem. We show how to incorporate the probabilistic
reliability notion into existing results and provide new guarantees for the desired level of constraint
violations. A simulation study is presented to illustrate the advantages of our proposed framework.
Keywords. Distributed MPC, Stochastic MPC, Scenario MPC, Plug-and-Play Framework.
1. Introduction
Stochastic model predictive control (SMPC) has attracted significant attention in the control
literature, due to its functionality to differently handle uncertain systems. SMPC takes into account
the stochastic characteristics of the uncertainties and thereby the system constraints are treated
in a probabilistic sense, i.e. using chance constraints [1, 2]. SMPC computes an optimal control
sequence that minimizes a given objective function subject to the uncertain system dynamics model
and chance constraints in a receding horizon fashion [3]. Chance constraints enable SMPC to offer
an alternative approach to achieve a less conservative solution compared to robust model predictive
control (MPC) [4], since it directly incorporates the tradeoff between constraint feasibility and control
performance.
Date: March 20, 2017.
This research was supported by the Uncertainty Reduction in Smart Energy Systems (URSES) research program
funded by the Dutch organization for scientific research (NWO) and Shell under the project Aquifer Thermal Energy
Storage Smart Grids (ATES-SG) with grant number 408-13-030.
The authors are with Delft Center for Systems and Control, Delft University of Technology, Mekelweg 2, 2628 CD,
Delft, The Netherlands. {v.rostampour, t.keviczky}@tudelft.nl.
1
2
V. ROSTAMPOUR AND T. KEVICZKY
Distributed MPC has been an active research area in the past decades, due to its applicability in
different domains such as power networks [5], chemical plants [6], process control [7], and building
automation [8]. For such large-scale dynamic systems with state and input constraints, distributed
MPC is an attractive control scheme. In distributed MPC one replaces large-scale optimization
problems stemming from centralized MPC with several smaller-scale problems that can be solved
in parallel. These problems make use of pieces of information from other subsystems to design a
distributed MPC. In the presence of uncertainties, however, the main challenge in the formulation of
distributed MPC is how the controllers should exchange pieces of information through a communication scheme among subsystems (see, e.g. [9], and references therein). This highlights the necessity
of developing distributed control strategies to cope with the uncertainties in subsystems while at the
same time minimizing information exchange through a communication framework.
To handle uncertainties in distributed MPC, some approaches are based on robust MPC [10, 11].
Assuming that the uncertainty is bounded, a robust optimization problem is solved at each sampling
time, leading to a control law that satisfies the constraints for all admissible values of the uncertainty.
The resulting solution using such an approach tends to be very conservative in many cases. Tubebased MPC, see for example [12] and the references therein, was considered in a plug-and-play (PnP)
decentralized setup in [13], and it has been recently extended to distributed control systems [14]
for a collection of linear stochastic subsystems with independent dynamics. While in [14] coupled
chance constraints were considered separately at each sampling time, in this paper we consider a
chance constraint on the feasibility of trajectories of dynamically coupled subsystems. Our approach
is motivated by [13] to reduce the conservativeness of the control design. Other representative
approaches for SMPC of a single stochastic system include affine parametrization of the control
policy [15], the randomized (scenario) approach [16–19], and the combined randomized and robust
approach [20–22]. None of these approaches, to the best of our knowledge, have been considered in
distributed control strategies.
This paper aims to establish some fundamental techniques for distributed SMPC by promoting
the scenario MPC technique to the distributed case in a more systematical approach. Scenario MPC
approximates SMPC via the so-called scenario (sample) approach [23, 24], and if the underlying
optimization problem is convex with respect to the decision variables, finite sample guarantees can
be provided. Following such an approach, the computation time for a realistic large-scale system of
interest becomes prohibitive, due to the fact that the number of samples to be extracted tends to
be very high, and consequently leads to a large number of constraints in the resulting optimization
problem. To overcome the computational burden caused by the large number of constraints in
[25, 26] a heuristic sampled-based approach was used in an iterative distributed fashion via dual
decomposition such that all subsystems collaboratively optimize a global performance index. In
another interesting recent work [27], a distributed multi-agent consensus algorithm was presented
to achieve consensus on a common value of the decision vector subject to random constraints. The
authors established a probabilistic bound on the tails of the consensus violation, similarly to [28]. In
addition, they studied a case where the constraints of agents are affected by a common uncertainty
source and derived a-posteriori probabilistic bounds on the number of shared scenarios based on the
results in [29] to reduce significantly the conservativeness of the resulting solution. We however note
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
3
that in most of the aforementioned references the aim to reduce communication among subsystems,
which we refer to as agents, has not been addressed.
Our work in this paper differs from the aforesaid references in two important aspects. A decomposition technique based on the large-scale system dynamics is employed to distribute the resulting
centralized scenario optimization problem at each sampling time and a novel communication scheme
is introduced to reduce the communication between the small-scale problems. The main contributions
of this paper are twofold:
a) We provide a technique to decompose the large-scale scenario program into distributed scenario
programs that exchange a certain number of scenarios with each other in order to compute local
decisions. We show the exactness of the decomposition with a-priori probabilistic guarantees
for the desired level of constraint fulfillment.
b) We develop an inter-agent soft communication scheme based on a set parametrization technique together with the notion of probabilistically reliable set to reduce the required communication between each subproblem. We show how to incorporate the probabilistic reliability
notion into existing results and provide new guarantees for the desired level of constraint
violations.
The structure of this paper is as follows: Section 2 describes a mathematical model of the control
system dynamics together with some preparatory results. In this section, we first formulate a SMPC
for a large-scale uncertain linear system with additive disturbance, then provide a reformulation,
namely the centralized scenario MPC, and provide a theoretical study on the connections of these
two control problems. In Section 3, we propose a decomposition technique to distributed scenario
MPC, and analyze the robustness of the obtained solution compared to the centralized scenario MPC
formulated in the previous section. Section 4 introduces two types of inter-agent communication
schemes between each subproblem, namely hard and soft communications. We then quantify the
robustness of the proposed schemes and provide a new theoretical guarantee. Section 5 provides
a summary on our developments to establish an applied PnP distributed SMPC framework, and
Section 6 presents a simulation study to illustrate the functionality of our theoretical achievements.
In Section 7 we conclude this paper with some remarks and future work.
Notations
R, R+ denote the real and positive real numbers, and N, N+ the natural and positive natural
numbers, respectively. We operate within n-dimensional space Rn composed by column vectors
Q
u, v ∈ Rn . The Cartesian product over n sets X1 , · · · , Xn is given by: ni=1 Xi = X1 × · · · × Xn =
{(x1 , · · · , xn ) : xi ∈ Xi }. The cardinality of a set A is shown by |A| = A. We denote a blockdiagonal matrix with blocks Xi , i ∈ {1, · · · , n}, by diagi∈{1,··· ,n} (Xi ), and a vector consisting of
stacked sub vectors xi , i ∈ {1, · · · , n}, by coli∈{1,··· ,n} (xi ).
Given a metric space ∆, its Borel σ-algebra is denoted by B(∆). Throughout the paper, measurability always refers to Borel measurability. In a probability space (∆, B(∆), P), we denote the
N -Cartesian product set of ∆ by ∆N with the respective product measure by PN .
4
V. ROSTAMPOUR AND T. KEVICZKY
x1
xi
u···
xi
y2 y···
y1
u1
yM
Agent M
xM
x···
u2
x2
Agent 1
x1
xM
Agent 2
x2
yi
Centralized
Scenario MPC
y···
ui
yj
uM
Agent i
xi
xj
uj
u···
xj
x···
x1
Agent j
xj
xM
Figure 1. Centralized scenario MPC that corresponds to the problem (7).
2. Problem Statement and Preparatory Results
This section provides an overview on the control problem statement. We first describe a large-scale
uncertain linear system dynamic together with input and state constraint sets and the control objective. We then formulate a centralized SMPC for such a large-scale control system problem. Finally
a tractable reformulation based on the scenario MPC [30] together with theoretical connections are
provided.
Consider a discrete-time uncertain linear system with additive disturbance in a compact form as
follows:
xk+1 = A(δk )xk + B(δk )uk + C(δk )wk ,
(1)
with a fixed initial condition x0 ∈ Rm . Here k ∈ T := {0, 1, · · · , T − 1} denotes instance of time,
xk ∈ X ⊂ Rm and uk ∈ U ⊂ Rp correspond to the state and control input, respectively, and wk ∈ Rn
is related to an additive disturbance. The system parameters A(δk ) ∈ Rm×m and the control input
B(δk ) ∈ Rm×p as well as C(δk ) ∈ Rm×n are random, since they are known functions of an uncertain
variable δk that influences the system parameters at each time step k.
Assumption 1. w := {wk }k∈T and δ := {δk }k∈T are defined on probability spaces (W, B(W), Pw )
and (∆, B(∆), Pδ ), respectively. w and δ are two independent random processes, where Pw and Pδ
are two different probability measures defined over W ⊆ Rq and ∆ ⊆ Rm , respectively, and B(·)
denotes a Borel σ-algebra.
The support sets W and ∆ of w and δ, respectively, together with their probability measures Pw
and Pδ are entirely generic. In fact, W, ∆ and Pw , Pδ do not need to be known explicitly. Instead,
the only requirement is availability of a "sufficient number" of samples, and will become concrete in
later parts of the paper.
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
5
The system in (1) is subject to hard constraints on the system state trajectories and control input.
Consider the state and control input constraint sets of the form
X := { x ∈ Rm : G x ≤ g } , U := { u ∈ Rp : H u ≤ h } ,
where G ∈ Rq×m , g ∈ Rq , and H ∈ Rr×p , h ∈ Rr . Keeping the state inside a feasible set X ⊂ Rm
for the entire prediction horizon steps maybe too conservative and results in a poor performance.
In particular, this is the case when the best performance is achieved close to the boundary of X ,
and thus, constraint violations will be unavoidable due to the fact that the system parameters are
imperfect and uncertain (1). To tackle such a problem, we will consider chance constraints on
the state trajectories to avoid violation of the state variables constraints probabilistically even if the
disturbance w or uncertainty δ has unbounded support. Notice that a robust problem formulation [4]
cannot cope with problems having an unbounded disturbance set.
The control task is to choose the control inputs u := {uk }k∈T from an input constraint set U
via a state feedback law φ : Rm → U with uk = φ(xk ). In order to find such a stabilizing fullinformation controller that leads to admissible control inputs and satisfies the state constraints, we
follow the traditional MPC approach. The design relies on the standard assumption of the existence
of a suitable pre-stabilizing control law, e.g., [13, Proposition 1]. To cope with the state prediction
under uncertainty and disturbance, we employ a parametrized feedback policy [15] for (1) and split
the control input, uk = Kxk + vk , with vk as a free correction input variable to compensate for
disturbances.
The control objective is to determine a receding horizon control, which minimizes a cumulative
quadratic stage cost of an infinite horizon that is as follows:
t
J∞
i
1X h >
Ruk ,
E xk Qxk + u>
:= lim
k
t→∞ t
(2)
k=0
1
p×p
2
where x := {xk }k∈T , Q ∈ Rm×m
0 , R ∈ R 0 , and it is assumed that (A, Q ) is detectable. Each stage
cost term is taken in expectation E[·], since the argument xk is a random variable. The formulated
cost function in (2) has a very large time horizon t ∈ T , i.e. the life cycle of the system [18], and
particularly, it is much larger than any practical prediction horizon T ∈ N (t >> T ). Therefore, the
control problem cannot be solved for the entire prediction horizon t ∈ T . We therefore recast J∞
into a finite horizon cost J(·) : Rm × Rp → R to be minimized at time k and is defined as
J(x, u) := E
"T −1
X
x>
k Qxk
+
u>
k Ruk
#
+
x>
T P xT
,
(3)
k=0
and P is the solution of discrete-time Lyapunov equation:
E[ Acl (δk )> P Acl (δk ) ] + Q + K > RK − P 0 ,
for the closed-loop system, e.g., Acl (δk ) = A(δk ) + B(δk )K.
(4)
6
V. ROSTAMPOUR AND T. KEVICZKY
Using v = {vk }k∈T , consider now the following stochastic control problem as follows:
min
v∈RT m
J(x, u)
(5a)
s.t. xk+1 = A(δk )xk + B(δk )uk + C(δk )wk ,
xk = x0 ,
(5b)
P[ xk+` ∈ X , ` ∈ N+ ] ≥ 1 − ε ,
(5c)
uk = Kxk + vk ∈ U ,
(5d)
∀k ∈ T ,
where ε ∈ (0, 1) is the admissible state constraint violation parameter of the large-scale system (1).
Remark 1. The objective function is assumed to be a quadratic function, however, this is not a
restriction and any generic convex function can be chosen instead. It is important to mention that
the parameters of constraint sets, X , U, and the objective function J(·) can be time-varying with
respect to each sampling time k ∈ T . For the clarity of our problem formulation, we assume timeinvariance.
Remark 2. The state trajectory xk+` , ∀` ∈ N+ , has a dependency on the random variables w and
δ, and thus, the chance constraint can be interpreted as follows: the probability of violating the state
constraint at the future time step ` ∈ N+ is restricted to ε, given that the state of the system in
(1) is measurable at each time step k ∈ T . Even though U and X are convex compact sets, due to
the chance constraint on the state trajectory, the feasible set of the optimization problem in (5) is a
non-convex set, in general.
Remark 3. Instead of the chance constraint on the state trajectory of form (5c), one can also bound
the average rate of state constraint violations [30]. Moreover, one can also define the cost function
(5a) as a desired quantile of the sum of discounted stage costs ("value-at-risk"), instead of the sum
of expected values.
Remark 4. Instead of a state feedback law, one can also consider a nonlinear disturbance parametrization feedback policy over the prediction horizon, similar to [22], using the scenario-based approximation. Such a parametrization does not affect the convexity of the resulting optimization problem [16].
To handle the chance constraint (5c), we recall a scenario-based approximation [30]. wk and
δk at each sampling time k ∈ T are not necessarily independent and identically distributed (i.i.d.).
Particularly, they may have time-varying distributions and/or be correlated in time. We assume that
a "sufficient number" of i.i.d. samples of the string of disturbance w ∈ W and δ ∈ ∆ can be obtained
either empirically or by a random number generator. We denote Sw := {w(1) , · · · , w(Ns ) } ∈ W Ns
and Sδ := {δ (1) , · · · , δ (Ns ) } ∈ ∆Ns as a set of given finite samples (scenarios).
Following the approach in [6], we approximate the expected value of the objective function empirically by averaging the value of its argument for some number of different scenarios, which plays a
tuning parameter role. Using Ns̄ as the tuning parameter, consider Ns̄ number of different scenarios
of w and δ to build S̄w = {w(1) , · · · , w(Ns̄ ) } ⊂ W Ns̄ and S̄δ := {δ (1) , · · · , δ (Ns̄ ) } ∈ ∆Ns̄ , respectively.
We then approximate the cost function empirically as follows:
" T −1
#
T
−1
X
X
X
(i) (i)
J(x, u) = E{w,δ}∈W×∆
V (xk (wk , δk ), uk ) ≈
V (xk (wk , δk ), uk ) , (6)
k=0
{w(i) ,δ (i) }∈S̄w ×S̄δ k=0
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
7
>
where V (xk (w, δk ), uk ) = xk (wk , δk )> Qxk (wk , δk ) + u>
k Ruk + xT (wk , δk ) P xT (wk , δk ). Note that
xk (wk , δk ) denotes the dependency of the state variables on the random variables.
We are now in a position to formulate a tractable version of the proposed stochastic control
problem in (5) using the following finite horizon scenario program:
min
v∈RT m
J(x, u)
(i)
(7a)
(i)
(i)
(i)
(i)
(i)
(i)
s.t. xk+1 = A(δk )xk + B(δk )uk + C(δk )wk ,
(i)
xk = x0 ,
(i)
xk+` ∈ X , ` ∈ N+ , ∀w(i) ∈ Sw , ∀δ (i) ∈ Sδ
(i)
(i)
uk = Kxk + vk ∈ U ,
∀k ∈ T .
(7b)
(7c)
(7d)
∗
The solution of (7) is the optimal control input sequence v ∗ = {vk∗ , · · · , vk+T
−1 }. Based on an MPC
∗
paradigm, the current input is implemented as uk := Kxk + vk and we proceed in a receding horizon
fashion. This means (7) is solved at each time step k by using the current measurement of the state
xk . Note that new scenarios are needed to be generated at each sampling time k ∈ T .
The key features of the proposed tractable optimization problem (7) are as follows:
• there is no need to know the probability measures Pw and Pδ explicitly, only the capability of
obtaining random scenarios is enough.
• formal results to quantify the robustness of the obtained approximations are available. In
particular the results follow the so-called scenario approach [24], which allow to bound apriori the violation probability of the obtained solution via (7).
In the following theorem, we restate the explicit theoretical bound of [24, Theorem 1] which
measures the finite scenarios behavior of (7).
Theorem 1. Let β ∈ (0, 1) and Ns ≥ N (ε, β, T m), where
(
)
TX
m−1
N i
N (ε, β, T m) := min N ∈ N
ε (1 − ε)N −i ≤ β .
i
i=0
v∗,
If the optimizer of the problem (7),
is applied to the discrete time dynamical system (1) for a
finite horizon of length T , then, with at least confidence 1 − β, the original chance constraint (5c) is
satisfied for all k ∈ T .
It was shown in [24] that the above bound is tight. The interpretation of Theorem 1 is as follows:
when applying v ∗ in a finite horizon control problem, the violation of the feasibility of the state
trajectory remains below ε with confidence 1 − β:
PNs Sw ∈ W Ns , Sδ ∈ ∆Ns : Vio(v ∗ ) ≤ ε ≥ 1 − β ,
with
Vio(v ∗ ) := P w ∈ W, δ ∈ ∆ : xk+` = Acl (δk )xk + B(δk )vk∗ + C(δk )wk ∈
/ X , ` ∈ N+ xk = x0 ,
where Acl (δk ) = A(δk ) + B(δk )K. It is worth to mention that the proposed hard constraint on the
control input in (7d) is also met in a probabilistic sense, due to the nature of the scenario approach
that appears in the proposed optimization problem (7).
8
V. ROSTAMPOUR AND T. KEVICZKY
Remark 5. One can obtain an explicit expression for the desired
number
of scenarios Ns as in [31],
e 1
1
where it is shown that given ε, β ∈ (0, 1) then Ns ≥ e−1 ε T m + ln β . It is important to mention
that Ns is used to construct the sets of scenarios, Sw , Sδ to obtain a probabilistic guarantee for the
desired level of feasibility. The number of scenarios Ns̄ can be interpreted as a tuning variable to
approximate the objective function empirically.
We formulated a large-scale SMPC (5c) together with a tractable reformulation based on the
proposed centralized scenario MPC (7). Figure 1 shows a pictorial representation of (7) as a largescale network of interconnected agents to summarize this section. In the following section, we will
provide a distributed framework to solve the proposed problem in (7) by decomposing the large-scale
system dynamics (1).
3. Distributed Scenario Model Predictive Control
In this section, we first describe a decomposition technique to partition the large-scale system
dynamics in (1) that we will get benefit in the subsequent sections. We then present the theoretical
links to the results that we provided in the previous section.
Consider the system (1) that is assumed to be decomposable into M subsystems and let M =
{1, 2, · · · , M } be the set of subsystem indices (for more details on the decomposition techniques
of a large-scale system, we refer interested reader to [32, Chapter 3]). The state variables xk ,
control input signals uk and the additive disturbance wk , can be considered as xk = coli∈M (xi,k ),
uk = coli∈M (ui,k ), and wk = coli∈M (wi,k ), respectively, where xi,k ∈ Rmi , ui,k ∈ Rpi , wi,k ∈ Rni ,
P
P
P
and i∈M mi = m, i∈M pi = p, i∈M ni = n. The following assumption is important in order to
be able to partition the system parameters.
Assumption 2. It is assumed that the control input and the disturbance variables of the subsystems
are decoupled,. e.g. ui,k and wi,k only affect subsystem i ∈ M for all k ∈ T .
We are now able to decompose the large-scale system matrices A(δk ) ∈ Rm×m , B(δk ) ∈ Rm×p , as
well as C(δk ) ∈ Rm×n as follows:
A11 (δk ) · · · A1m (δk )
B11 (δk ) · · ·
0
C1 (δk )
.
.
..
..
..
..
..
A(δk ) =
, B(δk ) = ..
, C(δk ) = .. ,
.
.
.
.
.
Am1 (δk ) · · · Amm (δk )
0
· · · Bmm (δk )
Cm (δk )
where Aij (δk ) ∈ Rmi ×mj , Bii (δk ) ∈ Rmi ×pi , and Ci (δk ) ∈ Rmi ×ni . Define the set of neighboring
subsystems of subsystem i as follows:
Ni = j ∈ M\i Aij (δk ) 6= 0 ,
(8)
where 0 denotes a zero matrix with proper dimension. Consider now a large-scale network that
consists of M interconnected subsystems, and each subsystem can be described by a linear uncertain
discrete-time invariant system with additive disturbance of the form
x
i,k+1 = Aii (δk )xi,k + Bii (δk )ui,k + qi,k
,
(9)
P
qi,k
=
Aij (δk )xj,k + Ci (δk )wi,k
j∈Ni
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
9
where for each subsystem i, xi,k ∈ Xi ⊆ Rmi , ui,k ∈ Ui ⊆ Rpi , and wi,k ⊆ Rni . We define the state
and control input constraint sets of each subsystem i ∈ M as the following form:
Xi := { x ∈ Rmi : [G]i x ≤ [g]i } , Ui := { u ∈ Rpi : [H]i u ≤ [h]i } ,
Q
Q
such that X = i∈M Xi , and U = i∈M Ui . Following Assumption 2, one can consider a linear
feedback gain matrix Ki for each subsystem i ∈ M such that K = diagi∈M (Ki ). Using Ki in each
subsystem, we assume that there exists Pi for each subsystem i ∈ M such that P = diagi∈M (Pi ) to
preserve the condition in (4). Consider now the objective function of each subsystem i ∈ M in the
following form:
"T −1
#
X
>
>
>
Ji (xi , ui ) := E
xi,k Qi xi,k + uk Ri ui,k + xi,T Pi xi,T ,
k=0
i ×pi
i ×mi
where Qi ∈ Rm
, Ri ∈ Rp 0
such that Q = diagi∈M (Qi ), and R = diagi∈M (Ri ). Note that
0
xi = colk∈T (xi,k ) and ui = colk∈T (ui,k ) such that x = coli∈M (xi ) and u = coli∈M (ui ).
Remark 6. For sake of simplicity of the mathematical notations, a decomposition of multiplicative
uncertainty δk is not considered. We however note that such a decomposition is straightforward by
considering δi,k for each subsystem i ∈ M. This leads to Aii (δi,k ), Bii (δi,k ), Ci (δi,k ), and an effect
on the state coupling matrices between subsystems Aij (δi,k ) for all j ∈ Ni in each subsystem i ∈ M.
Using v i = colk∈T (vi,k ) such that v = coli∈M (v i ), we decompose the proposed formulation in (7)
using the following finite horizon scenario program for each subsystem i ∈ M:
min
v i ∈RT mi
s.t.
Ji (xi , ui )
(10a)
(i)
(i)
(i)
(i)
(i)
(i)
xi,k+1 = Aii (δk )xi,k + Bii (δk )ui,k + qi,k ,
(i)
(i)
xi,k+` ∈ Xi , ` ∈ N+ , ∀wi ∈ Swi , ∀δ (i) ∈ Sδ
(i)
(i)
ui,k = Ki xi,k + vi,k ∈ Ui ,
(i)
xi,k = xi,0 ,
∀k ∈ T ,
(10b)
(10c)
(10d)
Q
(Ns )
Ns
(1)
where wi = colk∈T (wi,k ) ∈ Wi such that W = i∈M Wi . Swi := {wi , · · · , wi i } ∈ Wi i
denotes a set of given finite samples (scenarios) of disturbance in each subsystem i ∈ M, such that
Q
Sw = i∈M Swi . Note that we use a parenthesis to refer to each scenario of the random variables,
e.g.(i), whereas without the parenthesis refers to each subsystem i ∈ M.
Remark 7. The proposed constraint (10c) represents an approximation of the following chance
constraint on the state of each subsystem i ∈ M:
P[ xi,k+` ∈ Xi , ` ∈ N+ ] ≥ 1 − εi ,
(11)
where εi ∈ (0, 1) is the admissible state constraint violation parameter of each subsystem (9). One
can also consider αi = 1 − εi as the desired level of state feasibility parameter of each subsystem (9).
In the following proposition, we provide a connection between the proposed optimization problem
in (10) and the optimization problem in (7).
Proposition 1. Given Assumption 2, the proposed optimization problem in (10) is an exact decomposition of the optimization problem in (7).
10
V. ROSTAMPOUR AND T. KEVICZKY
Proof. Following the proposed structure of decomposition, any optimizer of each subsystem v ∗i yields
a feasible pair of the state and control input variables of its subsystem {x∗i , u∗i } ∈ Xi × Ui such that
Q
Q
Xi = k∈T Xi , and Ui = k∈T Ui . Therefore, the collection of the optimizers v ∗ = coli∈M (v ∗i ) will
yield the collection of feasible pairs of the state and control input variables of their subsystem:
)
) (
(
Y
Y
(12)
Ui ,
{x∗ = coli∈M (x∗i ) , u∗ = coli∈M (u∗i )} ∈ X =
Xi × U =
i∈M
i∈M
which eventually yields a feasible point for the optimization problem in (7). It is straightforward
to use the above relation and to show that any optimizer of the optimization problem in (7) v ∗
also yields a feasible solution for the proposed optimization problem in (10). We then have to show
that both optimization problems will have the same performance index in terms of their objective
function values. Due to the proposed decomposition technique, it is easy to see that the objective
function in (7) can be formulated as additive components such that each component represents the
objective function of each subsystem i ∈ M, and thus:
X
J(x∗ , u∗ ) =
Ji (x∗i , u∗i ) .
i∈M
The proof is completed.
The following theorem can be considered as the main result of this section to quantify the robustness of the solutions obtained by (10).
P
Theorem 2. Let εi , βi ∈ (0, 1) for all i ∈ M be chosen such that ε =
i∈M εi ∈ (0, 1) and
P
∗
∗
β = i∈M βi ∈ (0, 1). If v = coli∈M (v i ) , the collection of the optimizer of the problem (10) for all
subsystem i ∈ M, is applied to the discrete time dynamical system (1) for a finite horizon of length
T , then, with at least confidence 1 − β, the original chance constraint (5c) is satisfied for all k ∈ T .
Proof. The proof is a direct result of combining Proposition 1 with Theorem 1, and uses arguments
similar to the ones adopted in [33, Theorem 1].
The interpretation of Theorem 2 is as follows. In the proposed distributed scenario program (10),
each subsystem i ∈ M can have a desired level of constraint violation εi and a desired level of
confidence level βi . To keep the robustness level of collection of the solutions in a probabilistic sense
(5c) for the discrete-time dynamical system (1), these choices have to follow a certain design rule,
P
P
e.g. ε = i∈M εi ∈ (0, 1) and β = i∈M βi ∈ (0, 1). This yields a fixed ε , β for the large-scale
system (1) and the individual εi , βi for each subsystem i ∈ M.
Remark 8. An important key feature of the proposed distributed scenario program in (10) compared
to the optimization problem in (7) is as follows. Using the proposed distributed framework, we
decompose a large-scale scenario program (7) with Ns number of scenarios into M = |M| smallscale scenario programs (10) with Nsi number of scenarios. This yields a significant reduction in the
computation time complexity of scenario programs similar to (7) by using the proposed distributed
scenario program (10).
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
Scenario
MPC 2
y2
u2
Sx1
Agent 2
x2
Scenario
y··· MPC ...
Sx···
u···
y1
Sx2
Scenario
Agent 1
MPC 1 u1
x1
Sxi
SxM
Sxi
Sx1
yM
Scenario
Agent M
MPC M u
xM
M
Sx···
Scenario
MPC ...
u···
Sxj
Agent i
xi
ui
yi
Scenario
MPC i
Sxj
SxM
y···
11
Agent j
xj
uj
y
j
Scenario
MPC j
Figure 2. Distributed scenario MPC which corresponds to the problem (10).
(i)
Using the subsystem dynamics in (9), one can substitute qi,k in the proposed scenario optimization
problem (10) with the following relation:
X
(i)
(i) (i)
(i)
(i)
(i)
(i) (i)
qi,k =
Aij (δk )xj,k + Ci (δk )wi,k := f (δk , wi,k , xj,k ) .
(13)
j∈Ni
(i)
δk
(i)
and wi,k are the local scenario of random variables that are available in each subsystem by
(i)
definition wi ∈ Swi and δ (i) ∈ Sδ . Hence, the only information that the subsystem i ∈ M needs
(i)
(i)
is a Nsi number of samples of the state variable xj = colk∈T (xj,k ) ∈ Xj from all its neighboring
subsystems j ∈ Ni at each sampling time k ∈ T .
Remark 9. It is important to note that even though the proposed distributed scenario program in
(10) yields a reduction of computation time complexity (see Remark 8), it however requires more
communication between each subsystem, since at each sampling time k ∈ T all neighboring agents
(Ns )
Ns
(1)
j ∈ Ni of agent i should send a set of scenarios of the state variable Sxj := {xj , · · · , xj i } ∈ Xj i
to the agent i ∈ M.
To summarize this section, we present a network of agents that are dynamically coupled with
their own local scenario MPC in Figure 2. In the next section, we will propose a novel inter-agent
information exchange scheme to provide a more flexible together with less conservatism framework
to exchange information between agents.
4. Inter-Agent Information Exchange Scheme
This section first describes the information exchange between agents and then proposes a setbased information exchange scheme which will be called as a soft communication protocol later
12
V. ROSTAMPOUR AND T. KEVICZKY
Agent j
xj
uj
y
j
(i)
xj
Sxj
Scenario
MPC j
Bj
B̃j
Agent i
xi
ui
yi
Scenario
MPC i
Figure 3. Pictorial representation of the proposed inter-agent soft communication
(i)
scheme. Sxj is the set of Ñsi scenarios of xj , Bj is the parametrized set used in the
optimization problem (14), and B̃j is the solution of the optimization problem (14).
in this section. We finally provide the theoretical results to quantify robustness of the proposed
information exchange scheme between agents.
When the proposed distributed framework (10) is applied to the large-scale scenario program (7),
all neighboring agents j ∈ Ni of agent i ∈ M should send a set of scenarios of the state variable
(Ns )
(1)
Sxj := {xj , · · · , xj i } to the agent i at each sampling time k ∈ T . It is of interest to address the
issue of how an agent j ∈ Ni is going to send the contents of Sxj to the agent i ∈ M .
We propose the following two schemes:
a) Following our proposed setup in (10) to achieve a probabilistic guarantee for the obtained
solution, agent i ∈ M requests from its neighboring agents to send the complete set of data
Sxj , element by element. It is important to mention that the number of required samples
Nsi , is chosen according to Theorem 2 in order to have a given probabilistic guarantee for the
optimizer v i . We refer to this scheme as a hard communication protocol between agents. Its
advantage is that it is simple and transmits exactly the contents of Sxj , but due to possibly high
values of Nsi it may turn out to be too costly in terms of required communication bandwidth.
b) To address this shortcoming, we propose another scheme, where instead agent j ∈ Ni sends
a suitable parametrization of a set that contains all the possible values of its privatized data
with a desired level of probability (the level of reliability) α̃j . By considering a simple family
of sets, for instance boxes in Rmj , communication cost can so be kept down at reasonable
levels. We refer to this scheme as a soft communication protocol between agents (see Fig. 3).
We now describe the soft communication protocol in more details. The neighboring agent j ∈ Ni
has to first generate Ñsi samples of xj in order to build the set Sxj . It is important to notice that in
the soft communication protocol the number Ñsi of samples generated by agent j may be different
than the one needed by agent i, which is Nsi , as will be remarked later. Let us then introduce
Bj ⊂ Rmj as a bounded set containing all the elements of Sxj . We assume for simplicity that Bj is
an axis-aligned hyper-rectangular set. This is not a restrictive assumption and any convex set could
have been chosen instead as described in [33]. We can define Bj := [−bj , bj ] as an interval, where
the vector bj ∈ Rmj defines the hyper-rectangle bounds.
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
13
Consider now the following optimization problem that aims to determine the set Bj with minimal
volume:
min
kbj k1
bj ∈Rmj
(l)
(l)
(14)
s.t.
xj ∈ [−bj , bj ] , ∀xj ∈ Sxj ,
l = 1, · · · , Ñsi
where Ñsi is a number of samples xj ∈ Sxj that neighboring agent j has to take in account in
order to determine Bj . If we denote by B̃j = [−b̃j , b̃j ], the optimal solution of (14) computed by
the neighbor agent j, then for implementing the soft communication protocol, the agent j needs to
communicate only the vector b̃j along with the level of reliability α̃j to the agent i.
Definition 1. We call a set B̃j is α̃j −reliable if
h
i
P xj ∈ Xj : xj ∈
/ [−b̃j , b̃j ] ≤ 1 − α̃j ,
(15)
and we refer to α̃j as the level of reliability of the set B̃j .
We now provide the following theorem to determine α̃j as the level of reliability of the set B̃j .
Theorem 3. Fix β̃j ∈ (0, 1) and let
v
u
u β̃j
Ñ
−m
s
α̃j = i t
j
.
Ñs
(16)
i
mj
We then have
h
i
o
n
Ñs
Ñs
/ [−b̃j , b̃j ] ≤ 1 − α̃j ≥ 1 − β̃j .
PÑsi {x1j , · · · , xj i } ∈ Xj i : P xj ∈ Xj : xj ∈
(17)
Proof. Equation (17) is a direct result of the scenario approach theory in [23], if β̃j is chosen such
that
Ñsi Ñsi −mj
α̃j
≤ β̃j .
(18)
mj
Considering the worst-case equality in the above relation and some algebraic manipulations, one can
obtain the above assertion.
Theorem 3 implies that given an hypothetical new sample xj ∈ Xj , agent j ∈ Ni have a confidence
of at least 1 − β̃j that the probability of xj ∈ B̃j = [−b̃j , b̃j ] is at least α̃j . Therefore, one can rely
on B̃j up to α̃j probability.
Remark 10. The number of samples Ñsj in the proposed formulation (14) is a design parameter
chosen by the neighboring agent j ∈ Ni . We however remark that one can also set a given α̃j as the
desired level of reliability and obtain from (16) the required number of samples Ñsi .
Remark 11. It is worth to mention that in the proposed optimization problem (14), the performance
index is considered to be the minimal volume of the parametrized set Bj . The reason of such a choice
is to determine a less conservative (minimal volume) set that is α̃j −reliable. We however note that
one can also consider different performance indices, which might not yield the minimal volume, but
14
V. ROSTAMPOUR AND T. KEVICZKY
it should be taken into account that such a solution will be more conservative. Moreover, the choice
of different types (hyper-rectangular, ellipsoidal, and etc.) of the parametrized set Bj in (14) may
also introduce some level of conservativeness.
Remark 12. After receiving the parametrization of B̃j and the level of reliability α̃j , agent i can
(i)
then obtain the samples needed for computing qi,k by either locally generating Nsi samples, drawing
them uniformly from B̃j , or taking the worst-case of B̃j . It is important to notice that in this way,
we decoupled the sample generation of agent j ∈ Ni from the agent i ∈ M.
When an agent i ∈ M and its neighbor j ∈ Ni adopt the soft communication scheme, there is
an important effect on the probabilistic feasibility of agent i, following Remark 7. Such a scheme
introduces some level of stochasticity on the probabilistic feasibility of agent i, due to the fact that
the neighboring information is only probabilistically reliable. This will affect the local probabilistic
robustness guarantee of feasibility as it was discussed in Theorem 2 and consequently in Theorem 1.
To accommodate the level of reliability of neighboring information, we need to marginalize a joint
cumulative distribution function probability (cdf) of xi and the generic sample xj ∈ Xj appearing
in Theorem 3. We thus have the following theorem, which can be regarded as the main theoretical
result of this section.
Theorem 4. Given α̃j ∈ (0, 1) and a fixed αi ∈ (0, 1), the state trajectory of a generic agent i ∈ M
is probabilistically ᾱi –feasible for all wi ∈ Wi , δ ∈ ∆, i.e.,
where ᾱi = 1 −
1−αi
α̃i
P [xi,k+` ∈ Xi , ` ∈ N+ ] ≥ ᾱi ,
Q
such that α̃i = j∈Ni (α̃j ).
(19)
Proof. The proof consists of two important steps. We first provide an analytical expression for the
robustness of the solution in agent i by taking into account the effect of just one neighboring agent
j ∈ Nj , and then extend the obtained results for the case when the agent i interacts with more
neighboring agents, e.g. for all j ∈ Nj .
Following Remark 7, we have the following updated situation:
h
i
αi ≤ P xi ∈ Xi , xj ∈ B̃j ,
which is a joint probability of xi ∈ Xi and xj ∈ B̃j . Such a joint probability can be equivalently
written as a joint cumulative distribution function (CDF):
h
i
α ≤ P xi ∈ Xi , xj ∈ B̃j
Z Z
(20)
=
p(xi , xj ) dxi dxj
Xi
B̃j
= Fxi , xj (Xi , B̃j ) ,
where p(xi , xj ) is the joint probability density function (PDF) of xi and xj . Our goal is to calculate:
Z
P [ xi ∈ Xi ] =
p(xi ) dxi
Xi
= Fxi (Xi ) ,
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
15
where p(xi ) is the PDF of xi . To Transform the joint CDF into the marginal CDF of xi , one can
take the limit of the joint CDF as B̃j approaches Rmj :
P [ xi ∈ Xi ] = Fxi (Xi )
=
=
lim Fxi , xj (Xi , B̃j )
B̃j →Rmj
(21)
lim Fxi | xj (Xi | B̃j )Fxj (B̃j )
B̃j →Rmj
= Fxi (Xi )
lim Fxj (B̃j ) ,
B̃j →Rmj
where the last equality is due to the fact that xi and xj are conditionally independent.
To determine
lim Fxj (B̃j ), one can calculate:
B̃j →Rmj
lim Fxj (B̃j ) =
B̃j →Rmj
Z
p(xj )dxj
Rmj
Z
=
Z
p(xj )dxj +
R
mj
\B̃j
p(xj )dxj
(22)
B̃j
h
i
h
= P xj ∈
/ B̃j + P xj ∈ B̃j
i
= (1 − α̃j ) + α̃j
=1,
where p(xj ) is the PDF of xj , and the last equality is a direct result of Theorem 3. We now put all
the steps together as follows:
h
i
αi ≤ P xi ∈ Xi , xj ∈ B̃j
= Fxi , xj (Xi , B̃j )
≤ Fxi (Xi )
lim Fxj (B̃j )
B̃j →Rmj
= P [ xi ∈ Xi ]
Z
Z
p(xj )dxj +
Rmj \B̃j
≤
Z
p(xj )dxj + P [ xi ∈ Xi ]
Rmj \B̃j
p(xj )dxj
B̃j
Z
p(xj )dxj
B̃j
= (1 − α̃j ) + α̃j P [ xi ∈ Xi ] ,
where the first inequality and equality is due to (20), the second inequality is due to (21), the second
and
h last equality is due
i to (22), and the last inequality is considering the worst-case situation, e.g.
P xi ∈ Xi xj ∈
/ B̃j = 1.
By rearranging the last equation in above result:
αi − (1 − α̃j )
1 − αi
=1−
= ᾱi ≤ P [ xi ∈ Xi ] .
α̃j
α̃j
(23)
16
V. ROSTAMPOUR AND T. KEVICZKY
This completes the proof of first part. We now need to show the effect of having more than one
neighboring agent. To this end, the most straightforward step, in order to extend the current results,
is to use the fact that all neighboring agents are independent from each other. We therefore can apply
the previous results for a new situation where the agent i with the probabilistic level of feasibility ᾱi
have another neighboring agent ν ∈ Ni with α̃ν the level of reliability of B̃ν . By using Equation (23),
we have the following relations for j, ν ∈ Ni
1−αi
1
−
1
−
α̃j
1 − ᾱi
1 − αi
1−
=1−
=1−
≤ P [ xi ∈ Xi ] .
(24)
α̃ν
α̃ν
α̃j α̃ν
By continuing the similar arguments for all neighboring agents, one can obtain ᾱi = 1 −
Q
P [ xi ∈ Xi ] such that α̃i = α̃1 · · · α̃j α̃ν · · · α̃|Ni | = j∈Ni (α̃j ). The proof is completed.
1−αi
α̃i
≤
Remark 13. Following the statement of Theorem 4, it is straightforward to observe that if for all
j ∈ Ni , α̃j → 1 then ᾱi → αi . This means that if the level of reliability of the neighboring information
is one, P xj ∈ B̃j : ∀j ∈ Ni = 1, then, the state feasibility of agent i will have the same probabilistic
level of robustness as the hard communication scheme, P xi ∈ Xi ≥ αi .
Remark 14. Following the results in Theorem 4 and taking into account the statement of Theorem 2,
the proposed soft communication scheme introduces some level of stochasticity on the feasibility of the
large-scale system as in (5c). In particular, εi ∈ (0, 1) the level of constraint violation in each agent
Q
i ∈ M will increase, since it is proportional with the inverse of j∈Ni (α̃j ) ∈ (0, 1), and therefore,
P
ε = i∈M εi ∈ (0, 1) will also increase.
We summarize this section by mentioning that Figure 3 depicts a conceptual representation of
the proposed soft communication scheme between two neighborhood agents. In the next section,
we provide an operational framework that uses our developments in preceding sections in a more
applied framework namely plug-and-play (PnP) operations.
5. Plug-and-Play Operational Framework
This section presents a detailed steps of the PnP framework of our proposed approach. It is
important to note that the PnP operations are implemented in an offline mode, similar to [13]. We
first provide an algorithm to design the distributed scenario MPC for a large-scale dynamical system
(1). We then develop two algorithms for the plug-in and plug-out operations of each agent i ∈ M,
respectively.
Algorithm 1 summarizes our proposed approach to distributed scenario MPC. It is assumed here
that agents communicate with each other by using our proposed scheme in Section 4, namely the soft
inter-agent communication scheme. We however note that in case of the hard communication scheme,
each agent needs to generate Nsi samples and send exactly all the samples to all the neighboring
agent j ∈ Ni . In another words, the following changes have to be made in Algorithm 1. Ñsi will be
substituted by Nsi in Step 7 and Step 8 will be removed. Steps 9 and 10 will send and receive the
exactly Nsi samples, respectively, and Step 11 will be also removed.
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
17
Algorithm 1 Distributed Scenario MPC
Decompose the large-scale dynamical system (1) into M agents as the proposed form in (9)
2: Determine the index set of neighboring agents Ni for each agent i ∈ M
3: For each agent i ∈ M do
4:
fix xi,k = xi,0 ∈ Xi , εi ∈ (0, 1), and βi ∈ (0, 1) such that
X
X
βi ∈ (0, 1)
(25)
εi ∈ (0, 1) , β =
ε=
1:
i∈M
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
i∈M
determine Ns̄i ∈ (0, +∞) to approximate the objective function as in (6), and Nsi following
Theorem 2 to approximate the chance constraint (11) in an equivalent sense
generate Ns̄i , Nsi scenarios of wi , δ to determine the sets of S̄wi , S̄δ and Swi , Sδ
generate Ñsi scenarios of xi using the dynamical system of agent i in form of (9)
determine set B̃i by solving the optimization problem in (14)
send the set B̃i to all neighboring agents j ∈ Ni
receive the set B̃j from all neighboring agents j ∈ Ni
generate the samples of xj by drawing them uniformly from B̃j , ∀j ∈ Ni
solve the proposed optimization problem in (7) and determine a feasible solution v ∗i
∗ into the uncertain subsystem (9)
apply the first input of solution ui,k = Ki xi,k + vi,k
measure the state and substitute it as the initial state of the next step xi,0
k ← k + 1.
goto Step (6)
End for
Remark 15. In Algorithm 1 it is assumed that the feedback control gain matrices Ki for all agent
i ∈ M are given, following our discussion in (4). In addition, the coupling terms Aij (δ) are assumed
to be known between each agent i ∈ M and its neighboring agents j ∈ Ni .
Algorithm 2 and Algorithm 3 are presented to summarize the steps that are needed for plug-in
and plug-out operations, respectively.
Algorithm 2 Plug-in Operation
Add the number of new subsystems into the previous number of subsystems, e.g. one additional
agent M ← M + 1 such that |M| = M + 1
2: Update the index set of neighboring agents Ni for each agent i ∈ M
3: Goto Step (3) of Algorithm 1
1:
Remark 16. When an agent is plugged-in or plugged-out as in Algorithm 2 and Algorithm 3, respectively, one can redesign Ki to may improve the local control performance of each agent i ∈ M.
Remark 17. In a plugged-in or plugged-out operation as in Algorithm 2 and Algorithm 3, respectively, all agents i ∈ M have to update their εi with βi to respect the condition in Theorem 2 as it is
shown in (25) to achieve the desired level of constraint feasibility for the large-scale system (1).
18
V. ROSTAMPOUR AND T. KEVICZKY
Algorithm 3 Plug-out Operation
Remove the number of new subsystems from the previous number of subsystems, e.g. one
exclusion agent M ← M − 1 such that |M| = M − 1
2: Update the index set of neighboring agents Ni for each agent i ∈ M
3: Goto Step (3) of Algorithm 1
1:
6. Numerical Study
This section presents a simulation study to illustrate the functionality of our proposed framework.
In order to numerically investigate our proposed approach, we simulate the thermal energy demands
of three buildings modeled using realistic parameters and the actual registered weather data in the
city center of Utrecht, the Netherlands, where these buildings are located and they had been equipped
with aquifer thermal energy storage (ATES) systems [33].
6.1. Simulation Setup
A less well-known sustainable energy storage technology is ATES which is used to store large
quantities of thermal energy in aquifers enabling the reduction of energy usage and CO2 emissions
of the heating and cooling networks in buildings [34]. An ATES system consists of two wells (warm
and cold water wells) and it is considered as a heat source or sink, or as a storage for thermal energy
that operates in a seasonal mode.
Following the developed model in [33], each ATES system of each building is modeled using firstorder difference equations, leading a large-scale network of interconnected agents (buildings) that
are dynamically coupled via the state variables of ATES systems, which are the volume of water and
the thermal energy content of each well. In order to prevent overlap of nearby systems, there are
h [m], r c [m], of each well of an ATES system with its
constraints on a growing thermal radius, ri,k
i,k
h + r c ≤ d for each agent i ∈ M and its neighboring agent j ∈ N .
neighboring agents, e.g. ri,k
ij
i
j,k
In [33, Corollary 1], it was shown that such a constraint on the growing thermal radius can be
replaced with a constraint on the dynamically coupled state variables (volume of water) of the ATES
systems. We refer interested readers to [33] for the detailed explanations on the case study.
We here take some steps forward to model a more realistic ATES system by introducing additive
disturbance and multiplicative uncertainty sources into the deterministic dynamical model of [33,34].
It has been reported in [35, 36] that the ambient temperature of water in aquifer is changing by the
long life-time usage of ATES systems. We capture such a behavior by introducing an additive
unknown (disturbance) source of energy which yields a time correlation function via the initial value
of energy content variable of an ATES system. In addition to this, the system parameters of an
ATES system are a function of the stored water temperature in each well, e.g. see [33, Figure 2].
We therefore introduce a small level of perturbation as an uncertainty source in the parameters of
the ATES system dynamics.
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
(a)
100
19
c
h
≤ 65
+ r2,k
r1,k
50
Distance [m]
0
Mar/11 Apr May Jun
100
DeSMPC
CSMPC
DSMPC
Jul Aug Sep Oct Nov Dec Jan/12
Feb Mar
DSMPCS-0.85
c
h
(b) r2,k + r3,k ≤ 70
DSMPCS-0.50
50
0
Mar/11 Apr May Jun
Jul
Aug Sep Oct Nov Dec Jan/12 Feb Mar
(c)
100
c
h
≤ 70
+ r1,k
r3,k
50
0
Mar/11 Apr May Jun
Jul
Aug Sep Oct Nov Dec Jan/12 Feb Mar
Time
Figure 4. A-posteriori feasibility validation of the obtained results. The "red",
"blue", "magenta", "solid green", and "dashed green" lines are related to the obtained results via DeSMPC, CSMPC, DSMPC, DSMPCS−0.85, and DSMPCS−0.50,
respectively. The "dotted black" lines are related to the upper bound values.
To generate random scenarios from the additive disturbance, we built a discrete normal process
such that the nominal scenario is 10% of the amplitude of the energy content in a deterministic
ATES system model and varies within 5% of its nominal scenario at each sampling time. As for
the multiplicative uncertainty, we generate a random variable from a Gaussian distribution with a
mean value 0, variance 0.3 and a maximal magnitude of 0.03 at each sampling time. The simulation
environment was MATLAB with YALMIP as the interface [37] and Gurobi as the solver.
6.2. Simulation Results
Simulation results are delivered using a hierarchy control scheme in which the upper layer is used
to coordinate between ATES systems to meet the dynamically coupled constraint in a weekly-based
time steps with a three months prediction horizon, whereas in the lower layer, we solved a SMPC
for individual buildings to keep track of the desired temperature of each building using the proposed
formulation in [33, Problem 11]. We here focus on the results obtained in the upper layer, since the
lower layer problems are decoupled and the results were similar to the one in [33, Figure 6].
To illustrate the advantages of our proposed framework, we simulate four problem formulations,
namely a centralized SMPC (CSMPC) using the formulation in (7), a distributed SMPC (DSMPC)
via the distributed scenario program in (10), and DSMPC with the proposed soft communication
scheme with 0.85−reliability (DSMPCS−0.85) as described in Definition 1 and DSMPCS−0.50, both
20
V. ROSTAMPOUR AND T. KEVICZKY
(a)
c
h
≤ 65
+ r2,k
r1,k
(b)
c
h
≤ 70
+ r3,k
r2,k
70
60
Distance [m]
Jun/11
Jul
75
Aug
70
65
12/04/11 12/11 12/18 12/25 01/01/12 01/08 01/15 01/22 01/29 02/05
(c)
80
c
h
≤ 70
+ r1,k
r3,k
70
60
May/11
Jun
Jul
Time
Figure 5. Zoom-in of the critical time periods in Fig. 4.
following Algorithm 1. For comparison purposes, we also present the results obtained via decoupled
SMPC (DeSMPC), where the coupling constraints are relaxed. The simulation study is carried
out for one year using actual weather conditions from March 2011-March 2012 and the proposed
formulations in a closed-loop control system framework.
In Fig. 4 we show an a-posteriori feasibility validation of the dynamically coupled constraints
c
h
h + rc
between each agent i = 1, 2, 3, and neighboring agents, e.g. r1,k
2,k ≤ 65, r2,k + r3,k ≤ 70,
h + rc
and r3,k
1,k ≤ 70. The "red" line is related to the results obtained via DeSMPC, the "blue"
line shows the results obtained via CSMPC, the "magenta" presents the results obtained by using
DSMPC, the "solid green" and "dashed green" lines show the results obtained via DSMPCS−0.85
and DSMPCS−0.50, respectively. The "dotted black" lines are related to the upper bound values
of these three coupling constraints. Fig. 5 focuses on the critical time periods in Fig. 4, where both
neighboring agents are injecting thermal energies with different pump flow rates. Fig. 5(a) shows
the results from mid-May to mid-August 2011, Fig. 5(b) presents the results of December 2011 to
February 2012, and Fig. 5(c) depicts the results of mid-April to mid-July 2011.
As a first desired achievement using our proposed frameworks, one can clearly see in Fig. 4 the
dynamically coupled states trajectories are feasible in a probabilistic sense, since the agent operations
are quite close to the upper bounds in the critical time periods compared to DeSMPC that violates
the constraints. Strictly speaking, using our proposed framework one can achieve the maximum usage
of the aquifer (subsurface) to store thermal energy without affecting the neighboring thermal storage.
This is a direct result of Theorem 1, such that the obtained solutions via our proposed formulations
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
21
have to be probabilistically feasible, that can be clearly seen in Fig. 5, since the trajectories are on
the upper bounds almost surely.
It is worth to mention that Fig. 5 illustrates our other two main contributions more precisely: 1)
the obtained results via CSMPC (blue line) and DSMPC (magenta line) are practically equivalent
throughout the simulation; this is due to Proposition 1 and Theorem 2. Actually, the obtained
solutions via DSMPC are slightly more conservative compared to the obtained results via CSMPC,
and this is a direct consequence of Theorem 2. In fact the level of violation in CSMPC is considered
to be ε = 0.05 and leading to εi = 0.0167 for all agents due to Theorem 2, which is more restrictive.
2) using the proposed soft communication scheme yields less conservative solutions as explicitly
derived in Theorem 4, and can be clearly seen in Fig. 5 with the obtained results via DSMPCS−0.85
(solid green) and DSMPCS−0.50 (dashed green). Following Theorem 4 the new violation level using
DSMPCS−0.85 is ε̄i = 0.0231, and using DSMPCS−0.50 is ε̄i = 0.0668. It is important to notice
that the violation level of global chance constraint will increase to ε̄ = 0.0702 and ε̄ = 0.2004 using
DSMPCS−0.85 and DSMPCS−0.50, respectively, following Remark 14.
7. Concluding Remarks and Future Work
In this paper we presented a rigorous approach to distributed stochastic model predictive control
(SMPC) using the scenario-based approximation. We then provided a novel inter-agent soft communication scheme to minimize the amount of information exchange between each subsystem. Using a
set-based parametrization technique, we introduced a probabilistically reliable notation and quantified the robustness of the obtained distributed SMPC solution via the integrated soft communication
scheme. The probabilistic guarantees of the proposed distributed SMPC framework coincide with
its centralized counterpart.
Our current research direction concentrates on extending the proposed framework to cope with
the case of a common uncertainty source. It is important to point out that the discussion on
scenario optimization was limited to constraint satisfaction probability. An interesting future work
is to analyze the performance index by means of optimality of the scenario program solutions in
distributed setting by exploiting the results in [38]. Finally, another interesting future research
direction is to investigate the possibility of extending our proposed frameworks to the fault detection
and isolation problems following the results in [39, 40].
References
[1] B. Kouvaritakis, M. Cannon, and V. Tsachouridis, “Recent developments in stochastic mpc and sustainable
development,” Annual Reviews in Control, vol. 28, no. 1, pp. 23–35, 2004.
[2] A. Mesbah, “Stochastic model predictive control: An overview and perspectives for future research,” IEEE Control
Systems, vol. 36, no. 6, pp. 30–44, 2016.
[3] J. A. Primbs, “A soft constraint approach to stochastic receding horizon control,” in Conference on Decision and
Control. IEEE, 2007, pp. 4797–4802.
[4] A. Bemporad and M. Morari, “Robust model predictive control: A survey,” in Robustness in identification and
control. Springer, 1999, pp. 207–226.
22
V. ROSTAMPOUR AND T. KEVICZKY
[5] A. N. Venkat, I. A. Hiskens, J. B. Rawlings, and S. J. Wright, “Distributed mpc strategies with application to
power system automatic generation control,” IEEE transactions on control systems technology, vol. 16, no. 6, pp.
1192–1206, 2008.
[6] V. Rostampour, P. Mohajerin Esfahani, and T. Keviczky, “Stochastic nonlinear model predictive control of an
uncertain batch polymerization reactor,” IFAC Conference on Nonlinear Model Predictive Control (NMPC),
vol. 48, no. 23, pp. 540–545, 2015.
[7] B. E. Ydstie, “New vistas for process control: Integrating physics and communication networks,” AIChE Journal,
vol. 48, no. 3, pp. 422–426, 2002.
[8] Y. Ma, S. Richter, and F. Borrelli, “Dmpc for building temperature regulation,” Control and Optimization with
Differential-Algebraic Constraints, vol. 23, p. 293, 2012.
[9] J. Lavaei and A. G. Aghdam, “Control of continuous-time lti systems by means of structurally constrained
controllers,” Automatica, vol. 44, no. 1, pp. 141–148, 2008.
[10] A. Richards and J. P. How, “Robust distributed model predictive control,” International Journal of control, vol. 80,
no. 9, pp. 1517–1531, 2007.
[11] C. Conte, M. N. Zeilinger, M. Morari, and C. N. Jones, “Robust distributed model predictive control of linear
systems,” in European Control Conference (ECC). IEEE, 2013, pp. 2764–2769.
[12] M. Cannon, B. Kouvaritakis, S. V. Rakovic, and Q. Cheng, “Stochastic tubes in model predictive control with
probabilistic constraints,” IEEE Transactions on Automatic Control, vol. 56, no. 1, pp. 194–200, 2011.
[13] S. Riverso, M. Farina, and G. Ferrari-Trecate, “Plug-and-play decentralized model predictive control for linear
systems,” IEEE Transactions on Automatic Control, vol. 58, no. 10, pp. 2608–2614, 2013.
[14] L. Dai, Y. Xia, Y. Gao, and M. Cannon, “Distributed stochastic mpc of linear systems with additive uncertainty
and coupled probabilistic constraints,” IEEE Transactions on Automatic Control, vol. PP, no. 99, pp. 1–8, 2016.
[15] P. Hokayem, E. Cinquemani, D. Chatterjee, F. Ramponi, and J. Lygeros, “Stochastic receding horizon control
with output feedback and bounded controls,” Automatica, vol. 48, no. 1, pp. 77–88, 2012.
[16] M. Prandini, S. Garatti, and J. Lygeros, “A randomized approach to stochastic model predictive control,” in
Conference on Decision and Control. IEEE, 2012, pp. 7315–7320.
[17] G. C. Calafiore and L. Fagiano, “Stochastic model predictive control of lpv systems via scenario optimization,”
Automatica, vol. 49, no. 6, pp. 1861–1866, 2013.
[18] G. Schildbach and M. Morari, “Scenario mpc for linear time-varying systems with individual chance constraints,”
in American Control conference (ACC), 2015. IEEE, 2015, pp. 415–421.
[19] M. Lorenzen, F. Allgöwer, F. Dabbene, and R. Tempo, “Scenario-based stochastic mpc with guaranteed recursive
feasibility,” in Conference on Decision and Control (CDC). IEEE, 2015, pp. 4958–4963.
[20] K. Margellos, P. Goulart, and J. Lygeros, “On the road between robust optimization and the scenario approach
for chance constrained optimization problems,” Trans. on Automatic Control, vol. 59, no. 8, pp. 2258–2263, 2014.
[21] X. Zhang, K. Margellos, P. Goulart, and J. Lygeros, “Stochastic model predictive control using a combination of
randomized and robust optimization,” in Conference on Decision and Control. IEEE, 2013, pp. 7740–7745.
[22] V. Rostampour and T. Keviczky, “Robust randomized model predictive control for energy balance in smart
thermal grids,” in European Control Conference (ECC). IEEE, 2016, pp. 1201–1208.
[23] G. C. Calafiore and M. C. Campi, “The scenario approach to robust control design,” Transactions on Automatic
Control, vol. 51, no. 5, pp. 742–753, 2006.
[24] M. C. Campi and S. Garatti, “The exact feasibility of randomized solutions of uncertain convex programs,” SIAM
Journal on Optimization, vol. 19, no. 3, pp. 1211–1230, 2008.
[25] Y. Long, S. Liu, L. Xie, and K. H. Johansson, “A scenario-based distributed stochastic mpc for building temperature regulation,” in Automation Science and Engineering (CASE), 2014 IEEE International Conference on.
IEEE, 2014, pp. 1091–1096.
[26] S. Liu, A. Sadowska, H. Hellendoorn, and B. De Schutter, “Scenario-based distributed model predictive control
for freeway networks,” in International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2016,
pp. 1779–1784.
[27] K. Margellos, A. Falsone, S. Garatti, and M. Prandini, “Distributed constrained optimization and consensus in
uncertain networks via proximal minimization,” arXiv preprint arXiv:1603.02239, 2016.
DISTRIBUTED SMPC FOR LARGE-SCALE UNCERTAIN LINEAR SYSTEMS
23
[28] G. C. Calafiore and D. Lyons, “Random convex programs for distributed multi-agent consensus,” in European
Control Conference (ECC). IEEE, 2013, pp. 250–255.
[29] M. C. Campi, S. Garatti, and F. A. Ramponi, “Non-convex scenario optimization with application to system
identification,” in Conference on Decision and Control (CDC). IEEE, 2015, pp. 4023–4028.
[30] G. Schildbach, L. Fagiano, C. Frei, and M. Morari, “The scenario approach for stochastic model predictive control
with bounds on closed-loop constraint violations,” Automatica, vol. 50, no. 12, pp. 3009–3018, 2014.
[31] T. Alamo, R. Tempo, A. Luque, and D. R. Ramirez, “Randomized methods for design of uncertain systems:
Sample complexity and sequential algorithms,” Automatica, vol. 52, pp. 160–172, Feb. 2015.
[32] J. Lunze, Feedback control of large-scale systems. Prentice-Hall London, 1992.
[33] V. Rostampour and T. Keviczky, “Probabilistic energy management for building climate comfort in smart
thermal grids with seasonal storage systems,” International Federation of Automatic Control (IFAC) World
Congress, 2017. [Online]. Available: https://arxiv.org/pdf/1611.03206.pdf
[34] V. Rostampour, M. Jaxa-Rozen, M. Bloemendal, and T. Keviczky, “Building climate energy management in smart
thermal grids via aquifer thermal energy storage systems,” Energy Procedia, vol. 97, pp. 59–66, 2016.
[35] M. Jaxa-Rozen, M. Bloemendal, V. Rostampour, and J. Kwakkel, “Assessing the sustainable application of aquifer
thermal energy storage,” in European Geothermal Congress, 2016.
[36] V. Rostampour, M. Bloemendal, M. Jaxa-Rozen, and T. Keviczky, “A control-oriented model for combined
building climate comfort and aquifer thermal energy storage system,” in European Geothermal Congress, 2016.
[37] J. Löfberg, “Yalmip: A toolbox for modeling and optimization in matlab,” in Computer Aided Control Systems
Design, International Symposium on. IEEE, 2004, pp. 284–289.
[38] P. Mohajerin Esfahani, T. Sutter, and J. Lygeros, “Performance bounds for the scenario approach and an extension
to a class of Non-Convex programs,” IEEE Trans. Automat. Contr., vol. 60, no. 1, pp. 46–58, Jan. 2015.
[39] P. M. Esfahani and J. Lygeros, “A tractable fault detection and isolation approach for nonlinear systems with
probabilistic performance,” IEEE Transactions on Automatic Control, vol. 61, no. 3, pp. 633–647, 2016.
[40] V. Rostampour, R. Ferrari, and T. Keviczky, “A set based probabilistic approach to threshold design for optimal
fault detection,” to appear in American Control Conference (ACC), 2017.
| 3cs.SY
|
Distributable Consistent Multi-Graph Matching
Nan Hu
Qixing Huang
Boris Thibert
Stanford University
Stanford, CA, USA
Leonidas Guibas
arXiv:1611.07191v2 [cs.DS] 22 Feb 2018
[email protected], [email protected], [email protected], [email protected]
Abstract
our setting). Recent works on joint matching has shown
that the cycle consistency constraint can be translated into
a much more manageable constraint, i.e., the data matrix
that stores pair-wise maps in blocks is positive semidefinite
and low-rank [23, 15]. Based on connection, people have
formulated multiple-object matching as solving semidefinite programs (or SDP), which are convex relaxations of the
corresponding matrix recovery problem. These algorithms
achieved near-optimal exact recovery conditions [15, 4]. On
the other hand, solving semidefinite programs are computationally intensive. In a recent work, Zhou et al [33] attempts
to address the computational issue using alternating minimization, a fast solver for solving semidefinite programs.
In this paper, we propose a novel framework that utilizes the cycle-consistency constraint in a hierarchical manner for scalable multiple object matching. We show how
to apply this framework to extend data matrix based lowrank recovery algorithms in [15, 33]. In particular, instead
of jointly imposing the global consistency on all pair-wise
maps [15, 4, 33], we split the collection of graphs into overlapping subsets, and impose consistency within each subset.
We then impose consistency between maps across the subsets. Interestingly, we show that by combing these two consistency constraints together, we can guarantee global consistency under mild conditions (See Section 2). Yet computationally, such a decoupled approach yields significant performance gains, when compared with existing approaches.
In this paper we propose an optimization-based framework to multiple graph matching. The framework takes as
input maps computed between pairs of graphs, and outputs maps that 1) are consistent among all pairs of graphs,
and 2) preserve edge connectivity between pairs of graphs.
The central idea of our approach is to divide the input
graph into overlapping sub-graphs and enforce consistency
among sub-graphs. This leads to a distributed formulation,
which is scalable to large-scale datasets. We also present
an equivalence condition between this decoupled scheme
and the original scheme. Experiments on both synthetic
and real-world datasets show that our framework is competent against state-of-the-art global optimization-based techniques.
1. Introduction
Graph matching techniques have been widely used in
many fields of computer vision, including 2D and 3D image analysis, object recognition, biomedical identification,
and object tracking. Tremendous amount of effort has
been taken to find meaningful approximate isomorphism
between pair of graphs [25, 6, 9, 10, 27, 8, 13, 14]. Many
tasks, however, require solving the so-called multi-graph
matching problem, i.e., finding consistent maps among all
pairs of graphs within a collection. Examples include nonrigid structure from motion [1, 7] and shared object discovery [3]. In this context, a central task is how to utilize the
data collection as a regularizer to improve the maps computed between pairs of objects in isolation [15, 4, 33].
An generic constraint that one can utilize to improve
maps among a collection is the so-called cycle consistency
constraint, namely composition of maps along any two
paths sharing the same starting and end objects are identical. A technical challenge of utilizing this constraint is
that it is impossible to check all cycles for consistency due
to the fact that the number of paths increase exponentially
with the total number of objects (each object is a graph in
1.1. Related Work
Early works on multi-graph matching, e.g. [32, 17]
extend pairwise matching schemes to multi-graph setting
without explicitly considering consistency issue. Later on,
methods are proposed based on detecting inconsistent cycles as in [31, 21], and formulate multi-graph matching
as solving combinatorial optimizations, i.e., removing bad
maps to break all inconsistent cycles. On the other hand,
people have proposed to formulate non-convex optimization
problems by using the cycle consistency as an explicit constraint for either pixel-wise flow computation [33], sparse
feature matching [28], sparse shape modeling [5], or struc1
Gi
Gi
Gij
Gj
Gj
ture from motion [22]. These problems are, as a consequence, hard to solve and do not admit exact recovery conditions. Recent works [16, 15, 23] showed that consistent
maps could be extracted from the spectrum of a data matrix that encodes pair-wise maps in blocks. Along this line
of research, Huang and Guibas [15] proposed an elegant
solution by formulating the problem as convex relaxation
and discussed the theoretical conditions for exact recovery.
The result is further analyzed in [4] under the condition that
the underlying rank of the variable matrix is known or can
be reliably estimated. Yan et. al. [30, 29] also proposed
matrix factorization based methods to solve the consistency
problem. These methods, however, are not scalable to largescale datasets, due to the cost of solving semidefinite programs. Zhou et al. [33] enforce the positive semidefinite
constraint using explicit low-rank factorizations, leading
to improved computational efficiency. In contrast to these
methods, our approach opens a new direction to enforce the
cycle-consistency constraint, i.e., by splitting the datasets
into overlapping subsets. This leads to further improvements in terms of computational efficiency. Most recently,
Leonardos et. al. [18] proposed a distributed consensusbased algorithm as an extension of [23]. Their method,
however, cannot handle partial matches, which limited the
applicability to practical problems.
We organize the reminder of this paper as follows.
Firstly, we discuss the problem setup and analyze the conditions In Section 2. Secondly, we discuss the formulation of our approach In Section 3. In the following Section 4, we present an alternating direction method of multipliers (ADMM) for solving the induced optimization problem, leading to a parallel algorithm via generalized message
passing. Last but not the least, we demonstrate the effectiveness of our approach on both synthetic and real examples in
Section 5.
2. Consistency
In this section, we extend the cycle-consistency formulation described in [15] to the distributed setting. The
key result is a sufficient condition so that cycle-consistency
among sub-graphs induces global cycle-consistency.
We begin with introducing the notations that are necessary to formally state this sufficient condition. For simplicity, we assume maps between objects are given by permutations. However, the argument can be easily extended to
the case where objects are partially similar with each other.
Formally speaking, we consider a map graph G = (V =
(H1 , · · · , Hn ), E). The vertex set V consists of objects to
be matched, and each object Hi is given by m points (e.g.,
key points extracted from an image). The edgeset connects
a subset of pairs of objects. Each edge (i, j) ∈ E is associated with a permutation φij : Hi → Hj . We first define the
global consistency of φij , ∀(i, j) ∈ E:
violation edge
Gij
Gi Gi
Gj
(a) joint normal
Gj
Gi
Gij
Gj
(b) Counter joint normal
Figure 1: Examples of subgraphs that are (a) joint normal
and (b) not joint normal.
G
Definition 1 (Cycle Consistency). A map graph G = ij
{V, E} is cycle consistent if for every node vi and ievery
cycle vi − vi1 − · · · − vik − vi , the composite map along
this cycle is the identity map, i.e.
G
φii1 ◦ · · · ◦ φik i = identity.
Now we introduce an equivalent formulation of cycleconsistency with respect to subgraphs of G under proper
conditions of these subgraphs. Towards this end, we introduce two conditions among collection of subgraphs of G.
The first condition concerns a pair of sub-graphs:
Definition 2 (Joint Normal). Let Gi = {Vi , Ei }, Gj =
{Vj , Ej } be the two subgraphs of G = {V, E}. We say Gi
and Gj are joint normal if the vertex sets Vi \ Vj and Vj \ Vi
are not connected by any edge of E:
(s, t) ∈
/ E, ∀Hs ∈ Vi \Vj , Ht ∈ Vj \Vi
As illustrated in Fig. 1, two subgraphs are joint normal if their intersection subgraph is either (i) empty, or (ii)
connected and there is no edge between a vertex in one subgraph to a vertex in the other except possibly those in their
intersection subgraph. In contrast, the two subgraphs illustrated in Fig. 1 (b) are not joint normal since there exists
an edge that connects the non-overlapping sets of the two
subgraphs.
The second condition concerns a topological constraint
among all the sub-graphs. We state this condition using the
notation of simplicial complex as detailed below:
Definition 3 (Cover Complex). Let {Gi = (Vi , Ei ), 1 ≤
i ≤ K} be a set of sub-graphs that cover G, i.e., ∪K
i=1 Vi =
V. We define the cover complex K of these sub-graphs
{Gi = (Vi , Ei ), 1 ≤ i ≤ K} so that K collects every
subset (i1 , · · · , ik } ⊂ {1, · · · , K} if the intersections of
Vi1 , · · · , Vik is non-empty, i.e. ∩j Vij 6= ∅.
We now state the cycle consistent matching theorem that
relates the global cycle consistency and the cycle consistency on each sub-graph:
Gj
G
Gij
Theorem 1 (Cycle Consistent Matching). Let G be a map
i
j
graph, G1 , · · · , GK be a cover of G, and K be the cover
complex. Then, a matching on G is cycle consistent if
G
1. Gi is cycle consistent ∀i,
2. Gi and Gj are joint normal ∀(i, j) ∈ E,
3. K is simply connected.
Intuitively, the cover complex K is simply connected if
all cycles are triangulated, c.f. [12] for a more general definition in terms of topological space. This theorem states
that the cycle consistency property on each sub-graph would
be propagated to the global consistency if the cover comG
plex iK satisfies theGabove-mentioned
conditions. The proof
j
to Theorem 1 is left to the supplementary material.
Note that the 3nd condition in Theorem 1 is necessary.
2(a) provides a simple counter example, which satisfies the
1st and 2nd conditions in Theorem 1. The cover complex
K, however, is homologous to the Torus T 2 , which is not
simply connected. It is easy to see that the local-consistency
Gij does
(which is trivial as each sub-graph is given by an edge)
Gi three edges.Gj
not lead to the global consistency among these
Figure 2(c) provides another example to understand the
correctness of Theorem 1. In this case, there are four objects. It is clear that the cycle-consistency among all four
triple sub-graphs induce the global consistency. This argument aligns with Theorem 1 as K is simply connected. We
defer detailed explanations to the supplementary material.
Algorithm 1: Greedy Construction of K
Input : Map graph G = (V, E)
Number of cover nodes K
Output: Cover nodes {Gi }K
1
1 Compute a graph clustering of G to K clusters (graph
cut or K-means on graph embeddings)
2 Assign each of the cluster to Gi , i = 1, · · · , K
3 while condition not satisfied do
4
Assign Hj to Gi if Hj is neighboring to Gi in G or
within distance of to Gi in the embedding space
5
Build K from {Gi }K
i
6
Check if K is connected
7
Compute homology group using [34]
8
Check if H1 (K) is trivial
9
if Both conditions satisfied then
10
Break
11
end
12 end
To develop an algorithm based on Theorem 1, we proposed a greedy algorithm to construct K as in Algorithm
1. Note that a complex K is simply connected, if it is connected. 1) it is connected; 2) the 1-dimensional homology
an empty triangle cover complex
G(a)i Subgraphs thatGforms
j
G
(b) Subgraphs that form an solid triangle cover complexij
Gi
(c) Subgraphs that form an empty tetrahedron cover complex
Figure 2: Given local consistency, (a) is not globally consistent, while (b) and (c) are guaranteed to be globally consistent. (Left: the map graph with subgraphs circled out.
Right: the corresponding cover complex.)
group H1 (K) is trivial. Condition 1) could be easily verified
by any graph traverse algorithm (BFS/DFS), and condition
2) can be verified computationally as in [34]. We defer a
rigorous treatment to the supplemental material.
3. Distributed Optimization
In this section, we introduce the proposed distributed formulation of recovering cycle-consistent maps from noisy
pair-wise maps.
3.1. Formulation
Our formulation takes as input the pairwise base maps
φij . We follow the state-of-the-art work on convex relaxation of maps [15, 4, 33] to encode φij into data matrix form Xij . Different from Section 2, we relax φij to
be partial correspondence and possibly be soft maps, i.e.
Xij ∈ [0, 1]mi ×mj , where mi denote the number of vertices in Hi .
Let XV be the matrix form of matching for all the objects
Gj
in V, i.e.
I
X21
XV =
.
..
Xn1
X12
···
I
···
..
.
···
···
···
X1n
..
.
..
.
I
The goal of our formulation is to find cycle consistent matching matrix XV from the noisy input XV . To
achieve this goal, one observation is that the desired matching matrix XV is intrinsically structured [15]. We assume
there’s a hidden universal entity space with dimensionality m, i.e. there are totally m distinct entities for all the
objects Hi ’s in V. For each object Hi , we have an implicit matching matrix of the form AHi ∈ {0, 1}mi ×m ,
mapping a vertex from Hi to an entity in the universal
space. Let AV be the vertical concatenation of all AHi ’s.
We know that the matching matrix XV could be written as
XV = AV A>
V , which indicates the intrinsic low-rankness
and positive-semidefiniteness. This insight also hints the
authors in [15, 33] to find matrix factorization ways to reconstruct XV with such consistency property.
W.l.o.g. assume h(XV ) is an objective function that existing matrix factorization based method is using. Our distributed framework could be conveniently written as
P
min
i h (XVi )
(1)
i
s.t. XVi∩j
= XV j , ∀(i, j) ∈ E,
i∩j
i
where XVi∩j
is the matching matrix of Vi∩j in Vi , i.e. a submatrix of XVi by picking blocks that belongs to the matching graphs in Vi∩j . Each h(XVi ) indicates local consistency
i
in Vi , and the condition that XVi∩j
= XV j will guarani∩j
tee that the overlapping subgraph are consistent. In such a
manner, the consistency condition will propagate through
the overlapping sub-graph to each component Vi conceptually similar to our proof of Theorem 1.
In the state-of-the-art research work of [15] and [33], the
authors proposed different formulations of objective function h(XV ). We will use Zhou’s formulation in [33] to
demonstrate the extension in our framework in the main
paper, because of its competent performance and superior
computational speed. We will leave in our supplementary
material for the derivation using Huang’s formulation in
[15].
As in [33], h(XV ) can be written as
min hWV , XV i + λkXV k∗
s.t. XV 0,
XV(ii) = Imi , ∀i
XV(ij) = X>
V(ji) , ∀i 6= j
0≤X≤1
of all 1’s. The purpose of adding constant α is to impose L1
constraint on XV for sparsity. We use XV(ij) to denote the
(i, j)th block of the block matrix XV . As has been shown
in [33], the constraint XV 0 could be relaxed if λ is
big enough. Let Ci encodes the convex set induced by the
constraints for Vi , we could then simplify the notation for
our distributed problem as
P
min
i (hWVi , XVi i + λkXVi k∗ )
s.t. XVi ∈ Ci
(3)
i
XVi∩j
= XV j , ∀(i, j) ∈ E
i∩j
4. Alternating minimization
4.1. Algorithms
The nuclear norm minimization in (3) can be efficiently
optimized thanks to recent results on low-rank optimization
by a change of variables XVi = AVi B>
Vi as in [2, 11, 33],
together with the guarantee in [24] for nuclear norm
kXk∗ =
Let XVi and XiVi∩j be shortened as Xi and Xij respectively. Moreover, let Eij denote the selection matrix to extract the part of Xi that belongs to the set of Vi ∩ Vj , i.e.
Xij = E>
ij Xi Eij . Therefore, the condition on intersection
>
consistency becomes E>
ij Xi Eij = Eji Xj Eji .
We then finalize our formation of the problem in (3) as
P
λ
λ
2
2
min
i hWi , Xi i + 2 kAi kF + 2 kBi kF
s.t. Xi = Ai B>
i ,
(4)
>
E>
ij Xi Eij = Eji Xj Eji ,
Xi ∈ Ci ,
We apply ADMM to (4). The solver is summarized in
Algorithm 2 (c.f. supplementary material for details). The
constraints on X’s in convex set are handled implicitly and
updated in the alternating algorithm. Here, Yi ’s and Zij ’s,
are dual variables. The ADMM algorithm therefore updates
primal variables by minimizing L and dual variables by gradient descent.
In all our experiments, we set α = 0.1, λ = 50, µ = 64,
and β = 1. Ai and Bi have a least square style close-form
update, and Xi0 is the solution to the linear equation
µXi +2β
(2)
where h·, ·i is the matrix inner product, k · k∗ is the matrix
nuclear norm, and WV = α1−XV , and 1 denote the matrix
1
kAk2F + kBk2F .
A,B:AB> =X 2
min
X
j
>
>
Eij E>
ij Xi Eij Eij = µAi Bi −(Wi +Yi )
+
X
j
Eij (2βMkj→i − Zij + Zji )E>
ij
Furthermore, the update on Xi requires a projection onto
the convex set C, PC (·), i.e. PC (X0 ) is the solution to the
Algorithm 2: Distributed Graph Matching via ADMM
1
2
3
5
6
7
8
λ †
Ai ← (Xi + µ1 Yi )Bi (B>
i Bi + µ I)
λ †
Bi ← (Xi + µ1 Yi )Ai (A>
i Ai + µ I)
Xi ← PCi (Xi0 )
Yi ← Yik + µ Xi − Ai B>
i
k+1
Zij ← Zkij + β Mk+1
i→j − Mj→i
/* inter-node information exchange
9
10
11
5.1. Simulation
Input : Pairwise basemaps Xi
Output: Consistent matches Xi
Initialize Ai , Bi randomly, and set Yi , Zij to be 0
Wi = α1 − Xi
while not converged do
/* inner-node update
4
5. Experiments
*/
In this section, we present the evaluation experiments using synthetic data.
We followed the same experimental setup as in [4, 33].
Given an optimized matching matrix X∗ and the ground
truth mapping Xg , we access the quality of X∗ by measuring the intersection over union (or IOU) score:
1−
*/
E>
ji Xj Eji
to node i
node j send Mj→i =
end
Quantize Xi with a threshold of 0.5.
where τ (·) denotes the mapping induced from the matching matrix, and | · | denotes the set size. Note in our distributed setting, we could only partially recover X∗ given
X∗i . Therefore, our ground truth setting is also different
from [4, 33] in this regard.
5.1.1
problem
min kX − X0 k2F .
X∈C
This is essentially a linear programming problem, and can
be solved efficiently.
The key point in Algorithm 2 is the separation of innernode update and inter-node information exchange. It can
be seen that all the matrix computations are done in each
node separately, which indicates a distributed computation.
While after each iteration, adjacent nodes will need to exchange information by passing messages. Namely, for node
Vj and all the neighboring nodes Vi ’s, a message will be
send in the format of Mj→i = E>
ji Xj Eji . Note there actually will be no computation for messages. Recall that Eji
is just a sub-block extraction matrix, and the way Mj→i
is computed is simply by extracting the sub-block of the
matching matrix Xj that belongs to the intersection Vi ∩Vj .
Since all the computations are indeed done on each node individually, the proposed algorithm is essentially completely
distributed, and the only add-on is a syncing stage.
4.2. Complexity
The computational complexity of Algorithm 2 is dominated by the matrix multiplication. In our proposed algorithm, the complexity per iteration is controlled
by the leading node in G, i.e. O maxi (mi )2 maxi (ki ) , where mi is
the total number of points in all graphs in Vi . This is in
comparison with the state-of-the-art algorithm on consistent graph matching
P as in [4, 33], where
P the complexity per
iteration is O(( i mi )3 ) and O(( i mi )2 k) respectively.
Furthermore, in our experiment, we found the total number
of iterations to converge for our algorithm is comparable as
with the global method in [33].
|τ (X∗ ) ∩ τ (Xg )|
|τ (X∗ ) ∪ τ (Xg )|
Matching Errors
In the first experiment, we targeted to evaluate the matching performance between our algorithm, DMatch, and the
global algorithm, MatchALS, as in [33]. We considered the
following model to generate the testing examples. The total
number of graphs is denoted by n. The size of the universe
is fixed at r = 20 points. In each graph, a point is randomly
observed with a probability ρ0 . We simulated error corruption by randomly removing true mapping and adding false
ones with a corruption rate ρe .
We considered two settings in our experiments. In the
first setting, we constructed our cover graph by making a
sparse three way tree. This was done by randomly selecting
a subset of V as a common intersection Vc and then split
the rest evenly into the three cover nodes V10 , V20 , V30 . As a
consequence, each cover node is equal to Vi = Vc ∪ Vi0 .
In the second setting, we increased the overlap density by
circularly adding one more split to each cover, i.e. Vi =
0
Vc ∪ Vi0 ∪ Vi+1
. We compared DMatch to MatchALS by
varying the parameters ρ0 , ρe , and n. For both algorithms,
we used k = 2r and λ = 50.
Figure 3 shows matching errors under various configurations, for both DMatch and MatchALS. In general, lowering
input error and increasing observation ratio or increasing
the total number of objects will improve the matching performance, i.e. with a lower matching error. In addition, we
could see that increasing the overlap between cover nodes
would have a positive impact on the recovery (comparison
between Figure 3(a) and Figure 3(b)).
Furthermore, in a comparison between Figure 3(b) and
Figure 3(c), it could be seen that when the cover is dense
enough, i.e. the size of the overlaps are sufficiently large,
the matching error would approach that from MatchALS,
which is the global recovery.
20
30
30
40
50
60
70
80
number of objects n
20
30
number of objects n
number of objects n
20
40
50
60
70
0.4
0.6
50
60
70
80
80
0.2
40
0.8
0.2
input error rate ;e
0.4
0.6
0.2
0.8
0.4
DMatch-sparse
0.6
0.8
input error rate ;e
input error rate ;e
MatchALS
DMatch-dense
0.2
0.2
0.3
0.3
0.3
0.4
0.5
0.6
0.7
0.8
observation ratio ;0
0.2
observation ratio ;0
observation ratio ;0
(a) number of objects n + input error rate ρe
0.4
0.5
0.6
0.7
0.4
0.6
0.8
0.2
input error rate ;e
5.2.1
0.5
0.4
0.6
0.7
0.2
0.8
0.4
0.6
0.8
input error rate ;e
MatchALS
DMatch-dense
(b) observation ratio ρ0 + input error rate ρe
Figure 3: Matching error comparison. Darker color means
lower matching error. The fixed parameter is set to ρ0 = 0.6
and n = 50 respectively.
5.1.2
Graph Covers
In the second experiment, we aimed to understand more on
the effects of graph covers. In comparison, we construct
a ground truth graph cover by selecting a sparse cover as
in Section 5.1.1. For every pair of graphs within the same
cover node, we set the error rate to be ρin , and for every
pair between different cover node, we set the error rate to
be ρout . The experiment is then conducted by comparing
DMatch to 1) using the ground truth cover and 2) using a
randomly constructed cover.
0.3
0.5
0.6
0.7
0.8
0.9
0.4
0.5
0.6
0.7
0.8
0.6
0.8
in-node error rate ;in
sparse
0.4
0.5
0.6
0.7
0.8
0.9
0.9
0.4
0.3
out-node error rate ;out
0.4
out-node error rate ;out
0.3
out-node error rate ;out
out-node error rate ;out
0.3
0.4
0.6
0.8
in-node error rate ;in
dense
(a) ground truth cover
0.4
0.5
0.6
0.7
0.8
0.9
0.4
0.6
0.8
in-node error rate ;in
sparse
CMU House Sequence
0.6
input error rate ;e
DMatch-sparse
5.2. Real Experiments
0.4
0.8
0.8
0.2
on ρout disappears as shown in Figure 4(b). One potential
explanation is that in this case the portion of out-node pairs
become lesser with denser cover. As a consequence, the
mixed error rate is dominated by ρin . In addition, we could
also see that with denser cover, the algorithm is more error
tolerant. This comes with a trade-off that on average the
size of each cover becomes bigger and the computational
cost also increases.
0.4
0.6
0.8
in-node error rate ;in
dense
(b) randomly constructed cover
Figure 4: Effects on recovery from the construction of graph
cover. Darker color means lower matching error. The fixed
parameter is set to ρ0 = 0.6 and n = 50.
Figure 4 shows the experimental results. In Figure 4(a),
we can see that using the ground truth cover, the matching performance does not depend much on ρout , because
we explicitly disregarded any information from ρout . While
on the randomly constructed cover, the two error rates are
mixed up. Specifically, the results favor more on small ρin
and small ρout at the same time. This situation, however, is
changed when the cover becomes denser. The dependency
In this part of the experiment, we are intended to test the
scalability of our distributed algorithm. We use the CMU
House sequence1 as the testing images. This sequence has
been widely used to test different graph matching algorithms. It consists of 110 frames, and there are 30 feature
points labeled consistently across all frames. We used the
geometry based constraint in pairwise matching as is done
in [14]. To construct a valid cover complex K, firstly we
computed all pairwise matches and for each match, the result comes with a matching matrix and the corresponding
matching score. We built a fully connected matching quality graph, where each vertex represents an image, and the
edge weight represents the matching score. Since we knew
the sequence was roughly generated by moving a camera in
a single dimension, we computed the Laplacian embedding
of the vertices onto 1 dimensional space using Fielder vector, and then we applied Algorithm 1 to build the cover complex. We compared our DMatch algorithm with the global
methods, MatchALS [33], MatchLift [15], Spectral [23],
and another distributed method, the Consensus algorithm
[18]. Besides running global algorithms on fully connected
map graph, we also ran them on the sparse map graph induced by our cover complex. The algorithm was implemented in a single laptop, we measured the time used in
each cover node separately and then take the max as the total computational time per iteration, where we assumed the
cost for messages passed between adjacent cover nodes was
negligible.
Table 1 shows the matching accuracy, timing and iterations used in these algorithms. In our test, Spectral method
appeared to be the fastest method among all methods in experiments. On the other hand, it has the largest reported
error. The consensus algorithm is not error-driven, hence
we used a preset 150 iterations to match the lowest number
of iterations among all experiments. In general, distributed
algorithms used less iterations to converge, and achieved at
least an order of magnitude speed-up compared with global
methods, while maintaining an error rate of 0 (almost 0 for
Consensus algorithm with k = 6). It can also be seen that
1 http://vasc.ri.cmu.edu//idb/html/motion/house/
index.html
Original
MatchALS
MatchLift
Spectral
MatchALS (k = 4)
MatchALS (k = 6)
MatchLift (k = 4)
MatchLift (k = 6)
DMatch (k = 4)
DMatch (k = 6)
Consensus (k = 4)
Consensus (k = 6)
Error Rate
0.1445
0.1031
0.1027
0.1277
0.0161
0.0648
0
0
0
0
0
0.0071
Iterations
266
1000
380
1000
1000
1000
203
150
150
150
Time
98.8
3791.1
0.6
103.5
268.2
4066.1
3972.3
28.9
7.6
28.3
18.7
Table 1: The error rate and the total computational time
(seconds) on CMU House sequence.
for distributed algorithms once we increased the number of
clusters from k = 4 to k = 6, both the number of iterations
and total computational time decreased. This, on the other
hand, proved our complexity analysis, since increasing the
number of clusters would in general reduce the number of
vertices in each cluster. Although both DMatch and Consensus algorithms achieved similar results in terms of accuracy and time, the latter requires knowledge of the number
of universal entities and has limitations dealing with partial matches. Another interesting point is at convergence,
the error of MatchALS on the sparse map graph induced
by our cover complex decreased compared with that on the
fully connected map graph. One reason for this could be
the graph cover structure grouped together images that have
high pairwise matching quality and explicitly disregard any
pairwise matches that are of low quality (covers are joint
normal), and as a consequence, it is more noise-proof. In
addition, we also extended the experiments for k = 8, 10
and found the running time is not reduced significantly on
these values as compared from k = 4 to k = 6. Since
we need to have a reasonable amount of overlap between
clusters to pass matching information around, increasing the
number of clusters does not necessarily reduce the max size
of the clusters.
5.2.2
Graffiti datasets
In this part of experimentation, we followed the procedure
as described in [33]. We used the benchmark datasets from
Graffiti datasets 2 . In each dataset there are 6 images of a
scene with various image transformations, including viewpoint change, blurring, and illumination variation etc.
To construct an affinity score matrix X, we employed
2 http://www.robots.ox.ac.uk/
html
˜vgg/data/data-aff.
the same procedure as in [33] for comparison purpose. We
first detected 1000 affine covariant feature [20] points in
each image of the dataset and computed their SIFT [19]
descriptors using VLFeat library [26]. The affinity scores
were computed as the inner product between every pair of
detected feature points on each pair of images. We excluded apparent mismatches be keeping only affinity scores
that are above the threshold 0.7. Furthermore, any potential matches that are indistinguishable was removed, i.e. if
the first and the second top matches were below the ratio
threshold 1.1, the candidate point was removed. Finally,
any feature point that has only one candidate match in the
dataset was also excluded.
In a comparison, to construct our cover graph, we first
built a matching quality graph, using the matching score
as the edge weight and used the Fiedler vector of the graph
laplacian as the embedding and applied Algorithm 1 to build
the cover complex.
To evaluate the performance, we used the ground truth
homography matrix given in the dataset, and adopted the
procedure used in [4]. For a testing point, we calculated the
true correspondence using homography and compared with
the matched correspondence. If they were within a predefined distance threshold, we deemed the matching is correct,
and otherwise, wrong. Then we swept along the threshold
dimension to draw an error curve that is dependent on the
threshold chosen.
We tested our DMatch algorithm against MatchALS
[33], MatchLift [15], Spectral [23] and the original pairwise
matchings. We ignored the Consensus algorithm [18] as it
cannot explicitly handle partial matches. Figure 5 shows
the curve for three datasets, Graffiti, Bikes, and Leuven.
Note that our DMatch will not give a full pairwise matching
between images, instead, we only have a matching when
the two images belong to the same cover node. Therefore,
we computed the one-hop composite match between image
pairs across different cover nodes3 . From the performance
curve, we can see that our DMatch performs very similar to
the best global methods in all datasets as shown in [33]. In
another word, DMatch could achieve computational gains
without loss of matching quality.
In Figure 6, we show the example matches between the
first and the fourth image for each dataset. The bottom
match is DMatch, the middle is MatchALS and the top one
is the original pairwise map. Clearly, our matching shows at
least as good as the results of MatchALS, where both corrected mismatches (reduced blue lines) and increased correct matches (denser yellow lines).
In our implementation we notice that the total number of
iterations to converge for both DMatch and MatchALS are
3 For each image i and j not in the same cover node, we loop through
all k 6= i, j and accumulate the composite matchings from i → k and
k → j.
1
0.9
0.8
0.7
0.6
0.5
Original
MatchALS
MatchLift
Spectral
DMatch
0.4
0.3
0.2
0.1
0.8
0.7
0.6
0.5
Original
MatchALS
MatchLift
Spectral
DMatch
0.4
0.3
0.2
0.1
0
Percentage of matches
1
0.9
Percentage of matches
Percentage of matches
1
0.9
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Threshold
(a) Graffiti
0.7
0.6
0.5
Original
MatchALS
MatchLift
Spectral
DMatch
0.4
0.3
0.2
0.1
0
0
0.8
0
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Threshold
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
Threshold
(b) Bikes
(c) Leuven
Figure 5: The performance curve on Graffiti, Bikes and Leuven datasets. The y-axis is the correct match ration and the x-axis
is the threshold value over the image width. We compare DMatch (red solid) with MatchALS [33] (blue dashed), MatchLift
[15], Spectral [23], and original pairwise matching (black dotted).
(a) Graffiti
(b) Bikes
(c) Leuven
Figure 6: Example of matching results. The bottom one is from DMatch, the middle one is from MatchALS, and the top
one is from original pairwise matching respectively. Yellow lines encode the correct matches, while blue lines are for wrong
ones.
roughly the same (around 60 iterations).
6. Conclusion
In this paper, we introduced a scalable framework for
establishing consistent matches across multiple graphs in
a distributed manner. We showed how to use our framework to extend state-of-the-art global methods. By running
an iterative optimization algorithm locally and exchange information in every iteration, our framework would achieve
local and global consistent matching at the same time. Furthermore, we theoretically proved the sufficient conditions
under which locally consistent matching would guarantee
global consistency. In our experiments, we showed that in
practice, the assumptions and the conditions in our theorem
could be relaxed without sacrificing performance. In addition, our proposed distributed framework achieved order of
magnitude improvement in speed. We believe this is a very
important first step for large scale exploration of images for
object matching as well as building 3D object models from
crowd-sourced collections. Future work includes matching
large collection of different deformable objects that have
high similarity and enough variance, e.g. a collection of
different dogs or cats.
References
[1] C. Bregler, A. Hertzmann, and H. Biermann. Recovering
non-rigid 3d shape from image streams. CVPR, 2000. 1
[2] R. Cabral, F. De la Torre, J. P. Costeira, and A. Bernardino.
Unifying nuclear norm and bilinear factorization approaches
for low-rank matrix decomposition. ICCV, 2013. 4
[3] X. Chen, A. Shrivastava, and A. Gupta. Enriching visual knowledge bases via object discovery and segmentation.
CVPR, pages 2035–2042, 2014. 1
[4] Y. Chen, L. Guibas, and Q. Huang. Near-optimal joint object
matching via convex relaxation. ICML, 2014. 1, 2, 3, 5, 7
[5] L. Cosmo, E. Rodola, A. Albarelli, F. Memoli, and D. Cremers. Consistent partial matching of shape collections via
sparse modeling. CGF, 2016. 1
[6] T. Cour, P. Srinivasan, and J. Shi. Balanced graph matching.
In NIPS’06, pages 313–320, 2006. 1
[7] Y. Dai, H. Li, and M. He. A simple prior-free method for
non-rigid structure-from-motion factorization. CVPR, 2012.
1
[8] A. Egozi, Y. Keller, and H. Guterman. A probabilistic approach to spectral graph matching. IEEE Transactions on
PAMI, 99(PrePrints), 2012. 1
[9] D. Emms, R. C. Wilson, and E. R. Hancock. Graph matching
using the interference of continuous-time quantum walks.
Pattern Recogn., 42(5):985–1002, May 2009. 1
[10] M. Gori, M. Maggini, and L. Sarti. Exact and approximate
graph matching using random walks. IEEE Trans. Pattern
Anal. Mach. Intell., 27(7):1100–1111, July 2005. 1
[11] T. Hastie, R. Mazumder, L. J, and Z. R. Matrix completion
and low-rank svd via fast alternating least squares. The Journal of Machine Learning Research, pages 3367–3402, Jan
2015. 4
[12] A. Hatcher. Algebraic Topology. Cambridge University
Press, 2002. 3
[13] N. Hu, R. Rustamov, and L. Guibas. Graph matching with
anchor nodes: A learning approach. CVPR, pages 2906 –
2913, 2013. 1
[14] N. Hu, R. Rustamov, and L. Guibas. Stable and informative
spectral signatures for graph matching. CVPR, pages 2313 –
2320, 2014. 1, 6
[15] Q.-X. Huang and L. J. Guibas. Consistent shape maps via
semidefinite programming. SGP, 32(5):177–186, 2013. 1, 2,
3, 4, 6, 7, 8
[16] V. G. Kim, W. Li, N. J. Mitra, S. DiVerdi, and T. A.
Funkhouser. Exploring collections of 3d models using fuzzy
correspondences. ACM Trans. Graphics, 31(4), 2012. 2
[17] J. Lee, M. Cho, and K. M. Lee. Hyper-graph matching via
reweighted random walks. CVPR, pages 1633 – 1640, 2011.
1
[18] S. Leonardos, X. Zhou, and K. Daniilidis. Distributed consistent data association. ICAR, 2017. 2, 6, 7
[19] D. G. Lowe. Distinctive image features from scale-invariant
keypoints. IJCV, 60(2):91?110, 2004. 7
[20] K. Mikolajczyk, T. Tuytelaars, C. Schmid, A. Zisserman,
J. Matas, F. Schaffalitzky, T. Kadir, and L. Van Gool. A
comparison of affine region detectors. IJCV, 65(1-2):43?72,
2005. 7
[21] A. Nguyen, M. Ben-Chen, K. Welnicka, Y. Ye, and
L. Guibas. An optimization approach to improving collections of shape maps.
Computer Graphics Forum,
30(5):1481–1491, 2011. 1
[22] D. Pachauri, R. Kondor, G. Sargur, and V. Singh. Permutation diffusion maps (pdm) with application to the image
association problem in computer vision. NIPS, 2014. 2
[23] D. Pachauri, R. Kondor, and V. Singh. Solving the multi-way
matching problem by permutation synchronization. NIPS,
2013. 1, 2, 6, 7, 8
[24] B. Recht, M. Fazel, and P. Parrilo. Guaranteed minimumrank solutions of linear matrix equations via nuclear norm
minimization. SIAM Review, 52(3):471–501, 2010. 4
[25] S. Umeyama. An eigendecomposition approach to weighted
graph matching problems. IEEE Trans. PAMI, 10(5):695–
703, 1988. 1
[26] A. Vedaldi and B. Fulkerson. VLFeat: An open and portable
library of computer vision algorithms, 2008. http://
www.vlfeat.org/. 7
[27] R. C. Wilson and P. Zhu. A study of graph spectra for comparing graphs and trees. Pattern Recogn., 41(9):2833–2841,
Sept. 2008. 1
[28] J. Yan, M. Cho, H. Zha, X. Yang, and S. Chu. Multi-graph
matching via affinity optimization with graduated consistency regularization. IEEE Trans. PAMI, 2015. 1
[29] J. Yan, J. Wang, H. Zha, X. Yang, and S. Chu. Consistencydriven alternating optimization for multigraph matching: A
unified approach. IEEE Trans Image Processing, 24(3):994–
1009, 2015. 2
[30] J. Yan, H. Xu, H. Zha, X. Yang, H. Liu, and S. Chu. A matrix decomposition perspective to multiple graph matching.
ICCV, 2015. 2
[31] C. Zach, M. Klopschitz, and M. Pollefeys. Disambiguating
visual relations using loop constraints. In Computer Vision
and Pattern Recognition (CVPR), 2010 IEEE Conference on,
pages 1426–1433, 2010. 1
[32] R. Zass and A. Shashua. Probabilistic graph and hypergraph
matching. CVPR, 2008. 1
[33] X. Zhou, M. Zhu, and K. Daniilidis. Multi-image matching
via fast alternating minimization. International Conference
on Computer Vision (ICCV), 2015. 1, 2, 3, 4, 5, 6, 7, 8
[34] A. Zomorodian and G. Carlsson. Computing persistent homology. Discrete and Computational Geometry, 33:249–
274, 2005. 3
| 8cs.DS
|
arXiv:1601.05033v4 [math.DS] 23 Jan 2018
VARIATIONAL ANALYSIS OF INFERENCE FROM
DYNAMICAL SYSTEMS
KEVIN MCGOFF AND ANDREW B. NOBEL
Abstract. We introduce and study a variational framework for
the analysis of empirical risk based inference for dynamical systems and ergodic processes. The analysis applies to a two-stage
estimation procedure in which (i) the trajectory of an observed
(but unknown) system is fit to a trajectory from a known reference system by minimizing cumulative per-state loss, and (ii) a
parameter estimate is obtained from the initial state of the best fit
reference trajectory. We show that the empirical risk of the best
fit trajectory converges almost surely to a constant that can be
expressed in a variational form as the minimal expected loss over
dynamically invariant couplings (joinings) of the observed and reference systems. Moreover, we establish that the family of joinings
minimizing the expected loss is convex and compact, and that it
fully characterizes the asymptotic behavior of the estimated parameters, addressing both identifiability and misspecification. The
two-stage estimation framework and associated variational analysis
apply to a broad family of empirical risk miminization procedures
for dependent observations. To illustrate this, we apply variational
analysis to the well studied problems of maximum likelihood and
non-linear regression, and then undertake an extended analysis of
system identification from quantized trajectories subject to noise,
a problem of interest in dynamics, where the models themselves
exhibit dynamical behavior across time.
1. Introduction
Minimization of empirical risk is a common approach to statistical
inference, having roots in maximum likelihood estimation from independent observations. Unlike maximum likelihood estimation, however, the form and motivation of empirical risk minimization remain
unchanged when considering stationary, dependent observations.
In this paper we introduce and study a variational framework for
the analysis of empirical risk based inference from dynamical systems
and ergodic processes. The variational analysis applies to a family of
procedures that can be decomposed into two stages: a tracking stage in
which a trajectory of a known reference system is fit to the trajectory
1
2
KEVIN MCGOFF AND ANDREW B. NOBEL
of an observed system by minimizing cumulative loss; and a translation
stage in which a parameter estimate is obtained by applying a continuous invariant map to the initial state of the best-fit reference trajectory.
We emphasize that tracking plus translation is not being proposed as
an inference procedure, but rather as a framework within which to
study empirical risk minimization. In Section 4 we show that the twostage inference framework encompasses the well studied problems of
maximum likelihood estimation and non-linear regression, as well as
new, more complex problems, including fitting dynamical models and
system identification from quantized trajectories.
This paper makes three main contributions. The first contribution
is the introduction of ideas and techniques from ergodic theory to the
study of empirical risk minimization through an analysis of two-stage
inference. The most important of these ideas is the notion of a joining,
which is a dynamically invariant coupling of two systems (see Section 2
for definition and further discussion). Our principal results concerning
two-stage inference are the following. First, the average loss of the optimal reference trajectory in the tracking stage converges almost surely
to a constant that can be expressed in variational form as the minimum
expected loss over all joinings of the observed and reference systems.
Second, the family of optimal joinings, namely those that achieve the
minimum expected loss in the variational problem, is non-empty, convex, and compact in the weak topology. Third, the family of optimal
joinings characterizes the limiting behavior of the estimates derived in
the translation stage.
Together, these results constitute a variational analysis of the twostage inference framework. In particular, they provide an explicit,
variational characterization of the asymptotic behavior of two-stage
inference that establishes a direct connection (not previously known
or understood) between minimum risk-based inference and an infinite
dimensional optimization problem in which the observation process is
projected, using a joining-based distortion, onto a family of processes
that is associated with the statistical models under study. This variational analysis is the second main contribution of the paper.
Variational analysis has a number of desirable properties. It requires
relatively mild assumptions (given in detail below). It readily accommodates model misspecification, as the observed and reference systems
need not be related to one another. It addresses the problem of identifiability in a direct way, by fully characterizing the limiting parameter
set, and it provides a systematic means of studying the limiting behavior and potential consistency of two-stage inference.
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 3
The third main contribution of the paper is an extended application
of variational analysis to the problem of system identification from
quantized trajectories with label noise, a problem that lies at the intersection of statistics and dynamical systems. In this case, analysis
requires results and constructions from the theory of joinings, illustrating the full power of the variational approach.
In the next two sections we describe the two-stage inference procedure and present the principal results underlying variational analysis.
Several applications of variational analysis to several existing and new
inference problems are given in Section 4, and our main application, the
analysis of system identification with quantized trajectories, is given in
Section 5.
2. The tracking stage
The tracking stage of the two-stage inference framework has three
basic components: an observed ergodic system, a reference topological
system, and an integrable loss function.
The observed dynamical system is a triple (Y, T, ν) consisting of a
non-empty Polish space Y, a Borel measurable map T : Y → Y, and a
Borel probability measure ν on Y that is invariant and ergodic under
T . Recall that ν is invariant under T if the action of T preserves ν in
the sense that ν(T −1 A) = ν(A) for each Borel set A ⊆ Y. Furthermore,
recall that ν is ergodic under T if T −1 A = A implies ν(A) ∈ {0, 1}, that
is, any set that is invariant under the action of T is negligible or has
full measure. The triple (Y, T, ν) is the standard model of a stationary
ergodic dynamical system [50, 58]. It is well known [7, 12] that any
stationary ergodic process can be expressed as such a triple (with Y
equal to a sequence space, T equal to the left shift transformation, and
ν equal to the process distribution) plus a univariate projection.
By contrast with the observed dynamical system, the reference system is a pair (X , S) consisting of a non-empty, compact metric space
X and a continuous map S : X → X . In the dynamics literature, the
pair (X , S) is referred to as a topological dynamical system [10, 58].
The loss ℓ : X × Y → R is assumed to be a lower semicontinuous function satisfying the envelope condition supx∈X |ℓ(x, y)| ≤ ℓ∗ (y) for some
ℓ∗ ∈ L1 (ν). In what follows, the conditions above will be referred to as
the standard assumptions.
Tracking. In the tracking problem we have access to a single trajectory y, T y, T 2 y, . . . of the observed system (Y, T, ν) with initial
state y drawn according to ν. Here T k denotes the k-fold composition
of T with itself and, by convention, T 0 is the identity. At time n we
4
KEVIN MCGOFF AND ANDREW B. NOBEL
observe the initial segment y, T y, . . . , T n−1 y of the trajectory and identify a corresponding initial state x of the reference system (X , S) that
Pn−1
minimizes the average loss n−1 i=0
ℓ(S i x, T i y). Our assumptions on
ℓ and (X , S) ensure that a minimizing initial condition exists.
The reader will note that there is an asymmetry in the specification
of the observed and reference systems: the observed system is equipped
with an invariant ergodic measure, while the reference system is specified without reference to any particular invariant measure. Indeed, we
view the reference system as a family of stationary dynamical systems
corresponding to the family of S-invariant measures. In more detail,
let M(X , S) be the family of all Borel probability measures on X that
are invariant under S in the sense that µ ◦ S −1 = µ. It can be shown
[29] that the family M(X , S) is non-empty. Importantly, every measure µ ∈ M(X , S) determines a stationary, but not necessarily ergodic,
dynamical system (X , S, µ). Let
S(X , S) = {(X , S, µ) : µ ∈ M(X , S)}
be the family of all such systems. As we show below, the asymptotic behavior of two-stage inference is determined by the relationship
between the observed system (Y, T, ν) and the family S(X , S). Quantifying this relationship requires two concepts. The first is the notion
of joining, which is a dynamically invariant couplings of two stationary systems; the second is a variationally defined distortion between
stationary systems.
Definition 2.1. Let S × T : X × Y → X × Y be the product transformation of S and T , defined by (S × T )(x, y) = (Sx, T y), and let
µ ∈ M(X , S). A Borel probability measure λ on X × Y is a said to be
a joining of the systems (X , S, µ) and (Y, T, ν) if λ is invariant under
S × T and the marginals of λ on X and Y are µ and ν, respectively.
A joining is a coupling of the measures µ and ν with the additional
property that the coupling is invariant (stationary) under the product
transformation S × T . The simple example given below gives an indication of the role that this invariance condition plays. Joinings were
introduced and first studied by Furstenberg [14], and they have played
an important role in ergodic theory since then, see [9, 15] for more
details.
For each µ ∈ M(X , S) let J (µ, ν) denote the family of all joinings
of (X , S, µ) and (Y, T, ν). Note that J (µ, ν) contains the product
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 5
measure µ ⊗ ν and is therefore non-empty. Define
[
J (S : ν) =
J (µ, ν),
µ∈M(X ,S)
to be the set of all joinings of the observed system with some system
in S(X , S). It is easy to see that J (S : ν) is just the set of Borel
probability measures λ on X × Y such that λ is invariant under S × T
and the Y-marginal of λ is ν.
Example 2.2. As a simple example of a joining, let X = Z/2Z with
S(x) = x + 1 mod 2, and let Y = Z/3Z with T (y) = y + 1 mod 3.
There is only one measure µ in M(X , S), which places measure 1/2
on both 0 and 1. Similarly, there is only one invariant and ergodic
measure ν for (Y, T ), which places measure 1/3 on 0, 1, and 2. To form
a joining of these systems, we need a probability measure on X × Y
that is invariant under S × T . As 2 and 3 are relatively prime, orbits of
the map S × T cycle among the pairs (0, 0), (1, 1), (0, 2), (1, 0), (0, 1),
(1, 2) in that order. Hence, there is only one invariant measure for
S × T , the uniform measure, which is the product joining of µ and ν.
Observe that in this case J (S : ν) contains only one element, despite
the fact that there are uncountably many couplings of µ and ν.
Our first principal result shows that the average risk in the tracking
problem has a limit, and that the limit has a simple variational form
involving the loss function ℓ and the family of joinings J (S : ν). The
proof of the theorem is given in Section 8.
Theorem 2.3 (Variational expression for limiting average risk). Under
the standard assumptions, for ν-almost every initial state y ∈ Y,
Z
n−1
1X
k
k
(2.1)
lim inf
ℓ(S x, T y) = inf
ℓ dλ := L(S : ν).
n x∈X n
λ∈J (S:ν)
k=0
The second infimum is finite, and is attained by some joining λ in
J (S : ν).
Remark 2.4 (Distortion and Projections). Note that L(S : ν)
R can be
written as inf µ∈M(X ,S) d(µ : ν) where d(µ : ν) = inf λ∈J (µ:ν) ℓ dλ defines a joining-based distortion between the stationary systems (X , S, µ)
and (Y, T, ν) under the loss function ℓ. Thus the limiting average risk
L(S : ν) in the tracking problem is the minimum distortion between
the observed ergodic system (Y, T, ν) and the family S(X , S) associated with the reference system (X , S). It is natural then to investigate
the joinings in J (S : ν) that achieve the minimum L(S : ν). These
6
KEVIN MCGOFF AND ANDREW B. NOBEL
joinings capture the joint behavior of the observed system (Y, T, ν) and
its projection onto the family S(X , S) under the distortion d(· : ·).
Recall that µ is said to be an extreme point of a convex family of
probability measures M if µ = tµ1 + (1 − t)µ2 with t ∈ (0, 1) and
µ1 , µ2 ∈ M implies µ1 = µ2 = µ. The proof of the following theorem
can be found in Appendix A.
Theorem 2.5 (Structure of optimal joinings). Under the standard assumptions, the set of optimal joinings
Z
n
o
Jmin (S : ν) = λ ∈ J (S : ν) : ℓ dλ = L(S : ν)
is non-empty, convex, and compact in the weak topology. Furthermore,
a joining λ is an extreme point of Jmin (S : ν) if and only if it is ergodic
under S × T .
Remark 2.6. It follows from Theorem 2.5 and the Krein-Milman theorem that there exists an optimal joining that is ergodic under S × T .
By considering the X -marginal of such a joining, we see that there exists an ergodic measure µ for the system (X , S) that can be optimally
joined with ν.
3. Translation
For a fixed sample size n the tracking problem is a special case of empirical risk minimization in which the observed sequence y, T y, . . . , T n−1 y
is fit using a family of sequences x, Sx, . . . , S n−1x indexed by initial
states x ∈ X of the reference system. In the translation stage of the
inference procedure, a parameter estimate is obtained from the initial
state of a reference trajectory that minimizes, or nearly minimizes, the
empirical risk.
Let Θ be a compact metrizable parameter space. A parameter map is
a continuous function ϕ : X → Θ that is invariant under the dynamics
of the reference system in the sense that ϕ ◦ S = ϕ. In what follows
these conditions on Θ and ϕ are included in the standard assumptions.
The invariance of ϕ ensures that the value θ = ϕ(x) is constant on the
trajectory x, Sx, S 2 x, . . ., and θ may therefore be viewed as a property
of the entire trajectory of x under S.
Definition 3.1 (ϕ-estimation). A sequence of measurable functions
θn : Y n → Θ, n ≥ 1, is an optimal ϕ-estimation scheme if θn = ϕ ◦ xn
where the functions xn : Y n → X are such that for ν-almost every y in
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 7
Y, with x̂n = xn (y, . . . , T n−1 y),
(3.1)
n−1
n−1
k=0
k=0
1X
1X
ℓ(S k x̂n , T k y) = lim inf
ℓ(S k x, T k y).
lim
n x∈X n
n n
Thus the estimate θ̂n = θn (y, . . . , T n−1y) is obtained by applying the
parameter map ϕ to an initial state x̂n of the reference system obtained
by minimizing (or nearly minimizing) the average loss with the observed
trajectory y, T y, . . . , T n−1 y.
We wish to characterize the limiting behavior of optimal ϕ-estimation
schemes {θ̂n }. To this end, for each θ ∈ Θ let Xθ = ϕ−1 {θ} be the set
of states in X that are mapped to θ, and let Sθ be the restriction of the
continuous map S to Xθ . It is easy to see that Xθ is a compact subset
of X that is invariant under Sθ , and therefore (Xθ , Sθ ) is a topological
dynamical system. Thus the parameter map ϕ gives rise to a family of
topological systems, indexed by the parameters θ ∈ Θ, each of which
can act as a reference system for tracking the observed system (Y, T, ν).
It follows from Theorem 2.3 that the limiting average loss of tracking (Y, T, ν) with the reference system (Xθ , Sθ ) is equal to L(Sθ : ν).
It is not difficult to show (see Lemma 9.2 below) that L(S : ν) =
minθ∈Θ L(Sθ : ν), and we therefore consider the set of parameters with
minimal limiting loss, namely,
(3.2)
Θmin = argmin L Sθ : ν .
θ∈Θ
It is straightforward to show that
(3.3)
Θmin = θ ∈ Θ : ∃ λ ∈ Jmin (S : ν) s.t. λ(Xθ × Y) = 1
is just the set of parameters θ whose associated states Xθ ⊆ X support
an optimal joining with the observed system (Y, T, ν). In other words,
Θmin is the set of parameters that can be optimally joined with the observed system, and it therefore serves as a natural limit set for optimal
ϕ-estimation schemes.
The next theorem is our principal result concerning two-stage inference. Its proof appears in Section 9. Here and throughout the paper,
we say that a sequence {θn }n≥1 converges to a set Θ0 if the distance of
θn to Θ0 converges to zero.
Theorem 3.2 (Convergence of Optimal ϕ-Estimators). Under the
standard assumptions Θmin is non-empty and compact. Moreover, if
(θn )n≥1 is an optimal ϕ-estimation scheme, then θ̂n = θn (y, . . . , T n−1y)
converges to Θmin for ν-almost every y ∈ Y. Conversely, for every
8
KEVIN MCGOFF AND ANDREW B. NOBEL
θ0 ∈ Θmin there exists an optimal ϕ-estimation scheme that converges
almost surely to θ0 .
Theorem 3.2 fully characterizes the limiting behavior of two stage
inference. In particular, it reduces questions about identifiability and
consistency to the analysis of the set Θmin , and the associated family
Jmin (S : ν) of optimal joinings. In this way, the theorem facilitates
the application of joining constructions and properties to the analysis
of empirical risk minimization in statistics and machine learning, see
Section 4 below. We emphasize that the theorem places no restrictions
on the relation between the observed and reference systems, which need
not be the same.
In summary, the variational analysis presented above provides a
starting point for the analysis of two-stage empirical risk minimization
that may involve dynamical models. With these results it is possible
to characterize the convergence of two-stage estimators in a variety of
applications, as illustrated in the following section. Next steps would
include rates of convergence or distributional convergence theorems.
Results of this type, while desirable, will involve substantially stronger
assumptions and substantial problem specific analyses, and are beyond
the scope of the present paper.
4. First examples of variational analysis
The two-stage inference framework may at first appear to be rather
restrictive, as it involves fitting deterministic observations to deterministic models in the absence of noise, a problem that is relatively rare in
statistical theory and practice. However, by appropriate choice of the
observed system (Y, T, ν), the reference system (X , S), and the loss
function ℓ(x, y), the framework and the results of Theorem 3.2 may
be applied to a wide variety of inference problems involving empirical risk minimization and dependent observations. In Sections 4.1 and
4.2 below, we show how maximum likelihood estimation and nonlinear
regression can be placed within the framework of two-stage inference,
and how Theorem 3.2 can be used to establish existing results on the
consistency of empirical risk based procedures for these problems.
While these initial applications are illuminating, the two-stage inference framework and Theorem 3.2 also provide powerful tools for
analyzing new, more complex, statistical problems in which the observations as well as the models of interest are dynamical in nature. In
Section 4.3 we briefly review recent work of McGoff and Nobel [39]
on fitting families of dynamical models, which is based in part on the
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 9
results of Theorem 3.2. The full generality of the two-stage framework and the power of the variational characterization is illustrated in
Section 5, where we describe and analyze the problem of system identification from quantized trajectories. These results are of independent
interest and, to the best of our knowledge, cannot be obtained by any
existing statistical methods.
4.1. Maximum likelihood estimation under ergodic sampling.
Let U be a Polish space, and let P = {fθ : θ ∈ Θ} be a family of density
functions fθ : U → [0, ∞) with respect to a fixed Borel measure Q on U.
Let Θ be a compact metric space and assume that (θ, u) 7→ fθ (u) is an
upper semi-continuous map from Θ × U to R. Suppose that we observe
the values of a stationary ergodic process U0 , U1 . . . ∈ U and wish to
identify a density fθ ∈ P that best approximates the one-dimensional
marginal distribution of the observed process in the sense that
(4.1)
E log fθ (U) = max
E log fθ′ (U),
′
θ ∈Θ
where U has the same distribution as U0 . Let θn : U n → Θ, n ≥ 1,
be measurable estimators that asypototically maximize the marginal
log-likelihood
n−1
(4.2)
n−1
1X
1X
lim
log fθ̂n (Ui ) = lim sup
log fθ (Ui ) wp1,
n n
n θ∈Θ n
i=0
i=0
where θ̂n = θn (U0 , . . . Un−1 ). The existence of measurable estimators
satisfying (4.2) follows from standard arguments, see Lemma 9.1. Note
that the common distribution of the observations Ui need not have a
density in P and need not be absolutely continuous with respect to the
reference measure Q.
The problem described above can be expressed as a two-stage inference procedure in the following way. To begin, we represent the
observed process {Ui }i≥0 as a shift-based system (Y, T, ν) following
the standard construction [7, 12] in which Y be the sequence space U N ,
T is the left-shift on Y, and ν is the measure on U N induced by {Ui }.
Let the state space X of the reference system be equal to the parameter
space Θ and, as the inference task involves no dynamics beyond those
of the observations Ui , let S be the identity map on Θ. Finally, let the
loss function ℓ : X × Y → R be defined by ℓ(θ, (ui )i≥0 ) = − log fθ (u0 ),
and let ϕ : X → Θ be the identity map. These correspondences are
detailed in Table 1. A direct application of Theorem 3.2 yields the
following classical result, which is similar to Theorem 5.14 of [56].
10
KEVIN MCGOFF AND ANDREW B. NOBEL
General setting MLE under ergodic sampling
X
Θ
S:X →X
Id : Θ → Θ
Y
UN
T :Y →Y
Left shift τ on U N
ν
Measure of process {Ui }i≥0
c:X ×Y →R
(θ, u) 7→ − log pθ (u0 )
Θ
Θ
ϕ:X →Θ
Id : Θ → Θ
Table 1. Correspondence between objects in the general setting and objects in MLE under ergodic sampling.
Theorem 4.1. If E supθ∈Θ | log fθ (U)| is finite, then θ̂n converges almost surely to the non-empty set Θ0 = argmaxθ∈Θ E log fθ (U).
Thus even in the misspecified setting, the empirical maximum likelihood estimators converge to the set of optimal parameters, namely
those that best approximate the observed process in the sense of (4.3).
If the supremum in the theorem fails to be measurable, one may replace
the expectation there by an outer expectation.
4.2. Nonlinear regression under ergodic sampling. Let U be a
Polish space, and let F = {fθ : θ ∈ Θ} be a family of functions
fθ : U → R indexed by a compact metric space Θ in such a way that
(θ, u) 7→ fθ (u) is a continuous map from Θ × U to R. Suppose that
we observe a stationary ergodic process (U0 , V0 ), (U1 , V1 ), . . . ∈ U × R
and wish to identify a function fθ ∈ F that best captures the marginal
relationship between U and V in the sense that
(4.3)
E ℓ0 (fθ (U), V ) = min
E ℓ0 (fθ′ (U), V ),
′
θ ∈Θ
where ℓ0 : R × R → [0, ∞) is a lower semicontinuous loss function. Let
θn : (U × R)n → Θ, n ≥ 1, be measurable estimators that asymptotically minimize the average loss, namely
n−1
(4.4)
n−1
1X
1X
lim
ℓ0 (fθ̂n (Ui ), Vi ) = lim inf
ℓ0 (fθ (Ui ), Vi ) wp1,
n n
n θ∈Θ n
i=0
i=0
where θ̂n = θ̂n ((U0 , V0 ), . . . (Un−1 , Vn−1)). This problem can readily be
expressed as a two-stage inference procedure; Table 2 gives the details.
The following result is an easy consequence of Theorem 3.2. The result
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 11
can also be established by arguments based on uniform laws of large
numbers.
Theorem 4.2. If E supθ∈Θ ℓ0 (fθ (U), V ) is finite then θ̂n converges almost surely to the set Θ0 = argminθ∈Θ E ℓ0 (fθ (U), V ).
General setting Regression under ergodic sampling
X
Θ
S:X →X
Id : Θ → Θ
Y
(U × R)N
T :Y→Y
Left shift τ on (U × R)N
ν
Measure of process (Ui , Vi )i≥0
ℓ:X ×Y →R
(θ, (u, v)) 7→ ℓ0 (fθ (u0 ), v0 )
Θ
Θ
ϕ:X →Θ
Id : Θ → Θ
Table 2. Correspondence between objects in the general setting and objects in the regression under ergodic
sampling section.
4.3. Fitting dynamical models. Here we provide a brief overview
of some recent results on fitting dynamical models to ergodic processes
[39]. These results use the variational analysis of the present work as
a starting point for a detailed analysis of a specific inference problem.
In contrast to the two previous examples (maximum likelihood estimation and nonlinear regression), the underlying model family in this
application is complex and captures interesting dynamical behavior.
A dynamical model consists of a continuous transformation f : K →
K on a compact metric space K and a continuous observation function g : K → Rd . The specific inference problem considered in [39]
involves fitting a family of dynamical models to observations from a
stationary ergodic process. Let Θ be a compact metric space, and
let D = {(fθ , gθ ) : θ ∈ Θ} be an indexed family of dynamical models on a common state space K such that both (θ, x) 7→ fθ (x) and
(θ, x) 7→ gθ (x) are continuous. Examples of natural families of dynamical models include the family of toral rotations and the logistic
family (see [25]). The family D is meant to capture regularities (deterministic patterns) of interest, and the goal of the inference is to
identify these regularities or patterns in an observed ergodic process
U = U0 , U1 , · · · ∈ Rd by fitting the values of the process with models
12
KEVIN MCGOFF AND ANDREW B. NOBEL
in D. This type of fitting occurs frequently in fields such as systems
biology and ecology, see for example [6, 35, 36, 41, 55].
The main results of [39] concern the asymptotic behavior of parameter estimates that arise from fitting dynamical models in D to U. In
particular, the results establish that if the model class D has low complexity, then empirical risk based fitting procedures are asymptotically
consistent in signal plus noise settings. Theoretical analysis of model
fitting begins with the variational analysis of Theorem 3.2 and makes
use of the theory and construction of joinings.
5. System identification from quantized trajectories
As a final application of the variational analysis techniques in Section
1, we consider the problem of system identification from quantized
trajectories. As in the fitting of dynamical models, the model family
in this problem is complex and captures dynamical behavior.
Discretizing the state space of a dynamical system is common in
both theoretical and applied settings. On the theoretical side, studying systems through their quantized trajectories has been a core idea
in ergodic theory since the theory of entropy was developed by Kolmogorov, Ornstein, and others (see the books [50, 58]). Moreover,
discretization of trajectories occurs frequently in applied settings as a
way of coarse graining the state space, accommodating computation,
and meeting data transmission and storage requirements. For example, many mathematical models of gene regulatory network dynamics
involve discretizing the gene expression levels [23]. Also, many numerical methods for analyzing dynamical systems begin by discretizing the
state space; examples include methods based on Conley index theory
[43, 42], Ulam’s method [5], or finite element method [11]. To our
knowledge, the results of this section represent the first detailed and
rigorous statistical analysis of inference for dynamical systems based on
discretized observations. Our analysis makes use of relatively independent joinings, as well as the classical result of Furstenberg concerning
the disjointness of zero entropy and Bernoulli systems (described in
detail below).
Let U be a Polish space and let R = {Rθ : θ ∈ Θ} be a family of
Borel measurable transformations Rθ : U → U indexed by a compact
metric space Θ in such a way that the map (θ, u) 7→ Rθ (u) is a Borel
measurable function from Θ × U to U. Let {A0 , A1 } be a known,
measurable partition of U, and let π : U → {0, 1} be the associated
label function defined by π(u) = j if u ∈ Aj . For each parameter θ ∈ Θ
and element u ∈ U there is an associated trajectory u, Rθ u, Rθ2 u, . . .
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 13
arising from repeated application of Rθ . Application of the map π gives
rise to a corresponding binary label sequence
lab(θ, u) = (π(u), π(Rθ u), π(Rθ2 u), . . .) ∈ {0, 1}N .
Of interest here is whether, and in what sense, we can estimate the
parameter θ from noisy observations of the label sequence lab(θ, u)
when the state u is drawn from an invariant ergodic measure for Rθ .
In more detail, we assume that observations take the form of a binary
stochastic process
(5.1)
Yk = π(Rθk0 U) ⊕ εk ,
k ≥ 0.
Here θ0 ∈ Θ is the parameter of the underlying transformation, U is a
U-valued random variable whose distribution is invariant and ergodic
under Rθ0 , {εk } is a sequence of independent Bernoulli(p) random variables independent of U, and ⊕ denotes addition modulo 2. Thus Yk is
equal to the label of Rθk0 U perturbed by noise. Standard arguments ensure that {Yk : k ≥ 0} is stationary and ergodic. Both the setting and
the results of this section may be generalized to allow arbitrary finite
measurable partitions and more general noise distributions (see [40]);
we consider the case of binary observations for clarity of presentation.
Let θn : {0, 1}n → Θ, n ≥ 1, be measurable estimators of θ0 that
approximately minimize average Hamming (0-1) risk,
n−1
(5.2)
1X
lim inf
I π Rθ̂kn u 6= Yk
n u∈U n
k=0
n−1
= lim inf inf
n
θ∈Θ u∈U
1X
I π Rθk u 6= Yk wp1,
n k=0
where θ̂n = θn (Y0 , . . . , Yn−1 ). The existence of measurable estimators
satisfying (5.2) is guaranteed by Lemma 9.1. We are interested in the
limiting behavior of θ̂n .
The estimation problem above can be expressed as a two stage inference procedure. As in the previous examples, the observation process
{Yk }k≥0 can be represented as a measure preserving system (Y, T, ν),
where Y is the sequence space {0, 1}N , T is the left-shift on {0, 1}N , and
ν is the measure of the process {Yk }k≥0 . Specification of the reference
system and loss require more care. Let the state space of the reference
system be the closure of all (parameter, label-sequence) pairs,
X = cl (θ, lab(θ, u)) : θ ∈ Θ, u ∈ U ⊆ Θ × {0, 1}N ,
where clA denotes the closure of A and we assume that {0, 1}N is
equipped with the usual product topology. Thus X is compact, and
14
KEVIN MCGOFF AND ANDREW B. NOBEL
we let the reference transformation S be the restriction to X of the
product idΘ × τ , where idΘ is the identity on Θ and τ is the left shift
on {0, 1}N . It is easy to see that X is invariant under S. We further
define the loss ℓ((θ, a), b) = I(a0 6= b0 ), and we let the parameter map
ϕ be the projection onto Θ, namely ϕ(θ, a) = θ. These correspondences
are summarized in Table 3.
Theorem 3.2 characterizes the limiting behavior of the parameter
estimates θ̂n . Let ν0 be the distribution of the true label process
{π(Rθk0 U) : k ≥ 0} on {0, 1}N . Note that ν0 is not equal to the process measure ν of {Yk } if the noise level p > 0. For each θ ∈ Θ let
Xθ = {a : (θ, a) ∈ X } be the θ-section of X , and define
(5.3)
Θ1 = θ ∈ Θ : ν0 (Xθ ) = 1
to be the set of parameters for which Xθ supports the true label process.
It is easy to see that θ0 ∈ Θ1 , so Θ1 is non-empty.
An equivalent way of viewing the parameter estimator θn is that it
picks a θ such that some sequence in Xθ provides the best Hamming
match to the observations Y0 , . . . , Yn−1. In other words, if {θn }n≥1
satisfies (5.2), then
n−1
n−1
1X
1X
lim inf
I ak 6= Yk = lim inf inf
I ak 6= Yk wp1
n a∈Xθ̂ n
n θ∈Θ a∈Xθ n
n
k=0
k=0
where θ̂n = θn (Y0 , . . . , Yn−1). Thus Θ1 is a natural identifiability class
in the absence of noise (p = 0). (See Remark 5.2 below for more
discussion of identifiability.)
In fact, Θ1 is also an identifiability class when noise is present (p > 0),
provided there are complexity constraints on the family of transformations R. To quantify these constraints, let L be the closure (in {0, 1}N )
of the set of all label sequences {lab(θ, u) : θ ∈ Θ, u ∈ U} of the
transformations in R. Let
1
(5.4)
h(R) = lim log # a0n−1 ∈ {0, 1}n : a ∈ L ,
n n
be the exponential growth rate of the number of distinct labeled trajectories of length n. A more detailed discussion of Θ1 and h(R) can
be found in Section 10, which also contains the proof of the following
theorem.
Theorem 5.1. Let {θn } be a sequence of estimators satisfying (5.2).
If either
(1) p = 0 or
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 15
General setting
Quantized observations
X
cl{(θ, lab(θ, u)) : θ ∈ Θ, u ∈ U}
S:X →X
idΘ × τ restricted to X
Y
{0, 1}N
T :Y →Y
Left shift τ on {0, 1}N
ν
Measure of process {Yk }k≥0
ℓ:X ×Y →R
((θ, a), b) →
7 I(a0 6= b0 )
ϕ:X →Θ
(θ, a) →
7 θ
Table 3. Correspondence between objects in the general setting and objects in the estimation of a transformation with quantized observations.
(2) 0 < p < 1/2 and h(R) = 0,
then θ̂n converges almost surely to Θ1 . Conversely, for every θ0 ∈ Θ1
there exists a sequence of estimators {θn } satisfying (5.2) that converges
almost surely to θ0 .
Note that the limit set Θ1 is the same for both the noisy and noisefree settings. Similar results hold when trajectories are quantized by
arbitrary finite partitions and subject to more general noise, see [40].
Note also that Theorem 5.1 holds without continuity assumptions on
the transformations Rθ or their indexing by θ. However, the topology
of Θ does play an important role in this result, as it affects the closure
operation that defines X , which in turn is used to define Θ1 .
Remark 5.2. The identifiability class Θ1 may be significantly larger
than the true parameter θ0 . For example, if θ1 is a parameter such that
Xθ1 = {0, 1}N , then Xθ1 supports all invariant probability measures, in
which case θ1 is certainly contained in Θ1 , regardless of the true value
of θ0 . Moreover, note that Xθ is defined after taking the closure to
form X . Thus, it may happen that Xθ contains binary sequences that
do not appear as the label sequence of any trajectory from Rθ , but
are instead limits of such trajectories. This closure may also cause
Θ1 to contain parameters other than θ0 . (See Proposition 10.5 for a
sufficient condition that guarantees that Θ1 = {θ0 }.) Theorem 5.1
nonetheless ensures convergence of empirical risk based estimates to
the identifiability class Θ1 , and the converse part of the theorem shows
that no smaller set would satisfy the same conclusion. Identification
of Θ1 can be carried out within the variational framework, as the next
example shows, but additional analysis is required.
16
KEVIN MCGOFF AND ANDREW B. NOBEL
5.0.1. Circle Rotations. As a non-trivial application of the results in
this section, consider the family R of circle rotations Rα : [0, 1) → [0, 1)
defined by Rα (x) = x+ α mod 1, with α ∈ Θ = [0, 1/2]. If α is rational
with reduced form m/n, then Rαn = Id, each orbit contains exactly n
distinct points, and each ergodic measure is supported on a single orbit.
On the other hand, if α is irrational, then it is known that Lebesgue
measure is the only ergodic Borel probability measure for Rα . Consider
the partition of [0, 1) into sets A0 = [0, 12 ) and A1 = [ 21 , 1). The proof
of the following result appears in Section 10.
Proposition 5.3. Let Yi = π(Rαi 0 U) ⊕ ǫi , where α0 ∈ [0, 1/2], the
distribution of U is invariant and ergodic under Rα0 , and {ǫi }i≥0 is
an independent sequence of Bernoulli(p) random variables that is independent of U. If p < 1/2 and the estimates θ̂n satisfy (5.2), then θ̂n
converges almost surely to α0 .
6. Related work
A number of recent papers have considered prediction and inference
from dynamical systems that evolve deterministically over time, e.g.,
[19, 20, 30, 32, 33, 37, 54]. While the variational framework considered
here focuses on the problem of parameter estimation, the recent work
cited above has focused on different aspects of statistical inference. For
example, Hang and Steinwart and Steinwart and Anghel [19, 20, 54]
give some results about forecasting dynamical systems with specified
mixing rates, and Lalley and Nobel [32, 33] establish both positive
and negative results for filtering problems in the context of certain
dynamical systems. For additional references and discussion, see the
recent survey on statistical inference for dynamical systems [38].
In several applied fields, there is interest in fitting parametrized families of dynamical systems to observations. Indeed, there are examples in ecology [36, 55], geophysical modeling [3, 22], systems biology
[6, 35, 41], and data assimilation [34]. As explained in greater detail
in [39], the variational approach taken here may be useful in analyzing
the fitting methods in settings such as these.
Although independence assumptions are common in the statistics
and machine learning literature, there has been long-standing interest,
both theoretical and applied, in the analysis of observations exhibiting
long-range dependence. Representative recent work can be found in [1,
2, 19, 31, 44, 59]. Most work in this area considers rates of convergence
and finite sample bounds. As noted at the end of Section 3, such
results in the general setting considered here would require substantial
additional analysis.
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 17
In ergodic theory, Ornstein and Weiss [49] considered finitary estimation of a stationary ergodic process from samples of the process.
They proposed a specific estimation scheme and characterized when
the scheme produces consistent estimates of the observed process. Interestingly, consistent estimation may be possible for restricted classes
of systems or processes, as we show in some of our results, despite the
fact that consistency is impossible for larger classes of processes (as
shown by Ornstein and Weiss). Other related work in ergodic theory
concerns finitary estimation of k-dimensional distributions for growing
k [24] and finitary estimation of isomorphism invariants [18, 46].
The minimal expected loss L(S : ν) and the set of optimal joinings
Jmin (S : ν) defined here have close analogies in the study of optimal
transport; see the book by Villani [57]. Indeed, one of the main goals
in optimal transport is to describe the properties of optimal couplings,
that is, couplings that achieve the infimum of the expected cost. Such
optimal couplings are analogous to the optimal joinings in Jmin (S : ν).
In some cases, notably in the case of Ornstein’s d-metric and its generalizations in ergodic theory and information theory (see the work of
Gray, Neuhoff and Shields [17] and the book of Gray [16]), the measures µ and ν are taken to be process measures, and the couplings are
required to be joinings. While similar in spirit, our results are distinct
from this previous work, since we consider the family of joinings between a topological dynamical system and a measure-preserving system
and we focus on applications to inference.
In the special case that the loss function does not depend on the observed trajectory, the tracking part of our two-stage procedure reduces
to the problem of ergodic optimization, which has received considerable
attention in the mathematical literature in recent years (see the survey
of Jenkinson [21] for a thorough introduction to the topic). For some
recent results, see the work of Quas and Seifken [52] and references
therein.
6.1. Organization of the rest of the paper. The next section provides some background notation and preliminary lemmas needed for
the proofs of the main results. Theorems 2.3 and 3.2 are established
in Sections 8 and 9, respectively. Section 10 contains the proofs of
consistency for the quantized observation problem presented above.
Appendix A contains material on the set of optimal joinings, including
the proof of Theorem 2.5.
18
KEVIN MCGOFF AND ANDREW B. NOBEL
7. Definitions and background
In this section we provide several preliminary definitions and facts
required for the principal results of the paper.
7.1. Dynamical systems and spaces of measures. All topological
spaces considered in this paper are Polish (separable and completely
metrizable). We endow any such space with its Borel σ-algebra and
suppress this choice in our notation. Let U be a Polish space. Following
standard notation, M(U) will denote the space of Borel probability
measures on U endowed with the usual weak topology, under which
M(U) is itself a Polish space. Recall that if R : U → U is a measurable
transformation, then M(U, R) denotes the set of measures µ ∈ M(U)
that are invariant under R. If h : U → V is measurable, then we define
the “push-forward” map mh (·) : M(U) → M(V) by mh (η) = η ◦ h−1 .
If U is a non-empty compact metric space and R is continuous, then we
refer to (U, R) as a topological dynamical system. It is well-known that
in this case, M(U, R) is non-empty and compact in the weak topology
(see [58]).
7.2. Product spaces and the shift map. The canonical projections
of a product space U × V onto its constituent sets will be denoted by
projU and projV . If λ is a measure on U × V, then its marginal distributions on U and V will be denoted by mU (λ) and mV (λ), respectively. In several places throughout the paper we will consider infinite
product spaces of the form U N , where U is a Polish space. In each
case U N is endowed with its product topology and associated Borel
sigma-field; elements of U N are denoted as sequences u = (ui )i≥0 . For
any product space U N the left-shift map τ : U N → U N is defined by
τ (u0 , u1 , . . .) = (u1 , u2 , . . .). Note that τ is continuous in the product topology. For any sequence (ul )l≥0 and 0 ≤ i ≤ j, we define
uji = (ui , . . . , uj ).
7.3. The process generated by a measure-preserving system.
Let the dynamical systems (X , S) and (Y, T, ν) satisfy the standard
assumptions (stated in Section 2). By definition of the left shift τ , any
probability measure ν̃ ∈ M(Y N , τ ) is the distribution of a one-sided
stationary process with values in Y. We will say that a measure ν̃ ∈
M(Y N , τ ) is generated by the system (Y, T, ν) if the one-dimensional
marginal distribution of ν̃ is ν, and if ν̃ is supported on trajectories of
T in the sense that
(7.1)
ν̃ {y ∈ Y N : yi+1 = T yi for i ≥ 0} = 1.
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 19
The following technical lemma, which relates J (S : ν̃) to J (S : ν),
will be used in several proofs.
Lemma 7.1. If ν̃ ∈ M(Y N , τ ) is generated by the system (Y, T, ν),
then mX ×Y (J (S : ν̃)) = J (S : ν).
Proof. Let ν̃ ∈ M(Y N ) be generated by (T, ν). Let λ̃ ∈ M(X × Y N )
have marginal distribution ν̃ on M(Y N ) and marginal distribution η
on X × Y. Let f : X × Y N → R be bounded and measurable, and
define an associated bounded, measurable function f0 : X × Y → R by
f0 (x, y) = f (x, (y, T y, . . .)). Using (7.1), one may verify that
Z
Z
(7.2)
f dλ̃ =
f0 dη,
and for λ̃-almost every (x, y),
(7.3)
f ◦ (S × τ )(x, y) = f0 (Sx, y1 ) = f0 ◦ (S × T )(x, y0).
Now suppose that λ̃ ∈ J (S : ν̃) has marginal distribution η on
X × Y. Let h : X × Y → R be a bounded measurable function, and
define f (x, y) = h(x, y0 ). It then
R follows from (7.2),
R (7.3), and the
invariance of λ̃ under S × τ that h ◦ (S × T ) dη = h dη. As h was
arbitrary, η is S × T invariant. Furthermore, it is easy to see that the
Y-marginal of η is ν. Thus, η ∈ J (S : ν).
To establish the other direction, suppose that η ∈ J (S : ν), and
define λ̃ ∈ M(X ×Y N ) to be the distribution of (X, Y0 , T Y0, . . .), where
(X, Y0 ) ∼ η. It is clear that mX ×Y (λ̃) = η and that the marginal
distribution ν̃ of λ̃ on Y N is generated by (T, ν). Moreover, it follows
from (7.2), (7.3), and the invariance of η under S ×T that λ̃ is invariant
under S × τ . Thus, λ̃ ∈ J (S : ν̃), as desired.
7.4. A genericity lemma. The following lemma is standard when U
is compact (e.g., see [10]). One may reduce the more general case
of interest here to the compact case using the regularity of µ, the
separability of C(K) for any compact K, and the pointwise ergodic
theorem. As the argument is straightforward, we omit the proof.
Lemma 7.2. Suppose U is a Polish space, equipped with the Borel σalgebra, R : U → U is measurable, and η is a Borel probability measure
on U that is ergodic and invariant with respect to R. Then there exists
a measurable set E ⊂ U such that η(E) = 1 and if x is in E, then for
each bounded continuous function f : U → R,
Z
n−1
1X
k
lim
f ◦ R (x) = f dη.
n n
k=0
20
KEVIN MCGOFF AND ANDREW B. NOBEL
8. The Tracking Theorem
This section is devoted to the proof of Theorem 2.3. We first establish
the finiteness of the optimal loss L(S : ν). Recall that under the
standard assumptions ℓ∗ ∈ L1 (ν) is a measurable upper bound on
supx |ℓ(x, y)|.
Lemma 8.1. Under the standard assumptions, L(S : ν) ∈ (−∞, ∞).
Proof. As X is non-empty and compact and S is continuous, there
exists at least one measure µ ∈ M(X , S). Thus µ ⊗ ν is in J
R (S : ν),
Rand∗ in particular, J (S : ν) is non-empty. By assumption, | ℓ dλ| ≤
ℓ dν < ∞, for each λ ∈ J (S : ν), and as this bound is independent
of λ, the lemma follows.
The proof of Theorem 2.3 relies on Kingman’s subadditive ergodic
theorem [26, 27, 28] and a weak compactness argument. We first establish the result when T is continuous and then deduce the general
result from this special case.
Proof of Theorem 2.3. We begin by establishing that, for ν-almost
every y,
(8.1)
Z
n−1
n−1
X
1X
1
k
k
k
k
lim inf
ℓ(S x, T y) = sup
ℓ S x, T y dν(y).
inf
n x∈X n
x∈X
n n
k=0
k=0
For each n ∈ N and y ∈ Y, define
Gn (y) = inf
x∈X
n−1
X
ℓ(S i x, T i y).
i=0
Note that the sequence (Gn )n≥1 is super-additive in the sense that
Gm+n (y) ≥ inf
x∈X
m−1
X
i
i
ℓ(S x, T y) + inf
i=0
x∈X
m+n−1
X
ℓ(S i x, T i y)
i=m
≥ Gm (y) + Gn (T m y).
By Kingman’s subadditive ergodic theorem applied to (−Gn )n≥1 , there
exists γ ∈ (−∞, ∞] such that for ν-almost every y,
Z
1
Gn (y)
Gn dν.
= γ = sup
(8.2)
lim
n
n
n n
This equation establishes the existence of the limit in (2.1) and the
equality in (8.1).
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 21
We now establish that L(S : ν) = γ. Let λ be any element of
J (S : ν). As λ is invariant under S × T , for each n ≥ 1,
Z
Z n−1
Z
1 X
1
i
i
(8.3)
ℓ dλ =
Gn dν.
ℓ(S x, T y) dλ ≥
n
n
i=0
R
It then follows from (8.2) that ℓ dλ R≥ γ. As λ ∈ J (S : ν) was
arbitrary, we conclude that γ ≤ inf λ ℓ dλ = L(S : ν), where the
infimum is taken over λ in J (S : ν).
To complete the proof,
we establish the existence of a joining λ ∈
R
J (S : ν) such that ℓ dλ ≤ γ. To do this, we construct a suitable
sequence of empirical measures on X × Y and then use a weak compactness argument to identify a limit λ with the desired properties.
Assume for the moment that T is continuous.
For each m ≥ 1, let Km ⊂ Y be a compact set such that ν(Km ) > 1−
1
.
Using the arguments above, Lemma 7.2, and the ergodic theorem,
m
one may identify a measurable set E ⊆ Y such that ν(E) = 1 and for
every y ∈ E, Equation (8.2) and each of the following relations holds
as n tends to infinity:
n−1
1X
δT k y converges weakly to ν;
n k=0
(8.4)
νn :=
(8.5)
νn (Km ) → ν(Km ) for each m ≥ 1;
Z
Z
∗
ℓ dνn →
ℓ∗ dν for each m ≥ 1.
(8.6)
ℓ∗ >m
ℓ∗ >m
Elements of the set E will be referred to as ν-generic points.
Let y be a ν-generic pointPin Y. By (8.2), there exists a sequence
n−1
c(S k xn , T k y) → γ. For each n ≥ 1,
(xn )n≥1 in X such that n−1 k=0
define the discrete measure
n−1
1X
λn =
δ k
k
n k=0 (S xn , T y)
R
on X × Y. Note that limn ℓ dλn = γ. We claim that the family
{λn : n ∈ N} is tight. To this end, let δ > 0 be given and choose
N > 1/δ. By definition, KN is compact and ν(KN ) > 1 − δ. As y ∈ E,
for all n sufficiently large, λn (X × KN ) = νn (KN ) > 1 − δ. As δ > 0
was arbitrary and X is compact, the claim follows.
Let λ be a weak limit of the family {λn : n ∈ N}. We claim that
λ is in J (S : ν) and that λ achieves the infimum in the definition of
L(S : ν) (see (2.1)). By passing to a subsequence if necessary, assume
22
KEVIN MCGOFF AND ANDREW B. NOBEL
that λn ⇒ λ. Let f be in Cb (X × Y). Under the assumption that S
and T are continuous, the composition f ◦ (S × T ) is in Cb (X × Y),
and weak convergence then implies that
Z
Z
n
1X
f (S i xn , T i y)
f ◦ (S × T ) dλ = lim f ◦ (S × T ) dλn = lim
n n
n
i=1
Z
Z
n−1
1X
i
i
= lim
f (S xn , T y) = lim f dλn = f dλ.
n n
n
i=0
As f ∈ Cb (X × Y) was arbitrary, it follows that λ is invariant under
S × T . In particular, λ ∈ M(X × Y, S × T ) and therefore mX (λ) is in
M(X , S). Furthermore, as y ∈ E, mY (λn ) = νn converges weakly to
ν, and therefore mY (λ) = ν. Thus, λ is in J (S : ν).
R
In light of the fact that limn ℓ dλn = γ, it suffices to show that
Z
Z
(8.7)
lim ℓ dλn ≥ ℓ dλ.
n
If the loss ℓ were bounded from below, this inequality would follow
from the Portmanteau Theorem for weak convergence, since we have
assumed that it is lower semicontinuous. For unbounded losses, we
appeal to a truncation argument. Though the details are somewhat
routine, we include them here for completeness. For m ∈ N, define the
truncated loss
ℓ(x, y), if |ℓ(x, y)| ≤ m
m,
if ℓ(x, y) ≥ m
ℓm (x, y) =
−m,
if ℓ(x, y) ≤ −m.
Note that |ℓm | ≤ |ℓ| and that ℓm → ℓ as m tends to infinity. The
integrability of ℓ with respect to λ follows from that of ℓ∗ with reRspect to ν, Rand the dominated convergence theorem then ensures that
ℓm dλ → ℓ dλ. Moreover, with νn defined as in (8.4), it follows from
the choice of y that
Z
Z
Z
∗
ℓ dνn =
ℓ∗ dν.
lim sup |ℓ − ℓm | dλn ≤ lim sup
n
n
ℓ∗ >m
ℓ∗ >m
In order to establish (8.7), let ǫ > 0 be fixed. By virtue of the results
in the previous paragraph, there exist integers m and n1 sufficiently
large that for each n ≥ n1
Z
Z
Z
|ℓ − ℓm | dλn < ǫ/3.
ℓm dλ − ℓ dλ < ǫ/3 and
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 23
Moreover, as λn ⇒ λ and ℓm is lower semi-continuous and bounded,
there exists n2 ≥ n1 such that for each n ≥ n2 ,
Z
Z
ℓm dλ − ℓm dλn < ǫ/3.
Combining the inequalities above, a straightforward bound shows that
Z
Z
ℓ dλ − ℓ dλn < ǫ
for n > n2 . As ǫ > 0 was arbitrary, the inequality (8.7) is established,
and we conclude that γ ≥ L(S : ν).
Suppose now that the transformation T is Borel measurable but not
continuous. Let ν̃ be the process measure on Y N generated by (Y, T, ν)
(see Section 7.3), and let ℓ̃ : X × Y N → R be defined by ℓ̃(x, y) =
ℓ(x, y0 ). Note that ℓ̃ is lower semicontinuous and that supx |ℓ̃(x, y)| is
bounded above by a ν̃-integrable function. As the left-shift τ : Y N →
Y N is continuous, we may apply the arguments above to the systems
(X , S) and (Y N , τ, ν̃) with loss ℓ̃. Equation (8.2) and inequality (8.3)
are the same for the original and shift systems. As for the inequality
γ ≥ L(S : ν), the arguments
that there is a joining λ̃ ∈
R
R above show
J (S : ν̃) such that γ = ℓ̃ dλ̃ = ℓ dmX ×Y (λ̃). By Lemma 7.1,
λ = mX ×Y (λ̃) is in J (S : ν). This establishes (2.1) and the existence
of a joining λ in J (S : ν) that achieves the infimum in the definition
of L(S : ν).
9. General results for inference
The present section is devoted to the proof of Theorem 3.2. We
begin with several preliminary lemmas, the first of which establishes
the existence of optimal tracking schemes. To ease the notation in the
proof of Theorem 3.2 below, we replace the functions xn in Definition
3.1 with functions fn , as in the following lemma.
Lemma 9.1. If the systems (X , S) and (Y, T, ν) and the loss ℓ satisfy
the standard assumptions of Section 1, then there exists a measurable
sequence of functions fn : Y n → X satisfying (3.1).
Proof. For each n ≥ 1, define ℓn : X × Y n → R by ℓn (x, y0n−1 ) =
Pn−1
k
k=0 ℓ(S x, yk ). Then it is easy to see that ℓn = sn ◦ ψn , where
ψn (x, y0n−1) = ((x, . . . , S n−1x), y0n−1 )
Pn−1
and sn (x0n−1 , y0n−1) =
k=0 ℓ(xk , yk ). Our assumptions on S and ℓ
ensure that ψn is continuous and that sn is lower semicontinuous, and
therefore ℓn is lower semicontinuous. It follows from [4, Proposition
24
KEVIN MCGOFF AND ANDREW B. NOBEL
7.33, p. 153] that there exists a Borel measurable function fn : Y n → X
such that
ℓn (fn (y0n−1), y0n−1) = inf ℓn (x, y0n−1).
x∈X
The definition of ℓn ensures that (fn )n≥1 satisfies the conclusions of the
lemma.
Lemma 9.2. Under the standard assumptions,
(9.1)
L(S : ν) = inf L(Sθ : ν)
θ∈Θ
where Sθ is the restriction of S to Xθ = ϕ−1 {θ}. Furthermore, the
infimum is attained.
Proof. Since (Xθ , Sθ ) is a subsystem of (X , S), it is immediate that
L(S : ν) ≤ L(Sθ : ν) for each θ ∈ Θ. Thus, it suffices to show
that L(Sθ : ν) ≤ L(S : ν) for some θ ∈ Θ. By Remark 2.6, there
exists an ergodic joining λ in Jmin (S : ν). Define h : X × Y → Θ
by h = ϕ ◦ projX and let η = λ ◦ h−1 be the push-forward measure
of λ on Θ. The assumption that ϕ ◦ S = ϕ ensures that η is ergodic
with respect to the identity transformation on Θ, and therefore η is
necessarily a point mass concentrated at some parameter θ ∈ Θ. In
particular,
λ(Xθ × Y) = 1, so that λ ∈ J (Sθ : ν). Thus L(Sθ : ν) ≤
R
ℓ dλ = L(S : ν), and the result follows.
Lemma 9.3. The set Θmin ⊆ Θ is non-empty and compact.
Proof. As the infimum in (9.1) is achieved, Θmin is non-empty. Since Θ
is compact by assumption, it suffices to show that Θmin is closed. Let
(θn )n≥1 be a sequence in Θmin that converges to a parameter θ ∈ Θ. It
follows from (3.3) that for each n, there is a joining λn ∈ Jmin (S : ν)
such that λn (ϕ−1 {θn } × Y) = 1. As Jmin (S : ν) is compact, the sequence (λn )n≥1 has a convergent subsequence. Passing to a subsequence
if necessary, suppose that λn converges to λ ∈ Jmin (S : ν).
Define h : X × Y → Θ by h = ϕ ◦ projX , and consider the associated
push-forward measures ηn = λn ◦ h−1 and η = λ ◦ h−1 on Θ. Note that
ηn ⇒ η, as λn ⇒ λ and h is continuous. Our choice of (θn ) ensures
that ηn = δθn ⇒ δθ , and as weak limits are unique, η = δθ . Thus
λ(ϕ−1 {θ} × Y) = 1, and as λ ∈ Jmin (S : ν) we conclude that θ is an
element of Θmin .
In the following proof, we make use of the ergodic decomposition of
an invariant measure and a related lemma, details of which may be
found in Appendix A. Note that for the sake of notation in the proof,
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 25
we denote the sequence of estimators by fn : Y n → X instead of by
xn : Y n → X .
Proof of Theorem 3.2. Let (θn )n≥1 be a sequence of estimators of
the form θn = ϕ ◦ fn , where {fn } is a sequence of measurable functions
satisfying (3.1), which exists by Lemma 9.1. As in the proof of Theorem
2.3, fix a set E ⊂ Y of ν-generic points having full measure. Let
y ∈ E, and then let θ̂n = θn (y, . . . , T n−1y). For n ≥ 1, define the
statePxn = fn (y, . . . , T n−1 y) and the associated empirical measure λn =
n−1
δ(S k xn ,T k y) on X × Y. By arguments identical to those in the
n−1 k=0
proof of Theorem 2.3, one may show that (λn )n is tight and that all of
its weak limit points are in Jmin (S : ν).
Let O ⊆ Θ be an open neighborhood of Θmin . Define the function
ψ : M(X ×Y) → [0, 1] by ψ(λ) = λ(ϕ−1 (O)×Y), and let V = ψ −1 ( 21 , 1].
As ϕ−1 (O) × Y is open, ψ is lower semi-continuous, and therefore V
is open in M(X × Y). We claim that Jmin (S : ν) ⊂ V . To see this,
first consider an ergodic optimal joining η ∈ Jmin (S : ν). Define the
factor map h : X × Y → Θ by h = ϕ ◦ projX . Since η is ergodic,
the push-forward measure η ◦ h−1 must be ergodic with respect to the
identity map on Θ, as in the proof of Lemma 9.2, and hence η◦h−1 = δθ
for some θ ∈ Θ. Thus, for every ergodic η there is a parameter θ ∈
Θmin such that η(ϕ−1{θ} × Y) = 1, and therefore η(ϕ−1 (Θmin ) × Y) =
1. Now consider anR arbitrary joining λ ∈ Jmin (S : ν) with ergodic
decomposition λ = η dξ. By Lemma A.3, ξ-almost every measure η
is an ergodic element of Jmin (S : ν). Thus, the above argument gives
that for ξ-almost every η, we have η(ϕ−1(Θmin ) × Y) = 1. It follows
that ψ(λ) = 1, and we conclude that λ is in V .
Since all the limit points of the family {λn }n≥1 are in the open
set V , there exists an integer n1 such that λn ∈ V for all n ≥ n1 .
By construction, λn (ϕ−1 (O) × Y) = δθ̂n (O). Thus for each n ≥ n1 ,
δθ̂n (O) = ψ(λn ) > 1/2, which implies that θ̂n is in O as desired.
Let us now show that any parameter in Θmin is the limit of an optimal
ϕ-estimation scheme. Let θ0 be any element of Θmin . We will show
that there exists a ϕ-optimal estimation scheme that converges almost
surely to θ0 . By Lemma 9.2, L(S : ν) = L(Sθ0 : ν), and it follows from
Remark 2.6 that there exists an ergodic joining λ ∈ Jmin (Sθ0 : ν). By
Birkhoff’s ergodic theorem, there exists a set E ⊂ Xθ × Y of λ measure
one such that if (x, y) ∈ E then
n−1
(9.2)
1X
lim
ℓ(S k x, T k y) =
n n
k=0
Z
ℓ dλ,
26
KEVIN MCGOFF AND ANDREW B. NOBEL
which is equal to L(S : ν) by construction. By the regularity of λ,
E contains a σ-compact set F of the same measure. The measurable
selection theorem of Brown and Purves [8, Theorem 1] then implies that
there is a measurable function f : Y → Xθ0 such that (f (y), y) ∈ F
for ν-almost every y ∈ Y. Let fn : Y n → X be given by fn (y0n−1) =
f (y0). By (9.2), the sequence {fn }n≥1 satisfies (3.1), and therefore the
constant estimator θ0 = ϕ ◦ fn is an optimal ϕ-estimation scheme. In
particular, there exists an optimal ϕ-estimation scheme that converges
to θ0 , as desired.
10. Quantized observations
This section is devoted to the proofs of our results concerning the
estimation of transformations from quantized observations. We refer
to the objects defined in Section 5 throughout this section.
Recall that the state space X of the reference system is defined to
be the closure of the set {(θ, lab(θ, u)) : θ ∈ Θ, u ∈ U} inside of
Θ × {0, 1}N . Hence X is compact and metrizable, and it is easy to see
that X is invariant under the map Id ×τ , where τ is the left-shift map
on {0, 1}N . Further recall that the family of transformations in R and
the topology on Θ enter indirectly through the definition of X , which
is used to define the target set Θ1 .
First we establish some additional notation and another interpretation of h(R), as the topological entropy of a dynamical system. Let L
be defined by
L = closure of (π(Rθk u))k≥0 : θ ∈ Θ, u ∈ U in {0, 1}N .
Since L is a closed subset of {0, 1}N , it is compact, and it is easy to
see that L is invariant under the left shift τ . Thus the pair (L, τ ) is a
topological dynamical system. Now we may observe that the quantity
h(R) defined in Section 5 is actually the topological entropy of the
system (L, τ ). See [58] for an introduction to topological entropy for
dynamical systems.
In what follows, we find it useful to state our results in terms of the
d-distance introduced by Ornstein [45, 47, 48] in the context of the
isomorphism theory for Bernoulli processes. For a fixed shift-invariant
measure ν on {0, 1}N and each θ ∈ Θ, define
d(θ : ν) = inf d(µ, ν),
µ
where the infimum is taken over all τ -invariant probability measures
µ such that µ(Xθ ) = 1. An application of Theorem 3.2 yields the
following result.
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 27
Theorem 10.1. Let (Bi )i≥0 be a stationary ergodic {0, 1}-valued process having distribution ν on {0, 1}N . If the sequence of estimators
(θn )n satisfies (5.2), then θ̂n (B0 , . . . , Bn−1 ) converges almost surely to
the set
Θ0 = argmin d(θ : ν).
θ∈Θ
Proof. By definition, Θmin = argminθ L(Sθ : ν). Using the correspondences in Table 3, it is easy to see that Sθ is the restriction of Id ×τ to
the set {θ} × Xθ , and therefore
Z
L(Sθ : ν) = inf
ℓ dλ =
inf
d(µ, ν) = d(θ : ν).
λ∈J(Sθ :ν)
µ∈M(Xθ ,τ )
Thus Θmin = argminθ d(θ : ν) = Θ0 , as desired.
By associating parameters θ with the set of τ -invariant measures on
Xθ , we may view the set Θ0 as the projection of the observation process
onto the parameter set Θ under the d-metric. The following lemma,
which characterizes when d(θ : η) equals zero, will be used to prove our
consistency results.
Lemma 10.2. If ν is a shift-invariant Borel probability measure on
{0, 1}N , then d(θ : ν) = 0 if and only if ν(Xθ ) = 1.
Proof. If ν(Xθ ) = 1, then clearly 0 ≤ d(θ : ν) ≤ d(ν, ν) = 0. Now
suppose that d(θ : η) = 0. By Theorem 2.3, the infimum defining
d(θ : η) = L(Sθ : η) is achieved, and it follows that there is a measure
µ ∈ M(Xθ , τ ) such that d(µ, ν) = 0. As d is a metric, µ = ν and
therefore ν(Xθ ) = 1.
Proof of Theorem 5.1, Part (1). By Theorem 10.1, the estimates
θ̂n (Y0 , . . . , Yn−1) converge almost surely to the set Θ0 = argminθ∈Θ d(θ :
ν0 ). Thus it suffices to show that Θ0 = Θ1 . As the measure ν0 of the
observation process is supported on Xθ0 , we have d(θ0 : ν0 ) = 0. Then
by Lemma 10.2, we see that Θ0 = {θ : d(θ : ν0 ) = 0} is identical to
Θ1 = {θ : ν0 (Xθ ) = 1}. The converse part of Theorem 5.1 is a direct
consequence of the converse part of Theorem 2.3.
Before turning to the proof of Part (2) of Theorem 5.1, we require
some additional definitions.
Remark 10.3. In the proof of Theorem 5.1 below, we make use of a standard construction, called the relatively independent joining, to “glue
together” two joinings along a common factor (see [9, Section 3.1]). In
more detail, suppose we have two measure-preserving system (Ui , Ri , ηi )
28
KEVIN MCGOFF AND ANDREW B. NOBEL
for i = 1, 2. Further suppose that these systems have a common factor (U, R, η), meaning that there exist measurable maps ψi : Ui → U
such Rthat η = ηi ◦ ψi−1 and ψi ◦ Ri = R ◦ ψi for i = 1, 2. Now let
ηi = ηi,u dη(u) be the disintegration of ηi over η, and define the measure λ on U1 × U2 by
Z
λ = η1,u ⊗ η2,u dη(u).
Then it is not difficult to
check that λ is a joining of (U1 , R1 , η1 ) and
(U2 , R2 , η2 ) such that λ (u1, u2 ) ∈ U1 × U2 : ψ1 (u1 ) = ψ2 (u2 ) = 1.
Remark 10.4. The proof of Theorem 5.1 also requires some elementary facts concerning the entropy of dynamical systems (see [58] for a
thorough treatment of the subject). Let Σ ⊂ {0, 1}N be a closed, shiftinvariant set of label sequences. The topological entropy htop (Σ, τ ) of
the system (Σ, τ ) is given in (5.4). If η is a shift-invariant Borel probability measure on such a set Σ, then the measure theoretic (KolmogorovSinai) entropy of the system (Σ, τ, η) is defined by
X
1
η([an0 ]) log η([an0 ]),
h(η) = lim −
n
n + 1 an ∈Σ
0
n
where Σn = {an0 ∈ {0, 1}n+1 : a ∈ Σ} and [an0 ] denotes the cylinder
set of sequences a ∈ {0, 1}N whose first n + 1 coordinates are an0 . The
well-known Variational Principle (see [58]) states that
htop (Σ, τ ) =
sup
h(η).
η∈M(Σ,τ )
Thus if htop (Σ, τ ) = 0, then h(η) = 0 for any measure η in M(Σ, τ ).
Proof of Theorem 5.1, Part (2). By Theorem 10.1, any sequence
of estimates satisfying (5.2) converges almost surely to the parameter
set Θ0 = argminθ∈Θ d(θ : ν), where ν is the measure of the observation
process (Yi )i≥0 , which involves errors. By contrast, we have Θ1 = {θ ∈
Θ : ν0 (Lθ ) = 1}, where ν0 is the measure for the error-free process
(π(Rθi 0 U))i≥0 . It therefore suffices to show that θ minimizes d(θ : ν) if
and only if ν0 (Xθ ) = 1. Additionally, with this relation established, the
converse part of Theorem 5.1 is a direct consequence of the converse
part of Theorem 2.3.
Let p = P ǫ0 = 1 , which is less than 1/2 by hypothesis. We claim
that for each θ in Θ,
(10.1)
d(θ : ν) ≥ p + d(θ : ν0 ) (1 − 2p).
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 29
To see this, fix θ ∈ Θ. Then Theorem 2.3 with (X , S) = (Xθ , τ ) ensures
that there is an optimal joining λ1 of θ and ν. In detail, there exists
λ1 ∈ M(Xθ × {0, 1}N , τ × τ ) such that its first marginal, µ say, is
supported on Xθ , its second marginal is equal to ν, and d(θ : ν) =
λ1 ({(a, b) : a0 6= b0 }).
Let η be the process measure for the noise process (ǫi )i≥0 , and let
λ2 denote the product measure ν0 ⊗ η. Define the maps ψ1 : {0, 1}N ×
{0, 1}N → {0, 1}N and ψ2 : {0, 1}N × {0, 1}N → {0, 1}N by ψ1 (a, b) = b
and ψ2 (v, ǫ) = (vi ⊕ ǫi )i≥0 . Note that ν is a common factor of λ1 and
λ2 under the maps ψ1 and ψ2 , respectively. Using the construction
discussed in Remark 10.3, one may construct a joining λ̃ of λ1 and λ2
such that
λ̃ (a, b, v, ǫ) : bi = vi ⊕ ǫi for all i ≥ 0
= 1.
Let (Ai , Bi , Vi , ǫi )i≥0 denote the multi-label process having distribution
λ̃. By construction, the following hold:
(1) (Ai )i≥0 ∼ µ and is supported on Xθ ;
(2) (Bi )i≥0 ∼ ν;
(3) (Vi )i≥0 ∼ ν0 ;
(4) (ǫi )i≥0 ∼ η is a copy of the i.i.d. noise process;
(5) (Ai , Bi )i≥0 ∼ λ1 , the joining of µ and ν;
(6) Bi = Vi ⊕ ǫi almost surely.
From these properties and elementary arguments we see that
d(θ : ν) = λ1 (A0 =
6 B0 ) = λ̃(A0 6= (V0 ⊕ ǫ0 ))
≥ λ̃ {ǫ0 = 1, A0 = V0 } ∪ {ǫ0 = 0, A0 6= V0 }
= λ̃ ǫ0 = 1, A0 = V0 + λ̃ ǫ0 = 0, A0 6= V0 .
As the measures µ and ν0 are supported on L, the assumption that
h(R) = 0 implies that h(µ) = 0 and h(ν0 ) = 0 (see Remark 10.4). Let
λ3 be the joining of µ and ν0 given by the marginal distribution of λ̃ on
(a, v). By a standard bound on entropy, h(λ3 ) ≤ h(µ) + h(ν0 ) = 0, and
therefore h(λ3 ) = 0. It follows from a classical result of Furstenberg [14,
Theorem I.2] that the only joining between the zero-entropy measure
λ3 and the i.i.d. measure η is the product (independent) joining λ3 ⊗ η.
Consequently, (A0 , V0 ) and ǫ0 are independent under λ̃. Therefore
λ̃ ǫ0 = 1, A0 = V0 = λ̃ ǫ0 = 1 λ̃ A0 = V0 = p 1 − λ̃(A0 6= V0 ) ,
30
KEVIN MCGOFF AND ANDREW B. NOBEL
and
λ̃ ǫ0 = 0, A0 6= V0 = λ̃ ǫ0 = 0 λ̃ A0 6= V0 = (1 − p) λ̃(A0 6= V0 ).
Combining the previous three displays gives
d(θ : ν) = λ1 A0 6= B0
≥ p 1 − λ̃(A0 6= V0 ) + (1 − p)λ̃(A0 6= V0 )
= p + λ̃(A0 6= V0 )(1 − 2p).
Under the joining λ̃, (Ai )i≥0 is distributed according to µ, which is
supported on Xθ , and (Vi )i≥0 is distributed according to ν0 . Thus
λ̃(A0 6= V0 ) ≥ d(θ : ν0 ), and the inequality (10.1) follows from the
previous display as p < 1/2.
It follows from (10.1) that d(θ : ν) ≥ p for all θ. We now show
that d(θ0 : ν) = p, from which it follows that Θ0 = {θ : d(θ : ν) =
p}. Let λ = ν0 ⊗ η ∈ M({0, 1}N × {0, 1}N ) and let ψ : {0, 1}N ×
{0, 1}N → {0, 1}N × {0, 1}N be defined by ψ(a, ǫ) = (a, (ai ⊕ ǫi )i≥0 ). It
is straightforward to show that λ4 = λ ◦ ψ −1 is a joining of ν0 with ν
such that λ4 ({(a, b) : a0 6= b0 }) = p, and therefore d(θ0 : ν) = p as
desired.
Let θ ∈ Θ0 . The arguments above show that d(θ : ν) = p, and then
it follows from (10.1) that d(θ : ν0 ) = 0. Hence ν0 (Xθ ) = 1 by Lemma
10.2, and we conclude that θ ∈ Θ1 . This shows that Θ0 ⊆ Θ1 . For
the reverse inclusion, we note that if ν0 (Xθ ) = 1, then (10.1) and the
joining λ4 can be used to show that d(θ : ν) = p (as we did for θ0 ),
which implies that θ is in Θ0 .
If the partition π does not not resolve differences between the generative transformation Rθ0 and other transformations on the support
of ν0 , then Θ1 may not be equal to {θ0 }. The following result provides
conditions under which Θ1 is a singleton.
Proposition 10.5. Suppose that for all θ′ 6= θ0 there exists a neighborhood O of θ′ and an integer N depending on θ0 and O such that for
all u, v ∈ U and all θ ∈ O, π(Rθk0 u) 6= π(Rθk v) for some k ∈ [0, N].
Then Θ1 = {θ0 }.
Proof. Suppose the hypotheses of the proposition hold, and let θ′ 6= θ0 .
/ Θ1 . Let O and N be
We will show that ν0 (Xθ′ ) = 0, and therefore θ′ ∈
as in the statement of the proposition. For θ ∈ Θ and u ∈ U define
C N (θ, u) = {a ∈ {0, 1}N : ak = π(Rθk u) for 0 ≤ k ≤ N},
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 31
and for Θ′ ⊆ Θ let C N (Θ′ ) be the union of C N (θ, u) over θ ∈ Θ′ and
u ∈ U. The cylinder sets C N (θ, u) are closed and open, and since N is
fixed, there are finitely many of them. As O is an open neighborhood
of θ′ it is clear that Xθ′ ⊂ C N (O). Moreover, as ν0 is supported on the
set of label sequences generated by θ0 and C N (θ0 ) is closed, supp(ν0 ) ⊂
C N (θ0 ). The hypotheses of the proposition imply that that C N (O) and
C N (θ0 ) are disjoint, and therefore Xθ′ is disjoint from supp(ν0 ).
Proposition 10.6. Under the hypotheses in Section 5.0.1, we have
Θ1 = {α0 }.
Proof. Recall that Θ = [0, 12 ]. We wish to apply Proposition 10.5. Let
α1 6= α2 and assume without loss of generality that α1 < α2 . Fix 0 <
ǫ < (α2 −α1 )/2 and let O = [α2 −ǫ, α2 +ǫ]∩Θ and N ≥ 3/2(α2 −α1 −ǫ).
Let u, v ∈ [0, 1) and α ∈ O. Define
1
.
k = inf j ≥ 0 : |(u + jα) − (v + jα1 )| ≥
2
Our choice of N ensures that N(α−α1 ) ≥ 3/2 ≥ 1/2+v −u. Therefore
1
u + Nα − (v + Nα1 ) ≥ ,
2
k
so that k ≤ N. We claim that π(Rα u) 6= π(Rαk 1 v). If k = 0 then
1
≤ |u − v| < 1, which implies that π(u) 6= π(v). Suppose that k ≥ 1.
2
Using the definition of k (twice) and the triangle inequality, we see that
1
≤ |(u + kα) − (v + kα1 )|
2
≤ |(u + (k − 1)α) − (v + (k − 1)α1 )| + |α − α1 |
1 1
< + = 1,
2 2
and therefore π(Rαk u) 6= π(Rαk 1 v). Now Proposition 10.5 yields the
result.
Proof of Proposition 5.3. We first show that h(R) = 0. For u in
[0, 1) and α in Θ, let π n (u, α) denote the element w = w0 . . . wn−1 of
{0, 1}n such that Rαk u ∈ Awk for k = 0, . . . , n − 1. Define
n
n
C(n) = w ∈ {0, 1} : ∃u ∈ [0, 1), ∃α ∈ Θ, w = π (u, α) ,
and note that
h(R) ≤ lim sup
n
1
log C(n).
n
32
KEVIN MCGOFF AND ANDREW B. NOBEL
It is known (see [13]) that C(n) ≤ Kn4 for some constant K, and it
then follows from the previous display that h(R) = 0. By Theorem
5.1, any estimates satisfying (5.2) converge almost surely to Θ1 , and
Θ1 = {α0 } by Proposition 10.6.
Acknowledgements
The authors thank Sayan Mukherjee and Karl Petersen for productive conversations. Both authors acknowledge the support of the National Science Foundation under grant DMS 1613261, and AN also
acknowledges support under grant DMS 1310002.
References
[1] Terrence M Adams and Andrew B Nobel. Uniform convergence of Vapnik–
Chervonenkis classes under ergodic sampling. The Annals of Probability,
38(4):1345–1367, 2010.
[2] Pierre Alquier and Olivier Wintenberger. Model selection for weakly dependent
time series forecasting. Bernoulli, 18(3):883–913, 2012.
[3] Andrew F Bennett. Inverse modeling of the ocean and atmosphere. Cambridge
University Press, 2005.
[4] Dimitri P Bertsekas and Steven E Shreve. Stochastic Optimal Control: The
Discrete–Time Case. Athena Scientific, 1996.
[5] Christopher Bose and Rua Murray. The exact rate of approximation in ulam’s
method. Discrete and continuous dynamical systems, 7(1):219–235, 2001.
[6] Chris A Brackley, Oliver Ebenhöh, Celso Grebogi, Jürgen Kurths, Alessandro
de Moura, M Carmen Romano, and Marco Thiel. Introduction to focus issue:
dynamics in systems biology, 2010.
[7] Leo Breiman. Probability. Society for Industrial and Applied Mathematics
(SIAM), Philadelphia, PA, 1992.
[8] L. D. Brown and R. Purves. Measurable selections of extrema. Ann. Statist.,
1(5):902–912, 09 1973.
[9] Thierry de la Rue. An introduction to joinings in ergodic theory. Discrete and
Continuous Dynamical Systems, 15(1):121–142, 2006.
[10] Manfred Denker, Christian Grillenberger, and Karl Sigmund. Ergodic theory
on compact spaces. Springer-Verlag, 1976.
[11] Gouri Dhatt, Emmanuel LefranÃ, and Gilbert Touzot. Finite element method.
John Wiley & Sons, 2012.
[12] Rick Durrett. Probability: theory and examples. Cambridge University Press,
2010.
[13] Frid, A. and Jamet, D. The number of binary rotation words. RAIRO-Theor.
Inf. Appl., 48(4):453–465, 2014.
[14] Harry Furstenberg. Disjointness in ergodic theory, minimal sets, and a problem
in diophantine approximation. Theory of Computing Systems, 1(1):1–49, 1967.
[15] Eli Glasner. Ergodic theory via joinings. American Mathematical Soc., 2003.
[16] Robert M Gray. Entropy and information theory. Springer Science & Business
Media, 2011.
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 33
[17] Robert M Gray, David L Neuhoff, and Paul C Shields. A generalization of
Ornstein’s d-bar distance with applications to information theory. The Annals
of Probability, pages 315–328, 1975.
[18] Yonatan Gutman and Michael Hochman. On processes which cannot be distinguished by finite observation. Israel Journal of Mathematics, 164(1):265–284,
2008.
[19] H Hang and I Steinwart. A Bernstein-type inequality for some mixing processes
and dynamical systems with an application to learning. Ann. Statist., 2016.
[20] H Hang, I Steinwart, Y Feng, and J Suykens. Kernel density estimation for
dynamical systems. arXiv:1607.03792, 2016.
[21] Oliver Jenkinson. Ergodic optimization. Discrete and Continuous Dynamical
Systems, 15(1):197, 2006.
[22] Eugenia Kalnay. Atmospheric modeling, data assimilation and predictability.
Cambridge university press, 2003.
[23] Guy Karlebach and Ron Shamir. Modelling and analysis of gene regulatory
networks. Nature Reviews Molecular Cell Biology, 9(10):770–780, 2008.
[24] Paul C. Shields Katalin Marton. Entropy and the consistent estimation of joint
distributions. The Annals of Probability, 22(2):960–977, 1994.
[25] A. Katok and B. Hasselblatt. Introduction to the modern theory of dynamical
systems. Cambridge University Press, 1995.
[26] JFC Kingman. The ergodic theory of subadditive stochastic processes. Journal
of the Royal Statistical Society. Series B (Methodological), pages 499–510, 1968.
[27] JFC Kingman. Subadditive ergodic theory. The Annals of Probability, pages
883–899, 1973.
[28] JFC Kingman. Subadditive processes. In Ecole d’Eté de Probabilités de SaintFlour V-1975, pages 167–223. Springer, 1976.
[29] Nicolas Kryloff and Nicolas Bogoliouboff. La théorie générale de la mesure
dans son application à l’étude des systèmes dynamiques de la mécanique non
linéaire. Annals of Mathematics, pages 65–113, 1937.
[30] Yury A Kutoyants. Statistical inference for ergodic diffusion processes. Springer
Science & Business Media, 2013.
[31] Vitaly Kuznetsov and Mehryar Mohri. Time series prediction and online learning. In 29th Annual Conference on Learning Theory, pages 1190–1213, 2016.
[32] Steven P. Lalley. Beneath the noise, chaos. The Annals of Statistics, 27(2):461–
479, 1999.
[33] Steven P. Lalley and A. B. Nobel. Denoising deterministic time series. Dyn.
Partial Differ. Equ., 3(4):259–279, 2006.
[34] Kody Law, Andrew Stuart, and Kostas Zygalakis. Data Assimilation. Springer,
2015.
[35] Benjamin Letham, Portia A Letham, Cynthia Rudin, and Edward P Browne.
Prediction uncertainty and optimal experimental design for learning dynamical systems. Chaos: An Interdisciplinary Journal of Nonlinear Science,
26(6):063110, 2016.
[36] Simon A Levin, Stephen R Carpenter, H Charles J Godfray, Ann P Kinzig,
Michel Loreau, Jonathan B Losos, Brian Walker, and David S Wilcove. The
Princeton guide to ecology. Princeton University Press, 2009.
34
KEVIN MCGOFF AND ANDREW B. NOBEL
[37] Kevin McGoff, Sayan Mukherjee, Andrew Nobel, and Natesh Pillai. Consistency of maximum likelihood estimation for some dynamical systems. The
Annals of Statistics, 43(1):1–29, 2015.
[38] Kevin McGoff, Sayan Mukherjee, and Natesh Pillai. Statistical inference for
dynamical systems: A review. Statist. Surv., 9:209–252, 2015.
[39] Kevin McGoff and Andrew Nobel. Empirical risk minimization and dynamical
models, 2016. arXiv:1611.06173.
[40] Kevin McGoff and Andrew Nobel. Optimal tracking for dynamical systems,
2016. arXiv:1601.05033v1.
[41] Kevin A McGoff, Xin Guo, Anastasia Deckard, Christina M Kelliher, Adam R
Leman, Lauren J Francey, John B Hogenesch, Steven B Haase, and John L
Harer. The local edge machine: inference of dynamic models of gene regulation.
Genome biology, 17(1):214, 2016.
[42] Konstantin Mischaikow and Marian Mrozek. Conley index. Handbook of dynamical systems, 2:393–460, 2002.
[43] Konstantin Mischaikow, Marian Mrozek, J Reiss, and Andrzej Szymczak. Construction of symbolic dynamics from experimental time series. Physical Review
Letters, 82(6):1144, 1999.
[44] Gusztáv Morvai and Benjamin Weiss. Prediction for discrete time series. Probability theory and related fields, 132(1):1–12, 2005.
[45] Donald Ornstein. Bernoulli shifts with the same entropy are isomorphic. Advances in Mathematics, 4(3):337–352, 1970.
[46] Donald Ornstein and Benjamin Weiss. Entropy is the only finitely observable
invariant. Journal of Modern Dynamics, 1(1):93–105, 2007.
[47] Donald S Ornstein. An application of ergodic theory to probability theory. The
Annals of Probability, 1(1):43–58, 1973.
[48] Donald S Ornstein. Ergodic theory, randomness, and dynamical systems. Yale
University Press, 1974.
[49] Donald S Ornstein and Benjamin Weiss. How sampling reveals a process. The
Annals of Probability, pages 905–930, 1990.
[50] Karl E Petersen. Ergodic theory, volume 2. Cambridge University Press, 1983.
[51] Robert R Phelps. Lectures on Choquet’s theorem, volume 1757. Springer Science & Business Media, 2001.
[52] Anthony Quas and Jason Siefken. Ergodic optimization of super-continuous
functions on shift spaces. Ergodic Theory and Dynamical Systems, 32(06):2071–
2082, 2012.
[53] Omri Sarig. Lecture Notes on Ergodic Theory. 2008.
[54] Ingo Steinwart and Marian Anghel. Consistency of support vector machines
for forecasting the evolution of an unknown ergodic dynamical system from
observations with unknown noise. The Annals of Statistics, pages 841–875,
2009.
[55] Peter Turchin. Complex population dynamics: a theoretical/empirical synthesis, volume 35. Princeton University Press, 2013.
[56] Aad W Van der Vaart. Asymptotic statistics, volume 3. Cambridge University
Press, 2000.
[57] Cédric Villani. Optimal transport: old and new, volume 338. Springer Science
& Business Media, 2008.
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 35
[58] Peter Walters. An introduction to ergodic theory, volume 79. Springer Science
& Business Media, 2000.
[59] A Zimin and C Lampert. Conditional risk minimization for stochastic processes. arXiv:1510.02706, 2016.
Appendix A. Structure of the set of optimal joinings
In this section, we investigate the structure of the set of optimal
joinings in Theorem 2.3 and provide the proof of Theorem 2.5. Our
results rely on a general version of the ergodic decomposition for invariant probability measures. The following version, a restatement of
[53, Theorem 2.5], is sufficient for our purposes.
Theorem (The Ergodic Decomposition). Suppose that R : U → U is
a Borel measurable map of a Polish space U and that λ ∈ M(U, R).
Then there exists a Borel probability measure ξ on M(U) such that
(1) ξ {η is invariant and ergodic for R} = 1
(2) If f ∈ L1 (λ), then f ∈ L1 (η) for ξ-almost every η, and
Z
Z Z
f dλ =
f dη dξ(η).
Whenever (2) holds, we write λ =
R
η dξ.
Remark A.1. Suppose f : U → V is a Borel measurable map
between
R
Polish spaces U and V and that λ ∈ RM(U) satisfies λ = η dξ. Then
one may readily check that mf (λ) = mf (η) dξ.
In the remainder of this section we assume that the systems (X , S)
and (Y, T, ν) and the loss ℓ satisfy the standard assumptions in Section
1. The following three lemmas will be used to prove Theorem 2.5.
R
Lemma A.2. If λ ∈ J (S : ν) has ergodic decomposition λ = η dξ,
then ξ-almost every η is in J (S : ν).
R
Proof. Let λ ∈ J (S : ν) have ergodic decomposition λ = η dξ. Then
ξ-almost every η is in M(X ×Y, S×T ), and for these measures mX (η) is
necessarily
in M(X , S). It follows from Remark A.1 that ν = mY (λ) =
R
mY (η) dξ. Since ν is ergodic, it is an extreme point of the convex set
M(Y, T ) (see [51, Proposition 12.4]), and therefore mY (η) = ν for
ξ-almost every η.
R
Lemma A.3. If λ ∈ Jmin (S : ν) has ergodic decomposition λ = η dξ,
then ξ-almost every η is ergodic and contained in Jmin (S : ν).
36
KEVIN MCGOFF AND ANDREW B. NOBEL
Proof. Ergodicity of ξ-almost every η follows from the definition of the
ergodic decomposition. By assumption, ℓ ∈ L1 (λ), and the ergodic
decomposition yields
Z
Z Z
ℓ dλ =
ℓ dη dξ(η).
R
By
R Lemma A.2, ξ-almost every η ∈ J (S : ν)
R and inRthis case ℓ dλ ≤
ℓ dη, as λ ∈ Jmin (S : ν). It follows that ℓ dλ = ℓ dη for ξ-almost
every η.
Lemma
A.4. The functional φℓ : J (S : ν) → R defined by φℓ (λ) =
R
ℓ dλ is lower semi-continuous.
If the loss ℓ is bounded from below, then the conclusion of this lemma
follows immediately from the Portmanteau Theorem for weak convergence. In the general case, the result may be established by a truncation argument; as the argument is very similar to that in the proof of
Theorem 2.3, we omit the details.
Proposition A.5. The set of measures J (S : ν) is compact in the
weak topology.
Proof. We address the general case of measurable T . If T is continuous,
then the proof may be simplified. Let ν̃ ∈ M(Y N ) be the process
measure generated by (T, ν) (see Section 7.3), and let J (S : ν̃) be the
associated set of joinings.
We claim that J (S : ν̃) is compact in M(X × Y N ). To see this, note
that the direct product of S with the left-shift τ on Y N is continuous,
since each of these maps is continuous. It then follows that the induced
push-forward map mS×τ from M(X × Y N ) to itself is continuous in the
weak topology. The set of fixed points of any continuous map is closed.
Hence the set C1 = M(X × Y N , S × τ ) of fixed points of mS×τ is closed.
Additionally, the continuity of the projection mY N : M(X × Y N ) →
M(Y N ) ensures that C2 = m−1
{ν̃} is also closed, and therefore our set
YN
of interest J (S : ν̃) = C1 ∩ C2 is closed. Now let ǫ > 0. Since {ν̃} is
tight, there exists a compact set K ⊂ Y N such that ν̃(K) > 1 − ǫ. By
assumption X is compact, and therefore X × K is compact. Also, for
any λ ∈ J (S : ν̃), we have λ(X × K) = ν̃(K) > 1 − ǫ. Since ǫ > 0 was
arbitrary, we see that J (S : ν̃) is tight. Then, since it is both tight
and closed, we may conclude that it is compact.
As the projection (x, y) 7→ (x, y0 ) is continuous, the induced pushforward map mX ×Y : M(X × Y N ) → M(X × Y) is continuous. By
Lemma 7.1, mX ×Y (J (S : ν̃)) = J (S : ν), and therefore J (S : ν)
VARIATIONAL ANALYSIS OF INFERENCE FROM DYNAMICAL SYSTEMS 37
is compact, as it is the image of a compact set under a continuous
map.
Proof of Theorem
R 2.5. Let φℓ : J (S : ν) → R be the functional
defined by φℓ (λ) = ℓ dλ. The convexity of Jmin (S : ν) follows from
the fact that φℓ (tλ1 + (1 − t)λ2 ) = tφℓ (λ1 ) + (1 − t)φℓ (λ2 ). As J (S : ν)
is compact in the weak topology (Proposition A.5) and φℓ is lower
semicontinuous (Lemma A.4), the set Jmin (S : ν) where φℓ attains its
minimum is compact.
It remains to identify the extreme points of Jmin (S : ν). Any ergodic
measure in the (convex) set M(X ×Y, S ×T ) is an extreme point of this
set (see [51, Proposition 12.4]), and it follows that any ergodic measure
λ ∈ Jmin (S : ν) is an extreme point of Jmin (S : ν).R Suppose now that
λ is an extreme point of Jmin (S : ν), and let λ = η dξ be its ergodic
decomposition. By Lemma A.3, ξ-almost every η is in Jmin (S : ν), and
as λ is an extreme point of Jmin (S : ν), it follows that ξ-almost every
η equals λ. Then it follows from the ergodic decomposition that λ is
ergodic.
Kevin McGoff, Department of Mathematics, University of North
Carolina at Charlotte, Charlotte, NC 28223
E-mail address: [email protected]
URL: https://clas-math.uncc.edu/kevin-mcgoff/
Andrew B. Nobel, Department of Statistics and Operations Research, University of North Carolina at Chapel Hill, 308 Hanes Building, Chapel Hill, NC 27599
E-mail address: [email protected]
| 10math.ST
|
Node.DPWS – High performance &
scalable Web Services for the IoT
Konstantinos Fysarakis
Dept. of Electronic & Computer Engineering
Technical University of Crete
Chania, Crete, Greece
[email protected]
Damianos Mylonakis
Dept. of Computer Science
University of Crete
Heraklion, Crete, Greece
[email protected]
Charalampos Manifavas
Dept. of Informatics Engineering
Technological Educational Institute of Crete
Heraklion, Crete, Greece
[email protected]
Ioannis Papaefstathiou
Dept. of Electronic & Computer Engineering
Technical University of Crete
Chania, Crete, Greece
[email protected]
1
Abstract. Interconnected computing systems, in various forms, are expected to permeate our lives,
realizing the vision of the Internet of Things (IoT) and allowing us to enjoy novel, enhanced services that
promise to improve our everyday lives. Nevertheless, this new reality also introduces significant
challenges in terms of performance, scaling, usability and interoperability. Leveraging the benefits of
Service Oriented Architectures (SOAs) can help alleviate many of the issues that developers,
implementers and end‐users have to face in the context of the IoT. This work presents Node.DPWS, a
novel implementation of the Devices Profile for Web Services (DPWS) based on the Node.js platform.
Node.DPWS can be used to deploy lightweight, efficient and scalable Web Services over heterogeneous
nodes, including devices with limited resources. The performance of the presented work is evaluated on
typical embedded devices, including comparisons with implementations created using alternative DPWS
toolkits.
Keywords: Services Architectures; Development tools; Software libraries; Ubiquitous computing;
Pervasive computing; DPWS; Node.js;
1. Introduction
The advent of the pervasive presence of computing devices, interconnected to form the IoT, brings forth
various challenges. End‐users do not typically possess the skills to configure and setup the devices that
may be found in smart environments; in large‐scale deployments, individually setting up devices is not
even feasible. From the perspective of developers and implementers, there is a need for rapid
development and deployment, while simultaneously tackling issues of scaling and inherent limitations in
terms of resources (CPU, memory, power etc.). These are exacerbated by interoperability concerns, as
smart devices are offered in variety of hardware platforms and often operate over heterogeneous
networks. Moreover, while existing networking mechanisms are updated and adapted to efficiently
handle the vast population of these resource‐constrained devices (e.g. IETF’s work on 6LoWPAN [1]
standards to enable IPv6 connectivity over IEEE 802.15.4 networks), higher level, machine to machine
interactions, are often required to make use of the devices’ full potential. The above have compelled
researchers and developers alike to focus on mechanisms that guarantee interoperability, providing
seamless access to the various devices and their functional elements. SOAs evolved from this need to
have interoperable, cross‐platform, cross‐domain and network‐agnostic access to devices and their
services. This approach has already been successful in business environments, as SOAs allow
stakeholders to focus on the services themselves, rather than the underlying hardware and network
technologies. The DPWS [2] specification enables the adoption of a SOA‐based approach on embedded
devices with limited resources, allowing system owners to enjoy these benefits across the
heterogeneous systems of the IoT ecosystem.
This work presents Node.DPWS, a novel implementation of the DPWS specification using the Node.js
platform (http://nodejs.org/), also referred to as Node, a JavaScript‐based runtime environment
designed to maximize throughput and efficiency. The Node.DPWS libraries leverage the benefits of both
DPWS and Node, allowing the creation of high performance, scalable and lightweight DPWS devices on
2
the heterogeneous, often resource‐constrained, platforms typically found in smart environments.
Moreover, the libraries are easy to use and the devices can be defined using a minimum of code,
reducing the development effort.
The article begins with an introduction to DPWS, including its key characteristics and advantages. The
related works section (sidebar) details the profile’s current status in the industry and research
communities and also lists all available alternatives to Node.DPWS (i.e. other libraries and toolkits that
can be used to implement DPWS devices). Details about Node.DPWS follow, with a brief introduction to
Node.js and a description of the presented DPWS libraries and their features, also providing sample
code to demonstrate the implementation approach. The performance of Node.DPWS is evaluated on
actual embedded platforms, including a comparison to identical devices developed using the most
mature DPWS toolkit currently available. The article concludes with some pointers to future
enhancements to Node.DPWS.
2. The Devices Profile for Web Services (DPWS)
DPWS was introduced in 2004 by a consortium led by Microsoft and is now an OASIS open standard (at
version 1.1 since July 2009). The DPWS specification defines a minimal set of implementation constraints
to enable Web Service messaging on resource‐constrained devices. It employs similar messaging
mechanisms as the Web Services Architecture (WSA, [3]), with restrictions to complexity and message
size, allowing the provision of totally platform‐ and language‐neutral services, similar to those offered by
traditional web services. As it is built around the OASIS Web Services (WS‐) stack, it features superior
scaling and seamless integration capabilities, both in the context of enterprise environments (e.g. plant
floors or enterprise‐wide information systems) and the Internet (e.g. regular W3C‐specified Web
Services). The SOA‐based approach of DPWS can act as an enabler, by providing real‐time monitoring
and control features usable over the whole smart infrastructure. Allowing the migration of low level
information (e.g. sensing data) into higher level contexts (e.g. business operations or knowledge
extraction), new types of services are made possible; these are expected to be vital for future end‐user
as well as enterprise deployments, in a number of industry domains [4].
The profile’s architecture includes hosting and hosted services. A single hosting service is associated
with each device while the same device may accommodate various hosted services. The latter represent
the device’s various functional elements and rely on the hosting service for discovery. Thus, a
multifunctional device integrated into e.g. a smart home or an enterprise environment, will have a single
hosting service but may feature various hosted services (e.g. a temperature service, a door control
service, a movement‐sensing service etc.). As discovery services are included, the device can advertise
its presence on the network and search for other devices. Metadata exchange mechanisms provide
dynamic access to services hosted on a device. Additionally, publish and subscribe eventing mechanisms
allow clients to subscribe to services provided by devices.
DPWS was originally conceived and introduced as a successor to Universal Plug and Play (UPnP), but due
to the lack of backward compatibility such a transition has not taken place yet. Instead, nowadays DPWS
3
is actively pushed by industry stakeholders as the solution of choice for large‐scale deployments, while
UPnP is mostly targeted to home environments (printers, home entertainment etc.) [5]. Also, like UPnP,
DPWS is natively integrated into the various versions of the Windows operating system, from Windows
Vista onwards.
The main disadvantage of DPWS, at its current state, is the existence of numerous specifications
(protocols, bindings etc.) which have not been consolidated yet.
3. Background Material
The deployment and orchestration of web services on heterogeneous embedded devices is a very active
research area, following the success of some early research efforts, like the “Service Infrastructure for
Real time Embedded Networked Applications” (SIRENA, ITEA2) project which proved the feasibility and
advantages of integrating web services, via DPWS, across business segments, including critical sections
(e.g. the production floor [6]). Its follow‐up projects, "Service‐Oriented Device & Delivery Architecture"
(SODA, ITEA2) and "Service‐Oriented Cross‐layer Infrastructure for Distributed Smart Embedded
Devices" (SOCRADES, FP6), built upon the work of SIRENA and focused on providing a secure scalable
ecosystem and adding more sophisticated features into the SOA‐enabled devices to serve the
requirements of future manufacturing [7]. More recent research efforts include the “ArchitecturE for
Service‐Oriented Process” (IMC‐AESOP, FP7) and “Web of Objects” (WoO, ITEA2) projects. Researchers
in AESOP proposed a system‐of‐systems approach for monitoring and control, based on a SOA for very
large scale (i.e. up to tens of thousands of devices) distributed systems in process control applications.
WoO promoted the use of DPWS to build a secure, context‐aware service infrastructure for smart
objects, focusing on the interoperability of devices & services through the use of semantics.
As many industry leaders from a variety of sectors (electronics, power, automation, enterprise, home
etc.) have been involved in the above research efforts, it is evident that there is significant industry
backing of DPWS and its use in future applications and products. Moreover, the use and benefits of
DPWS have been studied extensively by researchers in the context of various applications areas, which,
among others, include automotive and railway systems [8], industrial automation [9], eHealth [10],
smart cities [11], smart homes [12] and wireless sensor networks [13]. All of the above are promising
indicators for the future of the specification and its wider adoption.
3.1. DPWS implementations
A survey on alternative to Node.DPWS APIs for DPWS development reveals a plethora of available
solutions with diverse characteristics: the libraries included in Microsoft’s .NET Micro Framework
(http://netmf.codeplex.com/releases/view/82448), WS4D’s toolkits which include WS4D‐uDPWS
(http://code.google.com/p/udpws/), WS4D‐JMEDS (http://sourceforge.net/projects/ws4d‐javame/),
WS4D‐Axis2
(http://trac.e‐technik.uni‐rostock.de/projects/ws4d‐axis2)
and
WS4D‐gSOAP
(https://trac.e‐technik.uni‐rostock.de/projects/ws4d‐gsoap/), and, finally, SOA4D’s solutions which
include
DPWS‐Core
(https://forge.soa4d.org/projects/dpwscore/)
and
DPWS4J
4
(https://forge.soa4d.org/projects/dpws4j/). Information for the identified DPWS implementations is
aggregated in Table 1.
Table 1. Overview of DPWS Toolkits
.NET Micro
Framework
WS4D‐
uDPWS
WS4D‐
JMEDS
C#
C
Java
Language
CPU/OS
DPWS 1.0
DPWS 1.1
IPv4
IPv6
ARM
√
√
√
‐
PC
(VM),
TelosB,
AVR
Raven
‐
√
√
√
Java SE,
Java
CDC/CLDC,
Android
WS4D‐
Axis2
Java
(Apache
Axis2)
Java SE
WS4D‐
gSOAP
DPWS‐
Core
DPWS4J
C
C
Java
Linux,
Linux,
Windows,
Windows
ARM
Java SE,
Java
CDC
√
√
√
√
√
√
√
√
‐
Partial
Partial
‐
√
√
√
√
‐
Partial
√
‐
Apache
License Apache 2.0 FreeBSD
EPL
GPL/LGPL
LGPL
LGPL
2.0
Active
Yes
No
Yes
No
No
Yes
No
Nevertheless, when focusing on key features such as code portability, deployment on heterogeneous
platforms, support for IPv6 (necessary in the context of the IoT) and active development and support of
the tools, the valid options are actually fewer. With the above considerations in mind, WS4D‐JMEDS
currently stands out as the most attractive choice. It is the most advanced work of the WS4D initiative,
providing a mature and feature‐rich platform which is being constantly updated and improved. It is,
therefore, the toolkit that will be used as a benchmark for the Node.DPWS implementation presented in
this work.
4. About Node.js
Node.js is a relatively new platform, introduced in May 2009 (in version v0.10.31, as of August 2014). It
is an evented server‐side implementation based on Google’s V8 JavaScript engine. Both Node and the
V8 engine are mostly implemented in C and C++, but Node’s wrapper enhances the engine’s basic
features by allowing server‐side deployment of JavaScript programs and the use of various C libraries,
system calls, binary data manipulation and request handling. The core concept behind its development
was to create the building block for lightweight and scalable servers, providing an evented, non‐blocking
infrastructure to develop highly concurrent applications.
Node handles network input/output (I/O) operations in an evented, non‐blocking fashion, while file I/O
operations are handled asynchronously. This differentiates Node from typical implementations which
follow the threaded model, where for each new connection a thread is created and which have inherent
5
issues with scaling. In Node, each new connection is only a small heap allocation. Moreover, Node’s
executing thread cannot be blocked; in situations where this would normally happen (e.g. waiting for
data from a remote database), the thread’s runtime is utilized to serve other requests. The above result
in very fast applications which also scale well, even in the case of resource‐constrained devices (i.e. with
no multi‐core processors or large amounts of memory) like those typically embedded in smart devices.
More details on Node’s characteristics, as well as sample code, can be found in [14].
Research [15] indicates that while Node.js offers significant benefits in terms of I/O operations’
performance and resource utilization, it is not as good at serving large static files. This does not harm its
applicability in developing fast, scalable network applications, and it is expected to improve as the
platform matures. Moreover, it is not an issue in the context of typical IoT applications, where devices
are expected to transmit low level information (e.g. sensing data) and receive commands to invoke
operations on their functional elements (e.g. turn ON/OFF, set thresholds, rotate etc.). Finally, some
concerns regarding the security of Node.js applications can be attributed to the lack of a stable version
(as the project is relatively new) and most can be avoided when following security‐conscious
programming practices [16]. The platform is not inherently insecure, and thus can safely be used in
production‐grade deployments.
As Node.js addresses many of the issues with developing applications requiring real‐time and
lightweight communications, it has quickly gained the support of the developers’ community (there is a
variety of libraries already available for or compatible with Node) and of major stakeholders in the
industry [17], including companies such as Google, Microsoft and Yahoo!. Popular websites such as
Wikipedia, LinkedIn, eBay and Microsoft’s Azure cloud platform already make use of Node.js, even
though it has not reached a stable (1.0) version yet; an indicator that there is a strong demand for
Node’s features and that its user base will continue to grow over time.
5. The Node.DPWS libraries
Node’s characteristics are a good match for event‐driven DPWS‐based web services deployable on
embedded devices present in smart enterprise, industrial, domestic and other ubiquitous‐computing‐
enhanced environments. Thus it is an attractive solution for implementing the DPWS specification,
which could potentially deliver highly scalable DPWS devices, able to handle many clients concurrently,
while having very low resource consumption. Node.DPWS provides such an implementation of the
DPWS specification using Node.js.
The developer is responsible for describing the device’s attributes, its supported services, operation and
events, and the libraries will properly advertise them and match them to requests. A key characteristic
of Node.DPWS is its ease of use. Operations can be defined with just a few lines of code, and the
developer does not have to deal with any low level aspects of the implementation, such as the
communication protocols. To highlight these characteristics, the source code of a DPWS device is
provided in Figure 1.
6
Figure 1. Creating a simple DPWS device using Node.DPWS
7
In this example Node.DPWS is used to expose a simple service which provides clients with temperature
readings (e.g. ones provided by a smart sensor). This temperature operation is defined in a few lines of
code; input/output types of the operation are specified and then a handler is provided. The latter will be
called whenever an invocation to this operation is made. More complex operations can also be
developed in an equally simple manner. In the example of the smart sensor above, a more complex
operation could allow clients to subscribe to get temperature readings at set intervals or when certain
events occur (following the WS‐Eventing specification [18]) and even control the device (setting
temperature thresholds, turning it off etc.).
In the proposed solution, nodes support auto‐discovery by implementing WS‐Discovery [19] which
defines a multicast discovery protocol to locate services, with the main mode of discovery being a client
looking for one or more target services. Apart from discovering devices, the developed library allows for
answering to discovery requests. The developer must provide the details of her device, which the library
will then forward to requesting nodes whose queries match the device. The library supports both DPWS
version 1.0 (2006) and 1.1 (2009).
5.1. Performance evaluation
In order to assess the performance of Node.DPWS and how it compares to alternative solutions, we
focused on creating and examining the behavior of a simple DPWS device featuring a “GetTemperature”
operation which, when invoked, returned an integer value. Three versions of the device were
developed: the Node.DPWS one and two versions using WS4D‐JMEDS (i.e. the most appealing
alternative implementation currently available), one compiled using Java Standard Edition (SE) and the
other following the Java Connected Device Configuration (CDC), which is part of the Java Platform Micro
Edition (Java ME) and designed for handheld and embedded systems (i.e. by design better‐suited to
resource‐constrained devices).
The devices were deployed on Beaglebone (http://beagleboard.org/bone) platforms, equipped with a
720MHz ARM Cortex‐A8 processor, 256MB of RAM and the Arch Linux ARM (http://archlinuxarm.org/)
operating system, interconnected via wired Ethernet, to minimize the effects of the network on the
reported performance figures. The test‐bed also featured a client application developed to discover and
query the DPWS devices, recording response times, for benchmarking purposes.
A total of 500 requests were issued from the benchmarking client (running on a desktop PC) to each of
the three DPWS devices, while various aspects of their performance were being monitored and
recorded. Figure 2 presents the response time (i.e. the time that the client had to wait before it got a
response to its GetTemperature invocation). As is evident from the graph, the Node.DPWS device
responded significantly faster than the WS4D‐JMEDS‐based implementations. Random spikes on the
response times can be attributed to the triggering of “housekeeping” operations on the target
platforms. Averaging the response times over the 500 requests reveals that each test client got its
response from the WS4D‐JMEDS SE device after about 72.48ms. The CDC‐compiled device responded
faster, with an average response time of 51.99ms. Still, the Node.DPWS device performed significantly
8
better than both the WS4D‐JMEDS devices, featuring an average response time of 24.44ms, i.e. 66.3%
and 53% faster than the SE and CDC variants respectively. Figure 3 better depicts this difference, along
with other performance aspects which were observed during benchmarks. The Node.DPWS device was
able to reply to twice the number of requests per second than the second best performing device, which
is the WS4D‐JMEDS CDC one. As a result of the above response time gap, the 500‐request benchmark
was completed in just 12.22 seconds when querying the Node.DPWS device.
1
16
31
46
61
76
91
106
121
136
151
166
181
196
211
226
241
256
271
286
301
316
331
346
361
376
391
406
421
436
451
466
481
496
200
180
160
140
120
100
80
60
40
20
0
Node.DPWS
WS4D‐JMEDS SE
WS4D‐JMEDS CDC
Figure 2. Response time (in ms) for 500 requests.
72,48
51,99
40,92
36,24
26,00
24,44
19,23
13,80
12,22
Node.DPWS
WS4D‐JMEDS SE
Avg. Response Time (ms)
Requests / sec
WS4D‐JMEDS CDC
Runtime (sec)
Figure 3. Performance aspects while handling 500 benchmark requests
9
Moving to the performance impact on the DPWS device itself, the load on the ARM processor and
memory consumption were recorded during benchmarking. The devices did not have significant
differences in terms of CPU load; the average load while handling the test requests was recorded at
90.4%, 91.9% and 96.7% for Node.DPWS, WS4D‐JMEDS SE and WS4D‐JMEDS CDC respectively. The
Node.DPWS device also demonstrated better behavior in terms of the memory consumed during
benchmarks. Its memory footprint was measured at 26440 bytes on average, which was 10% lower than
its WS4D‐JMEDS CDC counterpart (which occupied 29387 bytes) and 18% lower than the WS4D‐JMEDS
SE one (which consumed an average of 32280 bytes).
6. Conclusions
In this article we validated the benefits of leveraging the relatively novel Node.js platform to implement
the DPWS specification in the form of Node.DPWS; an easy to use and lightweight set of libraries for
creating and deploying DPWS devices on heterogeneous systems with limited resources. The
performance assessment revealed that Node.DPWS outperforms the most attractive alternative
currently available, the WS4D‐JMEDS toolkit. The enhanced performance, scaling and ease‐of‐
development characteristics of Node.DPWS show that there is significant room for improvement in the
DPWS‐related tools currently available to developers. It is, therefore, worthwhile to pursue further work
on the Node.DPWS implementation, in order to enrich its libraries with more features, e.g. to extend the
support to other WS‐* related protocols, starting with the WS‐Security [20] specification. This will also
allow us to study the API’s behavior in more complex scenarios. Finally, work is under way to provide a
website along with proper documentation, to facilitate the use and expansion of Node.DPWS (and the
associated Node.js and DPWS technologies) by the research and development community.
7. References
[1] G. Mulligan, “The 6LoWPAN architecture,” in Proceedings of the 4th workshop on Embedded
networked sensors ‐ EmNets ’07, 2007, p. 78.
[2] “Devices profile for web services, version 1.1,” 2009. [Online]. Available: http://docs.oasis‐
open.org/ws‐dd/dpws/1.1/os/wsdd‐dpws‐1.1‐spec‐os.pdf. [Accessed: 23‐Aug‐2014].
[3] D. Booth, H. Haas, F. McCabe, E. Newcomer, M. Champion, C. Ferris, and D. Orchard, “Web Services
Architecture,” W3C Working Group, 2004. [Online]. Available: http://www.w3.org/TR/ws‐arch/.
[Accessed: 23‐Aug‐2014].
[4] P. Spieß and S. Karnouskos, “Maximizing the business value of networked embedded systems
through process‐level integration into enterprise software,” in 2007 2nd International Conference
on Pervasive Computing and Applications, ICPCA’07, 2007, pp. 536–541.
[5] T. Nixon, “UPnP Forum and DPWS Standardization Status,” Rally Technologies, Windows Device and
Storage
Technologies
Group,
2008.
[Online].
Available:
http://download.microsoft.com/download/f/0/5/f05a42ce‐575b‐4c60‐82d6‐
208d3754b2d6/UPnP_DPWS_RS08.pptx. [Accessed: 23‐Aug‐2014].
10
[6] F. Jammes, A. Mensch, and H. Smit, “Service‐oriented device communications using the devices
profile for Web services,” in Proceedings ‐ 21st International Conference on Advanced Information
Networking and Applications Workshops/Symposia, AINAW’07, 2007, vol. 2, pp. 947–955.
[7] S. Karnouskos, D. Savio, P. Spiess, D. Guinard, V. Trifa, and O. Baecker, “Real‐world Service
Interaction with Enterprise Systems in Dynamic Manufacturing Environments,” in Artificial
Intelligence Techniques for Networked Manufacturing Enterprises Management SE ‐ 14, L.
Benyoucef and B. Grabot, Eds. Springer London, 2010, pp. 423–457.
[8] V. Venkatesh, V. Vaithayana, P. Raj, K. Gopalan, and R. Amirtharaj, “A Smart Train Using the DPWS‐
based Sensor Integration,” Res. J. Inf. Technol., vol. 5, no. 3, pp. 352–362, Mar. 2013.
[9] T. Cucinotta, A. Mancina, G. F. Anastasi, G. Lipari, L. Mangeruca, R. Checcozzo, and F. Rusina, “A
Real‐Time Service‐Oriented Architecture for Industrial Automation,” IEEE Trans. Ind. Informatics, vol.
5, no. 3, pp. 267–277, Aug. 2009.
[10]S. Pöhlsen, S. Schlichting, M. Strähle, F. Franz, and C. Werner, “A DPWS‐Based Architecture for
Medical Device Interoperability,” in World Congress on Medical Physics and Biomedical Engineering,
September 7 ‐ 12, 2009, Munich, Germany SE ‐ 22, vol. 25/5, O. Dössel and W. Schlegel, Eds.
Springer Berlin Heidelberg, 2009, pp. 82–85.
[11]K. Fysarakis, K. Rantos, O. Sultatos, C. Manifavas and I. Papaefstathiou, “Policy‐based Access Control
for DPWS‐enabled Ubiquitous Devices,” in 19th IEEE International Conference on Emerging
Technologies and Factory Automation (ETFA 2014), Barcelona, Spain, Sept. 16‐19, 2014.
[12]A. Müller, H. Kinkelin, S. K. Ghai, and G. Carle, “A secure service infrastructure for interconnecting
future home networks based on DPWS and XACML,” in Proceedings of the 2010 ACM SIGCOMM
workshop on Home networks ‐ HomeNets ’10, 2010, p. 31.
[13]O. Dohndorf, J. Krüger, H. Krumm, C. Fiehe, A. Litvina, I. Luck, and F. J. Stewing, “Towards the web of
things: Using DPWS to bridge isolated OSGi platforms,” in 2010 8th IEEE International Conference on
Pervasive Computing and Communications Workshops, PERCOM Workshops 2010, 2010, pp. 720–
725.
[14]S. Tilkov and S. Vinoski, “Node.js: Using JavaScript to build high‐performance network programs,”
IEEE Internet Comput., vol. 14, pp. 80–83, 2010.
[15]I. K. Chaniotis, K.‐I. D. Kyriakou, and N. D. Tselikas, “Is Node.js a viable option for building modern
web applications? A performance evaluation study,” Computing, Mar. 2014.
[16]A. Ojamaa and D. Karl, “Security Assessment of Node.js Platform,” in Information Systems Security,
2012, pp. 35–43.
[17]“Projects,
Applications,
and
Companies
Using
Node.”
[Online].
Available:
https://github.com/joyent/node/wiki/Projects,‐Applications,‐and‐Companies‐Using‐Node.
[Accessed: 23‐Aug‐2014].
[18]D. Box, L. F. Cabrera, C. Critchley, F. Curbera, D. Ferguson, S. Graham, D. Hull, G. Kakivaya, A. Lewis,
B. Lovering, P. Niblett, D. Orchard, S. Samdarshi, J. Schlimmer, I. Sedukhin, J. Shewchuk, S.
Weerawarana, and D. Wortendyke, “Web Services Eventing (WS‐Eventing),” W3C Member
Submission, vol. 2009. pp. 1–34, 2006.
11
[19]T. Nixon, A. Regnier, V. Modi, and D. Kemp, “Web Services Dynamic Discovery (WS‐ Discovery),
version 1.1,” OASIS Standard Specification, 2009.
[20]K. Lawrence, C. Kaler, A. Nadalin, R. Monzilo, and P. Hallam‐Baker, “Web Services Security: SOAP
Message Security 1.1,” OASIS Standard Specification, 2006.
12
| 6cs.PL
|
arXiv:1707.06658v4 [cs.LG] 29 Nov 2017
RAIL: Risk-Averse Imitation Learning
Anirban Santara∗
IIT Kharagpur
Abhishek Naik∗ Balaraman Ravindran
IIT Madras
[email protected]
{anaik,ravi}@cse.iitm.ac.in
Dipankar Das
Dheevatsa Mudigere
Sasikanth Avancha
Parallel Computing Lab - Intel Labs, India
Bharat Kaul
{dipankar.das,dheevatsa.mudigere,sasikanth.avancha,bharat.kaul}@intel.com
Abstract
Imitation learning algorithms learn viable policies by imitating an expert’s behavior
when reward signals are not available. Generative Adversarial Imitation Learning
(GAIL) is a state-of-the-art algorithm for learning policies when the expert’s
behavior is available as a fixed set of trajectories. We evaluate in terms of the
expert’s cost function and observe that the distribution of trajectory-costs is often
more heavy-tailed for GAIL-agents than the expert at a number of benchmark
continuous-control tasks. Thus, high-cost trajectories, corresponding to tail-end
events of catastrophic failure, are more likely to be encountered by the GAILagents than the expert. This makes the reliability of GAIL-agents questionable
when it comes to deployment in risk-sensitive applications like robotic surgery and
autonomous driving. In this work, we aim to minimize the occurrence of tail-end
events by minimizing tail risk within the GAIL framework. We quantify tail risk by
the Conditional-Value-at-Risk (CV aR) of trajectories and develop the Risk-Averse
Imitation Learning (RAIL) algorithm. We observe that the policies learned with
RAIL show lower tail-end risk than those of vanilla GAIL. Thus the proposed
RAIL algorithm appears as a potent alternative to GAIL for improved reliability in
risk-sensitive applications.
1
Introduction
Reinforcement learning (RL) [Sutton and Barto, 1998] is used to learn an effective policy of choosing
actions in order to achieve a specified goal in an environment. The goal is communicated to the agent
through a scalar cost and the agent learns a policy that minimizes the expected total cost incurred over
a trajectory. RL algorithms, along with efficient function approximators like deep neural networks,
have achieved human-level or beyond human-level performance at many challenging planning tasks
like continuous-control [Lillicrap et al., 2015, Schulman et al., 2015] and game-playing [Silver
et al., 2016, Mnih et al., 2015]. In classical RL, the cost function is handcrafted based on heuristic
assumptions about the goal and the environment. This is challenging in most real-world applications
and also prone to subjectivity induced bias. Imitation learning or Learning from Demonstration
(LfD) [Argall et al., 2009, Schaal, 1997, Atkeson and Schaal, 1997, Abbeel and Ng, 2011, 2004, Ng
et al., 2000] addresses this challenge by providing methods of learning policies through imitation
of an expert’s behavior without the need of a handcrafted cost function. In this paper we study the
reliability of existing imitation learning algorithms when it comes to learning solely from a fixed set
of trajectories demonstrated by an expert with no interaction between the agent and the expert during
training.
∗
Authors contributed equally as a part of their internship at Parallel Computing Lab - Intel Labs, India.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Figure 1: Histograms of the costs of 250 trajectories generated by the expert and GAIL agents at
high-dimensional continuous control tasks, Hopper-v1 and Humanoid-v1, from OpenAI Gym. The
inset diagrams show zoomed-in views of the tails of these distributions (the region beyond 2σ of the
mean). We observe that the GAIL agents produce tails heavier than the expert, indicating that GAIL
is more prone to generating high-cost trajectories.
Imitation learning algorithms fall into two broad categories. The first category, known as Behavioral
Cloning [Pomerleau, 1989, Bojarski et al., 2016, 2017], uses supervised learning to fit a policy function
to the state-action pairs from expert-demonstrated trajectories. Despite its simplicity, Behavioral
Cloning fails to work well when only a limited amount of data is available. These algorithms assume
that observations are i.i.d. and learn to fit single time-step decisions. Whereas, in sequential decision
making problems where predicted actions affect the future observations (e.g. driving), the i.i.d.
assumption is violated. As a result, these algorithms suffer from the problem of compounding error
due to covariate shift [Ross and Bagnell, 2010, Ross et al., 2011]. Approaches to ameliorate the
issue of compounding error like SMILe [Ross and Bagnell, 2010], SEARN [Daumé et al., 2009],
CPI [Kakade and Langford, 2002] suffer from instability in practical applications [Ross et al., 2011]
while DAGGER [Ross et al., 2011] and AGGREVATE [Ross and Bagnell, 2014] require the agent to
query the expert during training which is not allowed in our setting of learning from a fixed set of
expert demonstrations. Another drawback of Behavioral Cloning is that it does not allow the agent to
explore alternate policies for achieving the same objective that might be efficient in some sense other
than what the expert cared for.
The second category of algorithms is known as Inverse Reinforcement Learning (IRL) (Russell
[1998], Ng et al. [2000], Abbeel and Ng [2011]). It attempts to uncover the underlying reward
function that the expert is trying to maximize from a set of expert-demonstrated trajectories. This
reward function succinctly encodes the expert’s behavior and can be used by an agent to learn a
policy through an RL algorithm. The method of learning policies through RL after IRL is known as
Apprenticeship Learning (Abbeel and Ng [2004]). IRL algorithms find reward functions that prioritize
entire trajectories over others. Unlike behavioral cloning, they do not fit single time-step decisions,
and hence they do not suffer from the issue of compounding error. However, IRL algorithms are
indirect because they learn a reward function that explains expert behavior but do not tell the learner
how to act directly (Ho and Ermon [2016]). The job of learning an actionable policy is left to RL
algorithms. Moreover, IRL algorithms are computationally expensive and have scalability issues in
large environments (Finn et al. [2016], Levine and Koltun [2012]).
2
The recently proposed Generative Adversarial Imitation Learning (GAIL) algorithm [Ho and Ermon,
2016] presents a novel mathematical framework in which the agent learns to act by directly extracting a
policy from expert-demonstrated trajectories, as if it were obtained by RL following IRL. The authors
show that unlike Behavioral Cloning, this method is not prone to the issue of compounding error and
it is also scalable to large environments. Currently, GAIL provides state-of-the-art performance at
several benchmark control tasks, including those in Table 1.
Risk sensitivity is integral to human learning [Nagengast et al., 2010, Niv et al., 2012], and risksensitive decision-making problems, in the context of MDPs, have been investigated in various fields,
e.g., in finance [Ruszczyński, 2010], operations research [Howard and Matheson, 1972, Borkar,
2002], machine learning [Heger, 1994, Mihatsch and Neuneier, 2002] and robotics [Shalev-Shwartz
et al., 2016, 2017, Abbeel et al., 2007, Rajeswaran et al., 2016]. [Garcıa and Fernández, 2015] give a
comprehensive overview of different risk-sensitive RL algorithms. They fall in two broad categories.
The first category includes methods that constrain the agent to safe states during exploration while
the second modifies the optimality criterion of the agent to embed a term for minimizing risk. Studies
on risk-minimization are rather scarce in the imitation learning literature. [Majumdar et al., 2017]
take inspiration from studies like [Glimcher and Fehr, 2013, Shen et al., 2014, Hsu et al., 2005] on
modeling risk in human decision-making and conservatively approximate the expert’s risk preferences
by finding an outer approximation of the risk envelope. Much of the literature on imitation learning
has been developed with average-case performance at the center, overlooking tail-end events. In
this work, we aim to take an inclusive and direct approach to minimizing tail risk of GAIL-learned
policies at test time irrespective of the expert’s risk preferences.
In order to evaluate the worst-case risk of deploying GAIL-learned policies, we studied the distributions (see Figure 1) of trajectory-costs (according to the expert’s cost function) for the GAIL agents
and experts at different control tasks (see Table 1). We observed that the distributions for GAIL are
more heavy-tailed than the expert, where the tail corresponds to occurrences of high trajectory-costs.
In order to quantify tail risk, we use Conditional-Value-at-Risk (CV aR) [Rockafellar and Uryasev,
2000]. CV aR is defined as the expected cost above a given level of confidence and is a popular
and coherent tail risk measure. The heavier the tail, the higher the value of CV aR. We observe
that the value of CV aR is much higher for GAIL than the experts at most of the tasks (see Table
1) which again suggests that the GAIL agents encounter high-cost trajectories more often than the
experts. Since high trajectory-costs may correspond to events of catastrophic failure, GAIL agents
are not reliable in risk-sensitive applications. In this work, we aim to explicitly minimize expected
worst-case risk for a given confidence bound (quantified by CV aR) along with the GAIL objective,
such that the learned policies are more reliable than GAIL, when deployed, while still preserving
the average performance of GAIL. [Chow and Ghavamzadeh, 2014] developed policy gradient and
actor-critic algorithms for mean-CV aR optimization for learning policies in the classic RL setting.
However these algorithms are not directly applicable in our setting of learning a policy from a set
of expert-demonstrated trajectories. We take inspiration from this work and make the following
contributions:
1. We formulate the Risk-Averse Imitation Learning (RAIL) algorithm which optimizes CV aR
in addition to the original GAIL objective.
2. We evaluate RAIL at a number of benchmark control tasks and demonstrate that it obtains
policies with lesser tail risk at test time than GAIL.
The rest of the paper is organized as follows. Section 2 builds the mathematical foundation of
the paper by introducing essential concepts of imitation learning. Section 3 defines relevant riskmeasures and describes the proposed Risk-Averse Imitation Learning algorithm. Section 4 specifies
our experimental setup and Section 5 outlines the evaluation metrics. Finally, Section 6 presents the
results of our experiments comparing RAIL with GAIL followed by a discussion of the same and
Section 7 concludes the paper with scope of future work.
2
Mathematical Background
Let us consider a Markov Decision Process (MDP), M = (S, A, T , c, p0 , γ), where S denotes
the set of all possible states, A denotes the set of all possible actions that the agent can take,
T : S × A × S → [0, 1] is the state transition function such that, T (s0 |s, a) is a probability
distribution over next states, s0 ∈ S given current state s ∈ S and action a ∈ A, c : S × A → R is the
cost function which generates a real number as feedback for every state-action pair, p0 : S → [0, 1]
gives the initial state distribution, and γ is a temporal discount factor.
3
A policy π : S × A → [0, 1] is a function such that π(a|s) gives a probability distribution over
actions, a ∈ A in a given state, s ∈ S. Let ξ = (s0 , a0 , s1 , . . . , sLξ ) denote a trajectory of length Lξ ,
obtained by following a policy π. We define expectation of a function f (·, ·) defined on S × A with
respect to a policy π as follows:
Lξ −1
X
Eπ [f (s, a)] , Eξ∼π
γ t f (st , at )
(1)
t=0
2.1
Generative Adversarial Imitation Learning
Apprenticeship learning or Apprenticeship Learning via Inverse Reinforcement Learning algorithms
[Abbeel and Ng, 2004] first estimate the expert’s reward function using IRL and then find the optimal
policy for the recovered reward function using RL. Mathematically, this problem can be described as:
RL ◦ IRL(πE ) = argmin max Eπ [c(s, a)] − EπE [c(s, a)] − H(π)
π∈Π
c∈C
(2)
where, πE denotes the expert-policy. c(·, ·) denotes the cost function. Π and C denote the hypothesis
classes for policy and cost functions. H(π) denotes entropy of policy π. The term −H(π) provides
causal-entropy regularization [Ziebart, 2010, Ziebart et al., 2008] which helps in making the policy
optimization algorithm unbiased to factors other than the expected reward.
[Ho and Ermon, 2016] proposed Generative Adversarial Imitation Learning (GAIL) which packs the
two step process of RL ◦ IRLψ (πE ) into a single optimization problem with special considerations
for scalability in large environments. The name is due to the fact that this objective function can be
optimized using the Generative Adversarial Network (GAN) [Goodfellow et al., 2014] framework.
The following is objective function of GAIL:
argmin
π∈Π
max
D∈(0,1)S×A
Eπ [log(D(s, a))] + EπE [log(1 − D(s, a))] − H(π)
(3)
Here, the agent’s policy, π, acts as a generator of state-action pairs. D is a discriminative binary
classifier of the form D : S × A → (0, 1), known as discriminator, which given a state-action pair
(s, a), predicts the likelihood of it being generated by the generator. A two-player adversarial game is
started, wherein the generator tries to generate (s, a) pairs that closely match the expert, while the
discriminator tries to correctly classify the (s, a) pairs of the expert and the agent. At convergence,
the agent’s actions resemble those of the expert in any given state.
The generator and the discriminator are assigned parameterized models πθ and Dw respectively.
The training algorithm alternates between a gradient ascent step with respect to the discriminator
parameters, w, and a policy-gradient descent step with respect to the generator parameters, θ.
Following the example of [Ho and Ermon, 2016] we use multi-layer perceptrons (neural networks
with fully-connected layers) [Haykin, 1998] to model both the generator and the discriminator.
3
Risk-Averse Imitation Learning
In this section, we develop the mathematical formulation of the proposed Risk-Averse Imitation
Learning (RAIL) algorithm. We introduce CV aR [Rockafellar and Uryasev, 2000] as a measure of
tail risk, and apply it in the GAIL-framework to minimize the tail risk of learned policies.
3.1
Conditional-Value-at-Risk
In the portfolio-risk optimization literature, tail risk is a form of portfolio risk that arises when the
possibility that an investment moving more than three standard deviations away from the mean
is greater than what is shown by a normal distribution [Investopedia, 2017]. Tail risk corresponds
to events that have a small probability of occurring. When the distribution of market returns is
heavy-tailed, tail risk is high because there is a probability, which may be small, that an investment
will move beyond three standard deviations.
Conditional-Value-at-Risk (CV aR) [Rockafellar and Uryasev, 2000] is the most conservative measure of tail risk [Dalleh, 2011]. Unlike other measures like Variance and Value at Risk (V aR), it can
4
be applied when the distribution of returns is not normal. Mathematically, let Z be a random variable.
Let α ∈ [0, 1] denote a probability value. The Value-at-Risk of Z with respect to confidence level α,
denoted by V aRα (Z), is defined as the minimum value z ∈ R such that with probability α, Z will
not exceed z.
V aRα (Z) = min(z | P (Z ≤ z) ≥ α)
(4)
CV aRα (Z) is defined as the conditional expectation of losses above V aRα (Z):
CV aRα (Z) = E [Z | Z ≥ V aRα (Z)] = min Hα (Z, ν)
(5)
ν∈R
where Hα (Z, ν) is given by:
Hα (Z, ν) , {ν +
3.2
1
E (Z − ν)+ }; (x)+ = max(x, 0)
1−α
(6)
RAIL Framework
We use CV aR to quantify the tail risk of the trajectory-cost variable Rπ (ξ|c(D)), defined in the
context of GAIL as:
Lξ −1
Rπ (ξ|c(D)) =
X
γ t c(D(st , at ))
(7)
t=0
where c(·) is order-preserving.
Next, we formulate the optimization problem to optimize CV aR of Rπ (ξ|c(D)) as:
min max CV aRα (Rπ (ξ|c(D))) = min max Hα (Rπ (ξ|c(D)), ν)
π
c
π,ν
(8)
c
Integrating this with the GAIL objective of equation 3, we have the following:
n
min max J = min max
− H(π) + Eπ [log(D(s, a))]
π,ν D∈(0,1)S×A
π,ν D∈(0,1)S×A
+ EπE [log(1 − D(s, a))] + λCV aR Hα (Rπ (ξ|c(D)), ν)
o
(9)
Note that as c(·) is order-preserving, the maximization with respect to c in equation 8 is equivalent
to maximization with respect to D in equation 9. λCV aR is a constant that controls the amount
of weightage given to CV aR optimization relative to the original GAIL objective. Equation 9
comprises the objective function of the proposed Risk-Averse Imitation Learning (RAIL) algorithm.
Algorithm 1 gives the pseudo-code. Appendix A derives the expressions of gradients of the CV aR
term Hα (Rπ (ξ|c(D))ν) with respect to π, D, and ν. When α → 0, namely the risk-neutral case,
CV aR is equal to the mean of all trajectory costs and hence, RAIL → GAIL. We use Adam
algorithm [Diederik Kingma, 2015] for gradient ascent in the discriminator and Trust Region Policy
Optimization (TRPO) [Schulman et al., 2015] for policy gradient descent in the generator. The CV aR
term ν is trained by batch gradient descent [Haykin, 1998].
4
Experimental Setup
We compare the tail risk of policies learned by GAIL and RAIL for five continuous control tasks
listed in Table 1. All these environments, were simulated using MuJoCo Physics Simulator [Todorov
et al., 2012]. Each of these environments come packed with a “true" reward function in OpenAI
Gym [Brockman et al., 2016]. [Ho and Ermon, 2016] trained neural network policies using Trust
Region Policy Optimization (TRPO) [Schulman et al., 2015] on these reward functions to achieve
state-of-the-art performance and have made the pre-trained models publicly available for all these
environments as a part of their repository [OpenAI-GAIL, 2017]. They used these policies to generate
the expert trajectories in their work on GAIL [Ho and Ermon, 2016]. For a fair comparison, we
use the same policies to generate expert trajectories in our experiments. Table 1 gives the number
of expert trajectories sampled for each environment. These numbers correspond to the best results
reported in [Ho and Ermon, 2016].
5
Algorithm 1 Risk-Averse Imitation learning (RAIL)
Input: Expert trajectories ξE ∼ πE , hyper-parameters α, β, λCV aR
Output: Optimized learner’s policy π
1: Initialization: θ ← θ0 , w ← w0 , ν ← ν0 , λ ← λCV aR
2: repeat
3:
Sample trajectories ξi ∼ πθi
1
4:
Estimate Ĥα (Dπ (ξ|c(D)), ν) = ν + 1−α
Eξi [(Dπ (ξ|c(D)) − ν)+ ]
5:
Gradient ascent on discriminator parameters using:
6:
∇wi J = Êξi [∇wi log(D(s, a))] + ÊξE [∇wi log(1 − D(s, a))]
+ λCV aR ∇wi Hα (Rπ (ξ|c(D)), ν)
KL-constrained natural gradient descent step (TRPO)
on policy parameters using:
∇θi J = E(s,a)∼ξi [∇θi logπθ (a|s)Q(s, a)] − ∇θi H(πθ )
+λCV aR ∇θi Hα (Rπ (ξ|c(D)), ν)
where Q(s̄, ā) = E(s,a)∼ξi [log(Dwi+1 (s, a))|s0 = s̄, a0 = ā]
Gradient descent on CVaR parameters:
∇νi J = ∇νi Hα (Rπ (ξ|c(D)), ν)
8: until i == max_iter
7:
Again, following [Ho and Ermon, 2016], we model the generator (policy), discriminator and value
function (used for advantage estimation [Sutton and Barto, 1998] for the generator) with multi-layer
perceptrons of the following architecture: observationDim - fc_100 - tanh - fc_100
- tanh - outDim, where fc_100 means fully connected layer with 100 nodes, tanh represents
the hyperbolic-tangent activation function of the hidden layers, observationDim stands for the
dimensionality of the observed feature space, outDim is equal to 1 for the discriminator and value
function networks and equal to the twice of the dimensionality of the action space (for mean and
standard deviation of the Gaussian from which the action should be sampled) for the policy network.
For example, in case of Humanoid-v1, observationDim = 376 and outDim = 34 in the policy
network. The value of the CV aR coefficient λCV aR is set as given by Table 1 after a coarse
hyperparameter search. All other hyperparameters corresponding to the GAIL component of the
algorithm are set identical to those used in [Ho and Ermon, 2016] and their repository [OpenAI-GAIL,
2017] for all the experiments. The value of α in the CV aR term is set to 0.9 and its lone parameter,
ν, is trained by batch gradient descent with learning rate 0.01.
5
Evaluation Metrics
In this section we define the metrics we use to evaluate the efficacy of RAIL at reducing the tail risk
of GAIL learned policies. Given an agent A’s policy πA we roll out N trajectories T = {ξi }N
i=1
from it and estimate V aRα and CV aRα as defined in Section 3.1. V aRα denotes the value under
Table 1: Hyperparameters for the RAIL experiments on various continuous control tasks from OpenAI
Gym. For a fair comparison, the number of training iterations and expert trajectories are same as
those used by [Ho and Ermon, 2016].
Task
Reacher-v1
HalfCheetah-v1
Hopper-v1
Walker-v1
Humanoid-v1
#training
iterations
200
500
500
500
1500
6
#expert
trajectories
18
25
25
25
240
λCV aR
0.25
0.5
0.5
0.25
0.75
Figure 2: Convergence of mean trajectory-cost during training. The faded curves corresponds to the
original value of mean trajectory-cost which varies highly between successive iterations. The data is
smoothened with a moving average filter of window size 21 to demonstrate the prevalent behavior and
plotted with solid curves. RAIL converges almost as fast as GAIL at all the five continuous-control
tasks, and at times, even faster.
which the trajectory-cost remains with probability α and CV aRα gives the expected value of cost
above V aRα . Intuitively, CV aRα gives the average value of cost of the worst cases that have a total
probability no more than (1 − α). The lower the value of both these metrics, the lower is the tail risk.
In order to compare tail risk of an agent with respect to the expert, E, we define percentage relativeV aRα as follows:
V aRα (A|E) = 100 ×
V aRα (E) − V aRα (A)
%
|V aRα (E)|
(10)
Similarly, we define percentage relative-CV aRα as:
CV aRα (A|E) = 100 ×
CV aRα (E) − CV aRα (A)
%
|CV aRα (E)|
(11)
The higher these numbers, the lesser is the tail risk of agent A. We define Gain in Reliability (GR) as
the difference in percentage relative tail risk between RAIL and GAIL agents.
GR-V aR = V aRα (RAIL|E) − V aRα (GAIL|E)
GR-CV aR = CV aRα (RAIL|E) − CV aRα (GAIL|E)
(12)
(13)
Table 2: Comparison of expert, GAIL, and RAIL in terms of the tail risk metrics - V aR0.9 and
CV aR0.9 . All the scores are calculated on samples of 50 trajectories. With smaller values of V aR
and CV aR, our method outperforms GAIL in all the 5 continuous control tasks and also outperforms
the expert in many cases.
Dimensionality
Observation Action
Reacher-v1
11
2
Hopper-v1
11
3
HalfCheetah-v1
17
6
Walker-v1
17
6
Humanoid-v1
376
17
Environment
Expert
5.88
-3754.71
-3431.59
-5402.52
-9839.79
VaR
GAIL
9.55
-1758.19
-2688.34
-5314.05
-2641.14
7
Ours
7.28
-3745.90
-3150.31
-5404.00
-9252.29
Expert
6.34
-2674.65
-3356.67
-2310.54
-4591.43
CVaR
GAIL
13.25
-1347.60
-2220.64
-3359.29
-1298.80
Ours
9.41
-3727.94
-2945.76
-3939.99
-4640.42
Table 3: Values of percentage relative tail risk measures and gains in reliability on using RAIL over
GAIL for different continuous control tasks.
V aR0.9 (A|E)(%)
CV aR0.9 (A|E) (%)
Environment
GR-VaR (%)
GR-CVaR (%)
GAIL
RAIL
GAIL
RAIL
Reacher-v1
-62.41
-23.81
38.61
-108.99
-48.42
60.57
Hopper-v1
-53.17
-0.23
52.94
-49.62
39.38
89.00
HalfCheetah-v1 -21.66
-8.20
13.46
-33.84
-12.24
21.60
Walker-v1
-1.64
0.03
1.66
45.39
70.52
25.13
Humanoid-v1
-73.16
-5.97
67.19
-71.71
1.07
72.78
6
Experimental Results and Discussion
In this section, we present and discuss the results of comparison between GAIL and RAIL. The
expert’s performance is used as a benchmark. Tables 2 and 3 present the values of our evaluation
metrics for different continuous-control tasks. We set α = 0.9 for V aRα and CV aRα and estimate
all metrics with N = 50 sampled trajectories (as followed by [Ho and Ermon, 2016]). The following
are some interesting observations that we make:
• RAIL obtains superior performance than GAIL at both tail risk measures – V aR0.9 and
CV aR0.9 , without increasing sample complexity. This shows that RAIL is a superior choice
than GAIL for imitation learning in risk-sensitive applications.
• The applicability of RAIL is not limited to environments in which the distribution of
trajectory-cost is heavy-tailed for GAIL. [Rockafellar and Uryasev, 2000] showed that if
the distribution of the risk variable Z be normal, CV aRα (Z) = µZ + a(α)σZ , where a(α)
is a constant for a given α, µZ and σZ are the mean and standard deviation of Z. Thus, in
the absence of a heavy tail, minimization of CV aRα of the trajectory cost aids in learning
better policies by contributing to the minimization of the mean and standard deviation of
trajectory cost. The results on Reacher-v1 corroborate our claims. Although the histogram
does not show a heavy tail (Figure 3 in Appendix B), the mean converges fine (Figure 2) and
tail risk scores are improved (Table 2) which indicates the distribution of trajectory-costs is
more condensed around the mean than GAIL. Thus we can use RAIL instead of GAIL, no
matter whether the distribution of trajectory costs is heavy-tailed for GAIL or not.
• Figure 2 shows the variation of mean trajectory cost over training iterations for GAIL and
RAIL. We observe that RAIL converges almost as fast as GAIL at all the continuous-control
tasks in discussion, and at times, even faster.
• The success of RAIL in learning a viable policy for Humanoid-v1 suggests that RAIL is
scalable to large environments. Scalability is one of the salient features of GAIL. RAIL
preserves the scalability of GAIL while showing lower tail risk.
RAIL agents show lesser tail risk than GAIL agents after training has been completed. However it
still requires the agent to act in the real world and sample trajectories (line 3 in Algorithm 1) during
training. One way to rule out environmental interaction during training is to make the agent act in a
simulator while learning from the expert’s real-world demonstrations. The setting changes to that of
third person imitation learning [Stadie et al., 2017]. The RAIL formulation can be easily ported to
this framework but we do not evaluate that in this paper.
7
Conclusion
This paper presents the RAIL algorithm which incorporates CV aR optimization within the original
GAIL algorithm to minimize tail risk and thus improve the reliability of learned policies. We report
significant improvement over GAIL at a number of evaluation metrics on five continuous-control
tasks. Thus the proposed algorithm is a viable step in the direction of learning low-risk policies by
imitation learning in complex environments, especially in risk-sensitive applications like robotic
surgery and autonomous driving. We plan to test RAIL on fielded robotic applications in the future.
Acknowledgments
The authors would like to thank Apoorv Vyas of Intel Labs and Sapana Chaudhary of IIT Madras for
helpful discussions. Anirban Santara’s travel was supported by Google India under the Google India
PhD Fellowship Award.
8
References
Pieter Abbeel and Andrew Y Ng. Apprenticeship learning via inverse reinforcement learning. In
Proceedings of the twenty-first international conference on Machine learning, page 1. ACM, 2004.
Pieter Abbeel and Andrew Y Ng. Inverse reinforcement learning. In Encyclopedia of machine
learning, pages 554–558. Springer, 2011.
Pieter Abbeel, Adam Coates, Morgan Quigley, and Andrew Y Ng. An application of reinforcement
learning to aerobatic helicopter flight. In Advances in neural information processing systems, pages
1–8, 2007.
Brenna D. Argall, Sonia Chernova, Manuela Veloso, and Brett Browning. A survey of robot learning
from demonstration. Robotics and Autonomous Systems, 57(5):469 – 483, 2009. ISSN 0921-8890.
doi: http://dx.doi.org/10.1016/j.robot.2008.10.024. URL http://www.sciencedirect.
com/science/article/pii/S0921889008001772.
Christopher G Atkeson and Stefan Schaal. Robot learning from demonstration. In ICML, volume 97,
pages 12–20, 1997.
Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon
Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, et al. End to end learning
for self-driving cars. arXiv preprint arXiv:1604.07316, 2016.
Mariusz Bojarski, Philip Yeres, Anna Choromanska, Krzysztof Choromanski, Bernhard Firner,
Lawrence Jackel, and Urs Muller. Explaining how a deep neural network trained with end-to-end
learning steers a car. arXiv preprint arXiv:1704.07911, 2017.
Vivek S Borkar. Q-learning for risk-sensitive control. Mathematics of operations research, 27(2):
294–311, 2002.
Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and
Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.
Yinlam Chow and Mohammad Ghavamzadeh. Algorithms for cvar optimization in mdps. In Advances
in neural information processing systems, pages 3509–3517, 2014.
Nivine Dalleh. Why is CVaR superior to VaR?(c2009). PhD thesis, 2011.
Hal Daumé, John Langford, and Daniel Marcu. Search-based structured prediction. Machine learning,
75(3):297–325, 2009.
Jimmy Ba Diederik Kingma. Adam: A method for stochastic optimization. arXiv:1310.5107 [cs.CV],
2015.
Chelsea Finn, Sergey Levine, and Pieter Abbeel. Guided cost learning: Deep inverse optimal control
via policy optimization. In International Conference on Machine Learning, pages 49–58, 2016.
Javier Garcıa and Fernando Fernández. A comprehensive survey on safe reinforcement learning.
Journal of Machine Learning Research, 16(1):1437–1480, 2015.
Paul W Glimcher and Ernst Fehr. Neuroeconomics: Decision making and the brain. Academic Press,
2013.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014.
Simon Haykin. Neural Networks: A Comprehensive Foundation. Prentice Hall PTR, Upper Saddle
River, NJ, USA, 2nd edition, 1998. ISBN 0132733501.
Matthias Heger. Consideration of risk in reinforcement learning. In Proceedings of the Eleventh
International Conference on Machine Learning, pages 105–111, 1994.
Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In Advances in Neural
Information Processing Systems, pages 4565–4573, 2016.
9
Ronald A Howard and James E Matheson. Risk-sensitive markov decision processes. Management
science, 18(7):356–369, 1972.
Ming Hsu, Meghana Bhatt, Ralph Adolphs, Daniel Tranel, and Colin F Camerer. Neural systems
responding to degrees of uncertainty in human decision-making. Science, 310(5754):1680–1683,
2005.
Investopedia. Definition of tail risk. http://www.investopedia.com/terms/t/
tailrisk.asp, 2017. Accessed: 2017-09-11.
Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In
ICML, volume 2, pages 267–274, 2002.
Sergey Levine and Vladlen Koltun. Continuous inverse optimal control with locally optimal examples.
arXiv preprint arXiv:1206.4617, 2012.
Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,
David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. CoRR,
abs/1509.02971, 2015. URL http://arxiv.org/abs/1509.02971.
Anirudha Majumdar, Sumeet Singh, Ajay Mandlekar, and Marco Pavone. Risk-sensitive inverse
reinforcement learning via coherent risk models. 2017.
Oliver Mihatsch and Ralph Neuneier. Risk-sensitive reinforcement learning. Machine learning, 49
(2-3):267–290, 2002.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare,
Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control
through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
Arne J Nagengast, Daniel A Braun, and Daniel M Wolpert. Risk-sensitive optimal feedback control
accounts for sensorimotor behavior under uncertainty. PLoS computational biology, 6(7):e1000857,
2010.
Andrew Y Ng, Stuart J Russell, et al. Algorithms for inverse reinforcement learning. In Icml, pages
663–670, 2000.
Yael Niv, Jeffrey A Edlund, Peter Dayan, and John P O’Doherty. Neural prediction errors reveal a
risk-sensitive reinforcement-learning process in the human brain. Journal of Neuroscience, 32(2):
551–562, 2012.
OpenAI-GAIL. Imitation learning github repository.
imitation.git, 2017. Accessed: 2017-06-27.
https://github.com/openai/
Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. In Advances in neural
information processing systems, pages 305–313, 1989.
Aravind Rajeswaran, Sarvjeet Ghotra, Sergey Levine, and Balaraman Ravindran. Epopt: Learning
robust neural network policies using model ensembles. 5th International Conference on Learning
Representations, 2016.
R Tyrrell Rockafellar and Stanislav Uryasev. Optimization of conditional value-at-risk. Journal of
risk, 2:21–42, 2000.
Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In Proceedings of the
thirteenth international conference on artificial intelligence and statistics, pages 661–668, 2010.
Stephane Ross and J Andrew Bagnell. Reinforcement and imitation learning via interactive no-regret
learning. arXiv preprint arXiv:1406.5979, 2014.
Stéphane Ross, Geoffrey J Gordon, and Drew Bagnell. A reduction of imitation learning and structured
prediction to no-regret online learning. In International Conference on Artificial Intelligence and
Statistics, pages 627–635, 2011.
10
Stuart Russell. Learning agents for uncertain environments. In Proceedings of the eleventh annual
conference on Computational learning theory, pages 101–103. ACM, 1998.
Andrzej Ruszczyński. Risk-averse dynamic programming for markov decision processes. Mathematical programming, 125(2):235–261, 2010.
Stefan Schaal. Learning from demonstration. In Advances in neural information processing systems,
pages 1040–1046, 1997.
John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel. Trust region
policy optimization. CoRR, abs/1502.05477, 2015. URL http://arxiv.org/abs/1502.
05477.
Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. Safe, multi-agent, reinforcement
learning for autonomous driving. arXiv preprint arXiv:1610.03295, 2016.
Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. On a formal model of safe and
scalable self-driving cars. arXiv preprint arXiv:1708.06374, 2017.
Yun Shen, Michael J Tobia, Tobias Sommer, and Klaus Obermayer. Risk-sensitive reinforcement
learning. Neural computation, 26(7):1298–1328, 2014.
David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,
Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering
the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.
Bradly C Stadie, Pieter Abbeel, and Ilya Sutskever. Third-person imitation learning. arXiv preprint
arXiv:1703.01703, 2017.
R.S. Sutton and A.G. Barto. Reinforcement Learning: An Introduction. A Bradford book. Bradford
Book, 1998. ISBN 9780262193986. URL https://books.google.co.in/books?id=
CAFR6IBF4xYC.
Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.
In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International Conference on, pages
5026–5033. IEEE, 2012.
Brian D Ziebart. Modeling Purposeful Adaptive Behavior with the Principle of Maximum Causal
Entropy. PhD thesis, Carnegie Mellon University, 2010.
Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, and Anind K Dey. Maximum entropy inverse
reinforcement learning. In AAAI, volume 8, pages 1433–1438. Chicago, IL, USA, 2008.
11
Appendix
A
Calculation of Gradients of the CVaR term
In this section we derive expressions of gradients of the CVaR term in equation 9 w.r.t. π, D, and ν.
Let us denote Hα (Dπ (ξ|c(D)), ν) by LCV aR . Our derivations are inspired by those shown by Chow
and Ghavamzadeh [2014].
• Gradient of LCV aR w.r.t. D:
∇D LCV aR = ∇D
=
ν+
1
Eξ∼π (Dπ (ξ|c(D)) − ν)+
1−α
1
Eξ∼π [∇D Dπ (ξ|c(D))1(Dπ (ξ|c(D)) ≥ ν)]
1−α
(A.1)
where 1(·) denotes the indicator function. Now,
∇D Dπ (ξ|c(D)) = ∇c Dπ (ξ|c(D)) ∇D c(D)
(A.2)
Lξ −1
∇c Dπ (ξ|c(D)) = ∇c
X
γ t c(st , at )
t=0
Lξ −1
=
X
γt
t=0
1 − γ Lξ
=
1−γ
(A.3)
Substituting equation A.3 in A.2 and then A.2 in A.1, we have the following:
∇D LCV aR
1
1 − γ Lξ
π
=
Eξ∼π
1(D (ξ|c(D)) ≥ ν)∇D c(D)
1−α
1−γ
(A.4)
• Gradient of LCV aR w.r.t. π:
∇π LCV aR
=
=
=
=
∇π Hα (Dπ (ξ|c(D)), ν)
π
1
+
∇π ν +
Eξ∼π (D (ξ|c(D)) − ν)
1−α
1
∇π Eξ∼π (Dπ (ξ|c(D)) − ν)+
1−α
1
Eξ∼π (∇π logP (ξ|π))(Dπ (ξ|c(D)) − ν)+
1−α
(A.5)
• Gradient of LCV aR w.r.t. ν:
∇ν LCV aR
= ∇ν ν +
1
Eξ∼π (Dπ (ξ|c(D)) − ν)+
1−α
1
= 1+
Eξ∼π ∇ν (Dπ (ξ|c(D)) − ν)+
1−α
1
= 1−
Eξ∼π [1(Dπ (ξ|c(D)) ≥ ν)]
1−α
12
(A.6)
B
Additional figures
Figure 3: Histogram of costs of 250 trajectories generated by a GAIL-learned policy for Reacher-v1.
The distribution shows no heavy tail. From Table 2 and Figure 2, we observe that RAIL performs as
well as GAIL even in cases where the distribution of trajectory costs is not heavy-tailed.
13
| 2cs.AI
|
LOCALLY COMPACT WREATH PRODUCTS
arXiv:1703.08880v2 [math.GR] 31 Mar 2018
YVES CORNULIER
Abstract. Wreath products of non-discrete locally compact groups are usually not locally compact groups, nor even topological groups. We introduce a
natural extension of the wreath product construction to the setting of locally
compact groups.
As an application, we disprove a conjecture of Trofimov, constructing compactly generated locally compact groups of intermediate growth without nontrivial compact normal subgroups.
1. Introduction
Let B, H be groups and let X be a H-set. The unrestricted wreath product
B¯≀X H is the semidirect product B X ⋊ H, where H permutes the copies in the
power B X . The (restricted) wreath product is its subgroup B ≀X H = B (X) ⋊H,
where B (X) is the restricted power. When X = H with action by left translation,
these are called the unrestricted and restricted standard wreath product. In
both cases, some authors also refer to the standard wreath product simply as
“wreath product”, and to the general case as “permutational wreath product”.
Originally the definition comes from finite groups, where the restricted/ unrestricted distinction does not appear. Specifically, the first occurring example was
probably the wreath product C2 ≀{1,...,n} Sn , where C2 is cyclic of order 2 and Sn is
the symmetric group. It is a Coxeter group of type Bn /Cn and thus a isomorphic
to the Weyl group in simple algebraic groups of these types.
An early use of general wreath products is the classical theorem [KK] that
every group that is extension of a normal subgroup B with quotient H embeds
into the unrestricted standard wreath product B¯≀H. See [CC, Th. 6.2] for a
topological version (with B compact and H discrete), as well as [Re] for a more
subtle generalization in the case of profinite groups.
In geometric group theory, the restricted wreath product occurs more naturally:
indeed it is finitely generated as soon as B, H are finitely generated and X has
finitely many orbits (while the unrestricted is uncountable as soon as X is infinite
and B nontrivial).
Still, the definition does not immediately generalize to locally compact groups.
Indeed, for X infinite, the power B X fails to be locally compact as soon as B is
Date: March 24, 2018.
2010 Mathematics Subject Classification. Primary 20E22, Secondary 20F69, 22D05, 22D10.
Supported by ANR 12-BS01-0003-01 GDSous.
1
2
YVES CORNULIER
noncompact, and the restricted power B (X) fails to be locally compact as soon as
B is nondiscrete. This bad behavior is essentially well-known. For instance for
the standard unrestricted wreath product, it was observed in [D] that, for B 6= 1
and H non-discrete the obvious product topology is a not even a group topology
on B ≀ H.
The purpose of this note is to indicate how the definition of wreath products
naturally extends, in the context of geometric group theory, to the setting of
locally compact groups. This is performed in §2.
This extension is natural even within the study of discrete groups. Let us
provide three illustrations.
• It is well-known that for any two finite groups F1 , F2 of the same cardinal n,
the groups F1 ≀ Z and F2 ≀ Z admit isomorphic (unlabeled) Cayley graphs, just
taking Fi ∪ {1Z } as generating subset. This means that these groups admit
embeddings as cocompact lattices in a single locally compact group, namely
the isometry group of this common Cayley graph. A natural explicit group
in which they indeed embed as cocompact lattices is the topological wreath
product Sn ≀Sn−1 Z (to be defined in §2), where Sk is the symmetric group on
k letters (see Example 2.7).
• Adrien Le Boudec [LB2] uses lattices in such wreath products to obtain two
quasi-isometric non-amenable finitely generated groups, one being simple and
the other having infinite amenable radical.
• It is a difficult question to determine which wreath products of discrete groups
G = B ≀H/L H have the Haagerup Property, assuming that B and H have
the Haagerup Property. It was proved in [CoSV] that this holds if L = 1.
Furthermore, assuming that L = N is normal, we can embed it diagonally into
H × B ≀H/N (H/N); thus if in addition H/N has the Haagerup Property, then
G also has the Haagerup Property. Considering topological wreath products
allows us to extend this result to the case when L is a commensurated subgroup:
Theorem 1.1. Assume that B, H have the Haagerup Property, that L is a
commensurated subgroup of H such that the relative profinite completion H ⋌ L
(which is non-discrete in general, see §3.2) has the Haagerup Property. Then
the wreath product B ≀H/L H also has the Haagerup Property.
(All relevant definitions are given in §3.) This is a particular case of Theorem
3.3, which applies to more general (non-discrete) groups. An instance where
it applies is when H = SL2 (Z[1/k]) and L = SL2 (Z[1/ℓ]), where ℓ divides k.
It is not covered by the previously known results (except in the trivial case
when k divides some power of ℓ, in which case Z[1/ℓ] = Z[1/k]). Let us also
mention that Theorem 3.3 also includes a statement about Property PW, a
combinatorial stronger analogue of the Haagerup Property.
Finally, using one instance of this wreath product construction, we obtain:
LOCALLY COMPACT WREATH PRODUCTS
3
Theorem 1.2 (See Theorem 5.1). There exists a totally disconnected, compactly generated, locally compact group that has subexponential growth and is not
compact-by-discrete (i.e., has no compact open normal subgroup).
This relies on the construction by Bartholdi and Erschler [BE] of some (discrete) wreath products of subexponential growth. This disproves a conjecture of
Trofimov [Tro1, (**), p. 120] about the structure of vertex-transitive graphs, see
§5.2. This conjecture is also the main subject of discussion in the more recent
[Tro3].
Outline.
• In §2 we introduce semirestricted wreath products, which is the promised
natural extension of wreath products to the setting of locally compact
groups.
• In §3 we prove a stability result for the Haagerup Property and its combinatorial strengthening Property PW, including Theorem 1.1 as a particular case.
• In §4, we describe the subgroup of bounded element and the polycompact
radical for arbitrary semirestricted locally compact wreath products. This
is used in a very particular case to obtain that this subgroup is trivial, so
as to prove that one group is not compact-by-discrete, in §5.
• In §5, we prove Theorem 1.2, explain why it disproves Trofimov’s conjecture, and ask some further open questions on locally compact groups of
intermediate growth.
• In §6, we extend some results of infinite presentability to the locally
compact setting, and we also consider them in the analogous context of
wreathed Coxeter groups.
• In §7, we present a variant of the construction of §2, relying on a commensurating action of the acting group.
Acknowledgement. I thank Adrien Le Boudec for his interest and motivating
discussions. I thank Pierre-Emmanuel Caprace for pointing me out Trofimov’s
conjecture as well as useful remarks. I thank the referee for a careful reading and
useful references.
2. Wreath products in the locally compact context
We wish to extend the wreath product B ≀X H to locally compact groups. In
all the following, B, H are groups and X is an H-set; the group H acts on B X
by h · f (x) = f (h−1 x).
We begin with the easier case when B is still assumed to be discrete; now H is
a locally compact group. On the other hand, we still assume that X is discrete:
X is a continuous discrete H-set. This means that one of the following equivalent
conditions is fulfilled:
4
YVES CORNULIER
• the action of H on X is continuous (that is, the action map H × X → X
is continuous)
• for every x ∈ X, the stabilizer Hx is open in H; F
• X is isomorphic as an H-set to a disjoint union i∈I H/Li , where (Li )i∈I
is a family of open subgroups of H.
Then the action of H on the discrete group B (X) (restriction of the action on
B X ) is continuous. Then the semidirect product B (X) ⋊ H is a locally compact
group for the product topology. Note that when H is non-discrete, this does not
include the standard wreath product; in this setting, the closest generalization is
the case when X = H/L with L compact open.
Now let us deal with the general case. We know that the restricted wreath
product behaves well when B is discrete and the unrestricted wreath product
behaves well when B is compact. The natural definition consists in interpolating
between restricted and unrestricted wreath products. We first define it with no
topological assumption:
Definition 2.1. Let B, H, X be as above and let A be a subgroup of B. We first
define the semirestricted power
B X,A = {f ∈ B X : f (x) ∈ A, ∀∗ x ∈ X},
where ∀∗ means “for all but finitely many”. The semirestricted wreath product is defined as the semidirect product
X,A
B ≀A
⋊ H.
X H = B
Remark 2.2. The semirestricted power is a particular case of the (semi)restricted
product, which underlies the classical notion of Adele group, and is also considered more generally (and with closer motivation) in [BCGM, §3.1], to notably
construct non-cocompact lattices in some metabelian locally compact groups.
An instance of semirestricted power (attributed to the author) appears in a
paper of Eisenmann and Monod [EM, §3]: namely, F is a finite perfect group
and K a nontrivial subgroup such that F is not normally generated by any
element of K, but that generates F normally. Then for X infinite, F X,K is
a perfect locally compact group, with no infinite discrete quotient, that is not
topologically normally generated by any element.
Locally compact wreath products B ≀H/L H with B discrete, but H arbitrary,
are mentioned in [GM, Theorem C].
An instance of semirestricted wreath product is mentioned in [LB1, Prop. 6.14].
Next, the general definition in the locally compact setting is when A is compact
open in B.
The following lemma is standard.
Lemma 2.3 ([CoH], Prop. 8.2.4). Let G be a group and H a subgroup. Let
T be a group topology on H. Suppose that every conjugation in G restricts to
a continuous isomorphism between two open subgroups of H. Then there is a
LOCALLY COMPACT WREATH PRODUCTS
5
unique topology on G making H open with the induced topology coinciding with
T.
Proposition 2.4. Suppose that B is a locally compact group and A a compact
open subgroup and X a (discrete) set. There is a unique structure of topological
group on B X,A that makes the embedding of AX a topological isomorphism to an
open subgroup. It is locally compact.
Suppose in addition that H is a locally compact group and that the H-action
on X is continuous (i.e., has open point stabilizers). Then there is a unique
X,A
structure of topological group on B ≀A
⋊ H that makes it a topological
X H = B
semidirect product.
Proof. The first fact follows from Lemma 2.3. Let us now check that H acts
continuously on B X,A . If hi → h, wi → w, then for i large enough, we can write
wi = αi w with αi ∈ AX , αi → 1. Also write hi = mi h, where mi → 1. Then
hi · wi = hi · (αi w) = (hi · αi )(hi · w)
(mi · (h · αi ))(mi · (h · w)).
We have νi = h · αi → 1, and hence mi · νi → 1: indeed, for all x ∈ X,
−1
mi · νi (x) = νi (m−1
i x); since mi → 1, for large i we have mi x = x, whence the
fact. Similarly mi · (h · w) tends to h · w and since B X,A is a topological group,
we obtain hi · wi → h · w.
Recall that a homomorphism between locally compact groups is copci if it is
continuous, proper, with cocompact image.
Proposition 2.5. Let B1 , B2 be locally compact groups with compact open subgroups A1 , A2 , and u : B1 → B2 be a continuous homomorphism mapping A1 into
A2 . This yields continuous homomorphisms
f : B1X,A1 → B2X,A2 ,
A2
1
f ′ : B1 ≀A
X H → B2 ≀X H.
Consider the induced map ū : B1 /A1 → B2 /A2 .
(1) Suppose that X 6= ∅. Then f is proper if and only if f ′ is proper, if and
only if ū is injective (that is, u−1(A2 ) = A1 ).
(2) Suppose that X is infinite. Then f has cocompact image if and only f ′
has cocompact image, if and only if ū is surjective (that is, the composite
map B1 → B2 → B2 /A2 is surjective).
(3) Thus for X infinite, f is copci if and only if ū is bijective.
Proof. It is enough to check everything for f . The inverse image of the compact
open subgroup A2 X is (u−1 (A2 ))X,A1 , and is compact if and only if u−1 (A2 ) = A1 .
This yields the first part.
For the second, if M is the projection in B2 /A2 of the image, then the projec(X)
tion in (B2 /A2)(X) = B2X,A2 /AX
. It has finite index only
2 of the image is M
X
when M = B2 /A2 , and since A2 is compact, conversely if M = B2 /A2 then
cocompactness follows.
6
YVES CORNULIER
Remark 2.6. Since A is a compact open subgroup of B, the unit connected com◦
ponent B ◦ = A◦ is compact. We readily see that (B ≀A
X H) is the unrestricted
◦¯
◦
◦ X
◦
wreath product B ≀X H ≃ (B ) × H .
Example 2.7. Fix any finitely generated group Γ. Let F be a finite group, and
fix a bijection of F with {1, . . . , n}. The left action of F on itself yields a homomorphism F → Sn , inducing a bijection F → Sn /Sn−1 , where Sn−1 is any
point stabilizer. Hence, by Proposition 2.5, it induces an embedding of F ≀ Γ into
Sn ≀Sn−1 Γ as a cocompact lattice.
It would thus be interesting to further investigate these groups Sn ≀Sn−1 Γ.
Actually, the isometry group H of the Cayley graph of Γ, with respect to some
finite generating subset S, can turn out to be larger than Γ, e.g., non-discrete
(e.g., when Γ is free over S and S contains at least two elements). Then, the
isometry group of the Cayley graph of F ≀ Γ (with respect to F ∪ S) includes a
S
larger subgroup including Sn ≀Sn−1 Γ, namely Sn ≀L n−1 H, where L is the (compact)
stabilizer of 1 ∈ Γ in the isometry group of the Cayley graph of Γ. We can expect
this to often coincide with the full isometry group of the given Cayley graph of
F ≀ Γ.
Note that for the wreath product C ≀ Z, this only yields an embedding into an
overgroup of finite index, while there are natural known non-discrete envelopes,
see Example 7.2 for q = 2.
The semirestricted wreath product construction preserves unimodularity:
Proposition 2.8. The semirestricted wreath product B ≀A
X H is unimodular as
soon as B and H are unimodular.
Proof. Since A is open in B and B is unimodular, A is unimodular and conjugation by B preserves locally its Haar measure around 1. It follows that conjugation
of AX by an element of the form δx (b) (mapping x to b and other elements of
X to 1) preserves locally the Haar measure of AX , and hence this holds for all
elements of B X,A . Hence B X,A .
Clearly H preserves the Haar measure of AX , and hence it locally preserves
the Haar measure of B X,A . Since H is unimodular, it follows that the semidirect
product is also unimodular.
Remark 2.9. Some authors, such as Klopsch [Kl, §4.3] refer to a possible notion
of profinite wreath product, defining it in a particular case.
Let us define it here, calling it compact wreath product; it will not be used
elsewhere in the paper and is distinct from the constructions we consider. Namely,
let B be an abelian compact group and H a profinite group. The compact wreath
product Bˆ≀H is the projective limit of B¯≀P , where P ranges over (Hausdorff) finite
quotients of P . It is mentioned in the particular case: B cyclic of order p and H
the p-adic group Zp in [Kl, §4.3]; the same example also occurs in [DDMS].
The main drawback of this construction, namely the restriction to B abelian,
is due to the fact that when P is a finite quotient of K and P ′ a quotient of P ,
LOCALLY COMPACT WREATH PRODUCTS
7
there is a canonical homomorphism from B ≀ P onto B ≀ P ′ only when B is abelian.
Its main advantage is that it does not refer to any choice of open subgroup in B
and does not require that H is discrete.
3. Haagerup and PW Properties
3.1. Definition of Haagerup and PW Properties. Recall that locally compact group has the Haagerup Property if the function 1 on G admits an approximation, uniformly on compact subsets, by continuous positive definite functions
vanishing at infinity. A locally compact group has the Haagerup Property if and
only if all its open, compactly generated subgroups have the Haagerup Property.
For these compactly generated subgroups, or more generally for σ-compact locally compact groups, the Haagerup Property is equivalent to the existence of a
proper continuous conditionally negative definite function, or equivalently of a
metrically proper affine isometric action on a Hilbert space. All these facts are
due to Akemann and Walter [AW].
Also recall that a locally compact group has Property PW if it admits a
continuous action on a discrete set X with a subset M ⊂ X such that the function
g 7→ ℓ(g) = #(M△gM) takes finite values and is proper. This notion has been
widely considered (at least for discrete group actions) before being given a name
in [CoSV] and being studied in [Co3] (where in particular it is checked that ℓ is
automatically continuous). Clearly Property PW implies σ-compactness. It is
equivalent to the existence of a metrically proper continuous action on a CAT(0)
cube complex.
3.2. Commensurated subgroups and relative profinite completion. Recall that two subgroups of a group are commensurate if their intersection has
finite index in both; a subgroup is commensurated if its conjugates are pairwise
commensurate. It is a classical observation that a subgroup L of a group H is
commensurated if and only if L has finite orbits on H/L.
Definition 3.1. Let H be a locally compact group and L be an open subgroup
of H. The relative profinite completion of (H, L) is the projective limit
H ⋌L= ←
lim
− H/M, where M ranges over the open finite index subgroups of L.
Endowed with the projective limit topology, this is a locally compact space
(regardless of L being commensurated) with continuous (left) H-action and a
canonical continuous H-equivariant map H → H ⋌ L. If, in addition, L is
commensurated, then there is a unique continuous group law making it a group
homomorphism.
The closure L ⋌ L of the image of L in H ⋌ L is an open subgroup and
L → L ⋌ L is the usual profinite completion of L. Note that the induced map
H/L → (H ⋌L)/(L⋌L) is a bijection, so we denote the latter as H/L if necessary.
8
YVES CORNULIER
Remark 3.2. a) The relative profinite completion was introduced by Belyaev [Be];
it is sometimes referred as Belyaev completion, or profinite completion of H localized at the subgroup L (all references I am aware of assume that H is discrete).
b) Let now again L be an open commensurated subgroup of H. A closely
related construction is the Schlichting completion H ⋌Schl L: this is the closure
of the image of H in the Polish group of permutations of H/L. Actually, this is
canonically the quotient of the relative profinite completion by the core of L⋌L in
H ⋌L (that is, the largest normal subgroup of H ⋌L included in the closure L⋌L
of the image of L). In particular, this canonical quotient map H ⋌ L → H ⋌Schl L
has a compact kernel. (See for instance [RW, §4] for these facts.)
c) The Schlichting completion is sometimes called relative profinite completion.
This choice is, in my opinion, confusing and we did not follow it, notably because
H ⋌ H is the profinite completion, while H ⋌Schl H is the trivial group. An object
called “relative profinite completion” should include the usual profinite completion as a particular case. The Schlichting completion is a practical construction,
but is a derived object of the more fundamental relative profinite (or Belyaev)
completion.
3.3. The stability results.
Theorem 3.3. Consider a semirestricted wreath product G = B ≀A
X H (H, B are
locally compact group, X an H-set with open point stabilizers, A is a compact
open subgroup of B).
1) Assume that B, H have the Haagerup Property. Assume that for every point
stabilizer L of X, L is a commensurated subgroup of H and the relative profinite
completion H ⋌L has the Haagerup Property. Then G has the Haagerup Property.
2) Assume that G is compactly generated (when X is non-empty and A 6= B,
this means that H, B are compactly generated and X has finitely many H-orbits).
Then the previous statement holds true when all occurrences of “Haagerup Property” are replaced with “Property PW”.
The proof will be by reduction to the case when X = H/L, with L compact
open. This latter case can be viewed as a locally compact extension of the case
of standard wreath products (slightly more general in the discrete case, where it
covers the case with finite stabilizers).
Lemma 3.4. Consider a semirestricted wreath product G = B ≀A
H/L H, with L
compact open in H. Assume that B, H have the Haagerup Property. Then G has
the Haagerup Property.
Proof. First recall that a H-invariant walling on H/L means an H-invariant
H/L
Radon measure µ on the locally compact space 2∗
= 2H/L r {∅, H/L}. To
such a walling, we can associate the pseudo-distance dµ on H/L defined by
dµ (x, x′ ) = µ M ⊂ H/L : M ⊢ {x, x′ } .
LOCALLY COMPACT WREATH PRODUCTS
9
Here ⊢ reads as “cuts” and M ⊢ N means that both intersections M ∩ N and
M ∩ N c (N c being the complement) are non-empty.
Let us now prove the result. Using that the Haagerup property is stable under
directed unions of open subgroups [CCJJV, Prop. 6.1.1], we first reduce to the
case when G is σ-compact. Namely, we need to show that every σ-compact open
subgroup U of the given semirestricted product G = B H/L,A ⋊H is included in an
open subgroup that is also, after modding out by a compact normal subgroup, a
semirestricted product of the same form, but in addition σ-compact. (Taking the
quotient by a compact normal subgroup does not matter because the Haagerup
property is clearly invariant under taking extensions by compact kernels.)
Denote by (B/A)(H/L) the set of functions from H/L to B/A with finite support, where the support is the set of points on which the value differs from the
basepoint of B/A (that is, the trivial left A-coset). The projection B → B/A
induces a continuous projection ρ : B H/L,A → (B/A)(H/L) . Also denote by π1 and
π2 the projections from B H/L,A ⋊ H to B H/L,A and H defined by π1 (f, h) = f ,
π2 (f, h) = h. These are continuous maps.
Since U is σ-compact and (B/A)(H/L) is discrete, ρ(π1 (U)) is countable, and
hence included in (B1 /A)(V /L) for some open, σ-compact subgroup B1 of B, including A, and some open, σ-compact subgroup V of H, including L ∪ π2 (U).
Then U is included in the open subgroup B1 ≀A
H/L V , and actually sits in a smaller
(V /L)
open subgroup, namely the semidirect product G1 = (B1
Y = H/L r V /L. Then
V /L,A
G1 = (B1
AH/L ) ⋊ V . Define
× AY ) ⋊ V.
Observing that Y is V -invariant, we deduce that AY is a compact normal subgroup of G1 and G1 /AY is isomorphic to the σ-compact semirestricted wreath
product B1 ≀A
V /L V . This terminates the proof of the reduction to the case when
H and B are σ-compact.
Since H has the Haagerup Property and is σ-compact, it admits a continuous
proper conditionally definite function. By averaging by the compact subgroup
L, we can choose it to be right
√ L-invariant. Thus it yields a proper H-invariant
kernel ψ on H/L. Set κ = ψ. By [CoSV, Proposition 2.8(iii)] (which is strongly
inspired by Robertson-Steger [RS, Proposition 1.4]), there exists an H-invariant
walling µ on H/L such that κ = dµ .
For f ∈ B H/L,A , define SuppA (f ) = {x ∈ H/L : f (x) ∈
/ A}. Define Dµ :
G × G → R as follows: for fi ∈ B H/L,A , hi ∈ H, i = 1, 2,
Dµ (f1 h1 , f2 h2 ) = µ M ⊂ H/L : M ⊢ SuppA (f1−1 f2 ) ∪ {h1 L, h2 L} .
The difference with the case of standard wreath products from [CoSV] is that
we have {h1 L, h2 L} instead of {h1 , h2 } and SuppA instead of Supp. The proof
follows, however, the same lines.
10
YVES CORNULIER
Then Dµ is well-defined, continuous, and left-invariant. It is well-defined because the set SuppA (f1−1 f2 )∪{h1 L, h2 L} is finite and the set of M cutting a given
H/L
finite subset is a compact open subset of 2∗ . It is continuous because it is locally constant. It is obviously left-invariant by B H/L,A . Finally, it is H-invariant,
by an immediate verification using that µ is H-invariant.
Recall that a map Ψ : Y × Y → R is measure-definite if it is L1 -embeddable,
in the sense that there is a map ϕ from Y to some L1 -space such that Ψ(y, y ′) =
kϕ(y) − ϕ(y ′)k for all y, y ′ ∈ Y (that is, Ψ is isometric, although not necessarily
injective). It is well-known that measure-definite implies conditionally negative
definite (see [RS], from which the terminology is borrowed).
We claim that Dµ is measure-definite, that is, isometrically embeddable into
L1 (this, regardless of the assumption that µ is H-invariant). This fact being
closed under combinations and pointwise limits, it is enough to check it when
µ = δ is a Dirac measure at some M ⊂ H/L. Here Dδ (f1 h1 , f2 h2 ) is 1 or 0
according to whether M cuts SuppA (f1−1 f2 ) ∪ {h1 L, h2 L}. Being a {0, 1}-valued
kernel, that Dδ is a pseudo-distance is equivalent to the condition that being at
Dδ -distance zero is an equivalence relation; the easy argument, already performed
in [CoSV], is left to the reader. Next, a {0, 1}-valued pseudo-distance is obviously
measure-definite.
If Dµ (1, f h) ≤ n, then µ{M : M ⊢ SuppA (f ) ∪ {L, hL}} ≤ n. If F is a
finite subset, then µ{M : M ⊢ F } ≥ supx,y∈F µ{M : M ⊢ {x, y}. Hence for all
x, y ∈ SuppA (f ) ∪ {L, hL}, we have dµ (x, y) ≤ n. In particular, SuppA (f ) ∪ {hL}
is included in the n-ball Bn of H/L, which is finite by properness.
To conclude, we combine with another affine action. Using that B is σ-compact,
we can fix one continuous proper affine isometric action of B on a Hilbert space H;
we can suppose that A fixes 0. Let ξ be the corresponding conditionnally
negative
L
H/L,A
2
definite function on B. Then let B
act on the ℓ -sum x∈H/L Hx of copies
H/L,A
of x, the x-th component in B
acting on the x-component of the direct sum
by the given action, and trivially on other components. The action extends to an
action of the semirestricted wreath product, H permuting theP
components. The
′
resulting conditionally negative definite function is ξ (f h) = x∈H/L ξ(f (x)); it
is continuous.
If ξ ′ (f h) ≤ n, then ξ(f (x)) ≤ n for all x. In particular, if both Dµ (1, f h) ≤ n
and ξ ′ (f h) ≤ n, then SuppA (f ) ⊂ Bn and in Bn , f takes values in the compact
subset ξ −1 ([0, n]) and h ∈ Bn . We see that this forces f h to belong in some
compact subset of G. Thus the continuous conditionally negative definite function
f h 7→ Dµ (1, f h) + ξ ′ (f h) is proper on G.
Lemma 3.5. Consider a semirestricted wreath product G = B ≀A
H/L H, with L
compact open in H. Assume that B, H have the Property PW. Then G has
Property PW.
LOCALLY COMPACT WREATH PRODUCTS
11
Proof. The proof could be expected to work along the same lines as for Property
PW. The problem is that for Property PW, we do not know if it is enough to have
a left-invariant continuous proper distance that is a sum of cut-metrics: Property
PW indeed requires that the decomposition under cut-metrics be invariant by
left-translation, which is not a priori clear. This means that we have to be more
explicit, essentially following the discrete case [CoSV], or rather its version in
terms of commensurating actions given in [Co3, Proposition 4.G.2].
Consider a continuous discrete H-set Y with commensurated subset M, in the
sense that ℓ0 (h) = M △ hM is finite for all h ∈ H. We will eventually assume
that ℓ0 is proper on H. We can suppose that M is L-invariant. For y ∈ Y , define
Wy = {hL ∈ H/L : y ∈ hM} (indeed y ∈ hM is a right-H-invariant condition on
h). Let Z be the set of pairs (y, p), where y ∈ Y and p is a function H/L → B/A
with Supp(p) finite and included in the complement Wxc of Wx .
We want to define an action of B ≀A
H/L H on Z. We begin defining
h · (y, p) = (hy, h · p);
¯ W c p),
f · (y, p) = (y, f|
y
h ∈ H, f ∈ B H/L,A ,
f¯ being the image of f in (B/A)(H/L) . First, for (y, p) ∈ Z, observe that
c
SuppA (h · p) = h SuppA (p) ⊂ hWyc = Why
.
Thus h · (y, p) ∈ Z. That f · (y, p) ∈ Z is clear. That these maps define actions of
H and B H/L,A is clear; both actions have open stabilizers and thus are continuous.
To see that this extends to an action of the topological semidirect product, we
have to compute, for h ∈ H, f ∈ B H/L,A
h · f · h−1 · (y, p) = h · f · (h−1 y, h−1 · p)
¯ W c h−1 · p) = (y, h · f¯|W c p);
= h · (h−1 y, f|
h−1 y
h−1 y
we have h · f¯|W c−1 = (h · f )|hW c−1 = (h · f )|Wyc = (hf h−1 )|Wyc . Thus h · f · h−1 ·
h
y
h
y
(y, p) = (hf h−1 ) · (y, p) and this duly yields a continuous action of the semidirect
product.
Define N = M ×{1}, ℓ0 (h) = #(M △hM) and ℓ(g) = #(N △gN). Let us check
that ℓ takes finite values. By subadditivity, it is enough to check that ℓ takes finite
values on both H and B H/L,A . For h ∈ H, we have hN △ N = (hM △ M) × {1}
and hence ℓ is finite on H, where it coincides with ℓ0 . We have f · (y, p) ∈ N if
and only if y ∈ M and f¯Wyc p = 1. Thus (y, p) ∈ N r f −1 N if and only if y ∈ M,
p = 1, and f¯Wyc 6= 1. The latter means that Wyc ∩ SuppA (f ) 6= ∅. For a given
element hL ∈ SuppA (f ) ⊂ H/L, the condition hL ∈ Wyc means hL ∈
/ Wy , that
is, y ∈
/ hM. Thus we have
[
N r f −1 N = {(m, 1) : m ∈
M r hM.}
hL∈SuppA (f )
−1
−1
Thus f N r N = f (N r f N) is also finite. So ℓ takes finite values on both
B H/L,A and H, and hence on all of G, by sub-additivity.
12
YVES CORNULIER
We have (M △ hM) × {1} ⊂ N △ f hN. Thus ℓ(f h) ≥ ℓ0 (h). Also since for
each fixed hL ∈ SuppA (f ), the subset N r f −1 N includes (M r hM) × {1}, we
have ℓ(f ) ≥ #(M r hM).
We can assume from the beginning that M r hM and hM r M have the
same cardinal for all h (replace if necessary Y with Y × {0, 1} and M with
M × {0} ∪ M c × {1}). Under this assumption (made for convenience), we have
2ℓ(f ) ≥ ℓ(h) for all hL ∈ SuppA (f ). In other words, SuppA (f ) is included in the
set of hL such that ℓ0 (h) ≤ 2ℓ(f ). Assuming from now one that ℓ0 is proper, this
set is finite.
Let (fn hn ) be a sequence in G, with fn ∈ B H/L,A and hn ∈ H, such that ℓ(fn hn )
is bounded, the property ℓ(fn hn ) ≥ ℓ0 (hn ) and the properness of ℓ0 ensures that
(hn ) is bounded. Hence ℓ(fn ) ≤ ℓ(fn hn ) + ℓ(hn ) is also bounded, say by k0 . Let
F (k0 ) be the set of hL ∈ H/L such that ℓ0 (h) ≤ 2k0 ; it is finite by properness of
ℓ0 . Then SuppA (fn ) is included in F (k0 ).
If B is compact, this shows that ℓ is proper on G and we are done. In general, we
consider another, simpler commensurating action of G, so as to also make use of
the assumption that H has Property PW. Namely, start from a commensurating
action of B, say on Y ′ , commensurating a subset M ′ , such that the function
ℓ1 : b 7→ #(M ′ △ bM ′ ) is proper on B. We can suppose that M ′ is A-invariant.
Let B H/L,A act on Z ′ = Y ′ × H/L by f · (y, x) = (f (x)y, x) (we can think
of Z ′ as disjoint copies of Y ′ on which the factors in B H/L,A act separately).
This action has open stabilizers and hence is continuous. It commensurates the
subset N ′ = M ′ × H/L. The group H also acts on Z ′ by permuting copies; this
action is continuous and preserves M ′ ; clearly it extends the semirestricted wreath
′
′
′
′
′
product. For g ∈ B ≀A
H/L H, define ℓ (g) = #(M △ gM ). Then ℓ (f h) = ℓ (f )
P
for all f ∈ B H/L,A and h ∈ H. Precisely, ℓ′ (f h) =
γL∈H/L ℓ1 (f (γL)). In
−1
′
particular, if ℓ (f h) ≤ k1 , then f is valued in K1 = ℓ1 ([0, k1 ]), which is compact
by properness of ℓ1 .
Therefore, if ℓ(fn hn ) + ℓ′(fn hn ) is bounded, then (hn ) is bounded, and for some
k0 as above, SuppA (fn ) is included in F (k0) and f takes values in some compact
subset K1 . This means that fn hn stays in some compact subset of B ≀A
H/L H,
showing the properness of ℓ + ℓ′ .
Proof of Theorem 3.3. If X = ∅ or B = A, the quotient of G by the compact
normal subgroup B X is isomorphic to H, and the Haagerup/PW Property follows.
We can thus suppose, in each case, that X 6= ∅ and B 6= A.
Recall that a group has the Haagerup Property if and only if all its compactly
generated open subgroups do. So we can suppose that G is compactly generated
(in the PW case this is an assumption); in particular
Fn H, B are compactly generated and X has finitely
many H-orbits: X = i=1 Xi . Then G embeds as a
Q
closed subgroup in ni=1 B ≀A
Xi H, which reduces to the transitive case: X = H/L.
By assumption, L is commensurated.
LOCALLY COMPACT WREATH PRODUCTS
13
Consider the diagonal embedding G → (B ≀A
H/L (H ⋌ L)) × H. It is proper:
indeed, if Q is a compact neighborhood of 1 in H, the inverse image of the compact
neighborhood of (AH/L × (L ⋌ L)) × Q of 1 is included in the subset AH/L × Q.
(We use here that a continuous homomorphism between locally compact group
is proper if and only if the inverse image of some compact neighborhood of 1 is
compact.)
Therefore, it is enough to show that B ≀A
H/L (H ⋌L) has the Haagerup Property,
resp. Property PW. In other words, we are reduced to the case when L is compact.
This is the contents of Lemma 3.4 in the Haagerup case and Lemma 3.5 in the
PW case.
We leave the converse of Theorem 3.3, for the Haagerup Property, as a conjecture.
Conjecture 3.6. Under the assumptions that all point stabilizers are commensurated subgroups of H (and assuming A 6= B), the converse holds: if B ≀A
X H
has the Haagerup Property, then for every point stabilizer L of X, the relative
profinite completion H ⋌ L has the Haagerup Property.
Remark 3.7. Conjecture 3.6 can be reduced to the case when A = 1 and B is
discrete cyclic of prime order, and X = H/L is a transitive H-set.
Indeed, we can first choose b ∈ B r A and consider the closure of the subgroup
it generates, to assume that B has a dense cyclic subgroup. This implies that
either B is infinite cyclic, or compact abelian. Then, in the latter case AX is a
compact normal subgroup in the semirestricted wreath product, and hence we
can mod it out and still preserve the Haagerup Property. Therefore, we can
suppose A = 1. Thus we can suppose that B is discrete and cyclic (infinite or
prime order if necessary), and G = B ≀X H. Also, if the stabilizer condition fails
for some x, we can pass to the closed subgroup B ≀Y H, where Y is the H-orbit
of x.
Proposition 3.8. Conjecture 3.6 holds when L is a normal subgroup.
Proof. By Remark 3.7, we can assume that A = 1 and B is discrete abelian
nontrivial. By assumption, G = B (H/L) ⋊ H has the Haagerup Property. Thus
(B (H/L) ⋊H, B (H/L) ) has the relative Haagerup Property (assuming as we can that
G is σ-compact, this means that it admits a continuous affine isometric action on
a Hilbert space that is proper in restriction to B (H/L) . By [CoT, Corollary 5], it
follows that (B (H/L) ⋊ (H/L), B (H/L) ) also have the relative Haagerup Property.
Let ψ be a continuous conditionally positive definite function on G. If (hn ) is
any sequence in H/L leaving compact subsets and b is a nontrivial element in
B, and δ the function on H/L mapping 1 to b and other elements to 1, then
(H/L)
hn δh−1
tends to infinity. Since ψ is√proper on B (H/L) , we obtain that
n ∈ B
−1
ψ(hn δhn ) tends to infinity, and hence (since ψ is symmetric and subadditive)
that ψ(hn ) tends to infinity. Since this holds for every (hn ), we deduce that ψ is
14
YVES CORNULIER
proper on H. Actually, the argument also works for any sequence (fn hn ) with hn
leaving compact subsets of H; for a sequence fn hn with (hn ) bounded and (fn )
leaving compact subsets of B (H/L) , we directly use properness on B (H/L) to infer
that ψ(fn hn ) tends to infinity.
Remark 3.9. The converse of Theorem 3.3 in the PW case can also naturally be
asked; however I refrain to any conjecture since the analogues of the results of
[CoT] are not available in this case.
4. Polycompact and bounded radicals
Recall that in a locally compact group,
• W(G) denotes the polycompact radical, namely the subgroup generated
by all compact normal subgroups, which is also the union of all compact
normal subgroups.
• B(G) denotes the bounded radical, namely the union of all relatively
compact conjugacy classes (B stands for bounded). It is also sometimes
called “topological FC-center”.
We have W(G) ⊂ B(G) ⊃ Z(G), where the latter is the center. Beware that
W(G) and B(G) can fail to be closed, see [WY], and also precisely in the case of
semirestricted wreath products, see Examples 4.6 and 4.7.
Proposition 4.1. Consider a semirestricted locally compact wreath product G =
B ≀A
X H with A 6= B and X 6= ∅. Also
• Write X = X∞ ⊔ Xf , separating the union of all infinite and all finite
orbits.
• Let N be the kernel of the H-action on X. Define N ′ ⊃ N to be the
inverse image in H of W((H/N)δ ), where (H/N)δ denotes H/N with the
discrete topology (so that W((H/N)δ ) is the union of all finite normal
subgroups of H/N). Define N ′′ ⊃ N to be the set of elements of h acting
on X as a finitely supported permutation.
Define Ξ ⊂ H to be N if B is non-compact and Ξ = N ′ ∩ N ′′ if B is
compact.
• Define CoreB (A) to be the largest normal subgroup of B included in A.
Then
W(G) = (CoreB (A)X∞ × W(B)Xfin ,W(B)∩CoreB (A) ) ⋊ (W(H) ∩ Ξ);
B(G) = (CoreB (A)X∞ × B(B)Xfin ,B(B)∩CoreB (A) ) ⋊ (B(H) ∩ Ξ).
In particular, if X has no finite H-orbit, then
W(G) = CoreB (A)X ⋊ (W(H) ∩ N);
B(G) = CoreB (A)X ⋊ (B(H) ∩ N).
W(G) = W(B)(Xfin ) ⋊ (W(H) ∩ Ξ);
B(G) = B(B)(Xfin ) ⋊ (B(H) ∩ Ξ).
Proof. Modding out by the compact normal subgroup CoreB (A), we can reduce
to assume CoreB (A) = 1 and prove, in this case, that
LOCALLY COMPACT WREATH PRODUCTS
15
Let M be the group on the right side. We first check the inclusions ⊃ in both
cases. This follows from the following inclusions:
• W(B)(Xfin ) ⊂ W(G). For every compact normal subgroup K of B and
finite H-invariant subset F ⊂ X, W(B)F is a compact normal subgroup
and hence is included in W(G); it follows that the union of all these
subgroups (F ranging over finite H-invariant subsets of X and K among
all compact normal subgroups of B), which is precisely W(B)(Xfin ) , is
included in W(G).
• B(B)(Xfin ) ⊂ B(G). Let f belong to this subgroup. Then the support
of f being finite and included in Xfin , the union of H-orbits is a finite
H-invariant subset F . If I is the finite image of f and J is the union of
closures of conjugacy classes of I, then J is compact and the conjugacy
class of f is included in J F , which is compact.
• W(H) ∩ N ⊂ W(G): indeed this is the union of K ∩ N where K ranges
over compact normal subgroups of H. Since K ∩ N is also normal in G,
it is therefore included in W(G). Taking the union yields the inclusion.
• B(H) ∩ N ⊂ B(G): if f ∈ N, then it centralizes B X,A and hence its Hconjugacy class equals its G-conjugacy class. Hence if f ∈ B(H) ∩ N then
f ∈ B(G).
• if B is compact, W(H) ∩ N ′ ∩ N ′′ ⊂ W(G). Let h be an element in N ′ ∩
N ′′ ∩ W(H). Let M be the closure of the normal subgroup of H generated
by h. Since h ∈ N ′′ , M acts on X by finitely supported permutations.
Since h ∈ N ′ , the image of M in the group of permutations of X is finite.
Combining, the union Y of all supports of elements of M is a finite subset
of X. Since M is normal in H, Y is H-invariant. Since h ∈ W(H), M
is compact. Hence B Y ⋊ M is a compact normal subgroup of G. Hence
h ∈ W(G).
• if B is compact, B(H)∩N ′ ∩N ′′ ⊂ B(G). We argue as in the previous case.
The difference is that M is not necessarily compact; however, M/(M ∩N)
is still finite. In B Y ⋊ M, modulo the compact normal subgroup B Y , the
G-conjugacy class of f has compact closure; hence this holds in B Y ⋊ M
and hence in G.
This shows the inclusions ⊃. We now have to prove the reverse inclusions.
Denote by π the projection G → H.
• W(G) ⊂ π −1 (W(H)) and B(G) ⊂ π −1 (B(H)) is clear.
• W(G) ⊂ B(G) ⊂ π −1 (N ′ ∩ N ′′ ). Indeed, let C be a conjugacy class
included in B(G). For g ∈ B X,A and f h ∈ C, we have g(f h)g −1 =
gf (hgh−1)−1 h. Define δx (b) be the function mapping x to b ∈ B and other
elements of X to 1, and apply this to g = δx (b). Then δx (b)f hδx (b)−1 =
δx (bf (x))δhx (b−1 ). Fix b ∈
/ A. Then, when f h ranges over C and x ranges
over elements such that hx 6= x, then the x-projection of δx (b)f hδx (b)−1
is b−1 , and thus x has to range over finitely many elements only. In other
16
YVES CORNULIER
words, the set Z of x such that h(x) 6= x for some f h ∈ C is finite. This
means that for every f h ∈ C, h acts as the identity on the complement
of Z. This shows that C ⊂ π −1 (N ′ ∩ N ′′ ).
• if B is non-compact, W(G) ⊂ B(G) ⊂ π −1 (N). We prove the complementary inclusion: suppose that f h ∈ G with h ∈
/ N. Fix x such that hx 6= x.
Then
p := [f h, δx (b)] = f hδx (b)h−1 f −1 δx (b−1 ) = f δhx (b)f −1 δx (b−1 )
= δhx (f (hx)bf (hx)−1 )δx (b−1 ).
So p(x) = b−1 . Since B is non-compact, this shows that these commutators, when b ranges over B, do not have a compact closure and thus
fh ∈
/ B(G).
Now we know that the projection to H is exactly W(H) ∩ Ξ, we only have
to show the reverse inclusions for f ∈ W(G) ∩ B X,A , namely W(G) ∩ B X,A ⊂
W(B)(Xfin ) and B(G)∩B X,A ⊂ B(B)(Xfin ) . This is equivalent to the three inclusions
below.
• W(G) ∩ B X,A ⊂ W(B)X,A and B(G) ∩ B X,A ⊂ B(B)X,A : both inclusions
are immediate.
• W(G) ∩ B X,A ⊂ B(G) ∩ B X,A ⊂ B (Xfin ) . Let f belong to B(G) ∩ B X,A .
Suppose by contradiction that f is not supported by Xfin . Then there
exists x ∈ X∞ such that f (x) 6= 1. Conjugating f by some element of
the form δx (b), we can suppose that f (x) ∈
/ A. Hence the H-conjugates
of f do not have compact closure, a contradiction. This shows that f is
supported by Xf .
To see that the last statement follows, we need to check that N ′ ∩ N ′′ = N in
case there is no finite orbit. Indeed, (N ′ ∩ N) is the union of all subgroups of the
form M where M includes N with finite index, M/N is normal in H/N and acts
on X with finite support. Given such an M, the union of all its supports is finite
and H-invariant, hence empty, so M = N and finally N ∩ N ′ = N.
Corollary 4.2. W(G) = 1 if and only if W(H) ∩Ξ = 1, (Xfin 6= ∅ ⇒ W(B) = 1),
and CoreA (B) = 1.
In case Xfin = ∅(6= X), this can be restated as: W(G) = 1 if and only if
W(H) ∩ N = 1 and CoreA (B) = 1.
Corollary 4.3. Assume that Xfin is infinite. Then W(G) is closed if and only if
W(B/CoreB (A)) = 1 and W(H) ∩ N is closed.
Proof. Since both properties are unchanged when modding out by the compact
normal subgroup CoreB (A), we can suppose that CoreB (A) = 1, and the result
then immediately follows.
Remark 4.4. Note that (W(G) is closed) ⇔ (B(G) is closed) holds for arbitrary
locally compact groups. Indeed ⇐ is stated in [Co2, Prop. 2.4(ii)], but the
converse ⇒ follows from [Co2, Prop. 2.4(iv)].
LOCALLY COMPACT WREATH PRODUCTS
17
Remark 4.5. The assumption only excludes two trivial cases, that is A = B or
X = ∅. Actually, W(G) = AX ⋊ W(H) if A = B and W(G) = W(H) if X = ∅,
and similarly for B(G).
Example 4.6. Let (Fi ) be a family of finite groups,
on
Q each Fi acting faithfully
F
some finite set Xi . Then the
compact
group
H
=
F
acts
on
X
=
X
.
Then
i
i
L
in this case, N ′ = N ′′ L
=
Fi .LSo if B is any nontrivial finite group, we have
W(B ≀X H) = B (X) ⋊
Fi =
B ≀ Fi . In this case W(G) is not closed. This
example is very similar to the classical original example of Wu-Yu [WY].
Example 4.7. Let B be a finite group and A a proper subgroup with trivial core
(e.g., B is non-abelian of order 6 and A has order 2). Let X be a faithful Z-set
with only finite orbits. Then for G = B ≀A
X Z, both W(G) and B(G) are equal to
(X)
B
(which is not closed).
5. Locally compact groups of intermediate growth
5.1. The construction. It is natural to wonder whether there exist CGLC (compactly generated locally compact) groups of intermediate growth that are not too
close to discrete groups.
This naturally led to the following question: is every CGLC group of subexponential growth compact-by-Lie? The point is that the answer is positive for
groups of polynomial growth [Los].
A construction of Bartholdi-Erschler [BE] along with the semirestricted wreath
product construction leads to a negative answer:
Theorem 5.1. There exists a totally disconnected CGLC group of intermediate
growth that is not compact-by-discrete.
Proof. Bartholdi and Erschler [BE] have shown that the first Grigorchuk group Γ,
which has intermediate growth, has a subgroup Λ of infinite index such that for
every finite group F , the wreath product ∆ = F ≀Γ/Λ Γ has intermediate growth
as well.
Then assume that n = |F | ≥ 3. Embed, by Proposition 2.5 the latter as a
Sn−1
cocompact lattice into the semirestricted wreath product G = Sn ≀Γ/Λ
Γ. Then
the latter has W(G) = 1 by the second case of Corollary 4.2 (using either that
W(Γ) = 1 or that Γ acts faithfully on Γ/Λ). Since ∆ has intermediate growth,
so does G.
We leave the following three questions open.
Question 5.2. Does there exist a totally disconnected CGLC group of subexponential growth G that satisfies one of the following
(1) G is non-compact and has no infinite discrete quotient (as a topological
group)?
(2) G is not commable to any discrete group?
18
YVES CORNULIER
(3) (Caprace [Ca, Question 3.9]) G is topologically simple and non-discrete?
(Wesolek [CaM2, Problem 20.9.6]) G is not Wesolek-elementary? (Wesolekelementary [Wes] means that it lies in the smallest isomorphism-closed
class of locally compact groups containing the trivial groups and stable
under taking directed unions and extensions with discrete or profinite
quotients.)
Recall that commable is the “equivalence relation” between locally compact
generated by the relations: G2 is isomorphic to a the quotient of a normal compact
subgroup of G1 , and: G2 is isomorphic to a closed cocompact subgroup of G1 .
Remark 5.3. 1) If G has polynomial growth, it is compact-by-discrete [Los] and
hence cannot fulfill any of the requirements (this also follows as a consequence of
Trofimov’s results on graphs [Tro2]).
2) If G is a non-compact totally disconnected CGLC group, by [CaM1, Theorem
A], either it has an infinite discrete quotient, or it has a non-compact non-discrete
topologically simple, compactly generated subquotient S. If G has intermediate
growth, then so does S; in particular a positive answer to (1) is equivalent to
the existence of a non-discrete CGLC, topologically simple group of intermediate
growth; thus a positive answer to (1) would also answer (3).
See §7 for an alternative construction for Theorem 5.1, yielding possible candidates for Question 5.2(2).
Remark 5.4. It is well-known that a discrete-by-compact CGLC group G is automatically compact-by-discrete. Indeed, let D be a cocompact normal discrete
subgroup. Being finitely generated, its centralizer C is open. Since clearly G◦ is
compact, C admits an open compact subgroup K, and hence KD is a cocompact
open subgroup, thus has finite index. So the intersection of all conjugates of K
is also open, and thus G is compact-by-discrete.
It is also well-known that this implication does not hold for arbitrary σ-compact
locally compact groups: the group of Example 4.6 is typical counterexample.
5.2. Trofimov’s conjecture. We say that a connected graph X essentially
includes a tree if there exists an injective, Lipschitz map from the regular
trivalent tree into X (Trofimov calls this “hyperbolic” but this differs from usual
terminology).
In the following, a graph is identified with its vertex set (so the only graph structure that matters is the given adjacency relation between vertices); in particular
a graph homomorphism is understood to be a map between vertex sets mapping
adjacent vertices to adjacent or equal vertices; it also means a 1-Lipschitz map.
We call it a graph-quotient homomorphism if every edge from Y is image of an
edge in X.
Let G be a group acting on a connected graph X. We say that the action
is block-discrete if there exists a graph Y and a continuous action of G on
X, a surjective G-equivariant graph-quotient homomorphism X → Y with finite
LOCALLY COMPACT WREATH PRODUCTS
19
fibers, such that, denoting by GY the image of G in Aut(Y ), the vertex stabilizers
of the GY -action on Y are finite.
Trofimov’s conjecture can now be stated:
Conjecture 5.5 (Trofimov [Tro1, Tro3]). Let G be a group acting vertex-transitively
on a connected graph X of finite valency. Then either the action is block-discrete,
or X essentially includes a tree.
In particular, it predicts that in the case X has subexponential growth, the
action is block-discrete. On the other hand, we have the following easy observation:
Fact 5.6. Let X be a connected graph of finite valency. Let G be a locally compact group acting properly vertex-transitively on X. Then the G-action is blockdiscrete if and only if G has a compact open normal subgroup.
Proof. We only prove the implication we need, leaving the converse to the reader.
Suppose that the action is block-discrete, and let Y be as in its definition. Since
the vertex stabilizers in Y include vertex stabilizers in X, the G-action on Y is
continuous. Since the G-action on X is proper and fibers are finite (and since
X → Y is a graph-quotient homomorphism), the G-action on Y is proper. Let
W be the kernel of the G-action on Y and let H be a vertex-stabilizer. Since
H/W is finite, H is open and W is a closed subgroup of H, the normal subgroup
W is open as well. By properness, W is compact, proving the implication.
We can conclude, as a Corollary of Theorem 5.1:
Corollary 5.7. Trofimov’s above conjecture is not true.
Proof. Let G be a totally disconnected CGLC group of intermediate growth and
no compact open subgroup, as asserted in the theorem. Let X be a Cayley-Abels
graph for G (see [CoH, §2.E]): this is a connected graph of finite valency on
which G acts continuously, properly and vertex-transitively. Since G is quasiisometric to X, the graph X has subexponential growth, and in particular does
not essentially include a tree in the above sense. By the fact above, the action is
not block-discrete. So it does not satisfy the conjecture.
Remark 5.8. Part of the discussion in [Tro3] is about when the above conjecture
is specified when the action of the vertex stabilizer Gv on the 1-sphere around
one vertex Gv is specified to be, modulo its kernel, a given finite permutation
group. We have not tried to describe this permutation group in this construction.
However, we can at least say something: we can arrange the counterexample so
that the group in this finite permutation group is a 2-group. Indeed, it is enough
to construct G as a 2-group. Using the notation from the proof of Theorem 5.1,
we can make a slight change in the construction and assume that G = F ≀LΓ/Λ Γ,
where F is a nontrivial finite 2-group and L is a nontrivial subgroup of F with
trivial core (e.g., F the dihedral group of order 8 and L a non-central cyclic
20
YVES CORNULIER
subgroup of order 2). Then W(G) = 1 by Corollary 4.2, and G has intermediate
growth for the same reason as in the proof of Theorem 5.1. (Alternatively, we
can use the construction of Proposition 7.4, which yields a 2-group.)
In [Tro3, Remark 2.3], Trofimov says that he would be “rather surprised if [his]
conjecture were proved in general”, but is optimistic about the case when the
local permutation group is primitive. We do not even know if our example can
be arranged to even yield transitive local permutation groups (beware that this
depends on the choice of Cayley-Abels graph, and forces studying the structure
of the Grigorchuk group action, rather than using it as a black box).
6. Presentability
The material of this section is essentially borrowed from an expunged part,
appearing in an earlier version of [BCGS] (Arxiv v2), exclusively for discrete
groups, in keeping with [BCGS].
We first introduce the following definition, which is [BCGS, Def. 5.9] in the
discrete case.
Definition 6.1. A CGLC group H is largely related if for every epimorphism
G ։ H of a compactly presented locally compact group G onto H with discrete
kernel, the kernel admits a non-abelian free quotient.
Definition 6.2. A family (Ni )i∈I of closed normal subgroups of a locally compact
group G is independent if Ni is not included in hNj : j ∈ I r {i}i for any i ∈ I,
or equivalently if the map from 2I to the space N (G) of closed normal subgroups
of G mapping J to hNj : j ∈ Ji is injective.
([BCGS, Def. 1.2]) A CGLC group G is INIP (infinitely independently presented) if for some/every compactly presented locally compact group G0 with a
quotient map G0 ։ G with discrete kernel, the kernel is generated by an infinite
independent family of G0 -normal subgroups.
Definition 6.3. Let H be a group and L ⊂ H a subgroup. Consider the equivalence relation on H: g1 ∼ g2 if g1 belongs to the same L-double coset as g2 or
g2−1 . Let Q be the quotient of H by this equivalence relation and Q∗ = Q r {L}
(observe that L is a single equivalence class).
Now assume that H is a locally compact group, L an open subgroup, B another
locally compact group and A a compact open subgroup of B. We need to define
a locally compact group G0 with a continuous quotient homomorphism with
discrete kernel G0 → B ≀A
H/L H. In case A = 1 (so B is discrete and this is a
usual wreath product), the definition of G0 is given by the amalgamated product
(B ×L) ∗L H. As an amalgam of two locally compact groups over a common open
subgroup, this is naturally a locally compact group. As an abstract group, this
is the quotient of the free product B ∗ H by the normal subgroup generated by
commutators [B, L]; however B ∗ H is not a locally compact group in a natural
LOCALLY COMPACT WREATH PRODUCTS
21
way unless H is discrete. In general, G0 is defined as the locally compact group
G0 = (ΠB,A,H/L ) ∗AH/L (A ¯≀ H),
where
ΠB,A,H/L = {f ∈ B H/L : f (H/L r {L}) ⊂ A}
(so ΠB,A,H/L ≃ B × AH/Lr{L} ).
Note that G0 is compactly generated as soon as H, B are. From the universal
property, there is a unique homomorphism G0 → B ≀A
H/L H mapping both factors
identically. It is continuous and surjective, and has discrete kernel N. (That it
is continuous with discrete kernel follows from the fact that it restricts to the
standard embedding of the open subgroup AH/L ⋊ L.)
Observe that in G0 , the normal subgroup Ng generated by [gBg −1, B], for
g ∈ H r L, only depends on the class of g in Q∗ . It is easy to see that the Ng
generate the kernel N.
Proposition 6.4. If B 6= A, then the family of normal subgroups (Ng ) of G0
is independent, when g ranges over Q∗ . In particular, if the double coset space
L\H/L is infinite and both B and H are compactly generated, then the semirestricted wreath product G = B ≀A
H/L H is INIP, and moreover it is largely related.
This extends a result of [Co1], where it was shown under the same assumptions
(and in the discrete case) that the wreath product B ≀G/H G is infinitely presented.
Proof. In G0 , denote by H and B the obvious copies of these groups. For h ∈ H
and x = hL ∈ H/L, define Bx = hBh−1 , and Ax = hAh−1 ⊂ Bx . Note that in
G0 , the subgroup U generated by all Bx is the quotient of their free product by
the relations [Ax , By ] for x 6= y ∈ H/L.
Observe that Q∗ is obtained from L\H/L by removing one point and modding
out an action of the cyclic group of order two (by inversion). So if L\H/L is
infinite, so is Q∗ . So once we will have proved that (Ng )g∈Q∗ is independent, it
will follow that the quotient G is INIP.
∗
Lift Q∗ to a subset
S of H. If I is a subset of Q , let NI be the normal subgroup
of H generated by g∈I Ng . Then NI is generated, as a normal subgroup of H,
S
by g∈I [BgH , B]. That (Ng ) is independent means that for every s ∈ Q∗ , the
group NQ∗ r{s} is not equal to NQ∗ . We will actually show that for every s ∈ Q∗ ,
the group NQ∗ /NQ∗ r{s} is non-trivial.
By a straightforward verification, NQ∗ r{s} is generated, as a normal subgroup
of U, by
[
γ[Bgξ , Bξ ]γ −1
g∈Q∗ r{s},γ∈H
(where ξ denotes the base-point in H/L), or equivalently by
[ [
(1)
[Bγgξ , Bγξ ].
g∈Q∗ r{s} γ∈H
22
YVES CORNULIER
Let Js be the normal subgroup of U generated by
[
Bx ,
x∈H/L\{ξ,sξ}
so U/Js is naturally identified with
Ξ = (Bξ ∗ Bsξ )/h[Aξ , Bsξ ], [Bξ , Asξ ]i ≃ (Bξ × Asξ ) ∗Aξ ×Asξ (Aξ × Bsξ ).
The kernel K of U/Js → Bξ × Bsξ is discrete and free: indeed it acts trivially
on the Bass-Serre tree for this amalgam decomposition. If |B/A| ≥ 3, it is nonabelian.
Observe that NQ∗ r{s} is included in Js : indeed it is generated by commutators
[Bx , By ] with {x, y} =
6 {ξ, sξ}, so each of these commutators is contained in Js .
Thus there is a natural epimorphism U/NQ∗ r{s} ։ Bξ ∗ Bsξ . It restricts to an
epimorphism
NQ∗ /NQ∗ r{s} ։ K.
If |B/A| = 2, we can fix the argument as follows: first in this case A is normal
and hence we can mod it out and hence suppose that B is discrete. Then we show
that if s, t are distinct in Q∗ , the group NQ∗ /NQ∗ r{s,t} surjects onto a non-abelian
free group, namely the kernel of the projection Bξ ∗ (Bsξ × Btξ ) ։ Bξ × Bsξ × Btξ .
This shows that, whenever B 6= A, for every subset I ⊂ Q∗ whose complement
contains at least two elements, NQ∗ /NI has a non-abelian free quotient.
Thus if Q∗ is infinite, and if P is compactly presented group with an epimorphism π onto B ≀A
G/H G, this epimorphism factors through the projection G0 /NI
for some finite I ⊂ Q∗ . So, the kernel of π admits NQ∗ /NI as a quotient and therefore possesses a non-abelian free group as a quotient. This shows that B ≀G/H G
is largely related.
We now turn to another similar example based on Coxeter groups.
Definition 6.5. Consider a Coxeter matrix on V i.e. a symmetric matrix
µ : V × V → {1, 2, 3, . . . , ∞} with with diagonal entries equal to 1 and nondiagonal entries in {2, 3, . . . , ∞}. It defines the Coxeter group with Coxeter
presentation
D
E
W (V, µ) = (wv )v∈V | (ws wt )µ(s,t) (s,t)∈V 2 .
Let a group H act on V . Now assume that µ is H-invariant, in the sense that
µ(gs, gt) = µ(s, t) for all g ∈ H and (s, t) ∈ V 2 . This induces a natural action
of H by automorphisms on W (V, µ), so that g · ws = wgs . The corresponding
semidirect product
W (V, µ) ⋊ H
is called a wreathed Coxeter group.
When H is locally compact and acts continuously on V (which is discrete),
that is, with open stabilizers, then the wreathed Coxeter group W (V, µ) ⋊ H is
a topological group (W (V, µ) being discrete and normal).
LOCALLY COMPACT WREATH PRODUCTS
23
This group was already considered, from a different perspective, in [CoSV],
when H is discrete.
If H acts with finitely many orbits on V and is finitely generated discrete (resp.
compactly generated), then the wreathed Coxeter group W (V, µ) ⋊ H is discrete
finitely generated (resp. compactly generated).
Theorem 6.6. Assume V 6= ∅. The wreathed Coxeter group G = W (V, µ) ⋊ H
is compactly presented if and only if V has finitely many H-orbits with compactly
generated stabilizers, H is compactly presented, and the set of pairs {(v, w) ∈
V 2 : µ(v, w) < ∞} consists of finitely many H-orbits.
In particular, if V = H/L and µ has no ∞ entry, then this holds if and only
if L\H/L is finite.
If the set of pairs {(v, w) ∈ V 2 : µ(v, w) < ∞} consists of infinitely many
H-orbits, then G is INIP.
Proof. We only sketch the proof. Fix v1 , . . . , vk representative of the H-orbits in
V , with stabilizers L1 , . . . , Lk . Let for i = 1, . . . , k, let Ci = hti i be a copy of the
cyclic group of order 2. Consider the amalgam G0 of all Ci × Li and G over their
intersections: it can be constructed iteratively:
G0 = ((· · · (G ∗L1 (C1 × L1 )) ∗L2 (C2 × L2 ) · · · ) ∗Lk (Ck × Lk ))
Then G0 is compactly presented as soon as H is compactly presented and all
Li are compactly generated. Then the wreathed Coxeter group is the quotient by the relations is the quotient by rg,h = (gti g −1htj h−1 )µ(gvi ,hvj ) whenever
µ(gvi , hvj ) < ∞. Actually two such relations rg,h and rg′ ,h′ are equivalent as soon
as (gvi , hvj ) = (g ′vi , h′ vj ). Hence if there are finitely many such orbits of pairs,
then G is compactly presented.
The converse follows the same lines as the case of wreath products, relying on
a well-known result of Tits, Theorem 6.7.
Because of the similarity with the proof of Proposition 6.4, we will prove the
result in a particular case that is enough to encompass all the differences, namely
the case when H = Z = hti = X (simply transitive action). The reader is invited
to prove the general case as an exercise.
So we have to prove that in the free product Γ = ht, w|w 2 = 1i, the family
of relators rn = (wtn wt−n )µ(0,n) , for n ∈ Z1 = {n ≥ 1 : µ(0, n) < ∞}, is
independent.
If p ∈ Z1 , let Γ[p] be the group obtained by modding out Γ by all relators
rn for n 6= p, and let µ′ be the matrix obtained from µ by replacing all entries
µ(n, n + p) by ∞. We see that in Γ[p] = W (V, µ′) ⋊ Z. By Tits’ theorem, wtp wt−p
has infinite order in Γ[p], so rp 6= 1 in Γ[p]. This proves independency of the family
of relators.
24
YVES CORNULIER
Theorem 6.7 (Tits). Given a Coxeter group generated by involutions (σs )s∈S ,
subject to relators (σs σt )µ(s,t) for all s, t (where µ(s, t)(s,t)∈S×S is a Coxeter matrix), the element σs σt has order exactly µ(s, t), and every subgroup generated by
a subset (σs )s∈T is a Coxeter group over this system of generators.
This follows from [Bki, V.§4.3 Prop. 4] and [Bki, IV.§1.8 Th. 2].
Example 6.8. The group Γ of permutations of Z generated by the transposition
0 ↔ 1 and the shift n 7→ n + 1, which is isomorphic to wreathed Coxeter group
Sym0 (Z) ⋊ Z, is INIP; here Sym0 (Z) denotes the group of finitely supported
permutations of Z. That the group Γ is infinitely presented is implicit in B.H.
Neumann [Neu], who expressed it as quotient of a finitely generated group by
a properly increasing union of finite normal subgroups. Half a century later, it
was mentioned by Stëpin [Ste] as an example of a finitely generated group that is
locally embeddable in finite groups in the sense of Maltsev (this also means:
approximable by finite groups for the topology of the space of marked groups)
but is not residually finite, a combination that cannot be achieved by finitely
presented groups.
7. Variants using commensurating actions
Let X be a set and M = (Mi )i∈I be partition of X (i.e., pairwise disjoint
and covering X). Let B be a locally compact group and A = (Ai )i∈I a family
of compactQopen subgroups. Consider the subgroup of B X generated by its
subgroups i∈I Ai Mi and B (X) . Denote it by B X,M,A .
For instance, if I is a singleton {1} and M1 = X, then this is precisely B X,A .
The main motivating case is when I has two elements.
Q
The B X,M,A is endowed with the group topology making i∈I Ai Mi a compact
open subgroup. It is standard (e.g., follows from Lemma 2.3) that this is welldefined.
Let now H be a locally compact group and assume that X is a continuous
discrete H-set.
Assume that the family M is uniformly commensurated by H, in the sense
that for every h ∈ H, we have (△ denoting symmetric difference):
X
#(Mi △ hMi ) < ∞.
i
Note that if H is finitely generated, then this forces all but finitely many of
the Mi to be H-invariant; this can be extended to the case when H is compactly
generated (see Proposition 7.5), but not in general. For instance, for H = Q2
and X = Q2 /Z2 , one can consider the family (modulo Z2 ) indexed by N: M0 =
Z2 /Z2 = {0}, Mi = (2−i Z2 r 2−i+1 Z2 )/Z2 for i ≥ 1.
Under the above assumptions, the action of H on B X preserves B X,M,A , and
is continuous. Therefore the semidirect product
B ≀M,A
H = B X,M,A ⋊ H
X
LOCALLY COMPACT WREATH PRODUCTS
25
is a locally compact group.
Let us focus on a specified case I = {1, 2}, B is a finite group, A1 = {1} and
A2 = B; we can only specify M = M2 since M1 is its complement. Then denote
B X,M,A = B (X,M ] : it is just the direct product B (XrM ) × B M . The assumption
of commensuration reduces to the requirement that M is commensurated by the
H-action: M △ hM is finite for all h ∈ H. Denote B ≀(X,M ] H = B (X,M ] ⋊ H
and call it half-restricted wreath product. This important particular case
was introduced by Kepert and Willis [KW]. It was used by Bhattacharjee and
Macpherson [BM] to exhibit a compactly generated totally disconnected locally
compact group that is uniscalar but has no open compact normal subgroup.
Example 7.1. Here is one particular case where a group naturally occurs and
actually turns out to be a half-restricted wreath product. Let K be the field
Fq ((t)) of Laurent series over the finite field Fq . Then the affine group K ⋊ K∗
over K can naturally be identified with
Fq ≀(X,M ] K∗ ,
where X = K∗ /K∗1 ≃ Z (quotient by the subgroups of elements of modulus 1)
and M is the image of the closed ball of radius 1 (corresponding in Z to the set
N of non-negative integers).
This shows that this construction can produce non-unimodular groups, in contrast with Proposition 2.8.
Example 7.2. Consider X = Z × {1, 2}, M1 = (N × {1}) ⊔ ((Z r N) × {2} and
M2 its complement. Choose B = Fq (finite field), A1 = B, A2 = {0}; finally let
Z act on X by n · (m, i) = (m + n, i). Then B ≀(X,M ] H can be identified with the
semidirect product (Fq ((t)) ×Fq ((t−1 ))) ⋊Z, where the positive generator of Z acts
by multiplication by t on both sides. This group naturally includes a cocompact
lattice isomorphic to the lamplighter group Fq ≀ Z.
We now address the description of W(G) and B(G). For simplicity, let us
reduce the study of W(G) and B(G) to the case with no finite orbit.
Proposition 7.3. Suppose that H has a single
on
T infinite orbit
T X. Let N be the
−1
kernel of the H-action on X. Define C = i∈B,b∈B gAi g = i∈I CoreB (Ai ).
Then, for G = B ≀M,A
H, we have W(G) = C X ⋊ (N ∩ W(H)) and B(G) =
X
C X ⋊ (N ∩ B(H)).
Proof. Since C is compact normal in B, C X is compact normal in G, and is
included in all the terms considered; hence we can suppose that C = 1.
The subgroup N of H is normal, and on N the H- and G-conjugacy classes
coincide. It immediately follows that N ∩ W(H) ⊂ W(G) and N ∩ B(H) ⊂ B(G).
Let π be the projection G → H. Clearly π(W(G)) ⊂ W(H) and π(B(G)) ⊂
W(F C).
Let us show that B(G) ∩ B X,M,A = 1. By contradiction, let f be a nontrivial element. Then c = f (x) 6= 1 for some x. Then there exists i such that
26
YVES CORNULIER
c∈
/ CoreB (Ai ). Conjugating if necessary, we can suppose c ∈
/ Ai . Then by transitivity, the H-conjugates of f do not remain in a compact subset, and this is a
contradiction.
Let us show that π(B(G)) ⊂ N. Let f h be an element of B(G) with h ∈ H and
f ∈ B X,M,A . If π(h) ∈
/ N, there exists x ∈ X such that h(x) 6= x. For any b ∈ B,
let δx (b) be defined as in the proof of Proposition 4.1. Then, writing c = f (x)
(f h)−1 δx (b)f hδx (b)−1 = h−1 δx (c−1 bc)hδx (b−1 ) = δh−1 x (c−1 bc)δx (b−1 );
this is a nontrivial element of B(G) ∩ B X,M,A , which is a contradiction.
Gathering everything, W(G) is a subgroup whose projection to H is equal to
W(H) ∩ N, it includes W(H) ∩ N, and its intersection with B X,M,A is trivial.
Hence W(G) = W(H) ∩ N. Similarly B(G) = B(H) ∩ N.
Let Γ be the first Grigorchuk group and Λ its subgroup as in Theorem 5.1.
Then the Schreier graph Γ/Λ is known to be 2-ended. Pick one half M. Denote
by C2 the cyclic group on 2 generators.
Proposition 7.4. The embedding C2 ≀Γ/Λ Γ → C2 ≀(Γ/Λ,M ] Γ has a dense image. In
particular, G = C2 ≀(Γ/Λ,M ] Γ has intermediate growth. It has W(G) = B(G) = 1,
and in particular is not compact-by-discrete.
Proof. The density is clear and the only nontrivial point is Bartholdi-Erschler’s
theorem that the left-hand discrete group has intermediate growth. It follows
that the right-hand group has subexponential growth (bounded above by that of
the discrete one). It does not have polynomial growth, since its quotient Γ does
not. It has B(G) = 1, by Proposition 7.3.
This example is motivated by Question 5.2(2): unlike the examples in the proof
of Theorem 5.1, which by construction have a cocompact lattice, these ones do
not a priori (note that this is not a single group: several examples are provided
in [BE] and also the choice of M rather or its complement could a priori matter).
Proposition 7.5. In the setting above (beginning of the section), if H is compactly generated (or more generally, has uncountable cofinality, in the sense
that it is not the union of a properly increasing sequence of subgroups), then all
but finitely many of the Mi are H-invariant.
Proof. SDefine X ′ = X × {i}, with the component-wise action of H. Define
M = i∈I Mi × {i}. Then M is commensurated by the H-action. Since G is
compactly generated, M intersects only finitely many orbits in a non-invariant
subset [Co3, Prop. 4.B.2]. Hence M ∩ (X × {i}) is H-invariant for all but finitely
many i, and this precisely means that Mi is H-invariant.
References
[AW]
C. Akemann, M. Walter. Unbounded negative definite functions. Canad. J. Math. 33
(4) (1981) 862–871.
LOCALLY COMPACT WREATH PRODUCTS
27
[BCGM] U. Bader, P-E. Caprace, T. Gelander, Sh. Mozes. Lattices in amenable groups.
ArXiv/1612.06220, 2016.
[BCGS] R. Bieri, Y. Cornulier, L. Guyot, R. Strebel. Infinite presentability of groups and
condensation. J. Inst. Math. Jussieu 13 (2014), no. 4, 811–848.
[Be]
V. Belyaev. Locally finite groups with a finite nonseparable subgroup (Russian).
Sibirsk. Mat. Zh. 34 (1993), no. 2, 23–41, 226, 233; English translation in Siberian
Math. J. 34 (1993), no. 2, 218–232 (1994)
[BE]
L. Bartholdi, A. Erschler. Growth of permutational extensions. Invent. Math. 189
(2012), 431–455.
[Bki]
N. Bourbaki. Élements de mathématique. Groupes et algèbres de Lie, chapitres 4 à
6. Hermann, Paris, 1968.
[BM]
M. Bhattacharjee, D. MacPherson. Strange permutation representations of free
groups. J. Aust. Math. Soc. 74 (2003), no. 2, 267–285.
[Ca]
P-E. Caprace. Non-discrete simple locally compact groups. 2016, to appear in the
Proceedings of the 7th European Congress of Mathematics.
[CaM1] P-E. Caprace, N. Monod. Decomposing locally compact groups into simple pieces.
Math. Proc. Cambridge Philos. Soc. 150(1) (2011) 97–128
[CaM2] P-E. Caprace and N. Monod. Future directions in locally compact groups. p343-355
in “New directions in locally compact groups”, Edts P-E. Caprace and N. Monod.
Cambridge Univ. Press, 2018.
[CC]
P-E. Caprace, Y. Cornulier. On embeddings into compactly generated groups. Pacific
J. Math. 269 (2014), no. 2, 305–321.
[CCJJV] P.-A. Cherix, M. Cowling, P. Jolissaint, P. Julg, and A. Valette. Groups with the
Haagerup property. Progress in Mathematics, vol. 197. Birkhäuser Verlag, Basel,
2001.
[Co1]
Y. Cornulier. Finitely presented wreath products and double coset decompositions.
Geom. Dedicata 122 (2006), 89–108.
[Co2]
Y. Cornulier. Commability and focal locally compact groups. Indiana Univ. Math. J.
64 (2015), no. 1, 115–150.
[Co3]
Y. Cornulier. Group actions with commensurated subsets, wallings and cubings.
ArXiv 1302.5982, 2016.
[CoH]
Y. Cornulier, P. de la Harpe. Metric geometry of locally compact groups. EMS Tracts
in Mathematics Vol. 25, European Math. Society, 2016.
[CoSV] Y. Cornulier, Y. Stalder, A. Valette. Proper actions of lamplighter groups associated
with free groups. C. R. Acad. Sci. Paris, Ser. I 346 (2008), no. 3-4, 173–176.
[CoSV] Y. Cornulier, Y. Stalder, A. Valette. Proper actions of wreath products and generalizations. Trans. Amer. Math. Soc. 364(6) (2012) 3159–3184.
[CoT]
Y. Cornulier, R. Tessera. A characterization of relative Kazhdan property T for
semidirect products with abelian groups. Ergodic Theory Dynam. Systems 31 (2011),
no. 3, 793–805.
[D]
J. Durbin. On locally compact wreath products. Pacific J. Math. 57 (1975), no. 1,
99–107.
[DDMS] J. Dixon, M. du Sautoy, A. Mann, D. Segal. Analytic pro-p groups. Second edition.
Cambridge Studies in Advanced Mathematics, 61. Cambridge University Press, Cambridge, 1999.
[EM]
A. Eisenmann, N. Monod. Normal generation of locally compact groups. Bull. Lond.
Math. Soc. 45 (2013), no. 4, 734–738.
[GM]
M. Gheysens, N. Monod. Fixed points for bounded orbits in Hilbert spaces. Ann. Sci.
ENS. 4e série 50 (2017), 131–156.
28
[KK]
[Kl]
[KW]
[LB1]
[LB2]
[Los]
[Mal]
[Neu]
[Re]
[RS]
[RW]
[Ste]
[Ti]
[Tro1]
[Tro2]
[Tro3]
[Wes]
[WY]
YVES CORNULIER
M. Krasner, L. Kaloujnine. Produit complet des groupes de permutations et problème
d’extension de groupes. III. Acta Sci. Math. Szeged 14 (1951), 69–82.
B. Klopsch. Abstract quotients of profinite groups, after Nikolov and Segal. In: Arbeitsgemeinschaft: Totally Disconnected Groups, Oberwolfach Reports No. 47/2014;
arXiv 1601.00343, 2016.
A. Kepert, G. Willis. Scale functions and tree ends. J. Aust. Math. Soc. 70 (2001),
no. 2, 273–292.
A. Le Boudec. Groups acting on trees with almost prescribed local action. Comm.
Math. Helv., Vol. 91, Issue 2 (2016), 253–293.
A. Le Boudec. Amenable uniformly recurrent subgroups and lattice embeddings.
arXiv:1802.04736, 2018.
V. Losert. On the structure of groups with polynomial growth. Math. Z. 195 (1987)
109–117.
A. I. Maltsev. Algebraicheskie sistemy (Russian). Posthumous edition. Edited by D.
Smirnov and M. Taiclin Izdat. “Nauka”, Moscow 1970. English translation: Algebraic
systems. Die Grundlehren der mathematischen Wissenschaften, Band 192. SpringerVerlag, New York-Heidelberg, 1973.
B.H. Neumann. Some remarks on infinite groups. J. London Math. Soc. 1(2) (1937)
120–127.
V. Remeslennikov. Imbedding theorems for profinite groups. Izvestiya Rossiiskoi
Akademii Nauk. Seriya Matematicheskaya 43 (1979), no. 2, 399–417.
G. Robertson, T. Steger. Negative definite kernels and a dynamical characterization
of property T. Ergodic Theory Dynam. Systems 18 (1998), no. 1, 247–253.
C. Reid, Ph. Wesolek. Homomorphisms into totally disconnected, locally compact
groups with dense image. arXiv:1509.00156.
A. M. Stëpin. Approximability of groups and group actions. Uspekhi Mat. Nauk 38
(1983), no. 6(234), 123–124 (Russian). English translation: Russian Math. Surveys
38 (1983), no. 6, 131–132.
J. Tits. Automorphismes à déplacement borné des groupes de Lie. Topology 3 Suppl.
1 (1964), 97–107.
V. Trofimov. Growth functions of permutation groups. (Russian) Groups and other
algebraic systems with finiteness conditions, 118–138, Trudy Inst. Mat., 4, ”Nauka”
Sibirsk. Otdel., Novosibirsk, 1984.
V. Trofimov. Graphs with polynomial growth. Mat. Sb. (N.S.) 123(165) (1984), no.
3, 407–421 (Russian). English translation: Math. USSR Sb. 51 (1985) 405–417.
V. Trofimov. Vertex stabilizers of graphs and tracks. I. European J. Combin. 28
(2007), no. 2, 613–640.
Ph. Wesolek. Elementary totally disconnected locally compact groups. Proc. Lond.
Math. Soc. (3) 110 (2015), no. 6, 1387–1434.
T. S. Wu and Y. K. Yu. Compactness properties of topological groups. Michigan
Math. J. 19(4) (1972), 299–313.
CNRS and Univ Lyon, Univ Claude Bernard Lyon 1, Institut Camille Jordan,
43 blvd. du 11 novembre 1918, F-69622 Villeurbanne
E-mail address: [email protected]
| 4math.GR
|
Deterministic Partially Dynamic Single Source
Shortest Paths in Weighted Graphs
arXiv:1705.10097v1 [cs.DS] 29 May 2017
Aaron Bernstein∗
May 30, 2017
Abstract
In this paper we consider the decremental single-source shortest
paths (SSSP) problem, where given a graph G and a source node s the
goal is to maintain shortest distances between s and all other nodes
in G under a sequence of online adversarial edge deletions. In their
seminal work, Even and Shiloach [JACM 1981] presented an exact
solution to the problem in unweighted graphs with only O(mn) total
update time over all edge deletions. Their classic algorithm was the
state of the art for the decremental SSSP problem for three decades,
even when approximate shortest paths are allowed.
The first improvement over the Even-Shiloach algorithm was given
by Bernstein and Roditty [SODA 2011], who for the case of an unweighted and undirected graph presented a (1 + )-approximate algorithm with constant query time and a total update time of O(n2+o(1) ).
This work triggered a series of new results, culminating in a recent
breakthrough of Henzinger, Krinninger and Nanongkai [FOCS 14], who
presented a (1 + )-approximate algorithm for undirected weighted
graphs whose total update time is near linear: O(m1+o(1) log(W )),
where W is the ratio of the heaviest to the lightest edge weight in the
graph. In this paper they posed as a major open problem the question
of derandomizing their result.
Until very recently, all known improvements over the Even-Shiloach
algorithm were randomized and required the assumption of a nonadaptive adversary. In STOC 2016, Bernstein and Chechik showed the
first deterministic algorithm to go beyond O(mn) total update time:
the algorithm is also (1 + )-approximate, and has total update time
Õ(n2 ). In SODA 2017, the same authors presented an algorithm with
total update time Õ(mn3/4 ). However, both algorithms are restricted
to undirected, unweighted graphs. We present the first deterministic
algorithm for weighted undirected graphs to go beyond the O(mn)
bound. The total update time is Õ(n2 log(W )).
∗
Department of Mathematics, Technical University of Berlin. Supported in part by the
Einstein Grant at TU Berlin. [email protected]
1
1
Introduction
The objective of dynamic graph algorithms is to handle an online sequence of
update operations while maintaining a desirable functionality on the graph,
e.g., the ability to answer shortest path queries. An update operation may
involve a deletion or insertion of an edge or a node, or a change in an edge’s
weight. In case the algorithm can handle only deletions and weight increases
it is called decremental, if it can handle only insertions and weight decreases
it is called incremental, and if it can handle both it is called fully dynamic.
In this paper we consider the problem of (approximate) single source
shortest paths (SSSP) in unweighted undirected graphs, in the decremental
setting. Specifically, given an undirected graph G with positive weights and
a source node s, our algorithm needs to preform the following operations: 1)
Delete(e) – delete the edge e from the graph 2) Increase-Weight(e) – increase
the weight of e 3) Distance(v) – return the distance between s and v, i.e.,
dist(s, v), in the current graph G.
Fully dynamic shortest paths has a very clear motivation, as computing
shortest paths in a graph is one of the fundamental problems of graph
algorithms, and many shortest path applications must deal with a graph that
is changing over time. The incremental setting is somewhat more restricted,
but is applicable to any setting in which the network is only expanding. The
decremental setting is often very important from a theoretical perspective,
as decremental shortest paths (and decremental single source shortest paths
especially) are used as a building block in a very large variety of fully dynamic
shortest paths algorithms; see e.g. [16, 17, 3, 2, 20, 1] to name just a few.
Decremental shortest paths can also have applications to non-dynamic graph
problems; see e.g. Madry’s paper on efficiently computing multicommodity
flows [18].
We say that an algorithm has an approximation guarantee of α if its
output to the query Distance(v) is never smaller than the actual shortest
distance and is not more than α times the shortest distance. Dynamic
algorithms are typically judged by two parameters: the time it takes the
algorithm to adapt to an update (the Delete or Increase-Weight operation),
and the time to process a query (the Distance operation). Typically one
tries to keep the query time small (polylog or constant), while getting the
update time as low as possible. All the algorithms discussed in this paper
have constant query time, unless noted otherwise. In the decremental setting,
which is the focus of this paper, one usually considers the aggregate sum of
update times over the entire sequence of deletions, which is referred to as
the total update time.
Related Work: The most naive solution to dynamic SSSP is to simply
invoke a static SSSP algorithm after every deletion, which requires Õ(m) time,
using e.g. Dijkstra’s algorithm. (The Õ notation suppresses polylogarithmic
2
factors.) For unweighted graphs, since there can be a total of m deletions,
the total update time for the naive implementation is thus Ω(m2 ).
For fully dynamic SSSP, nothing better than the trivial O(m) time per
update is known. That is, we do not know how to do better than reconstructing from scratch after every edge update. For this reason, researchers have
turned to the decremental (or incremental) case in search of a better solution.
The first improvement stems all the way back to 1981, when Even and
Shiloach [8] showed how to achieve total update time O(mn) in unweighted
undirected graphs. A similar result was independently found by Dinitz [7].
This was later generalized to directed graphs by Henzinger and King [9].
This O(mn) total update time bound is still the state of art, and there are
conditional lower bounds [19, 15] showing that it is in fact optimal up to log
factors. (The reductions are to boolean matrix multiplication and the online
matrix-vector conjecture respectively).
These lower bounds motivated the study of the approximate version
of this problem. In 2011, Bernstein and Roditty [6] presented the first
algorithm to go beyond the O(mn) bound of Even and Shiloach [8]: they
presented a (1 + )-approximate decremental
SSSP algorithm for undirected
√
unweighted graphs with O(n2+O(1/ log n) ) = O(n2+o(1) ) total update time.
Henzinger, Krinninger and Nanongkai [13] later improved the total update
time to O(n1.8+o(1) + m1+o(1) ), and soon after the same authors [11] achieved
a close to optimal total update time of O(m1+o(1) log W ) in undirected
weighted graphs, where W is the largest weight in the graph (assuming
the minimum edge weight is 1). The same authors also showed that one
can go beyond O(mn) total update time bound in directed graphs (with a
(1 + ) approximation) [12, 14], although the state of art is still only a small
improvement: total update time O(mn0.9+o(1) log W ).
However, every single one of these improvements over the O(mn) bound
relies on randomization, and has to make the additional assumption of a
non-adaptive adversary. In particular, they all assume that the updates
of the adversary are completely independent from the shortest paths or
distances returned to the user, i.e. that the updates are fixed in advance.
This makes these algorithms unsuitable for many settings, and also prevents
us from using them as a black box data structure. For this reason, it is
highly desirable to have deterministic algorithms for the problem.
Very recently, in STOC 2016, Bernstein and Chechik presented the first
deterministic algorithm to go beyond O(mn) total update time, again with
a (1 + ) approximation. The total update time is Õ(n2 ) [4]. They followed
this up with a second deterministic algorithm [5] that has total update
√
time Õ(n1.5 m) = Õ(mn3/4 ). Both algorithms rely on the same basic
technique, so this is currently the only know technique for deterministically
breaking through the O(mn) barrier. However, both results above were
limited to undirected, unweighted graphs. In this paper, we show that this
core technique can also be applied to weighted graphs.
3
Our Results:
Theorem 1.1 Let G be an undirected graph with positive edge weights subject
to a sequence of edge deletions and weight increases, let s be a fixed source, and
let W be the ratio between the largest and smallest edge weights in the graph.
there exists a deterministic algorithm that maintains (1 + )-approximate
distances from s to every vertex in total update time O(m log2 (n) log(nW ) +
n2 log(n) log(nW )−2 ). The query time is O(1). (Like most decremental
shortest paths algorithms, our result can very easily be extended with the same
update time to the incremental case, where the update sequence contains edge
insertions and weight increases.)
(Technical Note: In weighted graphs the number of updates can be very
large – i.e. if each update just increases some edge weight by some small –
and so in addition to the total update time above, the algorithm necessarily
requires an additional O(1) per update. This O(1) factor is present in all
dynamic algorithms, and is typically omitted.)
Our algorithm does not match the randomized state of the art of
Õ(m1+o(1) log(W )) [11], but it is optimal up to log factors for dense graphs,
and is the first deterministic algorithm to go beyond the O(mn) barrier in
weighted graphs. The algorithm is also much simpler than the randomized
algorithm for weighted graphs , and it does not incur the extra no(1) factor
present in the randomized algorithms. Thus, our algorithm is in fact faster
than all existing randomized
√ algorithms for the problem in dense weighted
2−1/ log(n)
graphs where m = Ω(n
).
As a final remark, we note that all previous deterministic algorithms to
go beyond the O(mn) bound [4, 5] have the strange drawback that there is no
obvious way to return an approximate path, only a distance. Our algorithm
unfortunately shares this drawback. The reason is that whereas in other
dynamic shortest path algorithms the distance returned corresponds to some
path in the graph, our distance involves an additive error that is bounded
through a structural claim about the non-existence of certain paths, but does
not itself correspond to an actual path (see Lemma 4.4 in [4], Lemma 5.3 in
[5], and Lemma 3.4 in our paper.)
Preliminaries: In our model, an undirected weighted graph is subject to
deletions and weight increases. All weights in the original graph are positive.
Let G = (V, E) always refer to the current versions of the graph. Let m refer
to the number of edges in the original graph, and n to the number of vertices.
Let w(u, v) refer to the weight of edge (u, v). Let us assume for simplicity
that the minimum weighted in the original graph is at least 1, and note that
since edge weights only increase, this will be true of all version of the graph.
Let W be the largest edge weight to ever appear in the graph. For any
pair of vertices u, v, let π(u, v) be the shortest u − v path in G (ties can be
4
broken arbitrarily), and let dist(u, v) be the length of π(u, v). Let s be the
fixed source from which our algorithm must maintain approximate distances,
and let refer to our approximation parameter; when the adversary queries
the distance to a vertex v, the algorithm’s guarantee is to return a distance
dist0 (v) such that dist(s, v) ≤ dist0 (s, v) ≤ (1 + O())dist(s, v). Note that
the (1 + O()) approximation factor can always be reduced to (1 + ) without
affecting the asymptotic running time by simply starting with a suitably
smaller 0 .
Given any two sets S, T we define the set difference S \ T to contain all
elements s such that s ∈ S but s ∈
/ T . We will measure the update time of
the dynamic subroutines used by our algorithm in terms of their total update
time over the entire sequence of edge changes. Note that although edges in
the main graph G are only being deleted, there may be edge insertions into
the auxiliary graphs used by the algorithm.
2
High Level Overview
Our algorithm extends the techniques in the STOC 2016 paper of Bernstein
and Chechik [4] from unweighted to weighted graphs. We now briefly summarize their approach for unweighted graphs, aiming in this overview for
total update time Õ(n2.5 ) instead of Õ(n2 ). We start with the well known
fact that the classic Even and Shiloach algorithm has total update time
√
O(n2.5 ) [8] if we only care about distances less than n. So the hard case
is long distances in a dense graph. The key observation of [4] is that these
two problematic poles cannot coexist. In particular, Bernstein and Chechik
√
showed that any shortest path contains at most 3 n vertices of degree more
√
than n. The basic reason is that if we look at every third high-degree
vertex on the shortest path, then these vertices must have mutually disjoint
neighborhoods, since otherwise there would be a shorter path between them
of length 2; the claim then follows because each high-degree vertex has at
√
least n neighbors, and there are only n vertices in total. Thus, Bernstein
and Chechik show that the total contribution of high-degree vertices to the
shortest path is small. They then argue that this allows us to effectively
√
”ignore” all vertices of degree ≥ n, at the cost of an additive error of only
√
O( n). (we do not go into details of this ignoring here.) This completed
√
their result because for distances less than O( n−1 ) the classic Even and
Shiloach has total update time O(n2.5 −1 ), whereas for longer distances we
√
can afford the O( n) additive error from ignoring high degree vertices (it is
subsumed into the (1 + ) multiplicative error), so the resulting graph only
O(n1.5 ) edges.
This technique has no easy extension to weighted graphs via scaling. To
√
see this, say there were two edge weights, 1 and n, and that every vertex
√
√
had very few incident edges of weight 1, but n edges of weight n. Then
5
it is easy to see that distances in the graph could still be as large as n (as
√
opposed to n in the unweighted case). But we cannot scale weights down,
or use some sort of hop-distance technique found elsewhere in the literature,
because although the heavy edges are the ones that cause the graph to be
dense, the shortest path could still contain many edges of weight 1. There is
thus no way in a weighted graph to ignore high-degree vertices.
To overcome this, we switch focus to a notion of degree that considers
edge weights. In particular, note that in the example above, although the
graph was dense, the total number of edges of weight 1 was small. The first
step in our result is to formalize this observation, and show that although
we cannot simply ignore vertices of high degree, for any vertex v we can
ignore its low-weight incident edges if there are many of them. The proof
of this is similar to the proof that we can afford to ignore dense vertices in
unweighted graphs, but because we have to deal with many edge different
weights at once, the analysis is more sophisticated.
The problem is that on its own ignoring large neighborhoods of low-weight
edges doesn’t help: the result is a graph that is guaranteed to only have a
small number of low-weight edges (and a medium number of medium-weight
edges), but the graph can still be very dense with high-weight edges. To
overcome this, we develop a variation on the standard Even and Shiloach
algorithm [8] which is more efficient at dealing with heavy edges but incurs
a (1 + ) approximation. The basic intuition is as follows, though the details
are somewhat complex. The classic algorithm guarantees that we only touch
an edge (u, v) when the distance dist(s, u) or dist(s, v) increases. But if
(u, v) has high weight, and dist(s, u) only increases by 1, then the distance
increase is insignificant with respect to the total weight on (u, v), and so we
can afford to ignore it. The number of times we touch edge (u, v) thus ends
up being proportional to −1 /w(u, v).
All in all, our algorithm is based on the framework of Bernstein and
Chechik for unweighted graphs [4], but requires significant modifications both
to the high-level orientation (we need a new notion of sparsity that depends
on edge weights), as well as to the technical details. The basic approach
introduced in [4] of ignoring certain classes of dense vertices has proved quite
versatile and powerful (see the two very different applications of it in [4]
and [5]), and is currently the only known approach for going beyond O(mn)
deterministically, so our extension to weighted graphs might be useful in
further applications of this approach. Also, our extension of the classic Even
and Shiloach algorithm is presented in extreme generality, and might prove
useful in other dynamic shortest path problems where one has a graph that
has more edges of high weight than of low weight.
6
3
The Threshold Graph
In our algorithm, each vertex will have a weight cutoff, denoted cut-off(v),
and will end up ”ignoring” the edges below that cutoff. As described in the
high-level overview above, the required degree to ignore edges goes up as the
edge weights go up.
Definition 3.1 Let the level of edge (u, v) be level(u, v) = blog(w(u, v))c.
For any vertex v, let I(v) contain all edges incident to v (I for incident), and
let Ii (v) contain all edges incident to v of level i or less. Now, for any positive
threshold τ (not necessarily integral), and any vertex v, we define cut-offτ (v)
to be the maximum index i such that Ii (v) contains at least τ 2i edges; if
no such index exists, cut-offτ (v) is set to 0. We say that an edge (u, v) is
τ -heavy if level(u, v) ≤ cut-offτ (u) OR level(u, v) ≤ cut-offτ (v); we say
that it is τ -light otherwise. Let heavy(τ ) be the set of all τ -heavy edges in G,
and let light(τ ) be the set of all τ -light edges in G. Let G[heavy(τ )] be the
graph (V, heavy(τ )). Note that the levels, cut-offs, heaviness, and lightness
are always defined with respect to the main graph G, never with respect to
auxiliary graphs. Since τ is usually clear from context, we often just write
cut-off(v), heavy, and light.
Definition 3.2 Given a graph G = (V, E) with |V | = n and |E| = m, and
a positive threshold τ , define the threshold graph Gτ = (Vτ , Eτ ) as follows
• Vτ contains every vertex v ∈ V .
• Vτ also contains an additional vertex c for each connected component
C in the graph G[heavy(τ )] = (V, heavy(τ ))
• Eτ contains all edges in light(τ ), with their original weights.
• For every vertex v ∈ V , Eτ contains an edge from v to c of weight
1/2, where c is the component vertex in Vτ \ V that corresponds to the
component C in G[heavy(τ )] that contains v.
For any pair of vertices s, t ∈ V define πτ (s, t) to be the shortest path from s
to t in Gτ , and define distτ (s, t) to be the weight of this path.
Lemma 3.3 For any i, Gτ contains at most n edges of weight 1/2 that are
not in E (one per vertex in V ), as well at most O(n · τ · 2i ) edges at level i.
Proof: An edge (u, v) ∈ E of level i is only in Eτ if (u, v) is τ -light, i.e. if
cut-off(u) < i AND cut-off(v) < i. But a vertex with cut-off less than i
has by definition less than 2i τ incident edges of level i.
2
The following Lemma is the weighted analogue of Lemma 4.4 in [4], but
the most direct extension of that proof to the weighted case would analyze
each of the log(W ) edge levels separately, and thus multiply the error by
7
log(W ), eventually leading to a log3 (W ) factor in the total running time
of the algorithm. To avoid this extra error, we need a somewhat more
sophisticated analysis than in Lemma 4.4 of [4]
Lemma 3.4 For any graph G = (V, E), any positive integer threshold τ ,
and any pair of vertices s, t ∈ V :
distτ (s, t) ≤ dist(s, t) < distτ (s, t) +
14n
τ
.
Proof: It is not hard to see that we have distτ (s, t) ≤ dist(s,T
t). Simply
consider the shortest s−t path π(s, t) ∈ G. All the edges in π(s, t) light(τ )
are also in Gτ . Otherwise, for any τ -heavy edge (u, v) on π(s, t), recall that
the weight w(u, v) in G is always at least 1, and note that because edge (u, v)
is τ -heavy, the vertices u and v must be in the same connected component
in G[heavy(τ )], and so there is a path of length 1 from u to v in Gτ : the
edge of weight 1/2 from u to the component vertex c, and then a second
edge from c to v.
We now prove that dist(s, t) ≤ distτ (s, t) + 14n
τ . Let πτ (s, t) be the
∗
shortest s − t path in Gτ . Consider the edge set E ⊆ E which contains all
the τ -light edges of πτ (s, t) (theseSedges are also
T in E), and all τ -heavy edges
in E. That is, E ∗ = heavy(τ ) (light(τ ) πτ (s, t)). Let G∗ = (V, E ∗ ).
We first observe that there exists an s − t path in G∗ . We construct this path
by following πτ (s, t). πτ (s, t) contains τ -light edges, which are by definition
in G∗ , as well as subpaths of length 2 of the form (v, c) ◦ (c, w), where v and
w are in the same connected component C in G[heavy(τ )]. But since v and
w are in the same connected component, there is a path of only heavy edges
connecting them, and all heavy edges are in G∗ .
Now, let π ∗ (s, t) be the shortest s − t path in G∗ . Let dist∗ (s, t) be the
length of π ∗ (s, t). Since G∗ is a subgraph of G, we know that dist(s, t) ≤
dist∗ (s, t). We now show that
dist∗ (s, t) < distτ (s, t) +
14n
τ
(1)
which will complete the proof of the lemma.
We prove Equation 1 by showing that all the heavy edges on π ∗ (s, t) have
∗
total weight at most 14n
τ : since all the light edges on π (s, t) are by definition
also on πτ (s, t), this proves the equation.
Let VH∗ be the set of vertices on π ∗ (s, t) that are incident to at least
one heavy edge on π ∗ (s, t). Now, for any vertex v ∈ VH∗ , let ball(v)
contain v and all vertices u ∈ V , such that there is an edge (u, v) with
level(u, v) ≤ cut-off(v). Note that by definition of cut-off(v), we have
|ball(v)| ≥ 2cut-off(v) τ .
8
Now, let P be an ordering of the vertices in π ∗ (s, t) in non-increasing
order of cut-off: so if i < j then cut-off(P [i]) ≥ cut-off(P [j]). We will now
construct a set IV of independent vertices in VH∗ as follows. We start with
IV empty. Now, go through the vertices in VH∗ according to their order in
P (so in non-increasing order of cut-off), and for each v do the following: if
ball(v) is disjoint from ball(u) for every u ∈ IV, add v to IV. Clearly if
v and w are independent then ball(v) and ball(w) are disjoint. But then
recalling that for any v, |ball(v)| ≥ 2cut-off(v) τ , and that the total number
of vertices in the graph is n, we have
X
2cut-off(v) ≤ n/τ
(2)
v∈IV
We will end up showing that each independent vertex v is ”responsible”
for at most 14 · 2cut-off(v) weight from heavy edges on π ∗ (s, t), which will
complete our proof. We say that a non-independent vertex u belongs to
independent vertex v if cut-off(v) ≥ cut-off(u) and ball(u) and ball(v)
are non-disjoint. By our independence construction, every non-independent
vertex u belongs to one or more independent vertices.
Claim 3.5 If v is independent, and u belongs to v, then the distance from v
to u along π ∗ (s, t) is at most 4 · 2cut-off(v) .
Proof: Say, for contradiction, that this was not the case.
T Then, since
u belongs to v, there must be some vertex z ∈ ball(v) ball(u). But
by definition of ball we have level(v, z) ≤ cut-off(v) and level(u, z) ≤
cut-off(u) ≤ cut-off(v), so by definition of level there is a path from u to v
through z of weight at most 4 · 2cut-off(v) , which contradicts π ∗ (s, t) being
the shortest path in G∗ .
2
Now, given any τ -heavy edge (x, y), let the dominant endpoint of (x, y)
be the endpoint that comes earlier in P : so in particular, if x is dominant
then cut-off(x) ≥ cut-off(y). We say that a τ -heavy edge (x, y) belongs to
independent vertex v if the dominant endpoint of (x, y) belongs to v. Note
that every τ -heavy edge in π ∗ (s, t) belongs to at least one independent vertex.
We now show that if v is independent, then the total weight of τ -heavy edges
that belong to v is at most 14 · 2cut-off(v) . Otherwise (for contradiction),
at least half that weight would come before or after v, so w.l.o.g, let us say
that there is at least 7 · 2cut-off(v) weight of edges that belong to v and
come after v on π ∗ (s, t) – i.e., are closer to t than v is. Let us consider the
τ -heavy edge (x, y) that belongs to v and is closest to t, and say that y is
closer to t than x. Note that the distance from v to y along π ∗ (s, t) is at
least 7 · 2cut-off(v) , so by Claim 3.5, y cannot belong to v. Thus, since edge
(x, y) belongs to v, x must be the dominant endpoint and belong to v. But if
x belongs to v and dominates y then cut-off(y) ≤ cut-off(x) ≤ cut-off(v),
and the only way edge (x, y) could be τ -heavy is if level(x, y) ≤ cut-off(v),
9
so w(x, y) ≤ 2 · 2cut-off(v) , so the distance from v to x along π ∗ (s, t) is
still at least 7 · 2cut-off(v) − 2 · 2cut-off(v) = 5 · 2cut-off(v) , which again
contradicts Claim 3.5. Thus, the total weight of heavy edges on π ∗ (s, t)
P
is at most 14 v∈IV 2cut-off(v) , which combined with Equation 2 proves
Equation 1, which proves the lemma.
2
Maintaining the threshold graph Gτ as G changes is easy, because the
only hard part is maintaining the connected components C in G[heavy(τ )],
and dynamic connectivity in undirected graphs is a well-studied problem
that admits deterministic O(log2 (n)) update time. The details are essentially
identical to those for unweighted graphs (see Lemmas 4.5 and 4.6 in [4]), so
we omit the proofs of the following two lemmas.
Lemma 3.6 Given a graph G subject to a decremental update sequence,
and a positive integer threshold τ , we can maintain the graph Gτ in total
time O(m log2 (n)). Moreover, the total number of edges of weight 1/2 ever
inserted into the graph is O(n log(n)), and the total number of edges (u, v)
that have level(u, v) = i when inserted is at most nτ 2i .
Lemma 3.7 Given a graph G subject to a decremental update sequence, a
positive integer threshold τ , and a pair of vertices u, v in G, the distance
distτ (u, v) in Gτ never decreases as the graph G changes.
4
An Extension of the Even and Shiloach Algorithm
In the paper of Bernstein and Chechik for unweighted graphs [4], the threshold
graph ended up being sparse, and so it was easy to efficiently compute
distances within it using the standard Even and Shiloach algorithm with
total update time O(mn) [8]. In our paper, however, the threshold graph
is only sparse with respect to low-weight edges; that is, because heaviness
and lightness is defined in terms of edge weights, the threshold graph can
have many edge of high weight. We now present a modification of the Even
and Shiloach algorithm that can deal more efficiently with edges of high
weight, which makes it perfectly suited to our threshold graph, which has
mostly high-weight edges. Like the standard Even and Shiloach algorithm,
our extension runs up to a certain depth bound d.
Definition 4.1 Given any number d, the function boundd (x) is equal to x
if x ≤ d, and to ∞ otherwise.
Note that to apply to the threshold-graph Gτ , the algorithm must be able
to handle insertions, as long as they do not change distances. (See Lemmas
3.6, 3.7). To this end we need the following definition:
10
Definition 4.2 Let G = (V, E) be a graph with positive weights ≥ 1, subject
a dynamic update sequence of insertions, deletions, and weight-increases. In
such a context, let A contain the set of ALL edges (u, v) to appear in G at
any point during the update sequence (A for all): if an edge (u, v) is deleted
and then inserted again, we consider this as two separate edges in A. For
every edge (u, v) ∈ A, let wo (u, v) (o for original) be the weight of (u, v)
when it first appears in A (this might be in the original graph itself ), and let
levelo (u, v) = blog(wo (u, v))c.
Lemma 4.3 Let G = (V, E) be an undirected graph with positive integer
edge weights, s a fixed source, and d ≥ 1 a depth bound. Say that G is subject
to a dynamic sequence of edge insertions, deletions, and weight increases,
with the property that distances in G never decrease as a result of an update.
Then, there exists an algorithm WSES(G, s, d) (stands for weight-sensitive
Even and Shiloach) that maintains approximate distances dist0 (s, v) with
0
dist(s, v) ≤ dist
P (s, v) ≤dboundd (dist(s, v))(1 + ), and has total update
time O(nd + (u,v)∈A wo (u,v) ).
Proof: Throughout the proof of this lemma, we will often rely on the
following function:
Definition 4.4 Given any positive real numbers β and x, let roundβ (x) be
the smallest number y > x that is an integer multiple of β.
Our algorithm WSES follows the basic procedure of the classic Even
and Shiloach algorithm [8] (denoted ES), with a few modifications. For this
reason we can describe many of the guarantees of ES without looking under
the hood for exactly how they are implemented, since we keep exactly the
same implementation. We only modify the classic algorithm in a few key
points. Recall the definition of A and wo (u, v) from Definition 4.2. Since
weights only increase, we always have w(u, v) ≥ wo (u, v).
The ES algorithm maintains for each vertex v a distance label δ(v) with
the property that dist(s, v) = δ(v). In particular, it maintains the invariant
that δ(s) = 0, and that if (u, v) ∈ E, then δ(v) ≤ δ(u) + w(u, v). Our
algorithm will also maintain a label δ(v) for every v, but since we only need
an approximation, we guarantee a weaker invariant:
Invariant 4.5 (Approximation) We always have δ(s) = 0. If p is the
parent of v in the tree then δ(v) ≥ δ(p) + w(p, v). Also, for all (u, v) ∈ E,
δ(v) ≤ δ(u) + w(u, v) + wo (u, v).
Approximation Analysis: We now show that as long as the Approximation
Invariant is preserved, we always have dist(s, v) ≤ δ(v) ≤ (1 + )dist(s, v).
The fact that δ(v) ≥ dist(s, v) follows trivially for the same reason as in
classic ES. To prove the other side, consider the shortest path π(s, v) in G.
Let the vertices in π(s, v) be s0 , s1 , ..., sq = v. We show by induction that
11
δ(si ) ≤ (1+)dist(s, si ). The claim is trivially true for i = 0 because δ(s) = 0
at all times. Now, say the claim is true for si , and note that dist(s, si+1 ) =
dist(s, si ) + w(si , si+1 ). Now, by the Approximation Invariant, we have
δ(si+1 ) ≤ δ(si )+w(si , si+1 )+wo (si , si+1 ) ≤ δ(si )+(1+)w(si , si+1 ) (the last
step follows from w(u, v) ≥ wo (u, v)). But then by the induction hypothesis
we have δ(si+1 ) ≤ (1 + )(dist(s, si ) + w(si , si+1 )) = (1 + )dist(s, si+1 ).
We must now show how to efficiently maintain the Approximation Invariant. The classic ES algorithm does O(1) time per update for basic setup,
and otherwise all the work comes from charging constant time operations
to various edges (u, v) and vertices v. The update time invariant in classic
ES is as follows: we only charge vertex v when δ(v) increases, and we only
charge edge (u, v) when δ(u) or δ(v) increases. Since weights are positive
integers, every δ(u) must increase by at least 1, and since we only go up
to distance d, each δ(v) increases at most d times, which yields O(md). To
improve upon this, our algorithm maintains a slightly different update time
invariant
Invariant 4.6 (Update-Time) A vertex v is only charged when δ(v) increases. An edge (u, v) is only charged when either roundwo (u,v) (δ(u)) or
roundwo (u,v) (δ(v)) increases.
Since for any β, roundβ (δ(v)) can increase at most d/β before it exceeds
d, it is clear that the Update-Time Invariant guarantees the total update
bound of the lemma.
Maintaining the Two Invariants: Now, let us recall how the standard
ES algorithm works. This algorithm only handles deletions, so that is what
we focus on first: we later show how to extend our algorithm to handle weightincreases and insertions that do not change distances. We will emphasize the
parts of classic ES that we will later change in our modified algorithm. First
off, each vertex v always maintains label information about its neighbors:
so in particular, for each edge (u, v), vertex v stores a copy δv (u) = δ(u)
(first-difference). The algorithm maintains a shortest path tree T from
the root s. We say an edge (u, v) is a tree edge if (u, v) ∈ T , and a non-tree
edge if (u, v) ∈ E \ T . Whenever a non-tree edge is deleted, shortest distances
do not change, so the algorithm does not have to do much. Now, consider the
deletion of a tree edge (u, v) where u is the parent of v in T . The algorithm
checks in constant time, by cleverly storing local label information δv (z), if v
has another edge to a vertex z with δ(v) = δv (z) + w(z, v). If yes, v can be
reconnected without a label increase, the edge (z, v) is added to the tree and
all distance labels remain the same so we are done. If not, label δ(v) must
increase, which in turn affects of the children of v, so The algorithm then
examines all the children of v and checks if they can be attached to the tree
without increasing their distance in a similar manner. (second-difference)
Some vertices will fail to be reattached with the same label, so the algorithm
12
keeps a list of all vertices that need to be re-attached at a higher label. Each
time the algorithm discovers that a label δ(v) must increase, it increases it
by 1 (which is the most optimistic new label it can get) and returns it to the
needs-reattachment list in at attempt to reattach v with this higher label.
The algorithm examines the vertices in the list in an increasing order of their
label. Note that the label of a vertex may be increased many times as a result
of an update. Whenever the label δ(v) of a vertex v changes, the algorithm
must adjust δu (v) for every edge (u, v) in the graph. (third-difference.)
Note that the algorithm above fails to satisfy the Update-Time Invariant,
because although a vertex v is only charged when δ(v) increases (otherwise
the search stops at v and can be charged to the edge that led to v), edge
(u, v) is charged whenever δ(u) or δ(v) increase even just by 1. In particular,
the invariant is not preserved because of the work in second-difference
and third-difference. The changes we implement are simple. We start
with first-difference: each vertex v will still store local label information
δv (u), but now it might be slightly inaccurate. In particular, we always have
Invariant 4.7 (Local-Information) δ(u) ≤ δv (u) ≤ roundwo (u,v) (δ(u)).
A vertex v will now choose its parent based on the slightly inaccurate
local labels δu (v) instead of the true label δ(u). Let p(v) be the parent of v
in the tree.
Invariant 4.8 (Parent-Choice)
p(v) = argmaxu∈neighbors(v) {δv (u) + w(u, v)} and δ(v) = δv (p(v)) +
w(p(v), v).
The Parent-Choice and Local-Information invariants combined clearly
guarantee the Approximation Invariant. We must now show how to maintain
these efficiently. This leads us to third-difference: when δ(v) changes,
we only update δu (v) for every vertex u when roundwo (u,v) (δ(v)) increases.
This ensures that we satisfy the Update-Time Invariant for this step of
the algorithm, while still ensuring that all local information adheres to the
Local-Information Invariant.
Note that to implement third-difference in the above paragraph
while maintaining the Update-Time invariant, we need a data structure for
every vertex v, such that whenever δ(v) increases, the data structure returns
all vertices u for which roundwo (u,v) (δ(v)) increased. This is easy to do
because wo (u, v) is completely fixed for each edge, i.e. it does not change
as the graph changes. Each vertex v will have d buckets, where bucket i
contains all edges (u, v) such that roundwo (u,v) (i − 1) < roundwo (u,v) (i).
This initialization requires O(nd) time to create d buckets for vertex, and
then each edge (u, v) is placed
v) buckets for vertices u and v,
P in d/wo (u,
d
leading to total time O(nd + (u,v)∈A wo (u,v)
), as desired. Once the buckets
13
are initialized, whenever δ(v) increases from i to j, δ(v) simply notifies all
edges in buckets i + 1, i + 2, ..., j
The last violation of the Update-Time Invariant was in second-difference.
When δ(v) changes, instead of processing all children u, we only need to process those for which δu (v) changes. But in our modified algorithm this only
occurs when roundwo (u,v) (δ(v)) increases, in keeping with the Update-Time
Invariant.
We have thus shown how to modify classical ES in a way that maintains
the Approximation Invariant and the Update-Time Invariant while processing
deletions. Increase-Weight(u, v) is processed in the same way: if (u, v) was
not a tree edge we do nothing, and otherwise if u was the parent of v, then
some invariants will be violated for v, and we fix these in exactly the same
way as for a deletion.
We now turn to processing insertions. In classical ES, it is trivial to
process an insertion of (u, v) that does decrease distances because the distance
labels do not change, so we just spend O(1) time fixing update δv (u) and
δu (v). In our algorithm, however, a strange difficulty arises: because our
labels are approximate, even if the insertion of (u, v) does not decrease
distances, it might nonetheless violate the Approximation Invariant: say that
dist(u) = dist(v) = 1000 but δ(v) = 1000(1 + ) while δ(u) = 1000 and we
insert and edge of weight 1. In this case we cannot afford to decrease δ(v)
because our update time analysis relied on non-decreasing labels. But note
that although v now violates the Approximation Invariant, δ(v) is still a
(1 + ) approximation to dist(s, v), because dist(s, v) never decreases. We
thus rely on the idea used in the ”Monotone” ES tree of Henzinger et al.
[10]: we simply never decrease distance labels. More formally, our distance
labels will satisfy the following relaxed Approximation Invariant: after every
update, for every vertex v, either δ(v) satisfies the standard Approximation
Invariant, or δ(v) = δ old (v), where δ old (v) was the distance label before
the update. By an induction on time, it is easy to see that with this relaxed
Approximation Invariant, we still always have that δ(v) ≤ (1 + )dist(s, v).
If δ(v) satisfies the approximation invariant, the proof is the same as in
the Approximation Analysis above. Otherwise, we have δ(v) = δ old (v), in
which case by our time induction we have δ old (v) ≤ (1 + )distold (s, v),
and since our updates are guaranteed not to decrease distances, we have
distold (s, v) ≤ dist(s, v) and we are done.
2
5
The decremental SSSP algorithm
We now put together all our ingredients to proving the main Theorem 1.1.
Lemma 5.1 For any threshold τ , and depth bound d = n−1 τ −1 , we can
maintain approximate distances dist0τ (s, v) in total time O(m log2 (n) +
14
n2 log(n)−2 ), with the property that dist0τ (s, v) ≤ (1+)boundd (distτ (s, v))+
d
Proof: First we use Lemma 3.6 to maintain the threshold graph Gτ in
d
O(m log2 (n)) total update time. Then, setting β = 2n
, for each edge
(u, v) ∈ Gτ , we round the edge weight w(u, v) up to the nearest multiple of β.
It is easy to see that since every shortest path in the graph contains at most n
edges, this weight change incurs an additive error of at most nβ = d/2. Since
all edge weights are now multiple of β, we can divide all edge weights by factor
of β without changing shortest paths, which results in a scaled graph Gsτ
with integral weights for which we have to maintain shortest distances up to
depth ds = d/β = 2n−1 . We now run WSES(Gsτ , s, ds ) in the scaled graph,
and then scale the resulting distances back up by β. Since WSES is a (1 + )
approximation, we get: dist0τ (s, v) ≤ (1 + )boundd (distτ (s, v) + d/2) ≤
(1 + )boundd distτ (s, v) + d.
P
s
Recall that the total update time of WSES(Gsτ , s, ds ) is O(nds + (u,v)∈A wod(u,v) ).
We know that O(nds ) = O(n2 −1 ). Now, Gsτ contains two types of edges:
edges between a vertex v ∈ V and a component vertex c ∈ Vτ \ C, and
τ -light edges from the original graph. By Lemma 3.6, the total number of
component edges ever inserted is O(n log(n)), and because of our scaling
they all have weight at least 1, so their total contribution to the sum is at
most n log(n)−1 · ds = O(n2 log(n)−2 ). For edges of the original graph,
let us look at the contribution of all edges (u, v) with levelo (u, v) = i. By
Lemmas 3.3 and 3.6, the total number of such edges is at most O(nτ 2i ).
Each such edge has wo (u, v) ≥ 2i , which has been scaled down by β in Gsτ ,
s
and so contributes d2βi = 2di to the sum. Thus in total we have that each
level i contributes a total of O(nτ d−1 ) = O(n2 −2 ).
We complete the proof by arguing that there are only log(n) contributing
levels. For simplicity, let us look at the graph Gτ before scaling. First off,
edges of weight more than d = n−1 /τ can be ignored, since we do not care
about distances greater than d (we are working with boundd ). Secondly,
from the definition of cut-off(v), an edge (u, v) of weight less than 1/(2τ ) will
always be heavy and so never appear in Gτ : if 2τ w(u, v) ≤ 2level(u,v) τ < 1,
then cut-off(u) and cut-off(v) are always at least as large as level(u, v), so
(u, v) will remain heavy. Thus, the only edges that appear in our graph have
weight between 1/(2τ ) and n/(τ ), which corresponds to log(n/) = O(log(n))
different levels of edge weights.
2
Lemma 5.2 For any distance bound d, with total update time O(m log2 (n) +
n2 log(n)−2 ) we can maintain approximate distances distd (s, v) to all vertices v such that
dist(s, v) ≤ distd (s, v) ≤ boundd (dist(s, v)) + 15d
Proof: Set τ = n/(d). By Lemma 5.1, in total update time O(m log2 (n) +
n2 log(n)−2 ) we can maintain values dist0τ (s, v) with distτ (s, v) ≤ dist0τ (s, v) ≤
15
(1 + )boundd (distτ (s, v)) + d. Now, By Lemma 3.4 we have dist(s, v) −
14n/τ ≤ distτ (s, v) ≤ dist(s, v). Plugging in our value τ = n/(d) we
get dist(s, v) − 14d ≤ distτ (s, v) ≤ dist(s, v). Thus, we have dist(s, v) −
14d ≤ dist0τ (s, v) ≤ boundd (dist(s, v)) + d. If we return distd (s, v) =
dist0τ (s, v) + 14d, we get the bound in the lemma.
2
It is now easy to prove our main Theorem 1.1. Observe that the maximum
distance we could possibly see is nW . Thus, for each i = 0, 1, 2, 3...dlog(nW )e,
we use Lemma 5.2 to maintain δi (v) = dist2i (s, v). By Lemma 5.2, the total
update time is the desired O(m log2 (n) log(nW ) + n2 log(n) log(nW )−2 ).
We then output as our final answer: dist0 (v) = mini {δi (v)}. Since for each
i we have δi ≥ dist(v), we have dist0 (v) ≥ dist(s, v). We now need to
show that dist0 (v) ≤ (1 + O())dist(s, v). It is not hard to see that for
i = dlog(dist(s, v))e, we end up with δi (v) ≤ dist(s, v) + 30dist(s, v), as
desired.
6
Conclusions
In this paper we presented the first deterministic decremental SSSP algorithm
for weighted undirected graphs that goes beyond the Even-Shiloach bound
of O(mn) total update time. Previously such a result was only known for
unweighted undirected graphs. The two main open questions are further
improving this total update time, and going beyond O(mn) deterministically
for directed graphs. Finally, we recall from the introduction that all existing
deterministic o(mn) results including ours are only able to return approximate
shortest distances, not the paths themselves.
16
References
[1] Ittai Abraham, Shiri Chechik, and Kunal Talwar. Fully dynamic allpairs shortest paths: Breaking the o(n) barrier. In Approximation,
Randomization, and Combinatorial Optimization. Algorithms and Techniques, APPROX/RANDOM 2014, September 4-6, 2014, Barcelona,
Spain, pages 1–16, 2014.
[2] Surender Baswana, Ramesh Hariharan, and Sandeep Sen. Maintaining
all-pairs approximate shortest paths under deletion of edges. In Proceedings of the Fourteenth Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA, pages 394–403. Society for Industrial and Applied
Mathematics, 2003.
[3] Aaron Bernstein. Fully dynamic (2 + epsilon) approximate all-pairs
shortest paths with fast query and close to linear update time. In
Proceedings of the 50th Annual IEEE Symposium on Foundations of
Computer Science, FOCS, pages 693–702, 2009.
[4] Aaron Bernstein and Shiri Chechik. Deterministic decremental single
source shortest paths: beyond the o(mn) bound. In Proceedings of the
48th Annual ACM Symposium on Theory of Computing (STOC), pages
389–397, 2016.
[5] Aaron Bernstein and Shiri Chechik. Deterministic partially dynamic
single source shortest paths for sparse graphs. In Proceedings of the
Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms,
SODA 2017, Barcelona, Spain, Hotel Porta Fira, January 16-19, pages
453–469, 2017.
[6] Aaron Bernstein and Liam Roditty. Improved dynamic algorithms for
maintaining approximate shortest paths under deletions. In Proceedings of the Twenty-Second Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA, pages 1355–1365, 2011.
[7] Yefim Dinitz. Dinitz’ algorithm: The original version and even’s version.
In Theoretical Computer Science, Essays in Memory of Shimon Even,
pages 218–240, 2006.
[8] Shimon Even and Yossi Shiloach. An on-line edge-deletion problem.
Journal of the ACM, 28(1):1–4, 1981.
[9] Monika Henzinger and Valerie King. Fully dynamic biconnectivity
and transitive closure. In 36th Annual Symposium on Foundations of
Computer Science, Milwaukee, Wisconsin, 23-25 October 1995, pages
664–672, 1995.
17
[10] Monika Henzinger, Sebastian Krinninger, and Danupon Nanongkai. Dynamic approximate all-pairs shortest paths: Breaking the o(mn) barrier
and derandomization. In Proceedings of the 54th Annual Symposium on
Foundations of Computer Science, FOCS, pages 538–547, 2013.
[11] Monika Henzinger, Sebastian Krinninger, and Danupon Nanongkai.
Decremental single-source shortest paths on undirected graphs in nearlinear total update time. In Proceedings of the 55th Annual Symposium
on Foundations of Computer Science, FOCS, pages 146–155, 2014.
[12] Monika Henzinger, Sebastian Krinninger, and Danupon Nanongkai.
Sublinear-time decremental algorithms for single-source reachability and
shortest paths on directed graphs. In Proceedings of the 46th Annual
ACM Symposium on Theory of Computing (STOC), pages 674–683,
2014.
[13] Monika Henzinger, Sebastian Krinninger, and Danupon Nanongkai. A
subquadratic-time algorithm for decremental single-source shortest paths.
In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA, pages 1053–1072, 2014.
[14] Monika Henzinger, Sebastian Krinninger, and Danupon Nanongkai. Improved algorithms for decremental single-source reachability on directed
graphs. In Proceedings of the 42nd International Colloquium, ICALP,
pages 725–736, 2015.
[15] Monika Henzinger, Sebastian Krinninger, Danupon Nanongkai, and
Thatchaphol Saranurak. Unifying and strengthening hardness for dynamic problems via the online matrix-vector multiplication conjecture.
In Proceedings of the Forty-Seventh Annual ACM on Symposium on
Theory of Computing (STOC), pages 21–30, 2015.
[16] Monika Rauch Henzinger and Valerie King. Maintaining minimum
spanning forests in dynamic graphs. SIAM J. Comput., 31(2):364–374,
2001.
[17] Valerie King. Fully dynamic algorithms for maintaining all-pairs shortest
paths and transitive closure in digraphs. In Proceedings of the 40th
Annual Symposium on Foundations of Computer Science, FOCS, pages
81–91, 1999.
[18] Aleksander Madry. Faster approximation schemes for fractional multicommodity flow problems via dynamic graph algorithms. In Proceedings
of the 42nd ACM Symposium on Theory of Computing, STOC 2010,
Cambridge, Massachusetts, USA, 5-8 June 2010, pages 121–130, 2010.
[19] Liam Roditty and Uri Zwick. On dynamic shortest paths problems.
Algorithmica, 61(2):389–401, 2011.
18
[20] Liam Roditty and Uri Zwick. Dynamic approximate all-pairs shortest
paths in undirected graphs. SIAM J. Comput., 41(3):670–683, 2012.
19
| 8cs.DS
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:1006.3215v1 [cs.AI] 16 Jun 2010
Solving Functional Constraints by Variable
Substitution
YUANLIN ZHANG
Department of Computer Science
Texas Tech University,
Lubbock, TX79409-3104, USA
(e-mail: [email protected])
ROLAND H.C. YAP
School of Computing
National University of Singapore
13 Computing Drive, 117417, Singapore
(e-mail: [email protected])
submitted 22nd June 2009; revised 20th October 2009; accepted 8th December 2009
Abstract
Functional constraints and bi-functional constraints are an important constraint class in
Constraint Programming (CP) systems, in particular for Constraint Logic Programming
(CLP) systems. CP systems with finite domain constraints usually employ CSP-based
solvers which use local consistency, for example, arc consistency. We introduce a new approach which is based instead on variable substitution. We obtain efficient algorithms for
reducing systems involving functional and bi-functional constraints together with other
non-functional constraints. It also solves globally any CSP where there exists a variable
such that any other variable is reachable from it through a sequence of functional constraints. Our experiments on random problems show that variable elimination can significantly improve the efficiency of solving problems with functional constraints.
To appear in Theory and Practice of Logic Programming (TPLP).
KEYWORDS: constraint logic programming, constraint satisfaction problem, functional
constraints, variable substitution, arc consistency
1 Introduction
Functional constraints are a common class of constraints occurring in Constraint
Satisfaction Problem(s) (CSP) (Stallman and Sussman 1977; Van Hentenryck et al. 1992;
Kirousis 1993). Roughly speaking, a constraint c(x, y) is functional if the value of
variable y is some function of the value of variable x (see Definition 1). Functional constraints arise in two ways, they may occur quite naturally since one may
have “functions” or “equations” in the constraint model. Functional constraints
also occur systematically in in Constraint Programming (CP) when the system is
a Constraint Logic Programming (CLP) system.
2
Y. Zhang and R. H. C. Yap
Functional constraints arise naturally in CLP in two ways. Firstly, the equations which arise from matching the head of a rule with an atom in the body are
functional constraints. Secondly, the basic (or primitive) constraints in a particular instance of a CLP language will often include functional constraints. Consider,
one of the most widely used and successful constraint domains for CLP, namely,
finite domains which we will call CLP(FD). The basic constraints in a CLP(FD)
system, for example, CHIP (Van Hentenryck et al. 1992), can express functional
constraints. An example would be the finite domain constraint, 3X + 2Y = 10,
with finite domain variables for X and Y .1 Matching the head and body, gives rise
to a number of equations, and in a FD system, the equations are functional constraints. For example, when matching p(Z 2 + 1) with a rule on p(X) where both X
and Z are finite domain variables, a functional constraint X = Z 2 + 1 is produced.2
We remark that in logic programming, the equations are solved by unification but
in the general setting, constraint solving over the particular domain is required.
Recognizing and exploiting functional constraints can facilitate the development of
more efficient constraint solvers for CLP systems.
Most work on solving functional constraints follows the approach in CSP which
is based on arc or path consistency (Van Hentenryck et al. 1992; David 1995). We
remark that, in many papers, “functional constraints” are actually what we call bifunctional constraints (Definition 2), a special case of functional constraints. In this
paper, we propose a new method — variable substitution — to process functional
constraints. The idea is that if a constraint is functional on a variable, this variable in
another constraint can be substituted away using the functional constraint without
losing any solution.3
Given a variable, the variable elimination method substitutes this variable in all
constraints involving it such that it is effectively “eliminated” from the problem.
This idea is applied to reduce any problem containing non-functional constraints
into a canonical form where some variables can be safely ignored when solving
the problem. We design an efficient algorithm to reduce, in O(ed2 ) where e is the
number of constraints and d the size of the largest domain of the variables, a general
binary CSP containing functional constraints into a canonical form. This reduction
simplifies the problem and makes the functional portion trivially solvable. When
the functional constraints are also bi-functional, then the algorithm is linear in the
size of the CSP.
Many CLP systems with finite domains make use of constraint propagation algorithms such as arc consistency. Unlike arc consistency, our elimination method
completely solves the functional portion of the problem, hence the functional constraints are eliminated and their consequences are incorporated into the reduced
problem. Our experiments show that the substitution based “global” treatment of
functional constraints can significantly speed up propagation based solvers.
1
2
3
Note that the examples which involve CLP use uppercase for variables as per the logic programming convention. In a more general context, we will use lowercase variables like x and y
for variables.
Notice that this is not a “linear” constraint in the arithmetic sense.
A preliminary version of this paper appeared in (Zhang et al. 2008).
Solving Functional Constraints by Variable Substitution
3
In the rest of the paper, background on CSPs and functional constraints is given
in Section 2. Variable substitution for binary functional constraints is introduced
and studied in Section 3. Section 4 presents several results on algorithms for variable
elimination in general CSPs containing functional constraints. Section 5 presents
an experimental study on the effectiveness of the variable elimination algorithm
and explains why functional elimination leads to a smaller problem with a reduced
search space. In Section 6, we extend binary functional constraints to non-binary
functional constraints and we discuss substitution for general problems where there
are non-binary constraints which may be in extensional and intensional form. Related work is discussed in Section 7, and the paper is concluded in Section 8.
2 Preliminaries
We begin with the basic concepts and notation used in this paper.
A binary Constraint Satisfaction Problem (CSP) (N, D, C) consists of a finite
set of variables N = {v1 , · · · , vn }, a set of domains D = {D1 , · · · , Dn }, where Di is
the domain of variable i , and a set of constraints each of which is a binary relation
between two variables in N .
A constraint between two variables i and j is denoted by cij . Symbols a and b
possibly with subscript denote the values in a domain. A constraint cij is a set of
allowed tuples. We assume testing whether a tuple belongs to a constraint takes
constant time. For a ∈ Di and b ∈ Dj , we use either (a, b) ∈ cij or cij (a, b) to
denote that values a and b satisfy the constraint cij . For the problems of interest
here, we require that for all a ∈ Di and b ∈ Dj , (a, b) ∈ cij if and only if (b, a) ∈ cji .
If there is no constraint on i and j, cij denotes a universal relation, i.e., Di × Dj .
A constraint graph G = (V, E) where V = N and E = {{i, j} | ∃cij ∈ C}. The
constraint graph is usually used to describe the topological structure of a CSP. A
solution of a constraint satisfaction problem is an assignment of a value to each
variable such that the assignment satisfies all the constraints in the problem. A
CSP is satisfiable if it has a solution. The solution space of a CSP is the set of all
its solutions. Two CSPs are equivalent if and only if they have the same solution
space. Throughout this paper, n represents the number of variables, d the size of
the largest domain of the variables, and e the number of constraints in C.
We need two operations on constraints in this paper. One is the intersection of
two constraints (intersection of the sets of tuples) that constrain the same set of
variables. The other operation is the composition, denoted by the symbol “◦” of
two constraints sharing a variable. The composition of two relations is:
cjk ◦ cij = {(a, c) | ∃b ∈ Dj , such that (a, b) ∈ cij ∧ (b, c) ∈ cjk }.
Composition is a basic operation in our variable substitution method. Composing
cij and cjk leads to a new constraint on variables i and k.
Example 1
Consider constraints cij = {(a1 , b1 ), (a2 , b2 ), (a2 , b3 )} and cjk = {(b1 , c1 ), (b2 , c2 ),
(b3 , c2 )}. The composition of cij and cjk is a constraint on i and k: cik = {(a1 , c1 ),
(a2 , c2 )}.
4
Y. Zhang and R. H. C. Yap
Definition 1
A constraint cij is functional on variable j if for any a ∈ Di there exists at most
one b ∈ Dj such that cij (a, b). cij is functional on variable i if cji is functional on
i. Given a constraint cij functional on variable j and a value a ∈ Di , we assume
throughout the paper that in constant time we can find the value b ∈ Dj , if there
is one, such that (a, b) ∈ cij .
A special case of functional constraints are equations. These are ubiquitous in
CLP. A typical functional constraint in arithmetic is a binary linear equation like
2x = 5 − 3y which is functional on x and on y. Functional constraints do not need
to be linear. For example, a nonlinear equation x2 = y 2 where x, y ∈ 1..10 is also
functional on both x and y. In scene labeling problems (Kirousis 1993), there are
many functional constraints and other special constraints.
When a constraint cij is functional on variable j, for simplicity, we say cij is
functional by making use of the fact that the subscripts of cij are an ordered pair.
When cij is functional on variable i, cji is said to be functional. That cij is functional
does not mean cji is functional. In this paper, the definition of functional constraints
is different from the one in (Zhang et al. 1999; Van Hentenryck et al. 1992) where
constraints are functional on each of its variables, leading to the following notion.
Definition 2
A constraint cij is bi-functional if cij is functional on variable i and also on variable
j.
A bi-functional constraint is called bijective in (David 1995). For functional constraints, we have the following property on their composition and intersection: 1)
If cij and cjk are functional on variables j and k respectively, their composition
remains functional; and 2) The intersection of two functional constraints remains
functional.
Example 2
The constraint cij = {(a1 , b1 ), (a2 , b1 ), (a3 , b2 )} is functional, while the constraint
cij = {(a1 , b3 ), (a2 , b1 ), (a3 , b2 )} is both functional and bi-functional. An example
of a non-functional constraint is cij = {(a1 , b1 ), (a1 , b2 ), (a2 , b1 ), (a3 , b2 )}.
In the remainder of the paper, rather than writing vi , we will simply refer to a
variable by its subscript, i.e. i rather than vi .
3 Variable Substitution and Elimination Using Binary Functional
Constraints
We introduce the idea of variable substitution. Given a CSP (N, D, C), a constraint
cij ∈ C that is functional on j, and a constraint cjk in C, we can substitute j by i
in cjk by composing cij and cjk . If there is already a constraint cik ∈ C, the new
constraint on i and k is simply the intersection of cik and cjk ◦ cij .
Solving Functional Constraints by Variable Substitution
5
Definition 3
Consider a CSP (N, D, C), a constraint cij ∈ C functional on j, and a constraint
cjk ∈ C. To substitute j by i in cjk , using cij , is to get a new CSP where cjk is
replaced by c′ik = cik ∩ (cjk ◦ cij ). The variable i is called the substitution variable.
A fundamental property of variable substitution is that it preserves the solution
space of the problem.
Property 1
Given a CSP (N, D, C), a constraint cij ∈ C functional on j, and a constraint
cjk ∈ C, the new problem obtained by substituting j by i in cjk is equivalent to
(N, D, C).
Proof
Let the new problem after substituting j by i in cjk be (N, D, C ′ ) where C ′ =
(C − {cjk }) ∪ {c′ik } and c′ik = cik ∩ (cjk ◦ cij ).
Assume (a1 , a2 , · · · , an ) is a solution of (N, D, C). We need to show that it satisfies
C ′ . The major difference between C ′ and C is that C ′ has new constraint c′ik . It
is known that (ai , aj ) ∈ cij , (aj , ak ) ∈ cjk , and if there is cik in C, (ai , ak ) ∈ cik .
The fact that c′ik = (cjk ◦ cij ) ∩ cik implies (ai , ak ) ∈ c′ik . Hence, c′ik is satisfied by
(a1 , a2 , · · · , an ).
Conversely, we need to show that any solution (a1 , a2 , · · · , an ) of (N, D, C ′ ) is
a solution of (N, D, C). Given the difference between C ′ and C, it is sufficient
to show the solution satisfies cjk . We have (ai , aj ) ∈ cij and (ai , ak ) ∈ c′ik . Since
c′ik = (cjk ◦cij )∩cik , there must exist b ∈ Dj such that (ai , b) ∈ cij and (b, ak ) ∈ cjk .
As cij is functional, b has to be aj . Hence, aj and ak satisfy cjk .
Based on variable substitution, we can eliminate a variable from a problem so
that no constraint will be on this variable (except the functional constraint used to
substitute it).
Definition 4
Given a CSP (N, D, C) and a constraint cij ∈ C functional on j, to eliminate j
using cij is to substitute j by i, using cij , in every constraint cjk ∈ C (except cji ).
We can also substitute j by i in cji to obtain c′ii and then intersect c′ii with the
identity relation on Di , equivalent to a direct revision of the domain of i with respect
to cij . This would make the algorithms presented in this paper more uniform, i.e.,
only operations on constraints are used. Since in most algorithms we want to make
domain revision explicit, we choose not to substitute j by i in cji .
Given a functional constraint cij of a CSP (N, D, C), let Cj be the set of all
constraints involving j, except cij . The elimination of j using cij results in a new
problem (N, D, C ′ ) where
C ′ = (C − Cj ) ∪ {c′ik | c′ik = (cjk ◦ cij ) ∩ cik , cjk ∈ C}.
In the new problem, there is only one constraint cij on j and thus j can be regarded
as being “eliminated”.
6
Y. Zhang and R. H. C. Yap
k
j
i
(a)
k
j
i
l
(b)
l
Fig. 1. (a): A CSP with a functional constraint cij . (b): The new CSP after eliminating the variable j using cij .
Example 3
Consider a problem with three constraints whose constraint graph is shown in
Figure 1(a). Let cij be functional which this is indicated by the arrow in the diagram.
The CSP after j has been eliminated using cij is shown in Figure 1(b). In the new
CSP, constraints cjk and cjl are discarded, and new constraints cik = cjk ◦ cij and
cil = cjl ◦cij are added. Note that the other edges are not directed as the constraints
cjk , cjl , cik , cil may not be functional.
The variable elimination involves “several” substitutions and thus preserves the
solution space of the original problem by Property 1.
Corollary 1
Given a CSP (N, D, C) and a functional constraint cij ∈ C, the new problem
(N, D, C ′ ) obtained by the elimination of variable j using cij is equivalent to
(N, D, C).
4 Elimination Algorithms for CSPs with Functional Constraints and
Non-Functional Constraints
We now extend variable elimination to general CSPs with functional and nonfunctional constraints. The idea of variable elimination (Definition 4 in Section 3)
can be used to reduce a CSP to the following canonical functional form.
Definition 5
A CSP (N, D, C) is in canonical functional form if for any constraint cij ∈ C
functional on j, the following conditions are satisfied: 1) if cji is also functional on
i(i.e., cij is bi-functional), either i or j is not constrained by any other constraint
in C; 2) otherwise, j is not constrained by any other constraint in C.
As a trivial example, a CSP without any functional constraint is in canonical
functional form. If a CSP contains some functional constraints, it is in canonical
functional form intuitively if for any functional constraint cij , there is only one
constraint on j. As an exception, the first condition in the definition implies that
when cij is bi-functional, one variable of {i, j} might have several bi-functional
constraints on it.
In a canonical functional form CSP, the functional constraints form disjoint star
graphs. A star graph is a tree where there exists a node, called the center, which we
Solving Functional Constraints by Variable Substitution
7
call the free variable, such that there is an edge between this center node and every
other node, which we call and eliminated variable. The constraint between the free
variable and eliminated variable is that it is functional on the eliminated variable.
In Figure 1(a), assuming cjk and cjl are functional on k and l respectively, then
there would be directed edges (arrows) from j to k and j to l. After eliminating j,
we get a star graph in Figure 1(b), since i will be the free variable at the center
of the star graph, with free variables k and l. Notice that before eliminating j,
Figure 1(a) is a star graph, but the constraints are not in a canonical form.
The constraint between a free variable i and an eliminated variable j is functional
on j, but it may or may not be functional on i. In the special case that the star graph
contains only two variables i and j and cij is bi-functional, one of the variables can
be called a free variable while the other is called an eliminated variable.
If a CSP is in canonical functional form, all functional constraints and the eliminated variables can be ignored when we try to find a solution for this problem.
Thus, to solve a CSP (N, D, C) in canonical functional form whose non-eliminated
variables are N E, we only need to solve a smaller problem (N E, D′ , C ′ ) where D′
is the set of domains of the variables N E and C ′ = {cij | cij ∈ C and i, j ∈ N E}.
Proposition 1
Consider a CSP P1 = (N, D, C) in a canonical functional form and a new CSP
P2 = (N E, D′ , C ′ ) formed by ignoring the eliminated variables in P1 . For any free
variable i ∈ N and any constraint cij ∈ C functional on j, assume any value of Di
has a support in Dj and this support can be found in constant time. Any solution
of P2 is extensible to a unique solution of P1 in O(|N − N E|) time. Any solution
of P1 can be obtained from a solution of P2 .
Proof
Let (a1 , a2 , · · · , a|N E| ) be a solution of (N E, D′ , C ′ ). Consider any eliminated variable j ∈ N − N E. In C, there is only one constraint on j. Let it be cij where i
must be a free variable. By the assumption of the proposition, the value of i in
the solution has a unique support in j. This support will be assigned to j. In this
way, a unique solution for (N, D, C) is obtained. The complexity of this extension
is O(|N − N E|).
Let S be a solution of (N, D, C) and S ′ the portion of S restricted to the variables
in N E. S ′ is a solution of (N E, D′ , C ′ ) because C ′ ⊆ C. S ′ can then be extended
to S by using the functional constraints on the values of the free variables in S ′ to
give unique values for the variables in N − N E.
Any CSP with functional constraints can be transformed into canonical functional
form by variable elimination using the algorithm in Figure 2. Given a constraint cij
functional on j, line 1 of the algorithm substitutes j by i in all constraints involving
j. Note the arc consistency on cik , for all neighbor k of i, is enforced by line 3.
Theorem 1
Given a CSP (N, D, C), Variable Elimination transforms the problem into a
canonical functional form in O(n2 d2 ).
8
Y. Zhang and R. H. C. Yap
algorithm Variable Elimination(inout (N, D, C), out consistent) {
L ← N;
while (There is cij ∈ C functional on j where i, j ∈ L and i 6= j){
// Eliminate variable j,
1.
C ← {c′ik | c′ik ← (cjk ◦ cij ) ∩ cik , cjk ∈ C, k 6= i} ∪ (C − {cjk ∈ C | k 6= i});
2.
L ← L − {j};
3.
Revise the domain of i wrt cik for every neighbour k of i;
if (Di is empty) then { consistent ← false; return }
}
consistent ← true;
}
Fig. 2. A variable elimination algorithm to transform a CSP into a canonical functional form.
Proof
Assume Variable Elimination transforms a CSP P1 = (N, D, C) into a new problem P2 = (N, D′ , C ′ ). We show that P2 is in canonical functional form. For any
constraint cij ∈ C ′ functional on j, there are two cases. Case 1: j ∈
/ L when the
algorithm terminates. Since j ∈ L when the algorithm starts and line 2 is the only
place where j can be removed from L, j must have been eliminated at certain step
of the while loop. In line 1 (the component after “∪”), all constraints on j (except
cij ) are removed. That is cij is the unique constraint on j. Case 2: j ∈ L when the
algorithm terminates. Since cij is functional on j, variable i is not in L when the
algorithm terminates (otherwise, j will be substituted by line 1 at certain step of
the while loop). Therefore, i is removed from L at certain step of the while loop. i is
not substituted using cki where k 6= j (otherwise cij 6∈ C ′ because of the elimination
of i). This implies that i was substituted using cji , and thus cji is functional on i
by the loop condition. Hence, cij is bi-functional, and i is not constrained by any
other constraints (thanks to line 1). Therefore, cases 1 and 2 show that P2 is in
canonical functional form.
Next, we prove the complexity of Variable Elimination. The algorithm eliminates any variable in N at most once because once it is eliminated it is removed
from L (line 2). Assume, before the algorithm, there is at most one constraint on
any pair of variables (otherwise, we take the intersection of all constraints on the
same pair of variables as the unique constraint). This property holds during the
elimination process because in line 1, the intersection in the component before “∪”
guarantees that we have only one copy of constraint on any two variables. So, for
each variable j and a constraint cij functional on j, there are at most n − 2 other
constraints on j. The variable j in those constraints needs to be substituted (line
1).
The complexity of the substitution j in each constraint is O(d2 ) which is the cost
of the composition of a functional constraint and a general constraint. Recall that,
for a functional constraint cij , given a value a ∈ Di , we can find its support in
Dj in constant time. To compose cij with a general constraint cjk , for each value
a ∈ Di , we find its support b ∈ Dj (in constant time). If we take each constraint
Solving
SolvingFunctional
FunctionalConstraints
ConstraintsbybyVariable
VariableSubstitution
Substitution
l1
i
......
..
j
......
..
k
...
......
......
......
......
.
.
.
.
.
.....
.....
.....
.....
.....
.....
.....
.
l2
(a)
..........
...
...
.
.
...
...
...
...
..
....
1 ........
j
j3...
9
i
...
...
...
...
...
...
...
..........
j2
......
..
......... 1....
...
......
...
.....
...
....
.
.
.
...
...
.
.
...
.
..
.
...
.
.
.
....
....
.
.
.
........
..
.
.
.
.
.
2.......
i
i3
(b)
Fig.3.3.(a)
(a)The
Theconstraint
constraint graph
constraints
ij and
jk . (b)
Fig.
graph ofofaaCSP
CSPwith
withfunctional
functional
constraints
cij and
cjk .A
directed graph.
(b) A directed graph.
as a matrix, the row of of
will be the row of
, which takes steps.
as a matrix, the row of b of cjkjkwill be the row of cjkjk◦ cij ,ij which takes d steps.
Therefore, the cost of computing
is
).
Therefore, the cost of computing cjkjk◦ cij ijis O(d2 ).
For
2 constraints, the elimination of (line 1) takes nd
). There are at most
For n−2 constraints, the elimination of j (line 1) takes O(nd2 ). There are at most
1 variables to eliminate and thus the worst case complexity of the algorithm is
n − 1 variables to eliminate and thus the worst case complexity of the algorithm is
).
O(n2 d2 ).
worthnoting
notingthat
thatthe
thevariable
variableelimination
eliminationalgorithm
algorithmis isable
abletotoglobally
globallysolve
solve
ItItis isworth
some
CSPs
containing
non-functional
constraints.
some CSPs containing non-functional constraints.
Example4 4
Example
Considera asimple
simpleexample
examplewhere
wherethere
thereare
arethree
threevariables
variablesi, i,
andk whose
whosedodoConsider
j,j,and
mains
are
and
the
constraints
are
+
1,
and
.
Note
that
mains are {1, 2, 3} and the constraints are i = j, i = k + 1, and j 6= k. Note that
althoughthe
theconstraints
constraintsare
arelisted
listedininananequational
equationalform,
form,the
theactual
actualconstraints
constraintsare
are
although
explicitand
anddiscrete,
discrete,thus
thusnormal
normalequational
equationalreasoning
reasoningmight
mightnot
notbebeapplicable.
applicable.By
By
explicit
eliminatingj using
usingcij ,ij cikikbecomes
becomes{(2,(21),1)(3,(32)},
2) and
, andthe
thedomain
domainofofi becomes
becomes
eliminating
. Thenon-functional
non-functionalconstraint
constraintcjkjkis isgone.
gone.The
Theproblem
problemis isinincanonical
canonicalfuncfunc{2, 3}. The
tional
form.
A
solution
can
be
obtained
by
letting
be
2
and
consequently
tional form. A solution can be obtained by letting i be 2 and consequently j ==2 2
andk ==1.1.
and
By
Bycarefully
carefullychoosing
choosingananordering
orderingofofthe
thevariables
variablestotoeliminate,
eliminate,a afaster
fasteralgorithm
algorithm
can
canbebeobtained.
obtained.The
Theintuition
intuitionis isthat
thatonce
oncea avariable
variablei is isused
usedtotosubstitute
substituteforfor
other
othervariables,
variables,i itself
itselfshould
shouldnot
notbebesubstituted
substitutedbybyany
anyother
othervariable
variablelater.
later.
Example
Example5 5
Consider
. .ItsItsconstraint
Considera aCSP
CSPwith
withfunctional
functionalconstraints
constraintscij ijand
andcjkjk
constraintgraph
graphis is
shown
in
Figure
3(a)
where
a
functional
constraint
is
represented
by
an
arrow.
shown in Figure 3(a) where a functional constraint is represented by an arrow.If If
weweeliminate
eliminatek and
andthen
thenj, we
, wefirst
firstget
getcjl1jland
andcjl2jl, and
, andthen
thenget
getcil1iland
andcil2il. Note
. Note
that
thatk is isfirst
firstsubstituted
substitutedbybyj and
andthen
thenlater
laterj is issubstituted
substitutedbybyi. If
. Ifweweeliminate
eliminate
j and
, and
andthen
thenk, we
, wefirst
firstget
getcikik
, andthen
thenget
getcil1il and
andcil2il. In
. Inthis
thisway,
way,wewereduce
reducethe
the
number
of
compositions
of
constraints.
number of compositions of constraints.
F
Given
D,D,
C),
E)E
Givena aCSP
CSPP ==(N,
(N,
C),P is isused
usedtotodenote
denoteitsitsdirected
directedgraph
graph(V,
(V,
where
j)j | cij ∈ C and
whereV = N and
andE = {(i, i,
andcij ijis isfunctional
functionalononj}. Non-functional
. Non-functional
ij
F
constraints
in
C
do
not
appear
in
P
.
A
subgraph
constraints in
do not appear in
. A subgraphofofa adirected
directedgraph
graphis isstrongly
strongly
connected
connectedif ifforforany
anytwo
twovertices
verticesofofthe
thesubgraph,
subgraph,any
anyone
oneofofthem
themis isreachable
reachablefrom
from
the
theother.
other.AAstrongly
stronglyconnected
connectedcomponent
componentofofa adirected
directedgraph
graphis isa amaximum
maximumsubsubgraph
that
is
strongly
connected.
To
describe
our
algorithm
we
need
the
following
graph that is strongly connected. To describe our algorithm we need the following
notation.
notation.
10
Y. Zhang and R. H. C. Yap
Definition 6
Given a directed graph (V, E), a sequence of the nodes of V is a functional elimination ordering if for any two nodes i and j, i before j in the sequence implies that
there is a path from i and j. A functional elimination ordering of a CSP problem
P is a functional elimination ordering of P F .
The functional elimination ordering is used to overcome the redundant computation shown in the example on Figure 3(a). Given a directed graph G, a functional
elimination ordering can be found by: 1) finding all the strongly connected components of G; 2) modifying G by taking every component as one vertex with edges
changed and/or added accordingly; 3) finding a topological ordering of the nodes
in the new graph; and 4) replacing any vertex v in the ordering by any sequence of
the vertices of the strongly connected component represented by v.
To illustrate the process, consider the example in Figure 3(b) which can be taken
as P F for some CSP problem P . All strongly connected components are {j1 , j2 , j3 },
denoted by c1 , and {i1 , i2 , i3 }, denoted by c2 . We construct the new graph by
replacing the components by vertices: ({c1 , c2 }, {(c1 , c2 )}). We have the edge (c1 , c2 )
because the two components are connected by (j2 , i2 ). The topological ordering of
the new graph is hc1 , c2 i. Now we can replace c1 by any sequence of j’s and c2 by
any sequence of i’s. For example, we can have a functional elimination ordering
hj3 , j2 , j1 , i2 , i3 , i1 i.
The algorithm Linear Elimination in Figure 4 first finds a functional elimination ordering O (line 1). The body of the while loop at line 4 is to process all the
variables in O. Every variable i of O is processed as follows: i will be used to substitute for all the variables reachable from i through constraints that are functional
in C 0 and still exist in the current C. Those constraints are called qualified constraints. Specifically, L initially holds the immediate reachable variables through
qualified constraints (line 8). Line 9 is a loop to eliminate all variables reachable
from i. The loop at line 11 is to eliminate j using i from the current C. In this
loop, if a constraint cjk is qualified (line 14), k is reachable from i through qualified
constraints. Therefore, it is put into L (line 15).
To illustrate the ideas underlying the algorithm, consider the example in Figure 3(b). Now, we assume the edges in the graph are the only constraints in the problem. Assume the algorithm finds the ordering given earlier: O = hj3 , j2 , j1 , i2 , i3 , i1 i.
Next, it starts from j3 . The qualified constraints leaving j3 are cj3 j2 only. So, the
immediate reachable variables through qualified constraints are L = {j2 }. Take
and delete j2 from L. Substitute j2 by j3 in constraints cj2 i2 and cj2 j1 . As a result,
constraints cj2 i2 and cj2 j1 are removed from C while cj3 j1 = cj3 j1 ∩ (cj2 j1 ◦ cj3 j2 ) and
new constraint cj3 i2 = cj2 i2 ◦ cj3 j2 is introduced to C. One can verify that both cj2 j1
and cj2 i2 are qualified. Hence, variables j1 and i2 are reachable from j3 and thus are
put into L. Assume j1 is selected from L. Since there are no other constraints on
j1 , nothing is done. Variable i2 is then selected from L. By eliminating i2 using j3 ,
ci2 i1 and ci2 i3 are removed from C and cj3 i1 and cj3 i3 are added to C. Constraint
ci2 i1 is qualified, and thus i1 is added to L. Note that ci2 i3 is not qualified because
it is not functional on i3 in terms of the graph. We take out the only variable i1 in
Solving Functional Constraints by Variable Substitution
11
algorithm Linear Elimination(inout (N, D, C)) {
1. Find a functional elimination ordering O of the problem;
2. Let C 0 be C; any cij in C 0 is denoted by c0ij ;
3. For each i ∈ N , it is marked as not eliminated ;
4. while (O is not empty) {
5.
Take and delete the first variable i from O;
6.
if (i is not eliminated ) {
8.
L ← {j | (i, j) ∈ C and c0ij is functional};
9.
while (L not empty) {
Take and delete j from L;
11.
for any cjk ∈ C − {cji } { // Substitute j by i in cjk ;
c′ik ← cjk ◦ cij ∩ cik ;
C ← C ∪ {c′ik } − {cjk };
14.
if (c0jk is functional) then
15.
L ← L ∪ {k};
}
16.
Mark j as eliminated ;
} // loop on L
}
} // loop on O
} // end of algorithm
Fig. 4. A variable elimination algorithm of complexity O(ed2 ).
L. After i1 is eliminated using j3 , ci1 i3 is removed from C, and constraint cj3 i3 is
updated to be cj3 i3 ∩ (ci1 i3 ◦ cj3 i1 ). Since ci1 i3 is qualified, i3 is added to L. One can
see that although i3 was not reachable when i2 was eliminated, it finally becomes
reachable because of i1 . All the variables in a strongly connected component are
reachable from the variable under processing if one of them is reachable. Now, take
i3 out of L, and nothing is done because there are no other constraints incident on
it. Every variable except j3 is marked as eliminated (line 16), the while loop on O
(line 4 and 6) terminates.
Theorem 2
Given a CSP problem, the worst case time complexity of
Linear Elimination is O(ed2 ) where e is the number of constraints and d the size
of the maximum domain in the problem.
Proof
To find a functional elimination ordering involves the identification of strongly
connected components and topological sorting. Each of the two operations takes
linear time. Therefore, line 1 of the algorithm takes O(n + e).
The while loop of line 4 takes O(ed2 ). Assume that there is a unique identification
number associated with each constraint in C. After some variable of a constraint is
substituted, the constraint’s identification number refers to the new constraint.
For any identification number α, let its first associated constraint be cjk . Assuming j is substituted by some other variable i, we can show that i will be never be
substituted later in the algorithm. By the algorithm, i is selected at line 5. Since
12
Y. Zhang and R. H. C. Yap
it is the first element of O now, all variables before i in the original functional
ordering have been processed. Since i is not eliminated, it is not reachable from any
variable before it (in terms of the original O) through qualified constraints (due to
the loop of line 9). Hence, there are two cases: 1) there is no constraint cmi of C
such that c0mi is functional on i, 2) there is at least one constraint cmi of C such
that c0mi is functional on i. In the first case, our algorithm will never substitute i by
any other variable. By definition of functional elimination ordering, case 2 implies
that i belongs to a strongly connected component whose variables have not been
eliminated yet. Since all variables in the component will be substituted by i, after
the loop of line 9, there is no constraint cmi of C such that c0mi is functional on i.
Hence, i will never be substituted again.
In a similar fashion, if variable k is substituted by l, l will never be substituted
later by the algorithm.
So, there are at most two substitutions occurring to α. Each of these substitutions
is a composition that involves a functional constraint. Hence, its complexity is O(d2 )
in the worst case as shown in the proof of Theorem 1.
Since there is a unique identification number for each constraint, the total number
of the unique identification numbers is e and thus the time taken by the while loop
at line 4 is O(ed2 ). In summary, the worst case time complexity of the algorithm is
O(ed2 ).
Before proving some properties of Linear Elimination, we first define trivially
functional constraints.
Definition 7
Given a problem P , let C 0 be the constraints before applying Linear Elimination
and C the constraints of the problem at any moment during the algorithm. A
constraint cij of C is trivially functional if it is functional and satisfies the condition:
c0ij is functional or there is a path i1 (= i), i2 , · · · , im (= j) in C0 such that, ∀k ∈
1..m − 1, c0ik ik+1 is functional on ik+1 .
Theorem 3
Algorithm Linear Elimination transforms a CSP (N, D, C) into a canonical functional form if all newly produced functional constraints (due to substitution) are
trivially functional.
The proof of this result is straightforward and thus omitted here.
Corollary 2
For a CSP problem with non-functional constraints and bi-functional constraints,
the worst case time complexity of algorithm Linear Elimination is linear to the
problem size.
This result follows the observation below. When the functional constraint involved
in a substitution is bi-functional, the complexity of the composition is linear to the
constraints involved. From the proof of Theorem 2, the complexity of the algorithm
is linear to the size of all constraints, i.e., the problem size.
Solving Functional Constraints by Variable Substitution
13
Corollary 3
Consider a CSP with both functional and non-functional constraints. If there is a
variable of the problem such that every variable of the CSP is reachable from it in
P F , the satisfiability of the problem can be decided in O(ed2 ) using Linear Elimination.
For a problem with the property given in the corollary, its canonical functional
form becomes a star graph. So, any value in the domain of the free variable is extensible to a solution if we add (arc) consistency enforcing during Linear Elimination.
The problem is not satisfiable if a domain becomes empty during the elimination
process.
5 Experimental Results
We experiment to investigate the effectiveness of variable elimination on problem
solving. In our experiments, a problem is solved in two ways: (i) directly by a general
solver; and (ii) variable elimination is applied before the solver.
There are no publicly available benchmarks on functional constraints. We test
the algorithms on random problems which are sufficiently hard so that we can
investigate the effect of different numbers of functional constraints and the effect
of constraint tightness.
We generate random problems hn, d, e, nf, ti where n is the number of variables,
d domain size, e the number of constraints, nf the number of functional constraints,
and t the tightness of non-functional constraints. The tightness r is defined as
the percentage of allowed tuples over d2 . There are nf functional constraints and
the rest of the binary constraints are non-functional. Each functional constraint
is constructed to have d allowed tuples. In the context of random problems, the
tightness factor of 1/d due to the functional constraints is rather tight. When we
increase nf, it can be the case that the search space is quickly reduced due to the
effect of these very tight constraints. Therefore the “hardness” of the problems drops
correspondingly when there is a significant increase of nf . As described below, we
try to counter this effect by removing problems which be solved too easily from the
benchmarks.
In the experiments, we systematically test benchmark problems generated using
the following parameters: n, d are 50, e varies from 100 to 710 with step size 122
(710 is ∼ 10% of the total number of possible constraints (1225)), nf varies from
2 to 12, and t varies from 0.2 to 1.0 with step size 0.05. When nf is small (for
example, 2), there are so many hard problems that we can only experiment with
a small portion of the problems because it is computationally infeasible. When nf
is large (for example, 12), even the most difficult problem instances from the set
of instances becomes easy and only a small number of backtracks is needed. These
instances can be solved too easily and thus are not very useful for benchmarking the
elimination algorithm (i.e., we will not expect the elimination algorithm to bring
any benefits to these instances). So, we do not include the cases with nf > 12.
When nf = 12, the most difficult problems we found are with e = 710. Table 1
14
Y. Zhang and R. H. C. Yap
tightness (t)
0.2 – 0.65
0.7
0.75
0.8
0.85
0.9 – 0.95
#bt
0
5.7
22.9
1023
0.2
0
Table 1. Hardness of the problem versus tightness
Table 2. The experimental results for random problems with n = d = 50.
e
nf
tightness
344
466
588
588
710
710
588
6
7
8
9
10
11
12
0.60
0.70
0.75
0.75
0.80
0.80
0.75
Elimination
cpu time (s) #backtracks
20042
9266
17922
10346
3039
481
24
No Elimination
cpu time (s) #backtracks
1.889e+06
9.362e+05
1.635e+06
5.679e+05
2.244e+05
26522
682
47781
35136
45464
21231
5771
959.7
57.9
5.381e+06
3.955e+06
4.386e+06
1.605e+06
5.146e+05
71258
2960
shows the hardness of problems instances, with nf = 12 and e = 710, in terms of
the number of backtracks #bt (average of 10 instances) needed. The hardness is
measured using an arc consistency solver without using the elimination algorithm.
When t is from 0.2 to 0.65, the problems are too easy (#bt is 0). For the most
difficult case of t being 0.8, #bt is still rather small (around 1000). On the other
hand, when nf is small, one can expect that the application of elimination may
not make much difference. Therefore, we do not include the cases when nf is small
either.
Due to the observations above, we evaluate the algorithm only on non-trivial
problem instances and where nf is not too tiny. For each nf (varying from 6 to 12),
the results of the most difficult problem instances discovered in the exploration
process above is shown in Table 2. The results were obtained on a DELL PowerEdge 1850 (two 3.6GHz Intel Xeon CPUs) with Linux. We implement both the
elimination algorithm and a general solver in C++. The solver uses the standard
backtracking algorithm armed with arc consistency enforcing algorithm after each
variable assignment. During the search, the dom/deg heuristic is used to select a
variable, and the value selection heuristic is in lexicographical order.4
In Table 2, the cpu time is the total time of twenty problem instances for a given
combination of e, nf and tightness, and the number of backtracks are their average.
For the problem instances used in Table 2, the time to transform the instances into
their canonical forms is negligible compared to the time needed for solving the
4
The dom/deg heuristic is a dynamic variable selection heuristic.
Solving Functional
Constraints
by Variable
Substitution 15
Solving Functional
Constraints
by Variable
Substitution
number of backtracks w/o elimination
1e+08
y=x
y = 5x
nf = 6
nf = 7
nf = 8
nf = 9
nf = 10
nf = 11
nf = 12
1e+07
1e+06
100000
10000
1000
100
10
1
1
10
100
1000 10000 100000 1e+06 1e+07
number of backtracks with elimination
Fig. 5. The number of backtracks needed with and without elimination.
Fig. 5. The number
of backtracks needed with and without elimination.
arereasons.
several reasons.
the of
number
of constraints
in
instance.instance.
There areThere
several
First, theFirst,
number
constraints
involved involved
in
the elimination
is relatively
small compared
to the
total of
number
of constraints
the elimination
is relatively
small compared
to the total
number
constraints
in the problems.
Second,
the algorithm
is as as
efficient
as the general
optimal arc
general arc
in the problems.
Second, the
algorithm
is as efficient
the optimal
consistency
algorithm
used
in
the
solver.
Thirdly,
the
elimination
is
applied
only
consistency algorithm used in the solver. Thirdly, the elimination is applied only
once to
reduce
the which
problem
canbefore
be done
the backtracking
search, while
once to reduce
the
problem
canwhich
be done
thebefore
backtracking
search, while
the arc consistency
to be
every
stepthe
during
the search,
the arc consistency
algorithmalgorithm
needs to needs
be called
at called
every at
step
during
search,
i.e. about
roughly
same
as the of
number
of backtracks.
i.e. roughly
theabout
samethe
as the
number
backtracks.
Theshow
results
show
the variable
elimination
can significantly
The results
that
the that
variable
elimination
can significantly
speed upspeed
the up the
probleminsolving
in both
termscpu
of both
and theof
number
of backtracks.
problem solving
terms of
time cpu
and time
the number
backtracks.
It reducesIt reduces
the of
number
of backtracks
by four
two times
to four
times
also the
reduces
the number
backtracks
by two to
and
also and
reduces
cpu the
timecpu time
correspondingly.
correspondingly.
The statistics
and of
number
of backtracks)
used in
Table
2 is for 20
The statistics
(cpu time(cpu
andtime
number
backtracks)
used in Table
2 is
for 20
problem instances
for each
value
of the parameters
selected parameters
row
in the table).
problem instances
for each value
of the
selected
(each row(each
in the
table).
notice
the hardness
of these instances
is not uniform,
some instances
We noticeWe
that
the that
hardness
of these instances
is not uniform,
i.e., somei.e.,
instances
are significantly
harder
theTo
others.
better visualize
the performance
are significantly
harder than
the than
others.
betterTo
visualize
the performance
of the of the
algorithms,
we
replot
the
same
data
from
Table
2
in
Figure
5.
Each
data
algorithms, we replot the same data from Table 2 in Figure 5. Each data point
in point in
5, represents
an instance
whose -coordinate
is the of
number
of backtracks
Figure 5,Figure
represents
an instance
whose x-coordinate
is the number
backtracks
with elimination
applied
its -coordinate
is that elimination.
without elimination.
Both
with elimination
applied while
its while
y-coordinate
is that without
Both
a log scale.
axis use aaxis
loguse
scale.
Theplot
scatter
plot in5 Figure
shows aperformance
similar performance
improvement
The scatter
in Figure
shows a5 similar
improvement
resulting resulting
from elimination.
that allabove
points
indicate
from elimination.
Note thatNote
all points
theabove
line ythe
= line
x indicate that
not that
us- not using elimination
requires
more backtracks.
We highlight
the instances
with
ing elimination
requires more
backtracks.
We highlight
the instances
with nf =
7 nf = 7
(the insymbol
in thewhich
graph),
which
shows
more results
extremeasresults
as elimination
can
(the × symbol
the graph),
shows
more
extreme
elimination
can
significantly
speed
up
or
slow
down
the
problem
solving.
The
slowing
down
is
an
significantly speed up or slow down the problem solving. The slowing down is an
interesting
of this experiment.
An explanation
is that eliminavariable eliminainteresting
discoverydiscovery
of this experiment.
An explanation
is that variable
16
16
. Zhang
R. H. C. Yap
Y. Zhang and
R. H.and
C. Yap
10000
cpu time
1000
100
10
1
w/ elimination
w/o elimination
0.1
0.7 0.71 0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79 0.8
tightness
The cpu
time for instances
50, =
nf 8=and
8 and
588 and
and tightness
Fig. 6. TheFig.
cpu6. time
for instances
with n with
= d = 50,= nf
e ==588
varying from 0 70 to 0 80.
tightness varying from 0.70 to 0.80.
tionthe
changes
the of
topology
of the which
problem,
the variable
dynamic variable
tion changes
topology
the problem,
maywhich
affectmay
the affect
dynamic
ordering heuristics
of the
general
It is well
known
that the performance
of a
ordering heuristics
of the general
solver.
It issolver.
well known
that
the performance
of a
constraint
solver
may
vary
significantly
with
a
different
variable
ordering.
constraint solver may vary significantly with a different variable ordering.
have idea
a better
ideathe
on elimination
how the elimination
To have aTo
better
on how
algorithm algorithm
performs, performs,
we look atwe look at
instances
with
various hardness.
We now
intoofthe
nfe =
and = 588
instances with
various
hardness.
We now zoom
into zoom
the case
nf case
= 8 of
and
= 8588
—
this
has
a
large
cpu
time
and
is
more
tight
from
the
the
experiments
— this has a large cpu time and is more tight from the the experiments in Table in Table
2. Wethat
remark
that our algorithm
performs in
similarly
in cases,
all these
cases,
2. We remark
our algorithm
performs similarly
all these
so we
justso we just
look
at the
of aone
specific
one results
with the
for all configurations
where
look at the
details
of details
a specific
with the
forresults
all configurations
where
the changes
tightnessfrom
changes
0 70
to 0a 80
with
a stepThe
of 0results
01. The
the tightness
0.70 from
to 0.80
with
step
of 0.01.
areresults
shownare shown
6 (cpu
Figure
7 (theofnumber
of backtracks).
in Figure.in6 Figure.
(cpu time)
and time)
Figureand
7 (the
number
backtracks).
Again, theAgain,
cpu the cpu
time
is
the
sum
of
the
cost
of
20
instances
per
parameter
setting
while
the number
time is the sum of the cost of 20 instances per parameter setting while the number
of backtracks
is their When
average.
the is
tightness
is 0.80
0 79 the
and problem
0 80 the problem
of backtracks
is their average.
theWhen
tightness
0.79 and
instancesvery
become
very
simple
with 200
less backtracks,
than 200 backtracks,
do not
instances become
simple
with
less than
we do not we
expect
theexpect the
elimination
improve
the performance
of the constraint
solver although
elimination
algorithm algorithm
to improvetothe
performance
of the constraint
solver although
reduces
theofnumber
of backtracks.
trivial problems,
elimination
it reducesitthe
number
backtracks.
For most For
non most
trivialnon
problems,
elimination
help tothe
improve
the (both
efficiency
cpu the
timenumber
and theofnumber
of backtracks)
does help does
to improve
efficiency
cpu (both
time and
backtracks)
significantly.
When
using 10 instances,
we also that
observed
when the is
tightness is
significantly.
When using
10 instances,
we also observed
whenthat
the tightness
73, the elimination
a worse performance
of the
generalsolver
problem
0.73, the elimination
leads to a leads
worsetoperformance
of the general
problem
in solver in
terms
of time
both and
cpu the
timenumber
and theofnumber
of backtracks.
terms of both
cpu
backtracks.
As observed,
when theofnumber
of functional
in theproblems
random problems
As observed,
when the number
functional
constraintsconstraints
in the random
theinstances
problem become
instances
become
(i.e.,backtracks
very few backtracks
increases, increases,
the problem
trivial
(i.e.,trivial
very few
are neededare needed
to solve
them).
This
makes
hardevaluate
to fullythe
evaluate
the elimination
To
to solve them).
This
makes
it hard
to itfully
elimination
algorithm.algorithm.
To
the inconsistency
potential inconsistency
byfunctional
the tight functional
which
reduce thereduce
potential
caused by caused
the tight
constraintsconstraints
which
the problems
easy,identity
we usefunctions
identity functions
special case
makes themakes
problems
easy, we use
(i.e., x = (i.e.,
y, a special, acase
Solving Functional
Constraints
by Substitution
Variable Substitution 17
Solving Functional
Constraints
by Variable
number of backtracks
1e+06
100000
10000
1000
100
10
w/ elimination
w/o elimination
0.7 0.71 0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79 0.8
tightness
7. Theofnumber
of backtracks
for instances
with
= 50,
and = 588
Fig. 7. TheFig.
number
backtracks
for instances
with n
= d = 50,
nfnf==88 and
and
tightness
varying
from
0
70
to
0
80.
e = 588 and tightness varying from 0.70 to 0.80.
Table
3. The experimental
for problems
random problems
with identity
functions
Table 3. The
experimental
results forresults
random
with identity
functions
(the
problem
are
= 100).
(the problem parameters are
n=
d = e =parameters
100).
nf
nf
tightness
10
20
30
40
50
60
10
0.04
20
0.04
30
0.08
40
0.19
50
0.20
60
0.25
Elimination
No Elimination
No Elimination
ightness Elimination
cpu time
(s) #backtracks
cpu time
(s) #backtracks
cpu time (s)
#backtracks
cpu time (s)
#backtracks
0.04
0.04
0.08
0.19
0.20
0.25
11.5
1.4
1.1
1.1
0.7
0.6
11.5779.6
1.4 4.5
1.1 0.2
1.1
0
0.7
0
0.6
0
779.6 34.6
4.5
9606.8
0.2 40.7
0177.9
0893.1
2.2
34.6
1870.1
1870.1
9606.8
131724.0
131724.0
40.7564.6
564.6
177.9
1275.7
1275.7
893.1
64163.2 64163.2
2.2 1.4
1.4
of functional
constraint),
of arbitrary
constraints.
We remark
of functional
constraint),
instead ofinstead
arbitrary
functionalfunctional
constraints.
We remark
that
we
could
also
have
used
a
permutation
form
of
the
identity
function
that we could also have used a permutation form of the identity function but that but that
would
have
the instance
problem construction
instance construction
more complex.
With identity
would have
made
the made
problem
more complex.
With identity
functions,
no
inconsistency
will
result
directly
from
functional
constraints.
The
functions, no inconsistency will result directly from functional constraints. The
experimental
results
given 3inshow
Table
3 show
that
we can
trivial problems
experimental
results given
in Table
that
we can
create
noncreate
trivialnon
problems
with
much
more
functional
constraints.
with much more functional constraints.
Forofthis
set of instances,
the elimination
up the solving
problembysolving by
For this set
instances,
the elimination
can speedcan
up speed
the problem
up
to
several
orders
of
magnitude
(in
terms
of
both
cpu
time
and theofnumber of
up to several orders of magnitude (in terms of both cpu time and the number
backtracks).
An important
observation
is that
when
is a significant
backtracks).
An important
observation
is that when
there
is athere
significant
amount amount
of
functional
constraints,
the
number
of
backtracks
needed
after
elimination
of functional constraints, the number of backtracks needed after elimination can be can be
18
Y. Zhang and R. H. C. Yap
as small as 0. However, without elimination, the general problem solving may need
a large number of backtracks (up to five orders of magnitude larger).
In summary, from our experiments, for non trivial random problems, elimination
can improve the efficiency of a general constraint solver by several times to several orders of magnitude. We also observe that the elimination could make a solver
slower possibly due to the change of the topological structure of the problems. However, the slowdown only occurs rarely in our experiments.
5.1 Search Space Reduction through Elimination
The experiments show that the elimination could reduce the search space significantly. In fact, we can show that the elimination can help reduce the size of the
search space (in terms of the current domains of the variables).
Proposition 2
Given a CSP problem P , let P ′ be the problem resulting from applying the elimination algorithm to P . After enforcing arc consistency on P ′ and P , for each variable
of P ′ , its domain in P ′ is a subset of that in P .
Proof
Instead of proving the original proposition, we prove the following claim: given a
CSP problem P = (V, D, C) and a constraint cij functional on j, let P ′ = (V −
{j}, D′ , C ′ ) be the problem resulting from the elimination of j. Any value, not from
Dj , removed by enforcing arc consistency on P will be removed by enforcing arc
consistency on P ′ . Let P 1 = (V, D1 , C) and P ′1 = (V − {j}, D′1 , C ′ ) be the result of
enforcing arc consistency on P and P ′ respectively. Equivalently, we will show that
there exists Dj′′ ⊆ Dj such that P ′′ = (V, D′1 ∪ {Dj′′ }, C) (i.e., “plug” the domains
of P ′1 to P ) is arc consistent. If this claim holds, the proposition holds by applying
the claim repeatedly (as the elimination proceeds).
Let the neighbors of j in P be i, k1 , . . . , km . For any constraint clk ∈ C −{cij , cjk1 ,
. . . , cjkm }, clk ∈ C ′ . Since clk is arc consistent in P ′1 , clk is arc consistent in P ′′ .
We next show cij ∈ C is arc consistent with respect to Di′1 and Dj , i.e., for any
value a ∈ Di′1 , there is a support in Dj . Assume there is no support in Dj , by
the definition of substitution, a has no support in any domain of kn (n ∈ 1..m),
which contradicts that cikn ∈ C ′ is arc consistent with respect to Di′1 and Dk′1n .
Furthermore, let the support of a in Dj be b. We claim b has a support with
respect to any constraint cjk (k ∈ {k1 , . . . , km }) in P ′1 . Otherwise, a has no support with respect to cik ∈ C ′ for some k ∈ {k1 , . . . , km }, by the definition of
substitution. It contradicts the fact that cik is arc consistent in P ′1 . Let Dj′′ =
Dj − {b | b has no support with respect to cji and Di′1 }. Clearly Dj′′ is not empty
(because cij is arc consistent in P ′1 ). It can be shown that cji is arc consistent over
′
domains Di1 and Dj′′ and cij is still arc consistent on these domains too. In other
words, cij is arc consistent in P ′′ .
′
Similarly, we can show that cjk1 , . . . , and cjkm are arc consistent over Dk1n (n ∈
k1 ..km ) and Dj′′ , and thus in P ′′ .
Solving Functional Constraints by Variable Substitution
19
Furthermore, after enforcing arc consistency on P and P ′ , for some variables of
P , its domain in P ′ is a proper subset of that in P . Consider the following example:
V = {x, y, z}, x, y, z ∈ {1, 2}, and cxy = {(1, 1), (2, 2)}, cyz = {(1, 2), (2, 1)}, czx =
{(1, 1), (2, 2)}. This problem is arc consistent and the domains of the variables are
{1, 2}. However, the problem P ′ resulting from variable elimination has an empty
domain.
Thus, we see that using the elimination algorithm together with arc consistency
for the non-functional constraints leads to a “higher amount” of consistency.
′
6 Beyond Binary Constraints
Our presentation so far is based on binary constraints. To model real life problems,
non-binary constraints are often useful. In this section, we discuss the potential
extension of the work reported in this paper. The first subsection is to generalize
substitution to non-binary constraints in extensional form, and the second proposes
an approach to processing non-binary constraints in intensional form.
6.1 Variable Elimination and Non-binary Functional Constraints
In this section, we discuss the treatment of the generalization of binary functional
constraints and substitution to non-binary constraints. We first generalize the functional property from binary constraints to non-binary constraints. Then, we show
how a variable is substituted by a set of variables.
A non-binary constraint is denoted by cS where S is the set of variables in the
constraint. A linear equation x + y + z = 8 with finite domains for {x, y, z} is a
non-binary constraint. We now define variable instantiations. An instantiation of
a set of variables Y is an assignment of values to the variables in Y . It is usually
denoted by a sequence. An instantiation is denoted by a character with a bar, for
example, ā.
Definition 8
A constraint cS is functional on j(∈ S) if for any instantiation ā of S − {j}, there
is at most one value of j such that this value and ā satisfy cS . A constraint cS is
functional if it is functional on some variable j ∈ S.
Example 6
Consider a constraint x + y + z = 8 with x, y, z ∈ {1, 2, 3}. Let ā = (1, 1) and
b̄ = (2, 3) be two instantiations of (x, y). For ā, no value for z can be found to
satisfy the constraint. For b̄, value 3 is the only value for z to satisfy the constraint.
It can be verified that the constraint is functional on z, and similarly on x and on
y.
Example 7
The constraint x2 + y 2 + z = 8 with x, y, z ∈ {−1, −2, −3, 0, 1, 2, 3} is functional on
z but not functional on x or y.
20
Y. Zhang and R. H. C. Yap
The idea of variable substitution is applicable to the functional non-binary constraints defined above. However, we need more generalised operations to implement
variable substitution. Below, we always take a constraint cS as a non-binary relation
whose tuples are given explicitly. It is also helpful to recall that a relation is simply a
set, and we can apply set operations like intersection to relations. For example, the
constraint x+ y + z = 8 with x, y, z ∈ {1, 2, 3} is taken as {(2, 3, 3), (3, 2, 3), (3, 3, 2)}
where each tuple is an instantiation of variables (x, y, z).
In the context of non-binary constraints, the composition of two constraints cS
and cT with respect to a variable i ∈ S ∩ T , denoted by “◦i ,” is defined below.
cS ◦i cT = {ā |
ā is an instantiation of S ∪ T − {i} and there exists a ∈ Di
such that (ā, a) satisfies both cS and cT .}
If cS is functional on j and j is a variable of constraint cT , to substitute the
variable j in cT in terms of cS is to replace cT by cS ◦j cT .
Definition 9
Consider a CSP (N, D, C) and two constraints cS and cT in C. Assume cS is
functional on j ∈ S ∩T . To substitute j in constraint cT using cS is to get a new CSP
(N, D, C ′ ) where C ′ = (C−{cT })∪{c′S∪T −{j} } and c′S∪T −{j} = cS∪T −{j} ∩(cS ◦j cT ).
The variable substitution preserves the solution of a CSP.
Property 2
Given a CSP (N, D, C), a constraint cS ∈ C functional on j, and a constraint
cT ∈ C where j ∈ T , the new problem obtained after j in cT is substituted using
cS is equivalent to (N, D, C).
Proof
Let the new problem after j in cT is substituted be (N, D, C ′ ) where C ′ = (C −
{cT }) ∪ {c′S∪T −{j} } and c′S∪T −{j} = cS∪T −{j} ∩ (cS ◦j cT ).
Assume ā is a solution of (N, D, C). We shall show that ā also satisfies C ′ .
Given a set of variables Y , āY will be used to denote the values in the solution ā
for the variables in Y . C ′ differs from C in that it has the new constraint c′S∪T −{j} .
It is known that āS ∈ cS , āT ∈ cT , and if there is cS∪T −{j} in C, āS∪T −{j} satisfies
cS∪T −{j} . The fact that c′S∪T −{j} = (cS ◦j cT ) ∩ cS∪T −{j} implies āS∪T −{j} ∈
c′S∪T −{j} (∈ C ′ ). Hence, c′S∪T −{j} is satisfied by ā.
Conversely, we need to show that any solution ā of (N, D, C ′ ) is a solution of
(N, D, C). Given the difference between C ′ and C, it is only necessary to show ā
satisfies cT . To facilitate the following proof, we write āS as (āS−{j} , āj ), āT as
(āT −{j} , āj ). We have (āS−{j} , āj ) ∈ cS and āS∪T −{j} ∈ c′S∪T −{j} . Assume, by con/ cT . Since āS∪T −{j} ∈ c′S∪T −{j} , there must exist b ∈ Dj
tradiction, (āT −{j} , āj ) ∈
such that b 6= āj , (āT −{j} , b) satisfies cT , and (āS−{j} , b) satisfies cS , contradicting
that cS is functional on j. So, āT , that is (āT −{j} , āj ), satisfies cT .
A CSP (N, D, C) with non-binary functional constraints can be reduced by variable substitution in a similar way as developed in this paper. In the non-binary case,
we note that the complexity of the algorithm is more expensive due to the composition operation (which is very close to the join operation in relational databases).
Solving Functional Constraints by Variable Substitution
21
6.2 Variable Elimination and Non-binary Constraints
Non-binary constraints such as arithmetic or global constraints are common in CP
systems. We discuss how variable elimination of functional constraints can be applied to these constraints. Non-binary constraints are either in extensional (defined
explicitly) or intensional (defined implicitly) form. To substitute a variable in an
extensional non-binary constraint, we can follow the definition given the previous
subsection.
In most existing CP systems, for intentional constraints, there are usually particular propagators with a specific algorithm associated with them. In this case,
the approach using composition is not directly applicable simply because it has
to interact with a constraint defined in terms of an arbitrary specific propagation algorithm. We sketch below an approach which allows variable elimination
to be employed with generic propagators. Assume we have a linear constraint c1 :
ax + by + cz < d and a constraint cwy functional on y. To substitute y in c1 , we
simply modify c1 to be ax + bw + cz < d and mark w as a shadow variable (w
needs special treatment by the propagator, which will be clear later). We call y the
shadowed variable. Assume we also have cuw functional on w. To eliminate w, c1 is
further changed to ax+bu+cz < d. Since w is a shadow variable, we generate a new
constraint cuy using cuw and cwy in a standard way as discussed in this paper. Now
u becomes the shadow variable while the shadowed variable is still y (variable w is
gone). Suppose we need to make c1 arc consistent. First “synchronize the domains”
of y and u using cuy , i.e., enforce arc consistency on cuy . (Note that due to elimination, cwy and cuw are no longer involved in the constraint solving.) Next, we enforce
arc consistency on c1 . During the process, since u is a shadow variable, all domain
operations are on y instead of u. After making c1 arc consistent, synchronize the
domain of y and u again. (If the domain of u is changed, initiate constraint propagation on constraints involving u.) This approach is rather generic: for any intensional
constraints, synchronize the domains of the shadow variables and shadowed variables, apply whatever propagation methods on the shadowed variables (and other
non-shadow variables), synchronize the domains of shadow variables and shadowed
variables again. In fact, the synchronization of the domains of the shadow and shadowed variables (for example, u and y above) seems be readily implementable using
the concept of views (Schulte and Tack 2005).
7 Related Work
We now discuss variable substitution in the context of CLP followed by related work
in variable substitution algorithms from other domains. Finally, the relationship
to functional, bi-functional and other variable elimination approaches in the CSP
literature.
22
Y. Zhang and R. H. C. Yap
7.1 CLP and Constraint Solving
Logic Programming and CLP (Jaffar and Maher 1994) systems often make use of
variable substitution and elimination. The classic unification algorithm discussed
below is a good example.
A more complex example is CLP(R) (Jaffar et al. 1992) which has constraints on
finite trees and arithmetic. Variables in arithmetic constraints are substituted out
using a parametric normal form which is applied during unification and also when
solving arithmetic constraints. Our approach is compatible with such CLP solvers
which reduce the constraint store to a normal form using variable substitution. We
remark that any CLP language or system which has finite domain constraints will
deal with bi-functional constraints simply because of the need to match an atom
in the body with the head of a rule. The question is how powerful is the approach
used. In this paper, we show that a variable substitution approach is more powerful
than just simple finite domain propagation on equations. The consistency of the
CSP is increased. Our experiments show that the time to solve the problem can be
significantly smaller.
7.2 Unification, Gaussian Elimination, and Elimination Algorithm for
Functional Constraints
The algorithm for unification of finite trees, the Gaussian elimination algorithm for
linear constraints and our algorithm for functional constraints share the same key
techniques: variable substitution and elimination. Such algorithms are also commonly used in CLP systems. We illustrate this by examples.
The first is the unification of finite trees or terms. Unifying two terms f (x, y, z)
and f (y, z, g(x)), where x, y and z are variables, results in three term equations:
x = y, y = z, z = g(x). These equations can be solved using a variable elimination
method. We select a variable and eliminate it from the system by substitution.
For example, we can select to eliminate x — substitute all x by y (using the first
equation x = y), which results in y = z, z = (g(y)). This process continues until
some constant symbols do not match in an equation, the left hand side variable
appears in a sub-term on the right hand side (or vice versa), or no new equation
can be produced and there is only one term equation left.
Our second example is equation solving for arithmetic over the real numbers. A
system of binary linear constraints on real numbers can be solved by the well known
Gaussian elimination method. A major step is to select a variable and eliminate it
using an equation. Specifically, to eliminate x using ax + by = c, we substitute all
x in the system by (c − by)/a.
Lastly, we look at the case of finite domain constraints considered in this paper.
Given a binary CSP, when we have a variable x and a general constraint cyx functional on x, variable x will be eliminated by substituting x in the remainder of the
constraints. The substitution here is achieved by (general) composition of relations.
One can show that the substitution of x in Gaussian elimination produces an equation (a constraint) which is the result of the composition of the involved equations
Solving Functional Constraints by Variable Substitution
23
(constraints). In other words, given cyx : x = (c−by)/a and cxz : a1 x+b1 z = c1 , the
equation a1 (c−by)/a+b1z = c1 is equal to the composition of cyx and cxz . Therefore,
the elimination method proposed in this paper can be regarded as a generalization
of Gaussian elimination for linear equations to functional constraints defined over
discrete domains.
To see the further similarity among these algorithms, let us examine the impact of
the variable elimination ordering on their efficiency. As shown in the earlier sections,
if all the constraints are known a priori, we can find a good ordering to make
the elimination algorithm more efficient. The same principle applies to unification
algorithm. Consider the set of term equations x = f (a, a), y = f (x, x), z = f (y, y).
Direct substitution using the ordering of x, y, z is more expensive than the ordering
z, y, x. In fact, given a term equation, it can be unified in linear time by finding a
good variable ordering (Paterson and Wegman 1978).
In a CLP solver, the constraints are added to the constraint store dynamically.
If the newly added constraint is a binary linear equation, it has been observed that
one can improve the efficiency by choosing properly a variable to eliminate from
the two involved in the equation (Burg et al. 1995). For example, a brand new
variable is preferred to an old one (a variable occurring previously in the constraint
store). With the new variable, no substitution is necessary. For elimination using
bi-functional constraints, the one involved in a lesser number of earlier constraints
will be eliminated (Zhang and Yap 2002). In the case of unification, when there
are several variables that can be eliminated, we choose the one that is involved in
less number of constraints too (for example, (Escala-Imas and Ghallab 1988)). The
variable selection idea, together with the disjoint set data structure and union-find
algorithm, had led to almost linear algorithms in (Escala-Imas and Ghallab 1988;
Zhang and Yap 2002).
7.3 Functional Constraints and Variable Elimination in CSP
We now discuss other work related to functional constraints from a CSP perspective.
Bi-functional constraints, a special case of functional constraints, have been studied in the context of arc consistency (AC) algorithms since Van Hentenryck et al.
(Van Hentenryck et al. 1992) proposed A worst case optimal AC algorithm with
O(ed) was proposed in (Van Hentenryck et al. 1992). (In many of the papers, bifunctional constraints were called functional constraints). The special properties of
bi-functional constraints were used to obtain the time complexity better than that
of the optimal AC algorithms such as AC2001/3.1 (O(ed2 )) (Bessiere et al. 2005)
for arbitrary binary constraints. A fast AC algorithm for a special class of increasing bi-functional constraints was also proposed in (Liu 1995). Here, our elimination
algorithm solves the consistency of functional constraints and variable substitution
incorporates the effect of the functional part of the problem into the rest of the
non-functional constraints. Thus, it gives a higher level of consistency as it achieves
global consistency for the functional constraints rather than local consistency like
arc consistency. At the same time, it may simplify the remainder of the constraints,
thus reducing the problem further.
24
Y. Zhang and R. H. C. Yap
A new type of consistency, label-arc consistency, was introduced in (Affane and Bennaceur 1996)
and they showed that bi-functional constraints with limited extensions to other
constraints can be (globally) solved, but no detailed analysis of their algorithms is
given. In (Zhang et al. 1999), we proposed a variable elimination method to solve
bi-functional constraints in O(ed). Bi-functional constraints also belong to the class
of “Zero/One/All” constraints which was shown to be one of the earliest classes
of tractable constraints (Cooper et al. 1994). The subclass of “One” constraints in
the “Zero/One/All” class corresponds to bi-functional constraints. What was not
realized in (Cooper et al. 1994; Zhang et al. 1999) was that because the concern
was the tractability of the class of Zero/One/All constraints, the importance of
variable substitution and class of functional constraints was missed. We also point
out that all the papers above deal with the special case of bi-functional constraints
rather than functional constraints.
David introduced pivot consistency for binary functional constraints in (David 1995).
Both pivot consistency and variable substitution are different ways of reducing a
CSP into a special form. However, there are some important differences between
pivot consistency and variable substitution in this paper. Firstly, the concept of
pivot consistency, a special type of directional path consistency, is quite complex.
It is defined in terms of a variable ordering, path (of length 2) consistency, and
concepts in directed graphs. As we show in this paper, Variable substitution is a
much simpler concept. It is intuitive and simple for binary CSPs, and it extends
also simply and naturally to non-binary CSPs. Secondly, by the definition of pivot
consistency, to make a CSP pivot consistent, there must be a certain functional
constraint on each of the non-root variables. Variable substitution is more flexible.
It can be applied whenever there is a functional constraint in a problem. Finally,
to reduce a problem, the variable elimination algorithm takes O(ed2 ) while pivot
consistency algorithm takes O((n2 − r2 )d2 ), where r is the number of root variables.
Another related approach is bucket elimination (Dechter 1999). The idea in common behind bucket elimination and variable substitution is to exclude the impact
of a variable on the whole problem. The difference between them lies in the way
variable elimination is performed. In each elimination step, substitution does not
increase the arity of the constraints while bucket elimination could generate constraints with higher arity (possibly with exponential space complexity). The former
may generate more constraints than the latter, but it will not increase the total
number of constraints in the problem.
Another methodologically related work is bucket elimination (Dechter 1999). The
common idea behind bucket elimination and variable substitution is to exclude the
impact of a variable on the whole problem. However, they also differ in many aspects. Bucket elimination deals with general constraints while variable substitution
is applicable only to functional constraints. Bucket elimination assumes a variable
ordering and eliminates the impact of a variable j on all relevant constraints that
involve variables before j. In contrast, variable substitution can be used to eliminate the impact of a variable on any number of relevant constraints. The ways to
eliminate a variable are different for the two methods. For example, consider the
CSP shown in Figure 8(a) where c12 , c23 , c34 , and c45 are functional.
Solving
Solving Functional
Functional Constraints
Constraints by
by Variable
Variable Substitution
Substitution
x
x..1........
x1
x
x
x2
x
x3
x3
x4
x4
x4
x5
x5
x5
x5
(a)
(b)
(c)
(d)
... ...1
...... ..........
.....
...
....
...
...
...
.
.
...
...
..... 2
...
.... .....
... ...
..
...
... ....
..
.... ...
.
3 ....
... ...
...
...
..
.....
....
...
..... .........
...
...
4
...
....
.....
.....
....
x
x
x
x
........
... ..1
.....
.......
...
.....
...
....
...
...
.
.. 2
.
.
.
..
....
.
..
....
...
.
..
.
... ...
.
.
...
.
... ...
.
.
.
.
.
.....
... ...
... ...
3
... ...
..... ...
............
..........
.....
...
...
...
2 .....
..
.
.
.
....
..........
Fig. 8. A CSP with variable ordering
Fig. 8. A CSP with
variable ordering x1 , x2 , · · · , x5 .
25
···,x
Assume there is a variable ordering , x · · · , x . The variables will be eliminated
Assume there is a variable ordering x1 , x2 , · · · , x5 . The variables will be eliminated
in the reverse of the variable ordering. When eliminating a variable, say variable
in the reverse of the variable ordering. When eliminating a variable, say variable x4 ,
bucket elimination considers constraints only involving variables before (including)
bucket elimination considers constraints only involving variables before (including)
and ignores other constraints. In this example, 45 is ignored while constraints
x4 and ignores other constraints. In this example, c45 is ignored while constraints
,
, the new ternary constraint
14 c24 , c34 are considered relevant. After eliminating
c14 , c24 , c34 are considered relevant. After eliminating x4 , the new ternary constraint
on , x , x is added to the problem.
c{1,2,3} on x1 , x2 , x3 is added to the problem.
In the variable substitution method, for variable
and the constraint 34 funcIn the variable substitution method, for variable x4 and the constraint c34 functional on it, we can choose to substitute
in one or some of the constraints 41 , c42
tional on it, we can choose to substitute x4 in one or some of the constraints c41 , c42
and 45 , depending on a specific setting (for example, a static CSP or incremental
and c45 , depending on a specific setting (for example, a static CSP or incremental
CSP). If we choose to substitute
in all these constraints, new binary constraints
CSP). If we choose to substitute x4 in all these constraints, new binary constraints
31 , c32 , c35 are added and old constraints 41 , c42 and 45 are discarded.
c31 , c32 , c35 are added and old constraints c41 , c42 and c45 are discarded.
This example shows that in each elimination step, bucket elimination generates
This example shows that in each elimination step, bucket elimination generates
constraints with higher arity than variable substitution while the latter generates
constraints with higher arity than variable substitution while the latter generates
more constraints than the former. However, the variable substitution method will
more constraints than the former. However, the variable substitution method will
not increase the total number of constraints in the problem (because every time
not increase the total number of constraints in the problem (because every time
a new constraint is added, an old one is discarded). In the case of bi-functional
a new constraint is added, an old one is discarded). In the case of bi-functional
constraints, it decreases the total number of constraints to after all variables are
constraints, it decreases the total number of constraints to n after all variables are
eliminated.
eliminated.
8 Conclusion
8 Conclusion
We have introduced a variable substitution method to reduce a problem with both
We
have introduced
a variable substitution
to reduce
a problem
both
functional
and non-functional
constraints. method
Compared
with the
previouswith
work
on
functional
and
non-functional
constraints.
Compared
with
the
previous
work
on
bi-functional and functional constraints, the new method is not only conceptually
bi-functional
and functional
new method
is not only
conceptually
simple and intuitive
but alsoconstraints,
reflects thethe
fundamental
property
of functional
consimple
and
intuitive
but
also
reflects
the
fundamental
property
of
functional
constraints.
straints.
For a binary CSP with both functional and non-functional constraints, an alFor a binary
CSP with
both functional
non-functional
constraints,
gorithm
is presented
to transform
it into aand
canonical
functional
form in anedal-).
gorithm
is
presented
to
transform
it
into
a
canonical
functional
form
O(ed2 ).
This leads to a substantial simplification of the CSP with respect to theinfunctional
This
leads to In
a substantial
the CSP
with respect
to the the
functional
constraints.
some cases,simplification
as one of ourofresults
(Corollary
2) shows,
CSP is
constraints.
In
some
cases,
as
one
of
our
results
(Corollary
2)
shows,
the
is
already solved. Otherwise, the canonical form can be solved by ignoring theCSP
elimialready
solved.
Otherwise,
the
canonical
form
can
be
solved
by
ignoring
the
eliminated variables. For example, this means that search only needs to solve a smaller
nated
variables.
Forone
example,
this means
that search
needs to solve a smaller
problem
than the
before variable
substitution
(oronly
elimination).
problem
than the one
before
Our experiments
show
thatvariable
variablesubstitution
elimination(or
canelimination).
significantly (in some cases
Our
experiments
show
that
variable
elimination
can
significantly
(in some
cases
up to several orders of magnitude) improve the performance
of a general
solver
in
26
Y. Zhang and R. H. C. Yap
up to several orders of magnitude) improve the performance of a general solver in
dealing with functional constraints. Our experiments also show some evidence that
although rarely, the elimination could slow down the general solver in a non trivial
way.
Acknowledgments
We thank Chendong Li for helping implement the elimination algorithm and carrying out some experiments in the earlier stage of this research, and Satyanarayana
Marisetti for writing the code for generating random functional constraints and the
functional elimination ordering. Portions of this work was supported by National
Univ. of Singapore, grant 252-000-303-112.
References
Affane, M. S. and Bennaceur, H. 1996. A labelling arc consistency method for functional constraints. In Proceedings of the Second International Conference on Principles
and Practice of Constraint Programming. Lecture Notes in Computer Science, vol. 1118.
Springer, 16–30.
Bessiere, C., Regin, J., Yap, R. H. C., and Zhang, Y. 2005. An optimal coarse-grained
arc consistency algorithm. Artificial Intelligence 165, 2, 165–185.
Burg, J., Stuckey, P. J., Tai, J., and Yap, R. H. C. 1995. Linear equation solving
for constraint logic programming. In Proceedings of the 12th International Conference
on Logic Programming. MIT Press, 33–47.
Cooper, M. C., Cohen, D. A., and Jeavons, P. G. 1994. Characterizing tractable
constraints. Artificial Intelligence 65, 2, 347–361.
David, P. 1995. Using pivot consistency to decompose and solve functional CSPs. Journal
of Artificial Intelligence Research 2, 447–474.
Dechter, R. 1999. Bucket elimination: A unifying framework for reasoning. Artificial
Intelligence 113, 1-2, 41–85.
Escala-Imas, G. and Ghallab, M. 1988. A practically efficient and almost linear
unification algorithm. Artificial Intelligence 36, 2, 249–263.
Jaffar, J. and Maher, M. J. 1994. Constraint Logic Programming. Journal of Logic
Programming 19/20, 503–581.
Jaffar, J., Michaylov, S., Stuckey, P. J., and Yap, R. H. C. 1992. The CLP(R)
language and system. ACM Transactions on Programming Languages and Systems 14, 3,
339–395.
Kirousis, L. M. 1993. Fast parallel constraint satisfaction. Artificial Intelligence 64, 1,
147–160.
Liu, B. 1995. Increasing functional constraints need to be checked only once. In Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence. Morgan
Kaufmann, 119–125.
Paterson, M. S. and Wegman, M. N. 1978. Linear unification. Journal of Computer
and System Sciences 16, 2, 158–167.
Schulte, C. and Tack, G. 2005. Views and iterators for generic constraint implementations. In Recent Advances in Constraints, Joint ERCIM/CoLogNET International
Workshop on Constraint Solving and Constraint Logic Programming. Lecture Notes in
Computer Science, vol. 3978. Springer, 118–132.
Solving Functional Constraints by Variable Substitution
27
Stallman, R. M. and Sussman, G. J. 1977. Forward reasoning and dependency-directed
backtracking in a system for computer-aided circuit analysis. Artificial Intelligence 9, 2,
135–196.
Van Hentenryck, P., Deville, Y., and Teng, C. M. 1992. A generic arc-consistency
algorithm and its specializations. Artificial Intelligence 57, 2-3, 291–321.
Zhang, Y. and Yap, R. H. C. 2002. Incrementally solving functional constraints. In Proceedings of the Eighteenth National Conference on Artificial Intelligence. AAAI press,
973–974.
Zhang, Y., Yap, R. H. C., and Jaffar, J. 1999. Functional elimination and 0/1/all constraints. In Proceedings of the Sixteenth National Conference on Artificial Intelligence.
AAAI Press, 275–281.
Zhang, Y., Yap, R. H. C., Li, C., and Marisetti, S. 2008. Efficient algorithms for
functional constraints. In Proceedings of the 24th International Conference on Logic
Programming. Lecture Notes in Computer Science, vol. 5366. Springer, 606–620.
| 6cs.PL
|
DISSIPATIVITY OF SYSTEM ABSTRACTIONS OBTAINED USING
APPROXIMATE INPUT-OUTPUT SIMULATION
arXiv:1711.07529v1 [cs.SY] 20 Nov 2017
ETIKA AGARWAL∗ , SHRAVAN SAJJA† , PANOS J. ANTSAKLIS∗ , AND VIJAY GUPTA
∗
Abstract. This work focuses on the invariance of important properties between continuous
and discrete models of systems which can be useful in the control design of large-scale systems and
their software implementations. In particular, this paper discusses the relationships between the
QSR dissipativity of a continuous state dynamical system and of its abstractions obtained through
approximate input-output simulation relations. First, conditions to guarantee the dissipativity of
the continuous system from its abstractions are provided. The reverse problem of determining the
Q, S and R dissipativity matrices of the abstract system from that of the continuous system is also
considered. Results characterizing the change in the dissipativity matrices are provided when the
system abstraction is obtained. Since, under certain conditions, QSR dissipative systems are known
to be stable, the results of this paper can be used to construct stable system abstractions as well. In
the second part of this paper, we analyze the dissipativity of the approximate feedback composition
of a continuous dynamical system and a discrete controller. We present illustrative examples to
demonstrate the results of this paper.
Key words. Dissipativity, passivity, abstraction, simulation
1. Introduction. Discrete event and hybrid system models for continuous systems are quite common. For example, such models are useful for sampled and quantized systems, and also in software implementations of continuous systems. Furthermore, such discrete models can be very useful in the control design of large dynamical
systems, especially when there are temporal logic performance specifications and verification of safety requirements, e.g., in robotic systems. This is the main motivation
behind studying system properties of interest that are present in both continuous and
discrete models of the same system.
These discrete models can be obtained using abstraction based approaches. See
for example, [1]-[7]. An abstracted system model approximates a continuous state
dynamical system by a system with a pre-order or equivalence relation between the
two systems. Control design of dynamical systems using abstraction based approaches
can be carried out efficiently based on two main factors [8], first, the possibility
of constructing symbolic or purely discrete abstractions of the original system and
second, the possibility to infer the behavior of the given continuous system based upon
its discrete abstraction. The idea of using discrete abstractions for control design is
motivated by the fact that control of a continuous system with a discrete controller
requires the use of a continuous to discrete and discrete to continuous conversion. This
set-up can be viewed as an interconnection of a continuous system with a software
system, as shown in Figure 1.
Dissipativity based approaches provide attractive alternatives to control analysis
and design of such systems. These dissipativity approaches have been used for systems
with control performance described in terms of stability and optimality requirements
[9][10] for continuous systems. Dissipativity is an energy based input-output property
of dynamical systems [11]. A special form of dissipativity is QSR dissipativity which
allows transparent relationships with several important concepts such as passivity
and L2 stability [12]-[14]. Further, QSR dissipativity is preserved over feedback and
parallel interconnections; and series interconnections under certain conditions [15].
∗ Department of Electrical Engineering, University of Notre Dame, Notre Dame, IN ([email protected], [email protected], [email protected])
† IBM Research, Bangalore, India ([email protected])
1
Hence, an added advantage of using control analysis and design techniques based on
QSR dissipativity is that they scale well.
More recently, [16][17] used dissipativity like concepts for compositional analysis of
interconnected systems with safety and temporal logic specifications. This opens up a
new application area for dissipativity theory where controllers can be designed to meet
temporal logic constraints (such as safety and reachability constraints) together with
traditional specifications such as passivity, stability and optimality. As a first step
towards this goal, in this paper, we analyze the dissipativity of system abstractions
and find relationships between the QSR dissipativity, and hence passivity of systems
and their approximately input-output similar abstractions [18][19].
Software
Symbol
Symbol
Continuous to discrete
Discrete to continuous
Physical System
Control input
Measurement
Discrete abstraction
Fig. 1. Finite state approximation of a continuous-time plant interacting with a finite state
controller (software).
There are several approaches to system abstraction. Different techniques preserve different properties of the system such as reachability [3] and compositionality
[4] while maintaining equivalence or pre-order in a certain sense between the two systems. One of the popular approaches for abstraction is using the notions of simulation,
bisimulation and their approximate versions [3][1] which have been used for both discrete and continuous time systems [1][5][6][7]. While these notions are composition
preserving in some sense [1], approximate simulation and bisimulation are not preserved for more general input-output interconnections such as cascade and feedback.
Tazaki et al [18] proposed a modification of the definitions of approximate simulation and bisimulation which, under mild conditions, are interconnection preserving.
This means that the abstraction of interconnected system is the interconnection of
system abstractions. However, [20] identified that stability is not preserved for simulation (or bisimulation) and additional continuity constraints need to be imposed on
the simulation relation. Prabhakar et al [21][22] also proposed some continuity constraints on simulation relation to preserve variants of stability during abstraction. An
alternate approach is offered in the present paper using dissipativity. In this work,
we are interested in inferring QSR dissipativity properties of discrete abstractions
from that of the corresponding continuous system. These discrete abstractions are
obtained using a slight modification of method in [7] and are related to the continuous
system through approximate input-output simulation relation. The definition of approximate input-output simulation relation used in this paper is a special case of the
interconnection-compatible approximate simulation relation in [18]. QSR dissipative
systems, under mild additional conditions, are stable. This allows us to obtain stable
and composable system abstractions without imposing any strict constraints on the
simulation relation. Related work on the passivity of bisimilar systems was done in
2
[23], where the problem of discretization of a continuous controller while preserving
the passivity of closed loop system was considered. Note that while [23] discusses the
passivity indices of approximate bisimulation of the controller, in the present paper,
we provide the relationship between the more general QSR dissipativity of systems
associated through approximate input-output simulation relations. We additionally
discuss the QSR dissipativity of an approximate feedback composition [25] of systems.
Moreover, as opposed to [23], we do not restrict the systems to be incrementally stable.
Some parts of our work are contained in [24], but have not been published previously.
The main contribution of this paper is to provide a relationship between the
QSR dissipativity of a system and its approximately input-output similar abstraction.
Briefly, the results of this paper enable us to determine the Q, S and R dissipativity
matrices of an abstract system by analyzing the dissipativity of the original continuous system and vice-versa. As a special case, these results also relate the passivity
levels and stability of two systems associated through an approximate input-output
simulation relation. We limit ourselves to the class of systems and abstract models
which allow for a definition of inner product between inputs and outputs of system.
An example of this is the abstraction used in [7]. This allows us to apply the standard
dissipativity definitions to abstractions of systems as well.
The concept of approximate input-output simulation relation is defined for both
the continuous and discrete system abstractions. It is possible to obtain a continuous
as well as discrete abstraction of the same dynamical system. The first result of this
paper, provides conditions under which the QSR dissipativity of an approximately
input-output similar continuous (continuous state) abstraction, implies the dissipativity of the original continuous dynamical system. We further consider the reverse
problem of characterizing the dissipativity of a discrete state abstraction when the
given continuous system is QSR dissipative. Two different cases are presented, (i)
when the system state is the measured output, and (ii) when system state is not measured, instead measured output is same as the system output. The main difference
in these two cases arises from the fact that the approximate input-output simulation
relation between a system and its abstraction holds with respect to measured output.
In both of these cases, the class of systems are incrementally forward complete and
the abstraction is obtained using the approach in [7]. Since both the approximate
input-output simulation and QSR dissipativity are composition preserving, composition of QSR dissipativities of abstractions is the same as the QSR dissipativity of the
abstraction of composition.
In the second part of this paper we analyze the QSR dissipativity of composition
of continuous state systems with discrete state systems, say software. Specifically, we
use the approximate feedback composition defined in [25]. We show that once two
systems are approximately feedback composable, then QSR dissipativity of even one
of those systems implies QSR dissipativity of the composition.
This paper is organized as follows. Section 2 describes the system model and
some important dissipativity notions and system relations. Section 3 addresses the
relationship between QSR dissipativity of approximately input-output similar systems. We also present results to quantify the dissipativity matrices of approximately
input-output similar systems for particular abstraction methods. Section 4 discusses
examples to illustrate the results. In Section 5, some results on the dissipativity of
approximate feedback composition of finite transition systems are provided. Finally,
Section 6 contains concluding remarks.
3
Table 1
Notation
Notation
k·k
k · k2
Z+
0
R
R+
R+
0
Rn
#»
0
P >Q
Meaning
infinity norm
Euclidean norm (or induced 2-norm if argument is a matrix)
set of nonnegative integers
set of real numbers
set of positive real numbers
set of nonnegative real numbers
Euclidean space of dimension n
zero vector of appropriate dimension
matrix (P − Q) is positive definite
2. Preliminaries. In this section we briefly explain few important notions of abstract systems and introduce the properties of dissipativity, passivity and incremental
forward completeness.
2.1. System Description. The system model we use in this work is that of
transition systems. The following definitions are standard and can be found in [7].
Definition 1. A transition system T = (X, U, −→, Ym , H) consists of:
• a set of states X;
• a set of inputs U;
• a transition map −→: X × U → X;
• a set of outputs Y ;
• and an output map H : X × U → Y .
If for any state x ∈ X and u ∈ U there exists at most one state x0 ∈ X such that
u
x −−−→ x0 then the system is deterministic. x0 is also known as the u-successor of x.
u
If the system is nondeterministic, then for a transition x −−−→ x0 the state x0 may
not be unique. In such a case x0 belongs to a set of all possible u-successors given by
P ostu (x) and we will use U (x) to denote the set of inputs u ∈ U for which P ostu (x)
is nonempty. Suppose the transition system is equipped with metrics dX : X × X −→
+
+
R+
0 , dU : U × U −→ R0 and dY : Y × Y −→ R0 representing the “distance” between
two elements of state space, input space and output space respectively. This transition
system is referred to as a metric transition system. Throughout this work, we
assume that the transition system allows for a notion of inner product between inputs
and outputs and the distance metric dX , dU and dY are infinity norms k · k.
Transition systems can be used to describe a large class of dynamical systems.
We restrict ourselves to continuous time dynamical systems of the form
(1)
Σ = (X, U, Ym , f, hm )
#»
where X = Rn is the state space; U ⊆ Rm : { 0 } ∈ U is the input space; Ym ⊆ Rp :
#»
{ 0 } ∈ Ym is the measured output space; f : X ×U → X is a Lipschitz continuous map
describing state transition and hm : X × U → Ym is the measured output map. At at
any time t ∈ R+
0 , the state, input and measured output of Σ are x(t) ∈ X, u(t) ∈ U ,
ym (t) ∈ Ym and the state and measured output evolve as ẋ(t) = f (x(t), u(t)) and
ym (t) = hm (x(t), u(t)). If ξ :]a, b[−−→ X is a solution of the differential equation
ẋ(t) = f (x(t), u(t)), then we will use ξ(t, x, u) to denote a unique point reached at
4
time t under the input signal u : [0, t] → U from an initial condition x ∈ X. The
transition system associated with Σ is then given by T (Σ) = (X, U, −→, Ym , hm )
where the state transition map is dictated by the differential equation ẋ = f (x, u).
We use notation T (Σ) and T interchangeably in this work. Also, for ease of notation,
we will often drop the time index when referring to the state, input and output of Σ.
For dissipativity analysis, we consider a separate system output dictated by system output space Y ⊆ Rm and output map h(x(t), u(t)) : X × U → Y . The measured
output hm (x(t), u(t)) can be different from system output h(x(t), u(t)). In this work,
we consider two different cases (i) when measured output is the system states, and
(ii) when measured output is same as the system output. The measured output space
Ym and output map hm take values accordingly.
We can also define a discrete time system Σd = (Xd , Ud , Ymd , fd , hmd ) where
Xd , Ud , Ymd , fd and hmd are state space, input space, measured output (or measurement) space, state transition map and measured output transition map respectively.
At any discrete time k the system state xd (k) ∈ Xd , input ud (k) ∈ Ud and output ymd (k) ∈ Ymd evolve in discrete time steps as xd (k + 1) = fd (xd (k), ud (k)) and
ymd (k) = hmd (xd (k), ud (k)) for all k ∈ Z+
0 . Similar to continuous time case, system
output (used for dissipativity) dictated by system output space Yd ⊆ Rm and output
map hd (xd (k), ud (k)) : Xd × Ud → Yd , can be different from measured output.
The following assumptions on system behavior are useful in deriving the main
results of this paper.
Assumption 2. (Incremental forward completeness [7]) The dynamical system
Σ is said to be incrementally forward complete if there exist continuous functions
+
+
+
+
+
α1 : R+
0 × R0 → R0 , and α2 : R0 × R0 → R0 , α1 (·, t), α2 (·, t) ∈ K∞ for every
t ≥ 0, such that for any two initial conditions x1 , x2 ∈ X, any input trajectories
v1 , v2 : [0, t] ∈ U , and for any t ∈ R+
0 the following bound holds:
(2)
kξ(t, x1 , v1 ) − ξ(t, x2 , v2 )k ≤ α1 (kx1 − x2 k, t) + α2 (kv1 − v2 k, t)
where ξ(t, xi , vi ) is the state trajectory of system with input vi and initial state xi .
It should be noted that incremental forward completeness is a weaker condition
than incremental stability for it does not require the system to be stable. It only
states that the distance between any two state trajectories is bounded.
Assumption 3. Assume that the operator from input u(t) to rate of change of
system output ẏ(t) has the finite L2 gain γ, that is
Z τ
Z τ
2
2
kẏ(t)k2 dt ≤ γ
ku(t)k22 dt
0
0
for any τ ≥ 0 and admissible input u(t).
Assumption 3 is an L2 gain condition which bounds the rate at which the output
y can change with respect to time.
We define transition systems Tτ (Σ) obtained after sampling Σ, and Tτ,µ,η (Σ)
obtained after appropriate sampling and quantization of Σ as follows.
Definition 4. [7] Let Σ be a dynamical system and the associated transition
uτ
system be T (Σ). For any τ > 0, the sampled transition system Tτ (Σ):=(Xτ , Uτ , −−→
,
τ
Ymτ , Hmτ ) is defined by:
• Xτ = X;
• Uτ = U ;
5
u
• xτ −−−τ−→ x0τ , if there exists a trajectory ξ : [0, τ ] −−−−→ ξ(τ, xτ , u) = x0τ
τ
where u : [0, τ ) → uτ , uτ ∈ Uτ ;
• Ymτ = Ym ;
• Hmτ (xτ , uτ ) = hm (xτ , uτ ) where xτ ∈ Xτ , uτ ∈ Uτ .
Tτ (Σ) evolves in discrete time and can be represented as a discrete time system
(like Σd ). As discussed before, for dissipativity analysis, we consider another output
(referred to as system output in this work) associated with Tτ (Σ). At any discrete
time step k, system output is described by the output space Y ⊆ Rm and output
function h(x(k), u(k)) = y(k) ∈ Y where x(k) ∈ Xτ and u(k) ∈ Uτ .
Definition 5. For any incrementally forward complete control system Σ, with
system states as measurement, i.e., hm (x, u) = x, and parameters τ > 0, η > 0, µ > 0
and design parameters θ1 , θ2 ∈ R+ , a countable transition system can be defined as,
uq
Tτ,µ,η (Σ) := (Xq , Uq , −−−−→, Ymq , Hmq ), where:
τ
• Xq = [X]η = x ∈ X|xi = ki η, ki ∈ Z, and i = 1, 2, . . . , n ;
• Uq = [U ]µ = u ∈ U |ui = ki µ, ki ∈ Z, and i = 1, 2, . . . , m ;
uq
• xq −−−−→ x0q , if kξ(τ, xq , uq ) − x0q k ≤ α1 (θ1 , τ ) + α2 (θ2 , τ ) + η/2;
τ
• Ymq = [X]η = x ∈ X|xi = ki η, ki ∈ Z, and i = 1, 2, . . . , n ;
• Hmq (xq , uq ) = xq where xq ∈ Xq , uq ∈ Uq .
α1 and α2 are functions from the definition of incremental forward completeness in
Assumption 2.
Here, Hmq (xq , uq ) = xq indicates that measured output is same as the system
states. At any discrete time instant k ∈ Z+
0 system output (used for dissipativity
analysis) of Tτ,µ,η (Σ) here is h(xq (k), uq (k)) where xq (k) ∈ Xq and uq (k) ∈ Uq . Note
that this set up is often useful when analyzing systems with state feedback.
We also define the sampled and quantized transition system for the case when
system output is the measurement.
Definition 6. For any incrementally forward complete control system Σ, with
system output as measurement, i.e., hm (x, u) = h(x, u) and parameters τ > 0, η > 0,
µ > 0 and design parameters θ1 , θ2 ∈ R+ , a countable transition system can be defined
uq
as Tτ,µ,η (Σ) := (Xq , Uq , −−−−→, Ymq , Hmq ), where:
τ
• Xq = [X]η = x ∈ X|xi = ki η, ki ∈ Z, and i = 1, 2, . . . , n ;
• Uq = [U ]µ = u ∈ U |ui = ki µ, ki ∈ Z, and i = 1, 2, . . . , m ;
uq
• xq −−−−→ x0q , if kξ(τ, xq , uq ) − x0q k ≤ α1 (θ1 , τ ) + α2 (θ2 , τ ) + η/2;
τ
• Ymq = [Y ]µ = y ∈ Y |yi = ki µ, ki ∈ Z, and i = 1, 2, . . . , m ;
• Hmq (xq , uq ) = hq (xq , uq ) : khq (xq , uq ) − h(xq , u)k ≤ µ/2 where u : [0, τ ) →
uq , xq ∈ Xq , uq ∈ Uq .
α1 and α2 are functions from the definition of incremental forward completeness in
Assumption 2.
As compared to Definition 5, since Definition 6 has system output as measurement, system output (used for dissipativity analysis) is also quantized and is described
by the output map Hmq : Xq × Uq → Ymq in Definition 6.
The transition system Tτ,µ,η (Σ) can be countably finite or infinite depending
on the size of state and input spaces. For most practical cases, the system states
and inputs are restricted due to the physical limitations of the system leading to a
countably finite Tτ,µ,η (Σ).
6
2.2. System Relations. Abstraction is an approach to reduce the complexity of
the description of dynamical systems. One of the popular approaches for abstraction is
using the notion of approximate simulation [1]. Since dissipativity is an input-output
property, we talk about a generalized notion of approximate input-output simulation
[19].
Consider two metric transition systems T1 and T2 . The approximate input output simulation relation can be defined as follows.
Definition 7. T2 is an approximate input-output simulation of T1 with precision
(u , y ) if there exists an approximate input-output simulation relation R ⊆ X1 × X2
such that for all x1 ∈ X1 , there exists x2 ∈ X2 such that (x1 , x2 ) ∈ R and, for all
(x1 , x2 ) ∈ R:
1. for all u1 ∈ U1 (x1 ) there exists u2 ∈ U2 (x2 ) such that dU (u1 , u2 ) ≤ u and
dY (H1 (x1 , u1 ), H2 (x2 , u2 )) ≤ y ,
2. for all u1 ∈ U1 (x1 ) there exists u2 ∈ U2 (x2 ) such that dU (u1 , u2 ) ≤ u and
u
u
x1 −−−1−→ x01 in T1 implies the existence of x2 −−−2−→ x02 in T2 such that
1
2
(x01 , x02 ) ∈ R.
(u ,y )
This is denoted as T1 IOS
T2 .
We can decompose the input and output vectors into two groups, internal and
external signals as in [18] to facilitate the discussion on system interconnection. It is
easy to see that in this case, Definition 7 becomes a special case of interconnectioncompatible approximate simulation in [18] where not only the internal inputs but
also external inputs of the two interconnection-compatible approximately similar systems are required to be close enough to each other. As such, under mild conditions
(Theorem 1 in [18]), we can interconnect abstractions (approximate input-output simulation) of systems to obtain an abstraction of interconnection of systems.
Similarly, we can also define the notion of approximate input-output alternating
simulation to take into account the non-deterministic nature of system trajectories.
Definition 8. [7] T2 is an approximate input-output alternating simulation of
T1 with precision (u , y ) if there exists an approximate input-output alternating simulation relation R ⊆ X1 × X2 such that for all x1 ∈ X1 , there exists x2 ∈ X2 such
that (x1 , x2 ) ∈ R and, for all (x1 , x2 ) ∈ R:
1. for all u1 ∈ U1 (x1 ) there exists u2 ∈ U2 (x2 ) such that dU (u1 , u2 ) ≤ u and
dY (H1 (x1 , u1 ), H2 (x2 , u2 )) ≤ y ,
2. for all u1 ∈ U1 (x1 ) there exists u2 ∈ U2 (x2 ) such that dU (u1 , u2 ) ≤ u and for
every x02 ∈ P ostu2 (x2 ) there exists x01 ∈ P ostu1 (x1 ) such that (x01 , x02 ) ∈ R.
(u ,y )
This is denoted as T1 IOAS
T2 .
The two notions of alternating approximate input-output simulation and approximate
input-output simulation coincide in the special case of deterministic systems.
2.3. Dissipativity. A dynamical system is said to be dissipative if it stores and
dissipates energy but does not generate any energy of its own. The notion of energy
mentioned here is general and is captured using the energy supply rate function. As described in [11], the supply rate function is a real valued function described on the input
Rt
and output space and is locally integrable, i.e., ω : U × Y −→ R, t01 |ω(u(t), y(t)|dt <
∞ ∀ t1 , t0 ∈ R+
0.
Definition 9. [11] A continuous time system Σ with output function y(t) =
h(x(t), u(t)), is said to be dissipative with respect to the supply rate function ω(u, y),
7
if there exists a nonnegative function V : X → R+ , called the storage function, such
that for all t1 ≥ t ≥ t0 ≥ 0, x(t0 ) ∈ X and u(t) ∈ U
Z t1
ω(u(t), y(t))dt ≥ V (x(t1 )) − V (x(t0 ))
(3)
t0
holds where x(t1 ) is the state at time t1 resulting from the initial condition x(t0 ) and
input function u(·).
Definition 9 discusses dissipativity of continuous time system. Similarly, we can
define dissipativity for discrete time system (or sampled transition system) as follows.
Definition 10. [29] A discrete time system Σd with output function y(k) =
h(x(k), u(k)), is said to be dissipative with respect to the supply rate function ω(u, y),
if there exists a nonnegative function V : Xd → R+ , called the storage function, such
that for all k ≥ k0 ≥ 0, x0 ∈ Xd and u(k) ∈ Ud
k−1
X
(4)
ω(u(i), y(i)) ≥ V (x(k)) − V (x(k0 ))
i=k0
holds where x(k) is state at k resulting from the initial condition x(k0 ) and input
function u(·).
We can obtain different dissipativity structures and system properties depending
on the choice of supply rate function.
Definition 11. (QSR dissipativity and passivity)
• A continuous time system Σ is said to be QSR dissipative if it is dissipative
with respect to the supply rate ω(u, y) = y T Qy + 2y T Su + uT Ru. Dissipativity
inequality is then given as
Z t1
(5)
(y T (t)Qy(t) + 2y T (t)Su(t) + uT (t)Ru(t))dt ≥ V (x(t1 )) − V (x(t0 ))
t0
where Q, S and R are matrices of appropriate dimensions.
• Similarly, a discrete time system Σd is said to be QSR dissipative if it is
dissipative with respect to the supply rate ω(u, y) = y T Qy + 2y T Su + uT Ru.
Dissipativity inequality is then given as
(6)
k−1
X
(y T (i)Qy(i) + 2y T (i)Su(i) + uT (i)Ru(i)) ≥ V (x(k)) − V (x(k0 ))
i=k0
where Q, S and R are matrices of appropriate dimensions.
• A continuous time system Σ (or discrete time system Σd ) is said to be input
feed-forward output feedback passive if it is QSR dissipative with Q = −ρI,
S = 21 I, R = −νI with ρ, ν ≥ 0.
In this work, we refer to Q, S, and R matrices as dissipativity matrices. A special
case of QSR dissipativity is input feed-forward output feedback passivity with ν and
ρ known as the input and output passivity index. In this work, we use the terms
passive and input feed-forward output feedback passive interchangeably.
Definition 12. (Quasi-dissipativity [27]) A continuous time system Σ (or discrete time system Σd ) is said to be quasi-dissipative with respect to ω(u, y) if there
exists a constant β ≤ 0 such that it is dissipative with respect to the supply rate
ω(u, y) − β.
8
The boundedness and stability of quasi-dissipative systems is discussed in [28].
Similar to dissipativity, quasi-dissipativity also can take different forms depending
on the structure of supply rate function ω(u, y). Quasi-QSR-dissipativity and quasipassivity are two special cases that are of interest to us. They can be defined similar to
QSR dissipativity and passivity described earlier. In this work, dissipativity for systems in Definition 5 and Definition 6 is discussed in the context of quasi-dissipativity
where, the presence of β on the right hand side of (8) indicates the energy generated
due to quantization process.
Remark 13. The definition of dissipativity here is independent of system representation, i.e., if Σ is dissipative then T (Σ) is also dissipative. Also note that while
transition system T (Σ) follows dissipativity definition for continuous time system, trajectories of transition systems in Definition 4, Definition 5 and Definition 6 evolve
in discrete time and hence follow the dissipativity definition for discrete-time systems.
Moreover, [29] showed that for discrete time systems, (4) holds if and only if
(7)
ω(u(k), y(k)) ≥ V (x(k + 1)) − V (x(k))
for all k ∈ Z+
0 , u(k) ∈ U and x(k) ∈ X. Equivalent condition for quasi-dissipativity
is,
(8) ω(u(k), y(k)) ≥ V (x(k + 1)) − V (x(k)) + β ∀ k ∈ Z+
0 , u(k) ∈ Ud , x(k) ∈ Xd .
3. Dissipativity of systems and their abstractions. In this section we discuss the relationship between the dissipativity properties of a continuous system and
its abstraction. We provide two main results. First, we analyze the dissipativity of a
continuous dynamical system when its approximate input-output simulation is QSR
dissipative. Secondly, we consider the reverse problem of determining the dissipativity
of a system abstraction. We provide conditions under which QSR dissipativity of a
continuous system implies QSR dissipativity of its discrete abstraction obtained using
the approach in [7].
3.1. Dissipativity of system from its QSR dissipative abstraction. Consider two continuous time systems Σ1 and Σ2 and corresponding transition systems
T1 (Σ1 ) and T2 (Σ2 ) . Let ui , yi and xi represent respectively the input, output and
states of Ti , i ∈ {1, 2} and both these systems allow a notion of inner product between their inputs and outputs. Suppose T2 is QSR dissipative with Q2 , S2 , R2 as
the dissipation matrices. If T2 approximately simulates T1 then the following theorem
gives the conditions under which T1 is QSR dissipative.
( , )
u y
Theorem 14. If T1 (Σ1 ) IOS
T2 (Σ2 ) and T2 (Σ2 ) is QSR dissipative, then
T1 (Σ1 ) is also QSR dissipative with matrices Q1 , S1 , R1 satisfying
λ(Q1 − Q2 ) − ζ1 kQ1 k22 − ζ3 ≥ 0
(9)
λ(R1 − R2 ) − ζ2 kS1 k22 − ζ4 kR1 k22 ≥ 0
S1 = S2
where ζ1 , ζ2 , ζ3 , ζ4 ∈ R+ are arbitrary non-zero constants, Q2 , S2 , R2 are the dissipativity matrices for T2 (Σ2 ) and λ(·) represents the smallest eigen value of the matrix
in discussion.
Proof. See Appendix.
9
Although the result of Theorem 14 is derived for continuous time systems, it can
be extended to discrete time dynamical systems as well. This result is general in
the sense that it is applicable irrespective of the method which is used to obtain the
approximate input-output simulation. We can also use it to compute upper bounds
on the passivity indices of transition system T1 .
( , )
u y
Corollary 15. If T1 (Σ1 ) IOS
T2 (Σ2 ) and T2 (Σ2 ) is passive with passivity
indices ρ2 , ν2 then T1 (Σ1 ) is also passive with passivity indices ρ1 , ν1 which satisfy
the following
(10)
ρ1 (1 + ζ1 ρ1 ) ≤ ρ2 − ζ3
ν1 (1 + ζ4 ν1 ) ≤ ν2 − ζ2 ,
where ζ1 , ζ2 , ζ3 , ζ4 ∈ R+ are arbitrary non-zero constants such that ρ2 − ζ3 > 0 and
ν2 − ζ2 > 0, and ρ2 , ν2 are the passivity indices for T2 (Σ1 ).
Proof. Use Definition 11 in Theorem 14 to obtain this result.
The result in Theorem 14 states that if condition (9) is met, then the QSR dissipativity of an approximate input-output simulation of transition system T1 (Σ1 ) implies
the QSR dissipativity of T1 (Σ1 ) itself. Note that the reverse of this result is not
true in general. This can be seen from the definition of approximate input-output
simulation. For every transition in T1 (Σ1 ) there exists a corresponding approximate
transition in T2 (Σ2 ). However, T2 (Σ2 ) can be a larger system in the sense that there
might be some transitions in T2 (Σ2 ) for which there is no corresponding transition
in T1 (Σ1 ). Therefore, from Theorem 14, QSR dissipativity of T2 (Σ2 ) implies the
QSR dissipativity of T1 (Σ1 ) and not the other way around in general. In the next
part of this section, we consider this reverse problem of determining the Q, S and R
dissipativity of an approximate input-output simulation from the QSR dissipativity
of the original incrementally forward complete continuous system under a particular
abstraction technique.
3.2. Dissipativity of abstraction of the QSR dissipative system. This
section considers the problem of determining dissipativity matrices of the abstraction
of a dissipative system. Zamani et al [7] showed that the approximate simulation
of incrementally forward complete systems can be computed using time and space
quantization. We make a slight modification to this procedure by introducing an extra
design parameter to obtain finite abstractions which are approximately input-output
similar to the original system. We then quantify the change in QSR dissipativity
of the system model under such abstraction. Since in this work the abstraction is
obtained with respect to measured output of the system, it makes sense to consider
two different cases, (i) when the measured output of system is same as system states
and, (ii) when measured output is same as the system output.
3.2.1. System state as measured output. In this section we discuss the dissipativity properties of approximate input-output simulation of sampled data systems.
The particular class of systems we address here are the ones where measured and
actual output of the system are different. For this purpose, we use states as the measured output of Tτ (Σ). However, for dissipativity analysis, we use an alternate output
corresponding to y = h(x, u).
There have been several approaches to obtain approximate simulation of systems.
Most of them concentrate on a restrictive class of systems. [7] introduced a new procedure for construction of abstractions for any non-linear sampled data system which
10
are incrementally forward complete. The approach discussed in [7] provides sufficient
conditions in terms of appropriate sampling time and quantization parameters to obtain countable transition systems guaranteeing approximate (alternating) simulation.
We use this technique to construct approximate input-output (alternating) simulation for sampled data systems Tτ (Σ) and analyze the dissipativity properties of thus
obtained abstract system.
Proposition 16. Consider a control system Σ in (1) whose states are the measured output. Given any desired precision parameters y > 0, u > 0, if Σ satisfies Assumption 2 then for any {τ, θ1 , θ2 , η, µ} > 0 satisfying η/2 ≤ y ≤ θ1 and
µ/2 ≤ u ≤ θ2 , we have:
(11)
( , )
( , )
u y
u y
Tτ,µ,η (Σ) IOAS
Tτ (Σ) IOS
Tτ,µ,η (Σ)
where Tτ (Σ) and Tτ,µ,η (Σ) are defined in Definition 4 and Definition 5 respectively.
Proof. Proof follows directly from Theorem 4.1 in [7]. An outline of the proof can
be found in Appendix.
We next analyze the QSR dissipativity of approximately input-output similar
system Tτ,µ,η (Σ) and consider the case where measured output is same as system
states.
Theorem 17. Consider a dynamical system Σ in (1) whose states are the measured output. Suppose Σ satisfies Assumptions 2 and 3 and it is QSR dissipative with respect to the output function y = h(x, u) and a storage function V (·) :
V (x1 ) − V (x2 ) ≤ Lkx1 − x2 k. Let Tτ (Σ) be the transition system corresponding to
Σ with a sampling time τ . If the input and state quantization parameters µ and η
are chosen such that Tτ,µ,η (Σ) in Definition 5 is (u , y ) - approximately input-output
similar to Tτ (Σ), then Tτ,µ,η (Σ) is quasi QSR dissipative with matrices Qτ,µ,η , Sτ,µ,η ,
Rτ,µ,η satisfying,
Qτ,µ,η ≥ Q + τ kQk2 (τ γ + 1)I
(12)
Sτ,µ,η = S
Rτ,µ,η ≥ R + τ γkSk2 I + τ γkQk2 (τ 2 γ + τ + γ)I
where Q, S, and R are the dissipativity matrices for Σ.
Proof. See Appendix.
3.2.2. System output as measured output. In the last section, we provided
results for the dissipativity properties of approximate input-output simulation for
the class of systems where measured output is same as the system states. We now
extend these results to the systems where measured output is same as the actual
system output. To do this, we make an additional assumption on the system output
to be Lipschitz continuous which means that output can not change abruptly. The
difference from previous section is that system output y = h(x, u) is also sampled and
quantized here. This can be useful for design of systems with output feedback.
Proposition 18. Consider a control system Σ in (1) whose measured output is
the same as system output. Given any desired precision parameters y > 0, u > 0,
if Σ satisfies Assumption 2 and the output function is Lipschitz continuous, i.e.,
kh(x1 , u1 ) − h(x2 , u2 )k ≤ K1 kx1 − x2 k + K2 ku1 − u2 k, then for any {τ, θ1 , θ2 , η, µ} > 0
satisfying K1 η/2 + (K2 + 1)µ/2 ≤ y , η/2 ≤ θ1 and µ/2 ≤ u ≤ θ2 , we have:
(13)
( , )
( , )
u y
u y
Tτ,µ,η (Σ) IOAS
Tτ (Σ) IOS
Tτ,µ,η (Σ)
11
where Tτ (Σ) and Tτ,µ,η (Σ) are defined in Definition 4 and Definition 6 respectively.
Proof. Proof follows directly from Theorem 4.1 in [7]. An outline of the proof can
be found in Appendix.
Theorem 19. Consider a dynamical system Σ in (1) whose measured output is
the same as system output. Suppose Σ satisfies Assumptions 2 and 3 and it is QSR
dissipative system with respect to the output function y = h(x, u) and a storage function V (·) : V (x1 ) − V (x2 ) ≤ Lkx1 − x2 k. Let Tτ (Σ) be the transition system corresponding to Σ with a sampling time τ . If the input and state quantization parameters
µ and η are chosen as per Proposition 18 so that Tτ,µ,η (Σ) in Definition 6 is (u , y ) approximately input-output similar to Tτ (Σ), then Tτ,µ,η (Σ) is quasi QSR dissipative
with matrices Qτ,µ,η , Sτ,µ,η , Rτ,µ,η satisfying,
Qτ,µ,η ≥ Q + τ kQk2 (τ γ + 1)I + (kQ + τ kQk2 (τ γ + 1)Ik22 )I
(14)
Sτ,µ,η = S
√
Rτ,µ,η ≥ R + τ γkQk2 (τ 2 γ + τ + γ)I + kSk22 I + (γ mτ µ + γ 2 τ )I
where Q, S, and R are the dissipativity matrices for Σ.
Proof. See Appendix.
Remark 20.
1. The inequalities in (9) restrict only the spectral radius of
the dissipation matrices. This gives some flexibility in choosing the actual
structure of the Q and R matrices.
2. Since passivity is a special case of QSR dissipativity, results guaranteeing the
passivity levels for abstract system can be derived from Theorems 17 and 19
in a straight forward manner.
3. It should be noted that the notion of approximate input-output simulation
here, under mild conditions, is composition preserving [18] in that the interconnection of two abstractions is the same as the abstraction of interconnection of two systems. Also, passivity is preserved over feedback and parallel
interconnections [11]. Therefore, the composition of passivity of abstractions
is the same as the passivity of abstraction of composition. This is an interesting result that can be used to reason about the passivity properties of
abstractions of large scale systems.
4. Symbolic models obtained using abstractions have been used for design and
analysis of control systems [8]. However, issues like stability are in general
difficult to address using symbolic models. Theorems 17 and 19 can be used
to infer dissipativity properties of symbolic models obtained using approximate input-output simulation based abstractions. Once the system design
is complete, the dissipativity properties of symbolic model can be translated
back to that of the original system using the discrete version of results in
Theorem 14. Since, QSR dissipativity implies system stability under certain conditions, apart from offering compositionality, these results on QSR
dissipativity can be used to guarantee stability of system abstractions as well.
4. Examples. In this section, we present two illustrative examples. We describe
two simple linear time invariant (LTI) systems and discuss how the ideas in this paper
can be used to obtain a symbolic model and carry out the dissipativity analysis. For
both these examples, we consider input feed-forward output feedback passivity which
is a special case of QSR - dissipativity.
12
Example 1. Consider an LTI system Σ : ẋ = −x + u, with the measured output
described by an identity map, i.e., ym = x. For dissipativity analysis, we consider
another output function y = h(x, u) = Cx + Du = x + u. It can be verified that
this system is input feed-forward output feedback passive with respect to an output
function y = x + u and V (x) = 21 xT P x = 12 xT (0.5154)x. The passivity index are
(0.25, 0.5).
Now we construct an approximately input-output similar symbolic model for Σ.
Based on the discussion on incremental forward completeness of linear systems in
Section V of [7], it is readily seen that Σ is incrementally forward complete, thus
we can apply Proposition 16. We work on the subset X = [−0.2, 0.2] of the state
space and subset U = [−0.1, 0.1] of the input space. To construct the symbolic model
Tτ,µ,η (Σ) of precision u = 0.1, y = 1, choose θ1 = 1, η = 0.1, θ2 = u = µ = 0.1
and τ = 0.2 so that the conditions in Proposition 16 are satisfied. Since µ = 0.1 and
τ = 0.2, the control input is piecewise constant signal of duration τ such that
{−µ, 0, µ} = {u−1 , u0 , u1 } = {−0.1, 0, 0.1} ∈ Uq ,
and the states of the symbolic system are described by
{−2η, −η, 0, η, 2η} = {−0.2, −0.1, 0, 0.1, 0.2} ∈ Xq .
u−1 , u0 , u1
u−1 , u0 , u1
u−1 , u0 , u1
u−1 , u0 , u1
u−1 , u0 , u1
η
2η
u1
−2η
u−1 , u0 , u1
−η
u0 , u1
0
u−1 , u0
u−1 , u0 , u1
u−1
Fig. 2. Symbolic model for Σ.
The transitions between states upon the action of a control input can be calculated
using the differential equation describing Σ. As seen in Figure 2, the symbolic model
Tτ,µ,η (Σ) is non-deterministic.
The effect of symbolic abstraction on the passivity properties of Σ using Theorem 17. It can be verified that the output y = x + u satisfies Assumption 3 for
γ = 1. Hence, using Theorem 17 we can state that Tτ,η,µ (Σ) is (ρτ,µ,η , ντ,µ,η )- input
feedforward output feedback quasi passive where
(15)
ρτ,µ,η = ρ − τ kρk2 (τ γ + 1) = 0.19
(16)
ντ,µ,η = ν − τ γk0.5k2 − τ γkρk2 (τ 2 γ + τ + γ) = 0.338
For the symbolic transition system, Theorem 17 can be alternatively verified by
checking if
`T o − ρτ,µ,η oT o − ντ,µ,η `T ` ≥ V̂ (p) − V̂ (q) − β
`
is satisfied for all transitions q −−−−→ p where p ∈ Post` (q), o = Cq + D`, V̂ (q) =
τ
1 T
T
T
T T
T
and β = Lη
2τ , i.e., q F q + ` Gq + q G ` + ` H` + τ β − 2 p P p ≥ 0 where
τ
τ
T
T
T
− ρτ,µ,η τ C C, G = 2 C − ρτ,µ,η τ D C, H = 2 (D + D ) − ρτ,µ,η τ DT D −
ντ,µ,η τ I. Candidate values for V̂ (·) and β are obtained from the proof of Theorem 17.
For the symbolic system, we assume that there are M quantized inputs denoted by
{`1 , `2 , . . . , `M } and there are N quantized states denoted by {q1 , q2 , . . . , qN }. All the
1 T
2τ q P q
F = 12 P
13
`j
transitions in the symbolic system can be represented by qi −−−−→ pji for i = 1, . . . N
τ
and j = 1, . . . , M, where pji represents the next state after time τ with an initial state
qi , under the action input `j . Hence, passivity verification would entail verification
of the inequality
(17)
1
qTi F qi + `Tj Gqi + qTi GT `j + `Tj H`j + τ β − (pji )T P (pji ) ≥ 0
2
for i = 1, . . . N and j = 1, . . . , M. In order to verify the above inequality for all
T
T
transitions in a systematic fashion, welet q̄ = q1 , · · · , qN , `¯ = `1 , · · · , `M
and
1
M
1 T
M T
1
M
¯ =
arrange vectors p̄ = p1 , · · · , pN , . . . , p̄ = p1 , · · · , pN
together as p̄
1
M T
p̄ , · · · , p̄
. Verification of (17) for i = 1, . . . N and j = 1, . . . , M would require
us to verify positivity of M N scalars. All these M N scalars will be arranged along
the diagonal of an M N × M N matrix, and this diagonal matrix would be checked
for its positive definiteness. This approach allows us to represent all the inequalities
together in a compact fashion. This compact representation will be achieved using
the Kronecker product as given by
P ASSIV E = IM ⊗ ((IN ⊗ q̄T )(IN ⊗ F )(IN ⊗ q̄)) + ((IN ⊗ `¯T )(IN ⊗ G)(IN ⊗ q̄)) ⊗ IM
¯ ⊗ IM + ((IM ⊗ `¯T )(IM ⊗ H)(IM ⊗ `))
¯ ⊗ IN
+ ((IN ⊗ q̄T )(IN ⊗ GT )(IN ⊗ `))
(18)
+ IM N ⊗
Ky
¯ T (IM N ⊗ P )p̄
¯≥0
− p̄
τ
¯ is not unique, we verify (18) for all possible
For the nondeterministic cases where p̄
¯
values of p̄ . Performing this test for our numerical example, we obtain the diagonal
¯ and it can verified that
elements of the PASSIVE matrix for two possible values of p̄
all the diagonal elements are positive, hence confirming the passivity of the symbolic
model.
This example demonstrates that the results in this paper can be used to avoid
large computations for determining passivity of discrete abstractions of continuous
systems.
Example 2. In this example we use the results of Theorem 19 to validate the passivity
performance of a plant connected with a controller implemented in software. Consider
a linear time invariant system,
(19)
ẋ = Ax + Bu,
−3.6
0.2
A=
2.4
0
0
0.2
−1
0
−0.6
0
2.4
0
−6
−4
1
0
−0.6
−4
−6
0.8
y = Cx + Du
0
0
1
,
−0.8
−2
0.1
0.4
B=
0.1 ,
0.5
0.1
C = BT ,
D = 0.2 .
It can be verified that the system in (19) is passive with passivity indices (0.15, 0.7).
The L2 gain that bounds the rate of change of output y is γ = CB = 0.44. Suppose
this system is connected in feedback to a passive LTI controller
(20)
ż = Ac z + Bc w,
Ac =
−2
−3
−1
,
−5
Bc =
v = Cc z + Dc w
0.1
,
0.2
14
Cc = 1
1 ,
Dc = 1
implemented in software. For a sampling time of 0.2 sec and the state, input and
output quantization value of 0.1, the passivity indices of this controller are (ρc , νc ) =
(0.0420, 0.8115). For the ease of analysis, we focus on the subset [−0.2, 0.2], [−0.1, 0.1]
and [−0.1, 0.1] of the state, input and output space respectively. The controller
symbolic model can be obtained by considering piecewise continuous control inputs
{u−1 , u0 , u1 } = {−0.1, 0, 0.1} and the symbolic states and outputs described by
{−x2 , −x1 , x0 , x1 , x2 } = {−0.2, −0.1, 0, 0.1, 0.2} and {y−1 , y0 , y1 } = {−0.1, 0, 0.1} respectively.
As shown in Figure 1, the system (19) is interacting with the software controller
through continuous to discrete and discrete to continuous conversion units. This
system can be analyzed by considering a discrete abstraction of the continuous plant.
Using Theorem 19 for passivity,
ντ,µ,η
ρτ,µ,η = ρ − τ ρ(τ γ + 1) − |1 − ρ + τ ρ(τ γ + 1)| = −0.7653
√
= ν − 1/2 − τ γρ(τ 2 γ + τ + γ) − (γ mτ µ + γ 2 τ ) = 0.1329
Clearly, this discrete abstraction of the plant is no longer output feedback passive
(ρτ,µ,η < 0). However, following Theorem 7 in [31], since ντ,µ,η > 0, ρc > 0 and
ρτ,µ,η + νc > 0, the closed loop system is passive with output passivity index of
0.0462.
5. Dissipativity of approximate feedback composition of systems. In
this section we discuss the dissipativity property of the approximate feedback composition of two transition systems as described in [25]. We show that once two transition systems are approximately feedback composable, then QSR dissipativity of one
of those transition systems implies QSR dissipativity of the entire composition.
Cyber physical systems can be constructed by interconnecting several individual
subsystems and this process for transition systems can be described using composition operations. It was shown in [25] that approximate feedback composition of two
transition systems can also be used to construct controllers for requirements such as
safety and reachability. Approximate feedback composition of two transition systems
is possible for state feedback if there exists an approximate alternating simulation
relation between the two systems that may be a plant and a controller. The idea of
supervisory control in [25] is that the controller restricts the behavior of the plant by
forcing it to simulate the controller. The controller selects an allowable input label,
the plant makes any transition having that input label, and the controller makes a
transition to maintain alternating simulation relation. This set up works if there is
room for the controller to select its input to properly navigate the plant behavior while
maintaining the approximate alternating simulation relation. In the original concept
of approximate alternating simulation in [25], the input sets of two systems which
are approximately alternating similar were different. This freedom of nonidentical
inputs of two transition systems along with the other conditions of approximate alternating simulation defined in [25] are captured by (u , y ) approximately input-output
alternating simulation in definition 4, with u 6= 0. Therefore, we can modify the definition of approximately feedback composable systems in [25] from using approximate
alternating similar systems to (u , y ) approximate input-output alternating similar
systems. This will be clear in the following definition.
Definition 21. A transition system T2 is said to be (u , y )-approximate feedback composable with system T1 if there exists an (u , y )-approximate input-output
(u ,y )
alternating simulation relation R from T2 to T1 , that is, T2 IOAS
T1 .
15
Let Ti := (Xi , Ui , −−−−→, Yi , Hi ), i = {1, 2} be two transition systems with a
τ
common time period τ and common input and output sets equipped with euclidean
norm as the metric. Let R be a (u , y ) - approximate input-output alternating
simulation relation from T2 to T1 . Let us define a feedback relation F ⊂ X1 × X2 ×
U1 × U2 defined by all the quadruples (x1 , x2 , u1 , u2 ) ∈ X1 × X2 × U1 × U2 for which
(x1 , x2 ) ∈ R and conditions 1 and 2 in Definition 8 are met.
The feedback composition of T2 and T1 with interconnection relation F, denoted
( , )
by T2 ×F u y T1 , is the transition system (X12 , U12 , −−−−→, Y12 , H12 ) consisting of
τ
• X12 = {(x1 , x2 ) ∈ (X1 × X2 ) dY (H1 (x1 , u1 ), H2 (x2 , u2 )) ≤ y }, where
dU (u1 , u2 ) ≤ u ;
• U12 = {(u1 , u2 )|dU (u1 , u2 ) ≤ u , u1 ∈ U1 , u2 ∈ U2 };
(u1 ,u2 )
• (x1 , x2 ) −−−−−→ (x01 , x02 ) if the following three conditions hold:
τ
u
1. x1 −−−1−→ x01 in T1 ;
τ
u
2. x2 −−−2−→ x02 in T2 ;
τ
3. (x1 , x2 , u1 , u2 ) ∈ F;
• Y12 = Y1 = Y2 ;
• H12 (x1 , x2 , u1 , u2 ) = 12 (H1 (x1 , u1 ) + H2 (x2 , u2 )).
( , )
This symmetrical choice of output allows T2 ×F u y T1 to be commutative. However, we can also choose an output for the composition as H12 (x1 , x2 , u1 , u2 ) =
H1 (x1 , u1 ) or H12 (x1 , x2 , u1 , u2 ) = H2 (x2 , u2 ).
Before analyzing the dissipativity of the feedback composition, we present the
following result from [25]. Even though the results in [25] were derived for approximate
simulation relationships they also hold true for approximate input-output simulation
relationships.
Proposition 22. Consider approximate feedback composition of two (u , y ) ap(u ,y )
proximately input-output similar transition systems T1 and T2 , where T2 IOAS
T1 .
If we define the output of the composition as
1. h12 (x1 , x2 , u1 , u2 ) = 21 (h1 (x1 , u1 ) + h2 (x2 ), u2 ) then,
( , )
( , /2)
u y
(a) T2 ×F u y T1 IOS
T2 ,
(u ,y )
(u ,y /2)
(b) T2 ×F
T1 IOS
T1
2. h12 (x1 , x2 , u1 , u2 ) = h1 (x1 , u1 ), then
( ,y )
T2 ×F u
( , )
u y
T1 IOS
T2
3. h12 (x1 , x2 , u1 , u2 ) = h2 (x2 , u2 ), then
( ,y )
T2 ×F u
( , )
u y
T1 IOS
T1 .
Proof. This is a direct consequence of Proposition 11.8 in [25].
Based on Proposition 22 and Theorem 14 we obtain the following result.
Theorem 23. Let T1 :=(X1 , U1 , −−−−→, Y1 , H1 ) and T2 :=(X2 , U2 , −−−−→, Y2 , H2 )
τ
τ
be two transition systems with a common time period τ and common input and output
sets equipped with euclidean norm as the metric. Let T2 be (u , y ) - approximately
(u ,y )
input-output alternatingly similar to T1 , T2 IOAS
T1 . If T1 is QSR - dissipative with
( , )
respect to an output function h(x1 , u1 ) where x1 ∈ X1 and u1 ∈ U1 , then T2 ×F u y T1
16
is also QSR dissipative w.r.t. 12 (h1 (x1 , u1 ) + h2 (x2 , u2 )) where (x1 , x2 , u1 , u2 ) ∈ F,
and the dissipation matrices Q12 , S12 , R12 of the composed system satisfy
λ(Q12 − Q1 ) − ζ1 kQ12 k22 − ζ3 ≥ 0
λ(R12 − R1 ) − ζ2 kS12 k22 − ζ4 kR12 k22 ≥ 0
(21)
S12 = S1
where ζ1 , ζ2 , ζ3 , ζ4 ∈ R+ are arbitrary non-zero constants, Q1 , S1 , R1 are the dissipativity matrices for T1 . λ(·) and λ(·) represent the smallest and largest eigen values
of the concerned matrix.
( , )
Also, T2 ×F u y T1 is QSR-dissipative w.r.t. h2 (x2 , u2 ) where x2 ∈ X2 such that
(x1 , x2 , u1 , u2 ) ∈ X12 × U12 with the dissipation matrices Q12 , S12 , R12 as in (21).
Proof. See Appendix.
This result states that once two transition systems are approximately feedback
composable, then QSR dissipativity of one of those transition systems implies QSR
dissipativity of the composed transition system. Since passivity is a special case of
QSR dissipativity, Theorem 23 can be applied to design discrete supervisory controllers for plants, while preserving the passive nature of the interconnection.
The continuous-time system should be preceded by a sample and hold element to
convert the common quantized input symbol into a piecewise constant input. Under
this framework, the interconnected system is passive w.r.t. outputs
(i) h2 (x2 , u2 ), where x2 is a discrete state of the controller and
(ii) 21 (h1 (x1 , u1 ) + h2 (x2 , u2 )) where x1 is the discrete plant state and x2 is the
discrete controller state.
Once the interconnection is passive, we can guarantee the stability of this interconnection [11]. Further discussion on supervisory control using passivity is a subject of
our future work.
6. Conclusion. In this paper we characterized the dissipativity properties of a
system and its abstracted model. First we presented results to compute the Q, S,
and R dissipativity matrices of transition system from those of its approximate inputoutput similar abstraction. We also provided conditions determining the dissipativity
matrices of an abstract system from those of the corresponding incrementally forward
complete continuous system. Abstraction was obtained by an approximate inputoutput simulation of the continuous system. Further, we considered the approximate
feedback composition of a continuous QSR dissipative system with a finite state transition system that may be a symbolic controller. We showed that if one of the components in this interconnection is QSR dissipative, then the approximate feedback
composition is also QSR dissipative.
We can also use the results presented here to design symbolic controllers to passivate and hence stabilize a dynamical system. Although a lot of work has been done
to design continuous passivating controllers for a variety of dynamical systems, it is
challenging to design such continuous controllers when other non-traditional control
constraints, for example expressed in terms of temporal logic, need to be met along
with passivity specifications. In such cases, the importance of supervisory or discrete
controllers is more apparent. As a future work, we will concentrate on designing controllers that passivate the system along with ensuring other system specifications such
as safety and reachability.
Appendix.
17
Proof of Theorem 14.
Proof. Suppose at any time t, the input, output and state of transition system Ti (Σi ) be ui (t), yi (t) and xi (t) respectively. Define the supply rate function
ω2 (u2 (t), y2 (t)) = y2T (t)Q2 y2 (t) + 2y2T (t)S2 u2 (t) + uT2 (t)R2 u2 (t) and storage function
V (·). Since T2 is QSR dissipative with dissipativity matrices Q2 , S2 and R2 , r2 =
R t1
ω2 (u2 (t), y2 (t))dt ≥ V (x2 (t1 ))−V (x2 (t0 )). Consider another quadratic supply rate
t0
function ω1 (u1 (t), y1 (t)) = y1T (t)Q1 y1 (t) + 2y1T (t)S1 u1 (t) + uT1 (t)R1 u1 (t). Let this be
Rt
the quadratic supply rate for transition system T1 . Define r1 = t01 ω1 (u1 (t), y1 (t))dt.
T2 is (u , y ) - approximately input-output similar to T1 with an approximate
u2
−→
input-output simulation relation R. Hence we can always find a transition x2 −−−−
t1
x02
u1
in T2 for every transition x1 −−−−−→
t1
x01
in T1 such that dU (u1 , u2 ) = ku1 − u2 k =
k − ∆uk ≤ u and dY (h1 (x1 , u1 ), h2 (x2 , u2 )) = ky1 − y2 k = k − ∆yk ≤ y with
(x1 , x2 ) ∈ R and (x01 , x02 ) ∈ R. For ease of representation, we drop the time index
from input and output signals.
Z
r1 =
t1
((y2 − ∆y)T Q1 (y2 − ∆y) + 2(y2 − ∆y)T S1 (u2 − ∆u) + (u2 − ∆u)T R1 (u2 − ∆u))dt
t0
Z
t1
r1 − r2 =
(y2T (Q1 − Q2 )y2 + 2y2T (S1 − S2 )u2 + uT2 (R1 − R2 )u2 − 2y2T Q1 ∆y
t0
− 2∆y TS1 u2 − 2y2TS1 ∆u − 2uT2R1 ∆u + 2∆y TS1 ∆u + ∆uTR1 ∆u + ∆y TQ1 ∆y)dt
(22)
y
+ ζ1 kQ1 k22 hy2 , y2 it1 , 2h∆y, S1 u2 it1
ζ1
u
kS1 k22 + ζ3 hy2 , y2 it1 , 2hu2 , R1 ∆ui2
≤
ζ3
h∆y, Q1 ∆yit1
≥ λ(R1 )h∆u, ∆uit1 ,
2hy2 , Q1 ∆yit1 ≤
(23) 2hy2 , S1 ∆uit1
h∆u, R1 ∆uit1
y
+ ζ2 kS1 k22 hu2 , u2 it1
ζ2
u
≤
+ ζ4 kR1 k22 hu2 , u2 it1
ζ4
≥ λ(Q1 )h∆y, ∆yit1
≤
where λ(·) is the smallest eigen value of matrix under consideration and notation
Rt
hy, yit1 = t01 y T (t)y(t)dt.
Substituting the set of equations (23) in (22), we can show that if (9) is satisfied
then then r1 > β1 , where β1 = r2 − ζy1 − ζy2 − ζu3 kS1 k22 − ζu4 − max{0, λ(−Q1 )}y −
max{0, λ(−R1 )}u − max{0, −2h∆y, S1 ∆uit1 }. Since T2 is QSR dissipative, as per
Theorem 3.1.11 of [30], the available storage Sa2 (x2 (t0 )) =
sup −r2 (t1 ) < ∞,
u2 (·),t1 ≥t0
x2 (t0 ) = x2 ∀x2 ∈ X2 . In particular, this is true for any x2 : (x1 , x2 ) ∈ R. Clearly,
available storage Sa1 (x1 ) = sup −r1 (t1 ) < ∞, x1 (t0 ) = x1 ∀x1 ∈ X1 . Therefore,
u1 (·),t1 ≥0
from Theorem 3.1.11 of [30], T1 is QSR dissipative.
Proof of Proposition 16.
( , )
u y
Proof. We first show that Tτ (Σ) IOS
Tτ,µ,η (Σ). Let us define a relation R ⊂
Xτ × Xq such that (xτ , xq ) ∈ R if kxτ − xq k ≤ η/2. Note that for any xτ ∈ Xτ and
uτ ∈ Uτ , there always exist xq ∈ Xq and uq ∈ Uq such that kxτ − xq k ≤ η/2 ≤ y
and kuτ − uq k ≤ µ/2 ≤ u . This is possible because of the specific quantization which
allows xτ to be within η/2 radius of xq and uτ to be within µ/2 radius of uq . From
the definitions of output functions Hmτ (xτ , uτ ) = xτ and Hmq (xq , uq ) = xq , we have
kHmτ (xτ , uτ ) − Hmq (xq , uτ )k = kxτ − xq k ≤ y , hence condition (i) of Definition 7 is
satisfied.
18
u
Now if we consider the transition xτ −−−τ−→ x0τ in the transition system Tτ (Σ) ,
τ
then the distance between x0τ and ξ(τ, xq , uq ) can estimated based on the incremental
forward complete property of Σ,
kx0τ − ξ(τ, xq , uq )k ≤ α1 (η, τ ) + α2 (µ, τ ) ≤ α1 (y , τ ) + α2 (u , τ )
As mentioned earlier, due to the particular structure of quantization, for any x0τ ∈ Xτ
there always exists x0q ∈ Xq such that
kx0τ − x0q k ≤ η/2
(24)
. From the triangular inequality we have
kξ(τ, xq , uq ) − x0q k ≤ kξ(τ, xq , uq ) − x0τ k + kx0τ − x0q k
≤ α1 (y , τ ) + α2 (u , τ ) + η/2
≤ α1 (θ1 , τ ) + α2 (θ2 , τ ) + η/2
uq
which, by the Definition 5 of Tτ,µ,η (Σ) implies the existence of xq −−−→ x0q in
Tτ,µ,η (Σ). Therefore, from inequality (24) we conclude that (x0τ , x0q ) ∈ R and condi( , )
u y
tion (ii) in Definition 7 holds. Thus, Tτ (Σ) IOS
Tτ,µ,η (Σ).
(u ,y )
Along similar lines we can prove that Tτ,µ,η (Σ) IOAS
Tτ (Σ). The key is to
define R ⊆ Xq × Xτ such that (xq , xτ ) ∈ R if kxq − xτ k = 0 and notice that for every
xq ∈ Xq , we can choose xτ = xq , xτ ∈ Xτ which satisfies condition (i) of definition 4
(i.e., kxτ − xq k = 0 < y ). This is possible because Xq ⊆ Xτ .
For every uq ∈ Uq , choose uτ = uq , uτ ∈ Uτ (this satisfies kuτ − uq k = 0 < u ). Conu
sider the unique transition xτ −−−τ−→ x0τ = ξ(τ, xτ , uτ ) ∈ Postuτ (xτ ). The distance
τ
between x0τ and ξ(τ, xq , uq ) can be bounded using the incrementally forward complete
property of Σ, i.e.,
kx0τ − ξ(τ, xq , uq )k ≤ α1 (0, τ ) + α2 (0, τ )
(25)
( , )
u y
Proceeding same as the proof of Tτ (Σ) IOS
Tτ,µ,η (Σ), we can show that for every
0
0
xτ ∈ Postuτ (xτ ) there exists xq ∈ Postuq (xq ) such that (x0q , x0τ ∈ R). This is
( , )
u y
condition (ii) of Definition 8. Thus, Tτ,µ,η (Σ) IOAS
Tτ (Σ).
Proof of Theorem 17.
Proof. At any discrete time k, the state, input and output of Tτ,µ,η in Definition 5 are xq , uq and h(xq , uq ). Since the original continuous time system Σ is QSR
dissipative, inequality (5) holds for any input and all t1 > t0 > 0. Therefore, it will
be valid even if we substitute t0 = kτ, t1 = (k + 1)τ and u(t) = uq ∈ Uq ⊂ U for
kτ ≤ t ≤ (k + 1)τ where τ is the sampling time and x(t0 ) = xq ∈ Xq ⊂ X. Under
these conditions, system output of Σ is h(x(t), uq ). For simplicity of notation we
represent h(xq , uq ) as y(kτ ) and h(x(t), uq ) as y(t). Dissipativity inequality becomes,
Z
(k+1)τ
(26)
(y(t)T Qy(t) + y(t)T Suq + uTq Ruq )dt ≥ V (ξ(τ, xq , uq )) − V (xq )
kτ
19
Bounds for
R (k+1)τ
kτ
y(t)T Qy(t) dt
(k+1)τ
Z
Z
(y(t)T Qy(t) − y(kτ )T Qy(kτ )) dt =
kτ
kτ
(k+1)τ
Z
t
Z
kQT y(s)k2 kẏ(s)k2 ds dt ≤ 2
≤2
kτ
Z
(k+1)τ
Z
kτ
(k+1)τ
kτ
t
Z
d T
(y (s)Qy(s))ds dt
ds
kτ
Z
τ
kQT y(s)k2 kẏ(s)k2 ds dt
kτ
y(t)T y(t) dt + γ 2 τ uTq uq )
≤ τ kQk2 (
(27)
(k+1)τ
kτ
(k+1)τ
Z
Z
(y(t)Ty(t) − y(kτ )Ty(kτ )) dt ≤
kτ
(ky(t) − y(kτ )k22 + 2ky(kτ )k2 ky(t) − y(kτ )k2 )dt
kτ
Using ky(t) − y(kτ )k2 ≤
√
τ
qR
(k+1)τ
kτ
(k+1)τ
Z
(k+1)τ
y(t)T y(t) dt ≤ (τ 2 γ 2 + τ γ)
Z
kτ
ky(s)k22 ds ≤ τ γkuq k2 in the above equation,
(k+1)τ
uTq uq dt + (τ γ + 1)
kτ
Z
(k+1)τ
y(kτ )T y(kτ ) dt.
kτ
This can be used in (27) to compute bounds on the first term of (26),
Z
(k+1)τ
y(t)T Qy(t) dt ≤ y(kτ )T (τ Q + τ 2 kQk2 (τ γ + 1)I)y(kτ )
(28)
kτ
+ uTq (τ 3 kQk2 (τ γ 2 + γ) + τ 2 γ 2 kQk2 )Iuq
Bounds for
Z
R (k+1)τ
kτ
y(t)T Suq dt
(k+1)τ
y(t)T Suq dt −
kτ
(k+1)τ
y(kτ )T Suq dt
kτ
Z
(k+1)τ
Z
t
sZ
√
τ
kẏ(s)k22 ds ≤
kẏ(s)k2 dsdt ≤ τ τ kSuq k2
≤ kSuq k2
kτ
0
kτ
Z
(29)
Z
Z
(k+1)τ
uTq τ γkSk2 uq dt
kτ
(k+1)τ
y(t)T Suq dt ≤ uTq τ 2 γkSk2 uq + τ y(kτ )T Suq
⇒
kτ
uq
Bounds for V (ξ(τ, xq , uq )): Now we consider a transition xq −−−−→ x0q in Tτ,µ,η (Σ)
τ
and by Definition of Tτ,µ,η (Σ) we have kξ(τ, xq , uq )−x0q k ≤ α1 (θ1 , τ )+α2 (θ2 , τ )+η/2.
For Lipschitz continuous storage functions
(30)
V (ξ(τ, xq , uq )) ≥V (x0q )−L(kx0q −ξ(τ, xq , uq )k) ≥ V (x0q )−L(α1 (θ1 , τ )+α2 (θ2 , τ )+η/2)
Using (28), (11) and (30) in (26) gives the result.
Proof of Theorem Proposition 18.
( , )
u y
Proof. We first show that Tτ (Σ) IOS
Tτ,µ,η (Σ). Consider a relation R ⊂
Xτ × Xq such that (xτ , xq ) ∈ R if kxτ − xq k ≤ η/2. Note that for any xτ ∈ Xτ and
uτ ∈ Uτ , there always exist xq ∈ Xq and uq ∈ Uq such that kxτ − xq k ≤ η/2 and
kuτ − uq k ≤ µ/2 ≤ u . This is possible because of the specific quantization which
20
allows xτ to be within η/2 radius of xq and uτ to be within µ/2 radius of uq .
From the definitions of output functions we have,
khτ (xτ , uτ ) − hq (xq , uq )k = kh(xτ , uτ ) − h(xq , uq ) + h(xq , uq ) − hq (xq , uq )k
≤ K1 kxτ − xq k + K2 kuτ − uq k + kh(xq , uq ) − hq (xq , uq )k
≤ K1 η/2 + K2 µ/2 + µ/2 ≤ y
(31)
Hence condition (i) of Definition 7 is satisfied. Along the lines of proof of Proposition 16, it can be shown that condition (ii) of Definition 7 is also satisfied, proving
(u ,y )
Tτ (Σ) IOS
Tτ,µ,η (Σ).
Similar to the steps described above and proof of Proposition 16, we can show
(u ,y )
that Tτ,η,µ (Σ) IOAS
Tτ (Σ).
Proof of Theorem Theorem 19.
Proof. At any discrete time k, the state, input and output of Tτ,µ,η in Definition 6 are xq , uq and hq (xq , uq ). Since the original continuous time system Σ is QSR
dissipative, inequality (5) holds for any input and all t1 > t0 > 0. Therefore, it will
be valid even if we substitute t0 = kτ, t1 = (k + 1)τ and u(t) = uq ∈ Uq ⊂ U for
kτ ≤ t ≤ (k + 1)τ where τ is the sampling time and x(t0 ) = xq ∈ Xq ⊂ X. Under
these conditions, system output of Σ is h(x(t), uq ).
For simplicity of notation we represent h(xq , uq ) = y(kτ ), h(x(t), uq ) = y(t),
hq (xq , uq ) = ŷ(kτ ) and ∆y = y(kτ ) − ŷ(kτ ). Dissipativity inequality becomes,
Z
(k+1)τ
(y(t)T Qy(t) + y(t)T Suq + uTq Ruq )dt ≥ V (ξ(τ, xq , uq )) − V (xq )
(32)
kτ
In order to prove the dissipativity of the approximate input-output similar system
Tτ,µ,η , we find bounds on each term of equation (32).
R (k+1)τ
Bounds for kτ
y(t)T Qy(t)dt : From proof of Theorem 17
Z
(k+1)τ
y(t)T Qy(t)dt
kτ
≤ y(kτ )T (Qτ + τ 2 kQk2 (τ γ + 1)I)y(kτ ) + uTq (τ 3 kQk2 (τ γ 2 + γ) + τ 2 γ 2 kQk2 )Iuq
≤ ŷ(kτ )T (Qτ + τ 2 kQk2 (τ γ + 1)I)ŷ(kτ ) + 2ŷ(kτ )T (Qτ + τ 2 kQk2 (τ γ + 1)I)∆y
+ ∆y T (Qτ + τ 2 kQk2 (τ γ + 1)I)∆y + uTq (τ 3 kQk2 (τ γ 2 + γ) + τ 2 γ 2 kQk2 )Iuq .
(33)
2ŷ(kτ )T (Qτ + τ 2 kQk2 (τ γ + 1)I)∆y ≤ τ ∆y T ∆y + τ kQ + τ kQk2 (τ γ + 1)Ik22 ŷ(kτ )T ŷ(kτ )
µ2
+ τ kQ + τ kQk2 (τ γ + 1)Ik22 ŷ(kτ )T ŷ(kτ )
4
∆y T (Qτ + τ 2 kQk2 (τ γ + 1)I)∆y ≤ λ̄(Qτ + τ 2 kQk2 (τ γ + 1)I)∆y T ∆y
1
(35)
≤ mτ µ2 λ̄(Q + τ kQk2 (τ γ + 1)I)
4
≤ τm
(34)
where λ̄(·) is the largest eigen value of matrix under consideration. Using (34) and
(35) in (33),
Z
(k+1)τ
y(t)T Qy(t)dt ≤ ŷ(kτ )T (Q + τ kQk2 (τ γ + 1)I + (kQ + τ kQk2 (τ γ + 1)Ik22 )I)τ ŷ(kτ )
kτ
(36)
+ uTq (τ 3 kQk2 (τ γ 2 + γ) + τ 2 γ 2 kQk2 )Iuq +
21
1
mτ µ2 (1 + λ̄(Q + τ kQk2 (τ γ + 1)I)).
4
Bounds for 2
R (k+1)τ
kτ
y(t)T Suq dt
(k+1)τ
Z
(y(t)T Suq dt − ŷ(0)T Suq )dt≤
2
Z
kτ
(k+1)τ
((y(t) − ŷ(kτ ))T(y(t) − ŷ(kτ )) + (Suq )T(Suq ))dt
kτ
Z (k+1)τ
(y(t) − ŷ(kτ ))T (y(t) − ŷ(kτ ))dt + τ kSk22 uTq uq
≤
(37)
kτ
Z
(k+1)τ
(y(t) − ŷ(kτ ))T (y(t) − ŷ(kτ ))dt
kτ
(k+1)τ
Z
ky(t) − ŷ(kτ )k22 dt ≤
≤
Z
(ky(t) − y(kτ )k2 + ky(kτ ) − ŷ(kτ )k2 )2 dt
kτ
kτ
s
≤ τ(
(k+1)τ
Z
(k+1)τ
γ2
kτ
kuq k22 dt +
√
√
√
µτ (mµ + γ mτ )
mµ/2)2 ≤ (γ 2 τ 2 + γτ mτ µ)uTq uq +
.
4
Using this in (37),
(38)
√
Z (k+1)τ
√
µτ (mµ + γ mτ )
2
.
y(t)TSuq dt ≤ 2τ ŷ(kτ )TSuq + (γ 2 τ 2 + γτ mτ µ + τ kSk22 )uTq uq +
4
kτ
Similar to the proof of Theorem 14, V (ξ(τ, xq , uq )) can also be bounded. We can use
this bound on V (ξ(τ, xq , uq )), (36) and (38) to bound the terms in (32) and rearrange
the resulting equation to get (14).
Proof of Theorem 23.
Proof. Output of T1 is h1 (x1 , u1 ) and we consider two possible outputs of T12 =
( , )
T2 ×F u y T1 . From the definition of approximate feedback composition and Proposition 22, possible relations between T12 and T1 are given by
Case 1:
Case 2:
1
(h1 (x1 , u1 ) + h2 (x2 , u2 ))
2
h12 (x1 , x2 , u1 , u2 ) = h2 (x2 , u2 )
h12 (x1 , x2 , u1 , u2 ) =
( , /2)
u y
⇒ T12 IOS
T1
( , )
u y
⇒ T12 IOS
T1 .
Using Theorem 14 for both these cases gives the result.
REFERENCES
[1] A. Girard, and G. J. Pappas, “Approximation Metrics for Discrete and Continuous Systems,”
in IEEE Transactions on Automatic Control, vol. 52, no. 5, pp. 782-798, May 2007.
[2] P. Tabuada, “An Approximate Simulation Approach to Symbolic Control,” in IEEE Transactions on Automatic Control, vol. 53, no. 6, pp. 1406-1418, July 2008.
[3] R. Alur, T. Henzinger, G. Lafferriere, and G. J. Pappas, “Discrete Abstractions of Hybrid
Systems,” in Proc. IEEE, vol. 88, pp. 971-984, 2000.
[4] A. Girard, and G. J. Pappas, “Approximate Bisimulation: A Bridge Between Computer Science
and Control Theory,” in European Journal of Control, vol. 17, pp. 568-578, 2011.
[5] A. van der Schaft, “Equivalence of dynamical systems by bisimulation,” IEEE Trans. Autom.
Control, vol. 49, no. 12, pp. 2160-2172, 2004.
[6] A. Girard and G. J. Pappas, “Approximate bisimulations for nonlinear dynamical systems,” in
Proc. of the 44th IEEE Conf. on Decision and Control, pp. 684-689, 2005.
[7] M. Zamani, G. Pola, M. Mazo, and P. Tabuada, “Symbolic models for nonlinear control systems
without stability assumptions,” in IEEE Transactions on Automatic Control, vol. 57, no. 7,
pp. 1804-1809, July 2012.
22
[8] A. Girard, and G. J. Pappas, “Hierarchical control system design using approximate simulation,” in Automatica, vol. 45, no. 2, pp. 566-571, Feb. 2009.
[9] E. Agarwal, S. Sivaranjani, and P. J. Antsaklis, “Feedback passivation of nonlinear switched
systems using linear approximations,” in Indian Control Conference, pp. 12-17, Jan. 2017.
[10] M. Xia, A. Rahnama, S. Wang, and P. J. Anstaklis, “Performance optimization based on
passivation of systems with applications to systems with input/output delay,” in American
Control Conference, pp. 973-978, Aug. 2016.
[11] J. C. Willems, “Dissipative dynamical systems part I: General theory,” in Arch. Rational Mech.
Anal., vol. 45, no. 5, pp. 321-351, 1972.
[12] P. J. Antsaklis, B. Goodwine, V. Gupta, M. J. McCourt, Y. Wang, P. Wu, M. Xia, H. Yu,
and F. Zhu , “Control of cyberphysical systems using passivity and dissipativity based
methods,” in Eur. Journal of Control, vol. 19, no. 5, pp. 379-388, 2013.
[13] H. K. Khalil, Nonlinear Systems, Upper Saddle River, NJ: Prentice Hall, 2002.
[14] D. Hill and P. Moylan, “The stability of nonlinear dissipative systems,” in IEEE Tran. on Aut.
Control , vol. 21, no. 5, pp. 708-711,1976.
[15] J. Bao, and P. L. Lee, Process Control: The Passive Systems Approach, SpringerVerlag, Advances in Industrial Control, London, first edition, 2007.
[16] S. Coogan and M. Arcak, “A dissipativity approach to safety verification for interconnected
systems,” in IEEE Transactions on Automatic Control, vol. 60, pp. 1722-1727, June 2015.
[17] Eric S. Kim, Murat Arcak, and Sanjit A. Seshia, “A small gain theorem for parametric assumeguarantee contracts,” in Proc. of the 10th International Conference on Hybrid Systems:
Computation and Control, pp. 239-248, April 2017.
[18] Y. Tazaki and J. Imura,“Bisimilar finite abstractions of interconnected systems,” in 11th International Conference on Hybrid Systems: Computation and Control, pp. 514-527, 2008.
[19] Y . Tazaki and J. Imura, “Finite abstractions of discrete-time linear systems and its application
to optimal control,” in 17th IFAC World Congress, pp. 10201-10206, 2008.
[20] P. J. L. Cuijpers, “On bicontinuous bisimulation and the preservation of stability,” in Proc. of
the International Conference on Hybrid Systems: Computation and Control, pp. 676-679,
2007.
[21] P. Prabhakar, G. Dullerud and M. Viswanathan, ”Stability Preserving Simulations and Bisimulations for Hybrid Systems,” in IEEE Trans. on Automatic Control, vol. 60, no. 12, pp.
3210-3225, Dec. 2015.
[22] P. Prabhakar, G. Dullerud, and M. Viswanathan, “Pre-orders for reasoning about stability,”
in Proc. of the 15th ACM int. conf. on Hybrid Systems: Computation and Control, pp.
197-206, 2012.
[23] X. Xu, N. Ozay, and V. Gupta, “Dissipativity degradation in discrete control implementations:
an approximate bisimulation approach,” in IEEE Conference on Decision and Control, pp.
6817-6822, 2015.
[24] S. Sajja, V. Gupta and P. Antsaklis, “Passivity based supervisory control,” ISIS Technical
report available at http://www3.nd.edu/ isis/techreports/isis-2014-005.pdf.
[25] P. Tabuada, Verification and control of hybrid systems: a symbolic approach. Springer, 2009.
[26] P. Tabuada, G. J. Pappas, and P. Lima, “Composing abstractions of hybrid systems,” in Hybrid
Syst.: Comput. Control, C. Tomlin and M. R. Greenstreet, Eds. New York: SpringerVerlag, pp. 436-450, 2002.
[27] R. Lozano, B. Brogliato, O.Egeland, and B. Maschke. Dissipative Systems Analysis and Control:
Theory and Applications. Springer, London, 2nd edition, 2007.
[28] G. Polushin, and H.J. Marquez, “Boundedness properties of nonlinear quasi-dissipative systems,” in IEEE Transactions on Automatic Control, vol.49, no 12, pp. 2257-2261, Dec.
2004.
[29] C. Byrnes and W. Lin, “Losslessness, feedback equivalence, and the global stabilization of
discrete-time nonlinear systems,” in IEEE Transactions on Automatic Control, vol. 39,
no. 1, pp. 83-98, Jan. 1994.
[30] A. van der Schaft, L2 -Gain and Passivity Techniques in Nonlinear Control, 3rd edition,
Springer, London, 2017.
[31] F. Zhu, M. Xia, and P. J. Antsaklis, “Passivity analysis and passivation of feedback systems
using passivity indices,” in American Control Conference, pp. 1833-1838, 2014.
23
| 3cs.SY
|
Blind Identification of Invertible Graph Filters
with Multiple Sparse Inputs
arXiv:1803.04072v1 [eess.SP] 12 Mar 2018
Chang Ye, Rasoul Shafipour and Gonzalo Mateos
Dept. of Electrical and Computer Engineering, University of Rochester, Rochester, NY, USA
Abstract—This paper deals with problem of blind identification
of a graph filter and its sparse input signal, thus broadening the
scope of classical blind deconvolution of temporal and spatial
signals to irregular graph domains. While the observations are
bilinear functions of the unknowns, a mild requirement on
invertibility of the filter enables an efficient convex formulation,
without relying on matrix lifting that can hinder applicability
to large graphs. On top of scaling, it is argued that (non-cyclic)
permutation ambiguities may arise with some particular graphs.
Deterministic sufficient conditions under which the proposed
convex relaxation can exactly recover the unknowns are stated,
along with those guaranteeing identifiability under the BernoulliGaussian model for the inputs. Numerical tests with synthetic
and real-world networks illustrate the merits of the proposed
algorithm, as well as the benefits of leveraging multiple signals
to aid the (blind) localization of sources of diffusion.
Index Terms—Graph signal processing, network diffusion,
bilinear equations, blind deconvolution, convex optimization.
I. I NTRODUCTION
Network processes such as neural activities at different
regions of the brain [9], [10], vehicle trajectories over road
networks [4], or spatial temperature profiles measured by
a wireless sensor network [19], can be represented as signals supported on the nodes of a graph. Under the natural
assumption that the signal properties are influenced by the
graph topology (e.g., in a network diffusion or percolation
process), the goal of graph signal processing (GSP) is to
develop algorithms that exploit this relational structure. Accordingly, generalizations of fundamental signal processing
tasks have been widely explored in recent work; see [14]
for a comprehensive tutorial treatment. Notably graph filters –
which generalize classical time-invariant systems – were conceived as information-processing operators acting on graphvalued signals [18]. Mathematically, graph filters are linear
transformations that can be expressed as polynomials of the
so-termed graph-shift operator (Section II). The graph shift
offers an alegbraic representation of network structure and
can be viewed as a local diffusion operator. For the directed
cycle graph representing e.g., periodic temporal signals, it
boils down to the classical time-shift operator [18]. Given a
shift, the polynomial coefficients fully determine the graph
filter and are referred to as filter coefficients.
Problem outline and envisioned applications. In this paper,
we revisit the blind identification of graph filters with sparse
inputs, with emphasis on modeling diffusion processes and
Work in this paper was supported by the NSF award CCF-1750428. Author
emails: {cye7,rshafipo,gmateosb}@ur.rochester.edu.
localization of the sources of diffusion [21]. Specifically, given
observations of graph signals {yi }P
i=1 that we model as outputs of a diffusion filter (i.e., a polynomial in a known graphshift operator), we seek to jointly identify the filter coefficients
h and the input signals {xi }P
i=1 that gave rise to the network
observations. This inverse problem broadens the scope of
classical blind deconvolution of temporal or spatial signals to
graphs [1], [11]. Since the resulting bilinear inverse problem
is ill-posed, we assume that the inputs are sparse – a wellmotivated setting when few seeding nodes inject a signal that
is diffused throughout a network [21]. Accordingly, envisioned
application domains include environmental monitoring (where
are the heat or seismic sources?), opinion formation in social
networks (who started the rumor?), neural signal processing
(which brain regions were activated?), and epidemiology (who
is patient zero for the disease outbreak?).
Related work and contributions. Different from most existing works dealing with source localization on graphs, e.g.,
[16], [24], like [15] the advocated GSP approach is applicable
even when a single snapshot of the diffused signal is available.
Often the models of diffusion are probabilistic in nature, and
resulting maximum-likelihood source estimators can only be
optimal for particular (e.g., tree) graphs [16], or rendered scalable under restrictive dependency assumptions [5]. Relative
to [9], [15], the proposed framework can accommodate signals
defined on general undirected graphs and relies on a convex
estimator of the sparse sources of diffusion. Furthermore,
the setup where multiple output signals are observed (each
one corresponding to a different sparse input), has not been
thoroughly explored in convex-relaxation approaches to blind
deconvolution of (non-graph) signals, e.g., [1], [13]; see [22]
for a recent and inspiring alternative that we leverage here.
A noteworthy approach was put forth in [21], which casts
the (bilinear) blind graph-filter identification task as a linear
inverse problem in the “lifted” rank-one, row-sparse matrix
xhT . While the rank and sparsity minimization algorithms
in [17], [21] can successfully recover sparse inputs along
with low-order graph filters, reliance on matrix lifting can
hinder applicability to large graphs. Beyond this computational
consideration, the overarching assumption of [21] is that the
inputs {xi }P
i=1 share a common support. Here instead we show
how a mild requirement on invertibility of the graph filter
facilitates an efficient convex formulation for the multi-signal
case with arbitrary supports (Section III); see also [22] for a
time-domain precursor. In Section IV we take a closer look
at inherent scaling and (non-cyclic) permutation ambiguities
arising with some particular graphs. We also briefly comment
on identifiability under the Bernoulli-Gaussian model for the
inputs [12], and state deterministic sufficient conditions under
which the proposed convex relaxation can exactly recover
the unknowns. Numerical tests with synthetic graphs and
a structural brain network corroborate the effectiveness of
the proposed approach in recovering the sparse input signals
(Section V). Concluding remarks are given in Section VI.
II. P RELIMINARIES AND P ROBLEM S TATEMENT
Consider a weighted and undirected network graph G =
(V, A), where V is the set of vertices with cardinality |V| =
N , and A ∈ RN ×N is the symmetric graph adjacency matrix
whose entry Aij denotes the edge weight between nodes i and
j. As a more general algebraic descriptor of network structure,
one can define a graph-shift operator S ∈ RN ×N as any
matrix having the same sparsity pattern as G [18]. Accordingly,
S can be viewed as a local diffusion (or averaging) operator.
Common choices are to set it to either A (and its normalized
counterparts) or variations of adjacency and Laplacian matrices [6], [14]. Since S is real and symmetric, it is diagonalizable
so that S = VΛVT , with Λ = diag(λ1 , . . . , λN ). Lastly, a
graph signal x : V 7→ RN is an N -dimensional vector, where
entry xi represents the signal value at node i ∈ V.
A. Graph-filter models of network diffusion processes
Let y be a graph signal supported on G, which is generated
from an input graph signal x via linear network dynamics of
the form
Q∞
P∞
l
y = α0 l=1 (I − αl S)x =
(1)
l=0 βl S x.
While S encodes only one-hop interactions, each successive
application of the shift in (1) diffuses x over G. Indeed, any
process that can be understood as the linear propagation of a
seed signal through a static graph can be written in the form
in (1), and subsumes heat diffusion, consensus and the classic
DeGroot model of opinion dynamics as special cases [3].
The diffusion expressions in (1) are polynomials on S
of possibly infinite degree, yet the Cayley-Hamilton theorem asserts they are equivalent to polynomials of degree
smaller than N . Upon defining the vector of coefficients
h := [h0 , . . . , hL−1 ]T and the shift-invariant graph filter
H := h0 IN +h1 S+h2 S2 +. . .+hL−1 SL−1 =
L−1
X
hl Sl , (2)
l=0
PL−1
l
the signal model in (1) becomes y =
l=0 hl S x := Hx,
for some particular h and L ≤ N . Due to the local structure
of S, graph filters represent linear transformations that can
be implemented in a distributed fashion [20], e.g., via L − 1
successive exchanges of information among neighbors.
Leveraging the spectral decomposition of S, graph filters
and signals can be represented in the frequency domain.
Specifically, let us use the eigenvalues of S to define the
N × L Vandermonde matrix ΨL , where Ψij := λj−1
. The
i
frequency representations of a signal x and filter h are defined
as x̃ := VT x and h̃ := ΨL h, respectively. The latter follows
since the output y = Hx in the frequency domain is given by
ỹ = diag ΨL h VT x = diag h̃ x̃ = h̃ ◦ x̃.
(3)
This identity can be seen as a counterpart of the convolution
theorem for temporal signals, where ỹ is the elementwise
product (◦) of x̃ and the filter’s frequency response h̃ := ΨL h.
B. Problem formulation
For given shift operator S and filter order L, suppose
we observe P output signals collected in a matrix Y =
[y1 , . . . , yP ] ∈ RN ×P such that Y = HX, where X =
[x1 , . . . , xP ] ∈ RN ×P is sparse having at most S N
non-zero entries per column. The goal is to perform blind
identification of the graph filter (and its input signals), which
amounts to estimating sparse X and the filter coefficients
h up to scaling and (possibly) permutation ambiguities; see
Section IV. Sparsity is well motivated when the signals in Y
represent diffused versions of a few localized sources in G,
here indexed by supp(X) := {(i, j) | Xij 6= 0}. Moreover,
the non-sparse formulation is ill-posed, since the number of
unknowns N P + L in {X, h} exceeds the N P observations
in Y.
All in all, using (3) the diffused source localization task can
be stated as a feasibility problem of the form
find {X, h}s. to Y = Vdiag ΨL h VT X, kXk0 ≤ P S, (4)
where the `0 -(pseudo) norm kXk0 := |supp(X)| counts the
non-zero entries in X. In words, the goal is to find the solution
to a system of bilinear equations subject to a sparsity constraint
in X; a hard problem due to the non-convex `0 -norm as well as
the bilinear constraints. To deal with the latter, building on [22]
we will henceforth assume that the filter H is invertible.
III. C ONVEX R ELAXATION FOR I NVERTIBLE F ILTERS
Here we show how to efficiently tackle the blind graph
filter identification problem, through a convex relaxation of
(4) when the diffusion filter is invertible.
To that end, note P
from (3) that graph filter H is invertible
L−1
if and only if h̃i = l=0 hl λli 6= 0, for all i = 1, . . . , N . In
words, the frequency response of the filter should not vanish
at the graph frequencies {λi }. In such case one can show
that the inverse operator G := H−1 is also a graph filter on
G, which can be uniquely represented as a polynomial in the
shift S of degree at most N − 1 [18, Theorem 4]. To be more
specific, letPg ∈ RN be the vector of inverse-filter coefficients,
N −1
i.e., G = l=0 gl Sl . Then one can equivalently rewrite the
generative model Y = HX for the observations as
X = GY = Vdiag(g̃)VT Y,
(5)
where g̃ := ΨN g ∈ RN is the inverse filter’s frequency
response and ΨN ∈ RN ×N is Vandermonde. Naturally,
G = H−1 implies the condition g̃ ◦ h̃ = 1N on the frequency
responses, where 1N denotes the N × 1 vector of all ones.
Leveraging (5), one can recast (4) as a linear inverse problem
min kXk0 , s. to X = Vdiag(g̃)VT Y, X 6= 0.
{X,g̃}
(6)
This approach is markedly different from the matrix lifting
technique used in [21] to handle the bilinear equations in (4).
The `0 norm in (6) makes the problem NP-hard to optimize.
Over the last decade or so, convex-relaxation approaches to
Algorithm 1 Iteratively-reweighted `1 minimization for (7)
1: Input: Matrix Y T V
V, δ > 0 and > 0.
2: Initialize t = 0, w(0) = 1N P and X(0) = 0.
3: repeat
4:
Solve
g̃(t+1) = argmin w(t) ◦ (YT V V)g̃
1
g̃
s. to
5:
6:
7:
8:
9:
1TN g̃
= 1.
Form X(t+1) = (YT V V)g̃(t+1) .
(t+1)
1
, i = 1, 2, ..., N P .
Update wi
= [vec(X(t+1)
)]i +δ
t ← t + 1.
until kX(t+1) − X(t) k1 /kX(t) k1 ≤
ˆ := g̃(t+1) and X̂ := X(t+1)
return g̃
tackle sparsity minimization problems have enjoyed remarkable success, since they often entail no loss of optimality.
Accordingly, we instead: (i) seek to minimize the `1 -norm
convex
surrogate of the cardinality function, that is kXk1 =
P
|X
ij |; and (ii) express the filter in the graph spectral
i,j
domain as in (5) to obtain the cost
kXk1 = kGYk1 = kVdiag(g̃)VT Yk1 = k(YT V
V)g̃k1 ,
where denotes the Khatri-Rao (i.e., columnwise Kronecker)
product. This suggests solving the convex `1 -synthesis problem (in this case a linear program), e.g., [23], namely
b̃ = argmin k(YT V
g
g̃∈RN
V)g̃k1 ,
s. to 1TN g̃ = 1.
(7)
b̃ = 0, it also serves
While the linear constraint in (7) avoids g
to fix the scale of the solution.
As a result, under the pragmatic assumption that the diffusion filter is invertible, one can readily use e.g., an off-the-shelf
interior-point method or a specialized sparsity-minimization
algorithm to solve (7) efficiently. Different from the solvers
in [17], [21], the aforementioned algorithmic alternatives are
free of expensive singular-value decompositions per iteration.
We have found that overall performance can be improved
via the iteratively-reweighted `1 -norm minimization procedure
tabulated under Algorithm 1; see also [2] for a justification of
such refinement. In any case, notice that once the frequency
b̃ of the inverse filter is recovered, one can readily
response g
reconstruct the sources via X̂ = (YT V V)g̃ as well as the
filter H, if desired.
In the next section we will take a closer look at the inherent
ambiguities associated with the bilinear model Y = HX,
some of which are unique to the network setting dealt with
here. These are of course important to delineate the scope of
identifiability (i.e., uniqueness) results. We will complete our
discussion with deterministic sufficient conditions under which
the convex relaxation (7) is tight.
IV. I DENTIFIABILITY AND E XACT R ECOVERY
To establish further connections with blind deconvolution
of periodic discrete-time signals, recall these can be viewed
as graph signals supported on the directed cycle graph (whose
Fig. 1. Toy undirected graph (left) used to illustrate the symmetric permutation
ambiguity between nodes 2 and 4. The fourth eigenvector v4 of S = A
(center) has the problematic form. Then if {X0 , h̃0 } satisfies the bilinear
equations Y = Vdiag(h̃)VT X, so does {PX0 , diag(p)h̃0 } for the shown
permutation matrix P (right) and p = [1, 1, 1, −1, 1, 1, 1]T .
circulant adjacency matrix is diagonalized by the DFT basis) [18]. In this special case, the blind identification task is
known to suffer from unavoidable scaling and circulant-shift
ambiguities; see e.g., [1], [22]. Here we examine more general
symmetric permutation ambiguities arising with unweighted
graphs, and briefly outline a relevant identifiability result as
well as preliminary exact recovery conditions for (7).
A. Permutation ambiguities for some unweighted graphs
In solving the bilinear inverse problem formulated in Section II, for some particular graphs in addition to scaling we
may also encounter (non-cyclic shift) permutation ambiguities.
We can resolve the scaling ambiguity by e.g., a fortiori setting
kg̃0 k1 = 1 as in the experiments of Section V, where g̃0 is the
ground-truth frequency response of the inverse filter. Inspired
by the identifiability studies for sparsity-constrained bilinear
problems [12], here we examine said permutation ambiguities
for unweighted graphs with shift S = VΛVT .
Let {X0 , h̃0 } collect the ground-truth sparse input signals
and the filter’s frequency response, respectively. Let u(i,j) ∈
(i,j)
RN be a unit-norm
vector with zero entries except for ui
=
√
(i,j)
−uj
= 1/ 2. As we show next, a permutation ambiguity
arises if, say, the kth eigenvector of S (i.e., the kth column of
V) has the form u(i,j) . Indeed, in that case one could introduce
a binary signed vector p ∈ {−1, 1}N with a single negative
entry pk = −1, to construct another solution of the form
X1 := PX0 ,
h̃1 := diag(p)h̃0 ,
(8)
where P = IN − 2u(i,j) (u(i,j) )T = Vdiag(p)VT is a
symmetric permutation matrix that interchanges the signal
values at nodes i (xi ) and j (xj ) when applied to the graph
signal x. It is immediate that the pair in (8) satisfies the
generative model Y = HX = Vdiag(h̃)VT X. So, if u(i,j)
is an eigenvector of S then we can not distinguish the values
at nodes i and j and the problem remains non-identifiable.
To exemplify this situation, consider the toy graph illustrated
in Fig 1-(left). One can consider the adjacency matrix as the
shift (S = A) and denote the corresponding eigenvectors as
V = [v1 , · · · , v7 ]. Fig 1-(center) shows that v4 = u(2,4) . Then
it follows that for the matrix P in Fig 1-(right) and the vector
p = [1, 1, 1, −1, 1, 1, 1]T , one can construct another solution
{X1 , h̃1 } =
6 {X0 , h̃0 } using (8). In other words, nodes 2 and
4 are indistinguishable.
While it is challenging to obtain a formal characterization
of problematic graphs, in practice we have encountered issues
with dense networks as well as with some very sparse graphs.
For (continuous-valued) weighted graphs such ambiguities
effectively disappear. Before moving on to issues of exact
recovery, a remark on identifiability of (4) for a simple but
widely adopted (random) sparsity model is in order.
Remark 1 (Identifiability for Bernoulli-Gaussian model)
Because of its analytical tractability, the Bernoulli-Gaussian
model is widely adopted to describe and generate random
sparse matrices such as X ∈ RN ×P (we also use it for
the simulations in Section V). Sparse matrices adhering to
the model are X = Ω ◦ R, where Ω ∈ RN ×P is an i.i.d.
Bernoulli matrix with parameter θ (i.e., P[Ωij = 1] = θ),
and R ∈ RN ×P is an independent random matrix with
i.i.d. symmetric random variables drawn from a standard
Gaussian distribution. Under the Bernoulli-Gaussian model,
[12, Proposition 40] asserts that problem (6) is identifiable
(up to scaling and symmetric permutation ambiguities) with
probability at least 1 − exp(−cθP ), for N1 < θ < 14 and
P > cN log(N ), where c > 0 is a sufficiently large constant.
B. Exact recovery conditions
Suppose that (6) is identifiable and let {X0 , g̃0 } be the
solution. The following proposition (that relies heavily on
[23, Theorem 1]) offers sufficient conditions under which the
convex relaxation (7) succeeds in exactly recovering {X0 , g̃0 }.
Proposition 1 Let I := supp(vec(X0 )) index the non-zero
entries of vectorized X0 , and let I c be the complement of I.
Moreover, define Z := YT V V ∈ RN P ×N and let ZS be
the submatrix of Z with rows indexed by S ⊂ {1, 2, ..., N P }.
Then, the solution to (7) is unique and equal to g̃0 if the two
following conditions are satisfied:
C1) rank(ZI c ) = N − 1; and
C2) There exists a vector f ∈ RN P such that ZT f = γ1N for
some γ 6= 0, such that fI = sign(ZI g̃0 ) and kfI c k∞ < 1.
Proof: As per [23, Theorem 1], g̃0 is the unique solution of
(7) if ker(ZI c ) ∩ ker(1N ) = {0}. But since g̃0 ∈ ker(ZI c )
and g̃0 6∈ ker(1N ) because of the constraint in (7), then C1)
ensures said intersection is {0}. Optimality condition C2)
essentially requires 1N to belong to the set of subgradients
of kZg̃k1 at g̃0 ; see [23, Theorem 1] for further details.
Naturally, a more insightful exact recovery and sample
complexity result along the lines of the one in Remark 1
would be most valuable [i.e., when are C1)-C2) satisfied for
the Bernoulli-Gaussian model?], but left as future work.
V. N UMERICAL R ESULTS
We assess the performance of our proposed approach by
testing the iteratively-reweighted `1 -norm minimization procedure in Algorithm 1. The per-iteration sparse recovery
problems are solved using CVX [7].
Simulation setup. In all cases we consider undirected graphs
with graph-shift operator chosen as the normalized adjacency
1
1
matrix S = D− 2 AD− 2 , where D := diag(A1N ) is a
diagonal matrix of node degrees. The ground-truth sparse input
matrix X0 is drawn from a Bernoulli-Gaussian model as in
Remark 1, for varying parameters N , P , and sparsity level
(i.e., number of nonzero entries) S. Filter coefficients h0 are
generated according to h0 = (e1 + αb)/ke1 + αbk1 , where
e1 = [1, 0, · · · , 0]T ∈ RL is the first canonical basis vector and
entries of b ∈ RL are drawn independently from a standard
Gaussian distribution. Such a model for h0 is inspired by
[22], and we later corroborate that the recovery performance
improves as α decreases. Also note that h0 is normalized to
unit `1 -norm to fix the scale of the problem. Finally, given
X0 and H0 = Vdiag(ΨL h0 )VT , the N × P matrix of
observations is generated as Y = H0 X0 .
The relative recovery error eX = kX̂ − X0 k/kX0 k is
adopted as figure of merit to evaluate algorithmic performance.
We estimate the rate of successful recovery for synthetic and
real-world graphs under different parameters by defining a
successful recovery as one with eX < 0.01.
Random graphs. Consider Erdős-Rényi random graphs with
N = 50 nodes, where edges are formed independently
with probability p = 0.3. The rate of successful recovery
is estimated for realizations of random graphs which are
connected and do not give rise to permutation ambiguities (cf.
Section IV-A). Figures 2(a) and 2(b) depict the recovery rates
as a function of P and S for α = 0.1 and 0.3, respectively,
averaged over 100 realizations for (invertible) graph filters of
order L = 5. As expected, in both cases recovery is more
challenging for larger S and smaller P ; see the dark-gray area
of low-success probability around the top-left corner. Moreover, decreasing α makes successful recovery more likely.
For instance, for α = 0.1 [Fig. 2(a)] we can successfully
recover dense input signals with e.g., S ≈ N/2 = 25 and
only P = 10 observations. For (effectively) lower-order filters
resulting in more localized diffusion dynamics, one obtains
favorable recovery performance.
We also compare the proposed approach against its lessscalable, matrix lifting-based precursor in [21]. Figures 2(c)
and 2(d) respectively show the recovery rates for both methods
as a function of sparsity S and filter order L, for N = 50,
p = 0.3, P = 10, and α = 0.5 averaged over 20 realizations. Apparently, Algorithm 1 [Fig. 2(c)] can be successful
over a larger range of values of L. Moreover, it uniformly
outperforms the algorithm in [21, Problem (9)]; see Fig. 2(d).
Brain graph. We also consider a structural brain graph with
N = 66 nodes or neural regions of interest (ROIs), and
edge weights given by the density of anatomical connections
between regions [8]. The level of activity of each ROI can be
represented by a graph signal x, thus successive applications
of S model a linear evolution of the brain activity pattern.
Supposing we observe a linear combination (filter) of the
evolving states of an originally sparse brain signal, then blind
identification amounts to jointly estimating which regions
were originally active, the activity in these regions and the
coefficients of the linear combination.
We repeat the recovery-rate analysis performed for ErdősRényi graphs averaged over 20 realizations, and report the
results in Figs. 2(e)–(h). Figures 2(e) and 2(f) showcase
that our algorithm successfully identifies the initial excitation
regions as well as the diffusion coefficients over a broad
region in parameter space. By comparing Figs. 2(g) and 2(h),
1
40
1
40
0.8
0.8
30
0.6
20
0.4
0.2
10
10
20
30
40
30
0.6
20
0.2
10
0
10
20
40
30
20
1
50
0.8
40
0.6
30
0.4
20
0.2
10
30
0.6
4
40
40
50
0
0.2
2
4
6
0.6
0.2
(e)
30
0.2
2
40
50
0
(f)
4
6
6
1
6
0.8
0.6
4
0.8
0.6
4
0.4
2
0.2
2
4
(g)
0
(d)
0.4
20
0.4
2
0
1
0.8
10
0.6
4
(c)
1
10
0
0.8
0.4
2
(b)
50
20
30
1
6
0.8
0.4
(a)
10
1
6
6
0
0.4
2
0.2
2
4
6
0
(h)
Fig. 2. Rate of recovery of X as a function of S (number non-zero entries in X) and P (number of observations) in N = 50-node Erdős-Rényi graphs with
p = 0.3 (edge existence probability) for (a) α = 0.1 and (b) α = 0.3, using Algorithm 1. Plots (e) and (f) are counterparts of (a) and (b), respectively, for
the structural brain network in [8]. Recovery rate in Erdős-Rényi graphs (N = 50, p = 0.3) as a function of S and L (filter order) using (c) Algorithm 1
and (d) the matrix-lifting approach of [21]. Plots (g) and (h) are counterparts of (c) and (d), respectively, for the aforementioned structural brain network.
it is apparent that also in this setting the proposed approach
outperforms the state-of-the-art method in [21], corroborating
the effectiveness of Algorithm 1.
VI. C ONCLUSION
We studied the problem of blind graph filter identification,
which extends blind deconvolution of time (or spatial) domain
signals to graphs. By introducing a mild assumption on
invertibility of the graph filter, we obtained a computationally
simpler convex relaxation for (diffused) source localization in
the multi-signal case. Ongoing work includes deriving suitable
graph-dependent conditions under which exact (and stable)
recovery can be guaranteed, even when only a fraction of
nodes is observed. This is a challenging problem, since the
favorable (circulant) structure of time-domain filters is no
longer present in the network-centric setting dealt with here.
R EFERENCES
[1] A. Ahmed, B. Recht, and J. Romberg, “Blind deconvolution using
convex programming,” IEEE Trans. Inf. Theory, vol. 60, no. 3, pp. 1711–
1732, 2014.
[2] E. J. Candes, M. B. Wakin, and S. P. Boyd, “Enhancing sparsity by
reweighted `1 minimization,” Journal of Fourier Analysis and Applications, vol. 14, no. 5-6, pp. 877–905, 2008.
[3] M. H. DeGroot, “Reaching a consensus,” Journal of the American
Statistical Association, vol. 69, pp. 118–121, 1974.
[4] J. A. Deri and J. M. F. Moura, “New York City taxi analysis with
graph signal processing,” in Proc. IEEE Global Conf. on Signal and
Information Process., Dec. 2016, pp. 1275–1279.
[5] S. Feizi, M. Médard, G. Quon, M. Kellis, and K. Duffy, “Network
infusion to infer information sources in networks,” arXiv preprint
arXiv:1606.07383 [cs.SI], 2016.
[6] A. Gavili and X.-P. Zhang, “On the shift operator, graph frequency,
and optimal filtering in graph signal processing,” IEEE Trans. Signal
Process., vol. 65, no. 23, pp. 6303–6318, 2017.
[7] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex
programming, version 2.1,” http://cvxr.com/cvx, Mar. 2014.
[8] P. Hagmann, L. Cammoun, X. Gigandet, R. Meuli, C. J. Honey, V. J.
Wedeen, and O. Sporns, “Mapping the structural core of human cerebral
cortex,” PLoS Biology, vol. 6, no. 7, p. e159, 2008.
[9] C. Hu, X. Hua, J. Ying, P. M. Thompson, G. E. Fakhri, and Q. Li,
“Localizing sources of brain disease progression with network diffusion
model,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 7, pp. 1214–
1225, 2016.
[10] W. Huang, L. Goldsberry, N. F. Wymbs, S. T. Grafton, D. S. Bassett,
and A. Ribeiro, “Graph frequency analysis of brain signals,” IEEE J.
Sel. Topics Signal Process., vol. 10, no. 7, pp. 1189–1203, Oct. 2016.
[11] A. Levin, Y. Weiss, F. Durand, and W. T. Freeman, “Understanding blind
deconvolution algorithms,” IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 33, no. 12, pp. 2354–2367, 2011.
[12] Y. Li, K. Lee, and Y. Bresler, “Identifiability in bilinear inverse problems
with applications to subspace or sparsity-constrained blind gain and
phase calibration,” IEEE Trans. Inf. Theory, vol. 63, no. 2, pp. 822–
842, Feb 2017.
[13] S. Ling and T. Strohmer, “Self-calibration and biconvex compressive
sensing,” Inverse Problems, vol. 31, no. 11, p. 115002, 2015.
[14] A. Ortega, P. Frossard, J. Kovac̆ević, J. M. F. Moura, and P. Vandergheynst, “Graph signal processing,” arXiv preprint arXiv:1712.00468
[eess.SP], 2017.
[15] R. Pena, X. Bresson, and P. Vandergheynst, “Source localization on
graphs via `1 recovery and spectral graph theory,” in Proc. IEEE Image,
Video, and Multidimensional Signal Process. Workshop, 2016, pp. 1–5.
[16] P. C. Pinto, P. Thiran, and M. Vetterli, “Locating the source of diffusion
in large-scale networks,” Physical Review Letters, vol. 109, no. 6, p.
068702, 2012.
[17] D. Ramı́rez, A. G. Marques, and S. Segarra, “Graph-signal reconstruction and blind deconvolution for diffused sparse inputs,” in Proc. Int.
Conf. Acoustics, Speech, Signal Process., Mar. 2017, pp. 4104–4108.
[18] A. Sandryhaila and J. M. Moura, “Discrete signal processing on graphs,”
IEEE Trans. Signal Process., vol. 61, no. 7, pp. 1644–1656, 2013.
[19] I. D. Schizas, G. Mateos, and G. B. Giannakis, “Distributed LMS for
consensus-based in-network adaptive processing,” IEEE Trans. Signal
Process., vol. 57, no. 6, pp. 2365–2381, Mar. 2009.
[20] S. Segarra, A. G. Marques, and A. Ribeiro, “Optimal graph-filter design
and applications to distributed linear network operators,” IEEE Trans.
Signal Process., vol. 65, no. 15, pp. 4117–4131, Aug 2017.
[21] S. Segarra, G. Mateos, A. G. Marques, and A. Ribeiro, “Blind identification of graph filters,” IEEE Trans. Signal Process., vol. 65, no. 5, pp.
1146–1159, Mar. 2017.
[22] L. Wang and Y. Chi, “Blind deconvolution from multiple sparse inputs,”
IEEE Signal Process. Lett., vol. 23, no. 10, pp. 1384–1388, 2016.
[23] H. Zhang, M. Yan, and W. Yin, “One condition for solution uniqueness
and robustness of both l1-synthesis and l1-analysis minimizations,” Adv.
Comput. Math., vol. 42, no. 6, pp. 1381–1399, 2016.
[24] P. Zhang, J. He, G. Long, G. Huang, and C. Zhang, “Towards anomalous
diffusion sources detection in a large network,” ACM T. Internet Techn.,
vol. 16, no. 1, p. 2, 2016.
| 7cs.IT
|
Reproducibility in Research: Systems, Infrastructure, Culture
Tom Crick1 , Benjamin A. Hall2 and Samin Ishtiaq3
1
arXiv:1503.02388v2 [cs.SE] 28 Jul 2017
Cardiff Metropolitan University, UK
2
University of Cambridge, UK
3
Microsoft Research Cambridge, UK
1
[email protected]
2
[email protected]
3
[email protected]
Abstract
The reproduction and replication of research results has become a major issue for a number of scientific disciplines. In computer science and related computational disciplines such as systems biology,
the challenges closely revolve around the ability to implement (and exploit) novel algorithms and models.
Taking a new approach from the literature and applying it to a new codebase frequently requires local
knowledge missing from the published manuscripts and transient project websites. Alongside this issue,
benchmarking, and the lack of open, transparent and fair benchmark sets present another barrier to the
verification and validation of claimed results.
In this paper, we outline several recommendations to address these issues, driven by specific examples
from a range of scientific domains. Based on these recommendations, we propose a high-level prototype
open automated platform for scientific software development which effectively abstracts specific dependencies from the individual researcher and their workstation, allowing easy sharing and reproduction of
results. This new e-infrastructure for reproducible computational science offers the potential to incentivise
a culture change and drive the adoption of new techniques to improve the quality and efficiency – and
thus reproducibility – of scientific exploration.
1
Introduction
Marc Andreessen (co-author of Mosaic, the first widely used web browser) boldly stated in 2011 that
“software is eating the world ” [1]. This is true: we live in a computational world, with our everyday
communications, entertainment, shopping, banking, transportation, national security, etc, all heavily datadriven and largely overtaken by software.
Andreessen’s statement is particularly true for science and engineering. A 2012 report by the UK’s Royal
Society stated that computational techniques have “moved on from assisting scientists in doing science, to
transforming both how science is done and what science is done” [2]. Many of the examples discussed in
this paper exploit a fundamental advantage of computer science and more generally, computational science:
the unique ability for researchers to share the raw outputs of their work as software and datafiles. New
experiments, simulations, models, benchmarks, even proofs increasingly cannot be done without software.
This software does not consist of simple hack-together, use-once, throw-away scripts; research software
repositories contain thousands, perhaps millions, of lines of code and they increasingly need to be actively
supported and maintained. More importantly, with reproducibility being a fundamental tenet of science,
they need to be open and re-useable.
However, if we closely analyse the scientific literature related to software tools it often does not appear
to be adhering to these rules [3, 4]. How many of them are open and available? How many explain their
experimental methodologies, in particular the basis for their benchmarking? In particular, can we (re)build
the code? [5] We, the authors, are perhaps as guilty as anyone in the past, where we have published
papers [6, 7] with benchmarks and promises of code to be released in the near future which depreciate as
you move onto the next project.
There are various reasons why the wider scientific community is in this position. We are currently
undergoing significant changes to models of academic dissemination, especially considering the wider open
1
research movement, with new models being proposed [8–10]. Now, numerous “high-impact” journals explicitly require that source code and data is made available online under some form of open source license,
but there still exists large disciplinary gaps. While these initiatives are great, they are often optional, seem
piecemeal, and do little to enable the verification and validation of scientific results at a later stage. Even
within the same field, there are different ideas of what defines reproducibility [11], as well as evidence of
“overturn bias” – replications that overturn original results are much easier to publish than those that
confirm original results [12].
Nevertheless, the reproduction and replication of reported scientific results has become a widely discussed
topic within the scientific community [13–15]. Whilst the increasing number of retractions of studies across
a variety of disciplines has drawn the focus of many commentators, automated systems, which allow easy
reproduction of results, offer the potential to improve the efficiency of scientific exploration and drive the
adoption of new techniques. However, just publishing (linked) scientific data is not enough to ensure the
required reusability [16]. There exists a wider socio-cultural problem that pervades the scientific community,
with estimates that as much as 50% of published studies, even those in top-tier academic journals, cannot
be repeated with the same conclusions by an industrial lab [17, 18]. There are numerous non-technical
impediments to making software maintainable and re-useable. The pressure to “make the discovery”, publish
quickly and move onto the next project disincentivises careful software curation and preservation. Releasing
code prematurely is often seen to give your competitors an advantage, but we should be shining light into
these “black boxes” [14]; in essence: better software, better research [19].
However, there is promising existing work in this area [20–23], with a variety of manifestos for reproducible research and community initiatives [24–29], top tips and “ten simple rules” [30–36], as well as analysis
of the wider legal, professional, ethical and risk perspectives [37, 38]. Things can, should and need to be
much better if we want to uphold and maintain the scientific tenets of openness and sharing. Building upon
previous work [39, 40], we present a call to action, along with a set of recommendations which we hope will
lead to better, more sustainable, more re-useable software, to move towards an imagined future practice and
usage of scientific software development. We also propose a high-level specification for a service that would
automate many of our recommendations.
2
2.1
We Need to Talk About Reproducibility
Can I Implement Your Algorithm?
Reproducibility is a fundamental tenet of high-quality research. Yet many descriptions of algorithms are too
high-level, too obscure, too poorly-defined to allow an easy re-implementation by a third party. A step in the
algorithm might say: “We pick an element from the frontier set” but which element do you pick? Will the
first one do? Why will any element suffice? Sometimes the author would like to give more implementation
detail but is constrained by an arbitrary page limit of a conference or journal paper. Sometimes the authors’
description in-lines other algorithms or data structures that perhaps only that author is familiar with.
Until recently, reproducibility was only discussed at conferences and workshops convened explicitly for
that purpose. This is changing, and a number of high-profile computer science venues such as the ACM
SIGPLAN conferences POPL and PLDI now explicitly acknowledge the importance of reproducibility, promoting community-driven reviewing and validation of software artefacts.
Recommendation I: We recommend that a paper must describe the algorithm in such a way that it is
implementable by any knowledgeable reader of that algorithm. The description is, of course, subjective, but
to help encourage better descriptions, we also recommend that — in addition to having incentives to support
sharing of computational artefacts — relevant scientific conferences develop special tracks for papers that
re-implement past papers’ algorithms, techniques or tools.
2.2
Set The Code Free
There can be no better proof of your algorithm working, than if you provide the source code of an implementation; software development is hard, but sharing and re-using code is relatively easy.
Many years ago, Richard Stallman (founder of the GNU Project and Free Software Foundation) postulated that all code would be free [41] and we would make our money by consulting on the code. As it
turns out, this is now the case for a significant part of the computing industry. There are, of course, hard
2
commercial pressures for keeping code closed-source. Even in the scientific domain, scientists and their
collaborators may wish to hold onto their code as a competitive advantage, especially if there exists larger
competitors who could use the available code to “reverse scoop” the inventors, charging into a promising
new research area opened by the inventors.
Closed source is one thing; licenses that deny the user from viewing, modifying, or sharing the source
are another thing. There are, however, even licences on widely adopted tools like Gaussian [42] (for computational chemistry) that prohibit even analysing software performance and behaviour. For example, a wide
variety of licenses exist for molecular dynamics software, with different degrees of openness e.g. Gromacs uses
the GNU Lesser General Public License (LGPL) [43], CHARMM and Desmond are Academic/Commercial
software licences [44, 45], Amber and NAMD are custom open-like licences. Z3 is an example from the
verification area: the code itself was only recently open sourced, but the previous MSR-LA license allowed
the source code to be read, copied, forked for academic use, providing researchers in the field substantial
flexibility [46].
Even ignoring licensing issues, sometimes the source is not made open because the author thinks that
it is not quite finished. You should follow the “release early, release often” mantra, as well as releasing
somewhere public like GitHub, where it is easy to share and fork. Your code is good enough [13].
Recommendation II: There is little doubt that, if scientific research wants to be open and free, then the
code that underlies it too needs to be open and free. Code that is available for browsing, modifying, and
forking, facilitates testing and comparison. We recommend that code be published under an appropriate
open source license [47]; while we defer legal discussion of the specifics of any particular licences, BSD and
Apache are good, flexible ones.
2.3
Be A Better Academic Citizen
If you have the appropriate knowledge, skills and experience, you can create better software. We have seen
the emergence of successful initiatives, such as the Software Sustainability Institute (http://www.software.ac.uk)
and the UK Community of Research Software Engineer (http://www.rse.ac.uk), in cultivating world-class
research through software, developing software skills and raising the profile of research software engineers.
Many scientists will not have had any formal, or even informal, training in scientific software development. Building upon the work of Software Carpentry (http://software-carpentry.org) and Data
Carpentry (http://www.datacarpentry.org), basic training in software engineering concepts like version
control (git, mercurial), unit testing (tests written to exercise the smallest testable parts of a system, like
a function exported from a module), regression testing (a test framework that ensures that previous results
are maintained over the changes in the source code), build tools (Make, scons), etc, can help improve the
quality of the software written enormously [48]. Interestingly, many of these concepts are taught to computer science undergraduates, but it could be argued that they are taught at the wrong time of their careers,
without the experience of complex, long-running projects.
Recommendation III: Software development skills should be regarded as fundamental literacies for scientists and engineers: we recommend that formal programming, data and computational skills are taught
as core at undergraduate and postgraduate level.
2.4
The Lingua Franca of Computational Research
There is no other scientific or technical field where its participants can just make up a non-principled artefact
like a programming language so easily. In a way, it shows how much of a “commons” computer science has
become, that anyone can create a new programming language, API, framework or compiler. This clearly
has its advantages and disadvantages.
High-level languages are generally more readable than their low-level relations. The “density” of a
program is often seen to be a good thing, but it is not always the case that a shorter Haskell program
(for example) is easier to maintain than a longer C++ one; what is important is the readability of the code
itself. A good example here is from the world of automatic theorem proving: the SSReflect language is much
more readable than the original, standard Coq language [49]. SSReflect uses mathematicians’ vernacular for
script commands, allows reproducibility of automatic proof-checking because parameters are named rather
3
than numbered. Even though these proof scripts are really only ever going to be run by a machine, they
seek to maintain the basic mathematical idea that a proof should be readable by another mathematician.
Many high-level programming languages impose constraints like types: that you can never add a number
and a string is the most basic example, but ML’s functors provide principled ways of plugging in components with their implementations completely hidden. Aggressive type checking avoids a subset of bugs which
can arise due to incorrectly written functions e.g. well publicised problems with a NASA Mars orbiter
(http://www.cnn.com/TECH/space/9909/30/mars.metric.02/). A further example is a pressure coupling
bug (http://redmine.gromacs.org/issues/14) in Gromacs [43], which arose due to the inappropriate
swapping of a pressure term with a stress tensor. A further extension of types, a concept called units of
measure that is implemented in languages such as F#, can deal with these kinds of bugs at compile time.
Similarly, problems found using in-house software for crystallography led to the retraction of five papers [50],
due to a bug which inverted the phases.
Recommendation IV: The use of a principled, high-level, typed programming language in which to write
your software helps hugely with the maintainability, robustness and openness of the software produced.
Even in web frontend work, you have choices: use Typescript or Flow rather than plain old Javascript; use
Hack rather than PHP.
2.5
Lineage (or: “Standing On The Shoulders Of Giants”)
Research software is not just software – it is the instantiation of novel algorithms and data structures (or
at least novel applications of data structures). Thus, lineage is important:
Recommendation V: Code should always include links to papers publishing key algorithms and the code
should include explicit relationships to other projects on the repository (i.e. Project B was branched from
Project A). This ensures that both the researchers and software developers working upstream of the current
project are properly credited, encouraging future sharing and development. Remember, the people who did
the research are not necessarily the same people as the developers and maintainers of the software, so it
is important to reward both appropriately with citations: take note of the FORCE11 Software Citation
Principles [51].
2.6
YMMV
Figure 1: #overlyhonestmethods on Twitter by @ianholmes
[source: https://twitter.com/ianholmes/status/288689712636493824]
The tweet in Figure 1 is satirical but worryingly true, highlighting the perils of reproducible research.
4
Often, the tool that the paper describes does not exist for download. Or runs only on one particular bespoke
platform. Or might run for the author, for a while, but will ‘bit-rot’ so quickly that even the author cannot
compile it the following year. Computational reproducibility would appear to be more straightforward than
replicating physical experiments, but the complex and rapidly changing nature of computer systems and
environments that are being used across different disciplines makes being able to reproduce and extend such
work a serious challenge [52].
Recommendation VI: You must provide the source code of the tool, but also with details of precisely
how you built and wrote the software. For example:
• you should provide the compiler and build toolchain;
• you should provide build tools (e.g. Makefiles/Ant/etc) and comprehensive build instructions;
• you should list or link to all non-standard packages and libraries that you use;
• you should note the specifics of the hardware and OS used.
This may appear to be significant extra overhead for researchers, but GitHub APIs, continuous integration servers, virtual machines and cloud environments can make it easier; see Section 3 for more on
this.
2.7
Data Representations and Formats
We often do not, and should not, care how things are stored on disk, what their precise representations are.
A common, constrained, standard representation is however good for passing tests or models around between
different tools. A properly described representation, like the SMT-LIB format (http://smt-lib.org) for
Satisfiability Modulo Theory (SMT) solvers, where both the syntax and semantics are well understood,
hugely aids developing tools, techniques and benchmarks.
Another example, from biology, is that of the standard representation of qualitative networks and Boolean
networks [53, 54]. These networks can be expressed in SMV format, but this would mean that standard
qualitative/Boolean network behaviours have to be hard-coded for each variable, introducing the possibility
for errors. In the BioModelAnalyzer tool [55], the JSON contains only the modifiable parameters limiting
the possibility for error; the SBML-Qual standard achieves a similar goal for logical models [56].
Recommendation VII: Avoid creating new representations when common formats already exist. Use
existing extensible internationally standardised representations and formats to facilitate sharing and re-use.
2.8
World Records
The benchmarks the tool describes are fashioned only for this instance of this time. They might claim to
be from the Microsoft Windows device driver set, but the reality is that they are stripped down versions
of the originals. Stripped down so much as to be useless to anyone but the author vs. the referee. It is
worse than that really: enough benchmarks are included to beat other tools. The comparisons are never
fair (especially for comparisons against your tool). If every paper has to be novel, then every benchmark,
too, will be novel; there is no monotonic, historical truth in new, synthetically-crafted benchmarks. It is as
if, in order to beat Usain Bolt’s 100m world record time, you make him race overweight and out of season,
with a winter overcoat and the wrong sized shoes. Given this setup, you could surely hope to beat his 9.58s
time on a shorter length track.
Recommendation VIII: Benchmarks should be public. They should allow anyone to contribute, implying that the tests are in a standard format. Further, these benchmarks must be heavily curated. Every
test/assertion should be justified. Papers should be penalised if they do not use these public benchmarks.
While there are some domains in which it may not be immediately possible to share full benchmarks sets,
this should be the exception (with justification) rather than the norm.
A good example of some of these points is the RCSB Protein Data Bank (http://www.pdb.org) and
Systems Biology Markup Language [56]. The software ones we know of, the SMT Competition (http:
5
//smtcomp.sourceforge.net/2014/), SV-COMP (http://sv-comp.sosy-lab.org/2015/) and Termination
Problems Data Base (http://termination-portal.org/wiki/TPDB) are on that journey. Such repositories
would allow the tests to be taken and easily analysed by any competitor tool. Some communities go
further; the Critical assessment of methods of protein structure prediction and prediction of interactions
(CASP and CAPRI)[57,58] communities present a single-blind test of protein folding and docking algorithms
annually, allowing open competition on a level playing field. Similarly the DREAM challenges (http:
//dreamchallenges.org/) attempt to address large scale problems through open competition.
2.9
Test It To See
Some models may be chaotic and influenced by floating-point errors (e.g. molecular dynamics), further
frustrating testing. For example: Sidekick is an automated tool for building molecular models and performing simulations [59]. Each system is simulated from an different initial random seed, and under most
circumstances this is the only difference expected between replicas. However, on a mixed cluster with both
AMD and Intel microprocessors on the nodes, the difference in architecture was found to alter the number
of water molecules added to each system by one. This meant that the same simulation performed on different architectures would diverge. Similarly, in a different simulation engine, different neighbour searching
strategies gave divergent simulations due to the differing order in which forces were summed.
A further example is the handling of pseudo-random number generation in Avida [60], an open source
scientific software platform for conducting and analysing experiments with self-replicating and evolving computer programs. While it may initially appear attractive to develop bespoke random number generators
within a system for consistency or performance across platforms, this invariably adds complexity to your
system and may inhibit sharing and reproducibility.
Recommendation IX: Despite these challenges to testing, unshared code is ultimately untestable. Testing
new complex scientific software is difficult – until the software is complete, unit tests may not be available.
You should aim to re-use modules or repos (git submodules) from publicly-shared code; a corollary of Linus’s
Law (“given enough eyeballs, all bugs are shallow”) might be that shared code is inherently more test-able.
2.10
Welcome to Web 2.0
Virtual machines (VMs) in the cloud also make the testing of scaling properties more simple. If you have a
tool that you claim is more efficient, you could put together a cluster of slow nodes in the cloud to demonstrate how well the software scales for parallel calculations. Cloud computing is cheap, and getting cheaper.
Algorithms that used to require massive HPC resources can now be run cheaply by bidding on the VM spot
market. The web is a great leveller: use and share workflows and web services [61, 62].
Recommendation X: The web and the cloud really do open up a whole new way of working. Even
small, seemingly trivial features like putting up a web interface to your tool and its tests will allow users
who are not able to install necessary dependencies to explore the running of the tool [63]. Ultimately,
this can lead to making an “executable paper” appear on the Internet. The interactive Try F#(http:
//www.tryfsharp.org/Learn) and Z3 tutorials (http://rise4fun.com/Z3/tutorial/guide) are a great
start that begin to expose what can be done in this area.
3
A Model for Reproducible Research Software
Some of our Recommendations, such as “Be A Better Person” or “The Lingua Franca”, are abstract, airyfairy, pie-in-the-sky even. However, most of them can be concretely realised by a service for reproducibility.
This service provides a concrete implementation of free source code (“Set The Code Free”) that depends
on other free source code (“Lineage”) building (“YMMV ”, “Welcome to Web 2.0 ”) and running tests
contributed in public (“Data Representations”, “World Records”) in a completely reproducible regime.
The service we describe here can be seen as a specification. We have not built it, but many services like
travis-ci or Azure VSTS provide some of the mechanical parts of it. A service for reproducibility is intended
to play three important roles; it should:
i) Demonstrate that a piece of code can be compiled, run and behaves as described, without manual
intervention from the developer;
6
ii) Store and link specific artefacts with their linked publications or other publicly-accessible datasets;
iii) Allow new benchmarks to be added, by users other than the developer, to widen the testing and identify
potential bugs.
The whole premise of our previous paper [40] is that algorithms (and their implementations) and models
(sometimes also called benchmarks) are inextricably linked. Algorithms are designed for certain types of
models; models, though created to mimic some physical reality, also serve to express the current known
algorithms. An integrated autonomous open cloud-based service can make this link explicit.
By developing a cloud-based, centralised service, which performs automated code compilation, testing
and benchmarking (with associated auditing), we will link together published implementations of algorithms
and input models. This will allow the prototype to link together software and data repositories, toolchains,
workflows and outputs, providing a seamless automated infrastructure for the verification and validation of
scientific models and in particular, performance benchmarks. The program of work will lead the cultural
shift in both the short and long-term to move to a world in which computational reproducibility helps
researchers achieve their goals, rather than being perceived as an overhead.
A system as described here has several up-front benefits: it links research papers more closely to their
outputs, making external validation easier and allows interested users to explore unaddressed sets of models.
Critically, it helps researchers across computational science to be more productive, rather than reproducibility being an overhead on top of their day-to-day work. In the same way that tools such as GitHub make
collaborating easier while simultaneously allowing effortless sharing, we envisage our system being similarly usable for sharing and testing algorithms and their implementations, software, models and benchmarks
online.
Suppose you have come up with a better algorithm to deal with some standard problem. You write
up the paper on the algorithm, and you also push an implementation of your algorithm to the our cloud
environment’s section on this standard problem. The effect of pushing your implementation is to register
your program as a possible competitor in this standard problem competition. There exist several dozen
widely-agreed tests on this problem already on our cloud environment’s database. Maybe, after some
negotiation due to your novel approach to this standard problem, you add some of your own tests to the
database too.
Pushing your code activates the environment’s continuous integration system. The cloud pulls in all the
dependencies your code needs, on the platforms you specify, and runs all the benchmarks. This happens
every time you push. It also happens every time one of your dependencies (a library, a firmware upgrade
for your platform, a new API) changes too. This system (presented in Figure 2) would integrate with
publicly available source code repositories, automates the build, testing and benchmarking of algorithms
and benchmarks. It would allow testing models against competing algorithms, and the addition of new
models to the test suite (either manually or from existing online repositories).
Figure 2: Proposed reproducibility service workflow
If we are truly serious about addressing the systemic socio-technical issues in scientific disciplines that
are underpinned by leveraging software and computational techniques, then the proposal above would bring
together almost all of the points we have discussed in this paper to provide an open research infrastructure
for all. There are already several web services that already aim to do many of these things [22,64], so a service
that can integrate most if not all of these features is possible. Such a service would then allow algorithms
and models to evolve together, and be reproducible from the outset. Something more open and complete,
7
and stamped with the authority of the major domain conferences/journals/national academies, would mean
that your code would never ‘bit-rot’, and no one would have problems reproducing the implementation of
your published algorithm.
4
Next Steps
Following the proposal of such a system, the question becomes: how do we encourage widespread uptake,
or even standardisation? Such a service would appear to be non-trivial, given the large numbers of tools
and workflows that could potentially require to be supported by the service. After such a service has been
implemented, how do we ensure it is useful and usable for researchers. Furthermore, how do we make it
sustainable?
The benefits to the wider computational research community from a cultural change to favour reproducibility are clear and as such we should aim through software e-infrastructure and sharable, community
curated research workflows to mitigate these costs. Furthermore, we can reasonably expect the distinct
needs of specific research communities to evolve over time, and initial implementations of the platform may
require refinement in response to user feedback (supporting the critical cultural change by improving the
efficiency of researchers). As such, if the wider research community is to move to requiring reproducibility,
it seems most reasonable that this is staggered over a number of years to allow for both of these elements
to develop, until eventually all researchers are required to use the service.
The key question for different research communities then becomes: how to initialise this change? Such
a requirement creates a set of new costs to researchers, both in terms of time spent ensuring that their tools
work on the centralised system (in addition to their local implementation), but also potentially in terms of
equipment (in terms of running the system). Such costs may be easier to bear for some groups compared
to others, especially those with large research groups who can more easily distribute the tasks, and it is
important that the service does not present a barrier to early career researchers and those with efficient
budgets (this type of cost analysis is not unique to reproducibility efforts – it has been estimated that a
shift to becoming exclusively open access for a journal may lead to a ten-fold increase in computer science
publication costs [65]).
Nevertheless, this proposed new e-infrastructure could have a profound impact on the way that computational science is performed, repositioning the role of models, algorithms and benchmarks and accelerating
the research cycle, perhaps truly enabling a “fourth paradigm” of data intensive scientific discovery [66].
Ultimately though, continuing with an honest and open discussion of what reproducibility means for the
wider science research community is important: we all need to explicitly confirm that this is worthwhile and
commit to addressing it, or don’t bother doing it at all.
A Note on Re-Writing the WSSSPE Paper
Many of the ideas, comments — even attitudes — in this paper come from the authors’ experience in
programming, programming languages, software. We have started from the Marc Andreessen comment
that opens this paper. In editing this paper from its original WSSSPE workshop form, we realised that
one assumption that seems to run through the manuscript is that the behaviours we think are good are in
fact those that can be enforced in software. Take mutability of variables in programming as an example.
Mutability increases the scope for bugs, so modern programming languages like OCaml or C++14 enforce
immutability at the language or library level. But in fact immutability leads very naturally to state-less
or de novo build environments, and so to the guideline that “software must be compilable with de novo
continuous integration”. And, similarly, so does the issue of openly publishing your toolchain: it too must
be compilable in a from-scratch build environment to be of use to anyone else.
References
[1] M. Andreessen, “Why Software Is Eating The World,” The Wall Street Journal, August 2011. Available
online: http://online.wsj.com/news/articles/SB10001424053111903480904576512250915629460.
[2] Royal Society, “Science as an open enterprise,” 2012. Available from: https://royalsociety.org/
policy/projects/science-public-enterprise/report/.
8
[3] Editorial, “Devil in the details,” Nature, vol. 470, no. 7334, pp. 305–306, 2011.
[4] B. Alberts, R. J. Cicerone, S. E. Fienberg, A. Kamb, M. McNutt, R. M. Nerem, R. Schekman, R. Shiffrin, V. Stodden, S. Suresh, M. T. Zuber, B. Kline Pope, and K. Jamieson, “Self-correction in science
at work,” Science, vol. 348, no. 6242, pp. 1420–1422, 2015.
[5] C. Collberg and T. A. Proebsting, “Repeatability in Computer Systems Research,” Communications
of the ACM, vol. 59, no. 3, pp. 62–69, 2016.
[6] T. Crick, M. De Vos, M. Brain, and J. Fitch, “Generating Optimal Code using Answer Set Programming,” in Proceedings of 10th International Conference on Logic Programming and Nonmonotonic
Reasoning (LPNMR’09), vol. 5753 of Lecture Notes in Computer Science, pp. 554–559, Springer, 2009.
[7] J. Berdine, B. Cook, and S. Ishtiaq, “SLAyer: Memory Safety for Systems-Level Code,” in Proceedings
of the 23rd International Conference on Computer Aided Verification (CAV 2011), vol. 6806 of Lecture
Notes in Computer Science, pp. 178–183, Springer, 2011.
[8] D. De Roure, “Replacing the Paper: The Twelve Rs of the e-Research Record.” http:
//www.scilogs.com/eresearch/replacing-the-paper-the-twelve-rs-of-the-e-researchrecord/, November 2011.
[9] V. Stodden, P. Guo, and Z. Ma, “Toward Reproducible Computational Research: An Empirical Analysis
of Data and Code Policy Adoption by Journals,” PLoS ONE, vol. 8, no. 6, 2013.
[10] G. Fursin and C. Dubach, “Community-Driven Reviewing and Validation of Publications,” in Proceedings of the 1st ACM SIGPLAN Workshop on Reproducible Research Methodologies and New Publication
Models in Computer Engineering (TRUST’14), pp. 1–4, ACM Press, 2014.
[11] National Academies of Sciences, Engineering, and Medicine, Statistical Challenges in Assessing and
Fostering the Reproducibility of Scientific Results: Summary of a Workshop. The National Academies
Press, 2016.
[12] S. Galiani, P. Gertler, and M. Romero, “Incentives for Replication in Economics,” tech. rep., National
Bureau of Economic Research, July 2017. NBER Working Paper No. 23576.
[13] N. Barnes, “Publish your computer code: it is good enough,” Nature, vol. 467, no. 753, 2010.
[14] A. Morin, J. Urban, P. D. Adams, I. Foster, A. Sali, D. Baker, and P. Sliz, “Shining Light into Black
Boxes,” Science, vol. 336, no. 6078, pp. 159–160, 2012.
[15] L. N. Joppa, G. McInerny, R. Harper, L. Salido, K. Takeda, K. O’Hara, D. Gavaghan, and S. Emmott,
“Troubling Trends in Scientific Software Use,” Science, vol. 340, no. 6134, pp. 814–815, 2013.
[16] S. Bechhofer, I. Buchan, D. De Roure, P. Missier, J. Ainsworth, J. Bhagata, P. Couch, D. Cruickshank,
M. Delderfield, I. Dunlop, M. Gamble, D. Michaelides, S. Owen, D. Newman, S. Sufi, and C. Goble,
“Why linked data is not enough for scientists,” Future Generation Computer Systems, vol. 29, no. 2,
pp. 599–611, 2013.
[17] L. Osherovich, “Hedging against academic risk,” Science-Business eXchange, vol. 4, no. 15, 2011.
[18] T. Hesman Saey, “Repeat Performance: Too many studies, when replicated, fail to pass muster,”
Science News, vol. 187, no. 2, pp. 21–26, 2015.
[19] C. Goble, “Better Software, Better Research,” IEEE Internet Computing, vol. 18, no. 5, pp. 4–8, 2014.
[20] F. Chirigati, M. Troyer, D. Shasha, and J. Freire, “A Computational Reproducibility Benchmark,”
IEEE Data Engineering Bulletin, vol. 36, no. 4, pp. 54–59, 2013.
[21] V. Stodden and S. Miguez, “Best Practices for Computational Science: Software Infrastructure and
Environments for Reproducible and Extensible Research,” Journal of Open Research Software, vol. 2,
no. 1, pp. 1–6, 2014.
9
[22] V. Stodden, S. Miguez, and J. Seiler, “ResearchCompendia.org: Cyberinfrastructure for Reproducibility
and Collaboration in Computational Science,” Computing in Science & Engineering, vol. 17, no. 12,
2015.
[23] V. Stodden, M. McNutt, D. H. Bailey, E. Deelman, Y. Gil, B. Hanson, M. A. Heroux, J. P. Ioannidis,
and M. Taufer, “Enhancing reproducibility for computational methods,” Science, vol. 354, no. 6317,
pp. 1240–1241, 2016.
[24] S. Fomel and J. F. Claerbout, “Reproducible Research,” Computing in Science & Engineering , vol. 11,
no. 1, 2008.
[25] “Reproducible Research,” Computing in Science & Engineering, vol. 12, no. 5, pp. 8–13, 2010.
[26] I. P. Gent, “The Recomputation Manifesto.” Available from: http://arxiv.org/abs/1304.3674, April
2013.
[27] G. Fursin, R. Miceli, A. Lokhmotov, M. Gerndt, M. Baboulin, A. D. Malony, Z. Chamski, D. Novillo,
and D. Del Vento, “Collective mind: Towards practical and collaborative auto-tuning,” Scientific Programming, vol. 22, no. 4, pp. 309–329, 2014.
[28] D. Bailey, J. Borwein, and V. Stodden, “Set the Default to “Open”,” Notices of the AMS, 2013.
[29] D. James, N. Wilkins-Diehr, V. Stodden, D. Colbry, C. Rosales, M. R. Fahey, J. Shi, R. F. da Silva,
K. Lee, R. Roskies, L. Loewe, S. Lindsey, R. Kooper, L. Barba, D. H. Bailey, J. M. Borwein, Ó. Corcho,
E. Deelman, M. C. Dietze, B. Gilbert, J. Harkes, S. Keele, P. Kumar, J. Lee, E. Linke, R. Marciano, L. Marini, C. Mattmann, D. Mattson, K. McHenry, R. T. McLay, S. Miguez, B. S. Minsker,
M. S. Pérez-Hernández, D. Ryan, M. Rynge, I. S. Pérez, M. Satyanarayanan, G. S. Clair, K. Webster,
E. Hovig, D. S. Katz, S. Kay, G. K. Sandve, D. Skinner, G. Allen, J. Cazes, K. W. Cho, J. Fonseca,
L. Hwang, L. Koesterke, P. Patel, L. Pouchard, E. Seidel, and I. Suriarachchi, “Standing Together for
Reproducibility in Large-Scale Computing: Report on reproducibility@XSEDE,” tech. rep., XSEDE,
2014.
[30] A. Prlić and J. B. Procter, “Ten Simple Rules for the Open Development of Scientific Software,” PLoS
Computational Biology, vol. 8, no. 12, p. e1002802, 2012.
[31] H. Masum, A. Rao, B. M. Good, M. H. Todd, A. M. Edwards, L. Chan, B. A. Bunin, A. I. Su,
Z. Thomas, and P. E. Bourne, “Ten Simple Rules for Cultivating Open Science and Collaborative
R&D,” PLoS Computational Biology, vol. 9, no. 9, p. e1003244, 2013.
[32] G. Sandve, A. Nekrutenko, J. Taylor, and E. Hovig, “Ten Simple Rules for Reproducible Computational
Research,” PLoS Computational Biology, vol. 9, no. 10, p. e1003285, 2013.
[33] J. M. Osborne, M. O. Bernabeu, M. Bruna, B. Calderhead, J. Cooper, N. Dalchau, S.-J. Dunn, A. G.
Fletcher, R. Freeman, D. Groen, B. Knapp, G. J. McInerny, G. R. Mirams, J. Pitt-Francis, B. Sengupta,
D. W. Wright, C. A. Yates, D. J. Gavaghan, S. Emmott, and C. Deane, “Ten Simple Rules for Effective
Computational Research,” PLoS Computational Biology, vol. 10, no. 3, p. e1003506, 2013.
[34] A. Goodman, A. Pepe, A. W. Blocker, C. L. Borgman, K. Cranmer, M. Crosas, R. Di Stefano, Y. Gil,
P. Groth, M. Hedstrom, D. W. Hogg, V. Kashyap, A. Mahabal, A. Siemiginowska, and A. Slavkovic,
“Ten Simple Rules for the Care and Feeding of Scientific Data,” PLoS Computational Biology, vol. 10,
no. 4, p. e1003542, 2014.
[35] N. P. Chue Hong, T. Crick, I. P. Gent, L. Kotthoff, and K. Takeda, “Top Tips to Make Your Research
Irreproducible,” 2015. http://arxiv.org/abs/1504.00062.
[36] M. List, P. Ebert, and F. Albrecht, “Ten Simple Rules for Developing Usable Software in Computational
Biology,” PLoS Computational Biology, vol. 13, no. 1, p. e1005265, 2017.
[37] V. Stodden, “The Legal Framework for Reproducible Scientific Research: Licensing and Copyright,”
Computing in Science & Engineering, vol. 11, no. 1, 2008.
[38] C. N. Haas, “Reproducible Risk Assessment,” Risk Analysis, vol. 36, no. 10, pp. 1829–1833, 2016.
10
[39] T. Crick, B. A. Hall, and S. Ishtiaq, ““Can I Implement Your Algorithm?”: A Model for Reproducible
Research Software,” in 2nd International Workshop on Sustainable Software for Science: Practice and
Experiences (WSSSPE2), 2014.
[40] T. Crick, B. A. Hall, S. Ishtiaq, and K. Takeda, ““Share and Enjoy”: Publishing Useful (and Usable)
Scientific Models,” in Proceedings of the 7th IEEE/ACM International Conference on Utility and Cloud
Computing, pp. 957–961, 2014.
[41] R. M. Stallman, Free Software Free Society: Selected Essays of Richard M. Stallman. Free Software
Foundation, 2010.
[42] J. Giles, “Software company bans competitive users,” Nature, vol. 429, no. 6989, 2004.
[43] B. Hess, C. Kutzner, D. van der Spoel, and E. Lindahl, “GROMACS 4: Algorithms for Highly Efficient,
Load-Balanced, and Scalable Molecular Simulation,” Journal of Chemical Theory and Computation,
vol. 4, no. 3, pp. 435–447, 2008.
[44] B. R. Brooks, C. L. Brooks, A. D. Mackerell, L. Nilsson, R. J. Petrella, B. Roux, Y. Won, G. Archontis,
C. Bartels, S. Boresch, A. Caflisch, L. Caves, Q. Cui, A. R. Dinner, M. Feig, S. Fischer, J. Gao,
M. Hodoscek, W. Im, K. Kuczera, T. Lazaridis, J. Ma, V. Ovchinnikov, E. Paci, R. W. Pastor, C. B.
Post, J. Z. Pu, M. Schaefer, B. Tidor, R. M. Venable, H. L. Woodcock, X. Wu, W. Yang, D. M.
York, and M. Karplus, “CHARMM: The biomolecular simulation program,” Journal of Computational
Chemistry, vol. 30, no. 10, pp. 1545–1614, 2009.
[45] K. J. Bowers, E. Chow, H. Xu, R. O. Dror, M. P. Eastwood, B. A. Gregersen, J. L. Klepeis, I. Kolossvary, M. A. Moraes, F. D. Sacerdoti, J. K. Salmon, Y. Shan, and D. E. Shaw, “Scalable algorithms
for molecular dynamics simulations on commodity clusters,” in Proceedings of the 2006 ACM/IEEE
Conference on Supercomputing, IEEE Press, 2006.
[46] L. de Moura, “Releasing the Z3 source code,” 2012. Available online: http://leodemoura.github.io/
blog/2012/10/02/open-z3.html.
[47] “Open Source Licenses.” http://opensource.org/licenses.
[48] G. Wilson, “Software carpentry: Getting scientists to write better code by making them more productive,” Computing in Science & Engineering, vol. 8, no. 6, 2006.
[49] G. Gonthier, B. Ziliani, A. Nanevski, and D. Dreyer, “How to make ad hoc proof automation less ad
hoc,” Journal of Functional Programming, vol. 23, no. 4, pp. 357–401, 2013.
[50] G. Miller, “A Scientist’s Nightmare: Software Problem Leads to Five Retractions,” Science, vol. 314,
no. 5807, pp. 1856–1857, 2006.
[51] A. M. Smith, D. S. Katz, K. E. Niemeyer, and the FORCE11 Software Citation Working Group,
“Software Citation Principles,” PeerJ Computer Science, vol. 2, no. e86, 2016.
[52] C. Boettiger, “An introduction to Docker for reproducible research,” ACM SIGOPS Operating Systems
Review, vol. 49, no. 1, pp. 71–79, 2015. Special Issue on Repeatability and Sharing of Experimental
Artifacts.
[53] S. A. Kauffman, “Metabolic stability and epigenesis in randomly constructed genetic nets,” Journal of
Theoretical Biology, vol. 22, no. 3, pp. 437–67, 1969.
[54] M. A. Schaub, T. A. Henzinger, and J. Fisher, “Qualitative networks: a symbolic approach to analyze
biological signaling networks,” BMC Systems Biology, vol. 1, p. 4, 2007.
[55] D. Benque, S. Bourton, C. Cockerton, B. Cook, J. Fisher, S. Ishtiaq, N. Piterman, A. Taylor, and M. Y.
Vardi, “BMA: visual tool for modeling and analyzing biological networks,” in Proceedings of the 24th
International Conference on Computer Aided Verification (CAV 2012), vol. 7358 of Lecture Notes in
Computer Science, pp. 686–692, Springer, 2012.
11
[56] C. Chaouiya, D. Berenguier, S. M. Keating, A. Naldi, M. P. van Iersel, N. Rodriguez, A. Drager,
F. Buchel, T. Cokelaer, B. Kowal, B. Wicks, E. Goncalves, J. Dorier, M. Page, P. T. Monteiro, A. von
Kamp, I. Xenarios, H. de Jong, M. Hucka, S. Klamt, D. Thieffry, N. Le Novere, J. Saez-Rodriguez,
and T. Helikar, “SBML qualitative models: a model representation format and infrastructure to foster
interactions between qualitative modelling formalisms and tools,” BMC Systems Biology, vol. 7, 2013.
[57] J. Moult, K. Fidelis, A. Kryshtafovych, T. Schwede, and A. Tramontano, “Critical assessment of
methods of protein structure prediction (CASP) — round x,” Proteins: Structure, Function, and
Bioinformatics, vol. 82, pp. 1–6, 2014.
[58] M. F. Lensink, S. Velankar, and S. J. Wodak, “Modeling proteinprotein and proteinpeptide complexes:
Capri 6th edition,” Proteins: Structure, Function, and Bioinformatics, vol. 85, no. 3, pp. 359–377,
2017.
[59] B. A. Hall, K. B. A. Halim, A. Buyan, B. Emmanouil, and M. S. P. Sansom, “Sidekick for membrane
simulations: Automated ensemble molecular dynamics simulations of transmembrane helices,” Journal
of Chemical Theory and Computation, vol. 10, no. 5, pp. 2165–2175, 2014.
[60] C. Ofria and C. O. Wilke, “Avida: A Software Platform for Research in Computational Evolutionary
Biology,” Artificial Life, vol. 10, no. 2, pp. 191–229, 2004.
[61] T. Crick, P. Dunning, H. Kim, and J. Padget, “Engineering Design Optimization using Services and
Workflows,” Philosophical Transactions of the Royal Society A, vol. 367, no. 1898, pp. 2741–2751, 2009.
[62] S. Olabarriaga, G. Pierantoni, G. Taffoni, E. Sciacca, M. Jaghoori, V. Korkhov, G. Castelli, C. Vuerli,
U. Becciani, E. Carley, and B. Bentley, “Scientific Workflow Management – For Whom?,” in Proceedings
of 10th IEEE International Conference on e-Science (e-Science 2014), pp. 298–305, IEEE Press, 2014.
[63] B. A. Hall, E. Jackson, A. Hajnal, and J. Fisher, “Logic programming to predict cell fate patterns and
retrodict genotypes in organogenesis,” Journal of The Royal Society Interface, vol. 11, no. 98, 2014.
[64] N. D. Rollins, C. M. Barton, S. Bergin, M. A. Janssen, and A. Lee, “A Computational Model Library for
publishing model documentation and code,” Environmental Modelling & Software, vol. 61, pp. 59–64,
2014.
[65] M. Y. Vardi, “Openism, IPism, Fundamentalism, and Pragmatism,” Communications of the ACM,
vol. 57, no. 8, 2014.
[66] T. Hey, S. Tansley, and K. Tolle, eds., The Fourth Paradigm: Data-Intensive Scientific Discovery.
Microsoft Research, 2009.
12
| 5cs.CE
|
Stabilized Sparse Scaling Algorithms for Entropy Regularized
Transport Problems
Bernhard Schmitzer
arXiv:1610.06519v1 [math.OC] 20 Oct 2016
October 21, 2016
Abstract
Scaling algorithms for entropic transport-type problems have become a very popular numerical method, encompassing Wasserstein barycenters, multi-marginal problems, gradient
flows and unbalanced transport. However, a standard implementation of the scaling algorithm has several numerical limitations: the scaling factors diverge and convergence becomes
impractically slow as the entropy regularization approaches zero. Moreover, handling the
dense kernel matrix becomes unfeasible for large problems. To address this, we propose several modifications: A log-domain stabilized formulation, the well-known ε-scaling heuristic,
an adaptive truncation of the kernel and a coarse-to-fine scheme. This allows to solve larger
problems with smaller regularization and negligible truncation error. A new convergence
analysis of the Sinkhorn algorithm is developed, working towards a better understanding of
ε-scaling. Numerical examples illustrate efficiency and versatility of the modified algorithm.
Contents
1 Introduction
1.1 Motivation and Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Contribution and Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Notation and Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
2
4
5
2 Entropy Regularized Transport-Type Problems and Diagonal Scaling Algorithms
7
2.1 Transport-Type Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.2 Entropy Regularization and Diagonal Scaling Algorithms . . . . . . . . . . . . . .
9
3 Stabilized Sparse Multi-Scale
3.1 Log-Domain Stabilization .
3.2 Epsilon-Scaling . . . . . . .
3.3 Kernel Truncation . . . . .
3.4 Multi-Scale Scheme . . . . .
Algorithm
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
12
12
14
15
18
4 Analogy between Sinkhorn and Auction Algorithm
4.1 Auction Algorithm . . . . . . . . . . . . . . . . . . . . .
4.2 Asynchronous Sinkhorn Algorithm and Iteration Bound
4.3 Stability of Dual Solutions . . . . . . . . . . . . . . . . .
4.4 Application to Epsilon-Scaling . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
23
27
31
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Numerical Examples
5.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Efficiency of Enhanced Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Versatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
33
34
36
6 Conclusion
42
A Additional Proofs
A.1 Proof of Lemma 4.20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2 Proof of Theorem 4.16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
43
45
1
1.1
Introduction
Motivation and Related Work
Applications of Optimal Transport. Optimal transport (OT) is a classical optimization
problem dating back to the seminal work of Monge and Kantorovich (see monographs [53, 3, 43]
for introduction and historical context). The induced Wasserstein distances lift a metric from a
‘base’ space X to probability measures over X. This is a powerful analytical tool, for example to
study PDEs as gradient flows in Wasserstein space [29, 4]. With the increase of computational
resources, OT has also become a popular numerical tool in image processing, computer vision
and machine learning (e.g. [42, 40, 54, 15, 50, 51, 46, 36, 23, 25]).
Many ideas have been presented to extend Wasserstein distances to general non-negative
measures. We refer to [31, 17, 35, 18] and references therein for some context. A transport-type
distance for general multi-channel signals is proposed in [52].
Computational Optimal Transport. To this day, the computational effort to solve OT
problems remains the main bottleneck in many applications. In particular large problems, or
even multi-marginal problems, remain challenging both in terms of runtime and memory demand.
For the linear assignment problem and discrete transport problems there are (combinatorial)
algorithms based on the finite dimensional linear programming formulation by Kantorovich,
such as the Hungarian method [33], the auction algorithm [10], the network simplex [2] and
more [27]. Typically, they work for (almost) arbitrary cost functions, but do not scale well for
large, dense problems. On the other hand, there are more geometric solvers, relying on the polar
decomposition [14], that tend to be more efficient. There is the famous fluid dynamic formulation
by Benamou and Brenier [6], explicit computation of the polar decomposition [28], semi-discrete
solvers [37], and solvers of Monge-Ampère equation [9, 8] among many others. However, these
only work on very specific cost functions, most notably the squared Euclidean distance. In a
compromise between efficiency and flexibility, several discrete coarse-to-fine solvers have been
proposed that adaptively select sparse sub-problems [45, 38, 44].
Entropy Regularization for Optimal Transport. Entropy regularization of optimal transport is related to the Schrödinger problem and the lazy gas experiment [34]. In the limit of
vanishing regularization, primal and dual solutions of the entropic OT problem converge to particular solutions of the unregularized problem. A very general study of this limit is given in [34].
[16] provides a simpler and direct analysis for the 2-Wasserstein distance on Rd , the discrete case
is investigated in [20].
2
It has been observed that entropy regularization is also useful numerically [32] as it allows to
use the Sinkhorn scaling algorithm [48], also called iterative projective fitting procedure (IPFP),
to obtain an approximate solution to OT. More recently, practical applicability of Sinkhorn’s
algorithm to large problems has been popularized by [22] and today it is widely used, e.g. [49,
41, 52]. The algorithm starts from a kernel matrix (induced by the OT cost function) and
iteratively rescales its rows and columns with diagonal scaling factors. Throughout the algorithm
only the scaling factors need to be modified. For sufficiently large regularization this method
has several advantages: The algorithm is very simple (and apt for parallelized implementation
on GPUs), and its runtime is relatively low. And, as with discrete algorithms, its feasibility
does not depend critically on the cost function. Moreover, the technique has been extended
to more general transport-type problems, such as Wasserstein barycenters and multi-marginal
problems [7], gradient flows [39] and unbalanced transport problems [19], resulting in a family of
Sinkhorn-like diagonal scaling algorithms.
Convergence Speed of Sinkhorn Algorithm. In [26] the convergence rate of the Sinkhorn
algorithm is studied for positive kernel matrices, yielding a global linear convergence rate of
the marginals in terms of Hilbert’s projective metric. However, applied to entropy regularized
optimal transport, the contraction factor tends to one exponentially, as the regularization approaches zero, leading to a weak and rather impractical bound. In [30] the local convergence
rate of the Sinkhorn algorithm near the solution is examined, based on a linearization of the
iterations. This bound is tighter and more accurately describes the behaviour of the algorithm
close to convergence. But these estimates do not apply when one starts far from the optimal
solution, which is the usual case for small regularization parameters. In [32] a qualitative comparison is made between the Sinkhorn algorithm and the auction algorithm. In particular the
role of the entropy regularization parameter is related to the slack parameter ε of the auction
algorithm and it is pointed out that convergence of both algorithms becomes slower, as these
parameters approach zero (but small parameters are required for good approximate solutions).
For the auction algorithm this can provably be remedied by ε-scaling, where the ε parameter
is gradually decreased during optimization. Analogously, it is suggested to gradually decrease
entropy regularization during the Sinkhorn algorithm to accelerate optimization. Consequently,
in the following we will also refer to the entropy regularization parameter as ε and to the gradual
reduction scheme as ε-scaling. No quantitative analysis of the convergence speed or ε-scaling for
the Sinkhorn algorithm is given in [32].
Using entropy regularization to approximately solve the linear assignment problem is also
studied in [47].
Limitations of Entropic Transport. Despite its considerable merits, there are some fundamental constraints to the naive entropy regularization approach. The entropy term introduces
blur in the optimal coupling, increasing with regularization. Sometimes this blur may be beneficial, such as in machine learning related classification tasks [22]. But in the majority of
applications, blur is considered a nuisance (e.g. it quickly smears distinct features in gradient
flows), and one would like to run the scaling algorithm with as little regularization as possible.
However, a standard implementation has some major numerical limitations, becoming increasingly severe as the regularization approaches zero. The diagonal scaling factors diverge in the
limit of vanishing regularization, leading to numerical overflow and instabilities. Moreover, the
algorithm requires an increasing number of iterations to converge. In practice this can often
be remedied by ε-scaling, but its efficiency is not yet well understood theoretically. Therefore,
3
numerically this limit is difficult to reach. In addition, naively storing the dense kernel matrix
requires just as much memory as storing the full cost matrix in standard linear programming
solvers and multiplications with the kernel matrix become increasingly slow. Thus, smart tricks
to avoid storing of and multiplication by the dense kernel matrix have been conceived, such as
efficient Gaussian convolutions or approximation by a pre-factored heat kernel [49]. However,
these remedies only work for rather particular (although relevant) problems, and do not solve
the issues of blur and diverging scaling factors.
1.2
Contribution and Outline
Contribution. The contributions presented in this article are twofold. Throughout Sect. 3 we
propose an enhanced variant of the classical Sinkhorn algorithm to remedy the issues of entropy
regularized optimal transport discussed in Sect. 1.1. These adaptions also apply to more general
scaling algorithms for other transport-type problems, as presented in [19].
(i) We define a log-domain formulation of the scaling algorithm which avoids numerical overflow of the diagonal scaling factors but largely preserves the efficient matrix multiplication
structure. This allows to robustly run the algorithm at small regularization parameters.
(ii) The already well-known and widely used ε-scaling heuristic (e.g. [32]) is used to reduce the
number of required iterations.
(iii) We sparsify the kernel matrix by adaptive truncation, to reduce memory demand and
accelerate iterations. We quantify the error induced by truncation and propose a truncation
scheme which reliably yields small error bounds that are easy to evaluate.
(iv) Finally, a multi-scale representation of the transport problem is considered. This serves two
purposes: First, it allows for a more efficient computation of the truncated kernel. Second,
combining a coarse-to-fine approach with simultaneous ε-scaling drastically reduces the
number of variables during early stages of ε-scaling, without losing much precision.
We will point out how each modification builds on the previous ones (Remark 5.1). Combining
all four allows to solve a wide family of large transport-type problems with significantly less
runtime, memory and regularization, as compared to the naive algorithm.
In Sect. 4 we develop an alternative convergence analysis of the Sinkhorn algorithm, different
from the Hilbert metric approach given in [26]. Our eventual goal is a better theoretical understanding of the ε-scaling heuristic for the Sinkhorn algorithm. For this we revisit the analogy
between the Sinkhorn and the auction algorithm, discussed in [32], and make a more quantitative comparison. We introduce a slightly modified asymmetric variant of the Sinkhorn iterations,
that more closely mimics the behaviour of the auction algorithm and adapt the classical complexity analysis of the auction algorithm to this variant. The obtained asymptotic bound is in
good agreement with numerical experiments. We then prove stability of optimal dual solutions
of entropy regularized OT under changes of the regularization parameter. This complements
results of [20] and is a step towards proving the efficiency of ε-scaling for the Sinkhorn algorithm.
We study a simple counter example to show that additional assumptions on the problem will
be required for a full proof. To our knowledge, these are the first theoretical results towards
ε-scaling for the Sinkhorn algorithm.
Remark 1.1 (Comparison with [19]). The framework of diagonal scaling algorithms for entropic
transport-type problems, as summarized in Sect. 2, was introduced in [19]. [19] focusses on theoretical analysis of the algorithms in a continuous setting, establishing for example existence of
4
iterates and convergence. In this article we focus on efficient numerical implementation of discretized problems. For self-containedness, the log-domain stabilization was also described in [19],
but it only becomes applicable to larger problems in combination with the other modifications
of Sect. 3.
Remark 1.2 (Comparison with [26] and [47]). The projective Hilbert metric, used in [26] to
measure convergence, is very different from typical numerical choices, such as L1 or L∞ marginal
errors. Therefore, an iteration estimate based on [26] does not accurately reflect practical performance of the algorithm. The analysis of Sect. 4 seems to be more suitable for this aspect (see
also Remark 4.11 and Sect. 5.2).
In [47] a deformed Sinkhorn iteration is introduced, which in our context corresponds to
gradually decreasing ε during the iterations. Asymptotic convergence of the primal iterates to
an unregularized optimizer is shown, when ε is decreased slowly enough [47, Thm. 3.6]. However,
this estimate is based on the convergence analysis by [26] and therefore requires an exponential number of iterations. In addition, a log-domain stabilization of the Sinkhorn algorithm is
proposed. Compared to our approach it does however involve more numerical overhead (see
Sect. 3.1).
Remark 1.3 (Comparison with [44]). Intuitively, the kernel truncation and coarse-to-fine scheme
of Algorithm 3.4 are adaptions of the ideas in [44] to the diagonal scaling algorithms, but the
related analysis is rather different. On one hand, due to the ‘true’ sparsity of linear programming
(LP) solutions, in [44] convergence to the exact optimal solution can be shown. Whereas the
scaling algorithm only solves the approximate regularized problem (up to a virtually negligible
truncation error), and in pathological cases it might not fully converge (Example 3.5). However,
in our experiments non-convergence was no issue and the approximation quality can be considered sufficient (Fig. 4). We observe that the sparse LP solver typically requires less memory (see
Fig. 3). On the other hand, the scaling algorithms are much easier to adapt to new cost functions,
since no shielding neighbourhoods must be constructed, but mere (somewhat accurate) hierarchical lower bounds suffice. It is also much easier to generalize to other transport-type problems
which is non-trivial for the linear programming solver. Moreover, the scaling algorithms seem
to scale more favourably with problem size (Fig. 3). We believe that these advantages outweigh
the difficulties in many practical applications.
Outline. The paper is organized as follows: The introduction is concluded with notational
conventions and preliminaries (Sect. 1.3). Then, we start in Sect. 2 by recalling the framework
for transport-type problems, as well as the scaling algorithms for their entropy regularized variants, as presented in [19]. The various proposed adaptions to the basic scaling algorithm are
discussed in Sect. 3. The comparison between the Sinkhorn algorithm and the auction algorithm
is presented throughout Sect. 4. A numerical study of the efficiency of the various modifications
and comparison to the analysis of Sect. 4, as well as various examples to illustrate the versatility
of the diagonal scaling algorithms are given in Sect. 5. We conclude in Sect. 6.
1.3
Notation and Preliminaries
Throughout this article, we will consider transport problems between two discrete finite spaces
X and Y . For a discrete, finite space Z (typically X, Y or X × Y ) we identify functions and
measures over Z with vectors in R|Z| , which we simply denote by RZ . For v ∈ RZ , z ∈ Z we
write v(z) for the component of v corresponding to z (subscript notation is reserved for other
5
purposes). The standard Euclidean inner product is denoted by h·, ·i. The sets of vectors with
Z
positive and strictly positive entries are denoted by RZ
+ and R++ . The probability simplex over
Z
def.
Z is denoted by P(Z). We write R = R ∪ {−∞, +∞} for the extended real line and R for the
space of vectors with possibly infinite components.
For a, b ∈ RZ the operators and denote pointwise multiplication and division, e.g. a b ∈
def.
Z
R , (a b)(z) = a(z)·b(z) for z ∈ Z. The functions exp and log are extended to RZ by pointwise
def.
application to all components: exp(a)(z) = exp(a(z)). We write a ≥ b if a(z) ≥ b(z) for all
z ∈ Z, a ≥ 0 if a(z) ≥ 0 for all z ∈ Z (and likewise for ≤, > and <). For a ∈ R, aZ denotes the
vector in RZ with all entries being a. We write max a and min a for the maximal and minimal
entry of a.
def. P
For µ ∈ RZ and a subset A ⊂ Z we also use the notation µ(A) =
z∈A µ(z), analogous
to measures. We say µ ∈ RZ is absolutely continuous w.r.t. ν ∈ RZ
and
write µ ν when
+
[ν(z) = 0] ⇒ [µ(z) = 0]. This is the discrete special case of absolute continuity for measures.
def.
The set spt µ = {z ∈ Z : µ(z) 6= 0} is called support of µ.
The power set of Z is denoted by 2Z .
For a subset C ⊂ RZ the indicator function of C over RZ is given by
(
0
if v ∈ C ,
Z
ιC : R → R ,
v 7→
+∞ else.
In particular, for v, w ∈ RZ one finds ι{v} (w) = 0 if v = w and +∞ otherwise. For v ∈ RZ we
introduce the short notation
(
0
if w(z) ≤ v(z) for all z ∈ Z,
ι≤v : RZ → R ,
w 7→
+∞ else.
Moreover, we merely write ι+ for ιRZ .
+
The projection matrices PX ∈ RX×(X×Y ) and PY ∈ RY ×(X×Y ) are given by
(
(
0,
1
if
x
=
x
1 if y = y 0 ,
def.
def.
PX (x, (x0 , y 0 )) =
PY (y, (x0 , y 0 )) =
0 else.
0 else.
They act on some π ∈ RX×Y as follows:
X
(PX π)(x) =
π(x, y) = π({x} × Y ),
(PY π)(y) =
y∈Y
X
(1.1a)
π(x, y) = π(X × {y}).
x∈X
(1.1b)
That is, they give the X and Y marginal in the sense of measures. Conversely, for some v ∈ RX ,
w ∈ RY we find
(P>
X v)(x, y) = v(x) ,
(P>
Y w)(x, y) = w(y) .
(1.1c)
Definition 1.4 (Kullback-Leibler Divergence). For µ, ν ∈ RZ the Kullback-Leibler divergence
of µ w.r.t. ν is given by
P
z∈Z: µ(z) log µ(z) − µ(Z) + ν(Z) if µ, ν ≥ 0, µ ν ,
ν(z)
def.
µ(z)>0
KL(µ|ν) =
(1.2)
+∞
else.
6
We write KL∗ for the convex conjugate w.r.t. the first argument and find
X
KL∗ (α|ν) =
(exp(α(z)) − 1) · ν(z) .
(1.3)
z∈Z
Throughout the article the KL divergence plays a central role and is used on various different base
spaces. Sometimes, when referring to the KL divergence on a space Z, we will add a subscript
KLZ for clarification.
Definition 1.5 (KL Proximal Step). For a convex function f : RZ → R and a step size τ > 0
the proximal step operator for the Kullback-Leibler divergence is given by
prox1/τ f : RZ → RZ ,
µ 7→ argmin τ1 KL(ν|µ) + f (ν) .
(1.4)
ν∈RZ
A unique minimizer exists, if there is some ν ∈ RZ , ν µ such that f (ν) 6= ±∞. Throughout
this article we shall always assume that this is the case.
For Sect. 4 we require the following Lemma.
Lemma 1.6 (Softmax and Softmin). For a parameter ε > 0 and a ∈ RZ let
!
!
X
X
def.
def.
softmax(a, ε) = ε log
exp(a(z)/ε) , softmin(a, ε) = −ε log
exp(−a(z)/ε) .
z∈Z
z∈Z
For ε, λ > 0 and a, b ∈ RZ one has the relations
max(a) ≤ softmax(a, ε) ≤ max(a) + ε log |Z|,
(1.5a)
min(a) − ε log |Z| ≤ softmin(a, ε) ≤ min(a),
(1.5b)
min(a − b) − λ log |Z| ≤ softmax(a, ε) − softmax(b, λ) ≤ max(a − b) + ε log |Z|,
(1.5c)
min(a − b) − ε log |Z| ≤ softmin(a, ε) − softmin(b, λ) ≤ max(a − b) + λ log |Z|.
(1.5d)
Proof. The first line follows immediately from 0 ≤ exp(a(z)/ε) ≤ exp(max a/ε). Line three then
follows from min(a−b) ≤ max(a)−max(b) ≤ max(a−b). The second and fourth line are implied
by softmin(a, ε) = − softmax(−a, ε).
We assume that the reader has a basic knowledge of convex optimization, such as convex
conjugation, Fenchel-Rockafellar duality and primal-dual gaps (see for example [13, 5]).
2
2.1
Entropy Regularized Transport-Type Problems and Diagonal
Scaling Algorithms
Transport-Type Problems
In [19] a family of transport-type optimization problems with a common functional structure
was introduced, encompassing standard optimal transport, unbalanced transport formulations
and gradient flows. The general structure is given in the following definition.
7
Definition 2.1 (Generic Transport-Type Problem). For two convex marginal functions FX :
X×Y
the primal transport-type problem is
RX → R, FY : RY → R and a cost function c ∈ R
given by:
min E(π)
π∈RX×Y
def.
E(π) = FX (PX π) + FY (PY π) + hc, πi + ι+ (π)
with
(2.1a)
The corresponding dual problem is given by:
max
(α,β)∈(RX ,RY
)
J(α, β)
>
J(α, β) = −FX∗ (−α) − FY∗ (−β) − ι≤c (P>
X α + PY β)
def.
with
(2.1b)
>
The indicator function ι≤c (P>
X α + PY β) denotes the classical optimal transport dual constraint
α(x) + β(y) ≤ c(x, y) for all (x, y) ∈ X × Y (see Section 1.3).
This structure can be extended to multiple couplings to describe barycenter and multimarginal problems (see [7, 19] for details). The standard optimal transport problem is obtained
as a special case.
Definition 2.2 (Standard Optimal Transport). For marginals µ ∈ P(X), ν ∈ P(Y ) and a cost
X×Y
function c ∈ R
the standard optimal transport problem is obtained from Def. 2.1 by setting
def.
def.
FX = ι{µ} , FY = ι{ν} . The primal and dual functional are given by:
E(π) = ι{µ} (PX π) + ι{ν} (PY π) + hc, πi + ι+ (π)
(2.2a)
>
J(α, β) = hα, µi + hβ, νi − ι≤c (P>
X α + PY β)
(2.2b)
The set
Π(µ, ν) = {π ∈ RX×Y
: PX π = µ, PY π = ν}
+
def.
(2.3)
is called the couplings between µ and ν. If E(π) < ∞, then π ∈ Π(µ, ν).
Let π † and (α† , β † ) be a pair of primal and dual optimizers. If the optimal value is finite,
then the following relation holds (see e.g. [53, Thm. 5.10]):
[π † (x, y) > 0] ⇒ [α† (x) + β † (y) = c(x, y)]
(2.4)
The general framework of Def. 2.1 also allows the formulation of unbalanced transport problems, where the hard marginal constraints of Def. 2.2 are replaced by soft constraints. This
allows meaningful comparison between measures of different total mass. Such formulations were
studied for example in [35] (see also [19] for more context). A particularly relevant case for the
soft constraints is the Kullback-Leibler divergence.
Definition 2.3 (Unbalanced Optimal Transport with KL Fidelity). For marginals µ ∈ P(X),
X×Y
ν ∈ P(Y ), a cost function c ∈ R
and a weight λ > 0 the unbalanced transport problem with
KL fidelity is given by:
E(π) = λ · KL(PX π|µ) + λ · KL(PY π|ν) + hc, πi + ι+ (π)
∗
∗
J(α, β) = −λ · KL (−α/λ) − λ · KL (−β/λ) −
8
ι≤c (P>
X
α+
P>
Y
(2.5)
β)
(2.6)
When X = Y is a metric space with metric d, for λ = 1 and the cost function c = d2 , the square
root of the optimal value of (2.5) yields the so called Gaussian Hellinger-Kantorovich (GHK)
distance on RX
+ , introduced in [35]. Similarly, for the cost function
(
− log [cos(d(x, y))]2 if d(x, y) < π/2
def.
c(x, y) =
(2.7)
+∞
else.
one obtains the Wasserstein-Fisher-Rao (WFR) distance (or Hellinger-Kantorovich distance),
introduced independently and simultaneously in [31, 17, 35]. WFR is the length distance induced
by GHK [35].
2.2
Entropy Regularization and Diagonal Scaling Algorithms
Now we apply entropy regularization to the above transport-type problems (see Sect. 1.1 for
references) and replace the non-negativity constraint in (2.1a) by the Kullback-Leibler divergence.
For this we need to select a suitable reference measure ρ ∈ RX×Y
. We then replace the term
+
ι+ (π) in (2.1a) by ε · KL(π|ρ), where ε > 0 is a regularization parameter. Then one typically
‘pulls’ the linear cost term into the KL divergence:
hc, πi + ε KL(π|ρ) = ε KL(π|K) + ε · ρ(X × Y ) − K(X × Y )
where K ∈ RX×Y
with
+
def.
K(x, y) = exp(−c(x, y)/ε) · ρ(x, y) .
(2.8)
with the convention exp(−∞) = 0. K is called the kernel associated with c and the regularization
parameter ε. For convenience we formally introduce the function
getK : R++ → RX×Y ,
ε 7→ exp(−c/ε)
ρ.
(2.9)
We obtain the regularized equivalent to Def. 2.1.
Definition 2.4 (Regularized Generic Formulation).
min
π∈M+ (X×Y )
max
(α,β)∈(RX ,RY )
E(π)
J(α, β)
def.
E(π) = FX (PX π) + FY (PY π) + ε KL(π|K)
(2.10a)
def.
∗
∗
∗
>
>
with J(α, β) = −FX (−α) − FY (−β) − ε KL [PX α + PY β]/ε K
with
(2.10b)
Primal optimizers π † have the form
π † = diag(exp(α† /ε)) K diag(exp(β † /ε))
(2.11)
where (α† , β † ) are dual optimizers. Conversely, for dual optimizers (α† , β † ), π † constructed as
above is primal optimal [19].
Intuitively we see the relation between (2.1) and (2.10) as ε → 0. For example, the term
>
ε KL∗ [P>
X α + PY β]/ε K in (2.10b) can be interpreted as a smooth barrier function for the dual
>
constraint P>
X α + PY β ≤ c in (2.1b). We refer to Sect. 1.1 for references to rigorous convergence
results.
9
Remark 2.5 (Role of ρ). In a continuous setup, the introduction of ρ is clearly necessary, as
one requires a reference measure to take the KL divergence against. This role is sometimes
neglected in articles on the discrete Sinkhorn algorithm and simply set to ρ(x, y) = 1. However,
when discretizing an underlying continuous problem with increasingly higher resolution, it is
important to take into account the consistent discretization of ρ: as the number of discrete bins
increases, the mass of the coupling π will be split into more and more bins. If one keeps using
ρ(x, y) = 1, the mass of the reference measure increases, and – with all other parameters (such as
ε) being fixed – the regularization artifacts will increase with resolution. In unbalanced transport
problems this will lead to an incentive to increase the total mass of the coupling.
Under suitable assumptions problem (2.10b) can be solved by alternating optimization in α
and β (see [19] for details). For fixed β, consider the KL∗ -term:
X
>
K
= KL∗X (α/ε|K exp(β/ε)) +
K(x, y) (exp(β(y)/ε) − 1) .
KL∗X×Y [P>
α
+
P
β]/ε
X
Y
(x,y)∈X×Y
Note that the last term is constant w.r.t. α. Therefore, optimizing (2.10b) over α, for fixed β
corresponds to maximizing
JX (α) = −FX∗ (−α) − ε KL∗X (α/ε|K exp(β/ε)) ,
(2.12)
where K exp(β/ε) denotes standard matrix vector multiplication. The corresponding primal
problem consists of minimizing
EX (σ) = FX (σ) + ε KLX (σ|K exp(β/ε)) .
(2.13)
This is a proximal step of FX for the KL divergence with step size 1/ε (see Def. 1.5). So, by using
the PD-optimality conditions between (2.12) and (2.13) (see e.g. [5, Thm. 19.1]), for a given β
the primal optimizer σ † of (2.13) and the dual optimizer α† of (2.12) are given by
α† = ε log(σ †
σ † = proxε FX (K exp(β/ε)),
(K exp(β/ε))),
(2.14)
Analogously, partial optimization w.r.t. β for fixed α is related to KL proximal steps of FY .
Starting from some initial β (0) , we can iterate alternating optimization to obtain a sequence
β (0) , α(1) , β (1) , α(2) , . . . as follows:
def.
α(`+1) = ε log proxε FX (K exp(β (`) /ε)) [K exp(β (`) /ε)] ,
(2.15a)
def.
β (`+1) = ε log proxε FY (K > exp(α(`+1) /ε)) [K > exp(α(`+1) /ε)] .
(2.15b)
The algorithm becomes somewhat simpler when it is formulated in terms of the effective variables
def.
def.
u = exp(α/ε) ,
v = exp(β/ε) .
(2.16)
For more convenient notation we introduce the proxdiv operator of a function F and step size
1/ε:
proxdivε F : σ 7→ proxε F (σ)
σ
(2.17)
The iterations then become:
u(`+1) = proxdivε FX (K v (`) ) ,
def.
v (`+1) = proxdivε FY (K > u(`+1) ) .
def.
(2.18)
This reduces the number of evaluations of logarithm and exponential function. The primal-dual
relation (2.11) then becomes π † = diag(u† ) K diag(v † ), which is why u and v are often referred
to as diagonal scaling factors.
10
Remark 2.6. Throughout this article, we will refer to the arguments of the dual functionals
(2.1b) and (2.10b) as dual variables and denote them with (α, β). The effective, exponentiated
variables, introduced in (2.16), will be denoted by (u, v) and referred to as scaling factors.
For future reference let us state the full scaling algorithm.
Algorithm 2.1 (Scaling Algorithm).
1: function ScalingAlgorithm(ε,v (0) )
2:
K ← getK(ε)
3:
v ← v (0)
4:
repeat
5:
u ← proxdivε FX (K v)
6:
v ← proxdivε FY (K > u)
7:
until stopping criterion
8:
return (u, v)
9: end function
// compute kernel, see (2.9)
The stopping criterion is typically a bound on the primal-dual gap between dual iterates
(α, β) = ε log(u, v) and primal iterate π = diag(u) K diag(v), an error bound on the marginals
of π (for standard optimal transport) or a pre-determined number of iterations.
With alternating iterations (2.15) or (2.18) a large family of functionals of form (2.10a) can
be optimized, as long as the KL proximal steps of FX and FY can be computed efficiently. A
particularly relevant sub-family is, where FX and FY are separable and are a sum of pointwise
functions. Then the KL steps decompose into pointwise one-dimensional KL steps, see [19,
Section 3.4] for details.
Since Section 4 focusses on the special case of entropy regularized optimal transport, let us
explicitly state the corresponding functional and iterations.
Definition 2.7 (Entropic Optimal Transport). For marginals µ ∈ P(X), ν ∈ P(Y ) and a cost
X×Y
function c ∈ R
the entropy regularized optimal transport problem is obtained from Def. 2.4
def.
def.
by setting FX = ι{µ} , FY = ι{ν} (see Definition 2.2 for the unregularized functional). We find:
E(π) = ι{µ} (PX π) + ι{ν} (PY π) + ε KL(π|K)
>
J(α, β) = hα, µi + hβ, νi − ε KL∗ [P>
X α + PY β]/ε K
(2.19a)
(2.19b)
The proximal steps of FX and FY are trivial (if K has non-empty columns and rows) and we
recover the famous Sinkhorn iterations:
proxdivε FX (σ) = µ
u(`+1) = µ
σ,
proxdivε FY (σ) = ν
(K v (`) ) ,
v (`+1) = ν
σ,
(2.20a)
(K > u(`+1) ) .
(2.20b)
Remark 2.8 (Entropic OT and Alternating Projections). Minimizing the regularized primal
functional (2.19a) amounts to computing the KL projection of the kernel K onto the constraint
set Π(µ, ν), see (2.3). In addition to the interpretation as alternating maximization of the dual,
the Sinkhorn iterations can also be viewed as alternating primal KL projections onto the row
and column constraints. After each u and v iteration, we find respectively
PX diag(u(`+1) ) K diag(v (`) ) = µ ,
PY diag(u(`+1) ) K diag(v (`+1) ) = ν .
For more details we refer to [7].
11
3
Stabilized Sparse Multi-Scale Algorithm
Throughout this section we propose four adaptions to the Algorithm 2.1, that overcome the
limitations of a naive implementation outlined in Section 1.1.
3.1
Log-Domain Stabilization
When running Algorithm 2.1 with small regularization parameter ε, entries in the kernel K, and
the scaling factors u and v may become both very small and very large, leading to numerical
difficulties. However, under suitable conditions (see Section 1.1, in particular [20]), the optimal
dual variables (α, β) remain finite and have a stable limit as ε → 0. Therefore, we are looking
for a reformulation of the algorithm, such that it remains numerically applicable with small ε.
In [47] it was proposed to rewrite the Sinkhorn iterations directly in terms of the dual variables, instead of the scaling factors. For example, an update of α would be performed as follows:
ψ (`+1) (x, y) = −c(x, y) + β (`) (y)
(3.1a)
ψ̃ (`+1) (x, y) = ψ (`+1) (x, y) − max ψ (`+1) (x, y)
(3.1b)
y∈Y
X
α(`+1) (x) = ε log µ(x) − ε log
exp(ψ̃ (`+1) (x, y)/ε) · ρ(x, y) − max ψ (`+1) (x, y) (3.1c)
y∈Y
y∈Y
Subtracting the maximum from ψ (`+1) and adding it again later, avoids large arguments in the
exponential function. While this resolves the issue of extreme scaling factors, it perturbs the
simple matrix multiplication structure of the original Sinkhorn algorithm and requires many
additional evaluations of the exponential function and the logarithm in each iteration.
As an alternative (as also presented in [19], see Remark 1.1) we employ a relative, redundant
parametrization of the alternating dual optimization iterations. We write the scaling factors
(u, v), (2.16), as
u = ũ · exp(α̂/ε) ,
v = ṽ · exp(β̂/ε) .
(3.2)
Our goal is to formulate iterations (2.18) directly in terms of (ũ, ṽ), while keeping (α̂, β̂) unchanged during most iterations. The role of (α̂, β̂) is to occasionally ‘absorb’ the large values of
(u, v) such that (ũ, ṽ) remain bounded. This leads to two types of iterations: stabilized iterations, during which only (ũ, ṽ) are changed, and absorption iterations, during which (ũ, ṽ) are
absorbed into (α̂, β̂). In this way, we can combine the simplicity of the scaling algorithm in terms
of the scaling factor formulation with the numerical stability of the iterations in the log-domain
formulation.
Analogous to the function getK, (2.9), we define the stabilized kernel as
getK : RX × RY × R++ → RX×Y , (α, β, ε) 7→ diag(exp(α/ε)) getK(ε) diag(exp(β/ε)), (3.3a)
[getK(α, β, ε)](x, y) = exp − 1ε [c(x, y) − α(x) − β(y)] · ρ(x, y) .
(3.3b)
Note that the second line, (3.3b), should be used to numerically evaluate the stabilized kernel,
such that extreme values in (α, β) and c can cancel before exponentiation. With K = getK(ε)
and K = getK(α̂, β̂, ε) the following identities are readily verified:
K v = exp(−α̂/ε)
K > u = exp(−β̂/ε)
K ṽ ,
12
K> ũ .
(3.4)
Moreover, we introduce a stabilized version of the proxdiv operator:
proxdivε F : (σ, γ) 7→ proxε F (exp(−γ/ε) σ)
σ
(3.5)
Note that the regular version of the proxdiv operator, (2.17), is a special case of the stabilized
variant with γ = 0. We observe that
proxdivε F (K ṽ, α̂) = proxdivε F (K v)
proxdivε F (K> ũ, β̂) = proxdivε F (K > u)
exp(α̂/ε),
exp(β̂/ε) .
(3.6a)
(3.6b)
Now we formally state the stabilized variant of Algorithm 2.1.
Algorithm 3.1 (Stabilized Scaling Algorithm).
1: function ScalingAlgorithmStabilized(ε,α(0) ,β (0) )
2:
(α̂, β̂) ← (α(0) , β (0) )
3:
(ũ, ṽ) ← (1X , 1Y )
4:
K ← getK(α̂, β̂, ε)
5:
repeat
6:
while [kũk∞ ≤ τ ] ∧ [kṽk∞ ≤ τ ] do
7:
// stabilized iteration
8:
ũ ← proxdivε FX (K ṽ, α̂)
9:
ṽ ← proxdivε FY (K> ũ, β̂)
10:
end while
11:
// absorption iteration
12:
(α̂, β̂) ← (α̂, β̂) + ε · log(ũ, ṽ)
13:
(ũ, ṽ) ← (1X , 1Y )
14:
K ← getK(α̂, β̂, ε)
15:
until stopping criterion
16:
(α̂, β̂) ← (α̂, β̂) + ε · log(ũ, ṽ)
17:
return (α̂, β̂)
18: end function
Clearly, any subsequent application of stabilized iterations and absorption iterations in Algorithm 3.1 leads to an algorithm that is mathematically equivalent to Algorithm 2.1, in the sense
that it produces the same iterates (keep in mind (3.2–3.6)). But numerically, with finite floating
point precision, combining both types of iterations can make a significant difference. In practice
one can run several stabilized iterations in a row, occasionally checking whether (ũ, ṽ) become
too large or too small (see line 6), and perform an absorption iteration if required. This inflicts
less computational overhead than the direct log-domain formulation (3.1) and largely preserves
the simple matrix multiplication structure of the scaling algorithms.
In the definitions for the stabilized kernel, (3.3b), and proxdiv-operator, (3.5), there still
appear exponentials of the form exp(·/ε), which may explode as ε → 0. It is not immediately
clear, if the modified ‘stabilized’ iterations are in fact numerically stable. We now address this
question.
Let us first consider the case of standard optimal transport. Recall the primal interpretation
of the Sinkhorn iterations as iterative KL projections onto the row and column constraints
(Remark 2.8). Since the stabilized kernel is precisely the primal iterate at that step, it follows
that its entries are bounded from above, and therefore so are the exponents in (3.3b). Entries
13
for which the exponents are extremely small may numerically be truncated to zero. However, as
we will learn from Proposition 3.3, these truncation errors are negligible.
For the fixed marginal constraint FX = ι{µ} we find that the stabilized proxdiv-operator (3.5)
does not depend on γ and is equal to the standard proxdiv-operator of FX (see Def. 2.7):
proxdivε FX (σ, γ) = proxdivε FX (σ) = µ
σ.
This invariance stems from the ‘reparametrization invariance’ of optimal transport, as described
in the following Lemma. It implies that switching to the stabilized kernel corresponds to modifying the cost function in a way that does not affect the set of primal optimizers and merely
adds an offset to the dual optimizers, which corresponds to the relative parametrization (3.2).
Lemma 3.1 (Reparametrization of Optimal Transport). For given marginals µ ∈ P(X), ν ∈
|X|×|Y |
>
P(Y ), a cost function c ∈ R
, and two functions (α̂, β̂) ∈ (RX , RY ) let ĉ = c − P>
X α̂ − PY β̂.
For the marginals µ, ν and cost functions c and ĉ, consider the corresponding standard optimal
transport problems (Def. 2.2), and their regularized variants, Def. 2.7. Assume that the optimal
values are finite.
Then, (π † , (α† , β † )) are primal and dual optimizers of the (un-)regularized problem for cost
function c, if and only if (π † , (α† − α̂, β † − β̂)) are primal and dual optimizers for the (un-)
regularized problem with modified cost ĉ.
Proof. If the primal-dual gap for π † and (α† , β † ) vanishes on the problem with cost function c,
we find that it vanishes for π † and (α† − α̂, β † − β̂) on the problem with cost function ĉ, and vice
versa. This hinges on the fact that the marginals of feasible π † are fixed to µ and ν.
Now we turn to more general transport-type problems, as laid out in Definition 2.4. For
meaningful problems, FX and FY will penalize unbounded growth of mass in the primal iterates π
and Proposition 3.3 still holds. Consequently we can argue as above for the numerical robustness
of the stabilized kernel. In the general case there is no invariance as given in Lemma 3.1 and the
stabilized proxdiv-operator (3.5) depends on γ. In the examples studied in Section 5 and those
given in [19] we find however, that evaluation of the exponential exp(−γ/ε) can be avoided.
It should be noted that one will face similar questions when trying to generalize the maxargument trick in (3.1) to more general scaling algorithms.
3.2
ε-Scaling
It is empirically and theoretically well-known (cf. Section 1.1) that convergence of Algorithm
2.1 becomes slow as ε → 0. A popular heuristic remedy is the so-called ε-scaling, where one
subsequently solves the regularized problem with gradually decreasing values for ε. Let E =
(ε1 , ε2 , . . . , εn ) be a list of decreasing positive parameters. We extend Algorithm 3.1 as follows:
Algorithm 3.2 (Scaling Algorithm with ε-Scaling).
1: function ScalingAlgorithmεScaling(E,α(0) ,β (0) )
2:
(α, β) ← (α(0) , β (0) )
3:
for ε ∈ E do
// iterate over list, form largest to smallest
4:
(α, β) ← ScalingAlgorithmStabilized(ε,α,β)
5:
end for
6:
return (α, β)
7: end function
14
Note that the dual variable β is kept constant while changing ε, not the scaling factor v.
This is because the optimal dual variables (α, β) usually have a stable limit as ε → 0, while the
scaling factors (u, v) typically diverge (see Sect. 1.1 and also Proposition 4.22). So far, virtually
no theoretical results are available for ε-scaling of the Sinkhorn algorithm (e.g. Remark 1.2). We
work towards a complexity analysis in Sect. 4.
3.3
Kernel Truncation
Storing the dense kernel K and computing dense matrix multiplications during the scaling iterations (2.18) requires a lot of memory and time on large problems. For several problems with
particular structure, remedies have been proposed (Sect. 1.1). But these do not comprise nonstandard cost functions, as the one used for the Wasserstein-Fisher-Rao distance, (2.7). Moreover
they are not compatible with the log-stabilization (Section 3.1), thus a certain level of blur cannot
be avoided. We are looking for a more flexible method to accelerate solving.
For many unregularized transport problems the optimal coupling π † is concentrated on a
sparse subset of X × Y . In fact, this is the underlying mechanism for the efficiency of most
solvers discussed in Section 1.1.
For the regularized problems the optimal coupling will usually be dense. This is due to the
diverging derivative of the KL divergence at zero. However, as ε → 0, the optimal coupling
quickly converges to an unregularized solution (see Sect. 1.1, in particular [20, Thm. 5.8]). As
ε → 0, large parts of the coupling will approach zero exponentially fast.
So while we will not be able to exactly solve the full problem, by looking at suitable sparse
sub-problems, we may still expect to obtain a reasonable approximation.
Let us formalize the concept of a sparse sub-problem.
Definition 3.2 (Sparse Sub-Problems). Let FX and FY be marginal functions and c be a cost
function as in Definition 2.1 and let N ⊂ X × Y . We introduce:
(
(
c(x, y) if (x, y) ∈ N ,
K(x, y) if (x, y) ∈ N ,
def.
def.
ĉ(x, y) =
K̂(x, y) =
(3.7)
+∞
else.
0
else.
We call problems (2.1a) and (2.1b) with c replaced by ĉ the problems restricted to N . This
corresponds to adding the constraint spt π ⊂ N to the primal problem, and only enforcing the
constraint α(x) + β(y) ≤ c(x, y) on (x, y) ∈ N in the dual problem.
The entropy regularized variants of the restricted problems are obtained through replacing
K by K̂ in (2.10a) and (2.10b).
Clearly, when N is sparse, then so is K̂ and the restricted regularized problem can be solved
faster and with less memory. We now quantify the error inflicted by restriction.
Proposition 3.3 (Restricted Kernel and Duality Gap). Let ε > 0 and N ⊂ X × Y . Let E and
J be unrestricted regularized primal and dual functionals with kernel K, as given in Definition
2.4, and let Ê and Jˆ be the functionals of the problems restricted to N , with sparse kernel K̂
(see Def. 3.2).
Further, let (α, β) be a pair of dual variables, let u = exp(α/ε), v = exp(β/ε) be the corresponding scaling factors and let π = diag(u) K̂ diag(v) be the corresponding (restricted) primal
coupling.
Then we find for the primal-dual gap between π and (α, β):
X
ˆ β) +
E(π) − J(α, β) = Ê(π) − J(α,
u(x) K(x, y) v(y) .
(3.8)
(x,y)∈(X×Y )\N
15
Proof. For the primal score we find:
E(π) = FX (PX π) + FY (PY π) + ε
h
i
π(x,y)
π(x, y) log K(x,y)
− π(x, y) + K(x, y)
X
(x,y)∈X×Y
= Ê(π) + ε
X
π(x, y) log
(x,y)∈(X×Y )\N |
π(x,y)
K(x,y)
{z
=0
− π(x, y) +K(x, y)
}
Analogously, for the dual score we get:
J(α, β) = −FX∗ (−α) − FY∗ (−β) − ε
X
K(x, y) · (exp([α(x) + β(y)]/ε) − 1)
(x,y)∈X×Y
ˆ β) − ε
= J(α,
X
(x,y)∈(X×Y )\N
K(x, y) · exp([α(x) + β(y)]/ε) −1
|
{z
}
=u(x) v(y)
Together we obtain:
ˆ β) + ε
E(π) − J(α, β) = Ê(π) − J(α,
X
K(x, y) · (1 + u(x) v(y) − 1)
(x,y)∈(X×Y )\N
That is, the primal-dual gap for the original full functionals is equal to the gap for the
truncated functionals plus the ‘mass’ that we have chopped off by truncating K to K̂, when
using the scaling factors u and v. If some N were known, on which most mass of the optimal
π † is concentrated, it would be sufficient to solve the problem restricted to N , to get a good
approximate solution. The remaining challenge is, how to identify N without knowing π † before.
We propose an iterative re-estimation of N , based on current dual iterates and to combine
this with the log-stabilized iteration scheme (Section 3.1) and the computation of the stabilized
kernel, (3.3b). For a threshold parameter θ > 0 we define the following functions:
def.
getN (α, β, ε, θ) = {(x, y) ∈ X × Y : exp(− 1ε [c(x, y) − α(x) − β(y)]) ≥ θ}
(3.9)
(
exp(− 1ε [c(x, y) − α(x) − β(y)]) ρ(x, y) if (x, y) ∈ getN (α, β, ε, θ) ,
def.
[getK̂(α, β, ε, θ)](x, y) =
0
else.
(3.10)
The function getK̂ can be used instead of getK in Algorithm 3.1. We refer to this as absorption
iteration with truncation. For this combination one finds a simple bound for the primal-dual gap
comparison of Proposition 3.3:
Proposition 3.4 (Simple Duality Gap Estimate for Absorption Iterations with Truncation).
For a regularized problem as in Definition 2.4 with functionals E and J, let (u, v) be a pair of
diagonal scaling factors and (α, β) = ε · log(u, v), let (α̂, β̂) a pair of dual variables and (ũ, ṽ) a
pair of relative scaling factors such that
u = ũ · exp(α̂/ε) ,
v = ṽ · exp(β̂/ε) .
Let further N = getN (α̂, β̂, ε, θ), K = getK̂(α̂, β̂, ε, θ), let Ê and Jˆ be the functionals restricted
to N (see Definition 3.2) and let π = diag(ũ) K diag(ṽ). Then
ˆ β) + kũk∞ · kṽk∞ · θ · ρ(X × Y )
E(π) − J(α, β) ≤ Ê(π) − J(α,
16
(3.11)
Proof. By virtue of Proposition 3.3
ˆ β) +
E(π) − J(α, β) = Ê(π) − J(α,
X
u(x) K(x, y) v(y) .
(x,y)∈(X×Y )\N
For (x, y) ∈ (X × Y ) \ N one has exp(− 1ε [c(x, y) − α̂(x) − β̂(y)]) < θ and therefore
u(x) K(x, y) v(y) = ũ(x) exp − 1ε [c(x, y) − α̂(x) − β̂(y)] · ρ(x, y) · ṽ(y)
≤ ũ(x) ṽ(y) θ ρ(x, y) .
The result follows by bounding ũ(x) ≤ kũk∞ , ṽ(y) ≤ kṽk∞ and summing over (X × Y ) \ N .
This implies that in Algorithm 3.1 with truncation the additional duality gap error due to
the sparse kernel is bounded by
kũ(`) k∞ · kṽ (`) k∞ · θ · ρ(X × Y ) .
In particular, before every stabilized iteration the error is bounded by τ 2 · θ · ρ(X × Y ) and after
every absorption iteration it is bounded by θ · ρ(X × Y ). This bound is easy to evaluate and does
not require to sum over (X × Y ) \ N , as the exact expression in Proposition 3.3. We find that in
practice this truncation error bound can be kept much smaller than the remaining primal-dual
ˆ β).
gap Ê(π) − J(α,
We point out that in general the stabilized iteration scheme with truncation might not converge. An illustration is given in Example 3.5. However, as we can infer from Proposition 3.4,
if one regularly performs an absorption iteration before kũ(`) k∞ · kṽ (`) k∞ becomes too large,
the potential oscillations in the primal iterates and primal and dual functionals are numerically
negligible.
Example 3.5 (Non-Convergence of Adaptive Sparse Scheme). For some ε > 0, C > 0 let
1 1
>
>
0 C
1
1 >
µ=ν= 2 2 , c=
, ρ = 21 12 , α(0) = 0 −C , β (0) = 0 C .
C 0
2
2
For θ ∈ (exp(−2 C/ε), 1] we find getN (α(0) , β (0) , ε, θ) = {(1, 1), (1, 2), (2, 2)}. The corresponding
getK̂(α(0) , β (0) , ε, θ) does not have total support [48], since the entry (1, 2) does not lie on a
positive diagonal. Consequently, the dual iterates diverge during the Sinkhorn algorithm. After
` iterations one obtains (up to a constant shift):
α(`) = 0 −C + ε log(2`)
>
β (`) = 0 C − ε log(2` + 1)
,
>
.
So for ` ≥ 21 exp(2 C/ε) the initial situation is approximately reversed and getN (α(`) , β (`) , ε, θ) =
{(1, 1), (2, 1), (2, 2)}. So for a suitable frequency of re-estimating N , the algorithm can be made
to oscillate between these two states. When N is re-estimated sufficiently often, oscillations in
π (`) are small, and oscillations in (α(`) , β (`) ) only reflect the degeneracy of the unregularized dual
optimal set (cf. Example 4.21).
17
3.4
Multi-Scale Scheme
Finally, we propose to combine the stabilized sparse iterations with a hierarchical multi-scale
scheme, analogous to the ideas in [37, 45, 38].
This serves two purposes: First, a hierarchical representation of the problem allows to determine the truncated sparse stabilized kernel getK̂, (3.10), with a coarse-to-fine tree search,
without explicitly testing all pairs (x, y) ∈ X × Y . The second reason is to make the combination of ε-scaling (Algorithm 3.2) with the truncated stabilized scheme more efficient. For a
fixed threshold θ, while ε is large, the support of the truncated kernel getK̂ will contain many
variables. At the same time, due to the blur induced by the regularization, the primal iterates
will not provide a sharply resolved assignment. Solving the problems with large ε-value on a
coarser grid reduces the number of required variables, without losing much spatial accuracy. As
ε decreases, so does the number of variables in getK̂ (since the exponential function decreases
faster), and the resolution of X and Y can be increased. Therefore, it is reasonable to coordinate the reduction of ε with increasing the spatial resolution of the transport problem, until the
desired regularization and resolution are attained.
We will now briefly recall the hierarchical representation of a transport problem from [45].
Definition 3.6 (Hierarchical Partition and Multi-Scale Measure Approximation [45]). For a
discrete set X a hierarchical partition is an ordered tuple (X0 , . . . , XI ) of partitions of X where
X0 = {{x} : x ∈ X} is the trivial partition of X into singletons and each subsequent level is
generated by merging cells from the S
previous level, i.e. for i ∈ {1, . . . , I} and any x ∈ Xi there
exists some X̂ ⊂ Xi−1 such that x = x̂∈X̂ x̂. For simplicity we assume that the coarsest level is
the trivial partition into one set: XI = {X}. We call I S
> 0 the depth of X .
This implies a directed tree graph with vertex set Ii=0 Xi . For i, j ∈ {0, . . . , I}, i < j we
say x ∈ Xi is a descendant of x0 ∈ Xj when x ⊂ x0 . We call x a child of x for i = j − 1, and a
leaf for i = 0.
For some µ ∈ RX its multi-scale
S measure approximation is the tuple (µ0 , . . . , µI ) of measures
µi ∈ RXi defined by µi (X̂ ) = µ( x∈X̂ x) for all subsets X̂ ⊂ Xi and i = 0, . . . I.
For convenience we often identify X with the finest partition level X0 , the set of singletons,
and µ with µ0 .
Definition 3.7 (Hierarchical Dual Variables and Costs [45]). Let X and Y be discrete sets with
hierarchical partitions X = (X0 , . . . , XI ), Y = (Y0 , . . . , YI ) of depth I, let α ∈ RX and β ∈ RY
be functions over X and Y , and let c ∈ RX×Y be a cost function.
Then we define the extension α̂ = (α̂0 , . . . , α̂I ) of α onto the full partition X by
(
α(x)
if i = 0 and x = {x} for some x ∈ X,
α̂i (x) = max α(x) =
(3.12)
0
x∈x
maxx0 ∈children(x) α̂i−1 (x ) if i > 0,
for i ∈ {0, . . . , I} and x ∈ Xi and analogous for β̂ and β. Similarly, define an extension ĉ of c by
ĉi (x, y) =
min
(x,y)∈x×y
c(x, y)
(3.13)
for i ∈ {0, . . . , I}, x ∈ Xi and y ∈ Yi .
For i ∈ {0, . . . , I}, x ∈ x ∈ Xi , y ∈ y ∈ Yi we find
ĉi (x, y) − α̂i (x) − β̂i (y) ≤ c(x, y) − α(x) − β(y) .
Now we can implement a hierarchical tree-search for getN (and analogously getK̂).
18
(3.14)
Algorithm 3.3 (Hierarchical Search for getN ).
1: function getN (α,β,ε,θ)
2:
(α̂, β̂) ← hierarchical extensions of (α, β)
3:
N ← ScallCell(α̂,β̂,ε,θ,I,{X},{Y })
4: end function
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
// see (3.12)
// call on coarsest partition level
function ScanCell(α̂,β̂,ε,θ,i,x,y)
N0 ← ∅
// temporary variable for result
if ĉi (x, y) − α̂i (x) − β̂i (y) ≤ −ε · log θ then // if cell cannot be ruled out at this level
if i > 0 then
// if not yet at finest level, check on all children
for (x0 , y0 ) ∈ children(x) × children(y) do
N 0 ← N 0 ∪ ScanCell(α̂,β̂,ε,θ,i − 1,x0 ,y0 )
end for
else
// if at finest level, add variable
N 0 ← N 0 ∪ (x × y) // recall x = {x}, y = {y} for some (x, y) ∈ X × Y at i = 0
end if
end if
return N 0
end function
From (3.14) follows directly that Algorithm 3.3 implements (3.9).
In many applications the discrete sets X and Y are point clouds in Rd and the hierarchical
partitions are 2d -trees over X and Y (see e.g. [44]). The cost function c is often originally defined
on the whole product space Rd × Rd (such as the squared Euclidean distance). For the validity of
Algorithm 3.3 it suffices if ĉi (x, y) ≤ min(x,y)∈x×y c(x, y). This allows to avoid computing (and
storing) the full cost matrix c ∈ RX×Y and the explicit minimizations in (3.13). c and lower
bounds on ĉi can be computed on-demand directly using the tree-structure.
The second purpose of the multi-scale scheme is the combination with ε-scaling. As explained
above, the purpose is to reduce the number of variables while ε is large. For an illustration see
Fig. 1. For this, we divide the list E of regularization parameters ε into multiple lists (E0 , . . . , EI ),
with the largest values in EI and the smallest (and final) values in E0 , and sorted from largest
to smallest within each Ei . Then, for every i from I down to 0 we perform ε-scaling with list
Ei at hierarchical level i, using the dual solution at each level as initialization at the next stage.
The full algorithm, combining log-stabilization, ε-scaling, kernel truncation and the multi-scale
scheme, is sketched next.
Algorithm 3.4 (Full Algorithm).
1: function ScalingAlgorithmFull((E0 , . . . , EI ),θ)
2:
i=I
3:
(α, β) ← ((0), (0))
// initialize dual variables
4:
while i ≥ 0 do
5:
// solve problem at scale i with ε-scaling over Ei
6:
for ε ∈ Ei do
// iterate over list, from largest to smallest
7:
(α, β) ← ScalingAlgorithmStabilized(i,ε,θ,α,β)
8:
end for
9:
i←i−1
10:
if i ≥ 0 then
// refine dual variables
11:
(α, β) ← RefineDuals(i,α,β)
19
ε = 1280 h2 , |N | = 57659
ε = 80 h2 , |N | = 20060
ε = 5 h2 , |N | = 5263
i =4, |N | = 225
i =2, |N | = 1253
i =0, |N | = 5263
Figure 1: ε-scaling, truncated kernels and multi-scale scheme. X = Y is a uniform onedimensional grid, representing [0, 1], |X| = 256, h = 256−1 . µ and ν are smooth mixtures
of Gaussians. Top row Density of optimal coupling π † on X 2 for various ε. |N | is the number
of variables in the truncated, stabilized kernel for fixed θ = 10−10 . As ε decreases, so does |N |,
since π † becomes more concentrated. Bottom row Optimal couplings for same ε as top row, but
for different levels i of hierarchical partitions. i and ε were chosen to keep number of variables
per x ∈ X approximately constant. For high ε (and i) |N | is now dramatically lower. While π †
is ‘pixelated’ for high i, due to blur, it provides roughly the same spatial information as the top
row. Images in third column are identical.
end if
13:
end while
14:
return (α, β)
15: end function
12:
function RefineDuals(i,α,β)
17:
(α0 , β 0 ) ← (0Xi , 0Yi )
// initialize refined duals with 0
18:
for x ∈ Xi do
19:
α0 (x) ← α(parent(x))
20:
end for
21:
for y ∈ Yi do
22:
β 0 (y) ← β(parent(y))
23:
end for
24:
return (α0 , β 0 )
25: end function
Note: ScalingAlgorithmStabilized refers to calling Algorithm 3.1 for solving the problem
at scale i, with getK replaced by getK̂, (3.10), with threshold θ, implemented according to
Algorithm 3.3. Accordingly, two arguments i and θ were added.
16:
Remark 3.8 (Hierarchical Representation of FX , FY ). To solve the problem at hierarchical scale
20
i, not only do we need a coarse version of c, as given in (3.13). In addition we need hierarchical
versions of the marginal functions FX , FY , see (2.10). An appropriate choice is often clear from
the context of the problem. For example, for an optimal transport problem between µ and ν,
see Def. 2.7, we set FXi = ι{µi } , where µi is taken from the multi-scale measure approximation
of µ (see Def. 3.6). For the unbalanced transport problem with KL fidelity, Def. 2.3, we use
FXi = λ · KLXi (·|µi ).
This completes the modifications of the diagonal scaling algorithm. Their usefulness will be
demonstrated numerically in Sect. 5.
4
Analogy between Sinkhorn and Auction Algorithm
In this section we develop a new complexity analysis of the Sinkhorn algorithm and examine
the efficiency of ε-scaling. To this end we will compare the Sinkhorn algorithm for the entropy
regularized linear assignment problem with the auction algorithm. The similarity has already
been pointed out in [32]. We will revisit this analogy and give a more quantitative comparison.
In this section we only consider the standard Sinkhorn algorithm (as opposed to general
scaling algorithms), since the auction algorithm solves the linear assignment problem and assumptions on fixed marginals µ, ν are required for our analysis.
The auction algorithm is briefly recalled in Section 4.1. In Section 4.2 we introduce an asymmetric variant of the Sinkhorn algorithm, that is more similar to the original auction algorithm
and provide an analogous worst-case estimate for the number of iterations until a given precision
is achieved. A stability result for the dual optimal solutions under change of the regularization
parameter ε is given in Section 4.3 and we discuss how it relates to ε-scaling in Sect. 4.4.
4.1
Auction Algorithm
For the sake of self-containedness, in this section we briefly recall the auction algorithm and
its basic properties. Note that compared to the original presentation (e.g. [11]) we flipped the
overall sign for compatibility with the notion of optimal transport.
In the following we consider a linear assignment problem, i.e. an optimal transport problem
between two discrete sets X, Y with equal cardinality |X| = |Y | = N where the marginals
Y
µ ∈ RX
+ , ν ∈ R+ are the counting measures. For simplicity we assume that the cost function
X×Y
c∈R
is finite and non-negative.
The main loop of the auction algorithm is divided into two parts: During the bidding phase,
elements of X that are unassigned determine their locally most attractive counterpart in Y
(taking into account the current dual variables) and submit a bid for them. During the assignment
phase, all elements of Y that received at least one bid, pick the most attractive one and change
the current assignment accordingly. A formal description is given in the following.
Algorithm 4.1 (Auction Algorithm).
1: function AuctionAlgorithm(β (0) )
2:
π ← 0X×Y
// initialize primal variable: ‘empty’ coupling
3:
β ← β (0)
4:
while π(X × Y ) < N do
5:
// bidding phase
6:
B(y) ← ∅ for all y ∈ Y
// initialize empty bid lists
7:
for x ∈ {x0 ∈ X : π({x0 } × Y ) = 0} do
// iterate over unassigned x
21
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
y ← argminy0 ∈Y [c(x, y 0 ) − β(y 0 )]
// pick some element from argmin
α(x) ← c(x, y) − β(y)
// set dual variable
B(y) ← B(y) ∪ {x}
// submit bid to y, i.e. add x to bid list of y
end for
// assignment phase
for y ∈ {y 0 ∈ Y : B(y 0 ) 6= ∅} do
// iterate over all y that received bids
π(·, y) ← 0
// set column of coupling to zero
x ← argminx0 ∈B(y) [c(x0 , y) − α(x0 )]
// find best bidder, pick one if multiple
β(y) ← c(x, y) − α(x) − ε
// update dual variable
π(x, y) ← 1
// update coupling
end for
end while
return (π, (α, β))
end function
Remark 4.1. In the above algorithm, line 9 is usually replaced by α(x) ← miny0 ∈Y \{y} [c(x, y 0 )−
β(y 0 )], which in practice may reduce the number of iterations. It does not affect the following
worst-case analysis however, therefore we keep the simpler version.
We briefly summarize the main properties of the algorithm.
Proposition 4.2. When ε > 0 and with dual initialization β (0) = 0Y , Algorithm 4.1 has the
following properties:
(i) α is increasing, β is decreasing.
(ii) After each assignment phase one finds:
α(x) + β(y) ≤ c(x, y) ,
[π(x, y) > 0] ⇒ [α(x) + β(y) ≥ c(x, y) − ε] .
(4.1)
The latter property is called ε-complimentary slackness. The primal iterate satisfies
PX π ≤ µ ,
PY π ≤ ν .
(4.2)
(iii) The algorithm terminates after at most N · (C/ε + 1) iterations, where C = max c.
Sketch of Proof. We note first, that β(y) only changes upon accepting a bid, whereupon it decreases by at least ε, which implies that α is increasing. When x ∈ X submits a bid to y ∈ Y
which gets accepted, we find directly that both properties (4.1) are satisfied and π(x, y) is changed
back to 0, before the next change in α(x) or β(y) is made. When y accepts a bid from some
other x0 ∈ X, the dual constraint at (x, y) is still respected.
From the assignment phase we see that once π(X × {y}) = 1 for some y ∈ Y , it remains 1.
As long as π(X × {y}) = 0, this y has not received a bid and therefore β(y) = 0. So while the
algorithm has not yet terminated (and we execute the bidding phase), there must be at least one
element y ∗ ∈ Y with β(y ∗ ) = 0.
If for some y ∈ Y we have β(y) < −C, it will no longer receive bids, as it can no longer be
minimal in line 8, since c(x, y ∗ ) − β(y ∗ ) ≤ C < c(x, y) − β(y). So the total number of accepted
bids is bounded by (C/ε + 1) · N . Since at least one bid is accepted by some y ∈ Y per iteration,
this bounds the total number of iterations.
For more details see for example [11].
22
From ε-complimentary slackness (4.1) we directly deduce the following result.
Corollary 4.3. Upon convergence, the primal-dual gap of π and (α, β), cf. Def. 2.2, is bounded
by
hc, πi − (hµ, αi + hν, βi) ≤ N · ε .
If c is integer and ε < 1/N , then the final primal coupling is optimal.
Remark 4.4 (Extension to Unbounded Cost Functions). For cost functions that assign c(x, y) =
+∞ to some pairs, but still allow a feasible assignment, a more general estimate for number of
decreases of β is given in [12, Lemma 5]. In the particular case of the auction algorithm this
essentially leads to an additional factor N in the bound on the number of iterations.
Remark 4.5 (ε-Scaling for the Auction Algorithm). During the auction algorithm it may happen
that several elements in X compete for the same target y ∈ Y , leading to the minimal decrease
of β(y) by ε in each iteration. This phenomenon has been dubbed ‘price haggling’ [12] and
can cause poor practical performance of the algorithm, close to the worst-case iteration bound.
The impact of price haggling can be reduced by the ε-scaling technique, where the algorithm is
successively run with a sequence of decreasing values for ε, each time using the final value of β
as initialization of the next run (see also Algorithm 3.2). With this technique the factor C/ε in
the iteration bound can essentially be reduced to a factor log(C/ε). An analysis of the ε-scaling
technique for more general min-cost-flow problems can be found in [12].
4.2
Asynchronous Sinkhorn Algorithm and Iteration Bound
We now introduce a slightly modified variant of the standard Sinkhorn algorithm, derive an
iteration bound and make a comparison with the auction algorithm. We emphasize that this
modification is primarily made to facilitate theoretical study of the algorithm and do not advocate
its merits in an actual implementation (cf. Remark 4.14).
For µ ∈ P(X), ν ∈ P(Y ) and a cost function c ∈ RX×Y
we consider the entropic optimal
+
transport problem (Def. 2.7). The reference measure ρ for regularization, see (2.8), is chosen to
be the product measure ρ(x, y) = µ(x) · ν(y).
Let us state the modified Sinkhorn algorithm.
Algorithm 4.2 (Asynchronous Sinkhorn Algorithm).
1: function AsynchronousSinkhorn(ε,v (0) ,qtarget )
2:
K ← getK(ε)
3:
v = v (0)
4:
repeat
5:
u ← µ (K v)
6:
v̂ ← ν (K > u)
7:
v ← min{v, v̂}
8:
π ← diag(u) K diag(v)
9:
q ← π(X × Y )
10:
until q ≥ qtarget
11:
return (π, (u, v))
12: end function
23
// compute kernel
// element-wise minimum
The only differences to the standard Sinkhorn algorithm (given by Algorithm 2.1 with
proxdiv-operators (2.20)) lie in line 7 and in the choice of the specific stopping criterion (see
Remark 4.11 for a discussion). In the standard algorithm one would set v ← v̂. The modification
implies that v is monotonously decreasing, which implies the following result for Algorithm 4.2 in
the spirit of Proposition 4.2. This monotonicity is crucial for bounding the number of iterations.
Remark 4.6. Throughout this section, for clarity, we enumerate the iterates u, v, as well as
the auxiliary variables v̂, π and q in Algorithm 4.2, starting with the initial v (0) and proceeding with (u(1) , v (1) , v̂ (1) , π (1) , q (1) ), . . ., similar to formulas (2.18). Moreover, we introduce the
corresponding dual variable iterates (α(`) , β (`) , β̂ (`) ) = ε · log(u(`) , v (`) , v̂ (`) ).
Proposition 4.7 (Monotonicity of Asynchronous Sinkhorn Algorithm).
(i) u and α = ε log u are increasing, v and β = ε log v are decreasing, q is increasing.
(ii) PX π ≤ µ and PY π ≤ ν. We say π is sub-feasible.
(iii) There exists some y ∗ ∈ Y such that v(y ∗ ) = v (0) (y ∗ ) for all iterations.
Proof. By construction we have v (`+1) ≤ v (`) . Consequently K v (`+1) ≤ K v (`) and thus u(`+1) ≥
u(`) and eventually v̂ (`+1) ≤ v̂ (`) .
After updating u(`+1) the row constraints are satisfied. That is PX diag(u(`+1) ) K diag(v (`) ) =
µ. Since v (`+1) is only decreased (i.e. if the corresponding column constraint is violated from
above), afterwards the iterate π (`+1) is sub-feasible.
Since v̂ (`) is decreasing, it follows that if v (`) (y) = v̂ (`) (y) for some y ∈ Y , then v (k) (y) =
(k)
v̂ (y) for all k ≥ `. Let Y (`) = {y ∈ Y : v (`) (y) = v̂ (`) (y)}. Then Y (`) ⊂ Y (`+1) . Conversely, if
y∈
/ Y (`) , then v (`) (y) < v̂ (`) (y) and therefore v (`) (y) = v (0) (y).
def.
Let now q (`) (y) = v (`) (y) [K > u(`) ](y). If y ∈ Y (`+1) , then q (`+1) (y) = ν(y) ≥ q (`) (y) (as
q (`) (y) can never exceed ν(y)). If y ∈
/ Y (`+1) , then v (`+1) (y)P= v (`) (y) = v (0) (y) and since u(`) is
(`+1)
(`)
increasing, we find q
(y) ≥ q (y). We obtain q (`+1) = y∈Y q (`+1) (y) ≥ q (`) .
When Y (`) 6= Y , there exists some y ∗ ∈ Y with y ∗ ∈
/ Y (k) , v (k) (y ∗ ) = v (0) (y ∗ ) for k ∈
{1, . . . , `}. If Y (`) = Y , then v̂ (`) = v (`) ≤ v (`−1) . By construction one has (u(`) )> K v (`−1) =
µ(X) and (u(`) )> K v̂ (`) = ν(Y ) = µ(X). So if Y (`) = Y , in fact v (`) = v (`−1) . Consequently,
there exists some y ∗ ∈ Y with v (`) (y ∗ ) = v (0) (y ∗ ) for all iterations.
Let us further investigate the increments of the dual variable iterates α(`) = ε log(u(`) ).
Lemma 4.8 (Minimal Increment of α(`) ). For ` ≥ 1 have α(`+1) − α(`) , µ ≥ ε (1 − q (`) ).
Proof. Recall that π (`) = diag(u(`) ) K diag(v (`) ), and introduce π 0(`) = diag(u(`+1) ) K diag(v (`) ).
Consider the following evaluations of the dual functional:
J(α(`) , β (`) ) = hα(`) , µi + hβ (`) , νi − ε · π (`) (X × Y ) + ε · K(X × Y )
J(α(`+1) , β (`) ) = hα(`+1) , µi + hβ (`) , νi − ε · π 0(`) (X × Y ) + ε · K(X × Y )
Note that π (`) (X × Y ) = q (`) , π 0(`) (X × Y ) = 1 and since going from α(`) to α(`+1) corresponds
to a block-wise dual maximization have J(α(`+1) , β (`) ) ≥ J(α(`) , β (`) ). The claim follows.
With these tools we can bound the total number of iterations to reach a given precision.
24
Proposition 4.9 (Iteration Bound for the Asynchronous Sinkhorn Algorithm). Initializing with
β (0) = 0Y ⇔ v (0) = 1Y , for a given qtarget ∈ (0, 1) the number of iterations n necessary to achieve
q (n) ≥ qtarget is bounded by
n≤2+
C
.
ε · (1 − qtarget )
(4.3)
where C = max c. Moreover, u(`) , µ ≤ exp(C/ε) for all iterates ` ≥ 1.
Proof. Let us look at the first ‘bid’ α(1) . With c ≥ 0 we have
α
(1)
(x) = ε log
≥ ε log
µ(x)
[K v (0) ](x)
1
!
P
y∈Y
= ε log
P
y∈Y
ν(y)
1
ν(y) · exp(−c(x, y)/ε)
!
= 0.
(4.4)
By virtue of Proposition 4.7 q (`) ≤ q (n) for ` ≤ n. With Lemma 4.8 this implies
hα(n) − α(1) , µi ≥
n−1
X
`=1
ε · (1 − q (`) ) ≥ ε · (n − 1) · (1 − q (n) )
| {z }
≥1−q (n)
and with (4.4)
hα(n) , µi ≥ ε · (n − 1) · (1 − q (n) ) .
(4.5)
From Proposition 4.7 we know that there is some y ∗ ∈ Y with v (`) (y ∗ ) = 1 ≤ v̂ (`+1) (y ∗ ) for
all iterates ` ≥ 0. So
v̂ (`+1) (y ∗ ) =
1
ν(y ∗ )
=P
≥ 1,
1
(`+1)
>
∗
(`+1) (x)] µ(x)
∗
[K u
](y )
x∈X exp − ε [c(x, y ) − α
from which we infer exp(−C/ε) · exp(α(`+1) /ε), µ ≤ 1, i.e. u(`+1) , µ ≤ exp(C/ε). With
Jensen’s inequality we eventually find α(`+1) , µ ≤ C for ` ≥ 0.
C
(n) < q
Combining this with (4.5) we obtain n ≤ 1 + ε (1−q
target we have
(n) ) . So, as long as q
n < 1+
q (n)
C
ε (1−qtarget ) .
By contraposition we know that there is some n ≤ 2 +
C
ε (1−qtarget )
such that
≥ qtarget .
And finally, we formally establish convergence of the iterates.
Corollary 4.10 (Convergence of Asymmetric Algorithm). The iterates (u(`) , v (`) ) of the asymmetric Algorithm 4.2 converge to a solution of the scaling problem and q (`) → 1.
Proof. With the upper bound u(`) , µ ≤ exp(C/ε) (Proposition 4.9) we obtain the pointwise
lower bound v (`) (y) ≥ exp(−C/ε) for all ` ≥ 0. Since v (`) is pointwise decreasing, it converges
to some limit v (∞) ≥ exp(−C/ε) > 0.
The map f : v (`) 7→ v (`+1) is continuous for v (`) > 0. With v (`) → v (∞) and v (`+1) =
f (v (`) ) → v (∞) have f (v (∞) ) = v (∞) which implies that v (∞) (together with the corresponding
u(∞) = µ (K v (∞) )) solves the scaling problem. This implies convergence of q (`) to 1.
25
Remark 4.11 (On the Stopping Criterion). The criterion q ≥ qtarget is motivated by Lemma 4.8,
to provide a minimal increment of α during iterations. 1 − q measures the mass that is still missing and is equal to the L1 error between the marginals of π and the desired marginals µ and
ν. In pathological cases the dual variables (α, β) may still be far from optimizers, even though
q ≥ qtarget (see Example 4.12). In [26, Lemma 2] linear convergence of the marginals in the
Hilbert projective metric is proven. This is a stricter measure of convergence, less prone to ‘premature’ termination. However, for small ε the contraction factor is roughly 1 − 4 exp(−C/ε),
which is impractical. The scaling O(1/ε) predicted by Proposition 4.9 is consistent with numerical observations when one uses the L1 or L∞ marginal error as stopping criterion (Sect. 5.2).
Therefore we consider the q-criterion to be a reasonable measure for convergence, as long as one
keeps 1 − qtarget δ (Example 4.12).
Example 4.12. We consider the 1 × 2 toy problem with the following parameters:
>
>
ν = 1−δ δ ,
c= 0 C ,
K = 1 − δ δ · e−C/ε
µ= 1 ,
for some C > 0, δ ∈ (0, 1) and some regularization strength ε > 0. And we consider the scaling
factors (one for X, two choices for Y ):
>
u= 1 ,
v1 = 1 1
>
,
v2 = 1 eC/ε
>
Let πi = diag(u) K diag(vi ) and corresponding total masses qi , i = 1, 2. We find:
π1 = 1 − δ δ · e−C/ε ,
q1 = 1 − δ (1 − e−C/ε ),
π2 = 1 − δ δ ,
.
q2 = 1.
π2 and (α, β2 ) = ε log(u, v2 ) are primal and dual solutions. π1 is sub-feasible (see Proposition 4.7). For fixed ε > 0, as δ → 0, q1 tends to 1 (but is strictly smaller), i.e. the pair (u, v1 ) has
almost converged in the q-measure sense, but the distance between β1 = ε log v1 and the actual
solution β2 is C.
Remark 4.13 (Analogy to Auction Algorithm). For now assume |X| = |Y | = N and µ, ν are
normalized counting measures. Then lines 5 and 6 in Algorithm 4.2, expressed in dual variables,
become
α(x) ← softmin({c(x, y) − β(y)|y ∈ Y }, ε) + ε log N ,
β̂(y) ← softmin({c(x, y) − α(x)|x ∈ X}, ε) + ε log N .
These are formally similar to the corresponding lines 9 and 16 in Algorithm 4.1. We can interpret
line 5 in Algorithm 4.2 as x not just submitting a bid to the best candidate y, but to all
candidates, weighted by the attractiveness (recall that in the Sinkhorn algorithm, a change in
the dual variable directly implies a change in the primal iterate via (2.11)). Conversely, in line
7, y does not only accept the best bid, but bids from all candidates, again weighted by price. If
there are too many bids (i.e. if the column constraint would be violated and v̂(y) < v(y)), β(y)
decreases and thereby rejects superfluous offers.
Both algorithms are decidedly different from mere alternating optimizations of the unregularized dual transport problem (2.2b): In the auction algorithm there is an additional slack ε in the
β-update, in the asynchronous Sinkhorn algorithm the min is replaced by softmin, as implied by
the regularized dual (2.10b). These modifications are crucial for convergence of the algorithms.
In both algorithms the dual variables α and β (or the scaling factors u and v) are monotonous,
as is the mass of the intermediate primal coupling π, which is not primal feasible during the
26
iterations (Propositions 4.2 and 4.7). In both algorithms there is a minimal step size for one
of the dual variables (−ε for β in the auction algorithm, see Lemma 4.8 for the asynchronous
Sinkhorn algorithm). The minimal step size is proportional to ε, leading to an iteration bound
that scales as O(ε−1 ). Thus, in both cases, a small value of ε is required to obtain accurate
solutions (see Corollary 4.3 and Sect. 1.1), but the naive algorithms (without ε-scaling) require
more iterations as ε decreases.
Two important differences are, that the iteration bound for the auction algorithm depends
on N , while Proposition 4.9 does not. Moreover, the auction algorithm terminates after a
finite number of steps, whereas the asynchronous Sinkhorn algorithm in general only converges
asymptotically and the iteration bound diverges as qtarget → 1.
Remark 4.14 (Comparison with Symmetric Algorithm). As mentioned above, Algorithm 4.2
can be turned into the standard Sinkhorn algorithm by replacing line 7 by v ← v̂. The above
analysis cannot be extended to this case, as the monotonicity properties of Proposition 4.7 are
lost. But of course the symmetric algorithm still converges. While it cannot be proven that the
symmetric iteration is always more efficient than the asymmetric one (there are very particular
exceptions), in the overwhelming majority of practical cases, we observe that the symmetric
algorithm is faster.
Despite these differences, the asymmetric variant and the above analysis provide some insight
on why convergence of the Sinkhorn algorithm becomes slower as ε → 0.
In the context of the comparison to the auction algorithm, one can interpret the standard
(symmetric) Sinkhorn algorithm such that elements of y that do not receive sufficient bids
(i.e. where v̂ > v), submit ‘counter-bids’ to X, to obtain mass more quickly. Such a symmetrization is in principle also possible for the auction algorithm. But then similarly the complexity
analysis based on monotonous dual variables is no longer feasible, and the algorithm may even
run indefinitely (see ‘down iterations’ in [12]).
4.3
Stability of Dual Solutions
The main result of this section is Theorem 4.16, which provides stability of dual solutions to
entropy regularized optimal transport (Def. 2.7) under changes of the regularization parameter
ε. Its implications for ε-scaling are discussed in Sect. 4.4.
We consider a similar setup as in Sect. 4.2: µ ∈ P(X), ν ∈ P(Y ), c ∈ RX×Y
. Again, the
+
reference measure ρ for regularization, see (2.8), is chosen to be the product measure ρ(x, y) =
µ(x) · ν(y). For Theorem 4.16 we introduce an additional assumption on µ and ν, its necessity
is discussed in Example 4.21.
Assumption 4.15 (Atomic Mass). For µ ∈ P(X), ν ∈ P(Y ) there is a positive integer M < ∞
Y
such that µ = M −1 · r, ν = M −1 · s for r ∈ ZX
++ , s ∈ Z++ , where Z++ denotes the set of strictly
positive integers.
Theorem 4.16 (Stability of Dual Solutions under ε-Scaling). Let max{|X|, |Y |} ≤ N < ∞ and
let µ and ν satisfy Assumption 4.15 for some positive integer M < ∞. For two regularization
parameters ε1 > ε2 > 0, let (α1 , β1 ) and (α2 , β2 ) be maximizers of the corresponding dual regularized optimal transport problems (Def. 2.7) and let ∆α = α2 − α1 and ∆β = β2 − β1 . Then
max ∆α − min ∆α ≤ ε1 · N · (4 log N + 24 log M ),
(4.6a)
max ∆β − min ∆β ≤ ε1 · N · (4 log N + 24 log M ).
(4.6b)
27
Note that the bounds in (4.6) do not depend on a bound on the cost function c.
Remark 4.17 (Proof Strategy). The proof requires several auxiliary definitions and lemmas.
The estimate consists of two contributions: One stems from following paths within connected
components of what we call assignment graph (defined in the following Lemma), using the primaldual relation (2.11). This reasoning is analogous to the proof strategy for ε-scaling in the auction
algorithm (see [12]). However, between different connected components (2.11) is too weak to yield
useful estimates. So a second contribution arises from a stability analysis of effective diagonal
problems (in Lemmas 4.19 and 4.20).
Lemma 4.18 (Assignment Graph). For two feasible couplings π1 , π2 ∈ Π(µ, ν) and a threshold
M −1 ∈ R the corresponding assignment graph is a bipartite directed graph with vertex sets (X, Y )
and the set of directed edges
E = {(x, y) ∈ X × Y : π2 (x, y) ≥ µ(x) · ν(y)/M }
t {(y, x) ∈ Y × X : π1 (x, y) ≥ µ(x) · ν(y)/M }
where (a, b) ∈ E indicates a directed edge from a → b.
The assignment graph has the following properties:
(i) Every node has at least one incoming and one outgoing edge.
(ii) Let X0 ⊂ X, Y0 ⊂ Y such that there are no outgoing edges from (X0 , Y0 ) to the rest of the
vertices, then |µ(X0 ) − ν(Y0 )| < 1/M . This is also true when there are no incoming edges
from the rest of the vertices. If µ and ν are atomic, with atom size 1/M (see Assumption
4.15), then µ(X0 ) = ν(Y0 ).
(iii) Let µ and ν be atomic, with atom size 1/M . Let {(Xi , Yi )}R
i=1 be the vertex sets of the
strongly connected components of the assignment graph, for some R ∈ N (taking into account
R
the orientation of the edges). Then the sets {Xi }R
i=1 and {Yi }i=1 are partitions of X and
Y , and µ(Xi ) = ν(Yi ) for i = 1, . . . , R.
P
Proof. Assume, a node x ∈ X had no outgoing edge. Then y∈Y π2 (x, y) < µ(x)/M ≤ µ(x).
This contradicts π2 ∈ Π(µ, ν). Existence of incoming edges follows analogously.
Let X̂0 = X \ X0 , Ŷ0 = Y \ Y0 . If (X0 , Y0 ) has no outgoing edges, then
X
π2 (x, y) <
(x,y)∈X0 ×Ŷ0
X
(x,y)∈X̂0 ×Y0
X
µ(x) · ν(y)/M ≤
1
,
M
µ(x) · ν(y)/M ≤
1
.
M
(x,y)∈X0 ×Ŷ0
π1 (x, y) <
X
(x,y)∈X̂0 ×Y0
Since π1 , π2 ∈ Π(µ, ν), the first line implies µ(X0 ) < ν(Y0 ) + 1/M and the second line implies
ν(Y0 ) < µ(X0 )+1/M , i.e. |µ(X0 )−ν(Y0 )| < 1/M . With Assumption 4.15 for atom size 1/M , this
implies µ(X0 ) = ν(Y0 ). The statement about incoming edges follows from µ(X̂0 ) = 1 − µ(X0 )
and ν(Ŷ0 ) = 1 − ν(Y0 ).
Every node in (X, Y ) is part of at least one strongly connected component (containing at
least the node itself). If two strongly connected components have a common element, they are
identical. Hence, the strongly connected components form partitions of X and Y . For some
x ∈ X (or y ∈ Y ), let Xout ⊂ X and Yout ⊂ Y be the set of nodes that can be reached
28
from x, let Xin ⊂ X and Yin ⊂ Y be the set of nodes from which one can reach x and let
(Xcon = Xout ∩ Xin , Ycon = Yout ∩ Yin ) be the strongly connected component of x. Clearly
(Xout , Yout ) has no outgoing edges. Hence, by item (ii) one has µ(Xout ) = ν(Yout ). Moreover,
(Xout \ Xin , Yout \ Yin ) has no outgoing edges, hence µ(Xout \ Xin ) = ν(Yout \ Yin ), from which
follows that µ(Xcon ) = ν(Ycon ).
R
Lemma 4.19 (Reduction to Effective Diagonal Problem). Let {Xi }R
i=1 and {Yi }i=1 be partitions
R
of X and Y , for some R ∈ N, with µ(Xi ) = ν(Yi ) for i = 1, . . . , R. Let {yi }i=1 ⊂ Y such that
yi ∈ Yi . Let (α† , β † ) be optimizers for the dual entropy regularized optimal transport problem
(Def. 2.7) for a regularization parameter ε > 0.
Consider the following functional over RR :
Jˆ : RR → R,
β̂ 7→ −ε
R
X
h
i
exp − 1ε d(i, j) + β̂(i) − β̂(j)
i,j=1
where d ∈ RR×R with
h
i
X
†
†
†
†
1
d(i, j) = −ε log
exp − ε c(x, y) − α (x) − β (yi ) − β (y) + β (yi ) · µ(x) · ν(y)
.
x∈Xi
y∈Yj
(4.7)
ˆ Conversely, if β̂ †† is a maximizer
Then β̂ † ∈ RR , given by β̂ † (i) = β † (yi ), is a maximizer of J.
††
†
ˆ
of J, then there is a constant b ∈ R, such that β̂ (i) = β̂ (i) + b for all i ∈ 1, . . . , R.
Proof. We define the functional Jˆ : RK → R as follows:
α̃
−BX
ˆ
J : β̂ 7→ J
β̂
+
BY
β̃
where J denotes the dual functional of entropy regularized optimal transport (2.19b), and
• α̃ ∈ RX with α̃(x) = α† (x) + β † (yi ) when x ∈ Xi ;
• β̃ ∈ RY with β̃(y) = β † (y) − β † (yi ) when y ∈ Yi ;
• BX ∈ RX×R with BX (x, i) = 1 if x ∈ Xi and 0 else;
• BY ∈ RY ×R with BY (y, i) = 1 if y ∈ Yi and 0 else.
Then one has
†
α̃
α
−BX
=
+
β̂ † .
BY
β†
β̃
Since maximizing Jˆ corresponds to maximizing J over an affine subspace, clearly β̂ † is a maxiˆ Since Jˆ inherits the invariance of J under constant shifts, any β̂ †† of the form given
mizer of J.
above, is also a maximizer. Consequently, we may add the constraint β̂(1) = 0, which does not
change the optimal value. With this added constraint the functional becomes strictly convex,
which implies a unique optimizer. Hence, any optimizer of the unconstrained functional can be
written in the form of β̂ †† .
29
ˆ β̂). We find
Let us now give a more explicit expression of J(
D
E
>
ˆ β̂) = BY> ν − BX
J(
µ, β̂
−ε
R X
X
h
i
exp − 1ε c(x, y) − α̃(x) + β̂(i) − β̃(y) − β̂(j) · µ(x) · ν(y)
i,j=1 x∈Xi
y∈Yj
D
E
+ hµ, α̃i + ν, β̃ + ε · K(X × Y ) .
Note that the third line is constant w.r.t. β̂. Since µ(Xi ) = ν(Yi ) the linear term vanishes and
we can write
ˆ β̂) = −ε
J(
R
X
h
i
exp − 1ε d(i, j) + β̂(i) − β̂(j) + const
i,j=1
with coefficients d ∈ RR×R , as given above. The constant offset does not affect minimization.
Lemma 4.20 (Effective Diagonal Problem and Stability). For a parameter ε > 0 and a real
matrix d ∈ RR×R consider the following functional:
Jˆε,d (β) =
R
X
exp ([−d(i, j) − β(i) + β(j)] /ε)
(4.8)
i,j=1
Minimizers of Jˆε,d exist.
Let ε1 > ε2 > 0 be two parameters and d1 , d2 ∈ RR×R two real matrices. Let β1† and β2† be
minimizers of Jˆε1 ,d1 and Jˆε2 ,d2 , let ∆d = d2 − d1 , ∆β = β2† − β1† . Let the matrix w ∈ RR×R be
given by w(i, j) = max{−∆d(i, j), ∆d(j, i)}. Then
max ∆β − min ∆β ≤ maxdiam(w) + 2 ε1 R log R ,
where
maxdiam(w) = max
(k−1
X
wji ,ji+1 : k ∈ {2, . . . , R},
i=1
)
ji ∈ {1, . . . , R} for i = 1, . . . , k, all ji distinct. .
That is, maxdiam(w) is the length of the longest cycle-less path in {1, . . . , R} with edge lengths
w.
The proofs of Theorem 4.16 and Lemma 4.20 can be found in Appendix A.
Example 4.21 (Necessity of Atomic Mass Assumption). Assumption 4.15 is in fact necessary
for Theorem 4.16. This can be illustrated by the following 2 × 2 example. Let
>
0 C
1
e−C/ε
1
1
1
1 >
µ= 2 +δ 2 −δ ,
ν= 2 2 ,
c=
,
K = −C/ε
C 0
e
1
30
for a parameter δ ∈ [0, 12 ) and some regularization strength ε > 0. (For simplicity we have chosen
the counting measure for ρ in K.) The dual optimizers of the unregularized problem are given
by (to remove the ambiguity of constant shifts, we fix α(1) = 0):
α= 0 τ
>
β = 0 −τ
,
>
,
where for δ ∈ (0, 12 ) one has τ = −C, and for δ = 0 the whole set τ ∈ [−C, C] is optimal.
Denote by (αε , βε ) the dual optimizers of the regularized problem with regularization parameter
ε (where we fix again αε (1) = 0 for uniqueness). It is possible, but tedious, to find explicit
formulas for (αε , βε ). Instead we give a qualitative discussion which we consider to provide more
insight. According to [20, Prop. 3.2], as ε → 0, (αε , βε ) converges to the centroid of the set of
dual optimizers. In our example, we find
lim αε = 0 τ
ε→0
>
lim βε = 0 −τ
,
ε→0
>
,
with τ = −C if δ > 0 and τ = 0 if δ = 0. The limit changes abruptly, depending on the
value of δ. In the regime 0 < δ e−C/ε 1, the blur introduced by entropic smoothing is
small compared to the scale of the cost function, but large compared to the small asymmetry
introduced by δ > 0. So (αε , βε ) will be close to the τ = 0 solution. But when one further
decreases ε and comes into the regime 0 < e−C/ε δ 1, one approaches the τ = −C solution.
By making δ arbitrarily small (but non-zero), one can make the transition phase between the
regimes for ε arbitrarily short, during which αε (2) changes by approximately C. Consequently,
there can be no stability result independent of C for the dual solutions without quantizing δ.
4.4
Application To ε-Scaling
Assuming that we know the dual solution for some ε1 > 0, then Theorem 4.16 allows bound the
number of iterations of Algorithm 4.2 for some smaller ε2 ∈ (0, ε1 ), independent of bounds on
the cost function c.
Proposition 4.22 (Single ε-Scaling Step). Consider the set-up of Theorem 4.16. In particular,
let ε1 > ε2 > 0 be two regularization parameters, let (α1 , β1 ) and (α2 , β2 ) be corresponding
optimizers of (2.19b).
If Algorithm 4.2 is initialized with v (0) = exp(β1 /ε), for a given qtarget ∈ (0, 1) the number of
iterations n necessary to achieve q (n) ≥ qtarget is bounded by
n≤2+
ε1 N · (4 log N + 24 log M ) + log M
.
ε2
1 − qtarget
Proof. For the optimal scaling factor u1 of the ε1 -problem we find:
def.
u1 (x) = exp(α1 (x)/ε1 ) =
X
−1
exp − ε11 [c(x, y) − β1 (y)] ν(y)
y∈Y
⇒ u1 (x)−1 ν(y)−1 ≥ exp − ε11 [c(x, y) − β1 (y)]
31
for all (x, y) ∈ X × Y
(4.9)
With this we can bound the first iterate of the ε2 -run of the algorithm by:
u(1) (x) =
X
−1
exp − ε12 [c(x, y) − β1 (y)] ν(y)
y∈Y
−1
≥
X
u1 (x)−ε1 /ε2 ν(y)−ε1 /ε2 ν(y)
≥ u1 (x)ε1 /ε2 M −ε1 /ε2
y∈Y
where we have used ν(y) ≥ 1/M , Assumption 4.15. Eventually we find
α(1) (x) ≥ α1 (x) − ε1 log M.
By monotonicity of the iterates we have β2 ≤ β (`) ≤ β (0) = β1 and β (`) (y 0 ) = β1 (y 0 ) for a
suitable y 0 ∈ Y (Proposition 4.2). Consequently max ∆β = 0. Then, from Theorem 4.16, we
obtain β2 (y) − β1 (y) ≥ min ∆β ≥ −ε1 · A where A = N · (4 log N + 24 log M ). With this we can
bound the u-iterates:
−1
X
def.
u(`) (x) ≤ u2 (x) = exp(α2 (x)/ε2 ) =
exp − ε12 [c(x, y) − β2 (y)] ν(y)
y∈Y
≤
X
−1
exp − ε12 [c(x, y) − β1 (y)] ν(y) exp εε21 A
y∈Y
With convexity of s 7→ sε1 /ε2 and Jensen’s inequality we get
u(`) (x) ≤
X
−ε1 /ε2
exp − ε11 [c(x, y) − β1 (y)] ν(y)
exp εε21 A = u1 (x)ε1 /ε2 exp εε21 A
y∈Y
and finally
α(`) (x) ≤ α1 (x) + ε1 A
Now using Lemma 4.8 and arguing as in Proposition 4.9, we find that there is some n ≤ 2 +
ε1 A+log M
(n) ≥ q
target .
ε2 1−qtarget such that q
Let now C = max c for a cost function c ≥ 0, let ε̂ > 0 be the desired final regularization
parameter, pick some λ ∈ (0, 1) and let k ∈ N such that ε̂ · λ−k ≥ C. Let E = (ε̂ · λ−k , ε̂ ·
λ−k+1 , . . . , ε̂) be a list of decreasing regularization parameters.
Now we combine Algorithm 4.2 with ε-scaling, (cf. Algorithm 3.2). For ε = ε̂ · λ−k ≥ C,
1
according to Proposition 4.9 it will take at most 2 + 1−qtarget
iterations. It is tempting to deduce
from Proposition 4.22 that for each subsequent value of ε at most 2 + λ (1−qAtarget ) iterations are
required, with A = N · (4 log N + 24 log M ) + log M . For N > 1 the total number of iterations
would then be bounded by (2 + λ (1−qAtarget ) ) · (k + 1). For fixed λ the step parameter k scales
like log(C/ε̂). Consequently, the total number of iterations would be bounded by O(log(C/ε̂))
32
w.r.t. the cost function and regularization, which would be analogous to ε-scaling for the auction
algorithm (Remark 4.5).
There is an obvious gap in this reasoning: Theorem 4.16 assumes that β1 is known exactly,
while Algorithm 4.2 only provides an approximate result. From Example 4.12 we learn that in
extreme cases this difference can be substantial and disrupt the efficiency of ε-scaling. Thus,
additional assumptions on the problem are required to make the above argument rigorous.
However, as discussed in Remark 4.11, in practice we usually observe that approximate
iterates are sufficient and we can therefore hope that ε-scaling does indeed serve its purpose.
5
Numerical Examples
Now we present a series of numerical experiments to confirm the usefulness of the modifications
proposed in Sect. 3. We show that runtime and memory usage are reduced substantially. At
the same time the adapted algorithm is still as versatile as the basic version of [19], Algorithm
2.1. But Algorithm 3.4 can solve larger problems at lower regularization, yielding very sharp
results. We give examples for unbalanced transport, barycenters, multi-marginal problems and
Wasserstein gradient flows. The code used for the numerical experiments is available from the
author’s website.1
5.1
Setup
We transport measures on [0, 1]d for d ∈ {1, 2, 3}, represented by discrete equidistant Cartesian
grids. The distance between neighbouring grid points is denoted by h. For the squared Euclidean
distance cost function c(x, y) = |x − y|2 , x, y ∈ Rd , K is a Gaussian kernel with approximate
√
width ε. Therefore, it is useful to measure ε in units of h2 . For ε = h2 the blur induced by the
entropy smoothing is on the length scale of one pixel. With the enhanced scaling algorithm we
solve most problems in this section with ε = 0.1 · h2 , leaving very little blur and giving a good
approximation of the original unregularized problem (see Fig. 4).
Unless stated otherwise, we use the following settings: Test measures are mixtures of Gaussians, with randomized means and variances. The cost function is the squared Euclidean distance.
ρ is the product measure µ ⊗ ν for optimal transport problems and the discretized Lebesgue measure on the product space for problems with variable marginals. For standard optimal transport
the stopping criterion is the L∞ error between prescribed marginals (µ, ν) and marginals of the
primal iterate π (and likewise for Wasserstein barycenters). For all other models the primal-dual
gap is used. We set θ = 10−20 for truncating the kernel and τ = 102 as upper bound for (ũ, ṽ)
(cf. (3.10), Algorithm 3.1, line 6), implying a bound of 10−16 · ρ(X × Y ) for the truncation
error, which is many orders of magnitude below prescribed marginal accuracies or primal-dual
gaps. The hierarchical partitions in the coarse-to-fine scheme are 2d -trees, where each layer i is a
coarser d-dimensional grid with grid constant hi . For combination with ε-scaling (Algorithm 3.4)
we choose the lists Ei , i > 0, such that for the smallest εi in each Ei we have roughly εi /h2i ≈ 1.
On the finest scale, we go down to the desired final value of ε. All reported run-times were
obtained on a single core of an Intel Xeon E5-2697 processor at 2.7 GHz.
33
105
104
3000 h2 /ε
100 h2 /ε
10
runtime/s
iterations
10
4
3
102
101
103
10−1
100
100
101
10−1
ε/h2
(i) log-domain
100
101
ε/h2
(ii) + ε-scaling
(iii) + truncation
(iv) + multi-scale
Figure 2: Efficiency of enhancements: average number of iterations and runtime for different ε
and algorithms. X = Y are 2-d 64 × 64 grids. (i) log-domain stabilized, Algorithm 3.1, (ii) with
ε-scaling, Algorithm 3.2, (iii) with sparse stabilized kernel, (3.10), (iv) with multi-scale scheme,
Algorithm 3.4. (ii) and (iii) need same number of iterations, but the sparse kernel requires less
time. The naive implementation, Algorithm 2.1, requires same number of iterations as (i), but
numerical overflow occurs at approximately ε ≤ 3 h2 .
5.2
Efficiency of Enhanced Algorithm
The numerical efficiency of the subsequent modifications presented in Sect. 3, applied to the
standard Sinkhorn algorithm, is illustrated in Fig. 2. While the stabilized algorithm (i) is not
yet faster than the naive implementation, it can robustly solve the problem for all given values of
ε. The required number of iterations scales like O(1/ε), in good agreement with the complexity
analysis of Sect. 4.2. With ε-scaling (ii) the number of iterations is decreased substantially.
Replacing the dense kernel with the adaptive truncated sparse kernel (iii) does not change the
number of required iterations, but saves time and memory. With the multi-scale scheme the
required number of iterations is slightly increased, since the initial dual variables obtained at a
coarser level are only approximate solutions. But by reducing the number of variables during
the early ε-scaling stages, the runtime is further decreased (cf. Fig. 1). The combination of all
modifications leads to an average total speed-up of more than two orders of magnitude on this
problem type.
A runtime benchmark and study of the sparsity of the truncated kernel are given in Fig. 3.
The runtime scales approximately linear with |X| and for large problems the algorithm becomes
faster than the adaptive sparse linear programming solver [44]. The final number of variables
in the sparse kernel is comparable with the number of variables in [44], for higher values of ε,
during scaling, more memory is required (cf. Fig. 4). This underlines again the importance of
the coarse-to-fine scheme (Sect. 3.4). It should be noted, that Fig. 2 shows results for 64 × 64
images, the smallest image size in Fig. 3. For larger images the runtime difference between (i-iv)
would be even larger, but due to time and memory constraints, only (iv) can be run practically.
The impact of different final values for ε is outlined in Fig. 4. As expected, the number of
variables in the truncated kernel increases with ε. This leads to two competing trends in the
1
http://wwwmath.uni-muenster.de/num/wirth/people/Schmitzer/
34
10−1
2-d
3-d
10
101
|X| · 10−3
err=1E-5
err=1E-6
err=1E-7
LP
100
10−1
12/|X|
ε/h2 = 0.1
102
|N |/|X × Y |
runtime/s
103
104
−2
ε/h2 = 1.8
LP
10−3
10−4
105
104
|X|
105
|X|
Figure 3: Average runtime and sparsity of Algorithm 3.4 for transporting test-images of different
size (up to 5122 pixels for 2-d, 643 for 3-d). Stopping criterion: L∞ -marginal error, for different
accuracy limits, final ε = 0.1 · h2 . Performance of the adaptive sparse linear programming solver
[44] given for comparison (LP). As expected, runtime increases with required accuracy. The
runtime of the scaling algorithm scales more favourably (approximately linear) with |X| and is
competitive for large instances. The number of variables scales as O(1/|X|), suggesting that the
number of variables per x ∈ X is roughly constant. For the final ε = 0.1 · h2 , the sparsity of the
truncated kernel is comparable to [44]. For ε = 1.8 · h2 , the largest value in E0 (the list for the
finest scale), more variables are required.
overall runtime: For large ε, the kernel truncation is less efficient, leading to an increase with
ε. For small ε, the number of variables is very small, but more and more stages of ε-scaling are
necessary, increasing the runtime as ε decreases further. Convergence of the regularized optimal
dual variables to the unregularized optimal duals is exemplified in the right panel, justifying the
use of the approximate entropy regularization technique for transport-type problems. While one
may consider the dual sub-optimality at ε ≈ 30 h2 sufficiently accurate, we point out that the
corresponding primal coupling still contains considerable blur (cf. Fig. 1) and that due to less
sparsity the runtime is actually higher than for ε ≈ h2 .
As illustrated by Figs. 3 and 4, by choosing the threshold for the stopping criterion and the
desired final ε, one can tune between required precision and available runtime.
Remark 5.1 (Interplay of Modifications). The numerical findings presented in Figs. 2-4 underline how each of the modifications discussed in Sect. 3 builds on the previous ones and that all
four of them are required for an efficient algorithm. The log-domain stabilization is an indispensable prerequisite for running the scaling algorithms with small regularization. However, for
small ε, convergence tends to become extremely slow (cf. Fig. 2), therefore ε-scaling is needed
to reduce the number of iterations. For small ε, kernel truncation significantly reduces the number of variables and accelerates the algorithm (cf. Figs. 2 and 4). However, for large ε (which
must be passed during ε-scaling), far fewer variables are truncated and the algorithm cannot
be run on large problems. This can be avoided by using the coarse-to-fine scheme, completing
the algorithm. In principle it is possible, only to combine log-domain stabilization with kernel
truncation, and to skip ε-scaling and the coarse-to-fine scheme. While this tends to solve the
stability and memory issues, convergence is still impractically slow.
35
104
0.6
dual suboptimality
100
90
10
runtime/s
|N |/|X|
103
2
80
70
60
50
1
10 −1
10
100
101
ε/h
2
102
40 −1
10
100
ε/h
2
101
0.4
0.2
0.0 −1
10
100
101
ε/h
102
2
Figure 4: Different final values for ε in Algorithm 3.4. X = Y are 2-d 256 × 256 grids. Left
Number of variables in truncated kernel per x ∈ X. For ε = 0.1 · h2 only about 10 variables
per x ∈ X are required. As expected, this number increases with ε (cf. Fig. 1). Center For
large ε, the runtime decreases with ε, since the number of variables decreases (cf. left plot). For
smaller ε, the runtime increases again, since more stages of ε-scaling are required. Right The
optimal regularized dual variables were transformed into feasible unregularized dual variables, by
decreasing each α(x) until all dual constraints α(x) + β(y) ≤ c(x, y) were met, (2.2b). The suboptimality of these dual variables is shown. As expected (see Sect. 1.1) they converge towards
a dual optimizer. Absolute optimal value was between 100 and 400 for the used test problems,
i.e. for small ε, sub-optimality is small compared to total scale.
5.3
Versatility
The framework of scaling algorithms developed in [19], see Sect. 2, allows to solve more general
transport-type problems for which the enhancements of Sect. 3 still apply. We now give some
examples to demonstrate this flexibility. The scope of the following examples is similar to [19],
but with Algorithm 3.4 one can solve larger problems with smaller regularization.
KL Fidelity and Wasserstein-Fisher-Rao distance. For the marginal function FX (σ) =
λ · KLX (σ|µ) with a given reference measure µ ∈ RX
+ and a weight λ > 0, see Def. 2.3, one
obtains for the (stabilized) proxdiv operator
λ
λ
α
proxdivε FX (σ) = (µ ν) λ+ε ,
proxdivε FX (σ, α) = exp − λ+ε
(µ ν) λ+ε .
(5.1)
A proof is given in [19]. Compared to the standard Sinkhorn algorithm, the only modification
is the pointwise power of the iterates. As λ → ∞ the Sinkhorn iterations are recovered. Note
−α
that in the stabilized operator only the exponential exp λ+ε
needs to be evaluated, which
remains bounded as ε → 0. Algorithm 3.4 performs similarly with KL-fidelity as with fixed
marginal constraints, allowing to efficiently solve large unbalanced transport problems. Since
the truncation scheme can also be used with non-standard cost functions such as (2.7), this
includes in particular the Wasserstein-Fisher-Rao (WFR) distance.
Fig. 5 shows two geodesics for the WFR distance, to intuitively illustrate its properties. The
geodesics have been computed as weighted barycenters between their endpoints (see below). For
a direct dynamic formulation we refer to [31, 17, 35]. For the relation to the KL soft-marginal
formulation, Def. 2.3, see [35].
36
t = 0.0
t = 0.2
t = 0.4
t = 0.6
t = 0.8
t = 1.0
Figure 5: Geodesics for the Wasserstein-Fisher-Rao distance on [0, 1]2 , approximated by a 256 ×
256 grid, computed as barycenters between the end-points with varying weights. Top row Mass
can increase or decrease during transport to match the target density. Bottom row Mass that
must travel further, is decreased during transport to save cost.
Wasserstein barycenters. Wasserstein barycenters as a natural generalization of the Riemannian center of mass have been studied in [1]. The computation of entropy regularized Wasserstein
barycenters with a Sinkhorn-type scaling algorithm has been presented in [7], an alternative numerical approach can be found in [24]. The iterations can be considered as a special case of the
framework in [19]. Here, we very briefly recall the iterations. Derivations and proofs can be
found in [7].
We want to compute the (entropy regularized) Wasserstein barycenter of a tuple (µ1 , . . . , µn )
∈ RX×n over a common base space X = Y with metric d with non-negative weights (λ1 , . . . , λn )
that sum to one. The primal functional can be written as an optimization problem over a tuple
(πi )ni=1 = (π1 , . . . , πn ) ∈ R(X×X)×n of couplings, which requires a slight generalization of Def. 2.4,
see [19]. It is given by
E((πi )i ) = F1 ((PX πi )i ) + F2 ((PY πi )i ) +
n
X
λi KL(πi |K)
(5.2a)
i=1
where
F1 ((νi )i ) =
n
X
i=1
ι{µi } (νi ),
(
0
if ∃ σ ∈ RX s.t. [σ = νi ∀ i = 1, . . . , n],
F2 ((νi )i ) =
+∞ else.
and K is the kernel (2.8) over X × X for the cost c = d2 . When an optimizer (πi† )i is found,
the common second marginal of all πi† is the sought-after barycenter. A corresponding dual
functional is
J((αi )i , (βi )i ) = −F1∗ (−(λi αi )i ) − F2∗ (−(λi βi )i ) − ε
n
X
>
λi KL∗ ([P>
X αi + PY βi ]/ε|K) (5.2b)
i=1
with
F1∗ ((αi )i ) =
n
X
i=1
hµi , αi i ,
(
Pn
0
if
∗
i=1 βi = 0,
F2 ((βi )i ) =
+∞ else.
37
(5.3)
ε = 0.1 h2
multi-marginal
ε = 2 h2
Figure 6: Barycenters in Wasserstein space over [0, 1]2 , computed on 256×256 grids for ε = 0.1 h2 .
Left ‘barycentric triangle’ spanned by a ring, a diamond and a square (for weights see (5.5)).
Center Close-up of the λ = (1, 2, 1)/4 barycenter, also shown for ε = 2 h2 (as reported in [7]). The
ε = 0.1 h2 version is much sharper, revealing discretization artifacts. Right The same barycenter
on a 64 × 64 grid, computed via multi-marginal formulation.
Note that relative to the dualization (2.10a) → (2.10b) we use rescaled dual variables [(αi /λi ,
βi /λi )]i . Primal and dual iterates are related by πi = diag(exp(αi /ε))K diag(exp(βi /ε)). As
before, we introduce scaling factors [(ui , vi )]i , ui = exp(αi /ε), vi = exp(βi /ε) and optimize
(5.2b) by alternating optimization in α and β, expressed by (u, v). Each αi can be optimized
(`+1) def.
(`)
independently and one gets a standard Sinkhorn update for every ui : ui
= µi (K vi ).
(`+1) def. (`+1)
(`+1)
Optimization in (βi )i is more involved. One finds vi
= σ
(K > ui
) where σ (`+1) =
Qn
> (`+1) )λi is the geometric mean of all (K > u(`+1) ) , weighted by (λ ) . Note that the
i
i i
i=1 (K ui
i
prox and proxdiv steps of F2 must be computed w.r.t. the weighted KL divergence
KLλ ((νi )i |(µi )i ) =
n
X
λi KL(νi |µi )
i=1
to account for the weights (λi )i in the third term of the dual functional.
These iterations can be stabilized, analogous to Sect. 3.1, where one splits each (ui , vi ) into
pairs (ũi , ṽi ) and (α̂i , β̂i ) and introduces one stabilized kernel Ki per pair (α̂i , β̂i ). The stabilized
proxdiv step of F2 w.r.t. KLλ is given by
!
n
X
σ
proxdivε F2 ((νi )i , (βi )i ) =
where
σ = exp
λi (log νi − βi /ε) .
(5.4)
νi
i=1
The terms βi /ε in the expressionPfor σ are seemingly unstable in the limit ε → 0. However, the F2∗
term in the dual (5.2b) enforces ni=0 λi βi = 0. Therefore, only the deviation from this constraint
enters the exponential in (5.4). When one gradually approaches the optimal solution during εscaling, this deviation can be kept numerically small. Likewise, adaptive truncation applies
to each kernel separately and a multi-scale coarse-to-fine approach can be used, as outlined in
38
Sects. 3.3-3.4. While this means that one must store multiple kernels, due to sparsity this is no
issue. Often one already has to keep different kernels, when one uses different discrete spaces Xi
for each marginal, to better capture the respective supports of the marginals µi .
In Fig. 6 a barycentric triangle is shown, computed with Algorithm 3.4, for weights (λ1 , λ2 , λ3 )
as follows:
(0, 4, 0)/4
(1, 3, 0)/4
(2, 2, 0)/4
(3, 1, 0)/4
(4, 0, 0)/4
(0, 3, 1)/4
(1, 2, 1)/4
(2, 1, 1)/4
(3, 0, 1)/4
(1, 1, 2)/4
(2, 0, 2)/4
(5.5)
(0, 2, 2)/4
(0, 1, 3)/4
(1, 0, 3)/4
(0, 0, 4)/4
The log-domain stabilization allows to reach a lower final regularization ε as for example in [7].
In fact, the regularization can be made so small, that discretization artifacts become detectable.
We do not argue that this is visually more pleasing. However, it clearly gives a better approximation to the original unregularized barycenter problem and illustrates that with log-domain
stabilization entropy regularized numerical methods can produce sharp results. We conjecture
that there are more elaborate ways of removing the artifacts than the rather indiscriminate entropic blur. Truncation of the kernel is also efficient for the barycenter computation: On the
finest hierarchy level, the average worst-case sparsity of the kernels was 3.5 · 10−3 , for the final
ε it was 6.0 · 10−4 .
Wasserstein-Fisher-Rao barycenters. Similarly one can define barycenters for transport
distances with KL marginal fidelity, which includes the Gaussian Hellinger-Kantorovich (GHK)
distance and the Wasserstein-Fisher-Rao (WFR) distance (Def. 2.3). Primal and dual functional
are given by (5.2) with
F1 ((νi )i ) = Λ ·
n
X
λi KL(νi |µi ) ,
F2 ((νi )i ) = inf Λ ·
σ∈RX
i=1
n
X
λi KL(νi |σ) ,
i=1
where Λ > 0 is a global weight of the KL-fidelity. When a primal optimizer is found, the
minimizing σ in F2 yields the sought-after barycenter. We refer to [19] for details. Partial
optimization w.r.t α can again be done for each αi separately, obtaining KL fidelity updates for
each ui , as given by (5.1). For the v-update one finds
−Λ
ε+Λ
proxdivε F2 ((νi )i , (βi )i ) = νi
βi
exp − ε+Λ
Λ
ε
ε
X
βj
.
·
λj νjε+Λ exp − ε+Λ
j
A barycentric triangle for the WFR distance is displayed in Fig. 7, a more detailed closeup with comparison to other models is shown in Fig. 8. Each reference measure consists of
three spatially separated smaller objects with different mass. This implies that the Wasserstein
barycenters have to transfer mass between the different groups, whereas the WFR barycenters
can compensate the difference by creating or annihilating mass, resulting in more natural interpolations. Fig. 8 also shows a barycenter for the GHK distance, as computed with Gaussian
convolution, without log-domain stabilization. Since the GHK uses the squared Euclidean distance as cost function, one can use efficient numerical methods for Gaussian convolution to avoid
storing the dense kernel [49]. But this does not generalize to the cost function (2.7) for WFR
39
Figure 7: Entropic barycentric triangle for the Wasserstein-Fisher-Rao distance on [0, π]2 , approximated by a 256 × 256 grid, with ε = 0.1 h2 ≈ 8.5 · 10−6 , weights given by (5.5).
and neither can it be combined with log-domain stabilization. Since adaptive truncation of the
kernel and multi-scale approximation of cost functions are rather flexible w.r.t. the cost function,
Algorithm 3.4 can efficiently compute the WFR barycenter for small ε. Again, the barycenters
computed with Algorithm 3.4 exhibit virtually no blur.
Multi-Marginal Problems. In [7] it was shown that the Sinkhorn algorithm can also be
extended to multi-marginal problems and it is straightforward to adapt Algorithm 3.4 to this
setting. For n marginals the kernel becomes an n-dimensional matrix and (α, β) are replaced by
a list (α1 , . . . , αn ). Stabilization is then
P done for each dual variable and for truncation in (3.10)
one considers exp(−[c(x1 , . . . , xn ) − ni=1 αi (xi )]/ε) where c is the n-dimensional cost function.
In [1] it is shown that barycenters in Wasserstein space can also be computed via a multimarginal formulation. As a toy example we computed a weighted barycenter between three
64 × 64 images (where pixels with value zero were ignored). Projection of the optimal multidimensional coupling to the barycenter image is done as described in [7]. The result is shown in
Fig. 6. Cardinality of the full product space was 1.5 · 109 . The final sparsity of the truncated
multi-dimensional kernel was 1.2 · 10−3 , worst-case sparsity was 4.7 · 10−3 .
Unfortunately, the dimensionality of multi-marginal problems grows so quickly with image
40
Figure 8: Comparison of different barycenters models: Left Close-up of Wasserstein-Fisher-Rao
barycenter for λ = (1, 2, 1)/4 from Fig. 7, for ε = 0.1 h2 . Center Corresponding Wasserstein
barycenter between normalized reference measures for ε = 0.1 h2 . Right Corresponding Gaussian Hellinger-Kantorovich barycenter for ε = 6.55 h2 , as computed with Gaussian convolution
without log-domain stabilization.
size, that at this point this does not yet scale efficiently to larger problems. For now, tricks to
avoid storing the full kernel that are tailored to the structure of particular cost functions are
more viable, see for example [7]. Nonetheless, we consider this an interesting proof of concept
for the flexibility of Algorithm 3.4 and will further study this application.
Wasserstein Gradient Flows. In [39] diagonal scaling algorithms were extended to compute proximal steps for entropy regularized optimal transport to approximate gradient flows in
Wasserstein space (cf. Sect. 1.1). This was then subsumed into the general framework of [19].
Here we given an example for the porous medium equation, for more details we refer to [39, 19].
Let
X
X µ(x)
F : P(X) → R,
v(x) µ(x)
(5.6)
µ 7→
u L(x) L(x) +
x∈X
x∈X
where u(s) = s2 , L is the discretized Lebesgue measure on X ⊂ Rd and v : X → R is a
potential. Then, for some initial µ(0) ∈ P(X) and a time step size τ > 0 we iteratively construct
a sequence (µ(`) )` where µ(`+1) is given by the proximal step of F with step size τ w.r.t. the
entropy regularized Wasserstein distance on X from reference point µ(`) . Based on Def. 2.7,
µ(`+1) can be computed as follows:
def.
def.
π (`+1) = argmin ι{µ(`) } (PX π) + 2 τ · F (PY π) + ε KL(π|K) , µ(`+1) = PY π (`+1) , (5.7)
π∈P(X 2 )
where K is the kernel w.r.t. the squared Euclidean distance on X. Then introduce the timecontinuous interpolation
µ : R+ → P(X),
t 7→ µ(`)
where t ∈ [τ · `, τ · (` + 1)).
(5.8)
Consider now the limit (τ, ε) → 0 in a way such that ε| log ε| ≤ τ 2 . Then, up to discretization,
the function µ converges to a solution of the porous media PDE
∂t µ = ∆(µ2 ) + div(µ · ∇v) .
41
(5.9)
t = 1.0E-3
t = 2.0E-3
µ(t, (x1 , 0.5))
t = 0.0E-3
1.2
1.0
0.8
0.6
0.4
0.2
0.0
0.0
t = 0.0E-3
t = 2.0E-3
t = 4.0E-3
0.2
0.4
0.6
0.8
1.0
x1
t = 4.0E-3
t = 5.0E-3
0.20
µ(t, (x1 , 0.5))
t = 3.0E-3
0.15
ε = 1E-5
ε = 1E-4
0.10
0.05
0.00
0.10
0.15
x1
0.20
Figure 9: Left Entropic Wasserstein gradient flow for the porous media equation on [0, 1]2 ,
approximated by a 256 × 256 grid with ε = 10−5 ≈ 0.66 h2 , τ = 2 · 10−4 . The energy is given by
(5.6) with v((x1 , x2 )) = 100 · x1 if x = (x1 , x2 ) ∈ Ω, v(x) = +∞ otherwise and Ω = [0, 1]2 \ Ω̂
where Ω̂ is a ‘barrier’ indicated by the white rectangles. Top Right Cross section of density at
different times along x2 = 0.5. Bottom Right Close-up for t = 4 · 10−3 for different values of
regularization ε. For ε = 10−5 the compact support of µ, a characteristic feature of the porous
media equation, is numerically well preserved. Without log-domain stabilization, for ε = 10−4
the entropic blur quickly distorts this feature.
A proof is given in [16]. Problem (5.7) is an instance of Def. 2.4 and can be solved by alternating
dual optimization. The updates w.r.t. ι{µ(`) } are standard Sinkhorn iterations. The proximal
step of F can be shown to be
ε
4 τ µ(x)
[proxε F (ν)] (x) =
· L(x) .
(5.10)
W
·
· exp(−v(x)/ε)
4τ
ε L(x)
where W denotes the Lambert W function (or product logarithm). To avoid numerical issues,
we propose to use the asymptotic expansion W(exp(s)) = s − log(s) + log(s)/s + o(1) as s → +∞
[21] whenever −v(x)/ε 1. The same applies for the stabilized iteration.
A numerical example is shown in Fig. 9. As in the previous experiments, Algorithm 3.4 allows
to use log-domain stabilization on large problems, producing sharp results. In this example, the
compact support of the porous media equation is numerically well preserved.
6
Conclusion
Scaling algorithms for entropy regularized transport-type problems have become a wide-spread
numerical tool. Naive implementations have some severe numerical limitations, in particular for
small regularization and on large problems. In this article, we proposed an enhanced variant of
the standard scaling algorithm to address these issues: Diverging scaling factors and slow convergence are remedied by log-domain stabilization and ε-scaling. Required runtime and memory
are significantly reduced by adaptive kernel truncation and a coarse-to-fine scheme. A new convergence analysis for the Sinkhorn algorithm was developed. Numerical examples showed the
42
efficiency of the enhanced algorithm, confirmed the scaling predicted by the convergence analysis
and demonstrated that the algorithm can produce sharp results on a wide range of transporttype problems. Potential directions for future research are the more detailed study of ε-scaling,
a more systematic understanding of the stability of the log-domain stabilization and application
to larger multi-marginal problems.
Acknowledgements. Lénaïc Chizat, Luca Nenna and Gabriel Peyré are thanked for stimulating discussions. Bernhard Schmitzer was supported by the European Research Council (project
SIGMA-Vision).
A
A.1
Additional Proofs
Proof of Lemma 4.20
First, we establish existence of minimizers. For some ε > 0, d ∈ RR×R the functional β 7→ Jˆε,d (β)
is convex and bounded from below. Further, it is invariant under adding the same constant to all
components of β. Hence, the optimal value minβ Jˆε,d (β) is not changed by adding the constraint
β1 = 0. With this added constraint the functional becomes strictly convex and coercive in the
remaining variables, hence a unique minimizer exists. The full set of minimizers is then obtained
via constant shifts.
The first order optimality condition for the functional yields for the i-th component of β:
1
β(i) =
softmax(−d(i, j) + β(j), ε) + softmin(d(j, i) + β(j), ε) ,
j:j6=i
j:j6=i
2
where the subscript j : j 6= i denotes that softmax is taken only over components {1, . . . , R}\{i}.
Finiteness of d ensures that this expression is meaningful. Let i1 ∈ {1, . . . , R} be an index where
∆β is maximal, i.e. ∆β(i1 ) = max ∆β.
From the optimality conditions for βa (i1 ), a = 1, 2, and (1.5) we obtain:
1
†
†
†
βa (i1 ) =
softmax(−da (i1 , j) + βa (j), εa ) + softmin(da (j, i1 ) + βa (j), εa ) ,
j:j6=i1
2 j:j6=i1
1
max (−∆d(i1 , j) + ∆β(j)) + max (∆d(j, i1 ) + ∆β(j)) + (ε1 + ε2 ) · log R
∆β(i1 ) ≤
j:j6=i1
2 j:j6=i1
ε1 +ε2
≤ max (w(i1 , j) + ∆β(j)) + 2 log R
j:j6=i1
where w(i, j) = max{−∆d(i, j), ∆d(j, i)}. This implies there is some i2 ∈ {1, . . . , R} \ {i1 } with
∆β(i2 ) ≥ −w(i1 , i2 ) + ∆β(i1 ) −
ε1 +ε2
2
· log R .
We will call the index i2 a child of i1 . We now repeat this reasoning to derive lower bounds
for other entries of ∆β. For this we must ‘remove’ the index i2 from the problem, defining a
reduced problem. Let I1 = {i1 , i2 } and let I2 = {1, . . . , R} \ I1 . We will keep all variables of β
with indices in I2 , but describe all variables with indices in I1 by a single reduced variable. For
this we consider vectors in R1+|I2 | , where we index the entries by {i1 } ∪ I2 . One can think of this
as a vector in RR , where we have ‘crossed out’ entries corresponding to I1 and replaced them by
a single effective entry, indexed with i1 . For a = 1, 2 we consider the reduced functionals
Jˆa : β̂ 7→ Jˆεa ,da (β̃a + B β̂)
43
where β̃a ∈ RR is a constant offset, β̂ ∈ R1+|I2 | is the reduced variable and B ∈ RR×(1+|I2 |) is a
matrix that implements the parametrization. We set
(
†
†
1 if j ∈ I1 , k = i1 ,
βa (j) − βa (i1 ) if j ∈ I1 ,
β̃a (j) =
B(j, k) = 1 if j = k ∈ I2 ,
0
else,
0 else.
So the reduced functionals are given by
X
exp([−da (j, k) − β̃a (j) + β̃a (k)]/εa )
Jˆa (β̂) =
j∈I1 ,
k∈I1
+
X
exp([−da (j, k) − β̃a (j) − β̂(i1 ) + β̂(k)]/εa )
j∈I1 ,
k∈I2
+
X
exp([−da (j, k) − β̂(j) + β̃a (k) + β̂(i1 )]/εa )
j∈I2 ,
k∈I1
+
X
exp([−da (j, k) − β̂(j) + β̂(k)]/εa )
j∈I2 ,
k∈I2
=
X
exp([−Da (j, k) − β̂(j) + β̂(k)]/εa )
j∈{i1 }∪I2 ,
k∈{i1 }∪I2
2
with the reduced coefficient matrix Da ∈ R(1+|I2 |) with entries
softmin
d
(r,
s)
+
β̃
(r)
−
β̃
(s),
ε
a
a
a
a
r∈I1 , s∈I1
softmin
r∈I1 da (r, k) + β̃a (r), εa
Da (j, k) =
softmin
d
(j,
s)
−
β̃
(s),
ε
a
a
a
s∈I1
da (j, k)
if j = i1 , k = i1 ,
if j = i1 , k ∈ I2 ,
if j ∈ I2 , k = i1 ,
if j ∈ I2 , k ∈ I2 .
Consider the reduced variables β̂a† ∈ R1+|I2 | with entries
(
βa† (i1 ) if j = i1 ,
β̂a† (j) =
βa† (j) if j ∈ I2 .
Then βa† = β̃a +B β̂a† and therefore β̂a† are minimizers of Jˆa . Note also that β̂2† (j)− β̂1† (j) = ∆β † (j)
for j ∈ {i1 } ∪ I2 .
Using the optimality conditions for the reduced functionals and arguing as above, we find
∆β(i1 ) ≤ max(W (i1 , k) + ∆β(k)) +
k∈I2
ε1 +ε2
2
log R
where W (i1 , k) = max{−∆D(i1 , k), ∆D(k, i1 )} for k ∈ I2 and ∆D = D2 − D1 . With (1.5) we
find
−∆D(i1 , k) ≤ max(−∆d(j, k) − ∆β(j) + ∆β(i1 )) + ε2 log R ,
j∈I1
∆D(k, i1 ) ≤ max(∆d(k, j) − ∆β(j) + ∆β(i1 )) + ε1 log R
j∈I1
44
and eventually W (i1 , k) ≤ maxj∈I1 (w(j, k) − ∆β(j)) + ∆β(i1 ) + max{ε1 , ε2 } · log R. So there is
some index i3 ∈ I2 such that
2
∆β(i3 ) ≥ min (−w(j, i3 ) + ∆β(j)) − ε1 +ε
+ max{ε1 , ε2 } log R .
2
j∈I1
The index i3 will be called a child of the minimizing index j ∈ I1 on the r.h.s., (or one of the
minimizing indices). Then we add i3 to the set I1 and repeat the argument with the reduced
functional, to obtain an index i4 and repeat this until I1 contains all indices.
Since we assign every new index ik that is added to I1 as a child to one parent node in I1 ,
this also constructs a tree graph with root node i1 (finiteness of d and consequently D implies
that this graph is connected). For an index ik let (i1 , i2 , . . . , ik ) be the unique path from the root
to ik . Then
∆β(ik ) ≥ −
k
X
w(ij−1 , ij ) + ∆β(i1 ) − 2 (k − 1) ε1 log R
j=2
≥ − maxdiam(w) + ∆β(i1 ) − 2 ε1 R log R .
Since ∆β(i1 ) = max ∆β the result follows.
A.2
Proof of Theorem 4.16
Let π1 , π2 be the primal optimizers associated with (α1 , β1 ) and (α2 , β2 ) and consider the assignment graph for π1 and π2 and threshold 1/M (see Lemma 4.18). Let {(Xi , Yi )}R
i=1 be the
strongly connected components of the assignment graph. By virtue of Lemma 4.18, statement
iii, µ(Xi ) = ν(Yi ) for i = 1, . . . , R. Pick some representatives {yi }R
i=1 ⊂ Y such that yi ∈ Yi for
i = 1, . . . , R.
For a = 1, 2, let now Jˆa be the reduced effective diagonal functionals, defined in Lemma 4.19,
corresponding to spaces (X, Y ), marginals (µ, ν), parameters εa , cost c, the partitions given by
the strongly connected components and the representatives {yi }R
i=1 . Let da be the corresponding
†
effective coefficients (finite, since c is finite), let β̂a be two corresponding maximizers and let
∆d = d2 −d1 , ∆β̂ = β̂2† −β̂1† . By virtue of Lemma 4.20 one has max ∆β̂−min ∆β̂ ≤ maxdiam(w)+
2 ε1 R log R, where w ∈ RR×R with w(i, j) = max{−∆d(i, j), ∆d(j, j)}.
Now we derive some estimates on ∆d. Consider once more the assignment graph for π1 , π2
and threshold 1/M . For every edge y → x we have (using (2.11))
α1 (x) + β1 (y) − c(x, y) ≥ −ε1 log M .
Moreover, from the marginal conditions we find π2 (x, y) ≤ ν(y), which implies
α2 (x) + β2 (y) − c(x, y) ≤ ε2 log M.
Combining the two estimates, we obtain
def.
∆α(x) + ∆β(y) ≤ (ε1 + ε2 ) log M = L
Similarly, for edges x → y we obtain
∆α(x) + ∆β(y) ≥ −(ε1 + ε2 ) log M = −L .
45
Let now (y1 , x1 , . . . , yk ) be an alternating path in (X, Y ), then, by combining the above inequalities we find ∆β(yj+1 ) ≥ ∆β(yj ) − 2 · L for j = 1, . . . , k − 1 and eventually
∆β(yk ) − ∆β(y1 ) ≥ −2 · (k − 1) · L .
Similarly, for a path (x̂1 , ŷ2 , . . . , ŷk ) get
∆α(x1 ) + ∆β(yk ) ≥ −(2 k − 1) · L ,
and for a path (ŷ1 , x̂1 , . . . , ŷk , x̂k )
∆α(xk ) + ∆β(y1 ) ≤ (2 k − 1) · L.
Consider now a partition cell (Xi , Yi ) and let yi ∈ Yi be the selected ‘representative’, as
described above. For every y ∈ Yi there is a path to and from yi with at most 2(|Yi | − 1)
edges, for every x ∈ Xi there is a path to and from yi with at most 2 |Yi | − 1 edges. With
∆α̃(x) = ∆α(x) + ∆β(yi ) and ∆β̃(y) = ∆β(y) − ∆β(yi ) we therefore obtain
|∆α̃(x)| ≤ (2 |Yi | − 1) · L,
|∆β̃(y)| ≤ 2 (|Yi | − 1) · L.
We recall (4.7)
da (i, j) = softmin c(x, y) − α̃a (x) − β̃a (y) − εa log(µ(x) ν(y)), εa
x∈Xi ,
y∈Yj
and get
∆d(i, j) ≤ max −∆α̃(x) − ∆β̃(y) − ∆ε · log µ(x) ν(y) + ε1 · log |Xi | |Yj |
x∈Xi ,
y∈Yj
≤ 4 |Yi | L + ε1 · log |Xi | |Yj |
∆d(i, j) ≥ min −∆α̃(x) − ∆β̃(y) − ∆ε · log µ(x) ν(y) − ε2 · log |Xi | |Yj |
x∈Xi ,
y∈Yj
≥ −4 |Yi | L − ε2 · log |Xi | |Yj |
From this follows that w(i, j) ≤ 8 max{|Yi |, |Yj |} ε1 log M + 2 ε1 log N , which in turn implies
that maxdiam(w) ≤ 16 ε1 N log M + 2 ε1 R log N .
Recall that ∆β̂ = β̂2† − β̂1† , where β̂a† , a = 1, 2, are the optimizers of the effective diagonal
problems. Then from Lemma 4.20, and by bounding R ≤ N we obtain that
max ∆β̂ − min ∆β̂ ≤ ε1 N (4 log N + 16 log M )
and finally with max ∆β − min ∆β ≤ max ∆β̃ − min ∆β̃ + max ∆β̂ − min ∆β̂ we get
max ∆β − min ∆β ≤ ε1 N (4 log N + 24 log M ),
and analogously we get the equivalent bound for ∆α.
46
References
[1] M. Agueh and G. Carlier. Barycenters in the Wasserstein space. 43(2):904–924, 2011.
[2] R. K. Ahuja, T. L. Magnanti, and J. B. Orlin. Network Flows: Theory, Algorithms, and
Applications. Prentice-Hall, Inc., 1993.
[3] L. Ambrosio and N. Gigli. A user’s guide to optimal transport, 2011.
[4] L. Ambrosio, N. Gigli, and G. Savaré. Gradient Flows in Metric Spaces and in the Space of
Probability Measures. Lectures in Mathematics. Birkhäuser Boston, 2005.
[5] H. H. Bauschke and P. L. Combettes. Convex Analysis and Monotone Operator Theory in
Hilbert Spaces. CMS Books in Mathematics. Springer, 1st edition, 2011.
[6] J.-D. Benamou and Y. Brenier. A computational fluid mechanics solution to the MongeKantorovich mass transfer problem. Numerische Mathematik, 84(3):375–393, 2000.
[7] J.-D. Benamou, G. Carlier, M. Cuturi, L. Nenna, and G. Peyré. Iterative Bregman projections for regularized transportation problems. 37(2):A1111–A1138, 2015.
[8] J.-D. Benamou, F. Collino, and J.-M. Mirebeau. Monotone and consistent discretization of
the monge-ampere operator. https://arxiv.org/abs/1409.6694.
[9] J.-D. Benamou, B. D. Froese, and A. M. Oberman. Numerical solution of the optimal transportation problem using the Monge–Ampère equation. Journal of Computational Physics,
260(1):107–126, 2014.
[10] D. P. Bertsekas. A distributed algorithm for the assignment problem. Technical report, Lab.
for Information and Decision Systems Report, MIT, May 1979.
[11] D. P. Bertsekas. The auction algorithm: A distributed relaxation method for the assignment
problem. Annals of Operations Research, 14:105–123, 1988.
[12] D. P. Bertsekas and J. Eckstein. Dual coordinate step methods for linear network flow
problems. Mathematical Programming, Series B, 42:203–243, 1988.
[13] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004.
[14] Y. Brenier. Polar factorization and monotone rearrangement of vector-valued functions.
44(4):375–417, 1991.
[15] M. Burger, M. Franek, and C.-B. Schönlieb. Regularised regression and density estimation
based on optimal transport. Applied Mathematics Research eXpress, 3 2012.
[16] G. Carlier, V. Duval, G. Peyré, and B. Schmitzer. Convergence of entropic schemes for
optimal transport and gradient flows. http://arxiv.org/abs/1512.02783, 2015.
[17] L. Chizat, G. Peyré, B. Schmitzer, and F.-X. Vialard. An interpolating distance between
optimal transport and Fisher-Rao. http://arxiv.org/abs/1506.06430, 2015.
[18] L. Chizat, G. Peyré, B. Schmitzer, and F.-X. Vialard. Unbalanced optimal transport: Geometry and Kantorovich formulation. http://arxiv.org/abs/1508.05216, 2015.
47
[19] L. Chizat, G. Peyré, B. Schmitzer, and F.-X. Vialard. Scaling algorithms for unbalanced
transport problems. http://arxiv.org/abs/1607.05816, 2016.
[20] R. Cominetti and J. San Martin. Asymptotic analysis of the exponential penalty trajectory
in linear programming. Mathematical Programming, 67:169–187, 1992.
[21] R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and D. E. Knuth. On the
LambertW function. Advances in Computational Mathematics, 5(1):329–359, 1996.
[22] M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transportation distances.
pages 2292–2300, 2013. http://arxiv.org/abs/1306.0895.
[23] M. Cuturi and D. Avis. Ground metric learning. 15:533–564, 2014.
[24] M. Cuturi and A. Doucet. Fast computation of wasserstein barycenters. In International
Conference on Machine Learning, 2014.
[25] J. H. Fitschen, F. Laus, and G. Steidl. Transport between RGB images motivated by
dynamic optimal transport. Journal of Mathematical Imaging and Vision, 2016.
[26] J. Franklin and J. Lorenz. On the scaling of multidimensional matrices. Linear Algebra and
its Applications, 114–115:717–735, 1989.
[27] A. V. Goldberg and R. E. Tarjan. Finding minimum-cost circulations by successive approximation. Math. Oper. Res., 15(3):430–466, 1990.
[28] S. Haker, L. Zhu, A. Tannenbaum, and S. Angenent. Optimal mass transport for registration
and warping. 60(3):225–240, December 2004.
[29] R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker-Planck
equation. 29(1):1–17, 1998.
[30] P. A. Knight. The Sinkhorn-Knopp algorithm: Convergence and applications. SIAM. J.
Matrix Anal. & Appl., 30(1):261–275, 2008.
[31] S. Kondratyev, L. Monsaingeon, and D. Vorotnikov. A new optimal transport distance on
the space of finite Radon measures. http://arxiv.org/abs/1505.07746, 2015.
[32] J. Kosowsky and A. Yuille. The invisible hand algorithm: Solving the assignment problem
with statistical physics. Neural Networks, 7(3):477–490, 1994.
[33] H. W. Kuhn. The Hungarian method for the assignment problem. Naval Research Logistics,
2:83–97, 1955.
[34] C. Léonard. From the Schrödinger problem to the Monge–Kantorovich problem. Journal of
Functional Analysis, 262(4):1879–1920, 2012.
[35] M. Liero, A. Mielke, and G. Savaré.
Optimal entropy-transport problems and a new Hellinger-Kantorovich distance between positive measures.
http://arxiv.org/abs/1508.07941, 2015.
[36] J. Maas, M. Rumpf, C. Schönlieb, and S. Simon. A generalized model for optimal transport
of images including dissipation and density modulation. submitted, 2014.
48
[37] Q. Mérigot. A multiscale approach to optimal transport. Computer Graphics Forum,
30(5):1583–1592, 2011.
[38] A. M. Oberman and Y. Ruan. An efficient linear programming method for optimal transportation. http://arxiv.org/abs/1509.03668, 2015.
[39] G. Peyré. Entropic approximation of Wasserstein gradient flows. 8(4):2323–2351, 2015.
[40] J. Rabin, G. Peyré, and L. D. Cohen. Geodesic shape retrieval via optimal mass transport.
pages 771–784, 2010.
[41] J. Rabin and N. Papadakis. Convex color image segmentation with optimal transport
distances. In Scale Space and Variational Methods (SSVM 2015), pages 256–268, 2015.
[42] Y. Rubner, C. Tomasi, and L. J. Guibas. The earth mover’s distance as a metric for image
retrieval. 40(2):99–121, 2000.
[43] F. Santambrogio. Optimal Transport for Applied Mathematicians, volume 87 of Progress in
Nonlinear Differential Equations and Their Applications. Birkhäuser Boston, 2015.
[44] B. Schmitzer. A sparse multi-scale algorithm for dense optimal transport. Journal of Mathematical Imaging and Vision, 56(2):238–259, 2016.
[45] B. Schmitzer and C. Schnörr. A hierarchical approach to optimal transport. In Scale Space
and Variational Methods (SSVM 2013), pages 452–464, 2013.
[46] B. Schmitzer and C. Schnörr. Globally optimal joint image segmentation and shape matching
based on Wasserstein modes. Journal of Mathematical Imaging and Vision, 52(3):436–458,
2015.
[47] M. Sharify, S. Gaubert, and L. Grigori. Solution of the optimal assignment problem by
diagonal scaling algorithms. http://arxiv.org/abs/1104.3830v2, 2013.
[48] R. D. Sinkhorn and P. J. Knopp. Concerning nonnegative matrices and doubly stochastic
matrices. Pacific J. Math, 21(2):343—-348, 1967.
[49] J. Solomon, F. de Goes, G. Peyré, M. Cuturi, A. Butscher, A. Nguyen, T. Du, and L. Guibas.
Convolutional Wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (Proc. SIGGRAPH 2015), 34(4):66:1–66:11, 2015.
[50] J. Solomon, A. Nguyen, A. Butscher, M. Ben-Chen, and L. Guibas. Soft maps between
surfaces. Computer Graphics Forum, 31(5), 2012.
[51] J. Solomon, R. Rustamov, L. Guibas, and A. Butscher. Wasserstein propagation for semisupervised learning. In International Conference on Machine Learning, 2014.
[52] M. Thorpe, S. Park, S. Kolouri, G. K. Rohde, and D. Slepčev. A transportation Lp distance
for signal analysis. https://arxiv.org/abs/1609.08669, 2016.
[53] C. Villani. Optimal Transport: Old and New, volume 338 of Grundlehren der mathematischen Wissenschaften. Springer, 2009.
[54] W. Wang, D. Slepčev, S. Basu, J. A. Ozolek, and G. K. Rohde. A linear optimal transportation framework for quantifying and visualizing variations in sets of images. 101:254–269,
2012.
49
| 5cs.CE
|
Nez: Practical Open Grammar Language
Kimio Kuramitsu
arXiv:1511.08307v1 [cs.PL] 26 Nov 2015
Yokohama National University, JAPAN
[email protected]
http://nez-peg.github.io/
Abstract
tions. Developers use a formal grammar, such as LALR(k), LL(k),
or PEG, to specify programming language syntax. Based on the
grammar specification, a parser generator tool, such as Yacc[14],
ANTLR3/4[29], or Rats![9], produces an efficient parser code that
can be integrated with the host language of the compiler and interpreter. This generative approach, obviously, enables developers to
avoid error-prone coding for lexical and syntactic analysis.
Traditional parser generators, however, are not entirely free
from coding. One particular reason is that the formal grammars
used today lack sufficient expressiveness for many of the popular programming language syntaxes, such as typedef in C/C++
and other context-sensitive syntaxes[3, 8, 9]. In addition, a formal
grammar itself is a syntactic specification of the input language, not
a schematic specification for tree representations that are needed in
semantic analysis. To make up for these limitations, most parser
generators take an ad hoc approach called semantic action, a fragment of code embedded in a formal grammar specification. The
embedded code is written in a host language and combined in a
way that it is hooked at a parsing context that requires extended
recognition or tree constructions.
The problem with arbitrary action code is that they decrease
the reusability of the grammar specification[28]. For example, Yacc
cannot generate the Java version of a parser simply because Yacc
uses C-based action code. There are many Yacc ports to other
languages, but they do not reuse the original Yacc grammar without
porting of action code. This is an undesirable situation because the
well-defined grammar is demanded everywhere in IDEs and other
software engineering tools[18].
Nez is a grammar language designed for open grammars. In
other words, our aim is that once we write a grammar specification,
anyone can use the grammar in any programming language. To
achieve the openness of the grammar specification, Nez eliminates
any action code from the beginning, and provides a declarative
but small set of operators that allow the context-sensitive parsing
and flexible Abstract Syntax Tree construction that are needed for
popular programming language syntaxes. This paper presents the
implementation status of Nez with the experience of our grammar
development.
The Nez grammar language is based on parsing expression
grammars (PEGs), a popular syntactic foundation formalized by
Ford[8]. PEGs are simple and portable and have many desirable
properties for defining modern programming language syntax, but
they still have several limitations in terms of defining popular programming languages. Typical limitations include typedef-defined
name in C/C++ [8, 9], delimiting identifiers (such as <<END) of the
Here document in Perl and Ruby, and indentation-based code layout in Python and Haskell [1].
In the first contribution of Nez, we model a single extended
parser state, called a symbol table, to describe a variety of contextsensitive patterns appearing in programming language constructs.
The key idea behind the symbol table is simple. We allow any
Nez is a PEG(Parsing Expressing Grammar)-based open grammar language that allows us to describe complex syntax constructs
without action code. Since open grammars are declarative and
free from a host programming language of parsers, software engineering tools and other parser applications can reuse once-defined
grammars across programming languages.
A key challenge to achieve practical open grammars is the expressiveness of syntax constructs and the resulting parser performance, as the traditional action code approach has provided very
pragmatic solutions to these two issues. In Nez, we extend the
symbol-based state management to recognize context-sensitive language syntax, which often appears in major programming languages. In addition, the Abstract Syntax Tree constructor allows us
to make flexible tree structures, including the left-associative pair
of trees. Due to these extensions, we have demonstrated that Nez
can parse not all but many grammars.
Nez can generate various types of parsers since all Nez operations are independent of a specific parser language. To highlight this feature, we have implemented Nez with dynamic parsing, which allows users to integrate a Nez parser as a parser library
that loads a grammar at runtime. To achieve its practical performance, Nez operators are assembled into low-level virtual machine
instructions, including automated state modifications when backtracking, transactional controls of AST construction, and efficient
memoization in packrat parsing. We demonstrate that Nez dynamic
parsers achieve very competitive performance compared to existing
efficient parser generators.
Categories and Subject Descriptors D.3.1 [Programming Languages]: Formal Definitions and Theory – Syntax; D.3.4 [Programming Languages]: Processors – Parsing
General Terms Languages, Algorithms
Keywords Parsing expression grammars, Context-sensitive grammars, Packrat parsing
1.
Introduction
A parser generator is a standard approach to reliable parser development in programming languages and many parser applica-
[Copyright notice will appear here once ’preprint’ option is removed.]
Unpublished, draft paper to be revised. Comments and suggestions are welcome
1
2015/11/30
parsed substrings to be handled as symbols, a specialized word
in another context. The symbol table is a runtime, growing, and
recursive storage for such symbols. If we handle typedef-defined
names as symbols for example, we can realize different parsing
behavior through referencing the symbol table. Nez provides a
small set of symbol operators, including matching, containment
testing, and scoping. More importantly, since the symbol table
is a single, unified data structure for managing various types of
symbols, we can easily trace state changes and then automate
state management when backtracking. In addition, the symbol table
itself is simply a stack-based data structure; we can translate it into
any programming language.
Another important role of the traditional action code is the construction of ASTs. Basically, PEGs are just a syntactic specification, not a schematic specification for tree structures that represent
AST nodes. In particular, it is hard to derive the left-associative
pair of trees due to the limited left recursion in PEGs. In Nez, we
introduce an operator to specify tree structures, modeled on capturing in perl compatible regular expressions (PCREs), and extend
tree manipulations including tagging for typing a tree, labeling for
sub-nodes, and left-folding for the left-associative structure. Due
to these extensions, a Nez parser produces arbitrary representations
of ASTs, which would leads to less impedance mismatching in semantic analysis.
To evaluate the expressiveness of Nez, we have performed extensive case studies to specify various popular programming languages, including C, C#, Java8, JavaScript, Python, Ruby, and Lua.
Surprisingly, the introduction of a simple symbol table improves
the declarative expressiveness of PEGs in a way that Nez can recognize almost all of our studied languages and then produce a practical representation of ASTs. Our case studies are not complete,
but they indicate that the Nez approach is promising for a practical
grammar specification without any action code.
At last but not least, parsing performance is an important factor since the acceptance of a parser generator relies definitively on
practical performance. In this light, Nez can generate three types
of differently implemented parsers, based on the traditional source
generation, the grammar translation, and the parsing machine[11].
From the perspective of open grammars, we highlight the latter
parsing machine that enables a PCRE-style parser library that loads
a grammar at runtime. To achieve practical performance, Nez operators are assembled into low-level virtual instructions, including automated state modifications when backtracking, transactional
controls of AST construction, and efficient memoization in packrat
parsing. We will demonstrate that the resulting parsers are portable
and achieve very competitive performance compared to other existing standard parsers for Java, JavaScript, and XML.
The remainder of this paper proceeds as follows. Section 2 is an
informal introduction to the Nez grammar specification language.
Section 3 is a formal definition of Nez. Section 4 describes eliminating parsing conditions from Nez. Section 5 describes parser
runtime and implementation. Section 6 reports a summary of our
case studies. Section 7 studies the parser performance. Section 8
briefly reviews related work. Section 9 concludes the paper. The
tools and grammars presented in this paper are available online at
http://nez-peg.github.io/
2.
PEG
’ ’
[]
.
A
(e)
e?
e∗
e+
&e
!e
e1 e2
e1 /e2
AST
{e}
$(e)
{$ e }
#t
‘ ‘
Symbols
<symbol A>
<exists A>
<exists A x>
<match A>
<is A>
<isa A>
<block e>
<local A e>
Conditional
<on c e>
<on !c e>
<if c>
<if !c>
Description
Matches text
Matches character class
Any character
Non-terminal application
Grouping
Option
Zero-or-more repetitions
One-or-more repetitions
And-predicate
Negation
Sequencing
Prioritized Choice
Construction
Construction
Construction
Construction
Construction
5
5
5
5
5
Constructor
Connector
Left-folding
Tagging
Replaces a string
Action
Predicate
Predicate
Predicate
Predicate
Predicate
Action
Action
5
5
5
5
5
5
5
5
Symbolize Nonterminal A
Exists symbols
Exists x symbol
Matches symbol
Equals symbol
Contains symbol
Nested scope for e
Isolated local scope for e
Action
Action
Predicate
Predicate
5
5
5
5
e on c is defined
e on c is undefined
If c is defined
If c is undefined
each of which is defined by a mapping from a nonterminal A to a
parsing expression e:
A=e
Table 1 shows a list of Nez operators that constitute the parsing
expressions. All PEG operators inherit the formal interpretation
of PEGs[8]. That is, the string ’abc’ exactly matches the same
input, while [abc] matches one of these characters. The . operator
matches any single character. The e?, e∗, and e+ expressions
behave as in common regular expressions, except that they are
greedy and match until the longest position. The e1 e2 attempts
two expressions e1 and e2 sequentially, backtracking to the starting
position if either expression fails. The choice e1 / e2 first attempt
e1 and then attempt e2 if e1 fails. The expression &e attempts
e without any character consuming. The expression !e fails if e
succeeds, but fails if e succeeds. Furthermore information on PEG
operators is detailed in [8].
The expressiveness of PEGs is almost similar to that of deterministic context-free grammars (such as LALR and LL grammars). In general, PEGs are said to express all LALR grammar languages, which are widely used in a standard parser generator such
as Lex/Yacc[14]. In addition, PEGs have many desirable properties
that are characterized by:
A Taste of Nez
Nez and Parsing Expression
Nez is a PEG-based grammar specification language. The basic
constructs come from those of PEGs, such as production rules and
parsing expressions. A Nez grammar is a set of production rules,
Unpublished, draft paper to be revised. Comments and suggestions are welcome
Proc.
5
5
5
5
5
4
4
4
3
3
2
1
Table 1. Nez operators: ”Action” stands for symbol mutators and
”predicate” stands for symbol predicates.
This section is an informal introduction to the Nez grammar specification language.
2.1
Type
Primary
Primary
Primary
Primary
Primary
Unary suffix
Unary suffix
Unary suffix
Unary prefix
Unary prefix
Binary
Binary
• deterministic behavior, avoiding the dangling if-else problem
• left recursion-free, preserving operator precedence
2
2015/11/30
Expr = Prod {$left (’+’ #Add / ’-’ #Sub) $right(Prod)}*
Prod = Val {$left (’*’ #Mul / ’/’ #Div) $right(Val)}*
Val = { [0-9]+ #Int }
• scanner-less parsing, avoiding the extensive use of lexer hacks
such as in C++ grammars, and
• unlimited lookahead, recognizing highly nested structures such
as {an bn cn | n > 0}.
Nez inherits all these properties from PEGs, and has extended
features based on AST operators, symbol operators and conditional
parsing, as listed in Table 1.
2.2
1 + 2 * 3
1 * 2 + 3
#Mul
#Mul
$left
#Add
$left
AST Operators
#Int
[1]
The first extension of parsing expressions in Nez is a flexible
construction of AST representations. Each node of an AST contains
a substring extracted from the input characters. Nez has adopted
an PCRE-like capturing operator, denoted by {}, to capture the
substring. The productions Int and Long below are examples of
capturing a sequence of digits (as defined in NUM.)
$right
$right
$left
#Int
[3]
$right
#Int
[1]
#Add
$left
#Int
[2]
#Int
[1]
$right
#Int
[2]
Figure 1. Mathematical basic operators and example of ASTs
Note that the AST representation in Nez is a so-called common
tree, a sort of generalized data structure. However, we can map a
tag to a class and then map labels to fields in a mapped class, which
leads to automated conversion of concrete tree objects. Although
mapping and type checking are an interesting challenge, they are
beyond the scope of this paper.
To the end, we show an example of mathematical basic operators written in Nez. Figure 1 shows some of ASTs constructed
when evaluate the nonterminal Expr.
NUM = [0-9]+
Int = { NUM }
Long = { NUM } [Ll]
A string to be captured is the exact one matched with NUM. That
is, Long accepts the input 0L but captures the only substring 0. An
empty string may be captured by an empty expression.
Tagging is introduced to distinguish the type of nodes. We can
just add a #-prefixed tag to each of the nodes.
2.3
Int = { NUM #Int }
Long = { NUM #Long } [Ll]
Symbol Table
PEGs are very expressive, but they sometimes suffer from insufficient expressiveness for context-sensitive syntax, appearing even in
popular programming languages such as:
A backquoat operator ‘ ‘ is a string operator that replaces the
captured string with the specified string. Using the empty capture,
we are allowed to create an arbitrary node at any point of a parsing
expression.
• Typedef-defined name in C/C++ [8, 9]
• Here document in Perl, Ruby, and many other scripting lan-
guages
DefaultValue = { ‘0‘ #Int }
• Indentation-based code layout in Python and Haskell [1]
• Contextual keywords used in C# and other evolving languages[5]
The connector $(e) is used to make a tree structure by connecting two AST nodes in a parent-child relationship. The prefix
$ is used to specify a child node and append it to a node that is
constructed on the left-hand side of $(e). As a result, the tree is
constructed as the natural order of left-to-right and top-down parsing.
Basically, the AST operators can transform a sequence of nodes
into a tree-formed structure. We can specify that a subtree is nested,
flattened, or ignored (by dropping the connector). In the following,
we make the construction of a flattened list (#Add 1 2 3) and a
nested right-associative pair (#Add 1 (#Add 2 3)) for the same
input 1+2+3.
The problem with context-sensitive syntax is that PEGs are assumed stateless parsing and cannot handle state changes in the
parser context. However, most of the state changes above can
be modeled by a string specialization; the meaning of words is
changed in a certain context. We call such a specialized string a
symbol. Nez is designed to provide a symbol table and related symbol operators to manage symbols in the parser context.
To illustrate what a symbol is, let us suppose that the production
NAME is defined:
NAME = [A-Za-z] [A-Za-z0-9]*
List = { $(Int) (’+’ $(int))* #Add }
Binary = { $(Int) (’+’ $(Binary)) #Add }
The symbolization operator takes the form of <symbol A> to
declare that a substring matched at the nonterminal A is a symbol.
We call such a symbol an A-specialized symbol, which is stored
with the association of the nonterminal in the symbol table.
Here is a symbolization of NAME, where NAME first attempts to
match and, if matched, <match NAME> adds the matched symbol to
the symbol table.
The construction of the left-associative structure, however, suffers from the forbidden left-recursion (e.g, Binary = Binary
’+’ Int). Nez specifically provides a left-folding constructor
{$ ...} that allows a left-hand-constructed node to be contained
in a new right-hand node. Accordingly, we can make a precedencepreserving form of ASTs by folding from the repetition.
<symbol NAME>
Add = Int {$ ’+’ $(int) #Add}*
In Nez, a symbol predicate is defined to refer to the symbol by
the nonterminal name. The following <match NAME> is one of the
symbol predicates that exactly match the NAME-specialized symbol
for the input characters.
The connector and the left-folding operator can associate an
optional label for the connected child node. The associated label
follows the $ notion:
Add = Int {$left ’+’ $right(int) #Add}*
Unpublished, draft paper to be revised. Comments and suggestions are welcome
<match NAME>
3
2015/11/30
Here, we consider a simple case where XML tags need to match
closed tags with the same open tags.
USERTYPE = [A-Za-z_] !W*
W = [A-ZA-z_0-9]
S = [ \t\r\n]
<A><B> ... </B> </A>
TypeDef
= ’typedef’ S* TypeName S* <symbol USERNAME> S* ’;’
Briefly, an XML element can be specified as follows:
ELEMENT =
<symbol TAG> ELEMENT <is TAG>
TypeName
= BuildInTypeName / <isa USERNAME>
As described above, the symbol TAG enables us to ensure the
same name in both open and closed tags. However, the ELEMENT
involves nested symbolization inside, resulting in repeated symbolization. On the other hand, <is TAG> refers to the latest TAGspecialized symbol. As a result, the ELEMENT above can only accept the last tag, such as <A><B> ... </B> </B>, which is a not
desirable result.
The notation <block e> is used to declare a nested scope of symbols. That is, any symbols defined inside <block e> are not referable outside the block. Here is a scoped version of the ELEMENT,
which works as expected
BuiltInType
= ’int’ !W / ’long’ !W / ’float’ !W ...
Figure 2. Definition of typedef-name syntax with symbol operators
Compared to the nonterminal NAME, the result of <match NAME>
varies, depending on the past result at <symbol NAME>. That is, if
the <symbol NAME> accepts Apple previously, <match NAME>
only accepts ’Apple’. In this way, the symbol operators handle
context-depended parsing with the state changes in the symbol table.
The uniqueness of Nez is that the state changes are limited
to a single symbol table. That is, we are allowed to add multiple
symbols with <symbol NAME> or different kinds of symbols with
another nonterminal in the same table. Nez offers various types of
symbol predicates, which can apply different patterns of the symbol
references.
ELEMENT = <block
<symbol TAG> ELEMENT <is TAG> >
In Nez, we have adopted a single nested scope for multiple
nonterminal symbols. In other words, different kinds of symbols
are equally controlled with a single scope. One could consider that
there is a language that has a more complex scoping rule, while
our focus is only on names that directly influence the syntactic
analysis. One important exception is an explicit isolation of the
specific symbols by <local A e >. In this scope, all A-specialized
symbols before the local scope are isolated and not referable by the
subexpression e.
• <exists A> – checks if the symbol table has any A-specialized
symbols
2.4
• <exists A s> – checks if the symbol table has an A-specialized
symbol that is equal to the given string s
• <match A> – matches the last A-specialized symbol over the
input characters
• <is A> – compares the last A-specialized symbol with an A-
matched substring.
• <isa A> – checks the containment of an A-matched substring
in a set of A-specialized symbols stored in the table.
Note that the two symbol predicates <match A> and <is A>
are quite similar, but they significantly differ in terms of the input
acceptance. To illustrate the difference, suppose that the previous
<match NAME> accepts and stores the symbol ’in’. <match NAME>
accepts the input ’include’ (i.e., the input ’clude’ is unconsumed), while <is NAME> does not because it compares the stored
symbol ’in’ with the NAME-matched string ’include’.
Figure 2 shows a simplified example of the typedef-name syntax, excerpted from the C grammar. The production TypeDef describes the syntax of the typedef statement, which includes the
symbolization of USERTYPE. In TypeName, we first match builtin type names and then match one of the USERTYPE-specialized
symbols. As a result, we can express context-sensitive patterns in
TypeName.
2.3.1
Spacing = <if !IgnoreNewLine> [\n\r] / [ \t]
The conditions, as well as the symbols, are a global state across
productions. That is, the condition IgnoreNewLine affects all nonterminals and parsing expressions that involve the Spacing nonterminal. The expression <on IgnoreNewLine e> is used to declare
the static condition of the inner subexpression. The following are
the IgnoreNewLine version and non-IgnoreNewLine version of
the Expr nonterminal.
Scoping
In principle, symbols in the symbol table are globally referable
from any production. However, many programming languages have
their own scope rule for identifiers, which may require us to restrict
the scope of symbols in parallel with a scope construct of the language. Nez provides the explicit scope declaration for that purpose.
Unpublished, draft paper to be revised. Comments and suggestions are welcome
Conditional Parsing
The idea of conditional parsing is inspired by the conditional compilation, where the #ifdef ... #endif directives switch the
compilation behavior. Likewise, Nez uses <if ...> to switch the
paring behavior, but it differs in that the condition, as well as the
semantic predicate, can be controlled by <on ...> in the parser
context.
Nez supports multiple parsing conditions, identified by the userspecified condition name. Let c be a condition name. A parsing
expression <if c> e means that e is evaluated only if c is true. We
allow the ! predicate for the negation of the condition c. That is,
<if !c> is a syntax sugar of !<if c>. Two expressions e1 and e2 are
distinctly switched by <if c> e1 / <if !c> e2.
In addition, Nez provides the scoped-condition controller for arbitrary parsing expressions. <on c e> means that the subexpression
e is evaluated under the condition that c is true, while <on !c e>
means that the subexpression e is evaluated under the condition
that c is false.
Here is an example of conditional parsing; the acceptance of
Spacing below depends on the condition IgnoreNewLine:
<on IgnoreNewLine Expr>
<on !IgnoreNewLine Expr>
4
2015/11/30
A, B, C ∈ N
a, b ∈ Σ
x, y, z ∈ Σ∗
xy
e
A=e
T
[A, x]
S = [A, x][A0 , y]
e
Nonterminals
Characters
Sequence of characters
Concatenation of x and y
Parsing expressions
Production
Tree node
A-specialized symbol x
State, sequence of labeled symbols
empty string or sequence
::=
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Figure 3. Notations uses throughout this paper
Input: if(a > b) return a; else { return b; }
#If
#GreaterThan
#Variable
'a'
#Variable
'b'
#Return
#Return
#Variable
'a'
#Variable
'b'
The initial state of the AST node is an empty tree. new(x) is a
function that instantiates a new AST node including a given string
x. Let v be a reference of a node of ASTs. That is, we say v = v 0 if
v 0 is the mutated v with the following tree manipulation functions:
Note that conditional parsing is a Boolean version of the symbol
table. Let C be an empty expression (i.e., C = ’’). <on c e >
is a syntax sugar of <block <symbol C > e > and <if c > is a
syntax sugar of <exists C ’’>. However, the conditional parsing
can be eliminated from arbitrary Nez grammars, since the possible
states are at most 2. This is why Nez provides specialized operators
for conditional parsing. The elimination of parsing conditions is
described in Section 4.
• tag(v, #t) – replaces the tag of v with the specified #t
• replace(v, x) – replaces the string of v with the specified
string x
• link(v, v 0 ) – appends a child node v 0 to the parent node v
Next, we will turn to the model of the symbol table. Let [A, x]
denote a symbol x that is specialized by nonterminal A. The symbol table S is represented by a sequence of symbols, such as
[A, x][B, y].... Suppose two following tables:
Grammar and Language Design
This section describes the formal definition of Nez. Figure 3 is a
list of notations used throughout this section.
3.1
S1 = [A, x][B, y]
S2 = [A, x][B, y][A, z]
ASTs and Symbol Table
We start by defining the model of the AST representation and the
symbol table in Nez.
An AST is a tree representation of the abstract structure of
parsed results. The tree is ”abstract” in the sense that it contains no
unnecessary information such as white spaces or grouping parentheses. Based on the AST operators in Nez, the syntax of the AST
representation, denoted by v, is defined inductively:
The initial state of the symbol table is an empty sequence,
denoted by . We write [A, x] ∈ S for the containment of [A, x]
in S. In addition, we write S[A, z] to represent an explicit addition
of [A, z] into S. (SS 0 is a concatenation of S and S 0 ) That is, we
say that S1 [A, z] = S2 and S1 ⊂ S2 . Other operations for S are
represented by the following two functions:
• top(S, A) – a A-specialized symbol that is recently added in
the table. That is, we say top(S1 , A) = x and top(S2 , A) = z
v :== #t[v] | #t[’...’] | v v
where #t is a tag to identify the type of T and a captured string
written by ’...’. A whitespace concatenates two or more nodes
as a sequence. Note that we ignore the labeling of subnodes for
simplicity.
Here is an example of an AST tree, shown in Figure 4.
• del(S, A) – a new table that removes all A-specialized entries
from S.
3.2
Grammars
A Nez G = (N, Σ, P, es , T ) has elements, where N is a set of
nonterminals, Σ is a set of characters, P is a set of productions, es
is a start expression, and T is a set of tags.
Each production p ∈ P is a mapping from A to e; we write
as A = e, where A ∈ N and e is a parsing expression. Figure 5
is an abstract syntax of parsing expressions in Nez. Due to space
constraints, we focus on core parsing expressions. Note that e? is
the syntax sugar of e/, e∗ the sugar of A0 = eA0 /, and e+ the
sugar of ee∗ (as defined in [8]).
#If[
#GreaterThan[
#Variable[’a’]
#Variable[’b’]
]
#Return[#Variable[’a’]]
#Return[#Variable[’b’]]
]
Unpublished, draft paper to be revised. Comments and suggestions are welcome
: empty
: nonterminal
: character
: sequence
: prioritized choice
: repetition
: and predicate
: not predicate
: constructor
: left-folding
: connector
: tag
: string replacement
: symbol definition
: symbol existence
: symbol existence
: symbol match
: symbol equivalence
: symbol containment
: nested scope
: isolated local scope
Figure 5. An abstract syntax of Nez expressions
Figure 4. Pictorial notation of ASTs
3.
A
a
e e0
e / e0
e∗
&e
!e
{e}
{$ e }
$(e)
#t
‘x‘
hsymbol Ai
hexists A
hexists A xi
hmatch Ai
his Ai
hisa Ai
hblock ei
hlocal A ei
5
2015/11/30
{e}
(S, x, v) −
→ (S, x, v)
a
e
(S, xy, v) −
→ (SS 0 , y, v 0 )
a
{$ e }
(S, ax, v) −
→ (S, x, v)
A=e
(S, xy, T ) −−−−→ (SS 0 , y, new(x))
e
$(e)
(S, xy, v) −
→ (SS 0 , y, v 0 )
A
(S, xy) −→
e
(S, xy, v) −
→ (S 0 , y, v 0 )
(SS 0 , y, v 0 )
$(e)
(S, xy, v) −−−→ (S 0 , y, link(v, v 0 ))
e e0
{$ e }
e
(S, xy, v) −
→ (0 S, y, v 0 )
e0
e
(S, xyz, v) −
→ (SS 0 , yz, v 0 ) (SS 0 , yz, v 0 ) −→ (SS 0 S 00 , z, v 00 )
{$ e }
(S, xy, v) −−−−→ (0 S, y, link(new(x), v))
e e0
(S, x, v) −−→ (SS 0 S 00 , z, v 00 )
#t
e/e0
e
(S, xy, v) −
→
#t
(S, x, v) −−→ (S, x, tag(v, #t))
(SS 0 , y, v 0 )
‘x‘
e/e0
(S, xy, v) −−−→ (SS 0 , y, v 0 )
‘x‘
(S, x, v) −−→ (S, x, replace(v, x))
e0
e
(S, xy, v) −
→ • (S, xy, v) −→ (SS 0 , y, v 0 )
e/e0
(S, xy, v) −−−→
&e
Figure 7. Semantics of AST operators
(SS 0 , y, v 0 )
e
→ (SS 0 , y, v 0 )
(S, xy, v) −
hsymbol Ai
A
(S, xy, T ) −→ (S, y, T 0 )
&e
(S, xy, v) −−→ (SS 0 , xy, v 0 )
!e
hsymbol Ai
(S, xy, T ) −−−−−−−−→ (S[A, x], y, T 0 )
e
(S, x, v) −
→•
hblock ei
!e
(S, x, v) −→ (S, x, v)
e
(S, xy, T ) −
→ (SS 0 , y, T 0 )
hblock ei
(S, xy, T ) −−−−−−→ (S, y, T 0 )
Figure 6. Semantics of PEG operators
hlocal A ei
e
0
0
The transition form (S, xy, v) −
→ (S , x, v ) may be read: In
symbol table S, the input sequence xy is reduced to characters y
while evaluating e (i.e, e consumes x). Simultaneously, the semantic value v is evaluated to v 0 . The symbol table is unchanged if
e
(S, x, v) −
→ (SS 0 , y, v 0 ) and S 0 is empty. Likewise, the AST node
e
is not mutated if (S, x, v) −
→ (S 0 , y, v 0 ) and v = v 0 .
We write • for a special failure state, suggesting backtracking
to the alternative if it exists. Here is an example failure transition.
e
SĀ = del(S, A) (SA , xy, T ) −
→ (S 0 , y, T 0 )
hlocal A ei
(S, xy, T ) −−−−−−−→ (S, y, T )
hexists Ai
hexists Ai
(S, x, T ) −−−−−−−→ (S, x, T )
hexists A xi
∃[A, z] ∈ S
hexists A xi
a
(S, bx, v) −
→ • (a 6= b)
Figure 6 is the definition of the operational semantics of PEG
operators. Due to the space constraint, we omit the failure transition case. Instead, undefined transitions are regarded as the failure
transition. The semantics of Nez is conservative and shares with the
same interpretation of PEGs. Notably, the and-predicate &e allows
the state transitions of S 7→ SS 0 and v 7→ v 0 , meaning that the
lookahead definition of symbols and the duplication of local AST
nodes.
Figure 7 and Figure 8 show the semantics of AST operators and
symbol operators. The recognition of Nez is AST-independent, because any values v are not premise for the transitions. Accordingly,
an AST-eliminated grammar accepts the same input with the original Nez grammar.
Formally, the language generated by the expression e is L(S, e) =
e
{x|(S, xy) −
→ (S 0 , y)}, and the language of grammar L(G) is
es
L(G) = {x|(, xy) −→
(S, y)}. Note that the semantic value is
unnecessary in the language definition. As suggested in the study
of predicated grammars[29], the language class of L(G) is considered to be context-sensitive because predicates can check both the
left and right context.
4.
(S, x, T ) −−−−−−−−→ (S, x, T )
hmatch Ai
top(S, A) = z
hmatch Ai
(S, zx, T ) −−−−−−−→ (S, x, T )
his Ai
A
(S, zx, T ) −→ (S, x, T 0 ) top(S, A) = z
his Ai
(S, zx, T ) −−−−→ (S, x, T 0 )
hisa Ai
A
(S, zx, T ) −→ (S, x, T 0 ) ∃[A, z] ∈ S
hisa Ai
(S, zx, T ) −−−−−→ (S, x, T 0 )
Figure 8. Semantics of Nez’s symbol operators
is that conditional parsing can be replaced with symbol operators
and an empty symbol. In addition, we can eliminate conditions
from a Nez grammar, and Nez parsers usually run based on the
eliminated grammar. This section describes how to eliminate the
conditions.
For simplicity, we consider the a single parsing condition, labeled c. Let x be a Boolean variable such as x ∈ {c, !c}, where !c
stand for not c, and f (e, x) be a conversion function of an expression e into the eliminated one.
Elimination of Parsing Condition
In the previous section, we described the formal definition of Nez
grammar language without conditional parsing. One reason for this
Unpublished, draft paper to be revised. Comments and suggestions are welcome
∃z[A, z] ∈ S
6
2015/11/30
Here we write Ḡ for the eliminated grammar. Eliminating ccondition from G is a conversion from G into Ḡ, and defined: for
each production A = e in G, two new productions Ac and A!c are
added into Ḡ, as follows:
function calls. Backtracking is simply implemented over the call
stacks.
Notoriously, backtracking might cause an exponential time
parsing in the worst cases, while packrat parsing[7] is well established to guarantee the linear time parsing with PEGs. The idea
behind packrat parsing is a memoization whereby all results of
nonterminal calls are stored in the memoization table to avoid redundant nonterminal calls. Although the memoization table may
require some complexity for its efficiency, we use a simple and
constant-memory memoization table, presented in [21].
In addition to the standard PEG parser runtime, Nez parsers require two additional state management to handle ASTs and symbols. Note that we assume that conditions are eliminated upfront,
as described in Section 4.
The AST construction runtime provides a parser with APIs
that are based on AST operators in Nez. Importantly, Nez parsers
are speculative paring in a way that some of the AST operations
may be discarded when backtracking. To handle the consistency
of ASTs, we support transactional operations (such as commit
and abort) and all operations are stored as logs to be committed.
This transactional structure can be easily implemented with stackbased logging, and the AST construction is partially aborted at the
backtracking time. Note that efficient packrat parsing with ASTs
requires additional transactional management for ASTs. A detailed
mechanism is reported in [22].
The symbol table requires another state management. As with
the operation logs in the AST constructor, we use a stack-based
structure to control both the symbol scoping and backtracking consistency. The symbol table runtime provides a parser with APIs that
enable adding symbols, eliminating symbols, and testing symbols
to match the input.
Originally, we write the AST runtime and the symbol table runtime in Java. There is no use of functional data structure; instead,
both operation logs and symbols are stored in a linked list, which
is available in any programming language. Indeed, we have already ported Nez runtime into several languages, including C and
JavaScript. The C version of the AST runtime is at most 500 lines
in code and the symbol table runtime is at most 200 lines in code,
suggesting its high portability.
Ac = f (e, c) A!c = f (e, !c)
Now, we define the conversion function f (e, x) recursively:
• f (A, x) = Ac if x = c, or A!c if x =!c
• f (e1 e2 , x) = f (e1 , x)f (e2 , x)
• f (e1 /e2 , x) = f (e1 , x)/f (e2 , x)
• f (&e, x) = &f (e, x)
• f (!e, x) = !f (e, x)
• f (<if l c>, x) = if x = c, ! if x =!c
• f (<if l !c>, x) = ! if x = S! c, if x = c
• f (<on c e>, x) = f (e, c)
• f (<on !c e>, x) = f (e, !c)
• f (e, x) = e if e is none of the above
The number of productions in Ḡ is twice as many as that of G.
That means that a grammar in which n conditions are eliminated
results in O(2n ) productions in worst cases. In practice, we can
make some unification for the same production such that Ac = A!c ,
which may considerably reduce the number of productions in Ḡ.
Our empirical study (as described in Section 6) suggests that a
single grammar involves not so many conditions that it would cause
such an exponential increase of productions.
Note that the significant reason why we eliminate conditions is
that condition operators may cause the serious invalidation of packrat parsing. In general, packrat parsing works on the assumption
of stateless parsing[7], while Nez adds new states such as ASTs,
the symbol table, and parsing conditions. As we showed in Section 3, the state of ASTs does not influence any parsing behavior.
The symbol table usually causes state changes with some character
consumption, resulting in a fact that nonterminals rarely produce
different results in the same position. As Grimm pointed out in [9],
the flow-forward state change is not problematic in packrat parsing.
However, conditional parsing always causes state changes without
any character consumption. This makes it easier to make a situation
of different results of the same nonterminal in the same position. In
fact, packrat parsing does not work without eliminating conditions
from a Nez grammar.
5.
5.2
Parser Runtime and Implementation
One of the advantages in open grammar is the freedom from parser
implementation methods. In fact, the Nez parser tool, which we
have developed with the Nez language, can generate three types
of differently implemented parsers. First, as well as traditional
parser generators, Nez produces parser source code written in the
target language of the parser application. Second, Nez provides a
grammar translator into other existing PEG-based grammars, (such
as Rats!, PEGjs, or PEGTL). Third, Nez itself works as an efficient
parser interpreter that loads a grammar at runtime. In this section,
we briefly describe these three implementations.
5.1
Parser Runtime and Parser Generation
The parser runtime required in Nez is fundamentally lightweight
and portable. Nez parsers, as well as PEG parsers, can be implemented with recursive descent parsing with backtracking. All productions are simply implemented with parse functions that compute
on the input characters, and then nonterminal calls are computed by
Unpublished, draft paper to be revised. Comments and suggestions are welcome
Grammar Translation
A Nez grammar is specified with a declarative form of the AST
operators and symbol operators, which are performed as a kind of
action in the parsing context. These actions are limited in number
and can be statically translated into code fragments written in
any programming language. This means that a Nez grammar is
convertible into PEGs with embedded semantic action code.
Grammar translation is another approach to the Nez parser implementation. The advantage is that we enjoy optimized parsers that
existing PEG tools generate. On the other hand, impedance mismatch occurs between two grammars. In particular, many existing
PEG-based parser generators have their own supports for AST representations. For example, PEGjs[24] produces a JSON object as a
semantic value of all nonterminal calls. In those cases, the translation of AST operators is unnecessary in practice. The symbol operators are always translatable if the symbol runtime that run on a
host language is readily available.
Figure 9 shows an example of converting symbol operators into
semantic actions { ... } and semantic predicates &{ ... } .
The rollback of the symbol table at the backtracking time is automated before attempting alternatives. As shown, a Nez grammar is
recursively convertible into PEGs with action code using the Nez
runtime.
Currently, the Nez tool provides the grammar translation into
PEGjs, PEGTL, and LPeg, although the translation of the AST
7
2015/11/30
e / e0
<symbol A>
<block e>
<local A e>
<exists A>
<exists A x>
<match A>
<is A>
<isa l>
{symbolTable.startBlock()}
e { symbolTable.commit() }
/ {symbolTable.abort()} e0
{ symbolTable.add(A, capture(A)) }
{ symbolTable.startBlock() }
e
{ symbolTable.abortBlock() }
{ symbolTable.startBlock().mask(A) }
e
{ symbolTable.abortBlock() }
&{ symbolTable.count(A) > 0 }
&{ symbolTable.top(A) == x }
&{ match(symbolTable.top(A)) }
&{ symbolTable.top(A) == capture(A) }
&{ symbolTable.contains(A, capture(A)) }
τ (A = e, L)
=
τ (, L)
τ (a, L)
τ (A, L)
τ (e1 e2 , L)
=
=
=
=
τ (e1 /e2 , L)
=
τ (&e, L)
=
τ (!e, L)
=
AST
Symbol
Memo
L0
L0
L0
Figure 9. Implementing Nez symbol operators with action code
PEG
A
L0
L0
nop fail alt succ jump call ret pos back skip
byte any
tpush tpop tleftfold tnew tcapture ttag treplace tstart tcommit tabort
sopen sclose smask symbol exists isdef
match is isa
lookup memo memofail tlookup tmemo
τ ({e}, L)
=
τ ($(e), L)
=
L0
L0
Table 2. Instructions of the Nez parsing machine
τ (#t, L)
τ (hblock ei, L)
=
=
τ (hlocal A ei, L)
=
τ (hsymbol Ai, L)
=
τ (hexists Ai, L)
τ (hmatch Ai, L)
τ (his Ai, L)
=
=
=
L0
construction might be restricted due to the impedance mismatch.
Note that translating a Nez grammar into a LALR or LL grammar
is an interesting challenge, though beyond the scope of this paper.
5.3
L0
Virtual Parsing Machine
L0
The simplicity of PEGs makes it easier to achieve dynamic parsing in a way that a parser interpreter loads a grammar at runtime to
parse the input. Since dynamic parsing requires no source compilation process, dynamic parsing makes it easier to import the Nez
parser functionality as a parser library into parser applications. We
consider that dynamic parsing is better suitable for many use cases
of open grammars. Accordingly, the standard Nez parser is based
on dynamic parsing and then implemented on top of a virtual parsing machine, an efficient implementation of dynamic parsing.
A Nez parsing machine is a stack-based virtual machine that
runs with a set of bytecode instructions, specialized for PEG, AST,
symbol, and memoization operators. Table 2 is a summary of the
bytecode instructions. A parsing expression e in Nez is converted
to bytecode by using a compile function τ (e, L), where L is a label representing the next code point. Figure 10 shows an excerpt
of the definition of τ (e, L). The byte compilation is simply inductive, while Nez supports additional super instructions to generate
optimized bytecode.
Currently, the Nez parser tool includes a bytecode compiler and
a parsing machine, both written in Java. In addition, the parsing
machine is ported into C and JavaScript. Especially, the C version
of the Nez parsing machine is highly optimized with indirect thread
code and SEE4 instructions. The performance evaluation is studied
in Section 7.
6.
L0
nop
byte a
call A
τ (e1 , L0 )
τ (e2 , L)
alt L0
τ (e1 , L)
succ
τ (e2 , L)
pos
τ (e, L0 )
back
alt L
τ (e, L0 )
succ
fail
tnew
τ (e, L0 )
tcapture
tpush
τ (e, L0 )
tlink
tpop
ttag #t
sopen
τ (e, L0 )
sclose
sopen
mask A
τ (e, L0 )
sclose
pos
τ (A, L0 )
symbol
exists A
match A
pos
τ (A, L0 )
is A
Figure 10. Definition of a compile function of e with core instructions
Language
C
C#5.0
CoffeeScript
Java8
JavaScript
Konoha
Lua5
Python3
Ruby
#G
102
897
121
185
153
163
102
139
420
#P
102
1325
121
185
166
190
105
155
580
Symbols and Conditions
TypeName
AwaitKeyword, Global
Indent
InOperator
Offside
MultiLineBracketString
Indent, Offside
Delimiter, Primary, DoExpr
Table 3. Grammar summary: the typewriter font stands for a
symbol name and the italic font stands for a condition name
Case Studies and Experiences
of parser productions through eliminating the conditional operators. The same numbers of productions in #G and #P means no use
of conditional parsing. The column labeled ”Symbols and Conditions” indicates nonterminals and conditions that are used in the
grammar.
Here are some short comments on each of the interesting developed grammars.
We have developed many Nez grammars, ranging from programming languages to data formats. All developed grammars are available online at http://github.com/nez-peg/nez-grammar.
Table 3 shows a summary of the major developed grammars.The
column labeled ”#G” indicates the number of defined productions
in a grammar, while the column labeled ”#P” indicates the number
Unpublished, draft paper to be revised. Comments and suggestions are welcome
τ (e, L0 )
ret
8
2015/11/30
• C – Based on two PEG grammars written in Mouse. The se1000
• C# – Developed from scratch, referencing C#5.0 Language
100
Latency
[ms]
mantic code embedded to handle the typedef-defined name is
converted into the TypeDef symbol. The nested typedef statements are not implemented.
Specification (written in a natural language). The condition
AwaitKeyword is used to express contextual keyword await.
• Java8 – Ported from Java8 grammar written in ANTLR41 . Java
can be specified without any Nez extensions.
• JavaScript – Based on JavaScript grammar for PEG.js2 . We
1
simplify the grammar specification using the parsing condition
InOperator that distinguishes expression rules from the for/in
context.
parse
match
parse
match
parse
match
parse
match
nez-‐j
processing
252
91
jquery
37
27
java20
110
70
xmark
185
117
• Konoha – Developed from scratch. Konoha is a statically typed
scripting language, which we have designed in [20]. The most
syntactic constructs come from Java, but we use the parsing
condition SemicolonInsertion for expressing the conditional end of a statement. More importantly, Konoha is a language implementation that uses ASTs constructed by the Nez
parser.
cnez
58
13
9
4
25
8
41
9
nez-‐c
112
30
16
9
55
25
118
35
nez-‐js
1746
476
241
120
801
494
753
405
Figure 12. Performance comparison of differently implemented
Nez parsers: nez-j, nez-c, cnez, and nez-js
• Python3 – Ported from Python 3.0 abstract grammar3 . The
Intel Core i7-4770, 8GB of DDR3 RAM, and running on Linux
Ubuntu 14.04.1 LTS. All C programs are compiled with clang3.6, all Java parsers run on Oracle JDK 1.8, and all JavaScript
parsers run on node.js version 5.0. Tests are run several times, and
we record the average time for each iteration. The execution time
is measured in millisecond by System.nanoTime() in Java APIs
and gettimeofday() in Linux. Tested source files are randomly
collected from major open source repositories. To improve the time
accuracy, we eliminate small files from the data set.
All tested Nez grammars are available online as described in
Section 6. For convenience, we label tested grammars: java.nez,
js.nez, python3.nez, xml.nez, and xmlsym.nez. Note that xmlsym.nez is a symbol version of xml.nez, which checks whether the
closed tag is equal to the open tag, as described in Section 2. The
two tested grammars python3.nez and xmlsym.nez contain symbol
operators in a grammar.
Using Nez grammars, we generate Nez parsers for Java, C,
and JavaScript. The generated parsers are grouped by the base
implementation with the following labels:
Indent is used to express the indentation-based code block by
capturing white spaces.
• Ruby – Developed from scratch. The Delimiter is used to
express the context-sensitive delimiting identifier in the Here
document.
• Haskell – Postponed. We didn’t express the indent symbol for
Haskell’s code layout since the indentation starts in the middle
of expressions. This is mainly because of a limitation of a
symbol operator in Nez; if Nez provided Haskell-specialized
symbol operator such as <haskell-indent>, we could handle
the indentation-based code layout through the symbol tables.
Two findings are confirmed throughout our case studies. First,
many of the programming languages, as listed in Table 3, can not
be well expressed with pure PEG. The introduction of symbol tables significantly improve the expressiveness, although some corner cases might remain. Even in corner cases, we consider that the
symbol table approach is still applicable. Second, the parsing condition, although it does not directly improve the expressiveness of
PEGs, simplifies the specification task. The growth of #D/#G indicates that the grammar developer would specify as many productions as indicated at #G if the conditional parsing were not supported in Nez. Furthermore detailed discussions and earlier reports
on grammar developments with examples are available in [23].
All grammars we have developed include the AST construction.
At the moment, the only working example of a language implementation with Nez is with Konoha, a reimplemented version of [20] as
a JVM-based language implementation, including type checking
and code generation on top of the ASTs that are constructed by a
Nez parser. Since each AST node records the source location when
capturing a string, Konoha can carry out error reporting, although
it is very primitive.
7.
10
• nez-j - an original Nez parsing machine on JVM, incorporated
as a standard Nez parser,
• nez-c - a C-ported Nez parsing machine,
• nez-js - a JavaScript-ported Nez parsing machine, and
• cnez - generated C parsers from the Nez tool
Linear time paring is a primary concern of parser implementations. Figure 11 shows the parsing time plotted against file sizes in,
respectively, Java, JavaScript, and Python. The lines are the linear
regression line fit of nez-j, cnez, and nez-c parsers, suggesting the
paring time is linear. Since Nez parsers are integrated with packrat parsing, we have not observed any super-linear parser behavior
throughout this experiment.
To make detailed analysis, we have chosen some of tested
source files from the collection. The chosen data are labeled as:
Performance Study
This section is an experimental report on our performance and
comparative study. We start by describing common setups. All
tests in this paper are measured on DELL XPS-8700, with 3.4GHz
• processing – a JavaScript version of the Processing runtime
• jquery – a very popular JavaScript library. We used a com-
pressed source.
1 https://github.com/antlr/grammars-v4/blob/master/java8/Java8.g4
• java20 – 20 Java files collection, consisting of relatively large
2 https://github.com/pegjs/pegjs/blob/master/examples/javascript.pegjs
files (more than 10KB size). The parsing time is recorded as the
summation of all 20 parsing times.
3 https://docs.python.org/3.0/library/ast.html
Unpublished, draft paper to be revised. Comments and suggestions are welcome
9
2015/11/30
Java
(java.nez)
JavaScript
(js.nez)
100
nez-‐j
Latency
[ms]
Python
(python3.nez)
100
100
cnez
nez-‐c
nez-‐j
nez-‐j
cnez
nez-‐c
10
10
10
1
1
1
0.1
0.1
0.1
1
10
1
100
nez-‐c
0.01
0.01
0.01
cnez
10
100
1
1000
10
100
File
Size
[KiB]
Figure 11. Parsing time of the open source files for Java, JavaScript, and Python against the input size plotted as log-log base 10.
10000
Latency
[ms]
1000
100
10
1
processing
jquery
java20
parse
match
parse
match
match
cnez
58
13
9
4
25
8
nez-‐c
112
30
16
9
55
25
yacc
18
5
parse
match
parse
match
parse
nez-‐j
252
91
37
27
110
70
parse
match
parse
Rats!
antlr4
1598
1483
9167
584
579
2199
191
128
150
nez-‐js
1745
476
240
120
801
494
pegjs
1627
2318
Figure 13. Performance comparison of processing, jquery, and java20 in various parsers. The pegjs parser cannot end parsing process due
to its out of memory errors.
200
250
180
200
160
150
Latency
[ms]
140
Latency
[ms]
120
100
80
60
100
50
0
40
20
0
parse
match
cnez
nez-‐c
41
9
118
35
yacc
69
xerces-‐
c
85
49
libxml
nez-‐j
antlr4
74
31
185
117
195
non
(xml)
symbol
(xmlsym)
xerces-‐
java
75
56
parse
match
nez-‐j
185
117
217
158
non
(xml)
parse
match
cnez
41
44
9
11
parse
match
nez-‐c
118
35
137
38
symbol
(xmlsym)
Figure 15. Effects of symbol operators
Figure 14. Parsing time in xmark
2.5 times faster than nez-java, and almost 10 times faster than the
nez-js.
Next, we will turn to the performance comparison with existing
standard parsers. Here, we have chosen Yacc[14], ANTLR4[30],
Rats![9], and PEGjs[24], since they notably produce efficient
parsers and are accepted in many serious projects. We set up these
parser generators as follows.
• xmark – a synthetic and scalable XML files that are provided by
XMark benchmark program [32]. We used a newly generated
file in 10MB.
Figure 12 shows the parse time of Nez parsers. The data point
parse indicates the parsing time including the AST construction,
while match indicates the parsing time without the AST construction. The time scale is plotted as log base 10 due to the variety of
results. The cnez parsers are fastest as easily predicted. In comparisons among the parsing machines, the nez-c parser is almost
Unpublished, draft paper to be revised. Comments and suggestions are welcome
• yacc – a standard LALR parser generator for C. In this exper-
iment, we suffered from the limited availability of grammars,
and could only run a JavaScript parser without AST construction.
10
2015/11/30
• antlr4 – a standard parser generator based on the extended
Data-dependent grammars [3, 13] share similar ideas in terms of
recognizing context-sensitive syntax. Briefly, these grammars uses
a variable to be bound to a parsed result, which seems to be the
addition of symbols in Nez. Data-dependent grammars are more
expressive in a way that they express more complex condition such
as ([n>0] OCTET{n:=n1}), while Nez is more declarative and
leads to better readability. In addition, the better expressiveness of
data-dependent grammars suggests a future promising extension of
Nez, including numerical values in the symbol table.
Among many formal grammars such as LALR(k) and LL(k),
PEGs are simple and seemingly a suitable foundation for implementing portable parser runtime. In the reminder of this section,
we would like to focus on PEG-specific related work.
Extensions. With Rats!, Robert Grimm has shown an elegant integration of PEGs with action code despite its speculative
parsing[9]. Many other PEG-based parser generators[10, 24, 31]
have adopted the semantic action approach for expressing syntax
constructs that PEGs can not recognize. Yet, there have been several declarative attempts for PEG extensions. Adams has recently
formalized Indent-Sensitive CFGs [1] for the indentation-based
code layout and has extended the PEG-version of IS-CFGs [2].
Iwama et al have extended PEGs with the e U e operator to combine a black-boxed parser for a natural language[12]. Compared to
these existing extensions, Nez more broadly covers various parsing
aspects.
AST Construction. Traditionally, a grammar developer usually
writes action code to construct some forms of ASTs as a semantic value[15]. However, writing the AST construction is tedious,
and many parser generators have some annotation-based supports
to automate the AST construction. The underlying idea is to filter trees that are derived from structural parse results of nonterminal calls. However, the filtering approach is problematic in PEGs
since PEGs disallows the left-recursive structure of derived trees.
Recently, handling the left recursion in PEGs has been established
in [27], although another annotation is needed for preserving operator precedence. Nez allows users to specify an explicit structure
of ASTs (including tags and labels), resulting in a more general
string-to-tree transducer. Notably, a similar capturing idea appears
in LPeg[11], a PEG-based pattern matching, but Nez enables more
structural complexity for capturing structured data.
Parser Runtime. Nez parser runtime is based on many of previous works reported in the literature[9, 21, 26, 31]. In particular,
the state management in speculative parsing is built on the partial
transaction management of Rats!. The Nez virtual machine is said
to be an extended version of the LPeg machine. Notably, the Nez
machine has instruction supports for packrat parsing [7], which is
a major lack in the LPeg machine. The originality of Nez parsers
is that we combine these established techniques in a way that the
parser runtime is still simple and portable.
LL(k) with some packrat parsing feature. All tested grammars
are derived from the official ANTLR4 grammar repository.
• rats! – a PEG-based parser generator[9] for Java. Java gram-
mar is derived from a part of the xtc implementation, while
JavaScript parser is generated from a Nez grammar js.nez by
the grammar translation.
• pegjs – a PEG-based parser generator[24] for JavaScript.
JavaScript grammar is derived from its official site, and Java
grammar is translated from java.nez by Nez tools.
Figure 13 shows the parsing time of processing, jquery, and
java20 in each of the examined parsers. All Nez parsers show better performance with competitive parsers in each of the host environments. In ANTLR4 and Rats!, some of results are surprisingly bad. We run fairly in terms of the time measurement and the
JIT condition, but we are still not sure that Rats! and ANTLR4
parsers are best optimized in the experiment. As a result, we conclude that Nez parsers achieve competitive performance compared
to these practical parser generators. Likewise, we compare Nez
parsers for xmark with standard XML parsers such as libxml and
xerces, which are highly optimized by hand. Figure 14 shows the
comparison of parsing time in xmark. The data points match and
parse in XML parsers correspond to SAX and DOM parsers, respectively. Although we have observed the overhead of AST construction in Nez parsers, they are also competitive.
Finally, we focus on the performance effect of symbol operators. Figure 15 shows the performance comparison of xml.nez and
xmlsym.nez. We confirm that the costs of symbol operators, requiring <block>, <symbol>, and <is> at each closed tag, are small at
the acceptable level.
8.
Related Work
Developing parsers is ubiquitous in many applications, and handwritten parsers are obviously prone to errors. Parser generation
from a formal grammar specification has a long history in programming language research. A common key challenge is balancing the
expressiveness of formal grammars and how to implement them in
practical manners. In the early 1970s, Stephen C. Johnson developed Yacc[14], based on a LALR(1) grammar with embedded C
code to perfom user-defined actions. Since Yacc successfully hits
a sweet spot between the expressiveness and practical parser generation, the use of action code has been broadly adopted in most
of the major parser generators [16, 17, 29] across various grammar
foundations, including LL(k), GLR, and PEG.
In parallel, the problems of action code have been pointed out,
for example, the lack of grammar reuse[28], decreased maintainability [19], and ad hoc behavior[1]. There have been several attempts to decrease the disadvantage of action code. Terence Parr,
the author of ANTLR, has proposed a prototype grammar, inspired
by the revision control system[28]. Elkhound[25], for instance, allows the users to write action code written in multiple programming
languages. Despite these attempts, grammar reuse is still limited
and arbitrary action inevitably requires porting to other languages.
The idea of open grammars has been strongly inspired by the
article: Pure and declarative syntax definition: paradise lost and
regained[17]. SDF+ADF and Stratego[4, 6] intend their users to
write a syntactic analysis using an algebraic transformation. The
actions are not arbitrary, but they are expressive enough. However,
the actions differ from those required in Nez, since SDF is based
on Generalized LR[33], where actions are mainly used to resolve
grammar ambiguity. Nez, on the other hand, is unambiguous as a
PEG, and symbol operators are used for handling state changes.
Unpublished, draft paper to be revised. Comments and suggestions are welcome
9.
Conclusion
Nez is a simple, portable, and declarative grammar specification
language. As an alternative to action code, Nez provides a small
set of extended PEG operators that allow the grammar developer
to make AST constructions, as well as context-sensitive and conditional parsing. Due to these extensions, Nez can recognize a language that cannot be well expressed by pure PEGs. In addition, the
Nez parser runtime is also simple and portable. Based on the Javabased implementation, we have developed C and JavaScript ports,
suggesting that the Nez parser runtime is implementable in any
modern programming languages. More importantly, Nez parsers
achieve practical and competitive performance in each of the ported
environments.
This paper is an initial report on Nez and the idea of open grammars. Further evaluations are obviously necessary on the expres-
11
2015/11/30
siveness of Nez with extensive case studies. To achieve the aim of
open grammars, a lot of interesting challenges remain. Future work
that we will investigate includes grammar modularity and inference, a tree checker for better connectability of parser applications,
and a DFA-based machine for more efficient parsing. In the end,
we hope that Nez parsers will be available in many programming
languages. Our developed tools and grammars are available online
at http://nez-peg.github.io/.
[10] C. Hirsch and D. Frey. Parsing expression grammar template library.
https://code.google.com/p/pegtl/, 2014.
[11] R. Ierusalimschy. A text pattern-matching tool based on parsing
expression grammars. Softw. Pract. Exper., 39(3):221–258, Mar. 2009.
ISSN 0038-0644. . URL http://dx.doi.org/10.1002/spe.
v39:3.
[12] F. Iwama, T. Nakamura, and H. Takeuchi. Constructing parser for
industrial software specifications containing formal and natural language description. In Proceedings of the 34th International Conference on Software Engineering, ICSE ’12, pages 1012–1021, Piscataway, NJ, USA, 2012. IEEE Press. ISBN 978-1-4673-1067-3. URL
http://dl.acm.org/citation.cfm?id=2337223.2337353.
Acknowledgments
The C version of Nez parsers (cnez and virtual parsing machines) have been developed by Masahiro Ide and Shun Honda.The
JavaScript version has been developed by Takeru Sudo. Finally, the
anonymous reviewers are acknowledged for their helpful suggestions for improving an earlier draft of this paper.
[13] T. Jim, Y. Mandelbaum, and D. Walker. Semantics and algorithms
for data-dependent grammars. In Proceedings of the 37th Annual
ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages, POPL ’10, pages 417–430, New York, NY, USA, 2010.
ACM. ISBN 978-1-60558-479-9. . URL http://doi.acm.org/
10.1145/1706299.1706347.
References
[14] S. C. Johnson and R. Sethi. Unix vol. ii. chapter Yacc: A Parser
Generator, pages 347–374. W. B. Saunders Company, Philadelphia,
PA, USA, 1990. ISBN 0-03-047529-5. URL http://dl.acm.org/
citation.cfm?id=107172.107192.
[1] M. D. Adams. Principled parsing for indentation-sensitive languages:
Revisiting landin’s offside rule. In Proceedings of the 40th Annual
ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages, POPL ’13, pages 511–522, New York, NY, USA, 2013.
ACM. ISBN 978-1-4503-1832-7. . URL http://doi.acm.org/
10.1145/2429069.2429129.
[15] A. Johnstone and E. Scott. Tear-insert-fold grammars. In Proceedings
of the Tenth Workshop on Language Descriptions, Tools and Applications, LDTA ’10, pages 6:1–6:8, New York, NY, USA, 2010. ACM.
ISBN 978-1-4503-0063-6. . URL http://doi.acm.org/10.1145/
1868281.1868287.
[2] M. D. Adams and O. S. Ağacan. Indentation-sensitive parsing for
parsec. In Proceedings of the 2014 ACM SIGPLAN Symposium on
Haskell, Haskell ’14, pages 121–132, New York, NY, USA, 2014.
ACM. ISBN 978-1-4503-3041-1. . URL http://doi.acm.org/
10.1145/2633357.2633369.
[16] M. Jonnalagedda, T. Coppey, S. Stucki, T. Rompf, and M. Odersky.
Staged parser combinators for efficient data processing. In Proceedings of the 2014 ACM International Conference on Object Oriented
Programming Systems Languages & Applications, OOPSLA ’14,
pages 637–653, New York, NY, USA, 2014. ACM. ISBN 978-14503-2585-1. . URL http://doi.acm.org/10.1145/2660193.
2660241.
[3] A. Afroozeh and A. Izmaylova. One parser to rule them all. In 2015
ACM International Symposium on New Ideas, New Paradigms, and
Reflections on Programming and Software (Onward!), Onward! 2015,
pages 151–170, New York, NY, USA, 2015. ACM. ISBN 978-14503-3688-8. . URL http://doi.acm.org/10.1145/2814228.
2814242.
[17] L. C. Kats, E. Visser, and G. Wachsmuth. Pure and declarative syntax
definition: Paradise lost and regained. In Proceedings of the ACM
International Conference on Object Oriented Programming Systems
Languages and Applications, OOPSLA ’10, pages 918–932, New
York, NY, USA, 2010. ACM. ISBN 978-1-4503-0203-6. . URL
http://doi.acm.org/10.1145/1869459.1869535.
[4] M. Bravenboer and E. Visser. Concrete syntax for objects: Domainspecific language embedding and assimilation without restrictions.
In Proceedings of the 19th Annual ACM SIGPLAN Conference on
Object-oriented Programming, Systems, Languages, and Applications, OOPSLA ’04, pages 365–383, New York, NY, USA, 2004.
ACM. ISBN 1-58113-831-8. . URL http://doi.acm.org/10.
1145/1028976.1029007.
[18] L. C. L. Kats, R. Vermaas, and E. Visser. Integrated language definition testing: Enabling test-driven language development. In K. Fisher,
editor, Proceedings of the 26th Annual ACM SIGPLAN Conference on
Object-Oriented Programming, Systems, Languages, and Applications
(OOPSLA 2011), Portland, Oregon, USA, 2011. ACM.
[5] M. Bravenboer, E. Tanter, and E. Visser. Declarative, formal, and
extensible syntax definition for aspectj. In Proceedings of the 21st
Annual ACM SIGPLAN Conference on Object-oriented Programming
Systems, Languages, and Applications, OOPSLA ’06, pages 209–228,
New York, NY, USA, 2006. ACM. ISBN 1-59593-348-4. . URL
http://doi.acm.org/10.1145/1167473.1167491.
[19] P. Klint, T. van der Storm, and J. Vinju. On the impact of dsl tools
on the maintainability of language implementations. In Proceedings
of the Tenth Workshop on Language Descriptions, Tools and Applications, LDTA ’10, pages 10:1–10:9, New York, NY, USA, 2010. ACM.
ISBN 978-1-4503-0063-6. . URL http://doi.acm.org/10.1145/
1868281.1868291.
[6] M. Bravenboer, K. T. Kalleberg, R. Vermaas, and E. Visser. Stratego/xt
0.17. a language and toolset for program transformation. Sci. Comput.
Program., 72(1-2):52–70, June 2008. ISSN 0167-6423. . URL
http://dx.doi.org/10.1016/j.scico.2007.11.003.
[20] K. Kuramitsu. Konoha: implementing a static scripting language with
dynamic behaviors. In S3 ’10: Workshop on Self-Sustaining Systems,
pages 21–29, New York, NY, USA, 2010. ACM. ISBN 978-1-45030491-7. .
[7] B. Ford. Packrat parsing:: Simple, powerful, lazy, linear time, functional pearl. In Proceedings of the Seventh ACM SIGPLAN International Conference on Functional Programming, ICFP ’02, pages 36–
47, New York, NY, USA, 2002. ACM. ISBN 1-58113-487-8. . URL
http://doi.acm.org/10.1145/581478.581483.
[21] K. Kuramitsu. Packrat parsing with elastic sliding window. Journal of
Infomration Processing, 23(4):505–512, 2015. .
[22] K. Kuramitsu. Fast, flexible, and declarative consturction of abstract
syntax trees with PEGs. Journal of Infomration Processing, 24(1):(to
appear), 2016.
[8] B. Ford. Parsing expression grammars: A recognition-based syntactic
foundation. In Proceedings of the 31st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’04, pages
111–122, New York, NY, USA, 2004. ACM. ISBN 1-58113-729-X. .
URL http://doi.acm.org/10.1145/964001.964011.
[23] K. Kuramitsu and T. Matsumura. A declarative extension of parsing
expression grammars for recognizing most programming languages.
Journal of Infomration Processing, 24(2):(to appear), 2016.
[9] R. Grimm. Better extensibility through modular syntax. In Proceedings of the 2006 ACM SIGPLAN Conference on Programming
Language Design and Implementation, PLDI ’06, pages 38–51, New
York, NY, USA, 2006. ACM. ISBN 1-59593-320-4. . URL http:
//doi.acm.org/10.1145/1133981.1133987.
Unpublished, draft paper to be revised. Comments and suggestions are welcome
[24] D. Majda. Peg.js - parser generator for javascript, 2015.
http://http://pegjs.org/.
URL
[25] S. McPeak and G. Necula. Elkhound: A fast, practical glr parser
generator. In E. Duesterwald, editor, Compiler Construction, volume
12
2015/11/30
2985 of Lecture Notes in Computer Science, pages 73–88. Springer
Berlin Heidelberg, 2004. ISBN 978-3-540-21297-3.
[26] S. Medeiros and R. Ierusalimschy. A parsing machine for pegs. In
Proceedings of the 2008 Symposium on Dynamic Languages, DLS
’08, pages 2:1–2:12, New York, NY, USA, 2008. ACM. ISBN 9781-60558-270-2. . URL http://doi.acm.org/10.1145/1408681.
1408683.
[27] S. Medeiros, F. Mascarenhas, and R. Ierusalimschy. Left recursion in
parsing expression grammars. Sci. Comput. Program., 96(P2):177–
190, Dec. 2014. ISSN 0167-6423. . URL http://dx.doi.org/10.
1016/j.scico.2014.01.013.
[28] T. Parr. The reuse of grammars with embedded semantic actions. In
Proceedings of the 2008 The 16th IEEE International Conference on
Program Comprehension, ICPC ’08, pages 5–10, Washington, DC,
USA, 2008. IEEE Computer Society. ISBN 978-0-7695-3176-2. .
URL http://dx.doi.org/10.1109/ICPC.2008.36.
[29] T. Parr and K. Fisher. Ll(*): The foundation of the antlr parser
generator. In Proceedings of the 32Nd ACM SIGPLAN Conference on
Programming Language Design and Implementation, PLDI ’11, pages
425–436, New York, NY, USA, 2011. ACM. ISBN 978-1-4503-06638. . URL http://doi.acm.org/10.1145/1993498.1993548.
[30] T. Parr, S. Harwell, and K. Fisher. Adaptive ll(*) parsing: The
power of dynamic analysis. In Proceedings of the 2014 ACM International Conference on Object Oriented Programming Systems
Languages & Applications, OOPSLA ’14, pages 579–598, New
York, NY, USA, 2014. ACM. ISBN 978-1-4503-2585-1. . URL
http://doi.acm.org/10.1145/2660193.2660202.
[31] R. R. Redziejowski. Parsing expression grammar as a primitive
recursive-descent parser with backtracking. Fundam. Inf., 79(3-4):
513–524, Aug. 2007. ISSN 0169-2968. URL http://dl.acm.org/
citation.cfm?id=2367396.2367415.
[32] A. Schmidt, F. Waas, M. Kersten, M. J. Carey, I. Manolescu, and
R. Busse. Xmark: A benchmark for xml data management. In
Proceedings of the 28th International Conference on Very Large Data
Bases, VLDB ’02, pages 974–985. VLDB Endowment, 2002. URL
http://dl.acm.org/citation.cfm?id=1287369.1287455.
[33] E. Visser. Syntax Definition for Language Prototyping. PhD thesis,
University of Amsterdam, September 1997.
Unpublished, draft paper to be revised. Comments and suggestions are welcome
13
2015/11/30
| 6cs.PL
|
An Experimental Comparison of Uncertainty
Sets for Robust Shortest Path Problems
Trivikram Dokka and Marc Goerigk
arXiv:1704.08470v1 [math.OC] 27 Apr 2017
Department of Management Science, Lancaster University, Lancaster, United Kingdom
Abstract
Through the development of efficient algorithms, data structures and preprocessing techniques, real-world shortest path problems in street networks are now very
fast to solve. But in reality, the exact travel times along each arc in the network
may not be known. This lead to the development of robust shortest path problems,
where all possible arc travel times are contained in a so-called uncertainty set of
possible outcomes.
Research in robust shortest path problems typically assumes this set to be given,
and provides complexity results as well as algorithms depending on its shape. However, what can actually be observed in real-world problems are only discrete raw
data points. The shape of the uncertainty is already a modelling assumption. In
this paper we test several of the most widely used assumptions on the uncertainty set
using real-world traffic measurements provided by the City of Chicago. We calculate
the resulting different robust solutions, and evaluate which uncertainty approach is
actually reasonable for our data. This anchors theoretical research in a real-world
application and allows us to point out which robust models should be the future
focus of algorithmic development.
Keywords:
study
robust shortest paths, uncertainty sets, real-world data, experimental
1 Introduction
The problem of finding shortest paths in real-world networks has seen considerable algorithmic improvements over the last decade [BDG+ 16]. In the typical problem setup, one
assumes that all data is given exactly. But also robust shortest path problems have been
considered, where travel times are assumed to be given by a set of possible scenarios. In
[YY98], it was shown that the problem of finding a path that minimizes the worst-case
length over two scenarios is already weakly NP-hard. For general surveys on results in
robust discrete optimization, we refer to [ABV09, KZ16].
1
There are many possibilities how to model the scenario set that is used for the robust
optimization process (see, e.g., [MG04, Büs12]), and it is not obvious which is ”the
right” one. Part of the current research ignores the problem by simply assuming that
the uncertainty was ”given” in some specific form, while this does not happen in reality.
In fact, the starting point for all uncertainty sets is raw data, given as a set of observations of travel times. This data is then processed to fit different assumptions on the
shape and size of the uncertainty set, and preferences of the decision maker. So far, the
discussion of these uncertainty sets has been led by theoretical properties, such as the
computational tractability of the resulting robust model. We believe that this leads to a
gap in the literature, where models are not sufficiently underpinned by actual real-world
data to verify results. The purpose of this paper is to close this gap. We use real-world
traffic observations by the City of Chicago to create a selection of the best-known and
most-used uncertainty sets from the research literature. Using these uncertainty sets,
we calculate different robust solutions and compare their performance. This allows us
to determine which uncertainty sets are actually valuable for real-world robust shortest path problems. Our results give strong impulses for future research in the field by
pointing out which problems are most worthy to solve more efficiently.
In Section 2 we briefly introduce all six uncertainty sets used in this study, and discuss
the complexity of the resulting robust problems. The experimental setup and results are
then presented in Section 3, before concluding this paper in Section 4.
2 Uncertainty Sets for the Shortest Path Problem
Let a directed graph G = (V, A) with nodes V and arcs A be given. In the classic
shortest path problem, each arc e has some specific travel time ce ≥ 0. Given a start
node s and a target node t, the aim is to find a path minimizing the total travel time,
i.e., to solve
min{ccx : x ∈ X }
where X ⊆ {0, 1}n denotes the set of s-t-paths, and n = |A|. For our setting we assume
instead that a set R of travel time observations is given, R = {cc1 , . . . , c N } with c i ∈ Rn .
This is the available raw data. In the well-known robust shortest path problem we
assume that an uncertainty set U is produced based on this raw data, and solve the
robustified problem
min max cx ,
x ∈X c ∈U
that is, we search for a path that minimizes the worst-case costs over all scenarios. In the
following sections we detail different possibilities from the current literature to generate
U based on R. Each set is equipped with a scaling parameter to control its size.
2
2.1 Convex Hull
In this approach, also known as discrete uncertainty (see [YY98, KZ16]), we set U CH =
R. The resulting robust problem can then be written as
min z
s.t. z ≥ c ix
x∈X
∀i ∈ [N ]
Note that this problem is equivalent to using U CH = conv({cc1 , . . . , c N }). The problem
is known to be NP-hard already for two scenarios.
P
Scaling: Let ĉc be the average of {cc1 , . . . , c N }, i.e., ĉc = N1 i∈[N ] c i . For a given λ ≥ 0,
we substitute each point c i with ĉc + λ(cci −ĉc), and take the convex hull of the scaled data
points.
2.2 Intervals
j
j
i∈[n] [minj∈[N ] ci , maxj∈[N ] ci ].
minj∈[N ] cji . The resulting ro-
We set U I as the smallest hypercube containing all data, i.e., U I =
For ease of notation, we write ci := maxj∈[N ] cji and ci :=
bust problem is then
Q
min cx
s.t. x ∈ X
which is a classic shortest path problem. Robust shortest path problems with interval
uncertainty are therefore easy to solve, but frequently used, especially in the so-called
min-max regret setting (see [CG15b]).
h
i
Q
c +c
c −c c +c
c −c
Scaling: We use U I = i∈[n] i 2 i − λ i 2 i , i 2 i + λ i 2 i for some λ ≥ 0.
2.3 Ellipsoid
Ellipsoidal uncertainty sets were first proposed in [BTN98, BTN99] and stem from the
observation that the iso-density locus of a multivariate normal distribution is an ellipse.
We use an ellipsoid of the form U E = {cc : (cc − µ )tΣ −1 (cc − µ ) ≤ λ} with size parameter
λ ≥ 0 that is centered on ĉc. We create it using a normal distribution found as a
maximum-likelihood fit. Recall that the best fit of a multivariate normal distribution
µ, Σ ) with respect to data points c 1 , . . . , c N is given by
N (µ
µ = ĉc =
and
Σ=
1 1
(cc + . . . + c N )
N
1 X i
(cc − µ )(cci − µ )t
N
i∈[N ]
3
The resulting problem can then be formulated as
min ĉcx + z
s.t. z 2 ≥ λ x tΣx
x∈X
which is an integer second-order cone program (ISOCP), see [BTN98] for details. Due to
the convexity of the constraints, the problem can still be solved with little computational
effort by standard solvers.
2.4 Budgeted Uncertainty
This
approach was introduced in [BS03], and is based on interval uncertainty U =
Q
i∈[n] [ĉi , ci ]. To reduce the conservatism of this approach one assumes that only at
most Γ ∈ {0, . . . , n} many values can be simultaneously higher than the midpoint ĉc.
Formally,
X
U B = {cc : ci = ĉi + (ci − ĉi )δi for all i ∈ [n], 0 ≤ δ ≤ 1 ,
δi ≤ Γ}
i∈[n]
Using the dual of the inner worst-case problem, the following compact mixed-integer
program can be found:
X
min ĉcx + Γπ +
ρi
i∈[n]
s.t. π + ρi ≥ (ci − ĉi )xi
π, ρ ≥ 0
∀i ∈ [n]
x∈X
This approach has the advantage that probability bounds can be found that compare
favorably with those for ellipsoidal uncertainty [BS04], while this problem also remains
polynomially solvable by enumerating possible values for the π variable. This means
that O(n) many problems of the original type need to be solved. For these reasons, the
budgeted uncertainty approach has been very popular in the literature.
2.5 Permutohull
The final two uncertainty sets we consider were proposed in [BB09]. The original inspiration comes from risk measures; the authors show that any so-called distortion risk
measure leads to a polyhedral uncertainty set. A risk measure µ is a distortion risk
measure if and only if there exists q ∈ {qq 0 ∈ ∆N : q1 ≥ . . . qN }, where ∆N denotes the
N -dimensional simplex such that
X
x) = −
µ(x
qi (cc(i)x )
i∈[N ]
4
where the sorting (i) is chosen such that c (1)x ≥ . . . ≥ c (N )x .
The conditional value at risk CV aRα with α ∈ (0, 1] is a well-known distortion risk
measure. Intuitively, it is the expected value amongst the α worst outcomes. Using the
matrix
1 . . . N 1−2 N 1−1 N1
.. ..
..
..
..
. .
.
.
.
1
1
1
Q N := 0 0
N
−2
N
−1
N
1
1
0 . . .
0
N −1
N
1
0 ...
0
0
N
the jth column of QN induces the risk measure CV aRj/N . The corresponding polyhedra
are called the q -permutohull and defined as
X
Πq (cc1 , . . . , c N ) := conv
qσ(i)c i : σ ∈ SN
i∈[N ]
To find the resulting robust problem, we first consider the worst-case problem for fixed
x ∈ X.
X
max
qi (ccj x )pij
i,j∈[N ]
s.t.
X
pij = 1
∀j ∈ [N ]
pij = 1
∀i ∈ [N ]
i∈[N ]
X
j∈[N ]
pij ≥ 0
∀i, j ∈ [N ]
Dualising this problem then gives the robust counterpart
X
min
(vi + wi )
i∈[n]
s.t. vi + wj ≥ qi (ccj x )
v, w ≷ 0
∀i, j ∈ [N ]
x∈X
which is a mixed-integer program (note that this approach is actually the same as the
ordered weighted averaging method, see [CG15a]). The problem is NP-hard, as it contains the convex hull of {cc1 , . . . , c N } as a special case. Through the choice of q , there
are N possible sizes of this uncertainty.
5
2.6 Symmetric Permutohull
In the same setting as before, the symmetric permutohull was also introduced in [BB09].
By using the bN/2c + 1 columns of the matrix
1 2 2 ... 2
1 1 2 . . . 2
1 1 1 . . . 2
1
.. .. .. .. ..
Q :=
Q̃
. . . . .
N
1 1 1 . . . 0
1 1 0 . . . 0
1 0 0 ... 0
it was shown that the resulting polyhedra are symmetric with respect to ĉc. Note that
Q contains the min-max approach for N = 2 as a
these problems are also NP-hard, as Q̃
special case.
2.7 Summary of Uncertainty Sets
In total we described six methods to generate uncertainty set U based on the raw data
R. Figure 1 illustrates these sets using a raw dataset with four observations (shown
as red points). The complexity to solve the resulting robust models, as well as the
type of program with the numbers of additional variables and constraints compared
to the classic shortest path problem are shown in Table 1. While the robust model
Complexity
Model
Add. Const.
Add. Var.
U CH
NPH
IP
N
1
UI
P
LP
0
0
UE
NPH
ISOCP
1
1
UB
P
MIP
n+1
n
UPH
NPH
MIP
N2
2n
U SP H
NPH
MIP
N2
2n
Table 1: Uncertainty sets in this study.
with budgeted uncertainty sets can be solved in polynomial time using combinatorial
algorithms, we still used the MIP formulation for our experiments, as it was sufficiently
fast.
3 Real-World Experiments
3.1 Data Collection and Cleaning
We used data provided by the City of Chicago1 , which provides a live traffic data interface. We recorded traffic updates in a 15-minute interval over a time horizon of 24 hours
1
https://data.cityofchicago.org
6
1.5
1.5
1
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
-1.5
-1.5
-1
-0.5
0
0.5
1
1.5
2
-1.5
-1.5
2.5
(a) Convex hull with λ = 1 and λ = 0.5.
-1
-0.5
0
0.5
1
1.5
2
2.5
(b) Intervals with λ = 1 and λ = 0.5.
1.5
1.5
1
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
-1.5
-1.5
-1.5
-1.5
-1
-0.5
0
0.5
1
1.5
2
-1
-0.5
0
0.5
1
1.5
2
2.5
2.5
(d) Budgeted uncertainty with λ = 1, Γ = 1
and λ = 1, Γ = 1.5.
(c) Ellipsoid with λ = 3 and λ = 1.
1.5
1.5
1
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
-1.5
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
-1.5
-1.5
(e) Permutohull uncertainty for CV aR2/N
and CV aR3/N (i.e., q = ( 21 , 12 , 0, 0) and
q = ( 13 , 13 , 13 , 0).
-1
-0.5
0
0.5
1
1.5
2
2.5
(f) Symmetric permutohull uncertainty for
q = ( 21 , 12 , 0, 0) and q = ( 21 , 14 , 14 , 0).
Figure 1: Example uncertainty sets.
7
spanning Monday March 27th 2017 morning to Tuesday March 28th 2017 morning. A
total of 98 data observations were thus used.
Every observation contains the traffic speed for a subset of a total of 1,257 segments.
For each segment the geographical position is available, see the resulting plot in Figure 2a
with a zoom-in for the city center. The complete travel speed data set contains a total
of 54,295 observations. There were 1,027 segments where the data was recorded at least
once of the 96 time points. Nearly for 88% of the segments, speeds were recorded for at
least 50 records with only 1% (10 segments) where only one observation was recorded.
We used linear interpolation to fill the missing records keeping in mind that data was
collected over time. The data after removing missing records and filling missing values
can be found at www.lancaster.ac.uk/~goerigk/robust-sp-data.zip.
As segments are purely geographical objects without structure, we needed to create
a graph for our experiments. To this end, segments were split when they crossed or
nearly crossed, and start- and end-points that were sufficiently close to each other were
identified as the same node. The resulting graph is shown in Figure 2b; note that this
process slightly simplified the network, but kept its structure intact. The final graph
contains 538 nodes and 1308 arcs.
3.2 Setup
Each uncertainty set is equipped with a size parameter. For each parameter we generated
20 possible values:
• For U CH and U I , λ ∈ {0.1, 0.2, . . . , 2}.
• For U E , λ ∈ {0.2, 0.4, . . . , 4}.
• For U B , Γ ∈ {5, 10, . . . , 100}.
• For U P H , we used columns q 1 , q 3 , . . . , q 39 .
• For U SP H , we used columns q 1 , q 2 , . . . , q 20 .
Each uncertainty set is generated using only every second scenario (i.e., 48 out of 96),
but all 96 scenarios are then used to evaluate the solutions. Furthermore, we generated
200 random s − t pairs uniformly, and used each of the 6 · 20 methods on the same 200
pairs. Each of our 120 methods hence generates 200 · 96 = 19, 200 objective values.
It is highly non-trivial to assess the quality of these solutions, see [CG16b]. If one just
uses the average objective value, as an example, then one could as well calculate the
solution optimizing the average scenario case to find the best performance with respect
to this measure. To find a balanced evaluation of all methods, we used four performance
criteria:
• the average objective value,
• the average of the worst-case objective value for each s − t pair, and
8
42.05
42
41.95
41.9
41.85
41.8
41.75
41.7
41.65
-87.95 -87.9 -87.85 -87.8 -87.75 -87.7 -87.65 -87.6 -87.55
(a) Raw segments with zoom-in for the city center. In red are segments without data.
42.05
42
41.95
41.9
41.85
41.8
41.75
41.7
41.65
-87.95 -87.9 -87.85 -87.8 -87.75 -87.7 -87.65 -87.6 -87.55
9 zoom-in for the city center.
(b) Resulting graph model with
Figure 2: Chicago instance
• the average value of the worst 5% of objective values for each s − t pair (as in the
CVaR measure)
We also considered the average rank. To this end, we rank all 120 methods for each
specific combination of s − t pair and scenario. The best performing methods are ranked
at 1, the second-best at 2 etc. We then take the average rank over all 19, 200 observations.
However, this measure was strongly correlated with the average objective value and is
therefore not presented.
For all experiments we used a computer with a 16-core Intel Xeon E5-2670 processor,
running at 2.60 GHz with 20MB cache, and Ubuntu 12.04. Processes were pinned to
one core. We used CPLEX v.12.6 to solve all problem formulations.
3.3 Results
We present our findings in the two plots of Figure 3. In each plot, the 20 parameter
settings that belong to the same uncertainty set are connected by a line. They are
complemented with Figure 4 showing the total computation times for the methods over
all 200 shortest path calculations.
The first plot in Figure 3a shows the trade-off between the average and the maximum
objective value; the second plot in Figure 3b shows the trade-off between the average and
the average of the 5% worst objective values. Note that for all performance measures,
smaller values indicate a better performance – hence, good trade-off solutions should
move from the top left to the bottom right of the plots. In general, the points corresponding to the parameter settings that give weight to the average performance can be
on the left sides of the curves, while the more robust parameter settings are on the right
sides, as would be expected.
We first discus Figure 3a. In general, we find that most concepts to indeed present a
trade-off between average performance and robustness through their scaling parameter.
The symmetric permutohull solutions have the best average performance, while interval
solutions are the most robust. Interestingly, that even holds for interval solutions where
the scaling parameter is very small. The budgeted uncertainty does not give a good
trade-off between worst-case and average-case performance, which confirms previous
results on artificial data [CG16a]. Scaling interval uncertainty sets achieves better results
than using budgeted uncertainty. Solutions generated with ellipsoidal uncertainty sets
slightly outperform (dominate in the Pareto sense) solutions generated with permutohull
uncertainty. We also note that methods that are computationally more expensive tend
to achieve better average performance at the cost of decreases robustness. The simplest
and cheapest method, interval uncertainty, gives the most robust solutions. Solutions
using the convex hull of raw data tend to be outperformed by the approaches that process
data.
We now consider the results presented in Figure 3b. Here the average is plotted against
the average performance of the 5% worst performing scenarios, averaged over all s − t
pairs. We note that for interval uncertainty, these two criteria are connected, with the
best solutions for small parameter size dominating all solutions for larger parameter size.
10
12.4
Convex hull
Intervals
Ellipsoid
Budgeted
Permutohull
Symmetric Permutohull
12.2
12.0
Maximum
11.8
11.6
11.4
11.2
11.0
10.8
10.6
10.4
7.0
7.1
7.2
7.3
7.4
7.5
7.6
7.7
7.8
7.9
8.0
7.8
7.9
8.0
Average
(a) Average vs worst-case performance.
9.45
Convex hull
Intervals
Ellipsoid
Budgeted
Permutohull
Symmetric Permutohull
9.40
Average of Worst 5%
9.35
9.30
9.25
9.20
9.15
9.10
9.05
9.00
7.0
7.1
7.2
7.3
7.4
7.5
7.6
Average
(b) Average vs CVaR performance.
11
Figure 3: Performance
results.
7.7
10000
Convex hull
Intervals
Ellipsoid
Budgeted
Permutohull
Symmetric
Permutohull
Computation Time
1000
100
10
1
Small
Large
Uncertainty Size
Figure 4: Computation times.
For the permutohull and the ellipsoidal uncertainty solutions, the order slightly changed
with the former often dominating the latter. Permutohull solutions are designed to be
efficient for the CVaR criterion, and the best-performing solution with respect to this
aspect is indeed generated by this approach. However, also solutions with ellipsoidal,
interval and convex hull uncertainty perform well.
Regarding computation times (see Figure 4), note that the two polynomially solvable approaches are also the fastest when using Cplex; these computation times can
be further improved using specialized algorithms. Using the convex hull is faster than
using ellipsoids, which are in turn faster than using the symmetric permutohull. For
the standard permutohull, the computation times are sensitive to the uncertainty size;
if the q vector that is used in the model has only few entries, then computation times
are smaller. This is in line with the intuition that the problem becomes easier if less
scenarios need to be considered.
To summarize our findings in our experiment on the robust shortest path problem
with real-world data:
• Convex hull solutions are amongst the more robust solutions, but tend to be outperformed by the other approaches.
• Interval solutions perform bad on average, but are the most robust. Especially
when the scaling is small they can give a decent trade-off, and are easy and fast
to compute.
• Ellipsoidal uncertainty solutions have very good overall performance and represent
a large part of the non-dominated points in our results.
12
• We do not encourage the use of budgeted uncertainty for robust shortest path
problems. Scaling interval uncertainty sets gives better results and is easier to use
and to solve.
• Permutohull solutions offer good trade-off solutions, whereas symmetric permutohull solutions tend to be less robust, but provide an excellent average performance.
These methods are also require most computational effort to find.
In the light of these findings, the interval and discrete (=convex hull) uncertainty sets
that are widely used in robust combinatorial optimization do warrant research attention,
as they may not produce the best solutions, but are relatively fast to solve. However,
permutohull and ellipsoidal uncertainty tend to produce solutions with a better tradeoff, while being computationally more challenging. The algorithmic research for robust
shortest path problems with such structure should therefore become a future focus.
4 Conclusion
In this paper wo constructed uncertainty sets for the robust shortest path problem using
real-world traffic observations for the City of Chicago. We evaluated the model suitability
of these sets by finding the resulting robust paths, and comparing their performance using
different performance indicators.
Naturally, conclusions can only be drawn within the reach of the available data. In
our setting we considered solutions that are robust with respect to all possible travel
times within a day. A use-case would be that a path needs to be computed for a specific
day, but the precise hour is not known. Using different sets of observations will result
in solutions that are different in another sense, e.g., one could use observations over
different days during the morning rush hours, or observations that span work days and
a weekend. It is possible that these sets will provide different structure.
Finally, we have observed that using ellipsoidal uncertainty sets provides high-quality
solutions with less computational effort than for the permutohull. If one uses only the
diagonal entries of the matrix Σ , then one ignores the data correlation in the network.
For the resulting problem specialized algorithms exist, see, e.g. [Nik09]. In additional
experiments we found that even by using Cplex, computation times were considerably
reduced when only using the diagonal entries of Σ , but the solution quality remained
roughly the same.
References
[ABV09]
Hassene Aissi, Cristina Bazgan, and Daniel Vanderpooten. Min–max and
min–max regret versions of combinatorial optimization problems: A survey.
European journal of operational research, 197(2):427–438, 2009.
[BB09]
Dimitris Bertsimas and David B Brown. Constructing uncertainty sets for
robust linear optimization. Operations research, 57(6):1483–1495, 2009.
13
[BDG+ 16] Hannah Bast, Daniel Delling, Andrew Goldberg, Matthias MüllerHannemann, Thomas Pajor, Peter Sanders, Dorothea Wagner, and Renato F
Werneck. Route planning in transportation networks. In Algorithm Engineering, pages 19–80. Springer, 2016.
[BS03]
Dimitris Bertsimas and Melvyn Sim. Robust discrete optimization and network flows. Mathematical programming, 98(1):49–71, 2003.
[BS04]
Dimitris Bertsimas and Melvyn Sim. The price of robustness. Operations
research, 52(1):35–53, 2004.
[BTN98]
Aharon Ben-Tal and Arkadi Nemirovski. Robust convex optimization. Mathematics of operations research, 23(4):769–805, 1998.
[BTN99]
Aharon Ben-Tal and Arkadi Nemirovski. Robust solutions of uncertain linear
programs. Operations research letters, 25(1):1–13, 1999.
[Büs12]
Christina Büsing. Recoverable robust shortest path problems. Networks,
59(1):181–189, 2012.
[CG15a]
André Chassein and Marc Goerigk. Alternative formulations for the ordered
weighted averaging objective. Information Processing Letters, 115(6):604–
608, 2015.
[CG15b]
André Chassein and Marc Goerigk. A new bound for the midpoint solution
in minmax regret optimization with an application to the robust shortest
path problem. European Journal of Operational Research, 244(3):739–747,
2015.
[CG16a]
André Chassein and Marc Goerigk. A bicriteria approach to robust optimization. Computers & Operations Research, 66:181–189, 2016.
[CG16b]
André Chassein and Marc Goerigk. Performance analysis in robust optimization. In Robustness Analysis in Decision Aiding, Optimization, and
Analytics, pages 145–170. Springer, 2016.
[KZ16]
Adam Kasperski and Pawel Zieliński. Robust discrete optimization under discrete and interval uncertainty: A survey. In Robustness Analysis in Decision
Aiding, Optimization, and Analytics, pages 113–143. Springer, 2016.
[MG04]
Roberto Montemanni and Luca Maria Gambardella. An exact algorithm for
the robust shortest path problem with interval data. Computers & Operations
Research, 31(10):1667–1680, 2004.
[Nik09]
Evdokia Nikolova. High-performance heuristics for optimization in stochastic traffic engineering problems. In International Conference on Large-Scale
Scientific Computing, pages 352–360. Springer, 2009.
14
[YY98]
Gang Yu and Jian Yang. On the robust shortest path problem. Computers
& Operations Research, 25(6):457–468, 1998.
15
| 8cs.DS
|
Vsep-New Heuristic and Exact Algorithms for Computing Graph
Automorphism Group and Graph Isomorphism
Stoicho D. Stoichev
Department of Computer Systems, Technical University of Sofia
email: [email protected]
Abstract. Five new algorithms, named Vsep, are described. Four of them are for
determining the generators, orbits and order of an undirected graph automorphism group.
Vsep-e – exact, Vsep-orb and Vsep-sch – heuristic and Vsep-a automatically selects the
optimal version among Vsep-e, Vsep-orb and Vsep-sch. The fifth algorithm, Vsep-is, is for
finding an isomorphism between two graphs. Vsep-orb firstly finds heuristically the
generators and orbits and then uses the exact one on the orbital partition for determining
the order of the group. Vsep-sch differs from Vsep-orb in using the Schreier-Sims
algorithm for determining the order of the group. A basic tool of these algorithms is the
adjacency refinement procedure that gives finer output partition on a given input partition
of graph vertices. The refinement procedure is a simple iterative algorithm based on the
criterion of relative degree of a vertex toward a basic cell in the partition. A search tree is
used in the algorithms - each node of the tree is a partition. All nonequivalent discreet
partitions derivative of the selected vertices called a bouquet are stored in a coded form in a
hash table in order to reduce the necessary storage – this is a main difference of Vsep-e
with the known graph automorphism group algorithms. A new strategy is used in the exact
algorithm: if during its execution some of the searched or intermediate variables obtain a
wrong value then the algorithm continues from a new start point losing some of the results
determined so far. The new start point is such that the correct results can be obtained. The
proposed algorithms has been tested on the nauy&Traces benchmark graphs and compared
with Traces, and the results show that for some graph families Vsep-e outperforms Traces
and for some of the others Traces outperforms Vsep-e. The heuristic versions of Vsep are
based on determining some number of discreet partitions derivative of each vertex in the
selected cell of the initial partition and comparing them for an automorphism, i.e. their
search trees are reduced. The heuristic algorithms are almost exact and are many times
faster than the exact one. The heuristic algorithms are good choice for the user because of
their smaller running times. Several cell selectors are used in Vsep, some of them are
known and some are new. We also use a chooser of cell selector for choosing the optimal
cell selector for the manipulated graph. The experiments show that the running time of
Vsep algorithms does not depend on the vertex labeling.
Key words: graph, isomorphism, automorphism, group, stabilizer, exact algorithm,
heuristic algorithm, partition, numeration, generators, orbits and order of the graph
automorphism group.
-1-
1. Introduction
We assume some familiarity with the basics in the design and analysis of algorithms[1–
3], combinatorial algorithms[4,5], graph theory and group theory [6–10]. We consider
simple finite undirected graphs (without loops and multiple edges). The graph is denoted
by G(V,E), where : V={1,2,3, . . . ,n} is the set of vertices and E - the set of edges (v,
w), v, w V. The number |E| of the edges of G we denote by k, k=O(n2). Our algorithms
are applicable to any undirected graph including disconnected graphs but for them there is
more efficient algorithm we do not describe here. The set of all vertices adjacent to a
vertex x is denoted by Adj(x). We use a static list representation of a graph (by two static
arrays) because of its least required storage and fastest operation of finding all adjacent
vertices of a given vertex compared with the adjacency matrix and the dynamic list
representation.
An isomorphism [4–12] between two graphs G1(V1, E1) and G2(V2,E2) is called one-to-one
correspondence (mapping) yi = f(xi) between the vertices of the graphs (xiV1, yiV2,
i=1,2,...,n, n=|V1|=|V2|) such that two adjacent (nonadjacent) vertices from one of the
graphs correspond to two adjacent (nonadjacent) vertices from the another graph, i. e.
every edge (non-edge) (p, q ) from graph G1 corresponds to an edge (non-edge) (f(p),f(q))
from graph G2 and vice versa. So, the isomorphism preserves the adjacency relation of
vertices - this kind of bijection is commonly called "edge-preserving bijection".
Two graphs G1 and G2 are called isomorphic (G1G2) if there is at least one isomorphism
between them. Otherwise they are non-isomorphic, G1!G2.
An automorphism [4–11,13] of a graph is an isomorphism of the graph onto itself. Or, an
automorphism h of graph G(V,E) is called one-to-one correspondence yi=h(xi) between
the vertices of the graph (xi,yiV, i=1,2,...,n) that preserves the adjacency of the vertices, i.
e. there is unique corresponding edge (non-edge) (h(p),h(q))E to each edge (non-edge)
(p,q)E. A fixed point x of an automorphism h is called a vertex x for which x = h(x).
Trivial automorphism is an automorphism h0 if each its vertex is a fixed point, xi=h0(xi),
i=1,2,...,n and a non–trivial automorphism is an automorphism for which there is at least
one pair of vertices x,y such that y=h(x) x.
Two vertices xi and yi in the a graph G(V,E) are called similar (or symmetric) [6,7], xiyi,
when they are corresponding, yi=h(xi), in some automorphism h. Otherwise, if xi and yi are
not similar we use the notation xi! yi. An automorphism h can be presented by two
sequences Па and Пb of graph vertex labels:
a1, a2,..., ai,... an=Па
b1, b2,..., bi,... bn=Пb, where the corresponding vertices are ai and bi=h(ai),
i=1,2,...,n.
The automorphism h may be shortly written in the form Пb=h(Па). The automorphism can
be presented by n! pairs of rows - each row being derived from the other by transposing the
positions of the pairs of corresponding vertices. The corresponding pairs of vertices can be
set on to any place of the rows, but it is possible the place to depend on the sorting criterion
which does not depend on the vertex labeling. Each automorphism can be written uniquely
only with the permutation Пb if we assume that Пa = 1, 2, 3, … n. Even a simpler notation
-2-
called cycle notation [10] is often used. In a cycle (x1,x2, . . . ,xi, xi+1, . . . ,xp) xi maps to xi+1,
1 i p-1 and xp maps to x1. For example,
1 2 3 4 5 6 7 8
h=(1,8)(2,6,3,7)(4,5) =
8 6 7 5 4 3 2 1
Any automorphism can be written as the product of disjoint cycles and the product is
unique up to the order of the cycles [10]. The cycles of length 1 are omitted.
The operation function composition (or superposition) [4–11,13] of two automorphisms
......YX..... and ......YZ ..... is a consecutive execution of and , i. e. . ......ZX...... . This
operation is usually called multiplication and is denoted by juxtaposition γ=..
The set of all automorphisms of a given graph G(V,E) form a graph automorphism group
(under the operation function composition of automorphisms) denoted Aut(G), shortly
A(G) or A ) [4–11,13]. The trivial automorphism is the identity of the group – we denote it
by I. The number of the automorphisms in A(G), |Aut(G)|, is called an order of the
automorphism group. An order of an automorphism is the order of the cyclic group
generated by this automorphism. If the automorphism is written in cycle form then its order
is the least common multiple of the cycle lengths[10].
The subset gen(А)А, denoted gen(А) or <g1,g2,...,gd>, is called a generating set
[4,5,9,10,12] of the automorphism group if every automorphism hA can be written as a
finite product of elements gi , gi1 gen( A) . The generating sets are not unique.
The subset of the graph vertices y similar to vertex x in any automorphism hAut(G),
y=h(x), is called an orbit of x, denoted Orb(x,A): Orb(x,A)={h(x)|hAut(G)} and Orb(x,A)
is read as ‘orbit of a vertex x under a group A’. By Orb(A) we denote the set of all orbits of
the group A.
A graph with only one orbit of Aut(G) is called transitive, and a graph is called rigid or
asymmetric if each of its vertices is in a separate orbit, i.e. |Aut(G)|=1. A stabilizer
А(х1,х2,...,хi) or point-wise stabilizer [5,10,13] is the subgroup of the automorphism group
A that contains only the automorphisms with fixed points х1,х2,...,хi.
There is a theorem called orbit-stabilizer theorem[5–8] for computing the order |A| of the
automorphism group of graph G(V,E). We denote it O-S theorem. The theorem claims that
|A|=|A(х1)|*|Orb(х1,A)|=|A(х1)|*d, where Orb(х1,A)={ х1, х2,..., хd} is the orbit of vertex x1
under the group A, and A(x1) is a stabilizer of a representative x1 of this orbit. In other
words, the order of the graph automorphism group is equal to the product of the order of a
stabilizer of one of its vertices and the length of the orbit of this vertex. Vsep algorithms
and other known graph isomorphism and automorphism algorithms use the orbit-stabilizer
theorem.
The stabilizer Ai= А(х1,...,хi) is called an ancestor stabilizer to the stabilizer Aj= А(х1,...,хj)
if j>i and Aj is called a successor (or descendant) stabilizer to Ai. Obviously, the orbits of
the successor stabilizer are subsets of the orbits of the ancestor stabilizer and the order of
the successor stabilizer is a divisor of the order of the ancestor stabilizer.
Given a set SV a set-wise stabilizer [10,12], denoted Aut(G,S), is a subgroup of Aut(G)
defined by:Aut(G,S)={Aut(G) |x,(x)S}.
The graph isomorphism (GI) problem consists in deciding whether two given graphs are
isomorphic, i.e., whether there is an isomorphism between them. The graph isomorphism
-3-
problem belongs to the class of NP problems [4–9,11,12,14], and has been conjectured
intractable, although probably not NP-complete.
It is neither known that this problem has polynomial time complexity nor to belong to the
class of NP–complete problems. Its unknown complexity status is a theoretical and
practical challenge. Two problems are called polynomial-time equivalent if there is a
polynomial time algorithm that transforms one of the problems to another [4,7,9,10]. The
problems [3,9,10] that are polynomial-time equivalent to graph isomorphism are called
graph isomorphism complete. The following problems are graph isomorphism complete
[15,16]: given two graphs G1 and G2: existence of isomorphism of the graphs; determine
the isomorphism of the graphs if it exists; determine the numbers of the isomorphisms
from G1 to G2; given a graph G: determine the generating set gen(Aut(G)), determine the
orbits Orb(Aut(G)), determine the order |Orb(Aut(G))|. The problems of determining the
orbits and order of the graph automorphism group A are also called automorphism
partitioning and automorphisms counting, respectively. The generators, orbits and order of
a graph G automorphism group A we denote by GOO(A) or GOO(Aut(G)) and by GO(A) the generators and orbits of the group. A class of graphs is called graph isomorphism
complete if the recognition of isomorphism for graphs from this class is a graph
isomorphism complete problem. The following classes are graph isomorphism complete
[16]: connected graphs, graphs of diameter 2 and radius 1, directed acyclic graphs, regular
graphs, bipartite graphs without non-trivial strongly regular subgraphs, bipartite Eulerian
graphs, bipartite regular graphs, line graphs, chordal graphs, regular self-complementary
graphs, etc. However, there are special cases of the graph isomorphism problem with
polynomial-time complexity: planar graphs (linear time), trees have a particularly simple
algorithm, graphs of bounded degree, interval graphs, permutation graphs and convex
graphs. At present it is not known a polynomial time algorithm for solving the graph
isomorphism complete problems in the worst-case [14,17] – all known algorithms have
exponential or moderately exponential worst-case time complexity. Graph isomorphism
problems are of great practical and theoretical interest [3–5,12,14,17]. Recently, Laszlo
Babai has claimed that the Graph Isomorphism problem can be solved in quasipolynomial
time [18]. There are several practical algorithms (their names are shown below in
parentheses ) for graph isomorphism and graph automorphism group, due to Brendan
McKay (nauty), Adolfo Piperno (Traces), William Kocay (Groups&Graphs), Schmidt and
Druffel, Jeffrey Ullman; L.P. Cordella, P. Foggia C. Sansone and M.Vento (VF2), Tommi
Junttila and Petteri Kaski (bliss), Hadi Katebi, Karem A. Sakallah, and Igor L. Markov
(saucy), Jose Luis Lopez-Presa, Luis Nunez Chiroque, and Antonio Fernandez Anta
(conauto), G. Tener and N. Deo (nishe), Nechepurenko [4], etc. There is a comparison in
[13,19] on running times of the following tools: nauty, Taces, bliss, conauto and saucy.
Their running time on random graphs is quite well but a major problem of these algorithms
is their exponential time performance in the worst- case. There are two main
generalizations of the graph isomorphism: subgraph problem (given two graphs determine
if one of them is a subgraph to another) and largest common subgraph problem (given two
graphs determine the common subgraph to both that has the maximum number of vertices
or edges). These two problems are NP-complete and have many applications and the
algorithms for them use the graph isomorphism and automorphism algorithms as basic
-4-
tools. Another important problem related to graph isomorphism is [5,12]: Compute a
complete invariant (certificate, signature) f for G, i.e. for all graphs G and H, G H
f(G) = f(H) (graph certificate identifies a graph uniquely up to isomorphism).
The graph isomorphism related problems (graph isomorphism itself, GOO(Aut(G))),
subgraph isomorphism, largest common subgraph, graph certificate and canonization of a
graph) arise in such fields as mathematics, chemistry, information retrieval, linguistics,
logistics, switching theory, bioinformatics, and network theory [4,20].
Our goal is to develop exact and heuristic algorithms for determining GOO(Aut(G)), i.e. to
solve the three problems by one algorithm with time complexity as lower as possible. In
addition, our requirements to the heuristic algorithms are to give results equal to the results
of the exact algorithm with the probability close to 1. There are a few heuristic algorithms
for the graph isomorphism problem [21–24] – there is no access to their program codes.
We propose three new heuristic algorithms (Vsep-orb, Vsep-hway, Vsep-sch) for
GOO(Aut(G)) with much lower polynomial time complexity than the exact one. The
experiments show that they are many times faster than the exact algorithm - even for
difficult graphs with large sizes they give correct results. One of the first step in the exact
algorithm is a call to a heuristic procedure for determining a representative of one of the
smallest orbits of Aut(G) as a starting selected vertex – this way we speed up the exact
algorithm and reduce the required storage.
2. Partitions and a refinement procedure
An ordered partition (or simply partition) П or П(G)
П = C1C2…Ci…Cp = C1C2...Ci...Cp
of the vertices of graph G(V, E) [5,12,15] is a sequence of disjoint non-empty subsets of V
whose union is V. П = | x1,1, x1,2, …, x1,k1 | … | xi,1, xi,2, …, xi,ki | … | xp,1, xp,2, …, xp,kp| is a
detailed presentation of П, where Ci=| xi,1, xi,2, …, xi,ki |, xi,j ϵ V, i=.1,…,p, j=1,…,ki. The
subsets C1, C2,… , Cp, are called cells (classes, blocks).
We denote the number of the cells in a partition П by |П|. Two cells are called adjacent if
there is at least one edge between their vertices, i.e. cells Ci,Cj П are adjacent if there is at
least one edge (x,y), x Ci, y Cj. A cell with cardinality one is called trivial (or
singleton). The vertex of such a cell is said to be fixed by П or it is called a fixed point of
П. A partition, of which each cell is trivial is called discrete or numeration having in
mind that in fact it is a permutation that can be viewed as a graph vertices renumbering –
vertex i corresponds to vertex x that is on the position i in the partition. By NC(x, П) we
denote the index of the cell C of П that contains vertex x, x C. The position (index) of a
vertex x in the partition (or in the cell) we denote by pos(x). The relative degree ρ(x,Ci) of
a vertex хСj toward a cell Сi is equal to the number of vertices of cell Сi adjacent to
vertex x. We denote by ν(x, П) a cell-degree vector defined as ν(x, П)=( ρ(x,Ci), i=1,…,p),
p=|П| – it is a vector whose components are the relative degrees of x to each cell in П. We
call a partition П stable (or equitable) if the cell-degree vectors ν(x, П)= ν(y, П) for each
two vertices x, y Ci, where Ci is any cell in П. We say that the partition П2 is finer than
П1, written П2 ≤ П1, if for every cell Ci П2 there exists a cell Cj П1 such that Сi Cj. In
order to get a finer partition П2=D1D2...Dq when given П1=С1С2...Ср , a refinement
-5-
procedure (RP, П2= RP(П1)) is used, that assign to each vertex х V a sorting criterion
according by which the vertices of each class СiП1 are sorted out in increasing or
decreasing order of their criterions [4,5,12]. Often a sorting criterion is the relative degree
ρ(x,W) of any vertex х П toward some cell WΠ.. The refinement procedure that uses
this criterion is called adjacency refinement procedure. We use only this version of the
procedure. Example of another criterion for sorting is the number of the subgraphs of a
given type (for example a triangle) that contain vertex x.
Two partitions Π1 and Π2 of the vertices of graph G(V,E) are called compatible [40] if: (1)
|Π1| = |Π2 |= m; (2) if Π1 = W1W2…Wm and Π2 = U1U2…Um, then for all i ϵ [1:m], |Wi| =
|Ui|; (3) for all x,y ϵ V, NC(x,Π1) = NC(y,Π2) implies ν(x,Π1) = ν(y,Π2). Similar
definition is valid if Π1 and Π2 are for different graphs.
Two partitions П1=С1С2...Ср and П2=D1D2...Dq are called equivalent if there is an
automorphism αAut(G) such that NC(x,П1) = NC(y,П2) for each pair of vertices x, y=
α(x). In other words, the similar vertices are in cells with the same label of П1 and П2).
Obviously for the equivalent partitions П1 and П2 we have: p = q and |Ci| = |Di|, i[1:p].
We denote the equivalent partitions by П2 = α(П1) or П2 П1. Evidently the equivalence
relation is transitive. Two discrete partitions П1 = a1a2...ak...an, П2 = b1b2...bk...bn of graph
G(V,E) vertices are called equivalent if they form an automorphism h Aut(G), bk = h(ak),
k = 1,2, ... ,n. The testing if two discrete partitions П1 and П2 form an automorphism is a
basic operation in our algorithms.
Given sequence of equivalent discrete partitions П1, ... , Пm, we store one of them, for
example П1 and the orbits derived by the automorphisms Пi = i(П1), i=2, ... ,m.
A partition-wise stabilizer A(G, П) is defined by: A(G, П) = {Aut(G) | x,(x) Ci,
where Ci is any cell of П} [12]. This is a subgroup of Aut(G) such that each automorphism
αAut(G) belongs to Aut(G, П) if to any vertex x of any cell of П corresponds a vertex
y=α(x) from the same cell. The orbits of A(G, П) are subsets of the cells of П as we’ll see
below. If П= Пu then A(G, Пu) = Aut(G).Algorithms for determining the graph
automorphisms and isomorphisms use very often the refinement procedure - each cell of its
output partition contains at least one orbit of a graph automorphism group or its stabilizer.
It is still not known a refinement algorithm that gives output partition each cell of which
coincides with an orbit (orbit partition or automorphism partition) [20] on unit input
partition.
One of the most efficient RP is the RP with a base cell. The RP with a base cell sorts
(counting sort [1,2]) the vertices of any cell Cj according to their relative degree rdg(x,i)
toward a selected base cell Ci. The sorting continues until it reaches a partition П2 in which
there is no cell that can be divided into subcells toward any base cell – such partition is
called stable (equitable) [5,12] as we noted above: it holds the property П1=RP(П1).
Vertices in every cell of the stable partition have the same sorting criterion - in our case,
the same relative degree toward each cell. The base cell refinement algorithm has time
complexity О(к.log n), where k and n are respectively the number of edges and the number
of vertices of the graph [25,26]. The base cell RP refinement algorithm (Figure. 1) uses
the counting sort that does not use a comparison operation – it sorts integers (the relative
degree of the vertices) within the range 0 to some integer. In our algorithms for
-6-
GOO(Aut(G)) we use only the adjacency refinement procedure with a base cell (Figure. 1)
[25] and a W queue that contains the labels of all not selected cells as base cell. It differs
from the known refinement algorithms [5,12,20] in the way the base cell is selected. After
the current base cell ends sorting the adjacent cells as a new base cell is selected the first
new smallest subcell and if there is no such subcell the label of the new base cell is taken
from a W queue - the first cell label in W queue. There is a property that speeds up RP
procedure: the label of the new largest subcell of C cell (adjacent to B cell) is not included
in W queue if C cell label is not in W. There is a version of RP refinement algorithm with
base cell that always takes the new base cell from the queue.
The individualization-refinement operation (denoted IR), used in the known GA
algorithms, has two steps: individualization and refinement. Given an equitable partition π
and a vertex x at the individualization step a new partition π1 is obtained: the cell C(x) of π
with index i is divided into 2 subcells: {x}-the first subcell, with index i and the second
subcell {C(x)\{x}} with index i+1, other cells of π are not changed. At the refinement step
the partition π1 is refined with the refinement procedure RP obtaining a new equitable
partition π2=RP(π1) finer than π1. Given a partition π and a vertex x we denote by
IR(π,x,brcl) the resulting partition π2 from the application of IR operation on π and x, where
brcl is the number of cells of π2.
Input: graph G(V,E); П1 is the input partition on the graph vertices, W is a queue of
some cells of П1,BRCL-the number of cells of П1
Output: a better stable partition П2 ( П2 П1), BRCL-the number of cells of П2
S1: П2:= П1;
S2: Base cell В:= first cell in W; Delete the label B from W;
S3: repeat{each the loop execution is performed for different base cell B}
S4: Each cell СП2 adjacent to B is divided into subcells according to its relative
vertex degrees toward В;
S5: if there are new subcells from S4 then В:= the label of the subcell with
minimum length. If there are more cells with minimum length then the one with the
smallest label is chosen. Put into W the labels of the new subcells in П2, excluding
the label of one of the largest subcell if it is not in W; Delete the label B from W;
S6: if there are no new subcells from S4 then В:= first cell label in W;
S7: until there are no new subcells from S4 and W becomes empty
Figure. 1. RP refinement procedure with base cell B
Cells in our algorithms are not consecutively labeled by 1, 2, … . The label NC(Cj,П) of a
i 1
cell Cj in the partition П is determined by NC (Ci , П ) | C j ) 1 , i. e.
j 1
the label of the cell Ci (respectively of each of its vertices ) is the first vertex index in the
cell, or it is greater by 1 than the number of the vertices in all cells preceding Ci in the
partition. NC(х,П) denotes the label of the cell that contains vertex x. This way of labeling
is time saving because changing the labels of a given cell does not cause change of the
labels of other cells.
-7-
Two cells of a partition are called non-trivially joined (have non-trivial join, non-uniformly
joined) if the number of edges between them is greater than 0 and less than maximum
possible.
A channel of a cell C (new notion), denoted as Ch(C)), is the number of the edges
adjacent to any vertex in C cell. A channel of two cells C1,C2, denoted as Ch(C1,C2)), is the
number of all edges (x,y) between any vertex xC1 and any vertexC2. Example: Let
П=|1,8|4,5|2,3,6,7|=C1 C2 C3 be a partition on the vertices of the graph on Fig. 2. The
channels of the cells are: Ch(C1,C1)=0, Ch(C1,C3)=4, Ch(C1)=4, Ch(C3,C3)=2,
Ch(C3,C2)=4, Ch(C3)=10, Ch(C2,C2)=1, ChC(C2)=5. Only the edges of cells with nontrivial join to C are included in Ch(C). For example, the cells D5 and D6 of the partition
П=|1|8|5|4|6, 7|2, 3|=D1 D2 D3 D4 D5 D6 of the vertices of the above graph have channels
Ch(D5)=1, Ch(D6)=1 since the edges incident to vertices from D5 and D6 and connected to
vertices from the trivial cells D1 D2 D3 D4 are not included – only the edge (6,7) is included
in Ch(D5) and (2,3) in Ch(D6).
A channel vector of a cell C П, ChVect(C), is a vector VC whose component VC(i) is
equal to Ch(Ci,C) channel where Ci is a cell П with non-trivial join to C. A channel
graph of a partition П, ChG (П), is an weighted graph with loops: each vertex of ChG
corresponds uniquely to a cell of Π and its weight is the channel of the cell; each edge of
ChG corresponds to a channel of the corresponding cells of Π and the weight of this edge
is equal to the weight of this channel. Similar notion is a quotient graph in [27]. A selected
non-trivial cell, SC(П), of a partition П is the cell Cj, | Cj|>1, that is selected by a defined
criterion (often it is called target cell, for example in [12]). The procedure that finds the
target cell is called cell selector.
Fig. 2. GIRA graph
The selection of the target cell has strong influence on the search tree size (see the next
section) and consequently, on the running time of the GA algorithms.
In Vsep algorithms, at given an equitable partition π we use the following 5 cell selectors
(their numbering is denoted by izb):
izb=1, MXNACS1: (a) it selects the nontrivial cell C of π with maximal number of
adjacent non-trivially joined cells to C, and (b) if there are several such cells it selects
the first one of the smallest size;
izb=2, MXNACS2: it differs from (1) only in (b) selecting the first cell with largest size.
izb=3, MXVECTCHVAL: it selects the first nontrivial cell C of π with the largest
vector of relative degrees of C to other cells – the vector is considered as a number with
leftist least significant digit.
-8-
izb=4, MXPRCHVAL: it selects the first nontrivial cell C of π with largest channel.
izb=5, ICLMXBRCL: it selects the first non-trivial cell C with maximal number of cells
of the derived partition π(x), x ϵ π, obtained after IR.
To select the optimal cell selector we use a cell selector chooser (CSLCh) – it chooses the
cell selector by three criteria: a) the smallest maximal level LMAX of the first path in the
search tree., b) the smallest product of the sizes of all target cells of the first path in the
search tree and c) the smallest product of the numbers of the cells of all partitions on the
first path in the search tree.
The partitions in the developed algorithms can be:
1. Unit partition (denoted by Пu=V=|1, . . ., n|) - all its vertices are in one cell С1 and the cell
number is NC(i)=1, i=1, . . .,n.
2. Еquitable partition - it is obtained as a result of the execution of the refinement procedure
(RP) on given input partition (see the definition given above);
3. Transformed partition ПТ is obtained from a stable partition П by separating from SC(П)
given vertex x in a singleton cell. The difference between ПТ and П is that SC(П) with cell
label i, is divided into two cells in ПТ: cell С'={x} with label i containing vertex x and cell
С''= SC(П)\{x} with label i+1 containing the other vertices of SC(П). The transformation
operation is denoted by TR: ПТ=TR(x, П) and is called an individualization in many
references.
The following theorem holds for the RP.
Theorem 1
Given a graph G(V,E), two different equivalent partitions П1 and П2 on V and a non-trivial
automorphism fAut(G). Then, NC(x,Па) = NC(y,Пb) for each pair x, y=f(x), where
Пa=RP(П1), Пb=RP(П2).
In other words, RP gives output partitions where the similar vertices remain in cells with
the same label.
This theorem is equivalent to the Theorem 7.1 in [5] – for equivalent partitions under an
isomorphism of two graphs and the statement that the RP is invariant under an
automorphism [28]. This is the reason for not giving here the proof of the theorem.
Corollary 1 of Theorem 1 Let П be a partition on V of a graph G(V,E) with an
automorphism fAut(G) and let for each pair of similar vertices x, y=f(x) the property
NC(x,П) = NC(y,П) holds. Then, NC(x,Пс) = NC(y,Пс), where Пс=RP(П).
Corollary 1 means that the similar vertices in any cell of a given input partition remain
in one cell of the resulting partition of RP.
Proof Let П=П1=П2 in theorem 1. Then, we obtain Пс=Па=Пb=RP(П) and
NC(x,Пс)=NC(y,Пс) is obtained by replacing Пс in NC(x,Па)=NC(y,Пb).
Corollary 2 of Theorem 1 The refinement procedure RP is invariant under the vertex
orbits of a subgroup HА=Aut(G(V,E)): if Orb(x,H)C(x,П) for each vertex хV of a
given partition П, then Orb(x,H)C(x,Пc) for a partition Пc=RP(П). (In other words, if
-9-
each orbit of a given automorphism subgroup is a subset of a cell of the input partition П ,
then the same is true for the output partition Пc=RP(П). Obviously, this corollary holds for
any stabilizer А(х1,х2,...,хi) of the automorphism group.)
Proof The condition Orb(x,H)C(x,П) means that NC(x,П)=NC(y,П), y=f(x) for each f
Н, i. e. the condition of Corollary 1 of Theorem 1 holds for each fН and from it follows
NC(x,Пс)=NC(y,Пс), y=f(x) for each fН, i.e. Orb(x,H)C(x,Пс).
Corollary 3 of Theorem 1 Orb(x,A)C(x,П1) holds for each vertex xV of the initial
partition П1=RP(Пu) of the graph G(V,E) where Пu is the unit partition.
Proof Let H=A and П=Пu for which obviously Orb(x,A)C(x, Пu). Then,
Orb(x,A)C(x,П1) follows from Corollary 2 where П1=Пс=RP(Пu).
3. Auxiliary algorithms
3.1.
A1 algorithm for determining one discrete partition
The output of A1 algorithm (Figure 3) is a series of better partitions the last of which is
discrete on given input partition П. Using A1 we introduce new notions and prove a
property used for speeding up our algorithms.
The following basic operations are performed in the A1 algorithm: {П1=RP(П), L=1};
{ determine SC(ПL) by some cell selector; select any vertex ХL in the SC(ПL); L:=L+1;
ПL:=IR(ПL-1, XL-1)}; L=2, . . . , LK, where ПLK is discrete and LK<n.
Given any input partition П the initial partition П1=RP(П) is determined first. In most cases
we’ll consider that the starting input partition П is the unit partition, П= Пu.
A selection level is called the successive number L of the refinement procedure RP calls for
obtaining the partition ПL in A1 algorithm. ПL=П(х1,х2, ... , хL-1) denotes the partition that
is obtained by making L-1 selections х1,х2, ... , хL-1 starting with the partition П1. The
selected non-singleton cell in partition ПL is denoted by SC(ПL) or SC(х1,х2, ... , хL-1). Five
operations (rows 4, 5) are executed in the loop C: determining SC(ПL), a selection of
vertex xL in SC(ПL), a move to next level (a forward move, L+1), obtaining the
transformed partition ПLT:=TR(xL-1,ПL-1), ПL:=RP(ПLT) (refinement with RP) – the last two
operations are IR . Obviously, ПL is a stable partition. The determination of SC(ПL) is
made by some cell selector. This execution continues until a discrete partition ПLK is
obtained – its level of selection LK is called terminal or final. The process of obtaining the
sequence П1, П2, . . . , ПLK is called series of forward moves (SFM).
The selection of a vertex also is made by some criterion but here we consider (for
definiteness) the selection of the first vertex in SC( ПL).
The five instructions (lines 4, 5) we denote by CSVSIR (cell selection, vertex selection,
individualization and refinement) – cell selection in ПL-1, vertex Xl-1 selection in SC(ПL-1),
L=L+1, individualization of XL-1 (ПL:=IR(ПL-1,XL-1)) and refinement of ПLT.
Each partition ПL, L=1, 2, ... , LK is called a partition-child of ПL-1 (partition-parent) and
a successor of each of the partitions ПL-1, ПL-2, ... , П1, П0 (partitions-ancestors of ПL).
- 10 -
The selected cells SC (L), L=1, 2, . . ., L-1 for a given partition ПL are called supporting SC
for ПL, and the selected vertices х1,х2, ... , хL-1 of ПL are called base points or a base of ПL
(or supporting points).
A search tree (ST) is an oriented tree whose root represents the initial partition П1=RP(П).
Each node of the search tree corresponds to an equitable partition ПL. Each arc coming out
of a node ПL corresponds to a selected vertex from the SC(ПL). Thus, the partition ПL+1 of a
particular node q on a level L+1 could be obtained from the partition ПL (node p on a level
L) by selecting a vertex х SC (ПL), x being depicted on the edge p-q in ST. Let r =
|SC(ПL)|, then each node ПL=П(х1,х2, ... ,хL-1) with SC (ПL) = {a1a2 ... ar} has r nodeschildren of level L+1:
П(х1,х2, ... ,хL-1,a1), П(х1,х2, ... ,хL-1,a2), ... , П(х1,х2, ... ,хL-1,ar). Thus, each vertex in each
SC(ПL), for each level L is selected in ST. Only in П1 one vertex is selected – the first
vertex in SC(П1).
The just described search tree ST may be called a full search tree in contrast to the reduced
Input: a graph G(V,E), a partition П of V
Output: Series of better partitions П1<П2<…< ПL< ..< ПLK(discrete)
1: L:=1; ПL:=П; ПL:=RP(ПL); Output(ПL);
2: if ПL is discrete then message; return end if
3: repeat {loop C}
4: determine SC(ПL) by some cell selector; select a vertex ХL in the
SC(ПL)
5: L:=L+1; ПL:=IR(ПL-1, XL-1,brcl); Output(ПL);
6: until brcl=n
Figure 3. A1 algorithm
search tree RST, in which some vertices from SC(ПL) are not selected according to a
certain criterion. In the reduced search tree, for each orbit of а stabilizer AL-1=А(х1,х2, ...
,хL-1) only one representative of the orbit(AL-1)SC(ПL) is selected – this is the first
orbits vertex met during the traversal of the SC(ПL). This way of a selection of vertices
is used in Vsep algorithms. It is based on the theorems proved below.
The reduced search tree used in Vsep algorithms is not explicitly presented in the
algorithm – only the partitions of the path from the root to the current ПL are stored, i.e.,
the sequence П1, . . . , ПL.
3.2.
A2 algorithm for generating the full search tree
A2 algorithm (Figure 4) generates all partitions of the search tree on a given input
partition П of graph G(V,E) and the selection of a vertex Х1 SC(L) of the initial
partition П1 = RP(П). The leaf nodes of ST are discrete partitions. A new partition ПL (a
new node of ST) is obtained after each execution of the instruction 10. In A2, each of
the vertices in every SC(ПL) is successively selected and after that the algorithm
continues with a forward move until a discrete partition is obtained. There are two loops
- 11 -
in A2: the loop C1 (lines 4-15) performs a forward move and the loop C2 (lines 9-13)
performs the backward move. In the forward move, as in A1 algorithm, we begin from
the partition ПL (on a level L) and selected vertex xSC(ПL), and we obtain
successively the partitions ПL+1, ПL+2, ... , ПLK (discrete). The backward move (L:=L-1)
is made when all vertices in the SC(L) are selected. We introduce a new notion – a level
of the last backward move, denoted by LP and meaning the last value assigned to L by
the instruction 10 (L:=L-1), i. e. LP is the level to which the last backward move is
made. We call LP a working level and SC(ПLP) - a working cell since the main
operations of the algorithms are with the vertices in SC(ПLP). Level LP is a significant
variable for Vsep-e algorithm. A sequence of forward moves (SFM) starts from the
level LP. A2 algorithm (as well as Vsep-e algorithm) generates the search tree in
preorder [1,2]– first visiting the root (a vertex SC(ПL) then visiting the subtree of
each vertex SC(ПL+1) from left to right.
Input: a partition П of vertices of G(V,E) graph
Output: all partitions –nodes of the full search tree
1: L:=1;ПL= П; ПL:=RP(П); Output ПL;
2: if ПL is discrete then message ‘ the graph is rigid’; return end if
3: determine SC(П1) by the cell selector; select a vertex X1 SC(ПL);
4: do { C1: loop of the forward move}
5: L:=L+1; ПL:=IR(ПL-1, XL-1,brcl); Output ПL;
6: if ПL is not discrete then
7:
determine SC(ПL) by the cell selector; Х L:=first vertex in the SC(ПL)
8: else
9: repeat{ C2: loop of the backward move}
10:
L:=L-1;
11:
if L=1 then return end if;
12:
Х L:=first unselected vertex in SC(ПL);
13: until Х L 0;
14: end if
15: end do
Figure 4. A2 algorithm
A bouquet В(хL) or В(х1,х2, ... , хL) of a selected vertex хLSC(ПL) is called the set of
all mutually non-equivalent discrete partitions derived from the partition ПL and the
selection of хL, i.e. obtained from П(х1,х2, ... , хL)=ПL+1. Only the first numeration
obtained by Vsep-e algorithm is stored from the set of equivalent discrete partitions and
is used as a representative in В(хL). The bouquet В(х1,х2, ... , хL) consists of the
bouquets of the orbits
representatives of vertices in the SC(ПL+1), i. e. it consists of the bouquets В(х1,х2, ... ,
хL, a1L 1 ), В(х1,х2, ... , хL, a L21 ), . . . , В(х1,х2, ... , хL, a Lp1 ), where a1L 1 , a L21 , . . . , a Lp1
are orbits representatives of the stabilizer А(х1, х2, ... , хL) in SC(L+1). Given the
- 12 -
selected vertices х1, х2, . . . , хL-1, хL the following relation holds for the bouquets: В(х1)
В(х2) . . . В(хL-1) В(хL), i.e., the bouquet of a selected vertex хL is included
in the bouquets of all preceding selected vertices.
Theorem 2 Given a graph G(V,E), a partition П on V, an automorphism fAut(G, П),
NC(x,П) = NC(y,П) for each pair x, y=f(x) (a property- equal cell labels for each pair
of similar vertices), two partitions PL=П(а1, а2, . . . , аL-1 ), QL= П(b1, b2, . . . , bL-1 )
obtained by 2 series of executions of CSVSIR on the input partition П and
NC(ai)=NC(bi) for bi = f(ai), i=1,2, . . . , L-1 (equal cell labels of corresponding base
vertices). Then, NC(x, PL) = NC(y, QL) for each pair x, y=f(x), x PL, y QL (the
property holds for each pair of similar vertices).
Proof (by the mathematical induction): (1). The base case: By assumption the property
NC(x, П) = NC(y, П) holds for the initial partition П. In particular, NC(a1, П)=NC(b1,
П).
(2). The inductive step: Let the property holds for i-1: NC(x, Pi-1) = NC(y, Qi-1) for
each pair x, y=f(x). For PiТ and QiТ the property NC(x, PiТ) =NC(y, QiТ) holds since by
assumption NC(ai-1)=NC(bi-1) for ai-1,bi-1=f(ai-1) is true. For all other vertices of the cells
С(ai-1) and С(bi-1) the cell labels are equal to NC(ai-1)+1 – this does not violate the
property. For the vertices of other cells of Pi-1, Qi-1 there is no change of cell labels in
PiТ,QiТi. Hence, according to Corollary 1 of Theorem 1 the property NC(x, Пi) =NC(y,
Pi) also holds for Пi=RP(ПiТ), Рi=RP(РiТ).
Since both the base and the inductive step have been proved, it has now been proved
that NC(x, ПL) =NC(y, PL) holds for all vertices x and y=f(x), x PL, y QL.
Corollary of Theorem 2 Given a partition ПL=П(х1,х2, ... ,хL-1). Then, Orb(x, AL-1)C(x,
ПL) for each vertex хV of the graph G(V,E), L=1, . . . , LK. (In other words, the orbits
of AL-1 are subsets of the cells of ПL).
Proof Let bi=ai, i=1,2, . . . , L-1 in Theorem 2. Thus, these L-1vertices become fixed
points of automorphism f and PL=QL. From NC(x, ПL) = NC(y, ПL) for each pair x, y =
f(x) and for each fАL-1=А(а1, а2, . . . , аL-1) follows Orb(x,AL-1)C(x,ПL).
Applying the corollary for L=1 we have Orb(x, A0)= Orb(x, A(П)) C(x, П1). Applying
it to L=LK we obtain Orb(x,ALK-1)C(x,ПLK) and |Orb(x,ALK-1)||C(x,ПLK)|=1 since
ПLK is discrete, i.e., |Orb(x, ALK-1)|=1 and ALK-1={I}. Thus, for the series of partitions
П1, . . . , ПLK only the stabilizer ALK-1 is known – the stabilizer of the first selected
vertex XLK-1 in SC(ПLK-1) .
Theorem 3 Given a graph G(V,E), n=|V|, an automorphism fAut(G) with fixed points
а1, а2, ... , аj-1 and a partition Пj=П(а1, а2, ... , аj-1), obtained after (j-1) successive
applications of CSVSIR operation on а1, а2, . . . , аj-1 with starting input partition П and
vertices
p, q SC (Пj), q=f(p).
Then, each numeration
- 13 -
П′=П(а1, а2, ..., аj-1, p, аj+1, аj+2, ..., аLК-1)=|x1| x2| . . . | xi| . . . | xn|,
derivative of П(а1, а2, ... , аj-1,p) has a corresponding numeration
П″=П(а1, а2, ..., аj-1, q, bj+1, bj+2, ..., bLК-1)= |y1| y2| . . . | yi| . . . | yn|,
derived from П(а1, а2, ... , аj-1,q), such that yi=f(xi), i=1, 2, ... , n and
NC(xi,П′)=NC(yi, П″).
(Note: There is an equivalent theorem of B. McKay - theorem 2.15 in [12]).
.
Proof Let we consider two executions (labeled I and II) of the operation CSVSIR on
starting input partition П of graph G(V, E). The first j-1 selections а1, а2, ... , аj-1 are
equal for both executions and the resulting partitions are equal to Пj. Under the
conditions of the theorem, there are vertices p, q=f(p) in the cell SC (Пj).
Let the jth selection be p in the first execution, and the jth selection be q in the second
execution, i.e. the obtained partitions are: for execution I – ПI=П(а1, а2, ..., аj-1, p), for
execution II – ПII=П( а1, а2, ..., аj-1, q), for which the conditions of Theorem 2 hold:
аi=f(аi) for i = 1, 2, ..., j-1 and q=f(p). Therefore, according to Theorem 2, the similar
vertices are in cells with the same label, and thus the vertex аj+1 in SC (ПI) will
correspond to the vertex bj+1= f(аj+1) in SC (ПII). So, the selection аj+1 is possible in
execution I and the selection bj+1 = f(аj+1) is also possible in execution II. The conditions
of Theorem 2 also hold for these selections and at the (j+2)th selection similar vertices
can be selected again, i. e. аj+2, in the execution I and bj+2 = f(аj+2) in the execution II.
This process continues until the last selections аLК-1, bLК-1=f(аLК-1) in both executions
have been done – after these selections the partitions will be discrete. This means that
each numeration
П(а1, а2, ..., аj-1, p, аj+1, аj+2, ..., аLК-1)= |x1| x2| . . . | xi| . . . | xn|=П′
has a corresponding numeration
П(а1, а2, ..., аj-1, q, bj+1, bj+2, ... , bLК-1)= |y1| y2| . . . | yi| . . . | yn|=П″ , such that
NC(xi,П′)=NC(yi, П″), yi=f(xi), i=1, 2, ... , n.
There are four obvious corollaries of Theorem 3:
Corollary 1 of Theorem 3 Under the conditions of Theorem 3 the discrete partitions of
type П″, successors of the partition П(а1, а2, ..., аj-1, q), do not find new automorphisms
(new similar vertices). Consequently, it is not necessary to determine them if we
pre.liminarily know the discrete partitions of type П′ - successors of the partition
П(а1, а2, ..., аj-1, p).
Corollary 2 of Theorem 3 The bouquets of two similar vertices in a given SC(ПL) are
of the same size.
This statement is obvious because to each numeration of the one bouquet uniquely
corresponds a numeration of the other bouquet.
Corollary 3 of Theorem 3 To determine whether two vertices X and Y in SC(ПL) are
similar we need to know the bouquet of one of the vertices, say B(L, X), and generate
one numeration n1 derived of a selection Y in SC(L) and compare n1 with the
numerations B(L, X). Even more, B(L, X) should not contain equivalent numerations
- 14 -
because of the transitivity of the equivalence: if n1 is equivalent to one of them it is
equivalent to the another.
Corollary 4 of Theorem 3 The bouquet B(L, X) contains all bouquets derived from
each representative of an orbit in SC(ПL+1).
.Important conclusions follow from Theorem 3 and its corollaries. There are three
possibilities to determine whether two vertices X and Y in SC(ПL) are similar under
A(x1, . . . ,xL-1): (a) The bouquet B(L, X) of the vertex X should be stored and for the
vertex Y we should generate only one numeration and compare it with the
numerationsB(L, X) - this version is used in Vsep-e algorithm; (b) One numeration
should be stored for vertex X and the whole bouquet for vertex Y should be generated.
This version is used in Nauty [12] and in the most of the known algorithms; (c) Two
bouquets В(L, X) and В(L, Y) are partially generated and their numerations are
compared for determining an automorphism (with a certain probability) that maps X to
Y. This probability might be near to 1 if we choose an appropriate selection of the
bouquets size. This is the basis for the heuristic algorithms described in Section 5.
versio
n
a
b
NS-number
stored
numerations
m
1
of NG-number of
generated
numerations
m+q-1
m+m(q-1)=q.m
NC
number
comparisons
of numerations
c.m(q-1)
c.m(q-1)
of
Table 1
Let’s compare versions (a) and (b) (Table 1). Let m=|B(L, X)| and let’s consider that the
numerations of the bouquets are stored in a hash table with a maximum number c of
collisions of some hash function (characteristic of the numeration) we’ll explain below.
Let’s also consider the worst-case – a rigid regular graph for L=1 and |SC(П1)|=q – in this
case all vertices in SC(П1) are not similar each other. This is the worst-case since: (i) for
L=1 the bouquets have the larger size than the bouquets for L>1 and (ii) the bouquets for
rigid graphs are full – each vertex at each level is selected. The advantage of a version (b)
is a low storage – only one numeration is stored and the disadvantage of version (a) is the
large required storage – the whole bouquet of size m for the first vertex x1 SC(П1) is
stored. A version (a) is faster since the number of the generated numerations is smaller:
NG=m+q-1. In this case the bouquet of the first vertex x1 SC(ПL) is generated and stored
and for each of the other q-1 vertices only one numeration is generated - totally NG =m+q1. In case of version (b) for each vertex x SC(П1) all m numerations of bouquet В(1, x)
are generated, i.e. NG= m.q – we suppose that the size of each bouquet is m or m is the
largest size. Since m.q >> m+q-1 version (a) is many times faster than version (b). The
number (NC) of the comparisons of numerations is c.m(q-1) for both versions. In version
(a) one numeration for each vertex in SC(1) is compared with c.m numerations of B(X1). In
- 15 -
version (b) the only stored numeration n1 derived from the selection X1 is compared with
c.m numerations of each bouquet B(X), x SC(1), x≠x1.
Examples:
Graph A29_1 (rigid regular graph from [29]):n=29, m=14 (this size is for each vertex in
SC(П1), q=n=29; NG(a)=m+q-1=42, NG(b)=q*m=29*14=406 (in the brackets is the
numbering of the version). We see the big difference between the .numbers of
generated numerations of the two versions.
Graph G1275 (Rigid affine plane of order 25, received from R. Mathon in private
communication): n=1275, bipartite graph with k=625*26=650*25=16250; m=8322060;
q=625; NG(a)=m+q-1=8322060+625-1=8322684; NG( b)=q*m = 625*8322060
=5201287500. In this case the difference between NG (a) and NG (b) is impressive.
4. S-code of a partition and storing the bouquets
We propose new code, named S-code, of a partition of the graph vertices. The partition
code is a number depending on the labels, sizes of the partition cells and the number of the
edges between the cells. S-code is used for reducing the time of comparing the partitions in
the graph isomorphism and automorphism algorithms. The code of a given partition can be
computed directly from the partition and the graph representation or from the code of the
parent partition and the differences between the partition and its parent partition.
In our algorithms a large number of discrete partitions (numerations) of graph vertices
are generated and stored. The length of each partition is n (n is the number of the graph
vertices). One way of reducing storage requirements is the coding of partitions. To every
partition is assigned a code (a number, characteristic value). The codes of two partitions are
compared (instead of comparing the corresponding partitions) and if they are equal then the
partitions are compared to determine if they form an automorphism. In this case the
partitions have to be regenerated using the stored base B, p=|B| of the partition and
applying the IR operation p times. The cardinality p of the base is many times less than n.
Let’s consider the storing of a partition L obtained from the start partition 0 by applying
the IR operation successively. There are 3 ways of storing the partition L: a) storing the
partition itself, i.e. n numbers are stored; b) storing the base B(L) and the code c(L) , i.e.
p+1 numbers are stored – one for the code and p for the base. This way the amount of the
stored information is reduced from n to p+1 numbers, where p << n. In this case a
regeneration of L is made when using of L is needed; c) storing the code c(L) and a
polynomial code of B, i.e. only 2 numbers are stored but a regeneration of both L and B is
needed. In our implementations the version (b) is used.
The requirements for the code are:
i) The codes of the equivalent partitions have to be equal;
ii) The splitting ability of the code has to be maximal. This means that the number of not
equivalent partitions with equal codes have to be minimal (minimum collisions);
iii) The computation of the code should have minimal number of operations (easy to
compute);
We have examined a few versions of coding and the code with the best satisfaction of the
requirements is the following:
- 16 -
code( )
L(C( x)).L(C( y)) , where:
(4.1)
( x , y )E
- the adjacency refinement partition of the vertices of graph G(V,E), (x,y)- an edge of
the graph, C(x),C(y) – the cells of the vertices x,yV and L(C(x)), L(C(y)) – the labels of
the cells C(x),C(y). The label of a cell is the index of the first vertex in the cell representing
a partition as an array.
The following theorem proves that the requirement (i) holds for the code (4.1):
Theorem Given a graph G and 1 ≡ 2, then code(1) = code(2).
Proof Each edge (x,y)E(G) has unique image (α(x),α(y))E(G) for an automorphism α
defining the equivalency of 1,2. Moreover, C(x) = C(α(x)), C(y) = C(α(y)) – the similar
vertices are in namesake cells. Therefore,
L(C(x)) = L(C(α(x))), L(C(y)) = L(C(α(y))) and
L(C(x)).L(C(y)) { code(1) }= L(C(α(x))).L(C(α(y))) { code(2)}.
Consequently, code(1) = code(2), since the last equation holds for each
edge E(G).
Evidently, the time complexity of computing the code by (4.4.1) is T=k=O(n2)
multiplications (k-the number of the graph edges) since k= O(n2). The code of can be
computed directly by (4.4.1) or indirectly by the code of the parent partition of . The
maximal value of the code Max(Code()) is obtained for a discrete partition of a
complete graph on n vertices (since its number of vertices is largest) :
n 1
Max(code(π)) i (i 1 i 2... n 1 n) = ∑𝑛−1
𝑖=1 𝑖
1
i 1
(𝑛+𝑖+1)(𝑛−𝑖)
2
=
𝑛−1 3
𝑛−1 2
= ((𝑛2 + 𝑛) ∑𝑛−1
𝑖=1 𝑖 − ∑𝑖=1 𝑖 − ∑𝑖=1 𝑖 ) =
2
1
𝑛(𝑛−1)
𝑛2 (𝑛−1)2
𝑛(𝑛−1)(2𝑛−1)
𝑛(𝑛−1)(𝑛+1)(3𝑛+2)
= ((𝑛2 + 𝑛)
−
−
)=
(4.2)
2
2
4
6
24
The second multiplicand of the first expression in (4.4.2) is a sum of an arithmetic
progression.
Example: Let’s consider the graph in Figure 5 and the series of partitions and their codes:
Figure 5. G8 graph
π0=π(0) = | 1,2,3,4,5,6,7,8| = C10 ; Code(π0)=12; π1= π(1) = RP (|1| 2,3,4,5,6,7,8|) =
|1|7,8|5,6|3,4|2| = C11 C21 C31 C41 C51 ; Code(π1)=208; π2=π(1,7) = RP (|1|7|8|5,6|3,4|2|) =
|1|7|8|5,6|3,4|2|) = C12 C22 C32 C42 C52 C62 C72 Code(π2)=218; π3=π(1,7,5) =
RP(|1|7|8|5|6|3,4|2|) =|1|7|8|5|6|4|3|2|) = C13 C23 C33 C43 C53 C63 C73 C83 ; Code(π3)=234. We
illustrate only how Code(π1) is computed (Table 2).
- 17 -
1,2
1,3
1,4
2,3
2,4
3,5
4,6
5,7
5,8
6,7
6,8
7,8
Code (π1)
L(C(x))*
1*8
1*6
1*6
8*6
8*6
6*4
6*4
4*2
4*2
6*2
6*2
2*2
208
L(C(y))
=8
=6
=6
=48
=48
=24
=24
=8
=8
=12
=12
=4
edge
(x,y)
Table 2
5. The exact Vsep-e algorithm
In this and the following section we describe the proposed algorithms some of which
were preliminary presented in arXiv [30].
5.1. Basics of the algorithm
We need the following theorem for the reasoning of Vsep-e algorithm.
Theorem 4 Given: A=Aut(G(V,E)), an orbit Q=Orb(x1,A) of vetex x1V and a generating
set gen(A(x1)) of a stabilizer A(x1). Then, there is a tower of acsending subgroups of A
A(x1)=A(1) A(2) . . . A(i) . . . A(m)=A, such that: (a) A(i) is a proper subgroup of A(i+1),
1 i m-1; (b) gen(A(i))={ gen(A(i-1)),i}, 2 i m, where the autmorphism i is such that
xi=i(x1); (c) | A(i)| 2.| A(i-1)|, |Orb(x1, A(i))| 2.|Orb(x1,A(i-1))|, 2 i m; (d) m is the
minimal value of i such that Orb(x1, A(m))=Q.
{Note: Evidently, the consequence of (b) is gen(A)=gen(A(m)))={gen(A(x1)), 2, 3,…,m.
Theorem 4 can be considered as an extension or modification of Theorem 5 of C. M.
Hoffmann in [10], page 25}
Proof (inductive, it follows the proof of Theorem 5 in [10]). Let’s construct a tower of
ascending subgroups of A: A(x1)=A(1) A(2) . . . A(i) . . . A(m)=A, such that A(i) is a
proper subgroup of A(i+1), i=1,2, . . . ,m-1; m is finite since A is a finite group. For i=1 we
have A(1)= A(x1) and gen(A(1)) =gen(A(x1)), known. Assume inductively that A(i) is a proper
subgroup of A and let we have orbits Q= Orb(x1, A(i-1))≠Q= Orb(x1, A(i-1)), Q, QQ.
Then, we’ll have an automorphism iA\ A(i-1) such that xi=i(x1) and Orb(x1, gen(A(i)))
Q Q. Thus, i is a generator of A and A(i) since it does not belong to A(i-1), i.e.
gen(A(i))=gen(A(i-1))i and therefore |A(i)|2.| A(i-1)| (since at least one new coset appear in
the partitioning of A(i) into cosets of A(i-1)) and |Orb(x1,A(i))| 2.|Orb(x1,A(i-1))| (from the
orbit-stabilizer theorem: |Orb(x1,A(i))|= |A(i)|/.|A(x1)| ≥|Orb(x1,A(i-1))|=||A(i-1)|/A(x1)|).
The first value of i when Orb(x1,gen(A(i)))=Q and A-A(i)= is i=m and consequently
gen(A)=gen(A(m)))={gen(A(x1)), 2, 3, . . . , m}.
Theorem 4 gives us the idea how to find the generators of a group if we know an orbit Q
of the group and the generators of a stabilizer of a representative x1 of this orbit. This is
done by traversing the orbit step by step. At each step we find one new generator of a new
subgroup of A knowing the generators and the orbits of the previous subgrpoup of A.
Before the first step the previous subgroup is equal to A(x1) with its orbits and generators.
- 18 -
Visiting each vertex x of the orbit, xx1, we select x only if x is not similar to x1 under the
previous subgroup. Thus a new automorphism-generator that maps x to x1 and new orbits
for a new subgroup are found. The new subgroup is a proper supergroup of the previous
subgroup.. At each step the generators, orbits and order of the current group A (subgroup
of A) are defined by the position of the selected vertex x. This process stops when the orbit
of x1 under the new subgroup becomes equal to the given orbit Q. The described process is
presented in Table 3. The selected vertex xi, i=2, . . . ,m is the first vertex after x i-1 in Q that
is not similar to x1 under A(i-1). The generators of A(x1) are called proper generators of A
and the generators 2, 3, . . . , m - mutual generators of A. The following corollary is
obvious:
Selec- Visited Mutual
gen(A)
ted
vertices generators
vertex
x
x1
{x1}
gen(A(x1))
(proper generators of
A)
x2
{x1:x2} x2=2(x1) gen(A(x1)) 2
x3
{x1:x3} x3=3(x1) gen(A(x1)) 23
x4
{x1:x4} x4=4(x1) gen(A(x1))
234
...
...
...
...
xi
{x1:xi} xi=i(x1) gen(A(x1)) 2 . . .
i
...
...
...
...
xm
{x1:xm} xi=m(x1) gen(A(x1)) 2. . .
=Q
m
|Orb(x1,A)| |A|
A
1=|{x1}|
|A(x1)|
A(1)=
A(x1)|
2
4
8
(2)
2.|A(x1)| A
(3)
4.|A(x1)| A
(4)
8.|A(x1)| A
...
2i-1
...
2i1
.|A(x1)|
...
2m1
.|A(x1)|
...
|Q|2m-1
...
A(i)
...
A(m)=A
Table 3
Corollary of Theorem 4 The number of the mutual generators of the graph automorphism
group A toward the stabilizer A(x1) is m=1+log2|Orb(x1, A)|.
Based on Theorems 3 4 and O-S theorem we build a new algorithm called A3. Actually,
Theorem 4 tell us that there is a generator xi=i(x1) if xx1 is not similar to x1 under the
current group but according to Theorem 3 to determine this generator we should know the
bouquet B(x1) and one numeration derivative of x and to compare them. Knowing the
partition П1 and SC(П1) according to Theorem 4 to traverse the unknown Q orbit we
should traverse its superset, SC(П1) Q. Thus, we come to the idea of A3 algorithm
(Figure 6). We denote by FRPO(X) – all first representatives of the orbits of A positioned
in SC(П1) before the selected vertex X and by BFRPO(X) – the bouquets of FRPO(X).
Then, the bouquet B may be considered as a union of the bouquets BFRPO(X).
- 19 -
A3 algorithm determines GOO(A) of the graph automorphism group A=Aut(G,1) ) and
the bouquet of each representative of an orbit Orb(A)SC(1) given GOO(A(X1)) of a
stabilizer A(X1) of the first vertex X1SC(1) and the bouquet B(X1).
According to Theorem 3 and 4 we’ll determine GOO(A) and the bouquet of each
representative of an orbit Orb(A)SC(1) visiting each vertex xSC(1) positioned after
X1. Before the traversal of SC(1) we have GOO(A)=GOO(A(X1)). Each visited vertex
xSC(1) is selected if it is not similar to any previous vertexSC(1) under the current
group A. Then, we determine if there is an automorphism , x=(y), where y is one of the
roots of numerations in B. This is made (according to Theorem3) by comparing the first
numeration LK derived from the selection x with the numerationsB. Two cases are
Input: a graph G; a partition 1=RP() for a given input partition ; SC(1); a vertex XF
SC(1) and its index in SC(1); GOO(XF): the generating set gen(A(XF)) of a stabilizer A(XF),
the orbits of A(XF) and |A(XF)| and the bouquet B(XF); the bouquets BFRPO(XF), i.e., the
bouquet of each first representative y of an orbit of A in SC(1) with index (y) < index (XF); each
representative y and XF has no similar vertices in a position in SC(1) before it.
Output: The generators, the orbits and the order (shortly GOO) of the graph automorphism group
A=Aut(G, 1); the bouquet B of each representative of an orbit of A in SC(1).
1.
orbits of A:=orbits of A(XF); gen(A):=gen(A(XF)); |(A)|:= |A(XF)|
2.
X:=XF;
3.
Do
4.
select next X SC(1); if there is no selected vertex X then exit;
5.
determine a numeration n1 by SFM(X);
6.
compare the numeration n1 with the numerations of the BFRPO(X);
7.
if n1 is equivalent to some numeration from BFRPO(X), i.e. there is new automorphism
mapping X to some vertex from FRPO(X) then
8.
gen(A):= gen(A); recompute the orbits of A; |A|:=|Orb(XF)|.|A(XF)|
9.
else {n1 is not equivalent to any numeration from BFRPO(X), i.e. there is no new
automorphism mapping X to some vertex from FRPO(X)}
build the search tree ST(X) for determining GOO(A(X)) and the bouquet B(X)
10.
11.
end; {if}
12. end do
Figure 6. A3 algorithm
possible: (a) If there is an automorphism then it belongs to a new subgroup of A since it
unites the orbits of x and X1, i.e., gen(A): = gen(A) {}; (b) If LK is not equivalent to any
numeration B then x becomes a representative of a new orbit and a new search tree ST(x)
is built - it determines gen(A(x)) and B(x). If during the generation of ST(x) an
automorphism is discovered and if it unites the orbits of A then, it is also a generator for A
since A(x)A. Thus, after case (b) for the next selected vertex we should compare the
numeration LK derived from the selection of x with the numerations of all known
bouquets B.
To determine the orbit of the vertex X1 it is sufficient to do the above comparison only
with the bouquet B(X1) but we do this comparison with the numerations of all bouquets of
- 20 -
B since we store them and the earlier finding of a generator is better since we can use it
earlier. If the found automorphism , x=(y), yX1 unites orbits of the current A it is a
generator of A.
After each of both cases has been handled the traversal of SC(1) continue by selection of
a new vertex x. Thus, after the traversal of SC(1) all verticesOrb(X1, A) will be visited
and the bouquet of each first representative xiSC(1) of an orbit of A and GOO(A) will be
determined (according to Theorem 4).
During the execution of A3 algorithm (as in Theorem 4) a tower of ascending subgroups of
A are built implicitly: A(x1)=A(0) A(1) . . . A(i) . . . A(m)=A, such that A(i) is a proper
subgroup of A(i+1), 0 i m-1.
A3 algorithm is applied differently for L partition of level L=1 and L>1. If L=1 then
GOO(X1) and B(X1) are determined only for the first vertex X1 SC(1) and for other
orbit representatives in SC(1) only one numeration is generated and compared with
B(X1). This means that instruction in line 10 of A3 is not necessary if L=1. If L>1 then
GOO(XL) and B(XL) are determined for all orbit representatives XL SC(1) since they
are needed for A3 application to L=L-1.
A3 algorithm can be applied for determining any GOO(A(XL-1)) and B(XL-1), L=2, . . . ,
LK-1 under the following requirements:
R1. For L known are: ПL, SC(ПL), a vertex XFLSC(ПL); its index in SC(ПL, GOO( A( XFL )) ,
B( XFL ) , XFL is not similar to any previous vertex in SC(ПL) under A(XL-1), XL-selected
vertex, bouquets BFRPO(XL); GOO(XL-1)= GOO(XL) at the start;
R2. XL is selected in interval index(XFL)+1 to the end of SC(ПL), XL is selected as a first
not similar vertex after the current XL under the current A(XL-1). At the start XL is the
vertex with index=index(XFL)+1. This means that the vertices of Orb( XFL , A( X L1 )) in
SC(ПL) are after XFL . Thus, traversing SC(ПL) we’ll traverse Orb(XFL,A(XL-1))– this is the
requirement of Theorem 4;
R3. If XL is similar to some vertexFRPO(XL) then the discovered automorphism is a
generator of A(XL-1): gen(A):= gen(A). Otherwise, each new numeration in B(XL)
(instruction 10) belongs also to B(XL-1).
5.2. Vsep-e exact algorithm: cases CS1, CS2, CS3 and CS4
Let we have the starting series SFM1: П1, П2,…, ПLK. We can apply A3 algorithm
directly only to the partition ПLK-1 because for the other partitions the required input
variables are not known. For the partition ПLK-1 we have B(xLK-1) ={ПLK} and
gen(A(xLK-1))=, i.e., |A(xLK-1)|=1 and discrete orbits of A(xLK-1). After the application of
A3 to ПLK-1 we have determined correctly B(xLK-2) and GOOGA(A(xLK-2)). Then, A3 can
be applied to ПLK-2, i.e., a backward move is done from LK-1 to LK-2. Thus, applying A3
to the series ПLK-1, ПLK-2,…, П2, П1 we can determine GOO(A). The lowest level to which
a backward move has been made we denote by LMIN, i.e., LMIN is the level for which we
determine GOO(A(XLMIN-1)). In A3 algorithm the process of the backward moves is not
included and the instruction 10 is not revealed. All this is taken into account in PART1
(Figure 9) and PART2 (
Figure 11) algorithms called from the Vsep-e algorithm
- 21 -
(Figure 8) that determines GOO(A) of the partition-wise stabilizer A=Aut(G, П) given a
graph G(V,E) and the input partition П on V. S3.
Before calling PART1 and PART2 Vsep-e algorithm determines (step S2) the orbits of A
by an TREE4 heuristic algorithm (see section 7.1) and selects X1 SC(П1) - a
representative of one of the smallest orbit of the vertices in SC(П1). Experimental tests
show with very rare exceptions that if the staring vertex X1 is a representative of one of the
smallest orbit of A then the size of the bouquet B(X1) built by PART1 is the smallest and
the running time is minimal. PART1 algorithm (Figure 9) can be considered as an
application of A3 algorithm with added the backward moves and revealed instruction 10 –
all above requirements are implemented in it. PART1 algorithm determines GOO(A(x1)
and B(x1) given П1, SC(П1) and x1SC(П1). PART2 algorithm (the second part of Vsep-e
algorithm, line S5 in Figure 8) determines GOO(A) given GOO(A(x1) and B(x1) obtained
from PART1 algorithm.
PART2 algorithm can be considered also as an application of A3 algorithm to the partition
П1 with replacing the instruction 10 by determining one derivative numeration ПLK of each
selected vertex xSC(П1) if x is not similar to x1 under the current A. PART2 algorithm
may be also considered as a direct application of theorems 3 and 4. The satisfaction of the
requirements of these theorems guarantees .the correctness of PART2 At the start of
PART2 GOO(Aut(G):= GOO(Aut(G,x1)). Each vertex x in SC(П1) that is not similar to x1
under the current Aut(G) is selected (line B2) and a comparison (line B5) of the first
numeration LK derived from the selection x (line B4) with the numerationsB(x1) is
made. If there is an automorphism between some numeration B(x1) and LK then is a
generator for A since it unites the orbits of x and x1. In both cases (existence or
nonexistence of ) the traversing of SC(П1) continues until its end. When the traversal of
SC(П1) completes, the generators and the orbits of A are determined and the ‘orbitstabilizer’ theorem is applied for determining |A|=|Orb(x1, A)|.| A(x1)| (line B3).
We’ll describe PART1 algorithm considering an intermediate state of ST search tree
(Figure 7) being built by the algorithm during its execution. The series of partitions ПL,
L=2, . . . , LK-1 can be divided into three intervals: the first is from П1 to ПLMIN-1, the
second – from ПLMIN to ПLP and the third – from ПLP+1 to ПLK-1. The search tree is built in a
preorder: first visiting the root (a partition ПL) and then its subtrees (the partitions ПL+1
derived from each selected vertex) in a defined order.
Applying A3 algorithm for determining gen(A(xL)) we have:
gen ( A( xL )) gen ( A( x1L1 )) MG( x1L1 , xL1 ) gen ( A( xL1 )) ,
where gen( A( xL )) is the generating set of the current stabilizer A( xL ) , gen( A( x1L1 )) is the
generating set of the stabilizer A( x1L1 ) , MG( x1L1 , xL1 ) is th.e set of the mutual generators of
A(xL) and gen( A( xL1 )) is the generating set of the current stabilizer A( xL1 ) . The following
conditions hold for the intermediate state of PART1 algorithm execution (Figure 9) shown
as a search tree ST on Figure 7:
C1: On the current path of ST tree known are: L, ПL, SC(ПL), XL, L=1, 2, . . . ,LK-1, XL is
the current selected vertex in SC(ПL). The current numeration is ПLK-1=n1;
C2: LMIN, LP, LMINLP, LP2 are known;
C3: X1 is the first vertex SC(П1);
- 22 -
C4: Each selected vertex XL is not similar to any previous vertex in SC(ПL) under A(XL-1),
L=2, . . . ,LP-1;
C5: Each selected vertex XL is the first vertex in the SC(ПL) for L=LP+1, . . . ,LK-1; The
partitions in this section of the path are a result of the forward move SFM(XLP);
C6: Known are: GOO(XLMIN-1), B(XLMIN-1), FRPO(XLMIN);
C7: Known are the bouquets BFRPO(XL), L=1, 2, . . . , LP;
C8: Known are the orbits and orders |A(X)| of vertices X FRPO(XL) and X= XL,
L=LMIN,…,LP, under A(XL-1).
C9: On the current path known are the computed orbits Orbc(Xl) and computed orders
AC(Xl) for L=LMIN+1,…,LK-1 under the current A(Xlmin-1). For L=LMIN we have
Orbc(Xl)= Orb(Xl) and AC(Xl)= A(Xl) under the current A(Xlmin-1) since all generators
found so far have the same fixed points with A(Xlmin-1).
The action that follows the above state is a comparison of n1 numeration with the
numerations of BFRPO(XLP) for discovering a new generator of A(Xlmin-1).
All above conditions can be considered as an invariant for correctness of the loop C1 of
PART1algorithm.
Let’s now describe PART1 algorithm. It calls SFM1 (Figure 10) and COMP (Figure 13)
algorithm. At the start all of the searched variables are not known and for each partition
ПL, L=2, . . . , LK-1 we select the first vertex XLSC(ПL) and obtain the partition ПL+1, i.e.
the only action we do is a forward move (line I1) until a discrete ПLK is obtained. Thus, the
conditions R1 to R3 hold for only for ПLK-1.Let’s now consider the above requirements R1
to R3 for determining GOO(A(XLP-1))=GOO(A(x1, , . . . , xLP-1)) and the bouquet B(XLP1)=B(x1, , . . . , xLP-1) given the partitions П, П1, . . . , ПLP. For ПLP=П(x1, , . . . , xLP-1) also
SC(ПLP) and the selected vertex XLP are known. Besides, the requirements hold for the
position of the current vertex XLP in SC(ПLP). By the loop C1 (lines I1- I12) in PART1
algorithm each selected cell SC(ПLP) is visited and the following four basic steps are
performed:
A1 {Selection}: The selection of a vertex XLP in SC(ПLP) (line I3) is made starting from the
position next to the current XLP. The vertex XLP should not be similar to any previous
vertex in SC(ПLP) under the current A(XLMIN-1). For each level L the position i(XL) of the
selected vertex XL is stored and when a backward move to this level is performed then the
selection of a new vertex starts from the next position, i.e. i(XL)+1. If there is no selected
vertex in SC(ПLP), i.e. the SC(ПLP) has been traversed then, a backward move follows (step
A4). If there is a selected vertex XLP in SC(ПLP) then, step A2 follows.
A2 {Series of forward moves}: A series of forward moves SFM1 is performed determining
the partitions ПL=ПLP+1, . . . , ПLK with discrete ПLK (line I5). In each of these partitions the
selected vertex XL is the first vertex in SC(ПL). This way the requirements hold for the
orbit Orb(XL, A(XL-1)).
A3 {Comparison}: A check if there is a new automorphism that not belong to the current
subgroup of A(xLP-1) and maps the vertex X to any vertex from FRPO(X) is made, i.e. if
belongs to the next subgroup of A(xLP-1). This check is made by comparing ПLK with
BFRPO(X) (line I6, COMP algorithm).
A4 {Backward move}: After the traversal of SC(ПLP) is completed then
GOO(A(xLP-1)) and B(xLP-1) are determined and a backward move
- 23 -
LP:=LP-1 is made. Stop follows if LP=1. Otherwise, a selection of a new
vertex in SC(ПLP) is made applying the step A1 to it.
If there is an automorphism then, it is a generator: gen(A(XLP-1)) := gen(A(XLP-1)){}
and the orbits and the order of A(XLP-1) are recomputed. It is a generator also for A(X1):
gen(A(X1)):=gen(A(X1)){} and the orbits and the order of A(X1) are recomputed. If
there is no mapping x to a vertexFRPO(X) then a move back to LK-1 follows. This
way the building of the tree ST(x) starts from LK-1 performing the step A1 to SC(ПLK-1).
ST(x) is necessary since it determines the bouquet B(x) that belongs to B(XLP-1). After
ST(x) has been built we continue with a selection of a new vertex in SC(LP) applying the
step A1 to SC(ПLP). The search tree is built in preorder traversal: first visiting the root (ПLP
partition) and then its subtrees (the partitions ПLP+1 derived from each selected vertex) in a
defined order.
For the selected vertex XLP (line I3 – the start of ST(XLP) building) by SFM1 (line I5) is
built the first (leftmost) tree ST(XLP+1), ST(XLP+2), . . . , ST(XLK-1) for each previous
subtree. Each of these subtrees is built in backward ord..er. When the subtree ST(XLP+1) has
been built then the building of the subtree for the next selected vertex XLP+1 starts (XLP+1
should hold the requirements). When there is no selected vertex XLP+1 then a backward
move LP+1 to LP is made – this means that the ST(XLP) is built. If
LP=1 the algorithm stops.
5.2.1. Cases CS1 and CS3
Figure 7. Search tree of VSEP-e
L=1
L=2
L=3
x1
x2
x3
LMIN
LMIN-1
xLMIN-1
.
.
LP-1
.
.
LP
.
.
B(a1,1LP)
xLP
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
SC(1)
SC(2)
SC(3)
.
.
.
.
.
.
.
.
.
.
.
.
.
SC(LMIN-1)
.
.
.
.
.
.
.
.
.
.
B(a2,1LP)
xLP
2
.
.
.
.
xLMIN
.
.
.
.
.
.
xLP-1
.
.
.
.
xL
LP+1
.
.
.
.
LP+2
LK-1
LK
n1
i
P
.
.
.
.
.
.
.
.
.
.
.
.
SC(LMIN)
SC(LP-1)
SC(LP)
- 24 -
.
.
.
.
.
.
.
.
SC(LP+1) SC(LP+2) SC(LK-1)
Let us consider the cases when the numeration .ПLK (Figure 7) does not form an
automorphism with any numeration BFRPO(ПLP). Knowing that |A(XLP-1)| is correct and
supposing that each orbit Orb(XL, A(XL-1)) for L=LP, . . . , LK-1 is also correct, and
applying the Theorem ‘O-S’ we obtain
|A(XLP-1)|/|Orb(XLP, A(XLP-1))|/|Orb(XLP+1, A(XLP))|/ … /|Orb(XLK-1,A(XLK-2))| =
|АLK-1| = 1.
(5.2.1)
However, the orbits in (5.2.1) are unknown and consequently we cannot use it. Instead, the
sets W=COrb(XL, A(XLP-1)) of the vertices in SC(ПL) similar to XL under A(XLP-1) are
known, i.e.
W=COrb(XL, A(XLP-1))= Orb(XL, A(XLP-1)) SC(ПL).
(5.2.2)
We call the set W ‘computed orbit’ to distinguish it from the real orbit U=Orb(XL, A(XL)).
The following relation holds U=Orb(xL,A(xL-1))COrb(xL,A(xLP-1))=W,
(5.2.3)
since A(xL-1)) A(xLP-1).
Considering (5.2.3) and applying the Theorem ‘O-S’ for the computed orbits we obtain
|A(XLP-1)|/|COrb(XLP,A(XLP-1))|/|COrb(XLP+1,A(XLP-1))|/ … / |COrb(XLK-1,A(XLP-1))| =
|АLK-1| 1.
(5.2.4)
The computation on (5.2.4) is performed by SFM1 algorithm (Figure 10) called from the
instruction I5 of PART1. If the sign in (5.2.4) is = (i.e., the real and the computed orbit are
equal), then we call the orbit Orb(xL, A(xL-1)) separated. If the sign in (5.2.4) is < , then,
obviously, W=UR, |R|>1, i.e., W (and R) include vertices that belong to SC(ПL) but are
not similar to XL under A(XL-1). It can be proved that when W contains some vertex, then it
contains the whole orbit of this vertex under A(XL-1).
Thus, W can be considered as an union of orbits of .A(XL-1). We call this case nonseparation of orbits (denoted by NSO) and the orbit U is called non-separated (nonpartitioned). Since the representatives of the orbits belonging to R are not known we
cannot select them during the traversal of the SC(ПL). Thus, the search tree of such a
Input: Graph G(V,E) and a partition П on V
Output: Generators, orbits and order of the partition-wise stabilizer A=Aut(G, П), denoted GOO(A)
S1: Initialization: П1:=RP(П, NCL);{NCL-the number of the cells in П1}
S2: if NCL=n {П1is discrete} then message'Trivial group'); return end if;
S3: CSLCh {call cell selector chooser-finds the cell selector};
S4: Determine SC(П1) and х1:=first vertex in SC(П1);
S5: Use a filter that defines whether or not to call the following procedures that:
a) Give a better partition П1 obtained by S code sorting of the vertices in П1. Determine a new
SC(П1) and х1:=first vertex in SC(П1);
b) Determine: (i) the generators and orbits of A by TREE4 heuristic algorithm,
starting from discrete orbits; (ii) SC(П1); X1:=a representative of the smallest
orbit of the vertices in SC(П1); Put the vertex х1 on the first position in SC(П1);
S6: PART1: Determine GOO(A(х1)), B(х1) given П1, SC(П1), х1SC(П1)
S7: PART2: Determine GOO(A) given GOO(A(х1)), B(х1)
Figure 8. Vsep-e algorithm
- 25 -
vertex cannot be built and its bouquet will not be determined. This is an unallowable error
since these bouquets belong to B(XL-1), B(XL-2), . . . , B(X1) and they are needed (as we
know from A3 algorithm) for determining
GOO(A(XL-2)), GOO(A(XL-3)), . . . , GOO(A(X1)). If there is only one partition with nonseparated orbit, then the sign in (4.2.1.4) is <, since the length of the computed orbit is
greater or equal to the length of the real orbit – this is the condition to detect the presence
of NSO. We call this case CS3. If there is no NSO in any partition, i.e. each computed
orbit is equal to the real one, then the sign in (4.2.1.4) is = and the case is denoted by CS1.
If the case is CS1, then at the exit of COMP algorithm (line I6 in PART1) the invariant for
correctness holds for LP=LK-1: only LP is changed, LMIN remains the same. When the
case is CS3 we lose GOO(XLMIN-1) determined so far and the algorithm continues with a
new start point: LP=LK-1, LMIN=LP, XLMIN= XLK-1 is the first vertex in SC(LK-1) and
A(XLMIN-1)={I}, i.e., each vertex is put into a separate orbit. Obviously, the invariant for
correctness holds for CS3 case.
Input: П1, SC(П1), х1
Output: GOO(A(х1)), B(х1)
I1:SFM1:Determine ПL,|A(XL-1)|=1,SC(L),L=2, . . . ,LK given ,П1,SC(П1),X1;
LP:=LK-1; LMIN:=LP; gen(A(х1)):=; |A(х1)|:=1; B(XLK-1):={ПLK};
I2: do {loop C1: Visit vertices in SC(ПLP) }
I3: Select XLP in SC(LP);
I4: if XLP0 then {forward move}
I5: SFM1: Determine ПL, |A(XL-1)|, SC(L),L=LP+1, . . . ,LK given LP, ПLP ,
SC(ПLP), XLP;
I6:
COMP: Compare ПLK with the numerationsBFRPO(XLP) for determining an
automorphism A(XLP-1). Determine GOO(A(XLMIN-1)), LP;
I7: else {backward move: SC(LP) has been traversed}
I8:
LP:=LP-1;
I9: if LP<LMIN then LMIN:=LP; |Aut(XLMIN)|:= |Aut(XLMIN+1)| end if;
I10: if LP=1 then return end if;
I11: end if
I12:end do;{ loop C1}
Figure 9. PART1 algorithm (instruction S6 of Vsep-e algorithm (Figure 8))
Input: L=LP, ПL, SC(ПL), XL; Output: ПL, |A(XL-1)|, L=LP+1,… ,LK
1. do
2. L:=L+1; ПL:=RP(ПL-1,XL-1,BRCL);
3. if BRCL=n then return end if
2. 4. determine SC(L); XL:=first vertex in SC(L);
5. |A(XL)|:= |A(XL-1)|/|Orb(XL, A(XLMIN-1)SC(ПL)|
5. end do
Figure 10. SFM1 algorithm (instruction I5 of PART1 algorithm (Figure 9)
- 26 -
The operations in the case CS3 may be considered as an error correction of the incorrect
orbits of some A(XL-1) determined by the moment since the algorithm interrupts its current
execution and starts from the new start point for finding the correct orbits of A(XL-1) and
the bouquet В(ХL-1) .
3.2.1. Cases CS2 and CS4
Let’s consider the cases when there is an automorphism mapping XLP to some vertex
ULPFRPO(XLP), i.e. the numeration ПLK forms an automorphism with some
numerationB(ULP), XLP=α(ULP).
Input: L=1, П1, SC(П1), х1, GOO(A(х1)), B(х1)
Outut: GOO(A)
B1: do
B2: Selecte next vertex Х in SC(П1);
B3: if X=0 then |A|=|Orb(х1, A|.|A(х1)|; return end if;
B4: SFM1A: Determine ПLK given L=1, П1, SC(П1), X SC(П1)
B5: COMP: Compare ПLK with the numerationsbouquet В(x1) for an
automorphism . If there is then gen(A):=gen(A) and recompute the
orbits of A;
B6: end do;
Figure 11. PART2 algorithm (instruction S5 of Vsep-e algorithm (Figure 8))
Input: L=1, П1, SC(П1), X SC(П1)
Output: ПLK
1: do
2: L=L+1; ПL=IR(ПL-1,XL-1);
3: if NCL=n then return end if;
4: else determine SC(L); XL=first vertex in SC(L)
5: end do
Figure 12. SFM1A algorithm (instruction B4 of PART2 algorithm (
Figure 11))
Then, there is a possibility of NSO for some orbits of the vertices of the target cells of the
current path for the levels LMIN+1 L LP: we denote by CS2 the case when there is no
NSO and by CS4 the case when there is at least one case of NSO in this interval. The
automorphism is a generator of A(XLMIN-1) since A(XLP-1) A(XLMIN-1): gen(A(XLMIN-1))
:= gen(A(XLMIN-1)) {}. Thus, the Orb(A(XLMIN)) and |A(XLMIN)| are changed (line 11)
and we denote by +α any variable with changed value. Before each orbit and each order
of the current stabilizers are correct: |A(ZL)| =|A(XL-1)|/|Orb(ZL, A(XL-1))| for each
representative of orbit ZL SC(L), L=LMIN, . . . , LP.
The following actions are:
- 27 -
a) For the interval L=LMIN to LP (loop: lines 11 to 24) of the current ST path the
stabilizers of the selected vertices XL before and after are determined (line 13) ;
Input: ПL, SC(ПL), XL, |A(XL-1)|, L=2,3, . . . , LK; LP, BFRPO(XLP), LMIN,
GOO(A(XLMIN-1))
Output: GOO(A(XLMIN-1)), LP
1: Compare ПLK with the numerationsBFRPO(XLP)
2: if ПLK is not equivalent to any numerationBFRPO(XLP)
3: then {CS1 or CS3}
4: if |A(XLK-1)|=1
5:
then {CS1}
6:
LP=LK-1; B(XLP)={ПLK}; if LP<LMIN then LMIN=LP
7:
else {CS3}
8:
LP=LK-1; B(XLP)={ПLK}; LMIN=LP; A(XLMIN-1)={I};XLMIN:=the the first vertex in
SC(LMIN); |Aut(XLMIN)|:=1
9:
end if
10: else{CS2 or CS4: there is an automorphism , XLP=(ULP), between ПLK and some
numerationBFRPO(XLP) derivative of the vertex ULPFRPO(XLP)}
11: Determine GOO(A+(XLMIN-1)): gen(A+(XLMIN-1))=A(XLMIN-1) {};
determine orbits of A+(XLMIN-1) and
| A+(XLMIN-1)|=|A(XLMIN)||Orb(XLMIN, A+(XLMIN-1))|
12: for L=LMIN, LMIN+1, . . . , LP do
13:
|A(XL)|= |A(XL-1)|/|Orb(XL, A(XL-1))|; |A+(XL)|= |A+(XL-1)|/|COrb(XL,
A+(XLMIN-1))|;
14: if L>LMIN then
15:
for each vertex ZLFRPO(XL) do{check if |A(ZL) has changed after }
16:
If |A+ (ZL)|= |A(XL-1)|/|COrb(ZL, A(XLMIN-1))|< |A(ZL)|= |A(XL-1) |/ |Orb(ZL,
A(XL-1)) |
17:
then CS4=true
18:
end if
19:
if L=LP and index(ZL)= index(ZLp)
20:
then IULP:= index(ULP); RSTBULP:=|A(ULP)|
21:
end if
22:
end for {loop for from line 15}
23:
end if
24: end for {loop for from line 12}
25: if (not CS4) then return end if
26: if RSTBULP=1
27: then LMIN:=LP;L:=LP; gen(A(XLMIN-1)):={};orbits of A(XLMIN-1) :=cycles of ;
|A(XLMIN-1)|:= |Orb(ULP, A(XLMIN-1)) |;XFLMIN:=ULP;
|A(XFLMIN-1)| :=1; the execution continues by starting the selection with a vertex with
index(ULP)+1
28: else SFM(ULP);L:=LK-1;LP:-L;LMIN:=L;A(XLMIN):={I};the next selected vertex is
the first vertex in SC(LP)
29: end if
30:end if
Figure 13. . COMP algorithm (instruction I6 of PART1 algorithm (Figure 9))
- 28 -
b) For each representative ZLFRPO(XL), L>LMIN a check if |A(ZL)| has changed after
(line 16) is performed. This check is excluded for L=LMIN since all orbits and
c) stabilizers are correct – all so far found generators have the same fixed points with
A(XLMIN-1); If there is a change of A(ZL)| for some L then CS4=true. Also, the
index(ZLP) and |A(ULP)| are stored;
d) If CS4=false (line 25), i.e. the case is CS2 then an exit from COMP follows, no
changes of the variables computed so far are made and the next selection starts from
the current selected vertex in SC (LP).
e) In CS4 GOO(XLMIN-1) found so far are lost and the algorithm continue from a new
start point as in CS3 case. The invariant for correctness holds for the new state of the
ST tree. The main requirement that hold is that the selected vertices XL of the current
path (ULP for L=P) are not similar to the previous vertices in SC(L) under A(XL-1).
If the case is CS4 (lines 26 to 30) then there are possible two subcases: i) |A(ULP)|=1
and ii) |A(ULP)|>1. If |A(ULP)|=1 then LMIN:=LP, L:=LP and a new GOO(XLMIN-!)
determined (line 17) are determined – the new start point is index(ULP) in SC(LP).
The case when A(ULP)|>1 (line 28) means that there are generators we do not know –
that’s why we do a forward move FM(index(ULP)) to come to a new start point:
L=LK-1;LP=L, LMIN=L, A(XLMIN):={I};the next selected vertex is the first vertex
in SC(LP). From registered CS4 cases in experiments on benchmark graphs there are
only cases with |A(ULP)|=1.
There is another way for CS2|CS4 check (not shown in COMP procedure): the check
for difference of the new and old stabilizer is made only for the selected vertices XL
of the path. The experiments show that both ways work correctly.
The experiments on the benchmark graphs in [19] show that CS4 case occurs only for the
graphs B52 (Mathon doubling of b25-1 graph [18]), latin-16 and 24, and had-96.
5.3. Examples
Figure 14. G10 graph
5.3.1. Simple example
Let us consider the search tree traversal in Figure 15 (in preorder) for the graph of
Figure 14 . Starting from the partition П1= П()= RP(Пu)=|2,9|4,5,6,7|1,3,8,10| we do a series
of forward moves SFM: П2 = П(4) = |2|9|4|7|5|6|8,10|1,3|, П3=П(4,8) = |2|9|4|7|5|6|8|10|1,3|,
П4=П(4,8,1)=|2|9|4|7|5|6|8|10|1|3| - the first numeration n1. The selected cell SC(ПL) in this
SFM is the cell with the largest number of adjacency cells, the selected vertex is always the
first in SC(ПL) and the order of the stabilizers A(0), A(1), A(2),…, A(L), L=1,2,…,LK-1 is 1.
Then, a back move follows, L=L-1, L=3, vertex 3 is selected and П4 = П(4,8,1)=
|2|9|4|7|5|8|10|3|1| - n2, n2 =α1(n1), α1=(1,3). Then, again back move to L=3, where there is
no selected vertex in SC(3), new back move to L=2, |A(4,8)|=|Orb(1)|.|A(4,8,1)|=2,1=2 and
- 29 -
L=1
4
П( )=
2
П1
1
SNSC(
SC(2)
2)
L=2
2 П(4)=
П2
1
4
8
5
10
SC(3)
|A(4,8,1)|
SC(1)
SNSC(
1)
|A(4,8)|
|A(4)|
selected vertex in SC(2) is 10, П3=П(4,10) = |2|9|4|7|5|6|10|8|1,3|, П43 = П(4,10,1) =
|2|9|4|7|5|6|10|8|1|3| - n3, n3 =α2(n1), α2=(8,10). At this point, there are no selected vertices
in SC(3) and SC(2), that’s why back moves to L=3,2,1 are made and |A(4)| | =
|Orb(8)|.|A(4,8)|=2.2=4. Here PART1 ends and PART2 starts generating two SFM, for
vertices 5, 6 SC(1). The first SFM is: П(5)= |2|9|5|6|4|7|6|8,10|1,3|,
П(5,8)=|2|9|5|6|4|7|6|8|10|1,3|, П(5,8,1)= |2|9|5|6|4|7|6|8|10|1|3| - n4, n4 =α3(n1),
α3=(4,5).(6,7). The second SFM is: П(6)= |9|2|6|7|5|4|1,3|8,10|, П(6,1)=
|9|2|6|7|5|4|1|3|8,10|, П(6,1,8)= |9|2|6|7|5|4|1|3|8|10| - n5, n5 =α4(n1),
α4=(1,8)(2,9)(3,10)(4,6).(5,7). Orbits of A are: (2,9)(1,3,8,10)(4,5,6,7) and
|A|=|Orb(4)|.A(4)=4.4=16.
1
L=3
П(4,8)=
П3
1
n1 = |2|9|4|7|5|6|8|10|1|3| =
П(4,8,1)
|A|
6
α4 16
α3 8
α2 4
α1 2
1
7
|A(4,10,1)|
3
α1 = (1,3) = (n1 ~ n2);
orbits: (1,3)(2)(4)(5)(6)(7)(8)(9)(10)
L=3
П(4,10)=
П3
1
L=2
П(5)=П2
1
n3 = |2|9|4|7|5|6|10|8|1|3| = П(4,10,1)
3
α2 = (8,10) = (n1 ~ n3);
orbits: (1,3)(2)(4)(5)(6)(7)(8,10)(9)
L=3
П(5,8)=
П3
8
10
1
L=2
П(6)=П2
3
n4 = |2|9|5|4|6|7|8|10|1|3| =
П(5,8,1)
α3 = (4,5)(6,7) = (n1 ~ n4);
orbits: (1,3)(8,10)(4,5)(6,7)(2)(9)
L=3
П(6,1)=
П3
1
3
8
10
Figure 15. Search tree for the graph of Fig, 14
- 30 -
n5 = |9|2|6|7|5|4|1|3|8|10| =
П(6,1,8)
α4 = (2,9)(4,6)(5,7)(1,8) = (n1 ~ n5);
orbits: (1,3,8,10)(2,9)(4,5,6,7)
5.3.2. Example with all cases (CS1, CS2, CS3, CS4) (Figure 16 )
The example is for graph G=B52 (Mathon [29]), regular graph, n=52, degree=25,
|Aut(G)|=12, orbit lengths: 2*2+4*6+2*12; orbits:
(4,30)(24,50) (21,37, 12,38,11,47) (17,36,16,42,10,43) (8,44,39,18,13,34)
(26,35,28,2,9,52) (29,25,22,15,20,33,7,46,41,48,3,51) (40, 6,49,45,5,1,27,31,19,23,32,14).
We show in Figure 16 only the subtrees of the selections (4,11), (4,13), (4,14) and (4,24) of
the search tree. The first selected vertex in П1=|1,2, . . . ,52| is the vertex 4 since it is from
one of the smallest orbits – the orbits are found by the heuristic algorithm. We start the
consideration from the selections (4,11,51) – this is the numeration n52: it is not equivalent
to any numeration from B(4,2), B(4,3), B(4,10). Before these selections there were
determined the bouquets of the representatives B(4,2), |B(4,2)|=9, B(4,3), |B(4,3)|=37,
B(4,10), |B(4,10)|=15, i.e., totally 51 nonequivalent numerations. These bouquets are
derivatives of FRPO of the set {2, 3, 7, 9, 10} – these vertices precede the selected vertex
11 in SC(П2), П2=П(4). There are no bouquets for the vertices 7 and 9 since they are
similar to previous vertices in SC(П2) under A(4): 7-3, 9-2. We have LP=LMIN=2 at the
selections (4,11,51). Before these selections there are found 4 generators of A(4), its order,
orbits and some stabilizers. At the selections (4,11,51) the case CS3 has been discovered.
That’s why the selected vertex 51 in SC(П3), П3=П(4, 11) becomes a new starting start
point: all information about the stabilizer A(4) obtained so far is lost and
A(4)=A(4,11)=A(4,11,51)={I}, LP=LMIN=3, XFLMIN=51, |A(XFLMIN)|=1. The next
selected vertex in SC(П3) is 7. The numeration П(4, 11, 7) is equivalent to numeration
n52.
Thus a new generator α5 for A(4,11) and A(4) is found (this is case CS2), α5=
(1,6)(2,52)(3,41)(4)(5,45)(7,51)(8,39)(9)(10,43)(11)(12,21)(13,34)(14,23)(15,29)(16)(17,3
6)(18)(19,31)(20,22)(24)(25,33)(26,28)(27,32)(30)(35)(37)(38,47)(40,49)(42)(44)(46,48)(5
0). Then, we compute
|A(XLMIN-1)|=|A(XFLMIN)||Orb(XFLMIN, A(XLMIN))|, i.e., |A(4,11)|=|Orb(51,
A(4,11))|.|A(4,11,51)|=2*1=2. The next selected vertex in SC(П3) is 9 - the partition
П4=П(4, 11,9) is not discrete: |A(4, 11,9)|= |A(4, 11)|/Orb(9, A(4, 11))|=2/|{9}|=2/1=2.
Then, the next selected vertex in SC(П4) is 46 and П5=П(4, 11,9,46) is discrete
(numeration n53). The numeration n53 is not equivalent to the numeration n52 and |A(4,
11,9,46)|= |A(4, 11,9)|/Orb(46, A(4, 11,9))|=2/|{46,48}|=2/2=1 – this is CS1. The vertex 48
is not selected in SC(П4) since it is similar to the vertex 46 under A(4, 11,9). Then, a
backward move to L=3 and a selection of the vertex 10 are made. The partition n54 is
discrete and not equivalent to any numeration in B(4,11), |A(4,11,10)|= |A(4,11)|/|Orb(10,
A(4,11))|=2/2=1 (the case is CS1). The vertex 43 in SC(П3) is not selected because it's
similar to the vertex 10 under A(4, 11). The next selected vertex in SC(П3) is 48- the
partition П4=П(4,11,48) is not discrete, so we do forward move to L=4 and choose vertex
27 in SC(П4). The numeration П5=П(4,11,48,27)=n55 is not equivalent to any numeration
in B(4,11): |A(4,11,48,27)|=|A(4,11)|/|Orb(48,A(4, 11))|/Orb(27, A(4,11,48))| =
2/|{46,48}|/|{27,32}| =2/2/2=0.5<1. This is case CS3. So, the vertex 27 in SC(П(4,11,48))
becomes a new starting start point: all information about the stabilizer A(4,11) is lost,
A(4)=A(4,11)=A(4,11,48)=A(4,11,48,27)={I}, LP=LMIN=4, XFLMIN=27, |A(XFLMIN)|=
- 31 -
L=1
L=2
6
2
4
L=3
2
3
...
3
...
1
4
...
5
L=4
n1
n8
n9
n35
7
52
9
10
...
7
n36
n50
4
CS3
1
51
α5
5
n51
CS2
7
6
2
9
6
CS1
n52
CS1
n53
CS3
n54
CS1
n55
CS1
n56
CS1
n57
CS1
n58
CS1
n59
CS1
n60
46
48
2
7
1
10
43
10
27
11
32
8
12
9
2
11
1
48
7
13
43
14
18
15
24
10
6
2
42
6
46
7
α6
CS2
48
2
11
2
18
5
CS1
n61
CS1
n62
CS1
n63
CS1
n64
CS1
n65
32
27
2
12
1
32
46
14
2
24
5
27
32
27
2
12
9
4
2
13
1
3
5
...
2
47
5
31
6
α7
CS4
14
25
46
7
2
10
2
5
CS1
n66
CS1
n67
CS1
n68
CS1
n69
CS1
n70
51
48
2
8
1
51
9
1
31
14
11
1
18
48
13
1
23
34
27
12
Figure 16. The search tree ST for the graph B52
- 32 -
L=1
1
2
3
...
4
L=3
...
6
2
L=2
3
4
...
5
L=4
n1
n8
n9
n35
7
52
9
10
...
11
...
7
8
13
10
1
... ...
9
n36
n50
n51
n63
n64
n67
n70
4
14
CS1
n71
CS1
n72
CS1
n73
CS3
n74
CS1
n75
CS1
n76
CS1
n77
CS1
n78
CS1
n79
CS1
n80
CS1
n81
CS1
n82
51
5
7
6
9
7
10
8
43
9
48
10
42
11
18
12
32
13
46
14
24
15
27
12
11
4
18
α8
CS2
3
...
12
12
4
α9
CS2
19
31
...
12
23
14
6
4
24
2
38
5
CS1
n83
CS1
n84
CS1
n85
10
42
...
2
5
2
43
47
25
5
32
14
2
10
11
9
32
...
12
Figure 16A. The search tree ST for the graph B52 (continued)
|A(4,11,48,27)|=1. We omit the description of the next selections in SC(П3)= SC(П(4,11)).
We only mention the occurrence of CS2: the generator 6=5 that leads to gen(A(4,11))=
{6}, |A(4,11)|=2 and the orbits of A(4,11) equal to the cycles of A(4,11). After the
SC(П(4,11)) has been traversed a backward move to L=2 follows: LP=LMIN=2,
B(4,11)={n52,n52,...,n64}, B(4)={n1,n2,...,n63}), XFLMIN=11, |A(XFLMIN)|= |A(4,11)|=2.
The next selected vertex in SC(П2) is 13: the partition П3=П(4,13) is not discrete, we do
forward move to L=3 and choose the vertex 3 in SC(П3). The numeration П4=П(4,13,3)
- 33 -
(n64) is not equivalent to any numeration in B(4). The order of its stabilizer is
|A(4,13,3)|=|A(4)|/|Orb(13, A(4))|/|Orb(3,A(4,13))| =2/|{13,34}|/|{3}|
=2/2/1=1 (CS1). After the selection 47 in SC(П3) and 31 in SC(П4) we obtain the
numeration n65=П5=П(4,13,47,31) that is not equivalent to n65 and
|A(4,13,47,31)|=|A(4)|/|Orb(13,A(4))|/|Orb(47,A(4,13))|/|Orb(31,A(4,13,47))|=2/|{13,34}|/|
{47}|/|{31}|=2/2/1/1=1 (CS1). After the selection 14 in SC(П4) we have the numeration
П(4,13,47,14), LP=4, LMIN=2. Numeration П(4,13,47,14) is equivalent to previous one
n65. Thus, a new generator α7 = (1,49)
(2,9)(3,46)(4)(5,40)(6,45)(7,41)(8,44)(10)(11,38)(12,37)(13)(14,31)(15,33)(16,17)(18,34)(
19,32)(20,29)(21)(22,25)(23,27)(24)(26)(28 35)(30)(36)(39)(42,43)(47)(48,51)( 50)(52)
for A(4) is found and the new orbits are
Orb(A(4))={1,5,6,40,45,49}{2,9,52}{3,7,41,46,48,51}{4}{8,39,44} {10,42,43}{11,38,47}
{12,21,37}{13,18,34}{14,19,23,27,31,32}{15,20,22,25,29,33}{16,17,36}{24}{26,28,35}
{30}{50} and the order of the stabilizer A(4) is |A(4)|=|A(4,11)|.|Orb(11,A(4))|=2.3=6
(XFLMIN=13). Then, a check for CS2/CS4 follows (LMIN=2<LP=4). We check for NRO
for each vertex ZLFRPO(XL), L=LMIN+1,...,LP=3,4, i.e., if there are changes of the
orders |A(4,13,3)| and |A(4,13,47,31)|. The order of |A(4,13,3)| before α7 is |A-α7(4,13,3)| =
|A-α7(4)|/|Orb(13,A-α7(4))|/|Orb(3,A-α7(4,13))|=2/2/1=1 and after α7 it is
|A(4,13,3)|=|A(4)|/|COrb(13,A(4))|/|COrb(3,A(4))|=6/|13,18,34|/|3,46,51,48|=|6/3/4=0.5.
This difference (|A-α7(4,13,3)| =1 |A(4,13,3)| = 0.5) shows that the orbit
Orb(3,A(4,13))={3,46,51,48} under A(4,13) is incorrect, it is united orbit, i.e., the case is
CS4. (As we’ll see later, the correct orbits are {3,46}{51,48}). Hence, the check for
|A(4,13,47,31)| is not necessary. As the case is CS4 we set LP=4 (not changed), LMIN=LP,
gen(A(XLMIN-1)=gen(A(4,13,47))={α7}; Orb(A(4,13,47))=cycles of α7 and| A(4,13,47))|=2
(the least multiple of the cycle lengths of α7). We also set XFLMIN=31 and we start the
selection of a new vertex from the current XLP=14 and since it is the last vertex in SC(П4)
we make a move back to the level L=3 selecting the vertex 46. We omit the following
actions of the algorithm. We only mention the last generators
α8=(1,5)(2)(3,51)(4)(6,40)(7,46)(8)(9,52)(10,42)(11,47)(12)(13,18)(14,32)(15,22)(16,36)(1
7)(19,23)(20,33)(21,37)(24)(25,29)(26,35)(27,31)(28)(30)(34)(38)(39,44)(41,48)(43)(45,4
9)(50) (14,27,19) (15,20,25)(16,17,36)(22,29,33)(23,31,32)(24)(26,35,28). The orbits due
to the generators α8 and α9 are:
Orb(A(4))={1,5,6,40,45,49}{2,9,52}{3,7,41,46,48,51}{4}{8,39,44}{10,42,43}{11,38,47}
{12,21,37}{13,18,34}{14,19,23,27,31,32}{15,20,22,25,29,33}{16,17,36}{24}{26,28,35}
{30}{50} and |A(4)|=|Orb(XFLMIN,A(4))|.|A(4,XFLMIN)|=6.1=6, where XFLMIN=14. The last
generator found by PART2 is
α10=(1,27)(2,28)(3,29)(4,30)(5,31)(6,32)(7,33)(8,34)(9,35)(10,36)(11,37)(12,38)(13,39)
(14,40)(15,41)(16,42)(17,43)(18,44)(19,45)(20,46)(21,47)(22,48)(23,49)(24,50)(25,51)
(26,52). The orbits of A due to the generators α8, α9 and α10 are given at the beginning of
this section and |A|=|A(4)||Orb(4,A)|=6*2=12. Thus, the output is: |A|=12, Orb(A) and
generators α8, α9 and α10.
- 34 -
L=1
36
L=2
37
L=3
LK=4
38
12
1
12
12
15
14
12
Figure 17. The search tree for A50 rigid graph
5.3.3. Example of a search tree for a rigid graph
The example is for graph A50, Mathon [29], n=50, regular bipartite graph,
k=105=15*7+35*3, Error! Reference source not found..). The bouquet |B(1)|=168,
|SC(П2)|=14, each selected cell SC(П3) has size |SC(П2)|=12. The search tree is full.
5.4.
Correctness and analysis of the algorithm
We do not prove formally the correctness of the algorithm but it follows from the
description of the algorithm. If the bouquets of each representative of an orbit in SC(ПLP)
are correct and the whole orbit of each selected vertex in SC(ПLP) is traversed then
according to Theorem 4 GOO(A(XLP-1)) and B(A(XLP-1)) will be determined correctly. The
main problem is to guarantee the correctness of the bouquets - this is shown in the
description of the cases CS1, CS2, CS3 and CS4. The traversal of the whole orbit of each
selected vertex in SC(ПLP) is also guaranteed (Figure 9). The formal correctness prove of
the algorithms and their analysis will be presented in a new article..
6. Vsep-is algorithm for graph isomorphism(GI algorithm)
Vsep-is algorithm finds an isomorphism between two graphs G1(n1,k1), G2(n2,k2), if any.
It is based on Corollary 3 of theorem 3 and theorem 4. Firstly, it checks for some necessary
conditions for isomorphism: a) n1=n2?; b) k1=k2? c) equality of sorted degree sequences
of vertices of G1,G2;c) compatibility of π1=RP(G1,πu); π2=RP(G2,πu); d) equality of
indices of the target cells of π1and π2. If one of these conditions does not hold the graphs
are not isomorphic. Otherwise, the bouquet B(x1, G1) is determined by applying PART1 to
the first vertex x1SC(π1).
Then, by applying PART2 for each vertex ySC(π2,G2) is generated one numeration that
is compared with the numerations of the bouquet B(x1,G1) for an isomorphism, If there is
an isomorphism for the current numeration then a message ‘G1G2’, printing the
isomorphism and stop follows. If all numerations (instruction 6) derived from each vertex
- 35 -
ySC(π2,G2) do not form an isomorphism with the bouquet B(x1,G1) then a message that
the graphs are not isomorphic follows.
Input: Graphs G1(n1,k1), G2(n2,k2),
Output: G1 and G2 are isomorphic or no; print the isomorphism if any
1: if (n1≠n2 or k1≠k2 or ordered sequences of vertex degrees of G1,G2 are not equal)
then output ‘G1!G2’;return end if
2: Define π1=RP(G1,πu); π2=RP(G2,πu); if π1,π2 are not compatible then message
‘G1!G2’; return end if
3. Define SC(π1), SC(π2); if(index(SC(π1)) ≠ index(SC(π2)))then output ‘G1!G2’;
return end if
4. Apply PART1 to the first vertex x1SC(π1) to determine the bouquet B(x1,G1)
5. for each vertex ySC(π2,G2) do
determine one numeration num by applying SFM(y);
compare num with the bouquet B(x1,G1) for isomorphism;
if there is an isomorphism then output ‘G1G2’ and the isomorphism; return end if
end for
6. output ‘G1!G2’
Figure 18. Vsep-is algorithm
7. The heuristic algorithms (Vsep-orb, Vsep-hway, Vsep-sch)
They are based on Theorem 3. For determining whether two vertices x and y are similar
two partial bouquets are built for both vertices and then, some automorphisms between the
numerations of these bouquets are determined. To determine certainly that x and y are
similar one of the bouquets should be full. Consequently, the probability to find at least one
automorphism mapping x to y is less than 1 if we use the algorithm with partial bouquets.
This makes the algorithms inexact. Instead, less time is needed for bouquets building and
less storage is needed for them because of their smaller sizes. We describe two heuristic
algorithms Vsep-orb, and Vsep-hway in this section.
7.1.
TABLE1 and TREE4 procedures
These procedures are needed for the following description of the heuristic algorithms.
TABLE1 heuristic procedure generates in SC(П) numerations by the operations fork and
RRST. Each new numeration is compared with previously found numerations (stored in
HT) for an automorphism. If there is an automorphisn it is added to previously found set of
generators. If there is no an automorphism the numeration is put into HT. RRST tree is
defined by the parameters W (width) and D (depth): for each level L, 1< L ≤ 1+D, in
SC(ПL) the number of selected vertices is W. The experiments show that when the selected
vertices are evenly distributed in SC(ПL) of RRST the results are more correct compared
- 36 -
Input: G- a graph, П – a stable partition, W, D – a width and a depth of the reduced
regular search tree RRST;
Output: Generators and orbits of A=А(G, П) stabilizer, norb - a number of orbits;
Local variables: HT- a hash table of partial bouquets of generated numerations
T1: Initialization: Set Orb(A(G, П) to discrete orbits; gen( A):=; HT empty;
T2: Find SC(П) target cell;
T3: Select each vertex xSC(П) that is not similar to a previous vertex in SC(LB)
under current orbits. Generate a numeration n1 for x vertex by a series of forward
move and compare it for an automorphism with the numerations in HT .If there is
then: gen(A) :=gen(A) {} and determine new Orb(A(G, П) and norb. If there is no
then add n1 into HT. We call this operation a fork.
T4: A reduced regular search tree (RRST) is built for each representative of the
orbits in SC(LB) determined so far. The orbits are sorted in increasing order of their
lengths starting from the smallest one. Each representative of an orbit is a root of a
tree. The depth of RRST is D. Each node of the RRST corresponds to a selection level
L, 1 < L ≤ 1+D and a target cell SC(L). In each SC(ПL) a fixed number W of vertices
y are selected, then a forward move to next level L is made for each y. In the 1+D
level only a series of forward moves SFM is made for each selected vertex until a
discrete partition (numeration n1) is obtained. The number W may be regarded as outdegree of the RRST node. The vertices in SC (ПL) are selected successively by step
st= |SC(ПL)| / W ≥ 1 starting from the beginning of SC ПL), i.e. they are evenly
distributed in SC(ПL). To determine an automorphism and new orbits each
numeration n1 is compared with the numerations in HT table. If there is then:
gen(A) :=gen(A) {}, new Orb(A(G, П) and norb are .determined. If there is no
then add n1 into HT. The procedure stops when RRSTs are generated for all
representatives of the orbits in SC(ПL).
Figure 19. TABLE1 procedure
Input: graph G, equitable partition П, parameters mbrsyvp,wth1, b1a, dpth1, br2a
Output: GO(A(П)), norb – number of orbits of A(П)
1: DO wth=wth1, wth1+br1a
2: DO dpth=dpth1, dpth1+br2a
3: TABLE1(П, WTH, DPTH, norb)
4: IF (norb=1 or SC(П1) contains one orbit or norb remains unchanged after
mbrsyvp number of successive calls to TABLE1) then
5:
exit
6: end if
7: END DO
8:END DO
Figure 20. TREE4 procedure
- 37 -
with other ways of distribution. One call to TABLE1 procedure with fixed parameters W
and D usually is not sufficient to determine а correct GO(A). That’s why several calls to
TABLE1 with different parameters (W, D) are needed – this is done by TREE4 procedure
(Figure 20). The parameters of TREE4 have the following meaning: mbrsyvp – maximal
number of successive calls to TABLE1 with no changed norb, wth1 – the start value of W,
wth1+br1a - the last value of W, dpth1 – the start value of D, dpth1+br2a - the last value of
D. The parameters have default values but there is a possibility the user to change them.
7.2.
Vsep-orb algorithm
It has two main steps S1 and S2:
S1) Determine the orbits Orb(G, Π) by an heuristic TREE4 algorithm (Figure 20);
S2) Use PART1, PART2 algorithms to determine GOO(G, ΠORB), where ΠORB is a
partition which cells to the orbits found in S1 (orbit partition).
TREE4 does not guarantee the computation of the exact orbits of GOO(G, Π) and
consequently Vsep-orb is an heuristic algorithm - it does not guarantee the computation of
the exact GOO(G, Π). But if the parameters of Vsep-orb are selected in proper way the
probability of the exact computation of GOO(G, Π) is very close to 1. The experiments
confirm this.
7.3.
Vsep-sch algorithm
Vsep-sch algorithm has two main steps:
S1) Determine the base points, generators and orbits of Aut(G, Π) by TABLE1 heuristic
algorithm (fig. 18) applied with a parameter LB = 1;
S2) Determine the order |Aut(G, Π)| using Schreier-Sims algorithm [5,31,32] on the base
points and generators found in S1. Obviously, at this step the strong generators (see the
definition below) of Aut(G,Π) are determined.
Recall [32] that a base for a group A is a sequence of points B=[1,2,…, m] such that the
stabilizer 𝐴𝛽1,𝛽2 ,…,𝛽𝑚 = 𝐼. Schreier-Sims algorithm determines a chain of stabilizers
A=A(1) ≥ A(2) ≥ . . . ≥ A(m) ≥ A(m+1) = I, where 𝐴(𝑗) = 𝐴𝛽1 ,𝛽2,…,𝛽𝑖−1 .
Each stabilizer A(j) is represented by a series of all right cosets to A(j+1)
A(j) = A(j+1) \ A(j) = A(j+1) ∪ A(j+1) α2 ∪ A(j+1) α3 ∪ . . . ∪ A(j+1) αn
The collection of elements in A(j+1) \ A(j) , j=1,2, . . . ,m are called strong generators for A
toward B. The algorithm [33] is constructing a table T(n,m) of strong generators, where
T(i,j) is an element of the group A which fixes 1, 2 ,…,j-1 and maps j to i. If no such
element exists, then the entry is empty. The non-empty elements of the j-th column of T
form a set of coset representatives U(j) of A(j+1) in A(j).
The step S2 algorithm (fig. 21) uses a procedure called sift (fig. 22) that for a given
automorpism α compares (j) with y=α((j)), j=1,2,…,m: (a) if y= (j) then a pass to the
next base point (j+1) follows; (b ) if y ≠ (j) then a check if α belongs to the stabilizer
A(j) follows, i.e. if T(y,j) is empty or no. If T(y,j) is empty then T(y,j) = α, the number of
nonempty elements in column j of T is increased by 1, c(j)=c(j)+1, and the sift ends. If
- 38 -
T(y,j)=γ then a new automorphism α = γα-1 is determined and all above actions are applied
to α for the next base points (j+1). The procedure sift modifies the table T by inserting at
most one new coset representative.
Step S2 algorithm is with base points - that is the difference from FHL version of SchreierSims method in [33]. When the first 3 instructions of S2 algorithm are finished, the table T
should have the property that α ε A iff α can be expressed as a 1a2 . . . am , where ai is a
member of the j-th column of T. This is called canonical representation of α toward B. The
time complexity of step S2 is O(n6) [33,34]
Input: n, generators, B - base points, m==|B|
Local variables: T(n, m)-2 dimensional integer array, c – integer array of size m, c(j) is
the number of nonempty elements in column j of T; Output: Aut(G, Π)
1; Set each element of c to 1;Set each element T(i,j) empty;
2: sift(n, α,B,m,T,c,) {Pass each generator to the procedure sift};
3: sift(n,γ,B,m,T,c,) {Pass the product γ of each pair of representatives in T to the
procedure sift until no new element is inserted in T};
4: Determine the order |Aut(G, Π)|=∏𝑚
𝑗=1 𝑐(𝑗).
Figure 21. S2 step of Vsep-sch algorithm
Input: n,α, B, m, T, c; c-array of size m=|B|,c(j) is the number of nonempty elements
in column j of T;Output: T,c
1: for j:=1 to m do
2: y:= α(B(j))
3: if(y=B(j)) cycle
4: if (T(y,j) is empty) then
5: T(y,j) := α; c(j) := c(j)+1; return
6: else
7: γ := T(y,j); α := γα-1
8: end if
9: end for
Figure 22. sift procedure
8. Vsep-a automatic version of Vsep algorithm
Experiments show that Vsep-sch algorithm runs long for graphs with large |Aut(G)| (small
number of orbits), especially for transitive graphs. In these cases T table is large and full or
almost full and evidently this causes more computation. On the other hand, Vsep-orb runs
fast on such graphs. This is the reason for developing Vsep-a algorithm that chooses Vseporb when the number of orbits is small (≤ 3) and Vsep-sch otherwise. There are also other
criteria for this selection. Vsep-a algorithm also chooses Vsep-e algorithm for many graphs
on the base of some criteria, for example for some non-regular graphs.
- 39 -
9. Program implementation of Vsep algorithms
All proposed algorithms in this paper are implemented in Fortran programs that can be
compiled by any Compaq Fortran compiler. The program version published on author’s
web site section My Programs [35] is called VSEP_PUB4. It implements Vsep-e, Vseporb, Vsep-sch and Vsep-a version of VSEP algorithms. The user may choose any version
and set different parameters (see user guide).
10.Experimental results
In this section we present: a) experiments on almost all benchmark graphs from [19] that
compare the performance of VSEP_PUB4 program with Traces (nauty2.6r5 [19]) (Table
4) –one of the most competitive known tool for the worst cases; b) The running times of
VSEP_PUB4 for whole families of some benchmark graphs are given in Table 5.
Some of the experimental results from VSEP _PUB4 (as they are received from the
program) are in RESVSE4P_PUB4OBSHT file given in [35]. The most difficult graphs for
our algorithms are the graphs with |Aut(G)|=1 or with small |Aut(G)|. It is known that none
of the known algorithms outperform others for all graphs. For each algorithm there are
specific difficult and easy graph families. A graph family may be easy for one algorithm
and very difficult for another. The same is for Vsep and Traces. Even more, for Vsep
different cell selectors give different running times – none of the cell selectors outperform
others. The chooser of cell selectors does not always choose the optimal cell selector. For
each result we show the cell selector for which it is obtained. The experiments were carried
out on a laptop Dell, CPU: Intel(R) Core (TM) i5-3317U@ 1.7 GHz, Memory: 8 GB, OS:
64 bit Microsoft Windows 7 Profesional. For the experiments, we have used all the
benchmark graphs of nauty&Traces page [19] that includes a variety of graph families
with different characteristics. We show mostly the results for the graphs that are worst
cases for either of the compared tools and cell selectors for which they are obtained (only
for Vsep). It is evident (See Table 4) that Vsep outperforms Traces considerably for the
graphs tnn(39)_1014-1, chh_cc(7-7)_1078-1, f-lex-srg-10-1 and f-lex-srg-50-1 . On the
other side Traces outperforms Vsep-e considerably for the graphs of projective planes (pp16-14, 15, 22, pp-25-90, 116) , had-112, had-176 and latin-sw-112 but for the same
graphs Vsep-a outperforms Traces considerably. There are no essential differences
between Vsep and Traces on the other graphs on the Table 4! Traces is slow for the graphs
with large order of the automorphism group and large number of generators – my
experience show that this is may be due to the use of Schreier-Sims method. The main
disadvantage of Vsep-e is the storing of the whole bouquet of the first selected vertexmillions of words for some graphs. For almost all benchmark graph families the heuristic
versions are many times faster than the exact one and at correctly chosen parameters give
correct results.
- 40 -
G graph, n*, minval*, maxval*,
|Aut(G)|, norb*
Traces
T in seconds
vsep-e*
izb*,
bouquet
size*.par*
0.08
1.01, 1,1,nop
0.83, 4,1,nop
0.14, 1, 1,nop
vsep-a*
selected
version(orb for
vsep-orb,shc for
vsep-sch,e-for
vsep-e),izb*,par*
1.03,e,1,nop
0.80,e,4,nop
0.13,e,1,nop
0.11
0.14, 1,21,nop
0.14,e,1,nop
0.14
NA
1.17, 1,4,nop
1.06,4,4,nop
0.17, 1,1,nop
1.17,e,1,nop
1.07,e,1,nop
0.17,e,1,nop
0.02
0.11,3,def
0.11,e,1,def
0.05
85.75,1,32287,de 0.42.e,1,def
f
429.75,1,21210,d 1.88,e,1,def
ef
0.61,1,1,def
0.62,e,1,def
tnn(39)_1014-1, 1014,
4,312, 6.314790834154e174,12
cmz-50, 1200,2,5,
1.267650600228e32,8
ag-49,4851,49,50,
2.710632960000e10,2
cfi-200,2000,
3,3, 2.535301200456e30, 800
chh_cc(7-7)_1078-1
1078,3,45,4.907372642035e71,8
had-128,512,129,129,
1.073807313661e19,1
had-112,448,113,113, 1677312,1
5654.38
had-176,704,177,177. 15257088,1
0.14
had-256,1024,257,257,
1.401962828716e24,1
latin-30.dre,900 ,87,87, 43200.1
0.05
latin-sw-30-1,900 ,87,87, 1,900
0.17
had-sw-112,448,113,113, 2,224
1.61
lattice-30.dre, 900 , 58,58,
1.407181592771e65,1
10cube,1024,10,10, 3.715891200e9,1
0.03
1.33.e,1,def
0.05
0.71,1,1106,def
0.66,4,962,def
4.33, 1,812;
1,2,1,1,1
179.7,1,
519624;
1,2,1,1,1
0.09, 1 ,1,def
0.02
0.11, 1,1,def
0.11,orb, 1,def
paley-461.dre ,461,230,230,
1.06030e5,1
0.02
0.14, 1,1,def
0.15,orb, 1,def
pg2-49, 4902,50,50,
1.328752276992e14,1
pp-16-14,546,17,17, 2304,14
0.11
2.00, 1,21,def
2.00,orb,1,def
0.66
55.24, 1, 420960; 0.44,sch,1;
- 41 -
3.16,e,1;
1,2,1,1,1
16.69,sch,1,def
0.61,sch,
1,2,1,1,1
0.12,orb,1,def
G graph, n*, minval*, maxval*,
|Aut(G)|, norb*
pp-16-22,546,17,17, 9216,10
Traces
0.55
pp-25-1,1302,26,26, 609336000000,2 0.05
T in seconds
vsep-e*
izb*,
bouquet
size*.par*
3,3,2,1,1
8.50, 2, 52680;
3,3,2,1,1
15.4, 1,117264;
3,3,2,1,1
2.26,2,13336;
3,3,2,1,1
0.12,2,10;
3,3,2,1,1
242.6, 2,
522995,
3,3,2,1,1;
2244.03,1,
5016276,def
500.0, 2 ,
1058568,
3,3,2,1,1
40.0,,2, 73008
35211
5631.0, 2,
1975680;
2,5,2,2,1
5600.0, 4,
1975680
2,5,2,2,1
5529.05, 2,
1975680
2,5,2,2,1
vsep-a*
selected
version(orb for
vsep-orb,shc for
vsep-sch,e-for
vsep-e),izb*,par*
3,3,2,1,1
0.26,sch,1;
3,3,2,1,1
0.12,orb,1,
3,3,2,1,1
9.7,sch,1;
3,3,2,1,1
pp-25-90,1302,26,26,1000,40
42.62
pp-25-116,1302,26,26, 500,64
104.82
pp-27-10, 1514,28, 28, 122472,6
1.05
pp-49-1,4902,50,50,288120,10
236.87
mz-aug-50,1000,3,6,
101412048018258352119736256430
08,250
mz-aug2-50,1200,2,5,
507060240091291760598681282150
4,700
mz-50,1000,3,3,
0.00
0.22,1,1,def
0.22, orb,1,def
0.00
0.54,1,1,def
0.53,orb, 1,def
0.00
0.22,1,1,def
0.20,orb, 1,def
- 42 -
6.00,sch,1,
3,3,2,1,1
4.59,sch,2;
35211
117.4,1;
2,5,2,2,1;
90.1,2;
2,4,2,2,1
G graph, n*, minval*, maxval*,
|Aut(G)|, norb*
101412048018258352119736256430
08,250
f-lex-reg-50-5,1700,
8,30,16384,1016
f-lex-srg-10-1,790,16,75,16,,439
f-lex-srg-50-5,3950,
16,75,32768,2360
Traces
T in seconds
vsep-e*
izb*,
bouquet
size*.par*
vsep-a*
selected
version(orb for
vsep-orb,shc for
vsep-sch,e-for
vsep-e),izb*,par*
0.06
0.42,1,2,def
0.42,orb,1,def
NA
NA
NA
NA
0.17,sch,2;
22122
10.69,sch,2;
22122
Table 4. Experimental comparison of VSEP4P_PUB4 program with Traces. n - number
of vertices, norb-number of orbits, vsep-e - exact version, vsep-a – automatic version, izb the numbering of the chosen cell selector; minval, maxval - minimal and maximal degree
of a vertex of the graph, bouquet size – number of the stored numerations; PAR –
parameters mbrsyvp, wth1, dpth1, br1a, br2a; 4,5,2 ,1,1 are default parameters
(def=default),nop-no parameters since there is no call to tree4;NA-not attended
Graph family
Parameters: mbrsyvp, wth1, dpth1, br1a, br2a; 4,5,2
,1,1-default (def)
izb T[secs]-running
time
mathon
mathon
dobling
tnn
tnn
cmz
ag
chh
cfi
cfi
latin
lattice
hypercubes3
paley
ppsmall
pp16
pp16
pp16
pp25
pp27
pp27
Def
Def
1
1
0.1248
0.312
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
Def
33241
33221
33211
33211
Def
35211
1
4
1
1
1
1
4
1
1
1
1
1
1
1
1
1
2
2
3.97
3.31
2.37
0.71
2.19
38.94
36.77
3.79
0.99
0.25
1.90
0.56
7.28
7.22
7.19
746.11
60.66
57.60
- 43 -
Graph family
Parameters: mbrsyvp, wth1, dpth1, br1a, br2a; 4,5,2
,1,1-default (def)
pp27
35211
pp27
34342
pp27
34342
pp27
24342
pp27
34332
pp27
34341
pp27
24332
pp27
24331
pp27
24341
pp27
24321
pp27
24311
pp49
35221
pp49
25221
pp49
25221
had
45211
had
45211
had
36211
f-lex=srg
22122
f-lex=reg
Def
pg
Def
pg
34211
pg
33211
mz-aug
Def
mz-aug2
Def
total run time for all cases in the table
izb T[secs]-running
time
3
1
2
2
2
2
2
2
2
2
2
1
2
1
1
3
3
2
1
1
1
1
1
1
58.26
68.53
88.54
70.76
88.90
60.31
76.32
68.62
75.48
69.28
68.82
505.00
349.60
365.10
293.28
278.40
281.36
589.29
18.47
7.84
7.80
7.80
2.09
3.77
4478.15
Table 5. Experimental results from VSEP4P_PUB4 program for whole families of the
most of the benchmark graphs in [19], some of the results are in RESVSEP4OBSHT file
[35], izb is the numbering of the chosen cell selector
11. Concluding remarks and open problems
Five new algorithms, named Vsep, are described. Four of them are for determining the
generators, orbits and order of an undirected graph automorphism group:. Vsep-e – exact,
Vsep-orb and Vsep-sch – heuristic and Vsep-a automatically selects the optimal version
among Vsep-e, Vsep-orb and Vsep-sch. The fifth algorithm, Vsep-is, is for finding an
isomorphism between two graphs. A new approach is used in the exact algorithm: if during
its execution some of the searched or intermediate variables obtain a wrong value then the
algorithm continues from a new start point losing some of the results determined so far
(cases CS3, CS4). The new start point is such that the correct results are obtained.
A new code, named S-code, of a partition of the graph vertices is proposed. S-code is used
for reducing the time of comparing the partitions in the proposed algorithms.
The experiments show that the worst case time complexity of Vsep-e for an arbitrary graph
is exponential but for some classes it is polynomial. The main difference of the exact one
and the widely known tools is the storing of the whole bouquet of numerations of the first
- 44 -
selected vertex in the first level. Five cell selectors are used in the algorithms and a chooser
of optimal cell selector is presented. Some of the cell selectors are new, namely
mxvectchval and mxprchval.
Experimental comparison of the proposed algorithms with Traces algorithm is made - it
shows their worst and best cases. A disadvantage of Vsep-e algorithm is its higher
requirements for memory (for some worst cases several millions of numbers are stored.
The worst cases for the algorithm Vsep-e are the graphs with smaller order |Aut(G)|,
especially the rigid graphs. Vsep-orb and Vsep-sch heuristic algorithms are extremely fast
(with some exceptions) compared with the exact one and are almost exact - for all tested
thousands of graphs they give correct results. Practically, their requirements for memory
are very small. Heuristic algorithms are an important approach for solving so hard problem
as graph isomorphism is.
The future work on developing Vsep algorithms will include: a) search for a new cell
selector that will reduce the size of the search tree; b) search for a new chooser of a cell
selector; c) conduct a comparison of Vsep algorithms with other known GOO algorithms;
d) develop parallel versions of Vsep algorithms. Open problems are: i) Develop a new
chooser of cell selector that given a graph can determine the best cell selector (that yields a
minimal running time of the algorithm); ii) Develop an algorithm that given a graph and
the first numeration can determine the parameters mbrsyvp, wth1, dpth1, br1a, br2a of the
forest of reduced regular trees such that when used the heuristic versions of VSEP
algorithms will give correct results for a minimal running time; iii) Develop an algorithm
that given a graph on the base of some criteria to choose the optimal algorithm from the
known published algorithms.
Acknowledgements
The author thanks to Brendan McKay, Bill Kocay, Rudi Mathon, Vladimir Tonchev,
Ulrich Dempwolff, Gordon Royle, Adolfo Piperno, Jose Luis López-Presa, Petteri Kaski
and Hadi Katebi for the extensive discussions on the graph isomorphism problem and
exchanging graphs with interesting properties and papers on this problem. Special thanks
to my former diploma student Apostol Garchev who contributed in performing and
repeating some of the experiments on the developed algorithms. The author would like to
thank all colleagues and students who contributed to this study.
References
[1]
[2]
[3]
[4]
A. V Aho, J.E. Hopcroft, J.D. Ullman, The Design and Analysis of Computer
Algorithms, 1974.
C.C.E. Leiserson, R.R.L. Rivest, C. Stein, T.H. Cormen, Introduction to Algorithms,
Third Edition, 2009. doi:10.2307/2583667.
E. by J. van Leeuwen, Handbook of Theoretical Computer Science, Volume B
Formal Models and Semantics, MIT Press, 1994.
https://mitpress.mit.edu/books/handbook-theoretical-computer-science-0.
edited by M.I. Nechepurenko, Algorithms and Programs for Solving Problems on
- 45 -
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
Graphs and Networks, Nauka, Novosibirsk, 1990.
http://mt3.org/bloog/wodolaz/39.html.
K. D. L., D. R.Stinson, Combinatorial Algorithms, CRS Press (1999)., CRS Press,
1999.
F. Harary, Graph theory 1736–1936, Hist. Math. 4 (1977) 480–481.
doi:10.1016/0315-0860(77)90099-4.
F. Harary, E.M. Palmer, Graphical Enumeration, 1973.
http://oai.dtic.mil/oai/oai?verb=getRecord&metadataPrefix=html&identifie
r=AD0765600.
R.B. Ash, Abstract Algebra : The Basic Graduate Year, Nature. 169 (1952) 167–168.
doi:10.1038/169167b0.
J.L. Gross, J. Yellen, Handbook of graph theory, 2004. doi:10.1016/03043878(86)90037-4.
C.M. Hoffmann, Group-Theoretic Algorithms and Graph Isomorphism, Lecture No,
Springer-Verlag ., 1982. http://link.springer.com/book/10.1007/3-540-11493-9.
P.J. Cameron, Automorphisms of graphs, J. Math. Sci. 72 (1994) 3341–3343.
doi:10.1007/BF01261692.
B.D. McKay, Practical Graph Isomorphism, Congr. Numer. 30 (1981) 45–87.
http://users.cecs.anu.edu.au/~bdm/papers/pgi.pdf.
B.D. McKay, A. Piperno, Practical graph isomorphism, II, J. Symb. Comput. 60
(2014) 94–112. doi:10.1016/j.jsc.2013.09.003.
L. Babai, Moderately Exponential Bound for Graph Isomorphism, Proceeding FCT
’81 Proc. 1981 Int. FCT-Conference Fundam. Comput. Theory. (1981) 34–50.
http://dl.acm.org/citation.cfm?id=739270.
R. Mathon, A note on the graph isomorphism counting problem, Inf. Process. Lett. 8
(1979) 131–136.
http://www.sciencedirect.com/science/article/pii/0020019079900048.
K.S.B. and C.J. Colbourn, Problems Polynomially Equivalent to Graph Isomorphism,
1979. https://cs.uwaterloo.ca/research/tr/1977/CS-77-04.pdf.
V.N. Zemljachenko, N.M. Korneenko, R.I. Tishkevich, Graph isomorphism problem,
(1982) 83–158. doi:1007/BF02104746.
L. Babai, Graph Isomorphism in Quasipolynomial Time, arXiv. (2016) 89.
http://arxiv.org/abs/1512.03547.
B.D. McKay, A. Piperno, Nauty&Traces website, (n.d.). http://pallini.di.uniroma1.it/.
D.G. Corneil, C.C. Gotlieb, An Efficient Algorithm for Graph Isomorphism, J. ACM.
17 (1970) 51–64. doi:10.1145/321556.321562.
A.V.P. Faizullin,R.T., Heuristic algorithm for solving the graph isomorphism prob
lem, 2002. http://arxiv.org/pdf/math/0205220.pdf.
G.-H.A.W.-G.T. KING, A New Graph Invariant for Graph Isomorphism:Probability
Propagation Matrix, J. Inf. Sci. Eng. 15 (1999) 337–352.
http://www.iis.sinica.edu.tw/page/jise/1999/199905_01.pdf.
H. Ogata, W. Fujibuchi, S. Goto, M. Kanehisa, A heuristic graph comparison
algorithm and its application to detect functionally related enzyme clusters., Nucleic
Acids Res. 28 (2000) 4021–4028. doi:10.1093/nar/28.20.4021.
- 46 -
[24] H.B. Mittal, A Fast Backtrack Algorithm for Graph Isomorphism, Inf. Process. Lett.
29 (1988) 105–110. doi:10.1016/0020-0190(88)90037-3.
[25] S.D. Stoichev, Fast Adjacency Refinement Algorithm, in: Intern. Conf. Digit. Signal
Process., Limassol, Cyprus, 1995: pp. 870–875.
[26] J. Hopcroft, An n log n algorithm for minimizing states in a finite automaton,
Reproduction. (1971) 189–196.
http://oai.dtic.mil/oai/oai?verb=getRecord&metadataPrefix=html&identifier=AD071
9398.
[27] A. Piperno, Search Space Contraction in Canonical Labeling of Graphs (Preliminary
Version), arXiv. (2008) 1–21. http://arxiv.org/abs/0804.4881.
[28] D.G. Corneil, C.C. Gotlieb, An Efficient Algorithm for Graph Isomorphism, J. ACM.
17 (1970) 51–64. doi:10.1145/321556.321562.
[29] R. Mathon, Sample graphs for isomorphism testing, in: Proc. 9th S.-E.conf. Comb.
Graph Theory Comput. Boca - Raton, 1978: pp. 499–517.
[30] S.D. Stoichev, Vsep-New Heuristic and Exact Algorithms for Graph Automorphism
Group Computation, (n.d.). arXiv:1007.1726v4 [cs.DS].
[31] M. Jaggi, Implementations of 3 Types of the Schreier-Sims Algorithm, (2005).
http://www.m8j.net/data/List/Files-118/Documentation.pdf.
[32] E.A. O’Brien., Toward effective algorithms for linear groups, pp. 163–190, in: Finite
Geom. Groups, Comput. , Gruyter, Berlin (2006), Berlin, 2006: pp. 163–190.
https://www.math.auckland.ac.nz/~obrien/research/survey.pd.
[33] G.. Butler, Chapter 14. Complexity of the Schreier-Sims method, in: Lect. Notes
Comput. Sci. Vol. 559 1991 Fundam. Algorithms Permut. Groups, Springer, 1991.
http://link.springer.com/chapter/10.1007/3-540-54955-2_33#page-1.
[34] M. Furst, J. Hopcroft, E. Luks, Polynomial-time algorithms for permutation groups,
in: Found. Comput. Sci. …, 1980: pp. 36–41.
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4567802.
[35] S.D. Stoichev, S.D. Stoichev’s web page, (n.d.).
https://sites.google.com/site/stoichostoichev2/my-programs.
- 47 -
| 8cs.DS
|
THOMPSON’S GROUP F IS NOT LIOUVILLE
VADIM A. KAIMANOVICH
arXiv:1602.02971v3 [math.GR] 22 Mar 2016
Dedicated to Wolfgang Woess on the occasion of his 60 th birthday
Abstract. We prove that random walks on Thompson’s group F
driven by strictly non-degenerate finitely supported probability
measures µ have a non-trivial Poisson boundary. The proof consists in an explicit construction of two non-trivial µ-boundaries.
Both of them are described in terms of the “canonical” Schreier
graph Γ on the dyadic-rational orbit of the canonical action of F
on the unit interval (in fact, we consider a natural embedding of
F into the group P LF (R) of piecewise linear homeomorphisms
of the real line, and realize Γ on the dyadic-rational orbit in R).
However, the definitions of these µ-boundaries are quite different
(in perfect keeping with the ambivalence concerning amenability
of the group F ). The first µ-boundary is similar to the boundaries
of the lamplighter groups: it consists of Z-valued configurations
on Γ arising from the stabilization of logarithmic increments of
slopes along the sample paths of the random walk. The second
µ-boundary is more similar to the boundaries of the groups with
hyperbolic properties as it consists of sections (“end fields”) of the
end bundle of the graph Γ, i.e., of the collections of the limit ends
of the induced random walk on Γ parameterized by all possible
starting points. The latter construction is more general than the
former one, and is actually applicable to any group which has a
transient Schreier graph with a non-trivial space of ends.
Introduction
A. General setup. The group F introduced by Richard Thompson in 1965 is the group of the orientation preserving piecewise linear
dyadic self-homeomorphisms of the closed unit interval [0, 1]. Arguably,
the most important open question about Thompson’s group F is the
one about its amenability (see Cannon – Floyd [CF11] for its history).
Indeed, due to the plethora of rather unusual properties of this group
(described, for instance, by Cannon – Floyd – Parry in [CFP96]) either answer would imply very interesting consequences. This problem
The author was partially supported by funding from the Canada Research Chairs
program, from NSERC (Canada) and from the European Research Council within
European Union Seventh Framework Programme (FP7/2007-2013), ERC grant
agreement 257110-RAWG. A part of this work was conducted during the trimester
“Random Walks and Asymptotic Geometry of Groups” in 2014 at the Institut Henri
Poincaré, Paris.
1
2
VADIM A. KAIMANOVICH
has recently attracted a lot of attention, with an impressive number of
failed attempts to prove either amenability or non-amenability of the
group F .
We are inclined to believe that Thompson’s group F is non-amenable,
and several recent papers provide circumstantial evidence for that.
There are various computational experiments due to Burillo – Cleary –
Wiest [BCW07], Haagerup – Haagerup – Ramirez Solano [HHRS15],
Elder – Rechnitzer – van Rensburg [ERJvR15]. Besides that, Moore
[Moo13] proved that if the group F is amenable, then its Følner function must be growing very fast (note, however, that, as it was shown
by Erschler [Ers06b], even the groups of intermediate growth may have
the Følner function growing faster than any given function).
The solution of a conjecture of Furstenberg [Fur73] by Rosenblatt
[Ros81] and by Kaimanovich – Vershik [KV83] provides the following characterization of amenability in terms of the Liouville property.
A countable group is amenable if and only if it carries a non-degenerate
Liouville random walk (i.e., such that its Poisson boundary is trivial).
Therefore, a possible approach to proving non-amenability of a given
group consists in showing that there are no non-degenerate Liouville
random walks on it. Here we make the first step in this direction for
Thompson’s group F by showing that the random walk (F, µ) is nonLiouville for any strictly non-degenerate finitely supported measure µ.
In particular, the simple random walk on the Cayley graph of F determined by any finite generating set is non-Liouville.
Independently of the amenability issue, Thompson’s group F was
among the first examples of finitely generated groups with an intermediate (between 0 and 1) value of the Hilbert space compression which
were exhibited by Arzhantseva – Guba – Sapir [AGS06]. They proved
that for the group F this value is equal to 1/2 and asked about existence of a compression function strictly better than the square root
[AGS06, Question 1.4]. However, as it has been pointed out by Gournay [Gou14], in view of the results of Naor – Peres [NP08], if such a
function exists, then the group must be Liouville. In combination with
our result, it allowed Gournay to answer the above question of Arzhantseva – Guba – Sapir in the negative: the best Hilbertian equivariant
compression function for Thompson’s group F is (up to constants) the
square root function.
B. Main results. The overwhelming majority of the currently
known examples of an explicit non-trivial behavior at infinity for random walks on discrete groups falls into one of the following two classes.
The first class consists of the examples, for which this behaviour is due
to some kind of boundary convergence usually related to more or less
pronounced manifestations of hyperbolicity (the most representative
example being, of course, the word hyperbolic groups), see the papers
by Kaimanovich – Vershik [KV83], Kaimanovich [Kai85, Kai89, Kai94,
THOMPSON’S GROUP F IS NOT LIOUVILLE
3
Kai00], Kaimanovich – Masur [KM96], Karlsson – Margulis [KM99],
Karlsson – Ledrappier [KL07], Maher – Tiozzo [MT14]. For the examples from the second class a non-trivial behavior at infinity is provided by pointwise stabilization of random configurations in a certain
way associated with the sample paths of the random walk (this situation is exemplified by the lamplighter groups), see Kaimanovich – Vershik [KV83], Kaimanovich [Kai83, Kai91], Erschler [Ers04a, Ers04b,
Ers11], Karlsson –Woess [KW07], Sava [Sav10a], Lyons – Peres [LP15],
Juschenko – Matte Bon – Monod – de la Salle [JMBNdlS15].
In the case of Thompson’s group F , true to its ambivalent nature, we
actually construct asymptotic behaviours (µ-boundaries) of both types.
Note that although the descriptions of these µ-boundaries are quite different, currently we do not know anything about their mutual position
in the lattice of all µ-boundaries determined by a fixed step distribution µ. In particular, a priori it is not excluded that both these
µ-boundaries might in fact coincide with the full Poisson boundary.
Our main tool is the “canonical” Schreier graph Γ of the group F
(endowed with the standard generators) on the dyadic-rational orbit
of its canonical action on the unit interval. In principle one could
argue just in terms of this action. However, it turns out to be much
more convenient to “change the coordinates” and to realize Thompson’s
group F as a subgroup Fe ∼
= F of the group P LF (R) of piecewise linear
homeomorphisms of the real line. [Actually, the geometry of the graph
Γ completely described by Savchuk [Sav10b] is really begging for this
coordinate change.] Then one of the generators of F becomes just the
translation γ 7→ γ − 1 on R. The dyadic-rational orbit in the unit
interval becomes the dyadic-rational orbit in R, so that
1 we can identify
the vertex set of Γ with the dyadic-rational line Z 2 ⊂ R.
As we have already said, we construct two non-trivial µ-boundaries of
the group F . The first one is inspired by an analogy between Thompson’s group F ∼
= Fe and the lamplighter groups. Let fun(Γ, Z) (respectively, Fun(Γ, Z)) denote the additive group of finitely supported
(respectively, of all) Z-valued configurations on the graph Γ. We ase
sign to any element g of the
1 group F ⊂ P LF (R) a configuration
∼
Cg ∈ fun(Γ, Z) on Γ = Z 2 . The value of Cg at a point γ is the
logarithmic increment of the slope of g at γ (so that the support of Cg
is precisely the discontinuity set of the derivative of g). Or, in a somewhat different terminology, the support of the configuration Cg is the
set of the break points of g, and its values are the logarithms of the
jumps of g.
By the chain rule the sequence of the configurations Cgn along a
sample path (gn ) of the random walk (Fe, µ) satisfies a simple recursive
relation. This relation is completely analogous to that for the random
walks on the lamplighter groups. In precisely the same way as with
4
VADIM A. KAIMANOVICH
the lamplighter groups [KV83, Kai83], if the measure µ is finitely supported, then the transience of the induced random walk on Γ implies
that the sequence Cgn almost surely converges to a random limit configuration C∞ ∈ Fun(Γ, Z). Then it is not hard to verify that the limit
configuration C∞ can not be the same for all sample paths. Therefore,
the space Fun(Γ, Z) endowed with the arising hitting distribution is a
non-trivial µ-boundary.
This argument (once again, precisely in the same way as for the
lamplighter groups) is hinged on the transience of the induced random
walk on Γ. We establish
it by showing that for the simple random
walk on Γ ∼
= Z 21 there is a “drift” which forces the 2-adic norm to
go to infinity along the sample paths (this is the original argument
we referred to in [Kai04b]). The classical comparison technique then
leads to the transience of all random walks on Γ driven by strictly
non-degenerate probability measures µ on F ∼
= Fe.
Alternatively, the transience of the simple random walk on Γ can be
directly obtained from the explicit description of the geometry of the
graph Γ due to Savchuk [Sav10b]. Indeed, after removing from Γ a
countable set of rays isomorphic to Z+ the remaining skeleton Γ is a
tree roughly isometric to the standard binary tree. This fact readily
implies that Γ is transient.
In a recent preprint [Mis15] Mishchenko gives yet another proof of
the transience of the Schreier graph Γ, which is based on an explicit
Dirichlet norm estimate. He then notices that due to a specific geometry of Γ this transience implies existence of a non-trivial behaviour
at infinity for the simple random walk on Γ. Since it is induced by
the simple random walk on Thompson’s group F , the latter also has a
non-trivial behaviour at inifnity. This argument gives a different proof
of the absence of the Liouville property for the group F .
Mishchenko settles just for showing non-triviality of the Poisson
boundary by exhibiting a non-trivial finite partition of it. However,
his observation can be developed much further. Indeed, a transient finite range random walk on any graph necessarily converges in the end
compactification. If the graph is endowed with a transitive group action which commutes with the transition operator of the random walk,
then the dependence of this limit end on the starting point is equivariant, so that it is the same boundary behaviour of the random walk
on the group that is exhibited independently of the starting point in
the graph. However, this is not the case in our situation, and here,
in order to obtain a µ-boundary (i.e., an equivariant quotient of the
Poisson boundary), one has to consider the collections of the limit ends
of the induced random walk on Γ parameterized by all possible starting points. In other words, the resulting µ-boundary is realized on the
THOMPSON’S GROUP F IS NOT LIOUVILLE
5
space of sections of the end bundle
Γ × ∂Γ → Γ ,
(γ, ω) → γ
over the graph Γ (here ∂Γ is the space of ends of Γ). One can consider
these sections as “end fields” on Γ, or, in yet another termonilogy, as
elements of the space Fun(Γ, ∂Γ) ∼
= (∂Γ)Γ of ∂Γ-valued configurations
on Γ. By using certain self-similar features of the graph Γ (although its
group of automorphisms is trivial, it has a lot of partial isomorphisms
between its subsets), it is easy to check that this µ-boundary is nontrivial.
Our construction of a µ-boundary of the group F on the configuration
space Fun(Γ, Z) heavily used the specific features of this group. On
the contrary, the above construction of a µ-boundary on the space of
sections Fun(Γ, ∂Γ) of the end bundle over the Schreier graph Γ is
much more general. It produces a non-trivial µ boundary for a strictly
non-degenerate finitely supported step distribution µ on an arbitrary
finitely generated group G whenever there exists a Schreier graph Γ of
this group with the following two properties:
(i) the graph Γ is transient with respect to the simple random walk;
(ii) for the induced random walk (Γ, µ) its harmonic (hitting) distributions on the space of ends ∂Γ are not one-point measures.
Returning to Thompson’s group F ∼
= Fe, let us notice that the space
of ends of its canonical Schreier graph Γ splits into a disjoint union
∂Γ = ∂Γ ∪ ∂− Γ ∪ ∂+ Γ .
Here ∂Γ is the space of ends of the skeleton Γ, which can be identified
(1)
with the space Z2 = Z2 \ 2Z2 of 2-adic integers of norm 1. The sets
∂− Γ, ∂+ Γ are countably infinite and correspond to the rays attached to
the respective subsets of the skeleton. The components of the above decomposition, on which the hitting measures are actually concentrated,
can be further described in terms of the barycentres of the images of
the measure µ under its homomorphisms to Z. In particular, in the
centered case the hitting measures are supported by Fun(Γ, ∂Γ).
C. Structure of the paper. The paper has the following structure. In Section 1 we remind the background definitions concerning
the Poisson boundary of a random walk on a discrete group, the Liouville property and its links with amenability. We also remind how one
obtains a non-trivial behaviour at infinity for the lamplighter groups,
as this was our source of inspiration for constructing one of the µboundaries of Thompson’s group (the one realized on the configuration
space Fun(Γ, Z)).
Further, in Section 2 we introduce Thompson’s group F and describe
its realization as a subgroup Fe ∼
= F of the group P LF (R) of piecewise
linear homeomorphisms of the real line.
6
VADIM A. KAIMANOVICH
In Section 3 we establish the transience of random walks on the
Schreier graph Γ. First we do that for the simple random walk (Theorem 14), and then, by using a general comparison criterion (Theorem 16), for the random walk determined by any strictly non-degenerate
step distribution (Theorem 25).
In Section 4 we construct a non-trivial µ-boundary of the random
walks on Thompson’s group F ∼
= Fe, which is determined by stabilizing
Z-valued configurations on the Schreier graph Γ. It is realized on the
configuration space Fun(Γ, Z) (Theorem 35).
In Section 5 we establish several geometric properties of the Schreier
graph Γ and its space of ends.
These properties are then used in Section 6 in order to construct
another non-trivial µ-boundary of random walks on Thompson’s group
F ∼
= Fe. It is determined by the convergence of the induced random
walk on the Schreier graph Γ to its ends, and it is realized on the space
of sections Fun(Γ, ∂Γ) of the end bundle over Γ (Theorem 38). In
Theorem 39 we generalize this result to arbitrary groups which admit
a transient Schreier graph with a non-trivial space of ends. In Theorem 41 we further specify the components of the space Fun(Γ, ∂Γ),
which actually support the µ-boundary constructed in Theorem 38.
Finally, Section 7 contains a discussion of possible future developments and ramifications.
D. Historical comments. The main result of this paper (absence
of the Liouville property for Thompson’s group F ) was presented in
the talks “Boundary behaviour of the Thompson group” at the conference “Combinatorial, Geometric, and Dynamical Aspects of Infinite Groups” (1-6 June 2003, Gaeta, Italy) and at the special session
“Probabilistic and Asymptotic Aspects of Group Theory” of the AMS
meeting in Athens, Ohio (26-27 March 2004), and the slides of these
talks [Kai04b] were circulated at that time. However, it took quite
a while to complete the present version, and in the meantime some
of its ideas and tools were independently developed by other authors.
The geometry of the Schreier graphs of the group F corresponding to
its canonical action on the unit interval was completely described by
Savchuk [Sav10b, Sav15]; Mishchenko [Mis15] published a new proof
of the absence of the Liouville property for the group F based on the
analysis of Savchuk.
E. Acknowledements. I would like to thank the editors of the
present volume Tullio Ceccherini-Silberstein, Maura Salvatori and Ecaterina Sava-Huss for their infinite patience, understanding and cooperation during the preparation of this article. I am grateful to the anonymous referee for a very interesting and competent report. Anna Erschler and Yair Hartman made a number of valuable comments and
suggestions. Finally, last but not least, my thanks go to Wolfgang
THOMPSON’S GROUP F IS NOT LIOUVILLE
7
Woess, who is at the origin of this volume, and with whose friendship
and wisdom I have been honoured for many years.
1. Preliminaries
1.A. Random walks on groups. The (right) random walk (G, µ) on
a countable group G determined by a probability measure (the step
distribution) µ on G is the Markov chain on G with the transitions
µ(h)
g∼
gh ,
g∈G,
i.e., its transition probabilities πg are the translates
πg = gµ ,
g∈G,
of the measure µ. Any initial distribution θ on G determines the associated Markov measure Pθ on the space GZ+ of sample paths g =
(g0 , g1 , . . . ).
By P = Pδe we denote the probability measure on the path space
whose initial distribution is the point measure δe concentrated on the
identity e of the group G. For the random walk issued from the identity of the group (so that g0 = e), its position gn at time n is the
product gn = h1 h2 . . . hn of n independent identically µ-distributed increments hi , and the distribution of gn (the n-th marginal distribution
of the measure P) is the n-fold convolution of the measure µ.
1.B. Poisson boundary. The measure P# , whose initial distribution
is the counting measure # on the group G, is invariant with respect to
the time shift T on the path space. The Poisson boundary ∂µ G of the
random walk (G, µ) is the space of ergodic components of the time
shift T on the space (GZ+ , P# ), and any initial probability distribution θ on G determines the corresponding harmonic measure νθ on ∂µ G
as the image of the measure Pθ ≺ P# on the path space under the
quotient map GZ+ → ∂µ G. We emphasize that the Poisson boundary is
defined in the measure category only. Below, when talking about the
Poisson boundary ∂µ G, we shall always (unless otherwise specified) endow it with the measure ν = νδe , which is the harmonic measure of the
initial distribution δe .
The Poisson boundary is equipped with the natural (left) action of
the group G induced by the coordinate-wise left translations on the
path space, and for an arbitrary initial distribution θ
X
νθ = θ ∗ ν =
θ(g) gν .
g
Although the harmonic measure ν need not be quasi-invariant (if the
semigroup sgr µ generated by the support supp µ of the measure µ is
8
VADIM A. KAIMANOVICH
smaller than G), it is µ-stationary with respect to this action in the
sense that
X
µ∗ν =
µ(g) gν = ν .
g
1.C. Poisson formula. A function f on G is called µ-harmonic if it
is preserved by the Markov operator of the random walk (G, µ)
X
Pµ f (g) =
µ(h)f (gh) ,
h
i.e., if Pµ f = f . The Poisson formula
f (g) = hfb, gνi ,
g∈G,
(1)
establishes an isometric isomorphism between the space H ∞ (G, µ) of
bounded µ-harmonic functions f on the group G and the L∞ space on
the Poisson boundary ∂µ with respect to the quotient measure class
determined by the measure P# . It consists in assigning to a boundary
function fb the function f on G whose value at a point g ∈ G is the
result of the integration of the function fb against the translate gν of
the harmonic measure ν.
If g in the Poisson formula (1) is only allowed to take values in the
semigroup sgr µ, then (1) becomes an isometric isomorphism between
the space of bounded µ-harmonic functions on sgr µ and L∞ (∂µ G, ν).
This property uniquely characterizes the Poisson boundary. Namely,
any G-space (B, λ), for which formula (1) is an isometric isomorphism
between the spaces H ∞ (sgr µ, µ) and L∞ (B, λ), is isomorphic to the
Poisson boundary (∂µ G, ν) in the category of measure G-spaces. See
Kaimanovich – Vershik [KV83], Kaimanovich [Kai92, Kai00], Furman
[Fur02], Erschler [Ers10] for more background on the Poisson boundary.
1.D. Stability of the Liouville property. A random walk (G, µ) is
called Liouville if the harmonic measure ν is a point measure. The reason for this terminology is that in view of the Poisson formula (1) this
property means that there are no non-constant bounded µ-harmonic
functions on the semigroup sgr µ. One can easily see that the latter
property is actually equivalent to the absence of non-constant bounded
µ-harmonic functions on the whole group gr µ generated by supp µ as
well.
There are numerous examples showing that one can have both Liouville and non-Liouville random walks on the same group (e.g., see
Kaimanovich – Vershik [KV83], Kaimanovich [Kai83], Erschler [Ers04b,
Ers04a], Bartholdi – Erschler [BE11]), and it is still not clear how the
Liouville property for random walks on the same group depends on the
step distribution µ.
THOMPSON’S GROUP F IS NOT LIOUVILLE
9
The main open problem here is
whether all finitely supported symmetric measures µ with
gr µ = G on the same finitely generated group G are
either Liouville or non-Liouville simultaneously,
or, in a somewhat different form,
whether the Liouville property is stable with respect to
rough isometries.
In the absence of group invariance, for general graphs or Riemannian
manifolds, the corresponding couterexamples to the stability of the Liouville property were first constructed by Terry Lyons [Lyo87]. We believe that such counterexamples should exist in the group setup as well.
However, the pool of groups, for which the Liouville property has been
studied, still remains quite limited in spite of some significant recent
progress, see Bartholdi – Virag [BV05], Brieussel [Bri09], Bartholdi –
Kaimanovich – Nekrashevych [BKN10], Amir – Angel – Virág [AAV13],
Amir – Angel – Matte Bon – Virág [AAMBV13], Matte Bon [MB14],
Kotowski – Virág [KV15], Juschenko – Matte Bon – Monod – de la
Salle [JMBNdlS15].
1.E. Liouville groups. Because of the above there are different ways
of calling a group G Liouville depending on which class of step distributions µ on G one considers (e.g., all measures, just finitely supported
measures, finitely supported symmetric measures, etc.). According to
one popular definition a group G is called Liouville if the random walk
(G, µ) is Liouville for any (possibly degenerate) finitely supported symmetric probability measure µ on G, see Matte Bon [MB14]. Sometimes
one also talks about the Liouville property of a group G with respect
to a finite symmetric generating set K (in which case one takes for µ
the measure equidistributed on K), see Gournay [Gou14].
1.F. Amenability and the Liouville property. Liouville groups
(no matter which definition one takes) are always amenable, which
goes back to Furstenberg [Fur73]. The reason for that is that a random walk (G, µ) is Liouville if and only if the sequence of Cesaro averages of the convolution powers of µ is asymptotically invariant with
respect to the left action of the group gr µ on itself, see Kaimanovich –
Vershik [KV83]. Therefore, the Liouville property of the random walk
(G, µ) implies amenability of the group gr µ in a very constructive way.
Conversely, as it was first conjectured by Furstenberg in [Fur73], and
proved by Rosenblatt [Ros81] and Kaimanovich – Vershik [KV83], any
amenable group G carries a symmetric measure µ with supp µ = G
such that the random walk (G, µ) is Liouville. Note that there are
amenable groups G, for which such a measure µ can not be chosen
finitely supported, as it was shown by Kaimanovich [Kai83], or even to
have finite entropy, as it was shown by Erschler [Ers04b].
10
VADIM A. KAIMANOVICH
Therefore, a possible approach to proving non-amenability of a given
group G consists in showing that it carries no Liouville random walks
with non-degenerate step distributions µ (i.e., such that gr µ = G).
Actually, in view of the above result of Rosenblatt and Kaimanovich –
Vershik, it is enough to consider just the measures with sgr µ = G
(we shall call such measures strictly non-degenerate; sometimes they
are also called adapted), or even just with supp µ = G. Here me make
the first step in this direction for Thompson’s group F by showing that
random walks on F with finitely supported strictly non-degenerate step
distributions are non-Liouville.
1.G. Lamplighter groups. Our approach to the Liouville property
for the group F consists in using the transience of the induced random
walk on an auxiliary homogeneous space Γ of the group to prove that
certain configurations on Γ associated with the elements of F stabilize
along the sample paths of the random walk. It is precisely this idea
that was first used for exhibiting a non-trivial boundary behavior for
random walks on lamplighter groups, see Kaimanovich – Vershik [KV83]
and Kaimanovich [Kai83], and, to begin with, we shall outline it in the
lamplighter context.
Definition 2. The lamplighter group L(G) over a base group G is the
semi-direct product of the group G and the additive group fun(G, Z/2Z)
of finitely supported Z/2Z-valued configurations on G endowed with
the action of G by translations.
Remark 3. In the context of functional and stochastic analysis these
groups were first introduced by Vershik and the author [KV83] under
the name of the groups of dynamical configurations. However, this term
did not stick, and the current generally accepted standard is to call
them lamplighter groups. In the algebraic language L(G) is the wreath
product with the active group G and the passive group Z/2Z. Since our
purpose is just to outline the general idea, we are not discussing here
more general wreath products (for which see, for instance, Kaimanovich
[Kai91] and Erschler [Ers04a, Ers06a]).
As a set, L(G) is the usual product of G and fun(G, Z/2Z), so that
L(G) = {(g, Φ) : g ∈ G, Φ ∈ fun(G, Z/2Z)} .
However, the group operation in L(G) is “skewed” by using the left
action of G on fun(G, Z/2Z) by the group automorphisms
Tg Φ(h) = Φ(g −1 h)
(4)
induced by the left action of the group G on itself by translations, so
that the group multiplication in L(G) is
(g1 , Φ1 )(g2 , Φ2 ) = (g1 g2 , Φ1 + Tg1 Φ2 ) .
(5)
THOMPSON’S GROUP F IS NOT LIOUVILLE
11
The identity of L(G) is (e, ∅), where e is the identity of G, and ∅ (the
empty configuration defined as ∅(h) = 0 for any h ∈ G) is the identity
of fun(G, Z/2Z).
1.H. Stabilization of configurations. Given a probability measure µ
on the group L(G), the positions (gn , Φn ) of the corresponding random
walk at two consecutive time moments are related as
(gn+1, Φn+1 ) = (gn , Φn )(hn+1 , ϕn+1) ,
where (hi , ϕi ) are the increments of the random walk. Therefore, by
formula (5),
(
gn+1 = gn hn+1 ,
(6)
Φn+1 = Φn + Tgn ϕn+1 .
In particular, the first component gn ∈ G, i.e., the image of the random
walk (L(G), µ) under the group homomorphism
L(G) → G ,
(g, Φ) 7→ g ,
performs the quotient random walk on G driven by the image µ′ of the
measure µ under the homomorphism L(G) → G.
If the quotient random walk (G, µ′) on G is transient (i.e., its sample
paths almost surely go to infinity on G), and the support of the measure
µ is finite (which implies that there is a finite set A ⊂ G which contains
the supports of all the increments ϕn ), then formula (6) implies that
the configurations Φn almost surely stabilize as n → ∞. It means that
there exists a pointwise random limit
Φ∞ = lim Φn = ϕ1 + Tg1 ϕ2 + Tg2 ϕ3 + . . . ,
n→∞
which belongs to the space Fun(G, Z/2Z) of all (not necesssrily finitely
supported) Z/2Z-valued configurations on G.
It is easy to see that under natural conditions on the measure µ
the limit configuration Φ∞ can not be the same for a.e. sample path,
and therefore the Poisson boundary of the random walk (L(G), µ) is
non-trivial. This is the case if the measure µ is non-degenerate in
L(G), but actually this assumption can be siginificantly weakened,
see Kaimanovich [Kai91]. The map, which assigns to a sample path
(gn , Φn ) the associated limit configuration Φ∞ ∈ Fun(G, Z/2Z), is obviously L(G)-equivariant (with respect to the action on Fun(G, Z/2Z)
determined by the “configuration component” of formula (5)). Therefore, the configuration space Fun(G, Z/2Z) endowed with the resulting hitting distribution is a µ-boundary (an equivariant quotient of the
Poisson boundary).
12
VADIM A. KAIMANOVICH
2. Thompson’s group F as a subgroup of P LF (R)
2.A. The group F and its generators.
Definition 7. The Thompson group F is the group of the orientation preserving piecewise linear self-homeomorphisms g of the closed
unit interval [0, 1] that are differentiable except for finitely many break
points ti , which are dyadic rational numbers, and such that the slopes ai
are integer powers of 2:
g(t) = ai t + bi
see Figure 1.
on [ti , ti+1 ] ,
ki
ti = m ∈ Z 12 ,
2 i
ai = 2ni ,
bi ∈ Z
1
2
,
1
0
ti
ti+1
Figure 1. The graph of an element of Thompson’s group F .
We refer to Cannon – Floyd – Parry [CFP96] for the general background on the group F . In particular, it is finitely generated with the
generators
t
0 ≤ t ≤ 12
2 ,
1
A(t) = t − 41 ,
≤ t ≤ 34
2
2t − 1 , 3 ≤ t ≤ 1
4
and
t,
0 ≤ t ≤ 21
t + 1 , 1 ≤ t ≤ 3
2
4
B(t) = 2 14
3
7
t − 8 ,
≤
t
≤
4
8
2t − 1 , 87 ≤ t ≤ 1
,
THOMPSON’S GROUP F IS NOT LIOUVILLE
13
see Figure 2 (where the arrow indicates that the graph of B restricted
to the square [ 21 , 1] × [ 12 , 1] is precisely the graph of A rescaled by a
factor of 2).
A
B
1
1
3
4
5
8
1
2
1
2
1
4
0
1
2
3
4
1
0
1
2
3
4
7
8
1
Figure 2. The standard generators of Thompson’s group F .
The group F is finitely presented, and it is determined by the relators
[AB −1 , A−1 BA] ,
[AB −1 , A−2 BA2 ] ,
where [x, y] = xyx−1 y −1 denotes the usual group theory commutator.
The group F is also often described by the infinite presentation
hx0 , x1 , x2 , . . . |x−1
k xn xk = xn+1 for all k < ni ,
which can be obtained from the previous one by putting x0 = A and
xn = A1−n BAn−1 for n > 0.
2.B. Homomorphisms to Z. The abelianization Z2 of the group F
is freely generated by the images of the above generators A and B. By
χa and χb we denote the corresponding homomorphisms of F to Z, i.e.,
the homomorphisms χa and χb are defined by putting, respectively,
χa (A) = 1 ,
χa (B) = 0 ,
χb (A) = 0 ,
χb (B) = 1 .
and
Geometrically, −χa (g) (respectively, χa (g) + χb (g)) is the base 2 logarithm of the slope of the graph of g at the endpoint 0 (respectively, at
the endpoint 1).
14
VADIM A. KAIMANOVICH
2.C. Change of variables. The group F can also be realized as a
subgroup of the group of piecewise linear homeomorphisms of the real
line P LF (R) introduced by Brin – Squier [BS85] (also see Haagerup –
Picioroaga [HP11, Remark 2.5]).
More precisely, let τ : [0, 1] → R be defined by putting
1
τ
=n+1,
n ≤ −1
2n
(8)
1
n≥1
τ 1− n =n−1 ,
2
and by linear interpolation in between, see Figure 3.
0
−2
1
8
1
4
1
2
3
4
−1
0
+1
7
8
1
+2
Figure 3. The change of variables from [0, 1] to R.
Then one can easily see that τ is a bijection between
1 the set of
dyadic-rational numbers in the open unit interval
Z 2 ∩ (0, 1) and
1
the whole set of dyadic-rational numbers Z 2 . After this change of
coordinates the generators A and B of the group F (see Section 2.A)
take the form
e
A(γ)
= γ−1 ,
γ∈R,
and
γ ≤ 0,
γ ,
γ
e
B(γ) = 2 ,
0 ≤ γ ≤ 2,
γ − 1 , γ ≥ 2
Their inverses are, respectively,
and
e−1 (γ) = γ + 1 ,
A
γ∈R,
γ ≤ 0,
γ ,
−1
e (γ) = 2γ ,
B
0 ≤ γ ≤ 1,
γ + 1 , γ ≥ 1 ,
e and B
e themselves
see Figure 4, where the graphs of the generators A
are drawn with solid lines, and the graphs of their inverses are drawn
with dashed lines.
THOMPSON’S GROUP F IS NOT LIOUVILLE
eA
e−1
A,
−1
e B
e −1
B,
1
0
15
2
1
0
1
1
2
−1
Figure 4. The graphs of the standard generators of the
group Fe.
We shall denote the image of the group F under the change of coordinates (8) by Fe ∼
= F , i.e., Fe is the subgroup of P LF (R) generated by
e and B.
e
the transformations A
2.D. Intrinsic description of Fe. One can easily see that Fe, as a
subgroup of P LF (R), consists precisely of those transformations g ∈
P LF (R), for which the following three conditions are satisfied:
(i) The discontinuity points of the derivative g ′ are all dyadic rational;
(ii) The slopes of g are integer powers of 2;
(iii) For a sufficiently large M > 0 the transformation g has the form
g(γ) = γ + C−
for γ ∈ (−∞, −M) ,
and
g(γ) = γ + C+
for γ ∈ (M, ∞) ,
where the constants C− = C− (g) and C+ = C+ (g) are integers.
Clearly, the maps C± : Fe → Z are group homomorphisms, and for
the generators of the group Fe
whereas
e = C+ (A)
e = −1 ,
C− (A)
e =0,
C− (B)
e = −1 ,
C+ (B)
so that in terms of the homomorphisms χa , χb : F ∼
= Fe → Z introduced
in Section 2.B
C− = −χa
and C+ = −χa − χb .
16
VADIM A. KAIMANOVICH
3. Random walk on the Schreier graph Γ ∼
=Z
1
2
3.A. The dyadic-rational Schreier graph. For technical reasons
(since traditionally one considers right random walks on groups, for
which the increments are added on the right), from now on we shall
use the postfix notation for the group Fe. Thus, the group operation
in Fe will be defined as
(g1 g2 )(γ) = g2 (g1 (γ)) ,
(9)
(rather than the more common (g1 g2 )(γ) = g1 (g2 (γ)) in the prefix notation which we used in Section 2.A). Then Fe will naturally act on R
on the right as
γ ∈ R, g ∈ Fe .
γ.g = g(γ) ,
(10)
Our notation agrees with the one used when dealing with random walks
on groupoids (in particular, on the groupoids associated with group actions), see Kaimanovich [Kai05].
1
The set of dyadic rational
numbers
Z
⊂ R is obviously transitive
2
1
for this action, so that Z 2 can be endowed with the structure of the
Schreier graph Γ of the group Fe with respect to the generating set
e B,
e A
e−1 , B
e −1 } .
K = {A,
(11)
3.B. Simple random walk on the Schreier graph. The simple
random walk on the graph Γ has the transitions
µ(g)
γ∼
γ.g ,
(12)
where µ = µK is the probability measure equidistributed on the generating set K, i.e.,
γ∼
1
e
∼4 A(γ)
1
∼4 A
e−1 (γ)
1
e
∼4 B(γ)
1
4 e −1
∼ B (γ)
.
More concretely, depending on the position of the point γ these transitions are
THOMPSON’S GROUP F IS NOT LIOUVILLE
γ∼
γ∼
γ∼
1
∼2 γ + 1
12
∼ γ−1
1
∼4 γ + 1
1
∼4 γ − 1
1
4
γ
∼
2
14
∼ 2γ
1
2
∼ γ + 1
1
∼4 γ − 1
14 γ
∼ 2
,
γ ∈ (∞, 0] ∪ [2, ∞) ,
,
γ ∈ [0, 1] ,
,
γ ∈ [1, 2] .
17
We shall denote by Q the probability measure on the path space of the
simple random walk on Γ corresponding to the starting point 0.
Let |γ|2 be the usual 2-adic norm of a number γ ∈ Γ ∼
= Z 21 , i.e.,
|γ|2 = 2−n
for γ = (2m + 1)2n ,
and |0|2 = 0.
Theorem 13. For Q-a.e. sample path (γn ) of the simple random walk
(12) on Γ
|γn |2 −→ ∞ .
n→∞
As a consequence we immediately obtain
Theorem 14. The simple random walk on Γ is transient.
Proof of Theorem 13. Let us first of all notice that, outside of the interval (0, 2) the transition probabilities of the random walk (12) are
just
p(γ, γ ± 1) =
1
.
2
Therefore, the interval [0, 2) is recurrent by the classical properties
of the simple random walk on Z. Moreover, starting from any point
γ ∈ [1, 2) the random walk will eventually hit the interval [0, 1) (and
the corresponding hitting distribution is supported by the points γ − 1
and γ2 with the equal weights 12 ). Thus, the interval I = [0, 1) is
also recurrent. The transitions of the induced random walk on I are,
18
VADIM A. KAIMANOVICH
for γ ∈ 0, 12 ,
γ∼
1
∼4 γ
1
∼4 γ + 1 ∼
1
4
γ
∼
2
14
∼ 2γ
1
∼2 γ
12 γ
∼ 2+
1
2
,
i.e.,
3
∼8 γ
1
∼4 γ
2
γ∼
and, for γ ∈
1
1
8
γ
∼
+
1 2
4
∼ 2γ
,
1
,
2
γ∼
1
∼4 γ
1
4
∼
γ+1 ∼
1
4
γ
∼
2
1
∼4 2γ ∼
i.e.,
γ∼
,
1
2
1
∼2 γ
12 γ
∼ 2+
1
2
1
∼2 2γ − 1
12
∼ γ
1
∼2 γ
1
∼4 γ
2
1
8
γ
∼
+ 12
2
81
∼ 2γ − 1
.
,
THOMPSON’S GROUP F IS NOT LIOUVILLE
19
Therefore, for γ ∈ I \ 0, 21 the logarithm x = log2 |γ|2 of the 2-adic
norm changes as
3
8
∼3 x
1
8
, γ ∈ 0,
x∼
,
∼ x+1
2
1
4
∼ x−1
1
2
∼3 x
1
8
,1 ,
, γ∈
x∼
∼ x+1
2
1
8
∼ x−1
i.e., these transitions on Z+ have a uniform positive drift, which implies
the claim.
Remark 15. It might be interesting to study the properties of the induced random walk on I and of its stationary measure (which is most
likely unique).
3.C. Comparison criterion for transience of Markov chains.
We shall now use a comparison argument in order to deduce the transience of general (not necessarily reversible!) random walks on the
Schreier graph Γ from the transience just of the simple random walk
on Γ. Its idea goes back to Baldi – Lohoué – Peyrière [BLP77], and
it has been quite popular ever since (e.g., see Varopoulos [Var83],
Chen [Che91], Woess [Woe94] as well as the exposition in Woess’ book
[Woe00, Sections 2.C and 3.A]). For the sake of completeness we shall
prove it here in the generality sufficient for our purposes by slightly
modifying the arguments of Varopoulos from [Var83, Section 4].
Theorem 16. Let P and P ′ be two Markov operators on a countable state space X with the respective transition probabilities p(·, ·) and
p′ (·, ·), and such that
(i) P and P ′ have a common stationary measure m;
(ii) The operator P is reversible with respect to the measure m,
i.e., it is self-adjoint as an operator on the space L2 (X, m), or,
equivalently,
m(x)p(x, y) = m(y)p(y, x)
∀ x, y ∈ X ,
(iii) There exists ε > 0 such that P ′ ≥ εP , i.e.,
p′ (x, y) ≥ εp(x, y)
∀ x, y ∈ X .
Then the transience of the operator P implies the transience of the
operator P ′ .
20
VADIM A. KAIMANOVICH
Proof. First of all, let us notice that the operator
P − εP ′
Q=
1−ε
is Markov and preserves the measure m. Therefore, for any 0 < λ < 1
and any f ∈ L2 (X, m)
|λhf, Qf i| ≤ |hf, Qf i| ≤ kf k2 .
It implies that
h(I − λP ′ )f, f i ≥ εkf k2P,λ ≥ 0 ,
where h·, ·i denotes the scalar product on the space L2 (X, m),
hf, giP,λ = hf, (I − λP )gi
(17)
(18)
is the positive definite bilinear form on L2 (X, m) determined by the
operator I − λP , and k · k, k · kP,λ are the respective associated norms.
The Cauchy-Schwarz inequality for the form h·, ·iP,λ then implies that
for any f ∈ L2 (X, m)
h(I − λP ′ )−1 f, f i2 = h(I − λP ′)−1 f, (I − λP )(I − λP )−1f i2
= h(I − λP ′)−1 f, (I − λP )−1 f i2P,λ
≤ k(I − λP ′ )−1 f k2P,λ · k(I − λP )−1f k2P,λ .
By (17),
whereas
1
k(I − λP ′)−1 f k2P,λ ≤ h(I − λP ′ )−1 f, f i ,
ε
k(I − λP )−1 f k2P,λ = h(I − λP )−1 f, f i ,
so that
1
h(I − λP ′)−1 f, f i2 ≤ h(I − λP ′ )−1 f, f i · h(I − λP )−1 f, f i . (19)
ε
Now, if f = δo for a point o ∈ X, then
1
GP,λ (o, o) =
h(I − λP )−1 δo , δo i
m(o)
is the λ-Green kernel of the operator P at the point o, and the same
holds for the operator P ′ . Since λ < 1,
0 < GP ′ ,λ (o, o) < ∞ ,
whence by (19)
1
1
GP ′ ,λ (o, o) ≤ GP,λ (o, o) ≤ GP,1 (o, o) ,
ε
ε
which, by letting λ → 1, implies the inequality
1
GP ′ ,1 (o, o) ≤ GP,1 (o, o) .
(20)
ε
Therefore, the finiteness of GP,1 (o, o) (≡ the transience of P ) implies
the finiteness of GP ′ ,1 (o, o) (≡ the transience of P ′ ).
THOMPSON’S GROUP F IS NOT LIOUVILLE
21
Remark 21. We are not aware of any “elementary” proof of inequality (20).
Remark 22. The form (18) decomposes as
h·, ·iP,λ = DP + (1 − λ)h·, ·i ,
where DP is the Dirichlet form of the operator P .
3.D. General random walks. Given a probability measure µ on G
we shall denote by (Γ, µ) the induced random walk on Γ with the transitions
µ(g)
γ ∼ γ.g .
(23)
In other words, the measure Qγ on the space of paths of the induced
random walk issued from a point γ ∈ Γ is the image of the measure P
on the path space of the random walk (G, µ) (see Section 1.A) under
the map
g = (gn ) 7→ γ.g = (γ.gn ) .
(24)
Theorem 25. Let µ be a strictly non-degenerate probability measure
on the group Fe. Then the induced random walk (Γ, µ) is transient.
Proof. This is a direct application of Theorem 16. By the strict nondegeneracy of the measure µ, for any g ∈ Fe there exists n = n(g) such
that the n-fold convolution µ∗n (g) is strictly positive. Let
n = max{n(g) : g ∈ K} ,
and
δ = min{µ∗n(g) (g) : g ∈ K} > 0 ,
where K is the generating set (11). Then the measure
n
1 X ∗k
µ′ =
µ
n
k=1
has the property that
µ′ (g) ≥ εµK (g)
∀g ∈ K ,
(26)
with ε = δ/n > 0.
Let us now take for P (respectively, P ′ ) the Markov operator of
the simple random walk (Γ, µK ) (respectively, the operator of the random walk (Γ, µ′ )). If we take for m the counting measure on Γ, then
conditions (i) and (ii) of Theorem 16 are obviously satisfied, whereas
condition (iii) follows from inequality (26). Thus, Theorem 14 implies
transience of the random walk (Γ, µ′ ), and therefore of the random walk
(Γ, µ) as well.
Remark 27. One should be able to significantly relax the condition
sgr µ = Fe imposed on the measure µ in Theorem 25. We expect it to
hold just under very mild assumptions on the group gr µ.
22
VADIM A. KAIMANOVICH
4. Non-trivial behaviour at infinity determined by
stabilizing configurations
4.A. Groups of configurations on Γ. Let fun(Γ, Z) (respectively,
Fun(Γ, Z)) denote the additive group of finitely supported (respectively,
of all) Z-valued configurations on the graph Γ, cf.Section
1.G. By using
1
∼
e
the right action (10) of the group F on Γ = Z 2 we shall define the
associated left action of Fe on the configuration space Fun(Γ, Z) as
Sg ϕ(γ) = ϕ(γ.g) = ϕ(g(γ)) ,
g ∈ Fe, γ ∈ Γ .
(28)
4.B. Configurations associated with the elements of Fe. Given a
transformation g ∈ Fe, we shall denote by Cg ∈ fun(Γ, Z) the associated
configuration on Γ defined as
Cg (γ) = log2 g ′(γ + 0) − log2 g ′ (γ − 0)
∀γ ∈ Γ .
In other words, Cg (γ) is the difference between the base 2 logarithms
of the left and the right slopes of g at the point γ. The support of Cg
is precisely the set of break points of g, i.e., the set of discontinuity
points of the derivative g ′ , see Figure 5.
Informally we shall say that the value Cg (γ) is the logarithmic incre′ (γ+0)
ment of the slope of g at the point γ. The ratio gg′ (γ−0)
appears in the
literature on Thompson’s groups (e.g., see Liousse [Lio08]) under the
name of the jump of g at the break point γ, so that in these terms Cg (γ)
is the logarithm of the jump of g at γ.
C(γ2 ) = 1
C(γ1 ) = −2
γ1
γ2
Figure 5. The configuration on Γ ∼
= Z
with an element of Fe.
1
2
associated
Remark 29. Obviously, two different transformations g1 , g2 ∈ Fe give
rise to the same configuration if and only if g2 = g1 + C for C ∈ Z (for,
if the difference between two functions from Fe is a constant, then this
constant must be an integer, see Section 2.D). On the other hand, not
every configuration ϕ ∈ fun(Γ, Z) corresponds to an element g ∈ Fe.
There are two conditions, whose combination is necessary and sufficient
THOMPSON’S GROUP F IS NOT LIOUVILLE
for that. The first condition is
X
23
ϕ(γ) = 0
γ∈Γ
(for, any g ∈ Fe has slope 1 both at −∞ and at +∞, so that the sum
of logarithmic increments of the slope must be 0). It guarantees that
there exists a transformation g ∈ P LF (R) with
(
g(τ ) = τ + C−
at − ∞ ,
g(τ ) = τ + C+
at + ∞ ,
where the difference C+ − C− is uniquely determined by the configuration ϕ. Now, the second condition is C+ − C− ∈ Z (cf. Section 2.D).
For instance, the configuration ϕ = −δ0 + δ1 defined as
−1 , γ = 0 ,
ϕ(γ) = +1 , γ = 1 ,
0 , othwerwise
satisfies the first condition, but not the second one; the resulting transformation g ∈ P LF (R) is (up to an additive constant)
γ≤0,
γ ,
γ
g(γ) = 2 ,
0≤γ≤1,
γ − 1 , γ ≥ 1 .
2
4.C. Composition of configurations. By the chain rule
(g1 g2 )′ (γ) = g1′ (γ) · g2′ (g1 (γ))
(keep in mind that the group multiplication (9) in Fe is defined by using
the “inverse composition”). Thus,
Cg1 g2 (γ) = Cg1 (γ) + Cg2 (g1 (γ))
∀ g1 , g2 ∈ Fe, γ ∈ Γ ,
(30)
or, in terms of the action (28),
Cg1 g2 = Cg1 + Sg1 Cg2
∀ g1 , g2 ∈ Fe .
(31)
Remark 32. Formula (31) looks very similar to the “configuration component” of formula (5) for the multiplication in the lamplighter groups.
Note, however, that the actions T (4) and S (28) on the respective configuration spaces, which appear in these formulas, are quite different.
The action T is determined by the action of the group on itself onthe
left, whereas the action S is determined by the action on Γ ∼
= Z 21 on
the right. In particular, as a result of this
Tg δh = δgh
∀ g, h ∈ G
in the lamplighter setup of Section 1.G, whereas for the action (28)
∀ g ∈ Fe, γ ∈ Γ .
Sg δγ = δγ.g−1
24
VADIM A. KAIMANOVICH
Formula (31) allows one to define yet another (“skew”) left action of
the group Fe on the configuration space Fun(Γ, Z) as
(g, C) 7→ Cg + Sg C .
(33)
This action is similar to the natural action of the lamplighter group
L(G) on the configuration space Fun(G, Z/2Z) defined as
(g, Φ) C = Φ + Tg C .
It is this action that we mentioned at the end of Section 1.H saying
that it is determined by the “configuration part” of the multiplication
formula in the group L(G). Note, however, that in the lamplighter
case the component g of a group element (g, Φ) ∈ L(G) is responsible
just for “translating” a configuration C, whereas the component Φ is
responsible just for adding an “increment” to C. On the contrary, there
is no such “splitting” in the case of the action (33).
Lemma 34. The action (33) of the group Fe on Fun(Γ, Z) has no fixed
points.
Proof. Let C ∈ Fun(Γ, Z) be a fixed point of the action, i.e.,
C = Cg + Sg C
∀ g ∈ Fe .
e of the
Since the configuration CAe associated with the generator A
group Fe (see Section 2.C) is empty, the configuration C must be then
e
invariant with respect to the transformation SA determined by the
action S (28). In other words, C must be periodic on Γ ∼
= Z 21 with
period 1.
e we arrive at the condition
In the same way, by taking g = B
e
C = CBe + SB C ,
where
CBe = −δ0 + δ2
e
e
by the definition
1 of the generator B (see Section 2.C). Since B acts
trivially on Z 2 ∩ (−∞, 0], it implies that C can not be 1-periodic,
whence a contradiction.
4.D. Stabilization of configurations.
Theorem 35. Let µ be a finitely supported strictly non-degenerate
probability measure on the group Fe. Then for a.e. sample path (gn )
of the random walk (Fe, µ) the configurations Cn = Cgn pointwise converge to a limit Z-valued configuration C∞ ∈ Fun(Γ, Z), and the space
Fun(Γ, Z) endowed with the arising limit distribution λ is a non-trivial
µ-boundary. Therefore, the Poisson boundary (∂µ Fe, ν) itself is also
non-trivial.
THOMPSON’S GROUP F IS NOT LIOUVILLE
25
Proof. By formula (30), for any γ ∈ Γ
Cn+1 (γ) = Cn (γ) + Chn+1 (γ.gn ) ,
where, as usual, (hn ) is the sequence of increments of the random
walk (gn ). Since supp µ is finite, the supports of all configurations
Ch , h ∈ supp µ, are contained in a certain finite set A ⊂ Γ. Therefore,
by Theorem 25 the sequence (Cn (γ)) almost surely stabilizes for any
γ ∈ Γ, i.e., it converges to the limit configuration
C∞ = lim Cn = Ch1 + Sg1 Ch2 + Sg2 Ch3 + . . . .
n
Then the resulting measure ν on the space Fun(Γ, Z) is µ-stationary
with respect to the action (33), cf. Section 1.H for the lamplighter
case.
Now it remains to show that the limit configuration C∞ can not be
the same for a.e. sample path (gn ), i.e., that there does not exist a
configuration C ∈ Fun(Γ, Z) such that almost surely C∞ = C. Indeed,
since sgr µ = Fe, if this were the case, then the limit configuration
C would have necessarily been a fixed point of the action (33) of the
group Fe on Fun(Γ, Z). However, this is impossible by Lemma 34.
Remark 36. The limit distribution λ constructed in Theorem 35 is
concentrated on the subset of Fun(Γ, Z) which consists only of infinitely
supported configurations. The reason is that its complement fun(Γ, Z)
is countable, whereas it is well-known that any non-trivial µ-boundary
is purely non-atomic, see Kaimanovich [Kai95].
5. Geometry of the Schreier graph Γ
5.A. The graph Γ. Our proof of Theorem 14 on
1the
transience of the
∼
simple random walk on the Schreier graph Γ = Z 2 of the group Fe was
based on purely probabilistic considerations, and this is the argument
we had referred to in [Kai04b].
Shortly thereafter Savchuk independently analyzed the geometry of
the Schreier graph of the original Thompson group F (endowed with the
standard set of generators) on the dyadic-rational orbit in the interval
[0, 1], and obtained its complete description [Sav10b] (also see [Sav15]
for the Schreier graphs on the other orbits of the canonical action of the
group F ). Since the dyadic-rational orbit of our group Fe is precisely
the image of the dyadic-rational orbit of the original group F under
the coordinate change (8) (see Section 2.C), this Schreier graph is
isomorphic to our graph Γ.
Figure 6 is a somewhat modified version of the picture of the graph Γ
which first appeared in [Sav10b] (I am most grateful to Dmytro Savchuk
for sharing his graphics source files with me).
The labelling of the vertices
1 of Γ on our Figure 6 corresponds to
∼
e
the action of F on Γ = Z 2 rather than to the original action of
26
VADIM A. KAIMANOVICH
1
1
2
3
2
1
4
3
4
5
4
7
4
5
8
1
8
3
8
13
8
9
8
1
16
9
16
5
16
7
8
15
8
11
8
13
16
3
16
7
16
11
16
15
16
Figure 6. The Schreier graph Γ of Thompson’s group
F ∼
= Fe on the dyadic-rational orbit in R.
γ
γ−3
γ−2
γ−1
Figure 7. Negative type rays in the Schreier graph Γ.
γ
γ+1
γ+2
γ+3
Figure 8. Positive type rays in the Schreier graph Γ.
the group F on Z 21 ∩ (0, 1) as in [Sav10b]. Following [Sav10b], we
use three different colors for the vertices of the
1 graph Γ in order to
better exhibit
its structure: black for γ ∈ Z 2 ∩ (−∞, 0], grey for
1
γ ∈ Z 2 ∩ (0, 1), and white for γ ∈ Z 12 ∩ [1, ∞).
The solid (respectively, dashed) arrows represent the oriented edges
e (respectively, B),
e see Section 2.C.
in Γ corresponding to the generator A
As is customary for Cayley and Schreier graphs, labels are assigned to
oriented edges, so that the label of the same edge endowed with the opeA
e−1 , B,
e B
e −1 } inverse
posite orientation is the letter of the alphabet {A,
to the label of the original edge.
THOMPSON’S GROUP F IS NOT LIOUVILLE
27
L99
The “open arrows” ← (respectively, ←
−) on Figure 6 represent the
infinite rays of negative (respectively, positive) type, see Figure 7 and
Figure 8, respectively. These rays are the connected
1 components of
the subgraph
1 of Γ with the vertex set (−∞, 0] ∩ Z 2 (respectively,
[2, ∞) ∩Z 2 ) attached tothe starting points from the set (0, 1] ∩Z 21
(respectively, [1, 2) ∩ Z 12 ). The only vertex of Γ, to which two rays
are attached (both a negative type and a positive type ones), is 1. Note
that by the presence of these two families of rays the structure of the
graph Γ is really begging for applying the coordinate change (8).
5.B. The skeleton. The skeleton Γ of the Schreier graph Γ is the tree
obtained by removing from Γ all the rays of negative and positive types,
see Figure 9. The tree Γ is obviously roughly isometric to the usual
rooted binary tree. Indeed, let B be the binary tree which consists
just
of the grey vertices of Γ (i.e., of the points from (0, 1) ∩ Z 21 ). Then
Γ and B differ only by the presence of an additional edge (the one
between the vertices 1 and 12 ) and of additional white vertices which
subdivide some of the edges of B (into two halves.
1
1
2
3
2
1
4
3
4
5
4
7
4
5
8
1
8
1
16
11
8
13
8
9
8
9
16
5
16
7
8
3
8
13
16
3
16
15
8
11
16
7
16
15
16
Figure 9. The skeleton Γ of the Schreier graph Γ.
5.C. Transience of Γ. Since the rays removed when passing from Γ
to Γ are clearly recurrent for the simple random walk (we have already
referred to this property in the proof of Theorem 13), the transience
of the simple random walk on the Schreier graph is then a direct consequence of the classically known transience of the rooted binary tree
28
VADIM A. KAIMANOVICH
(e.g., see Woess [Woe00]). This argument is self-evident, once the geometry of the Schreier graph Γ has been exhibited. Although the statement about the transience of Γ does not explicitly appear in Savchuk’s
paper [Sav10b], its author was aware of it and mentioned it in some of
his talks given at that time.
5.D. Space of ends of Γ. Let ∂Γ denote the space of ends of the
graph Γ. It obviously splits into a disjoint union
∂Γ = ∂Γ ∪ ∂− Γ ∪ ∂+ Γ .
(37)
Here ∂− Γ and ∂+ Γ are the infinite sets of ends determined by the rays
of the negative and the positive types, respectively, and ∂Γ is the space
of ends of the skeleton Γ.
The sets ∂− Γ and ∂+ Γ can be identified with the respective subsets
Γ, to which the corresponding
rays are attached, i.e.,
of the skeleton
1
1
with Z 2 ∩ (0, 1] and with Z 2 ∩ [1, 2), see Section 5.A. Note that
the topology of the union (37) is the same as that of the end compactification of the skeleton Γ (with the only difference that the point 1 ∈ Γ
appears with multiplicity 2).
In what concerns ∂Γ, as a topological space it can be easily identified
(1)
with the space Z2 = Z2 \ 2Z2 of 2-adic integers of norm 1. For doing
that let us first notice that the downward branchings in the “grey”
binary tree B (see Section 5.B) consist in applying the maps
f0 : γ 7→
γ
2
and f1 : γ 7→
γ 1
+ ,
2 2
or, equivalently,
γ
1
+ε
with ε ∈ {0, 1} .
2
2
Therefore, any geodesic ray γ = (γ1 , γ2 , . . . ) in the tree B issued from
the point γ1 = 12 can be encoded by the corresponding sequence ε =
(ε1 , ε2 , . . . ), where εn are uniquely determined by the condition
fε : γ 7→
γn+1 = fεn (γn )
εn εn−1 εn−2
ε2
ε1
1
=
+
+
+ · · · + n−1 + n + n+1 ,
2
4
8
2
2
2
so that
|γn |2 = 2n
∀n ≥ 1 .
n≥1,
Then the sequence
|γn |2 · γn = 1 + ε1 · 2 + ε2 · 4 + · · · + εn−1 · 2n−1
obviously converges in the 2-adic topology to the 2-adic integer
1+
∞
X
n=1
ε n · 2n
THOMPSON’S GROUP F IS NOT LIOUVILLE
29
of norm 1. Conversely, any 2-adic integer of norm 1 gives rise to the
corresponding geodesic γ = (γ1 , γ2, . . . ) in the tree B, i.e., to an end
of Γ.
5.E. Sections of the end bundle. By
Fun(Γ, ∂Γ) ∼
= (∂Γ)Γ
we shall denote the space of ∂Γ-valued configurations on Γ, or, equivalently, of sections of the end bundle
Γ × ∂Γ → Γ ,
(γ, ω) → γ ,
over the graph Γ (cf. the definitions of the hyperbolic boundary bundles
and of the Poisson bundles in author’s papers [Kai04a] and [Kai05], respectively). One can also consider the configurations from Fun(Γ, ∂Γ)
as end fields on Γ. In the same way as the space Fun(Γ, Z) of Z-valued
configurations on Γ, the space Fun(Γ, ∂Γ) is endowed with the left
action (28) of the group Fe.
5.F. Partial isometries of Γ. Generally speaking, groups do not
act on their Schreier graphs by graph automorphisms (here and below
when talking about graph automorphisms or isomorphisms we mean
the maps which preserve the Schreier labelling of edges). In particular, in our case the graph Γ and its skeleton Γ are both rigid, i.e.,
their groups of automorphisms are trivial. However, in spite of that
the graph Γ still has certain symmetry properties sufficient for our purposes. Namely, it has a rather rich set of partial isomorphisms, i.e., of
graph isomorphisms between its subgraphs.
First of all, obviously, all the rays of negative (respectively, of positive) type are pairwise isomorphic. Further, for a tree T and any two
vertices o 6= o′ ∈ T let To→o′ denote the shadow of the vertex o′ as seen
from the vertex o, i.e., the subtree of T which consists of all vertices
x ∈ T such that o′ lies on the geodesic [o, x]. Then the skeleton Γ is selfsimilar in the sense that the shadows Γ1→ 1 and Γ1→γ are isomorphic for
2
any γ ∈ Z 12 ∩ (0, 1) (i.e., for any “grey” vertex γ ∈ Γ). If now Γ1→γ
denotes the subgraph of Γ obtained by reattaching all the negative and
positive type rays to the vertices of the shadow Γ1→γ ⊂ Γ, then the
subgraphs Γ1→γ ⊂ Γ are all pairwise isomorphic for γ ∈ Z 12 ∩ (0, 1).
6. Non-trivial behaviour at infinity determined by the
boundary of the Schreier graph
6.A. Mishchenko’s work. In a recent preprint [Mis15] Mishchenko
gives yet another proof of the transience of the Schreier graph Γ different from the proofs described in Section 3.B and Section 5.C (it
is based on an explicit estimate of the Dirichlet norm on Γ). He further notices that due to a specific geometry of Γ this transience implies
existence of a non-trivial behaviour at infinity (≡ non-triviality of the
30
VADIM A. KAIMANOVICH
Poisson boundary) for the simple random walk on Γ, and therefore for
the simple random walk on Thompson’s group F as well (cf. the discussion of the relationship between the simple random walks on Γ and on
F ∼
= Fe in Section 3.B). Thus, the Poisson boundary of the simple random walk on Thompson’s group F is non-trivial [Mis15, Theorem 2.5],
which provides another proof of our result on the absence of the Liouville property for the group F .
The aforementioned geometric argument in [Mis15] is based on the
following observation (Theorem 2.3), which we shall quote here in a
slightly modified form:
If a tree T has the property that there exists a vertex o ∈ T
of degree at least 2, and such that for any neighbour o′ of
o the shadow To→o′ is transient, then the Poisson boundary
of the simple random walk on T is non-trivial.
In fact, it is classically known (e.g., see Woess [Woe00]) that the
transience of the simple random walk on a connected graph implies
convergence of its sample paths to an end of this graph. Moreover,
the same is true for any bounded range random walk, and, as we have
already established in Theorem 25, the random walk on Γ determined
by any strictly non-degenerate probability measure µ on Fe is transient.
Yet another point is that the boundary behaviour provided by the
above argument from [Mis15] depends on the starting point γ ∈ Γ of
the induced random walk, so that the resulting quotient of the Poisson
boundary is not a µ-boundary. Actually, the quotient of the Poisson
boundary produced by [Mis15, Theorem 2.3] is finite (it can be identified with the set of neighbours of the vertex o), so that it can not
be a µ-boundary already for this reason (for, as we have already mentioned in Remark 36, any non-trivial µ-boundary is purely non-atomic
[Kai95]).
6.B. Non-trivial µ-boundary.
Theorem 38. Let µ be a strictly non-degenerate finitely supported
probability measure on the group Fe. Then for a.e. sample path g = (gn )
of the random walk (Fe, µ) and any point γ ∈ Γ ∼
= Z 12 the sequence
γ.gn converges to an end ωγ = ωγ (g) ∈ ∂Γ, which gives rise to a map
g 7→ (ωγ )γ∈Γ ∈ Fun(Γ, ∂Γ)
from the path space (FeZ+ , P) of the random walk (Fe, µ) to the configuration space Fun(Γ, ∂Γ), and the space Fun(Γ, ∂Γ) endowed with the
arising probability measure λ (the image of the measure P on the path
space under the above map) is a non-trivial µ-boundary.
Proof. As we have already noticed, Theorem 25 in combination with
the fact that the support of µ is finite implies the convergence
γ.gn → ωγ ∈ ∂Γ
for all γ ∈ Γ and a.e. sample path (gn ) .
THOMPSON’S GROUP F IS NOT LIOUVILLE
31
Thus, it only remains to show that the distribution of the configurations
(ωγ )γ∈Γ is not a point measure.
First of all let us notice that if the configuration (ωγ ) were the same
for a.e. sample path (gn ), then it would necessarily have been constant
e
(i.e., taking the same value at all points of Γ),
because the group F acts
1
on Fun(Γ, ∂Γ) by “changing” the Γ ∼
= Z 2 variable, and sgr µ = Fe.
Now let us assume that ωγ ≡ ω ∈ ∂Γ. It means that for any starting point γ ∈ Γ almost every sample path of the random walk (Γ, µ)
(23) converges to ω. Let us denote by Q# the measure on the space
of sample paths of this random walk whose initial distribution is the
counting measure # on Γ.
If ω ∈ ∂− Γ ∪ ∂+ Γ is the end corresponding to a ray of the negative
or of the positive type, then there is a Q# -non-negligible set of sample
paths which are entirely contained in this ray. However, since all the
rays of the same type are isomorphic (see Section 5.F), the same would
be true for any other ray of the same type, which means that with
positive probability the limit end would be different from ω.
If ω ∈ ∂Γ is a skeleton end, then, in the same way, the set of the sample paths entirely contained in the subgraph Γ1→ 1 is Q# -non-negligible.
2
In view of the self-similarity of the graph Γ (see Section 5.F) it implies
that the same would be true for any subgraph Γ1→γ . However, one can
obviously choose γ in such a way that ω ∈
/ ∂Γ1→γ .
The construction of a µ-boundary of the group F on the configuration
space Fun(Γ, Z) from Theorem 35 heavily used the specifics of this
group. On the contrary, the above construction of a µ-boundary on
the space of sections Fun(Γ, ∂Γ) of the end bundle over the Schreier
graph Γ is much more general.
Theorem 39. If a Schreier graph Γ of a finitely generated group G is
transient (with respect to the simple random walk on it), then for any
strictly non-degenerate finitely supported probability measure µ on G
either
(i) there exists an end ω ∈ ∂Γ such that for any γ ∈ Γ a.e. sample
path of the induced random walk (Γ, µ) issued from γ converges
to ω,
or
(ii) the space of sections Fun(Γ, ∂Γ) of the end bundle over Γ is
a non-trivial µ-boundary, so that the Poisson boundary of the
random walk (G, µ) is also non-trivial.
Proof. In the same way as in the proof of Theorem 25, the transience of
the simple random walk on Γ implies the transience of the induced random walk (Γ, µ) for any strictly non-degenerate probability measure µ
on G. Further, if µ is finitely supported, then a.e. sample path of the
induced random walk on Γ converges to a random end ω ∈ ∂Γ. Let us
32
VADIM A. KAIMANOVICH
denote by {κγ }γ∈Γ the family of the corresponding hitting distributions
on the space of ends ∂Γ. By the strict non-degeneracy of µ, all the
measures κγ are pairwise equivalent. Now, in case (i) all measures κγ
coincide with the delta-measure δγ . Otherwise, in case (ii), all measures κγ are not delta-measures, so that Fun(Γ, ∂Γ) endowed with the
arising limit distribution is a non-trivial µ-boundary in the same way
as in Theorem 38.
6.C. Convergence to components of the end space. The key
ingredient of our proof of Theorem 38 was the observation that for a
finitely supported measure µ on Fe the transience of the random walk
(Γ, µ) (23) implies the convergence of a.e. sample path to a random end
of the graph Γ (also see Theorem 39). By using the homomorphisms
χa , χb : F ∼
= Fe → Z
introduced in Section 2.B one can easily describe the components of
the decomposition (37) of the space of ends ∂Γ, on which the hitting
measures κγ of the random walk (Γ, µ) are actually concentrated.
Since the measure µ is finitely supported, the restriction of the random walk (Γ, µ) to any of the negative (respectively, positive) type
rays (see Section 5.A) coincides, outside of a finite neighbourhood
of ray’s origin, just with the usual translation invariant random walk
on Z. Here and below we identify the negative (respectively, positive) type rays in Γ with the negative (respectively, positive) integer ray
Z− = {. . . , −2, −1, 0} (respectively, Z+ = {0, 1, 2, . . . }) of Z. As it
follows from the description of the graph Γ (see Figure 6), the step
distribution of the induced random walk on Z− (respectively, on Z+ )
is (−χa )(µ) (respectively, (−χa − χb )(µ)), where by χ(µ) we denote
the image of the measure µ under a group homomorphism χ : Fe → Z.
Therefore, the drift of the induced random walk on Z− (≡ on the negative type rays) is the barycentre
(−χa )(µ) = −α ,
and the drift of the induced random walk on Z+ (≡ on the positive
type rays) is the barycentre
(−χa − χb )(µ) = −α − β ,
where
α = χa (µ) and β = χb (µ)
denote the barycentres (the expectations) of the measures χa (µ) and
χb (µ) on Z, respectively.
It is classically known that the recurrence properties of the random
walk on Z governed by a finitely supported step distribution σ are
completely determined by the barycentre σ. In particular, the ray Z−
THOMPSON’S GROUP F IS NOT LIOUVILLE
33
(respectively, Z+ ) is transient (i.e., the random walk escapes to infinity along this ray) if and only if the drift σ is negative (respectively,
positive). Therefore, we arrive at the following description of the components of the decomposition (37), on which the hitting measures κγ
are concentrated, in terms of the parameters α, β ∈ R (also see Figure 10):
Proposition 40. Under the conditions of Theorem 38
(i) If α > 0 and α + β < 0, then both the negative type and the
positive type rays are transient, and the hitting measures κγ are
concentrated on the union ∂− Γ ∪ ∂+ Γ;
(ii) If α > 0 and α + β ≥ 0, then only the negative type rays are
transient, and the hitting measures κγ are concentrated on ∂− Γ;
(iii) If a ≤ 0 and α + β < 0, then only the positive type rays are
transient, and the hitting measures κγ are concentrated on ∂+ Γ;
(iv) If a ≤ 0 and α + β ≥ 0, then neither the negative nor the
positive type rays are transient, and the hitting measures κγ are
concentrated on ∂Γ.
β
∂Γ
∂− Γ
α
∂+ Γ
∂− Γ ∪ ∂+ Γ
Figure 10. The domains in the parameter plane (α, β)
corresponding to different supports of the hitting measures κγ .
Proposition 40 allows us to make more precise the description of the
µ-boundary of the random walk (Fe, µ) obtained in Theorem 38.
Theorem 41. Under conditions of Theorem 38 the arising probability
measure λ is concentrated on
Fun(Γ, ∂− Γ ∪ ∂+ Γ) , if α > 0, and α + β < 0 ,
Fun(Γ, ∂ Γ) ,
if α > 0, and α + β ≥ 0
−
Fun(Γ, ∂+ Γ) ,
if a ≤ 0, and α + β < 0
Fun(Γ, ∂Γ) ,
if a ≤ 0, and α + β ≥ 0 .
34
VADIM A. KAIMANOVICH
7. Concluding remarks
7.A. Relaxing conditions on the step distribution. Throughout
the paper we have used the assumption that the measure µ on the group
Fe ∼
= F is finitely supported and strictly non-degenerate. It would be
interesting to see, to what extent our results can be carried over to more
general measures. [Let us remind once again that amenability of the
group F is equivalent to existence of a Liouville symmetric probability
measure µ with supp µ = F .]
Here one can try to relax both the conditions on how much the
measure µ is allowed to spread out:
finite support
↓
finite first moment
↓
finite entropy
↓
general measures
and the conditions on how non-degenerate the measure µ is allowed to
be:
sgr µ = Fe
↓
gr µ = Fe
↓
gr µ is a non-elementary subgroup of Fe .
The meaning of “non-elementary” of course needs to be specified in
our context (for instance, cf. Kaimanovich – Masur [KM96] for the
subgroups of the mapping class group, or Kaimanovich [Kai00] for the
subgroups of groups with hyperbolic properties).
As for the conditions on how much the measure µ is allowed to
spread out, one can expect that an extension to the class of measures
with a finite first moment should be quite feasible (cf. the case of the
lamplighter groups considered by the author [Kai91] and the case of hyperbolic graphs considered by Kaimanovich – Woess [KW92]). Notice,
however, that, once the finite first moment condition is dropped, the
“hyperbolic” and the “lamplighter” models drastically diverge. Boundary convergence in hyperbolic spaces does not require any moment
conditions, see Kaimanovich [Kai00], Maher – Tiozzo [MT14], whereas
for the lamplighter groups, due to their amenability, the situation is
completely different.
THOMPSON’S GROUP F IS NOT LIOUVILLE
35
For instance, there are examples of measures with infinite first moment on the lamplighter groups, for which the configurations do not
stabilize, but the Poisson boundary is still non-trivial, see Kaimanovich
[Kai83, Kai91]. In fact, Erschler proved in [Ers04b] that the Poisson
boundary on the lamplighter groups is non-trivial for all step distributions with finite entropy provided the quotient random walk is transient. She also described another family of groups (which she calls
groups of Baumslag type) with this property. These groups are quite
similar to the lamplighter groups, but the difference is that for the
Baumslag type groups certain infinitely supported configurations are
also allowed.
Outside of the class of measures with finite entropy the situation
becomes even more complex. There are examples of measures µ on the
lamplighter groups, for which the Poisson boundary is trivial in spite of
non-triviality of the Poisson boundary for the reflected measure µ̌(g) =
µ(g −1), see Kaimanovich [Kai83]. This is impossible for the measures
with finite entropy because of the entropy criterion of triviality of the
Poisson boundary, as in the finite entropy case the asymptotic entropies
of the original and of the reflected random walks are obviously the same.
7.B. Other subgroups of P LF (R). It would be interesting to understand, to what extent our results can be carried over to other subgroups
of P LF (R), for instance, to the group generated by the translation
γ 7→ γ + 1 and the transformation defined as
(
γ, γ≤0,
γ 7→
2γ , γ ≥ 0 .
Note that by a result of Brin – Squier [BS85] P LF (R) (and therefore
all its subgroups as well) does not contain non-abelian free subgroups.
As it has been pointed out by the referee, our “drifting” technique
used in the proof of Theorem 13 should be applicable to the “F -series”
of the so-called Thompson–Stein groups (they are a generalization of
Thompson’s groups introduced by Stein [Ste92]). In particular, it almost verbatim carries over to the groups F (p) introduced by Burillo –
Cleary – Stein [BCS01] (for these groups the break points are allowed
to be in Z[ 1p ] for an integer p > 1). An interesting feature of the family
{F (p)} is that all these groups embed one into the other, so that they
are all either amenable or non-amenable simultaneously.
We are not aware of any explicit description of the Schreier graphs of
the canonical action of any of the Thompson–Stein groups other than
Savchuk’s description of the Schreier graphs for the original Thompson’s group F [Sav10b, Sav15].
36
VADIM A. KAIMANOVICH
7.C. The lattice of µ-boundaries and the problem of identification of the Poisson boundary. Although the µ-boundaries constructed in Theorem 35 and Theorem 38 seem to be quite different,
currently we do not have any information about their mutual position
in the lattice of all µ-boundaries. In particular, a priori it is not excluded that both these µ-boundaries might in fact coincide with the
full Poisson boundary (which is the maximal element in the lattice
of µ-boundaries). If not, what are, for instance, their infimum and
supremum? Can one be just a quotient of the other one? And, of
course, how are they related to the full Poisson boundary? Will their
supremum, one of them, or maybe even both be the Poisson boundary? See author’s papers [Kai96, Kai00] for a discussion of the general
identification problem for the Poisson boundary.
7.D. Configuration spaces as µ-boundaries, and generating
partitions of the Poisson boundary. The µ-boundary from Theorem 38 seems to be the first example when a boundary behavior is
realized on the space of configurations on a single group orbit in such a
way that the group action consists of just permuting the configurations
without changing their point values.
There is a well-known construction in ergodic theory which allows
one to realize general group actions as actions on configuration spaces
by translations (e.g., see Cornfeld – Fomin – Sinai [CFS82]). Namely,
let (X, m) be a Lebesgue measure space endowed with a measure class
preserving action of a countable group G. Given an at most countable
measurable partition ξ, let Xξ be the corresponding countable quotient
space (i.e., the space of the elements of the partition ξ), and let x 7→
ξ(x) be the corresponding quotient map which assigns to any point
x ∈ X the element of the partition ξ which contains x. The function
Φx : G → Xξ ,
g 7→ ξ(g −1x) ,
can be then considered as an Xξ -valued configuration on G, and the
map x 7→ Φx is obviously G-equivariant if one endows the space of
configurations on G with the action (4). If the map x 7→ Φx separates
(mod 0) the points of X, i.e., if this map is an isomorphism between
the space (X, m) and the space of Xξ -valued configurations on G, then
the partition ξ is called generating.
Generating partitions for Z-actions are a classical tool of ergodic
theory. Moving to more general actions, it is, for example, easy to see
that if one takes the partition of the boundary of a free group into
the cylinder sets determined by the first letter of the infinite words
representing boundary points, then this partition is generating even in
the topological category. Therefore, this example provides a symbolic
realization of the action of the free group on its Poisson boundary
whenever the Poisson boundary can be identified with the space of
infinite words (for instance, for the step distributions with a finite first
THOMPSON’S GROUP F IS NOT LIOUVILLE
37
moment, see Kaimanovich [Kai00]). However, we are not aware of any
work on generating partitions for Poisson boundaries or their quotients
in general.
7.E. Poisson boundary and sections of boundary bundles over
Schreier graphs. The statement of Theorem 39 on the construction
of a non-trivial µ-boundary for a group G from the space of ends of its
transient Schreier graph Γ is, of course, very well known when one takes
for Γ just the group G (in which case the action of the group G on itself
extends to the space ∂G, and the limit configurations are equivariant
as functions from G to ∂G). However, it would be interesting to have
other examples in the situation when the stabilizers Gγ , γ ∈ Γ, are
sufficiently far from being normal. It is natural to impose the condition
that the Schreier graph is determined by a core-free subgroup H ⊂ G
(i.e., that the action of G on the homogeneous space Γ ∼
= H\G is
faithful), as otherwise one can always pass to the quotient of G by the
normal core of H. Note that for Thompson’s group F ∼
= Fe the latter
condition is obviously satisfied as the canonical action of Thompson’s
group is completely determined by its restriction to the dyadic-rational
orbit (cf. Section 2.C).
Instead of the end bundle one can use the Poisson bundle over the
graph Γ, i.e., the map Γ × ∂µ Γ → Γ, where ∂µ Γ denotes the Poisson
boundary of the induced random walk (Γ, µ) on the graph Γ. The
space of its sections is the space of configurations Fun(Γ, ∂µ Γ). Let
bnd = bnd(Γ,µ) denote the boundary map from the path space of the
random walk (Γ, µ) to the Poisson boundary ∂µ Γ. If ∂µ Γ is non-trivial,
then the projection
g = (gn ) 7→ {bnd(γ.g)}γ∈Γ ,
where g 7→ γ.g is the map (24) from the path space of the random
walk (G, µ) to the path space of the induced random walk (Γ, µ), makes
Fun(Γ, ∂µ Γ) a non-trivial µ-boundary. It would be interesting to find
conditions under which this µ-boundary coincides with the full Poisson
boundary of the random walk (G, µ).
References
[AAMBV13] Gideon Amir, Omer Angel, Nicolás Matte Bon, and Bálint
Virág, The Liouville property for groups acting on rooted trees,
arXiv:1307.5652, 2013.
[AAV13]
Gideon Amir, Omer Angel, and Bálint Virág, Amenability of linearactivity automaton groups, J. Eur. Math. Soc. (JEMS) 15 (2013),
no. 3, 705–730. MR 3085088
[AGS06]
G. N. Arzhantseva, V. S. Guba, and M. V. Sapir, Metrics on diagram
groups and uniform embeddings in a Hilbert space, Comment. Math.
Helv. 81 (2006), no. 4, 911–929. MR 2271228 (2007k:20084)
38
[BCS01]
[BCW07]
[BE11]
[BKN10]
[BLP77]
[Bri09]
[BS85]
[BV05]
[CF11]
[CFP96]
[CFS82]
[Che91]
[ERJvR15]
[Ers04a]
[Ers04b]
[Ers06a]
[Ers06b]
VADIM A. KAIMANOVICH
J. Burillo, S. Cleary, and M. I. Stein, Metrics and embeddings of
generalizations of Thompson’s group F , Trans. Amer. Math. Soc.
353 (2001), no. 4, 1677–1689. MR 1806724
José Burillo, Sean Cleary, and Bert Wiest, Computational explorations in Thompson’s group F , Geometric group theory,
Trends Math., Birkhäuser, Basel, 2007, pp. 21–35. MR 2395786
(2009c:20071)
Laurent Bartholdi and Anna Erschler, Poisson-Furstenberg boundary
and growth of groups, arXiv:1107.5499, 2011.
Laurent Bartholdi, Vadim A. Kaimanovich, and Volodymyr V.
Nekrashevych, On amenability of automata groups, Duke Math. J.
154 (2010), no. 3, 575–598. MR 2730578
Paolo Baldi, Noël Lohoué, and Jacques Peyrière, Sur la classification
des groupes récurrents, C. R. Acad. Sci. Paris Sér. A-B 285 (1977),
no. 16, A1103–A1104. MR 0518008
Jérémie Brieussel, Amenability and non-uniform growth of some directed automorphism groups of a rooted tree, Math. Z. 263 (2009),
no. 2, 265–293. MR 2534118 (2010g:43002)
Matthew G. Brin and Craig C. Squier, Groups of piecewise linear
homeomorphisms of the real line, Invent. Math. 79 (1985), no. 3,
485–498. MR 782231 (86h:57033)
Laurent Bartholdi and Bálint Virág, Amenability via random walks,
Duke Math. J. 130 (2005), no. 1, 39–56. MR 2176547 (2006h:43001)
J. W. Cannon and W. J. Floyd, What is . . . Thompson’s group?,
Notices Amer. Math. Soc. 58 (2011), no. 8, 1112–1113. MR 2856142
J. W. Cannon, W. J. Floyd, and W. R. Parry, Introductory notes on
Richard Thompson’s groups, Enseign. Math. (2) 42 (1996), no. 3-4,
215–256. MR 1426438 (98g:20058)
I. P. Cornfeld, S. V. Fomin, and Ya. G. Sinaı̆, Ergodic theory, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 245, Springer-Verlag,
New York, 1982, Translated from the Russian by A. B. Sosinskiı̆.
MR 832433 (87f:28019)
Mu Fa Chen, Comparison theorems for Green functions of Markov
chains, Chinese Ann. Math. Ser. B 12 (1991), no. 3, 237–242, A
Chinese summary appears in Chinese Ann. Math. Ser. A 12 (1991),
no. 3, 521. MR 1130250
Murray Elder, Andrew Rechnitzer, and Esaias J. Janse van Rensburg, Random sampling of trivial words in finitely presented groups,
Exp. Math. 24 (2015), no. 4, 391–409. MR 3383471
Anna Erschler, Boundary behavior for groups of subexponential
growth, Ann. of Math. (2) 160 (2004), no. 3, 1183–1210. MR 2144977
(2006d:20072)
, Liouville property for groups and manifolds, Invent. Math.
155 (2004), no. 1, 55–80. MR 2 025 301
, Isoperimetry for wreath products of Markov chains and multiplicity of selfintersections of random walks, Probab. Theory Related
Fields 136 (2006), no. 4, 560–586. MR 2257136
, Piecewise automatic groups, Duke Math. J. 134 (2006),
no. 3, 591–613. MR 2254627
THOMPSON’S GROUP F IS NOT LIOUVILLE
[Ers10]
39
, Poisson-Furstenberg boundaries, large-scale geometry and
growth of groups, Proceedings of the International Congress of Mathematicians. Volume II, Hindustan Book Agency, New Delhi, 2010,
pp. 681–704. MR 2827814 (2012h:60016)
[Ers11]
, Poisson-Furstenberg boundary of random walks on wreath
products and free metabelian groups, Comment. Math. Helv. 86
(2011), no. 1, 113–143. MR 2745278 (2011k:60013)
[Fur73]
Harry Furstenberg, Boundary theory and stochastic processes on homogeneous spaces, Harmonic analysis on homogeneous spaces (Proc.
Sympos. Pure Math., Vol. XXVI, Williams Coll., Williamstown,
Mass., 1972), Amer. Math. Soc., Providence, R.I., 1973, pp. 193–
229. MR 50 #4815
[Fur02]
Alex Furman, Random walks on groups and random transformations,
Handbook of dynamical systems, Vol. 1A, North-Holland, Amsterdam, 2002, pp. 931–1014. MR 1928529 (2003j:60065)
[Gou14]
Antoine Gournay, The Liouville property and Hilbertian compression,
arXiv:1403.1195, 2014.
[HHRS15]
Søren Haagerup, Uffe Haagerup, and Maria Ramirez-Solano, A computational approach to the Thompson group F , Internat. J. Algebra
Comput. 25 (2015), no. 3, 381–432. MR 3334642
[HP11]
Uffe Haagerup and Gabriel Picioroaga, New presentations of Thompson’s groups and applications, J. Operator Theory 66 (2011), no. 1,
217–232. MR 2806554 (2012f:46099)
[JMBNdlS15] Kate Juschenko, Nicolás Matte Bon, Monod Nicolas, and Mikael
de la Salle, Extensive amenability and an application to interval exchanges, arXiv:1503.04977, 2015.
[Kai83]
Vadim A. Kaimanovich, Examples of nonabelian discrete groups with
nontrivial exit boundary, Zap. Nauchn. Sem. Leningrad. Otdel. Mat.
Inst. Steklov. (LOMI) 123 (1983), 167–184, Differential geometry,
Lie groups and mechanics, V. MR 697250 (85b:60008)
, An entropy criterion of maximality for the boundary of ran[Kai85]
dom walks on discrete groups, Soviet Math.Dokl. 31 (1985), 193–197.
MR 86m:60025
, Lyapunov exponents, symmetric spaces and a multiplicative
[Kai89]
ergodic theorem for semisimple Lie groups, J. Soviet Math. 47 (1989),
2387–2398. MR 89m:22006
, Poisson boundaries of random walks on discrete solvable
[Kai91]
groups, Probability measures on groups, X (Oberwolfach, 1990),
Plenum, New York, 1991, pp. 205–238. MR 1178986 (94m:60014)
, Measure-theoretic boundaries of Markov chains, 0-2 laws
[Kai92]
and entropy, Harmonic analysis and discrete potential theory (Frascati, 1991), Plenum, New York, 1992, pp. 145–180. MR 94h:60099
[Kai94]
, The Poisson boundary of hyperbolic groups, C. R. Acad. Sci.
Paris Sér. I Math. 318 (1994), no. 1, 59–64. MR 1260536 (94j:60142)
, The Poisson boundary of covering Markov operators, Israel
[Kai95]
J. Math. 89 (1995), no. 1-3, 77–134. MR 96k:60194
[Kai96]
, Boundaries of invariant Markov operators: the identification problem, Ergodic theory of Zd actions (Warwick, 1993–1994),
London Math. Soc. Lecture Note Ser., vol. 228, Cambridge Univ.
Press, Cambridge, 1996, pp. 127–176. MR 97j:31008
40
[Kai00]
[Kai04a]
[Kai04b]
[Kai05]
[KL07]
[KM96]
[KM99]
[KV83]
[KV15]
[KW92]
[KW07]
[Lio08]
[LP15]
[Lyo87]
[MB14]
[Mis15]
[Moo13]
[MT14]
VADIM A. KAIMANOVICH
, The Poisson formula for groups with hyperbolic properties, Ann. of Math. (2) 152 (2000), no. 3, 659–692. MR 1815698
(2002d:60064)
, Boundary amenability of hyperbolic spaces, Discrete geometric analysis, Contemp. Math., vol. 347, Amer. Math. Soc., Providence, RI, 2004, pp. 83–111. MR 2077032 (2005j:20051)
, Boundary behaviour of the Thompson group, presentation
slides, 2004.
, Amenability and the Liouville property, Israel J. Math.
149 (2005), 45–85, Probability in mathematics. MR 2191210
(2007c:43001)
Anders Karlsson and François Ledrappier, Linear drift and Poisson
boundary for random walks, Pure Appl. Math. Q. 3 (2007), no. 4,
part 1, 1027–1036. MR 2402595
Vadim A. Kaimanovich and Howard Masur, The Poisson boundary
of the mapping class group, Invent. Math. 125 (1996), no. 2, 221–264.
MR 1395719 (97m:32033)
Anders Karlsson and Gregory A. Margulis, A multiplicative ergodic
theorem and nonpositively curved spaces, Comm. Math. Phys. 208
(1999), no. 1, 107–123. MR 1729880 (2000m:37031)
V. A. Kaimanovich and A. M. Vershik, Random walks on discrete
groups: boundary and entropy, Ann. Probab. 11 (1983), no. 3, 457–
490. MR 85d:60024
Mihal Kotowski and Bálint Virág, Non-Liouville groups with return
probability exponent at most 1/2, Electron. J. Probab. 20 (2015), no.
12, 1–12 (electronic).
Vadim A. Kaimanovich and Wolfgang Woess, The Dirichlet problem
at infinity for random walks on graphs with a strong isoperimetric
inequality, Probab. Theory Related Fields 91 (1992), no. 3-4, 445–
466. MR 1151805 (93e:60139)
Anders Karlsson and Wolfgang Woess, The Poisson boundary of
lamplighter random walks on trees, Geom. Dedicata 124 (2007), 95–
107. MR 2318539 (2009b:60246)
Isabelle Liousse, Rotation numbers in Thompson-Stein groups and
applications, Geom. Dedicata 131 (2008), 49–71. MR 2369191
Russell Lyons and Yuval Peres, Poisson boundaries of lamplighter groups: proof of the Kaimanovich-Vershik conjecture,
arXiv:1508.01845, 2015.
Terry Lyons, Instability of the Liouville property for quasi-isometric
Riemannian manifolds and reversible Markov chains, J. Differential
Geom. 26 (1987), no. 1, 33–66. MR 892030 (88k:31012)
Nicolás Matte Bon, Subshifts with slow complexity and simple groups
with the Liouville property, Geom. Funct. Anal. 24 (2014), no. 5,
1637–1659. MR 3261637
Pavlo Mishchenko, Boundary of the action of Thompson’s group F
on dyadic numbers, arXiv:1512.03083, 2015.
Justin Tatch Moore, Fast growth in the Følner function for Thompson’s group F , Groups Geom. Dyn. 7 (2013), no. 3, 633–651.
MR 3095713
Joseph Maher and Giulio Tiozzo, Random walks on weakly hyperbolic
groups, arXiv:1410.4173, 2014.
THOMPSON’S GROUP F IS NOT LIOUVILLE
[NP08]
[Ros81]
[Sav10a]
[Sav10b]
[Sav15]
[Ste92]
[Var83]
[Woe94]
[Woe00]
41
Assaf Naor and Yuval Peres, Embeddings of discrete groups and the
speed of random walks, Int. Math. Res. Not. IMRN (2008), Art. ID
rnn 076, 34. MR 2439557
Joseph Rosenblatt, Ergodic and mixing random walks on locally compact groups, Math. Ann. 257 (1981), no. 1, 31–42. MR 83f:43002
Ecaterina Sava, A note on the Poisson boundary of lamplighter random walks, Monatsh. Math. 159 (2010), no. 4, 379–396. MR 2600904
(2011b:60308)
Dmytro Savchuk, Some graphs related to Thompson’s group
F , Combinatorial and geometric group theory, Trends Math.,
Birkhäuser/Springer Basel AG, Basel, 2010, pp. 279–296.
MR 2744025 (2012b:20104)
, Schreier graphs of actions of Thompson’s group F on the
unit interval and on the Cantor set, Geom. Dedicata 175 (2015),
355–372. MR 3323646
Melanie Stein, Groups of piecewise linear homeomorphisms, Trans.
Amer. Math. Soc. 332 (1992), no. 2, 477–514. MR 1094555
Nicolas Th. Varopoulos, Brownian motion and transient groups,
Ann. Inst. Fourier (Grenoble) 33 (1983), no. 2, 241–261. MR 699497
Wolfgang Woess, Topological groups and recurrence of quasitransitive
graphs, Rend. Sem. Mat. Fis. Milano 64 (1994), 185–213 (1996).
MR 1397471 (97i:60092)
, Random walks on infinite graphs and groups, Cambridge
Tracts in Mathematics, vol. 138, Cambridge University Press, Cambridge, 2000. MR 2001k:60006
Department of Mathematics and Statistics, University of Ottawa,
585 King Edward, Ottawa ON, K1N 6N5, Canada
E-mail address: [email protected], [email protected]
| 4math.GR
|
Synkhronos: a Multi-GPU Theano Extension for
Data Parallelism
arXiv:1710.04162v1 [cs.DC] 11 Oct 2017
Adam Stooke
University of California, Berkeley
[email protected]
Pieter Abbeel
University of California, Berkeley
[email protected]
Abstract
We present Synkhronos, an extension to Theano for multi-GPU computations
leveraging data parallelism. Our framework provides automated execution and
synchronization across devices, allowing users to continue to write serial programs
without risk of race conditions. The NVIDIA Collective Communication Library
is used for high-bandwidth inter-GPU communication. Further enhancements to
the Theano function interface include input slicing (with aggregation) and input
indexing, which perform common data-parallel computation patterns efficiently.
One example use case is synchronous SGD, which has recently been shown to
scale well for a growing set of deep learning problems. When training ResNet-50,
we achieve a near-linear speedup of 7.5x on an NVIDIA DGX-1 using 8 GPUs,
relative to Theano-only code running a single GPU in isolation. Yet Synkhronos
remains general to any data-parallel computation programmable in Theano. By
implementing parallelism at the level of individual Theano functions, our framework uniquely addresses a niche between manual multi-device programming and
prescribed multi-GPU training routines.
1
Introduction
Theano [1] is the classic auto-differentiation Python package. It translates user’s computation
expressions into optimized code for fast CPU or GPU execution, and deep learning is among its most
common uses. Yet it lacks built-in support for multi-device parallel programming,1 a clear means to
faster computing. While significant speedups in multi-core CPU execution are possible,2 we focus on
use of GPUs.
1.1
Other Frameworks
More recent auto-differentiation frameworks do offer multi-GPU programming in Python, under
data or model parallelism. For example, in Tensorflow [2], perhaps the most similar to Theano,
users can program “multi-tower” computations3 with more or less arbitrary device placements for
computations. Other successful packages, such as PyTorch [3], Chainer [4], and MXNet [5] now
also include tools to automatically use multiple GPUs in training neural networks. Specifically, these
tools leverage the fact that SGD and its variants exhibit data parallelism, a simple and powerful motif
for scaling [6, 7]. Data parallelism simply requires that a computation can correctly be performed by
reducing (or gathering) the results of independent calls over arbitrary data subdivisions, and it has
wide applicability to computations amenable to GPU acceleration [8].
1
support for multi-GPU model parallelism is still experimental
In our experience with some BLAS routines (called by Theano), calling separate single-threaded routines on
each core in a data-parallel fashion can out-perform the multi-threaded routine using all cores on the full dataset.
3
https://www.tensorflow.org/tutorials/using_gpu
2
1.2
Our Framework
Although it shares many elements with those examples, our framework offers perhaps a unique
blend of generality and ease of use. It does so by implementing a level of abstraction that addresses
the gap left between manual programming of individual devices and pre-fixed multi-GPU training
routines. Synkhronos4 automatically coordinates multiple devices to operate as if one, but in a way
fully general to any data-parallel computation programmable in Theano.
Platoon5 is one previous Theano extension supporting asynchronous multi-GPU computation, wherein
workers are programmed separately from the master thread. In contrast, our framework implements
parallel execution automatically at the level of individual Theano functions, which prescribes synchronicity. This dramatically simplifies user development. Usage primarily entails a few line-for-line
code exchanges from Theano, and the user’s programming responsibilities remain entirely within the
original (serial) program.
At the same time, our framework is designed with scaling performance in mind. The multi-processed
implementation ensures concurrent operation6 for high numbers of devices, without requiring MPI.
Inclusion of a Numpy-like shared memory interface (also shareable with other Python processes)
and use of the NVIDIA Collective Communication Library (NCCL) combine to enable nimble
management of both GPU and CPU memories. Lastly, computations are organized to minimize
GPU-to-CPU transfers, wherever possible.
1.3
Paper Organization
In this paper we describe our framework from a systems perspective, as follows. First, relevant
background on Theano’s interface is provided, leading to responsibilities for Synkhronos. Then, we
describe the techniques used for employing multiple devices, including matters of setup, synchronization, and memory management. We also introduce additional interface features useful for common
data parallel compute patterns, including automated input slicing to avoid out-of-memory errors and
input indexing for efficient memory sharing and shuffling. Finally, we investigate performance results
in a representative case study before concluding with our future outlook. Readers already familiar
with Theano or other frameworks may wish to view directly the example code comparison provided
in Appendix A.
2
Background: Design Requirements
To formulate the functionalities required of Synkhronos, we first review relevant features of the
Theano interface it must interact with: functions and shared variables.
Theano Functions After constructing symbolic computation expressions using Theano, the user
builds functions by specifying the expressions to be computed–and the corresponding input variables–
to Theano’s “function” method. Theano compiles optimized code (for CPU or GPU), which the user
later executes by calling the function on data inputs. Example functions may compute a forward pass
through a neural network, or a gradient over a minibatch of inputs and labels.
Theano Shared Variables A Theano shared variable has a data array associated with it and can
serve as an implicit function input. Two reasons for using them include: 1) the data persists across
different functions, and 2) the data can be held on the GPU, eliminating latency of transmission from
the CPU during function calls. One common use is for the weights of a neural network layer.
Theano Updates Aside from returning outputs, Theano functions can also modify shared variables,
in an operation termed an “update”. It is common practice, for example, to build a function which
computes a gradient and immediately uses it to update network parameters in place.
Synkhronos Requirements In sum, Synkhronos must prepare Theano functions for each GPU
and make the relevant shared variables available on each device. Then it must provide means to 1)
4
https://github.com/astooke/Synkhronos
https://github.com/mila-udem/platoon
6
i.e., in the presence of Python’s Global Interpreter Lock
5
2
Figure 1: Synkhronos overview: automated worker management for synchronous computations and
communications.
run functions on all GPUs simultaneously and reduce the results and 2) modify each device’s shared
variables, including translating local updates into global ones.
3
Synkhronos Program Flow
Given what functionality Synkhronos must perform relative to Theano, we now describe our design
decisions by way of a program overview. Data parallelism and synchronicity are two defining
characteristics simplifying the design. Overall functionality is summarized in Figure 1, and brief
example code appears in Appendix A. Discussion on data management is deferred to the following
section.
3.1
Computation Setup
Fork Synkhronos automatically forks a separate Python process for each additional GPU. A barrier
across all processes guards the start and finish of any user call requiring action by workers, preventing
any race conditions. The user’s process acts as master and participates as a parallel worker; it remains
available for single-GPU use through Theano.
3
Build Symbolic Expressions After forking, the user builds expressions as usual with Theano (or
other extensions, e.g. Lasagne [9] for neural networks), and shared variables are thereby allocated on
the master GPU. This can comprise a majority of the user’s overall Theano code–Synkhronos leaves
it untouched.
Build Functions Synkhronos presents a “function” method, replacing Theano’s. Internally, it
builds underlying Theano functions. The main difference to the user is the ability to specify a
reduce/gather operation to use for each output. Thereafter, all functions are distributed to the worker
processes using Theano’s function serialization, via pickling. When unpickling, all involved shared
variables are automatically replicated on every GPU.
3.2
Running Computations
Function Calling Following distribution, the user calls Synkhronos functions just as Theano
functions, but now all GPUs participate in computation. A function call induces the sequence: 1)
data inputs are scattered equally (as possible) across workers, 2) each device calls the same Theano
function on its assigned data, and 3) results are reduced or gathered back to the master process and
returned to the user. For program clarity, function updates to shared variables are applied only locally
within each GPU.
Shared Variable Management A new element beyond single-device Theano programs is the
management of (replicated) shared variables across multiple memories. Synkhronos provides several
MPI-like colectives, such as broadcast and all-reduce, which can use NCCL (via the Pygpu package)
for high-bandwidth inter-GPU communication (including NVLink). CPU-based collectives are also
included, along with the means to get and set values on any individual GPU.
3.3
An Example: Synchronous SGD
Using SGD as an example, the computation sequence could be as follows: 1) a first Synkhronos
function takes in data and computes the gradient, storing it in a shared variable (local on each GPU),
2) this variable is all-reduced using Synkhronos, and finally 3) a second Synkhronos function applies
the update rule in each GPU using the combined gradient values. For convenience, such an adaptation
of all update rules included in Lasagne is provided , namely momentum [10], RMSProp [11], and
Adam [12], among others. These store the gradients of all variables into one (flattened) array for
faster inter-GPU communication.
4
Synkhronos Data Management
Synkhronos must transfer users’ data inputs, typically in the form of Numpy arrays, from the user
process to the worker processes. For fast performance, memory copies (within the CPU) and memory
transfers (CPU-GPU) should be kept to a minimum. Synkhronos includes special data objects and
communication collectives for this purpose.
4.1
Synkhronos Data Objects (for Function Inputs)
Input data is communicated to worker processes using operating system shared memory,7 exposed
to the user in a special data object. Every process has equal read-write access, so workers can
concurrently feed their Theano functions by excerpting their assigned data in parallel. Synkhronos
wraps each shared memory allocation in the Numpy array interface, allowing full array indexing for
user reading and writing. Writing an entire data set (or a large chunk) into such an array obviates
the need for any future memory copies, with the use of input indexing (see Section 5.2) if needed,
e.g. for minibatches. The lowest tensor dimension is taken to represent independent data points for
scattering inputs.8 Inputs designated for broadcast are simply used as is.
7
It
is
similar
to
a
Multiprocessing
RawArray
(see
section
17.2.2.6.1
at
https://docs.python.org/3.5/library/multiprocessing.html, except allocatable before or
after forking.
8
Under C-style memory layout this provides contiguous memory assignments.
4
The size of the underlying memory allocation can be made greater than the size of the outward facing
Numpy array, preventing reallocation when growing or shrinking an array later. Special methods
are provided for reshaping these arrays and freeing their memory. Conveniently, the Numpy array
sub-object can be passed to other user Python processes to be read/written as shared memory. In sum,
the aim is to present an interface which allows the user to optimize memory performance given the
multi-process context.
4.2
Scattering to GPUs (for Shared Variables)
Programs with re-used inputs are often more efficient when data is stored on the GPU over multiple
function calls, and the increased aggregate memory of multiple GPUs makes this feasible in a greater
number of cases.9 Synkhronos helps with its “scatter” collective, which evenly divides input arrays
into shared variable storage across GPUs. It uses the same scattering scheme as for explicit inputs,
i.e. by first tensor dimension. This is a convenient one-line replacement for setting multiple device
memories manually (also possible), and it can make use of Synkhronos data objects and input
indexing.
5
Synkhronos Extensions to Theano Function Interface
Synkhronos includes two extensions to the Theano interface for calling functions. These support
common data-parallel compute patterns, simplifying user code and improving program efficiency.
5.1
Automated Input Slicing & Aggregation
The first extension is automated input data slicing, which can be used to avoid out-of-memory errors
during computations too large to do in one call to a device. When slicing is used, workers compute
their results by aggregating over multiple calls to the underlying Theano function, each using a subset
of the worker’s assigned data. Slice results are aggregated in-place on the GPU. Worker results are
reduced once back to the master process. If the function includes updates, all slices are computed
using the original values, with updates accumulated and applied only once at the end. Automated
slicing also applies to implicit inputs, i.e. data stored in advance on the GPU in shared variables
(the memory requirement during computation is often much larger than the stored data). The user
provides the number of slices to use as an optional input at each function call.
5.2
(Parallel) Input Indexing
The other extension is input indexing, whereby the user can specify which data elements to use
during a function call. Designated indexes can make a slice or can be a list of (random) indexes.
This is efficient for shuffling, which necessitates a memory copy, as each work can excerpt its own
share of the indexes in parallel from the shared memory; no excess memory copies occur. Similar
functionality is provided for inputs stored on-GPU in shared variables, where the same or different
indexes can be applied to each device. Automated input slicing applies internally in each worker,
after input indexing has determined assignments.
6
Scaling Performance
We explored an example case study to answer whether Synkhronos can achieve linear scaling on a
modern supervised learning problem. We timed the training of a ResNet-50 [13] model using SGD
on an NVIDIA DGX-1 (8x Tesla P100 GPUs, NVLink). Several gigabytes of synthetic images were
generated as random single-precision arrays of dimension 3x224x224, and each timing run consisted
of the same number of epochs over the entire data set, lasting at least several minutes.
The base case was a Theano-only program running on a single GPU with the rest of the machine
idle, using a batch size of 64 images. We ran separate base timings for data shuffling or not. Results
from several multi-GPU configurations are shown in Figure 2, all obtained using Synkhronos. No
9
Our original motivation was a big-batch, hessian-free learning algorithm, which achieved super-linear
speedup by fitting each batch across multiple GPUs.
5
input slicing was used, but input indexing was, and all input data was passed as explicit inputs to the
function call (i.e., not stored in advance on the GPU). The network update scheme was as described
in Section 3.3.
Figure 2: Speedups of ResNet-50 training relative to single-GPU, Theano-only program, with
ablations.
6.1
End-to-End Training Timing
Batch - 64 The first test explored the scaling while keeping the algorithm perfectly unchanged.
The total batch size remained fixed at 64, so the amount of data processed by each GPU in each
minibatch scaled down with the number of GPUs. This yielded sub-linear scaling above 2 GPUs, as
GPU utilization decreased, with a speedup only slightly greater than 5x when using all 8 GPUs.
Batch - 64 × #GPU The remainder of the tests used an algorithm modified by scaling up the minibatch size so that each GPU processes 64 images at every call, following the technique successfully
applied in [14] to train to convergence on ImageNet in under one hour. This improved the scaling of
the fully-communicating algorithm up to 7.5x with shuffling and 7.6x without, as compared against
their own respective base cases–nearly linear, as seen in Figure 3. The base Theano code processed
1.75 minibatches per second, giving training speeds for 1-GPU Theano and 8-GPU Synkhronos of
roughly 110 and 830 images per second, respectively.
As shown in Figure 2, turning off the gradient all-reduction improved scaling only slightly, to 7.6x
with shuffling and nearly 7.7x without. This confirmed a small cost of communication compared to the
computation in this scenario (note that Synkhronos permits any intermediate level of communication
frequency).
Another view is to measure using the 2-GPU case as the baseline, i.e. as 2x speed, which highlights
scaling alone by de-emphasizing any fixed overhead in Synkhronos. Under this comparison, both
the shuffle and non-shuffle algorithms achieved >7.8x speeds using 8 GPUs. This indicates high
likelihood for good scaling up to greater numbers of GPUs.
6.2
Detailed Profiling
Lastly, we examined the main use case in more detail in an attempt to understand scaling imperfections.
These tests included full communication, data shuffling, and batch scaling, and they used CUDA
launch blocking to enforce CPU-GPU synchronization for proper profiling. Table 1 contains a
summary of the training loop’s main elements and their timings in a short run, comparing Theanoonly code against 8-GPU Synkhronos.
Starting from the bottom of the table, the all-reduce provided through NCCL added less than one
percent in overhead. The exact timing of shuffling and computing naturally varied from call to call,
leading to a slight slowdown from the straggler effect, in this case roughly 2%. Shuffling scaled
well, to 7.9x, leaving some uncertainty as to the source of diminished scaling in the end-to-end tests.
6
Figure 3: Speedup (near-linear) of ResNet-50 training relative to single-GPU, Theano-only program;
with batch scaling.
Further profiling gathered through Theano indicated a 1.8% time overhead for CPU-to-GPU data
transfer within the Theano function. This increased to 3.6% in the 8 GPU case, as expected on the
given hardware, accounting for much of the loss to 7.7x in the Theano function calls. Overall, we
measured near-linear scaling of 7.5x, matching the sum of these components’ contributions.
Table 1: Detailed profiling, ResNet-50 SGD training (CUDA launch blocking enabled)
Time (s)
Item
Theano
Synk-8
Scaling
395.1
382.5
12.6
–
–
52.6
49.5
1.6
1.0
0.46
7.5
7.7
7.9
–
–
Total
Theano Function
Shuffle
Straggler Effect
All-Reduce Gradient
7
Conclusion
We have presented Synkhronos, an extension to Theano for computing with multiple devices under
data parallelism. After detailing the framework and functionality, we demonstrated near-linear
speedup on a relevant deep learning example, training ResNet-50 with 8 GPUs on a DGX-1. The
design emphasizes easy migration from single- to multi-device programming by the user while
maintaining full generality to any data parallel computation. It includes flexible tools for efficient
memory management, yet currently remains limited to single-node computing. Lastly, because it is
written entirely in Python, the package is more widely accessible to modification by interested users.
We refer to the code repository10 for further examples. We hope that this package will accelerate
the work of researchers and developers who use it, and that it may contribute helpful concepts for
multi-device interfaces for other performance computing frameworks going forward.
Acknowledgments
We thank Frédéric Bastien, Pascal Lamblin, and Arnaud Bergeron for discussions and pointers
throughout development, and the entire Theano development team for their base of work. Thanks
10
https://github.com/astooke/Synkhronos
7
again to Frédéric and to Carlos Florensa, Peter (Xi) Chen, and Anushree Saxena for helpful comments
on the manuscript. The DGX-1 used for this research was donated by the NVIDIA Corporation.
Adam Stooke gratefully acknowledges the support the Fannie and John Hertz Foundation.
References
[1] Theano Development Team. Theano: A Python framework for fast computation of mathematical
expressions. arXiv e-prints, abs/1605.02688, May 2016. URL http://arxiv.org/abs/1605.
02688.
[2] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL https:
//www.tensorflow.org/. Software available from tensorflow.org.
[3] Pytorch Development Team. Pytorch. URL https://github.com/pytorch/pytorch.
[4] Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open
source framework for deep learning. In Proceedings of Workshop on Machine Learning Systems
(LearningSys) in The Twenty-ninth Annual Conference on Neural Information Processing Systems (NIPS), 2015. URL http://learningsys.org/papers/LearningSys_2015_paper_
33.pdf.
[5] Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing
Xu, Chiyuan Zhang, and Zheng Zhang. Mxnet: A flexible and efficient machine learning
library for heterogeneous distributed systems. CoRR, abs/1512.01274, 2015. URL http:
//arxiv.org/abs/1512.01274.
[6] W. Daniel Hillis and Guy L. Steele, Jr. Data parallel algorithms. Commun. ACM, 29(12):
1170–1183, December 1986. ISSN 0001-0782. doi: 10.1145/7902.7903. URL http://doi.
acm.org/10.1145/7902.7903.
[7] Guy E Blelloch. Vector models for data-parallel computing, volume 356. MIT press Cambridge,
1990.
[8] John Nickolls, Ian Buck, Michael Garland, and Kevin Skadron. Scalable parallel programming
with cuda. Queue, 6(2):40–53, March 2008. ISSN 1542-7730. doi: 10.1145/1365490.1365500.
URL http://doi.acm.org/10.1145/1365490.1365500.
[9] Sander Dieleman, Jan Schlüter, Colin Raffel, Eben Olson, Søren Kaae Sønderby, Daniel Nouri,
et al. Lasagne: First release., August 2015. URL http://dx.doi.org/10.5281/zenodo.
27878.
[10] Yurii Nesterov. A method of solving a convex programming problem with convergence rate
O(1/sqr(k)). Soviet Mathematics Doklady, 27:372–376, 1983. URL http://www.core.ucl.
ac.be/~{}nesterov/Research/Papers/DAN83.pdf.
[11] T. Tieleman and G. Hinton. RMSprop Gradient Optimization. URL http://www.cs.toronto.
edu/~{}tijmen/csc321/slides/lecture_slides_lec6.pdf.
[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR,
abs/1412.6980, 2014. URL http://arxiv.org/abs/1412.6980.
[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. CoRR, abs/1512.03385, 2015. URL http://arxiv.org/abs/1512.03385.
[14] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola,
Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training
imagenet in 1 hour, 2017.
8
A
Code Example
In the figures below we present a simple code example for running stochastic gradient descent in
Theano, which is revised into Synkhronos. In the Theano example, the functions build_cnn() and
setup_training() will contain other Theano code (or possibly Lasagne), which is left untouched
in the Synkhronos program. The fork() command automatically uses all GPUs if not otherwise
specified, and the distribute() command replicates all functions and Theano shared variables on
the worker GPUs. Remaining line numbers where Synkhronos is active are highlighted. The data()
command is used here to write the training data to operating system shared memory (once). The
main change in the training loop is the call to all-reduce the network parameters, which in the case of
simple SGD preserves the algorithm. Further examples and demonstrations can be found at the code
repository.
Figure 4: Theano program for single-GPU SGD.
Figure 5: Synkhronos program for multi-GPU SGD.
9
| 2cs.AI
|
Faster Algorithms for RNA-folding using the
Four-Russians method
Balaji Venkatachalam, Dan Gusfield, and Yelena Frid
arXiv:1307.7820v1 [q-bio.QM] 30 Jul 2013
Department of Computer Science, UC Davis
{balaji, gusfield, yafrid}@cs.ucdavis.edu
Abstract. The secondary structure that maximizes the number of noncrossing matchings between complimentary bases of an RNA sequence of
length n can be computed in O(n3 ) time using Nussinov’s dynamic programming algorithm. The Four-Russians method is a technique that will
reduce the running time for certain dynamic programming algorithms
by a multiplicative factor after a preprocessing step where solutions to
all smaller subproblems of a fixed size are exhaustively enumerated and
n3
solved. Frid and Gusfield designed an O( log
) algorithm for RNA folding
n
using the Four-Russians technique. In their algorithm the preprocessing
is interleaved with the algorithm computation.
We simplify the algorithm and the analysis by doing the preprocessing
once prior to the algorithm computation. We call this the two-vector
method. We also show variants where instead of exhaustive preprocessing, we only solve the subproblems encountered in the main algorithm
once and memoize the results. We give a simple proof of correctness and
explore the practical advantages over the earlier method.
The Nussinov algorithm admits an O(n2 ) time parallel algorithm. We
show a parallel algorithm using the two-vector idea that improves the
n2
time bound to O( log
).
n
We have implemented the parallel algorithm on graphics processing units
using the CUDA platform. We discuss the organization of the data structures to exploit coalesced memory access for fast running times. The ideas
to organize the data structures also help in improving the running time
of the serial algorithms. For sequences of length up to 6000 bases the
parallel algorithm takes only about 2.5 seconds and the two-vector serial
method takes about 57 seconds on a desktop and 15 seconds on a server.
Among the serial algorithms, the two-vector and memoized versions are
faster than the Frid-Gusfield algorithm by a factor of 3, and are faster
than Nussinov by up to a factor of 20.
1
Introduction
Computational approaches to find the secondary structure of RNA molecules are
used extensively in bioinformatics applications. The classic dynamic programming (DP) algorithm proposed in the 1970s has been central to most structure
prediction algorithms. While the objective of the original algorithm was to maximize the number of non-crossing pairings between complementary bases, the
dynamic programming approach has been used for other models and approaches,
including minimizing the free energy of a structure. The DP algorithm runs in
cubic time and there have been many attempts at improving its running time.
Here, we use the Four-Russians method for speeding up the computation.
The Four-Russians method, named after Aralazarov et al. [4], is a method to
speed up certain dynamic programming algorithms. In a typical Four-Russians
algorithm there is a preprocessing step that exhaustively enumerates and solves
a set of subproblems and the results are tabled. In the main DP algorithm,
instead of filling out or inspecting individual cells, the algorithm takes longer
strides in the table. The computation for multiple cells is solved in constant
time by utilizing the preprocessed solutions to the subproblems. The longer
strides to fill the table reduce the runtime by a multiplicative factor. The size
of the subproblems is chosen in a way that does not make the preprocessing too
expensive.
Frid and Gusfield [11] showed the application of the Four-Russians approach
for RNA folding. In their algorithm, the preprocessing is interleaved with the
algorithm computation. They fill out a part of the DP table and use these entries
to complete a part of the preprocessing. The preprocessed entries are used later
in the computation.
We show a simpler algorithm where all the preprocessing is completed before
the start of the main algorithm. This simplifies the correctness proof and the
runtime analysis. This approach helps in obtaining a log n factor improvement
for the parallel algorithm. In comparing various methods for RNA folding, Zakov
and Frid (personal communication) had independently observed that the algorithm in [11] could be modified to do the preprocessing at once. It is essentially
the idea as described here.
In this paper we explore the implications of the one-pass preprocessing idea.
This description of the algorithm leads naturally to two other variants. We empirically evaluate these variants and also the implementation of the parallel algorithm.
The parallel architecture of general-purpose graphical processing units (GPUs)
have been exploited for many real-world application in addition to applications
in gaming and visualization problems. GPUs have also been used to speed up
RNA folding algorithms [6,23,24]. Here we show how the Four-Russians method
allows an organization of the data structures for fast memory accesses. We also
describe the organization of the parallel hierarchy to exploit the inherent parallelism of the solution.
In the rest of the section, we describe the problem in relation to the other
problems in RNA folding. To keep the paper self-contained, we will first describe the two-vector algorithm, our application of the Four-Russians method to
the RNA folding problem. We will use that description to describe the original
Four-Russians method for RNA folding by Frid and Gusfield [11]. This discussion leads to two other variants where the preprocessing is done on demand,
instead of the exhaustive preprocessing in the two-vector method and the FridGusfield algorithm. In section 4 we discuss the O(n2 / log n) parallel algorithm.
We will then describe the implementation of a parallel algorithm using CUDA.
The final sections have discussion on empirical observations and conclusions. Due
to space limitations, this manuscript focuses mostly on the theoretical aspects
and describes the experimental results briefly. Detailed discussion can be found
in [26].
Related work The O(n3 ) dynamic programming algorithm due to Nussinov et
al. [21,20] maximizes the number of non-crossing matching complimentary bases.
There have been many methods since Zuker and Stiegler [31] that infer the
folding using thermodynamic parameters [25,19] which are more realistic than
maximizing the number of base pairs. These methods have been implemented in
many packages including UNAFold [18], Mfold [30], Vienna RNA Package [15],
RNAstructure [22].
Probabilistic methods include stochastic context-free grammars [10,9], the
maximum expected accuracy (MEA) method, where secondary structures are
composed of pairs that have a maximal sum of pairing probabilities, eg., MaxExpect [17], Pfold [16], CONTRAfold [8] which maximize the posterior probabilities of base pairs; and Sfold [7], CentroidFold [14] that maximize the centroid
estimator. There are also other methods that use a combination of thermodynamic and statistical parameters [2] and methods that use training sets of known
folds to determine their parameters, eg., CONTRAfold [8], and Simfold[3] and
ContextFold[28].
In addition to the Four-Russians method, other methods to improve the
running time include Valiant’s max-plus matrix multiplication by Akutsu [1]
and Zakov et al. [29]; and sparsification, where the branch points are pruned to
get an improved time bound [27,5].
CUDA, the programming platform for GPGPUs, has been used to solve
many bioinformatics problems. Chang, Kimmer and Ouyang [6] and Stojanovski,
Gjorgjevikj and Madjarov [24] show an implementation of the Nussinov algorithm on CUDA. Rizk et al. [23] describe the implementation for Zuker and
Stiegler method involving energy parameters. These methods are discussed in
section 5.2.
2
The Nussinov Algorithm
In this paper, we consider the basic RNA folding problem of maximizing the number of non-crossing complimentary base pair matchings. Complimentary bases
can be paired, i.e., A with U and C with G. A set of disjoint pairs is a matching. The pairs in a matching must not cross, i.e., if bases in positions i and
j are paired and if bases k and l are paired, then either they are nested, i.e.,
i < k < l < j or they are non-intersecting, i.e., i < j < k < l. The objective is
to maximize the number of pairings under these constraints.
The following algorithm, due to Nussinov [21] maximizes the number of noncrossing matchings. For an input sequence S of length n over the alphabet A,
C, G, U, the recurrence is defined as follows. Let D(i, j) denote the optimal cost
of folding for the subsequence from i to j. For all i, D(i, i − 1) = D(i, i) = 0 and
for all i < j:
b(S(i), S(j)) + D(i + 1, j − 1)
D(i, j) = max
(1)
maxi+1≤k≤j D(i, k − 1) + D(k, j)
where b(., .) = 1 for complimentary bases and 0 otherwise. The DP table is the
upper triangular part of the n × n matrix. The optimal solution is given by
D(1, n). The table can be filled column-wise from the first column till the nth .
There are other ways of filling the table too, eg., along the diagonals — the (i, i)diagonal first, (i, i + 1)-diagonal next and so on, until the last diagonal with one
entry, D(1, n). To allow for traceback we need to store the bases that are paired
to get the maximum value. Let D∗ (i, j) denote the corresponding indices. These
are obtained by substituting arg max in place of max in the above recurrence
and can be computed along with the max value.
The first part of the recurrence can be solved in constant time. The second
part is more expensive, incurring Θ(n) look ups and maximum computations.
There are O(n2 ) entries in the DP table and each cell can be computed in O(n)
time, giving an O(n3 ) time algorithm.
3
The Four-Russians Algorithms
In this section we discuss three variants of the Four-Russians algorithm. We will
first describe the two-vector approach. Since it is simpler than the other methods
we will use the description to discuss two other variants.
3.1
Two-vector algorithm
To apply the Four-Russians technique we start with the following observation:
Lemma 1. The values along a column from bottom to top and along a row from
left to right are monotonically non-decreasing. Consecutive cells differ at most
by 1.
Proof. Consider neighboring cells (i, j) and (i + 1, j). D(i, j) represents the solution of a longer sequence than D(i + 1, j). Therefore the former value should be
at least as large as the latter. Suppose D(i, j) differed from D(i + 1, j) by more
than one. Then we can remove any matching for i. This has at most one fewer
base pair matching and is a valid solution for the subsequence (i + 1, j) with a
larger value than its current value, contradicting the optimality of D(i + 1, j).
An analogous argument holds along the columns.
Once the cells D(i, l), D(i, l+1), . . . , D(i, l+q−1) are computed, for some l ∈
{i, . . . , j − q}, they can be represented by D(i, l) + V0 , D(i, l) + V1 , . . . , D(i, l) +
Vq−1 , where Vp = D(i, l + p) − D(i, l), for p ∈ {0, . . . , q − 1}. Let us define, v0 = 0
and vp = Vp − Vp−1 , for p ∈ {1, . . . , q − 1}. From lemma 1, vp ∈ {0, 1}, for all
p ∈ [0, q − 1]. Let v denote the binary vector v0 , v1 , . . . , vq−1 of differences and
let V denote the vector of running totals V0 , V1 , . . . , Vq−1 .
Since the vp ’s are defined from Vp ’s, the inverse function is well defined:
Pi
Vp = k=0 vk . Thus D(i, l) together with the vector v represents q consecutive
cells of the table.
Similarly, since the values are non-increasing down a column, D(i + l +
1, j), . . . , D(i + l + q, j) be represented by the pair D(i + l + 1, j), v̄, where
v̄ ∈ {0, −1}q . We call v the horizontal difference vector or the horizontal vector
and we call v̄ the vertical difference vector or the column vector. The corresponding vector of sums is denoted V̄ .
Consider q consecutive cells from l + 1 to l + q used in computing D(i, j):
D(i, j) ←
←
max
l+1≤k≤l+q
max
0≤k≤q−1
D(i, k − 1) + D(k, j)
(2)
D(i, l) + Vk + D(i + l + 1, j) + V̄k
← D(i, l) + D(i + l + 1, j) +
max
0≤k≤q−1
Vk + V̄k
(3)
As before, we use arg max in place of max to obtain D∗ (i, j), which facilitates
the traceback.
As noted above the second line of the recurrence (1), looping over elements,
is more expensive and we will use (3) instead of (2) to compute the D and D∗
values in the Four-Russians method. That is, we will use (3) for groups of q cells
each instead of one loop of (1). Since the V vectors are in bijection with the v
vectors, we will do the preprocessing using v. Let v and v̄ be the corresponding
vectors in (3). The following algorithm evaluates the max computation.
Input: horizontal difference vector v and vertical difference vector v̄
1: max-val ← 0 and max-index ← 0
2: sum1 ← 0 and sum2 ← 0
3: for k = 0 to q − 1 do
4:
sum1 ← sum1 + vi
5:
sum2 ← sum2 + v̄i
6:
if sum1 + sum2 > max-val then
7:
max-val ← sum1 + sum2
8:
max-index ← k
9:
end if
10: end for
11: return (max-val, max-index)
Using this instead of (2) is not advantageous in itself. However, if this algorithm is given as a black box, D(i, j) can be computed in constant time by
invoking the black box once. In the preprocessing stage, we will run the above
algorithm for all possible vector pairs of length q and store the results in table
R. Table R is indexed by a pair of numbers in the range [2q ] to represent the two
vectors (v, v̄). Since there are two entries in the table, the lookup is a constant
time operation. We will show later that this exhaustive enumeration is not too
expensive.
In the Nussinov algorithm described in the previous section, the recurrence
is evaluated using (2) and it takes O(q) time. In the Four-Russians method,
using the preprocessing step, the max computation is available through a table
lookup and the recurrence for q terms can be completed in constant time. This
reduction in the computation time is the reason for the speedup by a factor of
q.
a
i
b
c
j
...
a
..
.
..
.
...
...
b
..
.
c
Fig. 1: A diagrammatic representation of the two-vector method. The row and
column blocks are matched as labelled. The gray boxes and the gray dashes show
the initial value and difference vectors. The group of cells in b correspond to the
Four-Russians loop in lines 15–19 of Algorithm 1; the cells in a are used in the
loop in lines 9–11 and the cells in c form the loop in lines 12–14.
The two-vector method modifies the Nussinov algorithm as follows. All the
rows and columns of the table are grouped into groups of q cells each. The
recurrence over these q cells is computed in constant time using the preprocessing
table. The recurrence involves D(i, k − 1) + D(k, j), i.e., the value in the (k − 1)st
column is used with the k th row. Therefore the row and column groupings differ
by one. That is, the columns are grouped (0, 1, . . . , q − 1), (q, q + 1, . . . , 2q − 1)
etc. The rows are grouped (1, 2, . . . , q), (q + 1, q + 2, . . . , 2q) etc. This ensures
that the row and column groups are well characterized. That is, to fill the cell
(i, j), the k th group along row i needs to be combined with the k th group below
(i, j) in column j.
The cells of the table are filled in the same order as before. When the last
cell of a row- or a column- group is evaluated the corresponding row and column
vectors are computed and stored. To fill cell (i, j), we retrieve the first element
and the horizontal vector of the group from row i and the first element and
the column vector from the corresponding group in column j. The recurrence is
solved using (3) by a table lookup. The final value for D(i, j) is the maximum
Algorithm 1 Procedure for the two-vector Four-Russians speedup. The DP
table is filled column-wise.
1: R ← preprocess all pairs of vectors of length q
2: for j = 1 to n do
3:
D(j, j) ← 0
4:
for i = j − 1 down to 1 do
5:
D(i, j) ← b(S[i], S[j]) + D(i + 1, j − 1)
6:
Let (i, i) be in the I th group in row i.
7:
Let (i, j) be in the J th group horizontally in the ith row and J 0 th group
vertically in the j th column.
8:
Let iq be the right-most entry of group I and jq be the left-most entry
in group J
9:
for k = i + 1 to iq do
// For all cells in the first group
10:
D(i, j) ← max(D(i, j), D(i, k − 1) + D(k, j))
11:
end for
12:
for k = jq to j do
// For all cells in the last group
13:
D(i, j) ← max(D(i, j), D(i, k − 1) + D(k, j))
14:
end for
15:
for k = 1 to J − I do
// For all groups in between
16:
Let p be the left-most cell in the kth group to the right of I and q
be the top-most cell in the kth group below J 0 .
17:
Let vp and vq be the corresponding horizontal and vertical difference
vectors.
18:
D(i, j) ← max(D(i, j), D(i, p) + D(q, j) + R(vp , vq ))
19:
end for
20:
if i mod q = 0 then
// compute the vertical difference vector
21:
compute and store the v vector i/q th group for column j
22:
end if
23:
if j mod q = q − 1 then
// compute the horizontal difference
vector
24:
compute and store the v vector (j − 1)/q th group for row i
25:
end if
26:
end for
27: end for
value over all the groups. There might be residual elements in the row that do
not fall in these groups. There are at most 2q such elements. These are solved
separately using Nussinov’s method. Algorithm 1 has the algorithm listing and
Figure 1 describes the algorithm pictorially.
Runtime Analysis. In the precomputation phase, there are 2q q-length vectors
and 22q pairs of vectors. The precomputation takes O(q) time per vector pair.
Thus the total time for precomputation is O(q22q ).
The main algorithm: There are O(n2 ) cells and to fill each cell it takes
O(n/q + q) time. That is, it takes O(n/q) time to look up the initial value
and the difference vector and the R table lookups for the the O(n/q) groups. It
takes O(q) time for the residual elements. Thus it takes O(n2 × (n/q + q)) time
to fill the table. Every cell is involved in at most two vector computations, where
the difference to its neighbor is computed once for the row and for the column
vector. This takes an amortized O(n2 ) time which is dominated by the rest of
the algorithm.
When q = log n, the total time for the entire algorithm is O(log n 22 log n +
2
n + n2 × ( logn n + log n)) = O(n2 log n + n3 / log n) = O(n3 /logn).
3.2
Other Variants
FG Algorithm. Frid and Gusfield [11] first showed how the Four-Russians approach could be applied to the RNA-folding problem. We will call their algorithm the FG algorithm. FG and two-vector algorithms are variants of the same
idea. We will highlight the differences in preprocessing and the maximum value
computation by the Four-Russians technique. In particular, we will show the
maximum computation in step 18 of Algorithm 1.
After computing the q-contiguous cells of a group in a row, the value in the
initial cell D(i, p) and the horizontal difference vector vp are known. They run
the preprocessing algorithm in page 5 for this fixed vp vector together with all
possible vertical difference vectors. They add the value of D(i, p) to the maximum
and table the result. This preprocessing step is computed for every block of every
row. The preprocessing table R is indexed by row number, group number and a
vector (which is a potential column vector). The horizontal vectors need not be
stored.
To fill cell (i, j), they iterate over all groups and find the q-length column
vectors. The preprocessed value for this vector in the corresponding block is
retrieved from the table and the result is added to D(q, j).
The preprocessing is for horizontal vectors seen in the table. Since the horizontal vectors are not known beforehand, the precomputation cannot be done
prior to the main algorithm. Instead, it is interleaved with the computation of
the table. They fill part of the DP table and use the vectors to complete some
preprocessing, which in turn is used fill another part of the table and so on.
Since the preprocessing is done for every group of every row, the same horizontal vector can be seen multiple times in the table. This leads to duplicated
work and slower running time than the two-vector algorithm.
Memoization The two-vector method computes the preprocessing over all possible vector pairs and the FG method for only the horizontal vectors that are
seen in the table. Stated this way, a hybrid approach suggests itself.
In our next variants, we memoize the results for a pair of vectors. Like the twovector approach, the preprocessing is done only once for a vector pair and like the
FG algorithm, it is only for the vectors seen in the table and the preprocessing
is interleaved with the main algorithm. Since the preprocessing table is indexed
by two vectors, unlike the FG algorithm, the results are computed only once for
every vector seen.
In the partially memoized version, upon completion of elements of a group, if
a new horizontal vector is seen, we pair it with all possible 2q column vectors and
the results are tabled. In the completely memoized version, the result for a pair
of vectors is computed the first time the pair is observed and the result is stored
in the table. The result for future occurrences of the same pair are obtained by
a table lookup. the rest of the algorithm is identical to the two-vector method.
All these variants take O(n3 / log n) time but the memoized versions potentially store fewer vectors than the two vector method and will have a similar
worst-case runtime in practice as the two-vector method. But, as argued before,
the FG method does duplicated work and will be slower in practice.
4
Parallel Algorithm
The Nussinov DP algorithm can be parallelized with n processes to get an O(n2 )
parallel algorithm. We assign one parallel process to a column. In the ith iteration, each process computes the value for the ith diagonal entry. That is, the
successive diagonals are solved in iterations and in each iteration the entries of
the diagonal are solved in parallel. To compute the value for cell (i, j), the entries
in the row to its left and in the column below (i, j) are needed. Since these values
are computed in earlier iterations, each diagonal cell can be filled independent
of the other processes.
A process has to compute the value for O(n) cells and for each cell it needs
to access O(n) other cells. Thus the total computation takes O(n2 ) time with n
processes.
The parallel algorithm for process j for j = 1, 2, . . . , n:
1: D(j + 1, j) ← 0, D(j, j) ← 0
2: for i = j down to 1 do
3:
D(i, j) ← D(i + 1, j − 1) + b(S[i], S[j])
4:
for k = i + 1 to j do
5:
D(i, j) ← max{D(i, j), D(i, k − 1) + D(k, j)}
6:
end for
7:
Synchronize with other processes
8: end for
We will describe the use the two-vector Four-Russians method to obtain
an O(n2 / log n) algorithm below. The preprocessing step that enumerates the
solution for 2q × 2q difference vectors is embarrassingly parallel and we do not
discuss the parallel algorithm for it.
As before, we have n processes one for each column. Each process solves the
entries of the column from bottom to top. Instead of computing the maximum
over each cell in the inner loop (lines 3 – 5 in the parallel algorithm above), we
use the Four-Russians technique to solve q cells in one step by looking up the
table computed in the preprocessing step.
Let dH (i, j) be the horizontal difference vector for cells D(i, j), . . . , D(i + q −
1, j) and let dV (i, j) be the vertical difference for cells D(i, j), . . . , D(i + q − 1, j).
We modify the inner loop of the parallel algorithm as follows:
1: for k 0 = 0 to bj/qc − 1 do
2:
k = i + k0 ∗ q
D(i, j) = max{D(i, j), D(i, k) + D(k + 1, j) + R[dH (i, k)][dV (k + 1, j)]}
end for
for k = bj/qc × q to j do
D(i, j) ← max{D(i, j), D(i, k) + D(k + 1, j)}
end for
Compute the horizontal and vertical differences and store them in dH (i −
q + 1, j) and dV (i, j) respectively.
For each entry, the first loop takes O(n/q) time and the second loop takes
O(q) time. Since all the processes are solving the k th diagonal in the k th iteration,
all of them execute the same number of steps before synchronization. Note that
we compute the horizontal and vertical differences for every node, unlike in
section 3.1 where they are computed every q th cell, to ensure that every process
performs the same number of steps and simplify the analysis. The difference
vectors can be computed in O(q) time. These can also be computed in constant
time by shifting the previous difference vector and appending the new difference.
But we will not assume this simplification for the time bound computation.
Thus each entry can be computed in O(n/q + q) time. There are O(n) entries
for each process, thus the total time taken for all processes to terminate is
O(n2 /q + nq). With q = log n as before, this gives an O(n2 / log n) algorithm.
3:
4:
5:
6:
7:
8:
5
5.1
Parallel Implementation
GPU Architecture
Graphics processing units (GPUs) are specialized processors designed for computationally intensive real-time graphics rendering. Compute Unified Device Architecture (CUDA) is the computing engine designed by NVIDIA for their GPUs.
The programmer can group threads in a block, which in turn can be organized
in a grid hierarchy. Memory hierarchy includes thread-specific local memory,
block-level shared memory for all threads in the block and global memory for
the entire grid. The access times increases along the hierarchy from local to
global memory.
Since the access to global memory is slower (more clock cycles than local
memory access), it is efficient for the threads within a block to access contiguous
memory locations. Then the hardware coalesces memory accesses for all threads
in a block into one request. More specifically, in our application, if a matrix is
stored in row-major order and if the threads in a block access contiguous elements
of a row, then the accesses can be coalesced. However, accessing elements along a
column is inefficient as distant memory elements have to be fetched from different
cache lines.
Programs that observe the hardware specifications can exploit the optimizations in the system and are fast in practice. We designed the program that
exploits the parallel structure of the DP algorithm and the hardware features of
the GPU.
5.2
Related Work
As mentioned earlier, the cells of a diagonal are independent of one another
and can be computed in parallel. In Stojanovski et al. [24], elements of the
diagonal are assigned to a block of threads. This design does not handle memory
coalescence for either row or column accesses. Chang et al. [6] allocate an n × n
table and reflect the upper-triangular part of the matrix on the main diagonal.
Successive elements of a column are fetched from the row in the reflected part
of the matrix. When threads of a block are assigned to elements of a diagonal,
the successive column accesses for a thread are to consecutive memory cells.
However, this does not allow coalesced access for threads within a block. Rizk
and Lavenier [23] show an implementation for RNA folding under energy models.
They show a tiling scheme where a group of cells are assigned to a block of threads
to reuse the data values that are fetched from a column. In this paper, we show
that storing the row and column vectors in different orders for two-vector method
can further improve the efficiency.
5.3
Design of the Four-Russians CUDA Program
We briefly describe the design of the CUDA program; a longer discussion can be
found in [26].
We group cells together into tiles, where each tile is a composite of q × q cells.
The tiles along a diagonal can be computed independent of each other. Each tile
is assigned to a block of threads and computed in parallel. After all the entries
of the tile are computed, only the horizontal and vertical differences are stored.
To fill a tile, the horizontal differences of all the tiles to the left and vertical
differences from the tiles underneath are accessed. These difference vectors are
stored in different orders. The horizontal difference vectors of the rows of a tile
are stored in contiguous memory locations and the tiles are stored in row-major
order. The vertical difference vectors of the columns of a tile are stored together
and the tiles are grouped in column-major order.
To fill a tile, the horizontal difference vectors from a tile to the left are fetched.
When each thread retrieves one vector, the block of threads accesses contiguous
memory locations and the memory accesses are coalesced. Successive iterations
fetch the tiles along a row which are in contiguous memory locations. Similarly
the vertical differences of a tile below are accessed in one coalesced memory
access by the threads of the block.
6
Empirical Results
Prior to empirical evaluation, the FG algorithm was expected to be the slowest
due to redundant work. The memoized versions were expected to be faster than
the two-vector algorithm, as they preprocess only a subset of the 22q vectors
seen in the table.
We ran the programs on complete mouse non-coding RNA sequences. We
also tested the performance on random substrings on real RNA sequences and
random strings over A,C,G,U.
The FG algorithm, while faster than Nussinov, was the slowest among the
Four-Russians methods, as expected. The completely memoized version was
slower than the other two variants. This is because every lookup of the preprocessing table includes a check to see if the pair of vectors has already been
3
processed. There are 22q unique vector pairs but there are O( nq ) queries to the
preprocessing table and each query involves checking if the vector pair has been
2
processed plus the processing time for new pairs. There are O( nq2 ) vector pairs
in the table. For larger n (eg., n > 1000 and q = 8), all the 22q vectors are
expected to be present in the DP table. Generally, memoized subproblems are
relatively expensive compared to the lookup. Since the preprocessing here has
only q steps, the advantage of memoization is not seen.
The partially memoized version was slightly slower than the two vector algorithm. Again, the advantage of potentially less preprocessing than the two-vector
method is erased by the need to check if a vector has been processed. The twovector method was the fastest on all sequence lengths tested.
For short sequences the two vector method took negligible time (less than 0.2
seconds up to 1000 bases) and are not reported. For longer sequences, we noticed
that using longer vector lengths reduced the running time and the improvement
saturated beyond q = 8 or 9. Beyond this, the extra work in preprocessing
overshadowed the benefit. A similar trend was seen for the memoized versions
too. However, for the FG method q = 3 gave the best speedup and longer vector
lengths had a slower running time due to the extra preprocessing at every group.
All the programs were written in C++ compiled with the highest compiler
optimizations. We only discuss the experimental results on a desktop and two
GPU cards in this paper. Detailed notes on running times can be found in [26].
We measured the running times of the different versions of our serial algorithms on a desktop machine with a Pentium II 3GhZ processor and 1MB cache.
The running times of Nussinov and the speedups of various programs compared
to Nussinov are shown in the table below. For sequences of length 6000, the
two-vector method takes close to a minute on the desktop.
Speedup factors of the serial programs on the desktop
Time
Speedup
Length Nussinov Two-vector Partially Completely FG
(in secs)
Memoized Memoized
2000
16.5
7.7
7.3
5.6
3.0
3000
62.5
8.8
8.3
6.4
3.4
4000
196.6
11.9
11.4
8.8
4.7
5000
630.3
21.1
18.9
14.7
7.8
6150
1027.8
18.1
17.0
13.3
7.03
Fig. 2 shows the execution times on two GPU cards – GeForce GTX 550
Ti card with 1GB on-card memory and Tesla C2070 with 5GB memory. The
Fig. 2: Running time of the CUDA program on two GPUs. The programs run
twice as fast on the Tesla card than the GeForce card.
programs take about a second for sequences up to 4000 bases long, and takes
about 5 seconds and 2.5 seconds for sequences of length 6000. The running times
for various sequence lengths are shown in the table below.
Running times for the parallel program (in secs)
Length On GeForce
On Tesla
2000
0.20
0.14
3000
0.62
0.38
4000
1.36
0.74
5000
2.70
1.39
6000
4.97
2.50
7
Conclusions and Future Work
We described the two-vector method for using the Four-Russians technique for
RNA folding. This method is simpler than the Frid-Gusfield method. It also
n2
improves the bound of the parallel algorithm by a log n factor to O( log
n ). We
showed two other variants that memoize the preprocessing results. These methods are faster than Nussinov by up to a factor of 20 and the Frid-Gusfield method
by a factor of 3.
In the future, it will be interesting to see the application of the Four-Russians
technique for other methods that use energy models with thermodynamic parameters. The Frid-Gusfield method has been applied to RNA co-folding [12]
and folding with pseudoknots [13] problems; the application of the two-vector
method to those problems and its implications are also of interest. It will be interesting to compare our run time with the other improvements over Nussinov,
like the boolean matrix multiplication method [1].
Acknowledgements
The first-listed author thanks Prof. Norm Matloff for the opportunity to lecture
in his class; this project spawned out of that lecture. Thanks also to Prof. John
Owens for access to the server with a Tesla card. Thanks to Jim Moersfelder
and Vann Teves from Systems Support Staff for help in setting up the CUDA
systems.
References
1. T. Akutsu. Approximation and exact algorithms for RNA secondary structure
prediction and recognition of stochastic context-free languages. J. Comb. Optim.,
3(2-3):321–336, 1999.
2. M. Andronescu, A. Condon, H. H. Hoos, D. H. Mathews, and K. P. Murphy.
Efficient parameter estimation for RNA secondary structure prediction. Bioinformatics, 23(13):i19–i28, 2007.
3. M. Andronescu, A. Condon, H. H. Hoos, D. H. Mathews, and K. P. Murphy. Computational approaches for RNA energy parameter estimation. RNA, 16(12):2304–
2318, 2010.
4. V. Arlazarov, E. Dinic, M. Kronrod, and I. Faradzev. On economical construction
of the transitive closure of a directed graph (in Russian). Dokl. Akad. Nauk.,
194(11), 1970.
5. R. Backofen, D. Tsur, S. Zakov, and M. Ziv-Ukelson. Sparse RNA folding: Time
and space efficient algorithms. Journal of Discrete Algorithms, 9(1):12 – 31, 2011.
6. D.-J. Chang, C. Kimmer, and M. Ouyang. Accelerating the nussinov RNA folding
algorithm with CUDA/GPU. In ISSPIT, pages 120–125. IEEE, 2010.
7. Y. Ding and C. E. Lawrence. A statistical sampling algorithm for RNA secondary
structure prediction. Nucleic Acids Research, 31(24):7280–7301, 2003.
8. C. B. Do, D. A. Woods, and S. Batzoglou. CONTRAfold: RNA secondary structure
prediction without physics-based models. Bioinformatics, 22(14):e90–e98, 2006.
9. R. Dowell and S. Eddy. Evaluation of several lightweight stochastic context-free
grammars for RNA secondary structure prediction. BMC Bioinformatics, 5(1):71,
2004.
10. R. Durbin, S. R. Eddy, A. Krogh, and G. Mitchison. Biological Sequence Analysis:
Probabilistic Models of Proteins and Nucleic Acids. Cambridge University Press,
1998.
11. Y. Frid and D. Gusfield. A simple, practical and complete O(n3 )-time algorithm for
RNA folding using the Four-Russians speedup. Algorithms for Molecular Biology,
5:13, 2010.
12. Y. Frid and D. Gusfield. A worst-case and practical speedup for the RNA co-folding
problem using the four-russians idea. In WABI, pages 1–12, 2010.
13. Y. Frid and D. Gusfield. Speedup of RNA pseudoknotted secondary structure
recurrence computation with the Four-Russians method. In COCOA, pages 176–
187, 2012.
14. M. Hamada, H. Kiryu, K. Sato, T. Mituyama, and K. Asai. Prediction of RNA secondary structure using generalized centroid estimators. Bioinformatics, 25(4):465–
473, 2009.
15. I. L. Hofacker. Vienna RNA secondary structure server. Nucleic Acids Research,
31(13):3429–3431, 2003.
16. B. Knudsen and J. Hein. Pfold: RNA secondary structure prediction using stochastic context-free grammars. Nucleic Acids Research, 31(13):3423–3428, 2003.
17. Z. J. Lu, J. W. Gloor, and D. H. Mathews. Improved RNA secondary structure
prediction by maximizing expected pair accuracy. RNA, 15(10):1805–1813, 2009.
18. N. R. Markham and M. Zuker. UNAFold. Bioinformatics, 453:3–31, 2008.
19. D. H. Mathews, M. D. Disney, J. L. Childs, S. J. Schroeder, M. Zuker, and D. H.
Turner. Incorporating chemical modification constraints into a dynamic programming algorithm for prediction of RNA secondary structure. PNAS, 101(19):7287–
7292, 2004.
20. R. Nussinov and A. B. Jacobson. Fast algorithm for predicting the secondary
structure of single-stranded RNA. PNAS, 77(11):6309–6313, 1980.
21. R. Nussinov, G. Pieczenik, J. R. Griggs, and D. J. Kleitman. Algorithms for loop
matchings. SIAM Journal on Applied Mathematics, 35(1):68–82, 1978.
22. J. Reuter and D. Mathews. RNAstructure: software for RNA secondary structure
prediction and analysis. BMC Bioinformatics, 11(1):129, 2010.
23. G. Rizk and D. Lavenier. GPU accelerated RNA folding algorithm. In International
Conference on Computational Science, 2009.
24. M. Stojanovski, D. Gjorgjevikj, and G. Madjarov. Parallelization of dynamic programming in nussinov RNA folding algorithm on the CUDA GPU. In ICT Innovations, 2011.
25. I. Tinoco et al. Improved Estimation Of Secondary Structure In Ribonucleic-Acids.
Nature-New Biology, 246(150):40–41, 1973.
26. B. Venkatachalam, Y. Frid, and D. Gusfield. Faster algorithms for RNA-folding
using the Four-Russians method. UC Davis Technical report, 2013.
27. Y. Wexler, C. B.-Z. Zilberstein, and M. Ziv-Ukelson. A study of accessible motifs
and RNA folding complexity. Journal of Computational Biology, 14(6):856–872,
2007.
28. S. Zakov, Y. Goldberg, M. Elhadad, and M. Ziv-Ukelson. Rich parameterization
improves RNA structure prediction. In V. Bafna and S. C. Sahinalp, editors,
Research in Computational Molecular Biology (RECOMB), volume 6577, pages
546–562. Lecture Notes in Computer Science, Springer, 2011.
29. S. Zakov, D. Tsur, and M. Ziv-Ukelson. Reducing the worst case running times of
a family of RNA and CFG problems, using Valiant’s approach. In WABI, pages
65–77, 2010.
30. M. Zuker. Mfold web server for nucleic acid folding and hybridization prediction.
Nucleic Acids Research, 31(13):3406–3415, 2003.
31. M. Zuker and P. Stiegler. Optimal computer folding of large RNA sequences using
thermodynamics and auxiliary information. Nucleic Acids Research, 9(1):133–148,
1981.
| 5cs.CE
|
Waterfall: Primitives Generation on the Fly
Guido Chari
Diego Garbervetsky
arXiv:1310.2741v1 [cs.PL] 10 Oct 2013
Departamento de Computación, FCEyN, UBA
http://lafhis.dc.uba.ar
Abstract
Modern languages are typically supported by managed
runtimes (Virtual Machines). Since VMs have to deal
with many concepts such as memory management, abstract execution model and scheduling, they tend to be
very complex. Additionally, VMs have to meet strong
performance requirements.
This demand of performance is one of the main
reasons why many VMs are built statically. Thus, design
decisions are frozen at compile time preventing changes
at runtime. One clear example is the impossibility to
dynamically adapt or change primitives of the VM once
it has been compiled.
In this work we present a toolchain that allows for
altering and configuring components such as primitives
and plug-ins at runtime. The main contribution is Waterfall, a dynamic and reflective translator from Slang,
a restricted subset of Smalltalk, to native code. Waterfall generates primitives on demand and executes them
on the fly. We validate our approach by implementing
dynamic primitive modification and runtime customization of VM plug-ins.
1.
Introduction
Modern high-level languages usually rely on runtime
systems (Virtual Machines) which provide abstract execution models that promote portability, automatic
memory management and enforce certain security properties. VMs are complex pieces of software that historically, for efficiency reasons, were developed in static
low-level languages. Over the last years a new branch
of VM implementations written in high-level languages
appeared [8]. By relying on better abstractions the
[Copyright notice will appear here once ’preprint’ option is removed.]
1
Camilo Bruni
Marcus Denker
Stéphane Ducasse
RMoD, INRIA Lille - Nord Europe, France
http://rmod.lille.inria.fr
VMs would be easier to develop, debug and modify.
The downside of the gain in abstraction is a further
separation from the low-level execution model of the
underlying hardware. It requires substantial efforts to
make VMs written in high-level languages efficient [17].
Self-hosted VMs [7] are written in the same language
they support. A common way of tackling them was to
restrict the actual semantics for the code targeted to the
VM construction. This is the path taken in many implementations like Squeak [11] for Smalltalk and PyPy [17]
for Python. The code of these VMs is written in a subset of the language they provide. This subset typically
has the same syntax, but the semantics are restricted
in order to statically bind everything at compile time.
Essentially, this prevents polymorphism and dynamic
method dispatch which are substantial features of highlevel host languages. A quite similar approach is observed in popular research VMs like Jikes [2] or Jalapeño
[1] for Java.
VMs written in high-level languages have improved
considerably and it has become a prolific area [7, 20].
One promising technique is by promoting low-level highlevel programming frameworks [9] which aims at solving
the problem of the abstraction mismatch of metacircular or self-hosted VMs. With these elaborate tools and
others like runtime code specialization the performance
penalty can be considerably mitigated.
However we identified that most of these VMs still
suffer from an important limitation:
It is not possible to change or configure many of the
most important design decision at runtime. Key components are frozen at compile time and hidden at runtime
One example of this limitation are the primitive
methods. They are used by VMs to perform basic operations such as arithmetic operations and object allocation or to accomplish performance demanding task
more efficiently [10]. They live at VM-level and are statically bound at compile time. In certain cases a developer
may be interested in creating or adapting a primitive,
for instance for instrumenting an application. In some
2018/4/2
languages this can be achieved by exploiting reflective
capabilities [18]. But the overhead imposed by actual
implementations of reflection make them impractical for
performance demanding tasks [13].
Similarly, some VMs support the concept of plugins to enable the efficient implementation of recurrent
general operations (e.g, file access, floating point operations, compression, etc). Plug-ins also provide some
features that can not be implemented fully at languageside. Moreover, only after releasing, developers may
identify a hotspot in their application and improve its
performance by providing a plug-in for it. Unfortunately, in many cases this implies that every user of the
application will need to recompile its VM in order to
update the plug-in. Other option is to ship the plug-in
as a library and depend on the user’s platform compilers
or dynamic linkers. Clearly, giving end-users responsibilities, or worse, requiring them to update their VM
every time a new plug-in is released is undesirable.
To overcome this limitations we conceived a toolchain
which solves the aforementioned problems. Its main
component is Waterfall, a runtime translator for code
in a syntax equal the the language the VM supports.
We take the approach of high-level low-level programming one step further and use it dynamically at runtime to change VM behavior. Concretely, Waterfall enables to dynamically modify primitives and plug-ins
from language-side, outperforming existing approaches
that are purely reflective. We use a single-language
approach instead of using Foreign Function Interfaces
(FFI) [12, 14] to access external libraries. This enables
to debug, inspect and change the code at runtime with
the same tools the developers use for their general purpose tasks. Exploiting Waterfall capabilities, developers
and experienced users are able to tune their VMs at
runtime without the need of external tools.
The contributions of this paper are:
• A toolchain, written in the language of the VM, that
enables to compile and activate code at runtime.
• A proof of concept that uses the toolchain for adapt-
ing VM behavior at runtime, without the need of a
system restart, considerably outperforming the reflective solutions.
• An empirical validation demonstrating the approach
is feasible and the penalty in terms of performance
is reasonable penalty.
2.
Statically Defined Primitives
One of the main components of VMs are primitives.
Primitive routines provide the runtime with the capability to perform essential operations that the language
could not supply by its own. For instance, create objects or provide access to low-level structures. On the
2
other hand, they are also used to optimize some critical
bottlenecks [10, ch.3, p.52]
Each time the runtime activates a method that is
a primitive, it swaps the execution mode. The VM,
instead of interpreting bytecodes or performing message dispatching, directly executes the binary code of
the primitive. Primitives typically are already statically
compiled, further optimized and in general they are
written in the same language used to implement the
VM. Changing or extending primitives is profitable in
the case where core features of a language needs to be
inspected or modified. For instance, if a VM accesses
its instance variables through a primitive, immutability
could be easily and efficiently achieved by changing that
primitive.
The problem is that primitives are deeply coupled to
the VM building process which is a complex and timeconsuming task. Moreover, changing primitives may not
be possible since it requires access to the VM source
code, which is not always open. Even in that case, primitives are written on a different abstraction level than the
high-level language the VM supports. This is a complex
barrier for the common developer of the host language
since he needs to work in two abstraction levels at the
same time and deal with different development environments and tools [9]. As a consequence it is observed that
the set of available primitives is statically defined and
frozen at compile time, making them difficult or even
impossible to change.
Self-hosted VMs try to overcome this two-language
problem by implementing the VM using a language with
a similar syntax as the host language, mitigating several
of the problems mentioned previously. But, even though
the syntax is usually similar, or even equal, the developer still has to be aware that the semantics are different from the host language. For example, in Squeak
and PyPy the VM is implemented in a language that
is almost identical to the host language but eventually
the semantics are reduced to C-like expressions.
Another example similar to primitives in high-level
dynamic environments is the Pharo VM [4, 16] plug-in
infrastructure. Pharo provides a large number of plugins tailored towards specific but heterogeneous tasks
such as algebraic matrix operations, floating point precision computations, file management, etc. All these
tasks have in common that they are intensively used for
repetitive and performance demanding tasks that are
not fulfilled by the standard runtime execution model.
In essence, plug-ins consist of a set of primitives isolated
on separated modules.
Plug-ins are, in many cases, built and deployed with
the VM. This is unpleasant and involves exactly the
same limitations as static primitives in a larger scale.
The other scenario is to build them as dynamic linked
2018/4/2
libraries (DLLs), like modules outside VM, and link
them dynamically. In an case, plug-ins imply bigger
binary footprints, more complexity and overhead for
deployment. Plug-ins force users to work with VM binaries with a large amount of statically built code that
they perhaps never use. Concerning runtime adaption,
the only option to change and modify plug-ins is by using the second approach (DLLs) and by loading them
dynamically with external OS tools, possibly after unloading a previous version. Relying on external compilers brings its own set of problems. Additionally to
the given CPU architecture difference, compilers and
linkers differ on each operating system.
cally at code generation time. Hence Slang’s semantics
are closer to C than to Smalltalk.
In Cog, primitives are defined at VM-side using Slang
and they are frozen at compile time. Following a similar
approach as primitives the Cog VM can be customized
with plug-ins which are also written in Slang and use
the same compilation strategy already explained. Plugins can be seen as modules that encapsulate a set of
particular primitives. One of the main differences is
that plug-ins can be compiled in external files, like
independent libraries, and then linked to the VM at
runtime.
3.2
Below we summarize the limitations of current VMs,
concerning runtime adaptation, that we want to address:
1. Primitives are statically defined and frozen at compiled time.
2. Primitives can not be efficiently changed at runtime.
3. It is complex, heavy-weight and error prone to extend or upgrade VM modules at runtime.
3.
Context
Smalltalk is a good representative of a dynamic, objectoriented and reflective high-level language. We choose
Pharo1 [4, 16], an open-source Smalltalk-inspired environment, for developing the proof of concept used to
validate our approach. Pharo features a self-hosted VM
with primitives accessible at language-side. Moreover,
the VM provides a powerful plug-in infrastructure with
a simple interface, enabling its extension with efficient
low-level functionalities, which works very similar to
primitives.
In the rest of this section we provide context for understanding the Waterfall approach presented on next
section.
3.1
Waterfall relies on a framework called Benzo [5] that
provides dynamic high-level low-level programming
techniques. Benzo provides a language-side assembler, a
dynamic code generator and a set of generic primitives
for activating native code from language-side.
Instead of building a separate infrastructure for generating and activating native code we rely on this
generic framework. Benzo is based on 5 generic primitives to activate native code and access VM symbols.
Benzo uses its own language-side assembler to generate native code without external help. Once the native
code is ready Benzo uses a dedicated primitive to active
the code. This very basic interface is enough to open
doors for new language-side implementations of typical
VM-level tools. Benzo is already successfully used in
production for a language-side FFI implementation [6].
4.
The VM for Pharo is the Cog VM [15], which has a
unique construction process that was inherited from the
Squeak VM [11]. The language under which Cog is developed is a subset of Smalltalk known as Slang. This
has the advantage that Slang programs can be managed
and explored in the same way as any other code from
the host language. The main difference is that Slang
code is not executed by the runtime. Instead there is a
compiler that translates Slang to C, wherefrom a standard C compiler takes over. This is why Slang basically
has the same syntax as Smalltalk but it is semantically
constrained to expressions that can be resolved stati1 http://pharo.org
3
Waterfall Compiler
Waterfall’s main task is to translate Slang to native
code at runtime. The compiler was designed as a chain
of transformations that receives as input a high-level
representation of code (Slang) and returns as output
a lower-level one (binary code) going through different
intermediate representations in the process.
4.1
Pharo VM
Benzo Framework
General Architecture
A high-level graphical description of the processes involved in generating native code is shown in Figure 1.
The Waterfall toolchain begins by parsing Slang to an
Abstract Syntax Tree (AST) representation. Then the
AST representation is translated to native code enforcing C-like Slang semantics. Waterfall generates assembler instructions at language-side and executes them
using the Benzo framework. Waterfall can also translate the AST into an intermediate representation (IR)
featuring an hybrid IR between Three Address Code
(TAC) and Static Single Assignment form (SSA) as a
pre-step before native code generation.
Even though Slang is syntactically similar to Smalltalk it is actually closer to C. It is completely bound
2018/4/2
Slang
Benzo
Pharo
Parser
Dynamic
Assembly
Execution
AST
Syntactic
Transformation
ent type of calls (calling convention) which we describe
in more detail in Section 4.4.3.
PSlang AST
Move to
IR
Assembly
Nativization
IR
TAC/SSA
Figure 1: All Waterfall stages for Slang to native compilation.
at compile time and most of their constructs can be
directly translated to C. Considering that fact, the binary code that Waterfall generates follows similar code
conventions used by C compilers. Concretely, Slang variables and arguments are pushed on the native stack and
Slang messages are treated very similar to C function
calls. Waterfall also sees every variable just as a word
in memory leaving its semantics interpretation (type)
to the developers.
Figure 2 exposes a high-level diagram of the core
classes involved in the translation. Waterfall’s main
component is the ClosureNativizer that has as collaborators: a set of SendNativizers, a static set of
Primitives of the language and a chain of Converters.
Converter Hierarchy. Converter objects essentially
receive as input a representation of the code, apply transformations and emit a modified output. A
chain of these transformations defines the compilation
phases that Waterfall will actually execute. VisitorConverter is the class that represents the behavior for
walking through nodes in a visitor fashion. All transformations that traverse ASTs doing something different
for each kind of tree node are implemented as visitors
and collaborate with a VisitorConverter. With this
architecture it is quite simple to define new conversions
such as optimizations or further IRs, and link them to
the chain. As an example of the reuse of standard environment tools, Smalltalk2AST converter is instantiated with the Pharo standard parser. The converter
is named Smalltalk since their syntax is the same, but
Waterfall uses it for translating Slang.
4.2
Replacing Primitives with Generated
Code
Once Waterfall generated the code, there still remains
the step of actually installing and executing it at runtime. In the case of changing a primitive, the generated
code must be executed instead of the original static
primitive. Below we describe a summary of the main
steps involved in this task:
1. Exploiting reflective properties of the environment
with meta programming techniques, Waterfall identifies whenever a plug-in or primitive is called. In the
Pharo environment these methods are identified by
a special pragma as first statement.
2. The pragma is removed and instead Waterfall installs
code that invokes the generation of a new primitive
on demand from Slang code.
Figure 2: Waterfall high-level class diagram
Primitives. Certain operations at Slang-level can
not be expressed as calls to other Slang methods and
thus are defined as assembler templates. Much like in
the normal Pharo environment, Slang relies on certain
primitive operations that can not be expressed at the
same level. For instance, algebraic operations such as
+, −, bit manipulation or direct memory access operations have to be treated differently. Instead of generating their native code from a Slang method we inline an
already nativized template (instance of the Primitive
class).
SendNativizer Hierarchy. The SendNativizer
objects have the responsibility of handling the differ4
3. Finally, Waterfall relies on Benzo for executing the
native code. Subsequent calls to the same primitive
function will not execute step 2, avoiding the compilation overhead.
4.3
Replacing Plug-ins with Generated Code
There is only a marginal difference when using Waterfall for generating plug-ins since plug-ins are in some
way modules encapsulating a set of related primitives.
They are more general. For dealing with them there
exist a class PluginNativizer whose instances receive
as input (collaborators) a set of methods (primitives)
from a plug-in. Additionally PluginNativizer knows
the class where the new dynamically nativized primitives should be installed. This nativizer is responsible
for calling the translator on each of the methods and
2018/4/2
installing the code generated as new primitives on the
corresponding class.
With this approach it is quite simple to mark a plugin from language-side as dirty for later recompilation.
Concretely, marking it as dirty implies to trigger a
call to the nativizer with only the changed method as
input. Since the interface described above is written at
language-side, and Smalltalk is a reflective language,
all this behavior can be dynamically activated. Plugins developers can also define their plug-ins as language
functions that trigger the nativizer whenever called for
the first time. It is also easy to make changes to a plugin or create a different version and let the user decide
to use the one that better fits his requirements.
4.4
Compilation Steps
In the rest of this section we describe the standard steps
of transformations (Converters) needed for translating
Slang code to native instructions.
4.4.1
Transforming Slang to AST
This step benefits from the cooperation of high-level
tools at language-side provided by the Pharo environment. Concretely, since Slang is a subset of Smalltalk,
Waterfall uses the standard Smalltalk parser to generate
the AST. The parser itself is written in Smalltalk and
resides at language-side, which makes it an easy target
for debugging and possible extensions. Hence Waterfall
does not need a special parser.
4.4.2
Computing the Set of Reachable
Methods
Transforming AST to Native Code
The actual nativization step consists of a visitor that
receives an AST as input and returns a stream of bytes
which then is loaded into memory for execution. For
abstracting from the binary code generation, we use an
already developed abstract assembler called ASMJit2
that is an independent tool inside the Benzo framework.
The Slang AST has few kind of nodes. Each node
is translated into a set of native instructions. As an
example, we briefly explain how to translate nodes cor2 ASMJit:
4.4.4
https://code.google.com/p/asmjit/
5
Dynamically Executing Generated
Native Code
After generating the native instructions there still
reamins the task of activating the native code and passing it down to the VM. This is not directly possible since
in general VMs strongly separate the language from the
low-level environment. Due to this barrier it is simpler
to ensure portability and security properties for a VM.
However, in our case this poses a limitation since we
want to dynamically execute instructions at VM-level
which were created at language-side. As already explained, Waterfall is supported by Benzo framework for
overcoming this limitation.
5.
Implementation Details
The main goal of Waterfall is to provide support for
changing low-level (VM) behavior at runtime. Due to
the existing restrictions of the VM we apply certain
simplifications or rely on external tools to accomplish
our goal.
5.1
Before executing a Slang method Waterfall translates
its code into machine instructions. However looking at
an isolated method itself is not enough. Slang programs
can be decomposed in several methods so Waterfall also
has to translate all the methods reachable for all possible executions starting at the method under translation.
Essentially, every potentially executable method has to
be nativized. However it must remain clear that Waterfall currently does not support polymorphism, only
modularization.
4.4.3
responding to variable references (more details about
the translation of other nodes in section 5). The visitor
checks what kind of variable is referenced (i.e., temporary, argument, global, low-level symbol) and it finds its
memory address: for temporaries and arguments it will
be a native stack location, for VM globals the position
where they were loaded in memory. Finally, the translator interacts with ASMJit for pushing the gathered
memory address into the stack.
Resolving External Symbols
Code written by developers may reference other code or
data that already exists in the managed runtime, like
global variables or VM functions. In these cases, the
nativization procedure has to find the actual addresses
in memory where these static references lie and inline
them into the generated native code. All this is necessary because Slang eventually resides at VM-level when
executed and thus it is not possible to access global objects directly. While executing a primitive it is not possible to interact with the high-level environment. For
instance accessing an object using the standard Smalltalk way by sending a message would cause recursion
problems.
To access the position of VM internal symbols referenced in Slang methods, Waterfall relies on an existing
Benzo API to interface with C libraries which is based
on dlsym. We developed also a parser of the nm Unix
command for gathering the positions not accessible to
dlsym in the VM binary. As a consequence of the external tools selected, the complete functionality of the
compiler can be only obtained on Unix platforms. However, it should not be difficult to develop a parser for a
tool similar to nm on Windows or other platforms.
2018/4/2
5.2
Special Parameters
Pharo primitives receive their arguments on the stack.
The VM is responsible for pushing them right before
calling the primitive. Low-level primitive code special
has to be careful when accessing these arguments. The
Cog VM uses a moving Garbage Collector (GC) which
requires careful access to high-level objects when in VMlevel code. The GC is not aware of the C stack. Therefore, if a GC pass happens during primitive activation,
Waterfall pointers to language-side objects held on the
C stack are not updated resulting in dangling pointers
that would cause severe troubles. Waterfall simplifies
the access to high-level objects by using a single statically known position for all parameters. This way only
a single memory address has to be registered at the
garbage collector for not moving it. Since the VM is
single threaded and no two primitives can be executed
at the same time it is safe to rely on a single global
argument position.
5.3
Slang Purification
The complete Slang language supports special syntax
for inlining C expressions. That means actually that
it is at least as expressible as C. It also allows for
other expression such as type pragmas which contain
type information and are directly translated to C types.
Waterfall currently does not fully support these two
support special Slang expressions. However to validate
our approach we reimplement these two features in a
simple fashion.
In the case of types there exists a special converter
that walks over the pragmas of the nodes, parses them
and finally assigns basic type information to the variables that were reified during the previous transformations. The converter also marks if their arguments are
used as value or as reference. Concerning the inlined C
macros, we realized that they were mostly used to call
external functions so we decided to provide a special
language construction only for that cases.
We finally provide a special converter that performs
simple string substitutions for the Slang features we
implement differently.
5.4
Managing Stack Frames
The most complex node of the AST (Section 4.4.3) to
deal with by Waterfall is the one for messages in Smalltalk which is reduced to function calls in Slang. The
generation of native code in this case implies defining
a calling convention for argument passing and register
preserving (similar to C in this case): pushing correctly
the arguments, preserving the needed registers, calling to the right place and resuming control at return.
For each function call, a Context object is instantiated
which represents the stack frame and has a pointer to
6
its parent. This context is responsible for determining
the actual position on the stack for every reference to
a variable. Since Waterfall allows blocks (lambda functions), it must manage very carefully the context stack
since a variable reference could be in another context far
from the current one. Finally, as an example of the different calling conventions, calling VM functions implies
sticking to a C ABI, whereas for Slang internal calls
there is a receiver of the function (implicit parameter)
that is always pushed on the stack.
6.
Validation
We present two case studies: an essential primitive and
a language-side plug-in untied from the VM building
process. The first experiment validates that our solution efficiently addresses the first two limitations identified in Section 2 concerning primitives. The second
experiment shows how Waterfall overcomes the limitation regarding language-side VM extensions. We run all
the benchmarks for this paper with the SMark3 benchmarking tool. On each benchmark we measure 50 runs
and take the average time.
6.1
Essential Primitives Instrumentation
We distinguish two types of primitives: essential and
non-essential. Essential primitives are required for the
bootstrap and vital operations of the language, such
as creating a new object or activating a block. Such
primitives can not be easily implemented at languageside. The second category of primitives are mainly used
for optimization purposes and could be replaced by
language-side code.
In this particular case study, we focus on essential primitives. Instrumentation of essential primitives
is an error-prone task falling in many cases in nontermination due to recursive loops. A difficult candidate
is the basicNew primitive, which is responsible for instantiating new objects. Even a very simple instrumentation task such as printing the address in memory of
the created object is problematic. If during the printing
process another object is created, the very same instrumented basicNew primitive would be triggered.
Using reflective techniques it is possible to avoid this
loop. Essentially one would have to inspect the current
stack for previous activations of the modified primitive
before using it. If the primitive has been used before, the
code jumps to the original unmodified primitive. Thus,
breaking the recursive loop. However this approach
imposes a considerable overhead.
3 http://smalltalkhub.com/#!/~StefanMarr/SMark
2018/4/2
Average Time
Unmodified
Unsafe reflective instrumentation
Secure reflective instrumentation
Waterfall-based instrumentation
Waterfall-based unmodified
0.28
21.80
27.72
7.72
7.08
±
±
±
±
±
0.16
0.33
0.40
0.27
0.23
ms
ms
ms
ms
ms
Relative Time
w.r.t Waterfall Instr.
1.0×
≈ 78×
≈ 99×
≈ 28×
≈ 25×
−
≈ 2, 8×
≈ 3.6×
1.0×
−
Table 1: Slowdowns comparison for instrumentations of the essential primitive basicNew.
6.1.1
Experiment
We present the code of two approaches for instrumenting the object creation primitive: a pure language-side
solution and one translated and executed by Waterfall.
The language-side version of an instrumented basicNew looks as follows:
Class>>basicNew
| object |
object ← super basicNew.
FileStream stdout ifNotNil: [ :stream |
stream << object nbAddress << String lf ].
We make different comparisons, measuring the creation
of 100 and up to 1000 objects. We run the experiment
for: the standard primitive, a version with an unsafe
reflective instrumentation, a safe version of a reflective
instrumentation with a secure guard, a primitive compiled by Waterfall for creating objects and finally an
instrumented Waterfall version. Table 1 illustrates the
relative slowdown factors.
However, printing on the standard output might easily
fall in a recursive loop as described before. Hence the
safe version needs an additional recursion guard:
Class>>basicNew
RecursionGuard
ifStackContains: #basicNew
do: [ ↑ self unmodifiedBasicNew ].
FileStream stdout ifNotNil: [ :stream |
stream << self name << String lf ].
↑super basicNew
In Waterfall we define a new version of basicNew in
a small Slang method which itself is a wrapper around
the unmodified primitive:
WaterFall >> slangBasicNew
| oop value |
oop ← self stackAt: 0.
self
callVMFunction: #printOop
withArguments: { oop }.
↑self
callVMFunction: #primitiveNew
withArguments: {}.
Much like the reflective solution, the instrumented
Slang version of the primitive delegates the main functionality to the original one. The reflective version uses
a normal message send to call the original primitive. In
Slang, after calling a low-level function that is responsible for printing the object memory address, it performs
a function call to the VM-level basicNew function.
6.1.2
Results
In Figure 3 we compare the run times for different instrumentation approaches of the basicNew primitive.
7
Figure 3: Instrumenting the creation of objects with
reflection and with Waterfall
Some interesting conclusions can be inferred from
the chart. First of all, it can be seen that in most
cases the time for creating objects with the standard
Pharo primitive is almost negligible. This is because
creating objects is an essential task in an object-oriented
environment and thus the primitive is highly optimized.
Then it is also visible that just the overhead of calling
Waterfall is quite high. This is because Waterfall is still
a prototype and we have not focused on optimizations
yet. Actually, is not our goal to compare Waterfall with
a fully optimized C compiler.
It is also worth noting that the cost of instrumentation using Waterfall is negligible compared to the reflective techniques. As the chart exhibits, plain Waterfall
versions are similar in time to Waterfall instrumented
version while the table shows that reflective instrumentation imposes a slowdown factor of 78× and 99×. This
2018/4/2
is encouraging since it shows that if Waterfall is further
optimized it would get closer to the pure static times.
Finally, concerning the analysis only for instrumented versions, the chart and table also exhibit that
the reflective solutions are considerably slower than the
ones based on Waterfall (a slowdown factor is between
3 and 4 depending on the approach). We compare Waterfall with two reflective approaches for completeness
reasons but, since Waterfall avoids the recursive loop,
a fair comparison will be with the safer version. The
final cost of this operation is correlated with the size of
the call stack it must traverse. Since this slowdown factor of almost 4× was obtained using the benchmarking
framework that generates a pretty shallow call stack,
we conjecture that a real application may suffer more
overheads, favoring our approach of using Waterfall.
Moreover, the implementation in Waterfall of simple
optimizations like function inlining will surely enlarge
even more the differences.
6.2
Towards Dynamic Customizable VMs
In order to provide a stronger validation we choose
an interesting and general enough plug-in, frequently
used by most of the users. We basically strip it from
the VM. Then, we use Waterfall for dynamically and
lazily compiling and executing the functions which are
actually used.
6.2.1
Experiment
In Pharo all file related operations are delegated to a
plug-in named the File Plug-ins. We choose to experiment with this plug-in since it encompasses several
functionalities and is exhaustively used by most standard users. It also is one of the most complex plug-ins,
because it has a strong interaction with the OS.
The experiment consists on evaluating the feasibility
of extracting the functionality from VM-side and using
Waterfall for generating only the functionality required
on demand.
6.2.2
Results
Following the same methodology already explained for
the previous validation, we compare the execution time
for creating directories with the standard static plugin with the time it takes for creating them with the
function dynamically generated with Waterfall. Figure 4
shows the results. The X axis represents the number of
different directories the test created. Each one is a call
to the primitive. The Y axis exhibits the average time.
It can be clearly observed that concerning the performance is almost the same with the static as with
the dynamic approach. Also it is visible, and it seems
like a contradiction, that the growth is non-linear. For
instance, creating 2000 directories is more than twice
than creating 1000. This is because the more directories
8
Figure 4: Static vs. dynamic plug-in comparison for
creating files
created, the more the environment works, more objects
created and then more garbage collection cycles.
We noted that we obtained very good timing results.
We conjecture that this could be related to the fact
that the main cost of file operations is expended by
the OS. For other plug-ins, the Waterfall version may
expose some performance degradation. However, the
main focus of this experiment is to show the feasibility
of our approach on removing plug-ins from the VM.
File Plug-ins is an excellent candidate for that goal
since it is one of the most-used plug-ins. We managed
to move it completely to language-side and show that
it has no impact on the usability of the environment.
The discussion about efficiency was already presented
on the previous experiment.
7.
Related Work
An approach more similar to ours is QUICKTALK [3].
It was conceived to compile Smalltalk code directly to
binary code. Developers must include type annotations
in order to bound all method invocations at compile
time. This focus on performance and the development
of a complex compiler for a new Smalltalk dialect. In
contrast Waterfall creates a bridge between the VM
and the language-side. Unfortunately, this tool is not
available to perform an empirical evaluation.
Another noteworthy Smalltalk implementation is
Smalltalk/X4 that has excellent C integration built into
the language. Using dedicated syntax, C expressions can
be written inline. Smalltalk/X explicitly uses this approach to define primitives transparently at languageside (in C). Much like Waterfall primitives can be modified at runtime, however Smalltalk/X does not focus
on a one-language approach where the VM and the
dynamic primitives share the same language.
High-level low-level programming [9] encourages to
use high-level languages for system programming. In
4 http://www.exept.de/en/products/smalltalkx
2018/4/2
this work authors present a low-level framework which
is used as system interface for Jikes, an experimental
Java VM. Methods have to be annotated in order to
tell compiler to use the low-level functionality. Although
this work is an step forward the use of high-level languages to build system software, the strong separation
between low-level code and runtime does not allow for
reflective extensions of the runtime.
Pharo benefits from Squeak[11] an early self-hosted
VM for Smalltalk. More recent self-hosted approaches
include Klein [19] for Self, PyPy [17] for Python, Maxine
[20] for Java or Tachyon [7] for JavaScript. The Maxine
VM stands out as it truly focuses on productivity and
developer interaction. Maxine uses abstract and highlevel representations of VM-level concepts and consistently exposes them throughout the development process. Inspectors at multiple abstraction levels are readily
available while debugging, giving insights to the complete VM state. Compared to Maxine, Waterfall currently lacks the debugging tools which would enable
a truly seamless interaction with the low-level world.
However, Maxine focuses on Java, a language with inferior reflective capabilities compared to Pharo. Hence the
live interaction with the VM is only presented in the development phase and not exposed to the language-side.
Maxine would be an excellent candidate to implement
our approach for Java.
Apart from the VM community, Waterfall shares
many ideas with research conducted in the high-level
reflection domain. For instance, Röthlisberger et al. [18]
present a tool to mitigate runtime adaptions with unanticipated partial behavioral reflection . Built on top of
Smalltalk they present a tool to limit the computational
overhead that dynamic runtime reification introduces.
This contrast with Aspect Oriented Programming solutions where the runtime has to be modified upfront
to allow for partial behavior reflection. In the scope
of our work, the latter approach corresponds to static
changes at VM-level. Waterfall on the other hand allows
for unanticipated changes.
8.
perform reasonably well with respect to their statically
compiled counterparts.
We believe this approach provides a flexible mechanism for adapting and evolving VMs and enables developers to deploy them with a smaller footprints. VMs can
be later customized by users according to their needs,
at runtime and without resorting to external tools.
Concerning conceptual improvements, with Waterfall we encourage the use of high-level low-level programming at runtime. Furthermore we advocate the
importance of self-hosted VMs to control and modify
essential parts of them from language-side.
We also managed to enhance quality properties of the
environment by decoupling it from the low-level building infrastructure while modifying low-level behavior.
Operating systems, compilers and linkers tends to impose important constraints. For instance, with Waterfall
it can be completely avoided for altering some components at runtime the common low-level cycle: compile
→ link → save to file → load to memory. As a consequence, our approach also improves portability since the
compilation infrastructure is in general very dependent
on the platform.
8.1
Perspectives
Even current self-hosted or metacircular VMs freeze
many aspects at compile-time. We think VMs should be
easier to evolve and adapt dynamically, using reflective
runtime capabilities. At the same time, the VM performance should be comparable with the solutions written
in low-level languages. We believe the work presented in
this paper together with other techniques such as gradual typing and type inference, opens up new perspectives about the possibility of approaching our vision.
In this setting we plan to work on improving the
Waterfall back-end in order to produce more efficient
code. This includes powerful static and dynamic analysis techniques targeted specially for highly dynamic
environments.
References
Conclusions
In this work we present a toolchain that allows for altering and configuring components such as primitives
and plug-ins at runtime. The main component is Waterfall, a dynamic and reflective translator for Slang,
a restricted subset of Smalltalk. Waterfall is implemented completely at language-side and it is integrated
in Pharo. Waterfall generates primitives and plug-ins
on demand and executes them on the fly.
Even though Waterfall still lacks substantial optimizations it outperforms dynamic instrumentation of
primitives based on purely reflective approaches. It also
enables to have dynamically generated plug-ins which
9
[1] B. Alpern, C. R. Attanasio, A. Cocchi, D. Lieber,
S. Smith, T. Ngo, J. J. Barton, S. F. Hummel, J. C.
Sheperd, and M. Mergen. Implementing jalapeño in
java. SIGPLAN Not., 34(10):314–324, Oct. 1999.
[2] B. Alpern, S. Augart, S. M. Blackburn, M. Butrico,
A. Cocchi, P. Cheng, J. Dolby, S. Fink, D. Grove,
M. Hind, K. S. McKinley, M. Mergen, J. E. B. Moss,
T. Ngo, and V. Sarkar. The jikes research virtual machine project: building an open-source research community. IBM Syst. J., 44(2):399–417, Jan. 2005.
[3] M. B. Ballard, D. Maier, and A. W. Brock. QUICKTALK: a Smalltalk-80 dialect for defining primitive
methods. SIGPLAN Not., 21(11):140–150, June 1986.
2018/4/2
[4] A. Black, S. Ducasse, O. Nierstrasz, D. Pollet, D. Cassou, and M. Denker. Pharo by Example. Square Bracket
Associates, 2009.
[5] C. Bruni, G. Chari, S. Ducasse, and I. Stasenko. Reflective low-level programming. Submitted to AOSD’14,
2013.
[6] C. Bruni, L. Fabresse, S. Ducasse, and I. Stasenko.
Language-side foreign function interfaces with nativeboost. In International Workshop on Smalltalk Technologies 2013, 2013.
[7] M. Chevalier-Boisvert, E. Lavoie, M. Feeley, and B. Dufour. Bootstrapping a self-hosted research virtual machine for javascript: an experience report. In Proceedings
of the 7th symposium on Dynamic languages, DLS ’11,
pages 61–72, New York, NY, USA, 2011. ACM.
[8] B. Folliot, I. Piumarta, and F. Riccardi. Virtual Virtual
Machines. pages 1–4.
[9] D. Frampton, S. M. Blackburn, P. Cheng, R. J. Garner,
D. Grove, J. E. B. Moss, and S. I. Salishev. Demystifying magic: high-level low-level programming. In Proceedings of the 2009 ACM SIGPLAN/SIGOPS international conference on Virtual execution environments,
VEE ’09, pages 81–90, New York, NY, USA, 2009.
ACM.
[10] A. Goldberg and D. Robson. Smalltalk-80: the language
and its implementation. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1983.
[11] D. Ingalls, T. Kaehler, J. Maloney, S. Wallace, and
A. Kay. Back to the future: The story of Squeak, a practical Smalltalk written in itself. In OOPSLA’97: Proceedings of the 12th International Conference on ObjectOriented Programming, Systems, Languages, and Applications, pages 318–326. ACM Press, Nov. 1997.
Co., Inc., Boston, MA, USA, 1st edition, 1999.
[13] J. Malenfant, M. Jacques, and F. N. Demers. A tutorial on behavioral reflection and its implementation.
Proceedings of the Reflection ’96 Conference, 1996.
[14] J. S. Miller and S. Ragsdale. The Common Language
Infrastructure Annotated Standard. Addison-Wesley
Longman Publishing Co., Inc., Boston, MA, USA, 2003.
[15] E. Miranda. The Cog Smalltalk virtual machine. In
VMIL ’11: Proceedings of the 5th workshop on Virtual
machines and intermediate languages for emerging modularization mechanisms. ACM, 2011.
[16] Pharo Smalltalk environment. http://pharo.org.
[17] A. Rigo and S. Pedroni. Pypy’s approach to virtual
machine construction. In Companion to the 21st ACM
SIGPLAN symposium on Object-oriented programming
systems, languages, and applications, OOPSLA ’06,
pages 944–953, New York, NY, USA, 2006. ACM.
[18] D. Röthlisberger, M. Denker, and É. Tanter. Unanticipated partial behavioral reflection. In Advances in
Smalltalk — Proceedings of 14th International Smalltalk
Conference (ISC 2006), volume 4406 of LNCS, pages
47–65. Springer, 2007.
[19] D. Ungar, A. Spitz, and A. Ausch. Constructing a
metacircular virtual machine in an exploratory programming environment. In OOPSLA ’05: Companion to the 20th annual ACM SIGPLAN conference on
Object-oriented programming, systems, languages, and
applications, pages 11–20, New York, NY, USA, 2005.
ACM.
[12] S. Liang. Java Native Interface: Programmer’s Guide
and Reference. Addison-Wesley Longman Publishing
[20] C. Wimmer, M. Haupt, M. L. Van De Vanter, M. Jordan, L. Daynès, and D. Simon. Maxine: An approachable virtual machine for, and in, java. ACM Trans.
Archit. Code Optim., 9(4):30:1–30:24, Jan. 2013.
10
2018/4/2
| 6cs.PL
|
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER
RINGS
arXiv:1801.08748v1 [math.GR] 26 Jan 2018
ANASTASIA STAVROVA AND ALEXEI STEPANOV
Abstract. The paper studies the lattice of subgroups of an isotropic reductive group G(R)
over a commutative ring R, normalized by the elementary subgroup E(R). We prove the
sandwich classification theorem for this lattice under the assumptions that the reductive
group scheme G is defined over an arbitrary commutative ring, its isotropic rank is at least
2, and the structure constants are invertible in R. The theorem asserts that the lattice splits
into a disjoint union of sublattices (sandwiches) E(R, q) 6 · · · 6 C(R, q) parametrized by
the ideals q of R, where E(R, q) denotes the relative elementary subgroup and C(R, q) is the
inverse image of the center under the natural homomorphism G(R) → G(R/I). The main
ingredients of the proof are the “level computation” by the first author and the universal
localization method developed by the second author.
1. Introduction
Let G be a reductive group scheme over a unital commutative ring K in the sense of [DG70].
The famous result of J. Tits [Tit64] establishes that if K is a field, and G has no normal
closed connected smooth K-subgroups, then its group of K-points G(K) is very close to
being simple as an abstract group (except in a few cases where K = F2 or F3 ). Namely,
G(K) contains a “large” normal subgroup G(K)+ whose central quotient is simple. If K is
a finite field, the corresponding simple group G(K)+ /C(G(K)+ ) is a finite simple group of
Lie type, and in fact such groups constitute the largest family in the classification of finite
simple groups [Wil09].
If K is no longer a field, then any proper ideal q of K determines a normal subgroup
G(K, q) of G(K) called the congruence subgroup of level q; this subgroup is the kernel of the
natural homomorphism ρq : G(K) → G(K/q). Thus G(K) is no longer simple, and its lattice
of normal subgroups is at least as rich as the lattice of ideals in K. The goal of the present
paper is to show that, under the natural assumptions on G and K, essentially, all normal
subgroups of G(K) are congruence subgroups. In order to state our result in a precise form,
we need a few more definitions and notations.
We say that G has isotropic rank > n, if every semisimple normal K-subgroup of G contains
a n-dimensional split K-torus (Gm,K )n . If the isotropic rank is > 1, then G contains a pair
of opposite parabolic K-subgroups P ± [DG70], and one defines the elementary subgroup
EP (K) as the subgroup of G(K) generated by UP + (K) and UP − (K), where UP ± denotes
the unipotent radical of P ± . If, moreover, the isotropic rank of G is > 2, the main result
of [PS09] implies that E(K) = EP (K) is independent of the choice of P ± and normal in
G(K) (see § 3 for the details). Under these assumptions, if G = GLn , then the elementary
subgroup is the subgroup generated by the elementary transvections e + teij , 1 6 i 6= j 6 n,
t ∈ K. And if K is a field, then E(K) = G(K)+ is the above-mentioned group of J. Tits.
1991 Mathematics Subject Classification. 20G35.
Key words and phrases. isotropic reductive groups; parabolic subgroup; elementary subgroup; congruence
subgroup; unipotent element; generic element; universal localization; normal structure.
This publication is supported by Russian Science Foundation grant 17-11-01261. The first author is a
winner of the contest “Young Russian Mathematics”.
1
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
2
Assume that the isotropic rank of G is > 2. For any ideal q of K, let UP ± (q) be the kernel
of the natural homomorphism ρq restricted to UP ± (K). Denote by EP (K, q) the normal
closure of hUP + (q), UP − (q)i in E(K). Also, denote by C(K, q) the full preimage of the center
of G(K/q) under ρq .
For any maximal ideal m of K, denote by K/m the algebraic closure of the field K/m. The
group GK/m is a reductive algebraic group in the usual sense [Bor91], and thus has a root
system Φ in the sense of Bourbaki. Note that GK/m has no closed connected normal smooth
K-subgroups if and only if Φ is irreducible. The structure constants of Φ are, by definition,
the integers ±1, together with ±2, if Φ is of type Bn , Cn , F4 , or ±2, ±3, if Φ is of type G2 .
The main result of the present paper is the following theorem.
Theorem 1.1. Let G be a reductive group scheme over a ring K such that the isotropic rank
of its derived group scheme [G, G] is > 2. Suppose that for any maximal ideal m of K the
root system of GK/m is irreducible, and its structure contants are invertible in K. Then
(i) For any ideal q of K one has EP (K, q) = [G(K, q), E(K)]. In particular, EP (K, q) =
E(K, q) is independent of the choice of a parabolic K-subgroup P .
(ii) For any subgroup H 6 G(K) normalized by E(K), there exists a unique ideal q in K
such that
E(K, q) 6 H 6 C(K, q).
Since groups of points of reductive group schemes include, in particular, the linear matrix groups GLn (K), SLn (K), Sp2n (K), as well as spinor and special orthogonal groups
Spin(q)(K), SO(q)(K), where q is a non-degenerate quadratic K-form, the study of their
normal structure has a very long history. Thus, it would take a separate survey paper to
describe it in full, and below we only list a few milestone results.
• Simplicity of the groups PSLn (F ) was proved by Camille Jordan around 1870 (for
prime fields) and in early 1900 by Leonard Dickson (for all finite fields).
• A similar result for the GLn over a skew-field was obtained by J.Dieudonne in 1943.
In [Die48] he established the simplicity of split groups over arbitrary fields. The first
uniform proof of their simplicity was given by C. Chevalley [Che55].
• The problem over rings distinct from fields was first approached by J.Brenner [Bre38,
Bre44, Bre60] and later by W.Klingenberg [Kli61a, Kli61b, Kli63]. They studied split
groups of classical types over quotients of Z and general local rings respectively.
• As mentioned above, J. Tits established the simplicity for isotropic reductive groups
over fields in [Tit64].
• For the general linear group over an arbitrary ring, the normal structure theorem was
proved by H. Bass [Bas64] under the stable range condition. The result was generalized
to other quasi-split classical groups by H.Bass himself [Bas73] and A.Bak [Bak69].
• H. Bass, M. Lazard, J.-P. Serre, and J. Milnor [BLS64, BMJP67] elucidated the normal structure of SLn and Sp2n over rings of integers of algebraic number fields, and
stated the Congruence subgroup problem. It was solved for all split groups by H.
Matsumoto [Mat69].
• The stable range condition for GLn was removed by J. Wilson [Wil72] (n > 4) and
I. Golubchik [Gol73] (n > 3). Using the ideas of H. Bass and Suslin’s theorem on the
normality of the elementary group [Sus77], Z. Borewich and N. Vavilov [BV85] gave a
simpler prove of the Wilson–Golubchik theorem.
• In 1974 E. Abe and K. Suzuki proved the normal structure theorem for all Chevalley
groups over local rings [AS76].
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
3
• In 1976 M.S. Raghunathan established the Congruence subgroup problem for groups
of isotropic rank > 2 over a global field [Rag76]. In 1986 he improved this result,
weakening the isotropy conditions [Rag86].
• In 1979 G. Margulis [Mar79] proved his celebrated theorem on lattices in isotropic
groups over local fields. His work was based on earlier work of V. P. Platonov [Pla69].
• In 1980 R. Bix established the normal structure theorem for isotropic groups of type
1 28
E6,2 over local rings [Bix80].
• L. Vaserstein in [Vas81] combined the results of H. Bass, J. Wilson and I. Golubchik
proving the standard normal structure of the GLn under a local stable rank condition.
• After the result of G. Taddei on normality of the elementary group in a Chevalley group
L. Vaserstein [Vas86] proved the standard normal sructure of Chevalley groups of rank
> 2 over commutative rings provided that the structure constants were invertible. The
latter condition was removed by E. Abe in [Abe89].
• In 1988 L. Vaserstein proved the normal structure theorem for isotropic orthogonal
groups over commutative rings [Vas88].
• The result of E. Abe has three exceptions where the elementary group is not perfect:
types C2 and G2 if the gound ring has a residue field of 2 elements, and type A1 . All
the exceptions were considered by D. Costa and G. Keller in a series of papers [CK91a,
CK91b, CK99]. Of course, for type A1 the ground ring must be low-dimensional
with many units, as the description of the normal structure of SL2 (Z) seems to be an
unrealistic problem.
• In 2010 Z. Zhang [Zha10] established the normal structure theorem for even hyperbolic
unitary groups over a commutative form ring with invertible 2. The latter assumption
was removed by Hong You [You12]. A shorter proof was obtained recently by R.
Preusser [Pre18].
2. Principal notation
Let x, y, z be elements of an abstract group G. Denote by xy = y −1xy the element conjugate
−1
to x by y. Sometimes the same element is denoted by y x. The commutator x−1 y −1xy is
denoted by [x, y]. In the sequel we frequently use the following commutator identity, which
can be easily verified by a straightforward calculation. Let x, y, z be elements of an abstract
group G. Then
(2.1)
[x, yz]z
−1
−1
= (x−1 )z xy = [z −1 , x] · [x, y].
Let S be a subset of G. By hSi we denote the subgroup spanned by S. For subgroups X
and Y of G by X Y we denote the subgroup of G generated by xy for all x ∈ X and y ∈ Y . In
other words, X Y is the smallest subgroup containing X and normalized by Y . The mutual
commutator subgroup [X, Y ] is a subgroup of G, generated by all the commutators [x, y],
x ∈ X, y ∈ Y . The center of an abstract group G is denoted by C(G).
All rings and algebras are assumed to be commutative and to contain a unit. All homomorphisms preserve unit elements. The multiplicative group of a ring R is denoted by R∗ . As
usual, Spec R denotes the prime spectrum of R. For p ∈ Spec R denote by Rp = (R r p)−1 R
the localization of R at p and by k(p) the residue field Rp /pRp . The algebraic closure of k(p)
is denoted by k(p).
Let s ∈ R. The principal localization at the element s (i. e. the localization at the
multiplicative subset generated by s) is denoted by Rs . The localization homomorphism is
denoted by λp or λs respectively.
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
4
Throughout the paper K is a ring, R denotes a K-algebra and G stands for a reductive
group scheme over K in the sense of [DG70], unless explicitly stated otherwise.
For any ideal q of R we denote by ρq : R → R/q the reduction homomorphism, and, by
abuse of notation, the induced homomorphism G(R) → G(R/q). The principal congruence
subgroup G(R, q) is the kernel of ρq : G(R) → G(R/q), whereas the full congruence subgroup
C(R, q) is the inverse image of the center C(G(R/q)) of G(R/q) under this homomorphism.
3. Elementary subgroup of an isotropic reductive group
Let P be a parabolic subgroup of G in the sense of [DG70]. Since the base Spec K is
affine, the group P has a Levi subgroup LP [DG70, Exp. XXVI Cor. 2.3]. There is a unique
parabolic subgroup P − in G which is opposite to P with respect to LP , that is P − ∩ P = LP ,
cf. [DG70, Exp. XXVI Th. 4.3.2]. We denote by UP the unipotent radical of P .
Note that if L′P is another Levi subgroup of P , then L′P and LP are conjugate by an
element u ∈ UP (K) [DG70, Exp. XXVI Cor. 1.8]. Since our proofs in the present paper do
not depend on a particular choice of LP or P − , we do not pay attention to this choice.
Definition. The elementary subgroup EP (R) corresponding to P is the subgroup of G(R)
generated as an abstract group by UP (R) and UP − (R).
Definition. A parabolic subgroup P in G is called strictly proper, if it intersects properly
every normal semisimple subgroup of G.
The following theorem is the main result of [PS09].
Theorem 3.1 ([PS09, Theorem 1]). Assume that for every maximal ideal m of K every
normal semisimple subgroup of GKm contains (Gm,Km )2 . Then the subgroup EP (R) of G(R) =
GR (R) is the same for any strictly proper parabolic R-subgroup P of GR . In particular, EP (R)
is normal in G(R).
Definition. Under the assumptions of Theorem 3.1 we call EP (R) the elementary subgroup
of G(R) and denote it by E(R).
We also use the following theorems, which are the main results of [KS13] and [LS12]. We
denote by C(G) the group scheme center of G in the sense of [DG70], see also discussion
around Proposition 6.7 in [Mil12]. The very definition of C(G) implies that C(G)(R) 6
C(G(R)).
Theorem 3.2 ([KS13, Theorem 1]). Under the hypothesis of Theorem 3.1, one has
CG(R) (E(R)) = C(G)(R) = C(G(R)).
Theorem 3.3 ([LS12, Theorem 1]). Under the hypothesis of Theorem 3.1, assume, moreover,
that for any maximal ideal m of R, one has k(m) 6= F2 whenever the root system of Gk(m)
contains an irreducible component of type B2 = C2 or G2 . Then E(R) = [E(R), E(R)].
The following statement will be used twice to get rid of the center.
Lemma 3.4. Let H be a subgroup
of G(R), normalized by E(R). Under the assumptions of
Theorem 3.3 [H, E(R)], E(R) = [H, E(R)]. In particular, if [H, E(R)] 6 C(G(R)), then
H 6 C(G(R)).
Proof. By Theorem 3.3 the group E(R) is perfect. Using the Hall–Witt identity we get
E(R), H = [E(R), E(R)], H 6 [H, E(R)], E(R) .
The inverse inclusion is obvious. The second assertion follows immediately from the first one
and Theorem 3.2.
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
5
4. Root systems corresponding to parabolic subgroups
±1
Let S = (Gm,K )N = Spec(K[x±1
1 , . . . , xN ]) be a split N-dimensional torus over R. Recall
that the character group X ∗ (S) = HomK (S, Gm,K ) of S is canonically isomorphic to ZN . If
S acts K-linearly on a K-module V , this module has a natural ZN -grading
M
V =
Vλ ,
λ∈X ∗ (S)
where
Vλ = {v ∈ V | s · v = λ(s)v for any s ∈ S(K)}.
N
Conversely, any Z -graded K-module V can be provided with an S-action by the same rule.
Assume that S acts on G by K-group automorphisms. The associated Lie algebra functor
Lie(G) then acquires a ZN -grading compatible with the Lie algebra structure,
M
Lie(G) =
Lie(G)λ .
λ∈X ∗ (S)
We will use the following version of [DG70, Exp. XXVI Prop. 6.1].
Lemma 4.1. Let L = CG (S) be the subscheme of G fixed by S. Let Ψ ⊆ X ∗ (S) be a
K-subsheaf of sets closed under addition of characters. Then there exists a unique smooth
connected closed subgroup UΨ of G normalized by L and satisfying
M
(4.1)
Lie(UΨ ) =
Lie(G)λ .
λ∈Ψ
Moreover,
1.
2.
3.
4.
if 0 ∈ Ψ, then UΨ contains L;
if Ψ = {0}, then UΨ = L;
if Ψ = −Ψ, then UΨ is reductive;
if Ψ ∪ (−Ψ) = X ∗ (S), then UΨ and U−Ψ are two opposite parabolic subgroups of G with
the common Levi subgroup UΨ∩(−Ψ) ;
5. If 0 ∈
/ Ψ, then UΨ is unipotent.
Proof. The statement immediately follows by faithfully flat descent from the standard facts
about the subgroups of split reductive groups proved in [DG70, Exp. XXII]; see the proof
of [DG70, Exp. XXVI Prop. 6.1].
Definition. The sheaf of sets
Φ(S, G) = {λ ∈ X ∗ (S) r {0} | Lie(G)λ 6= 0}
is called the system of relative roots of G with respect to S.
Remark. Choosing a total ordering on the Q-space Q ⊗Z X ∗ (S) ∼
= Qn , one defines the
subsets of positive and negative relative roots Φ(S, G)+ and Φ(S, G)− , so that Φ(S, G) is a
disjoint union of Φ(S, G)+ , Φ(S, G)− , and {0}. By Lemma 4.1 the closed subgroups
UΦ(S,G)+ ∪{0} = P,
UΦ(S,G)− ∪{0} = P −
are two opposite parabolic subgroups of G with the common Levi subgroup CG (S). Thus, if
a reductive group G over R admits a non-trivial action of a split torus, then it has a proper
parabolic subgroup. The converse is true Zariski-locally, see Lemma 4.2 below.
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
6
Let P be a parabolic subgroup scheme of G over K, and let L be a Levi subgroup of P .
By [DG70, Exp. XXII, Prop. 2.8] the root system Φ of Gk(p) , p ∈ Spec K, is locally constant
in the Zariski topology on Spec K. The type of the root system of Lk(p) is determined by
a Dynkin subdiagram of the Dynkin diagram of Φ, which is also constant Zariski-locally on
Spec K by [DG70, Exp. XXVI, Lemme 1.14 and Prop. 1.15]. In particular, if Spec K is
connected, all these data are constant on Spec K.
Definition. Assume that the root system Φ of Gk(p) has the same type for all p ∈ Spec K.
Then we call Φ the absolute root system of G.
Lemma 4.2 ([Sta16, Lemma 3.6]). Assume that K is connected. Let L̄ be the image of L
under the natural homomorphism G → Gad ⊆ Aut(G). Let D be the Dynkin diagram of the
absolute root system Φ of G. We identify D with a set of simple roots of Φ such that Pk(p)
is a standard positive parabolic subgroup with respect to D. Let J ⊆ D be the set of simple
roots such that D r J is the subdiagram of D corresponing to Lk(p) . Then there are a unique
maximal split subtorus S ⊆ C(L̄) and a subgroup Γ 6 Aut(D) such that J is invariant under
Γ and for any p ∈ Spec R and any split maximal torus T ⊆ L̄k(p) the kernel of the natural
surjection
(4.2)
π
X ∗ (T ) ∼
= ZΦ −−→ X ∗ (Sk(p) ) ∼
= ZΦ(S, G)
is generated by all roots α ∈ D r J, and by all differences α − σ(α), α ∈ J, σ ∈ Γ.
In [PS09], we introduced a system of relative roots ΦP with respect to a parabolic subgroup
P . This system ΦP was defined independently over each member Spec K = Spec Ki of a
suitable finite disjoint Zariski covering
Spec K =
m
a
Spec Ki ,
i=1
such that over each Ki , 1 6 i 6 m, the root system Φ and the Dynkin diagram D of G is
constant. Namely, we considered the formal projection
πJ,Γ : ZΦ −→ ZΦ/hD r J; α − σ(α) | α ∈ J, σ ∈ Γi,
and set ΦP = ΦJ,Γ = πJ,Γ (Φ) r {0}. The last claim of Lemma 4.2 allows to identify ΦJ,Γ and
Φ(S, G) whenever Spec K is connected.
Definition. In the setting of Lemma 4.2 we call Φ(S, G) ∼
= ΦJ,Γ a system of relative roots
with respect to the parabolic subgroup P over R and denote it by ΦP .
If K is a field or a local ring, and P is a minimal parabolic subgroup of G, then ΦP is
nothing but the relative root system of G with respect to a maximal split subtorus in the
sense of [BT65] or, respectively, [DG70, Exp. XXVI §7].
We have also defined in [PS09] irreducible components of systems of relative roots, the
subsets of positive and negative relative roots, simple relative roots, and the height of a root.
These definitions are immediate analogs of the ones for usual abstract root systems, so we
do not reproduce them here.
We will need later the following two lemmas on relative roots.
Lemma 4.3. Let Φ be a root system, and let ΦJ,Γ = π(Φ) r {0} be a relative root system
with the canonical projection π : ZΦ → ZΦJ,Γ . Let α, β ∈ ΦJ,Γ be two simple relative roots
such that α + β ∈ ΦJ,Γ . Then for any j > 1 such that jβ ∈ ΦJ,Γ one has α + jβ ∈ ΦJ,Γ .
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
7
Proof. The inclusion α + β ∈ ΦJ,Γ is equivalent to the existence of two simple roots in π −1 (β)
and π −1 (α) respectively that are connected in the Dynkin diagram by a chain (possibly,
empty) of simple roots lying in π −1 (0). Then for any element µ ∈ π −1 (jβ) the set Sµ ⊆ Φ of
simple roots occurring in its decomposition is also connected to a simple root in π −1 (α) by
a chain of simple roots lying in π −1 (0), which allows to find a root ν ∈ π −1 (α) such that Sµ
and Sν are disjoint but adjacent subsets of the Dynkin diagram of Φ, and hence µ + ν ∈ Φ.
Then π(µ + ν) = α + jβ ∈ ΦJ,Γ .
Lemma 4.4. Let Φ be a root system with the scalar product ( , ) and a system of simple
roots Π, and let ΦJ,Γ = π(Φ) r {0} be a relative root system with the canonical projection
π : ZΦ → ZΦJ,Γ . Let β ∈ π(Π) ∩ ΦJ,Γ be a simple relative root. Then there is a proper
parabolic subset Σ of ΦJ,Γ that contains all α ∈ ΦJ,Γ such that α + β ∈
/ ΦJ,Γ . If Φ is simply
laced, then
X
Σ = {α ∈ ΦJ,Γ | (a,
b) > 0 for all a ∈ π −1 (α)}
b∈π −1 (β)
(4.3)
= {α ∈ ΦJ,Γ | (a,
X
b) > 0 for some a ∈ π −1 (α)}.
b∈π −1 (β)
Proof. If Φ is not simply laced, then there is a simply laced root system Φ′ with a system of
simple roots Π′ and a projection π ′ : ZΦ′ → ZΦ such that Φ = Φ′Π′ ,Γ′ for a suitable group of
automorphisms Γ′ of the Dynkin diagram of Φ′ . Then ΦJ,Γ = Φ′π′−1 (J),Γ′′ , where Γ′′ is a group
of automorphisms of the Dynkin diagram of Φ generated by Γ′ and the natural lifting of Γ.
This means that we can assume that Φ is simply laced from the start.
For any α ∈ ΦJ,Γ and any µ ∈ π −1 (α) by [PS09, Lemma 3] the set π −1 (α) is a union of
roots µ′ ∈ Φ such that σ(µJ ) = µ′J for some σ ∈ Γ, where µJ denotes the shape of µ with
respect to J in the sense of [PS09] or [ABS90]. By [ABS90, Lemma 1] the Weyl group of ΠrJ
acts transitively on the set of roots of a fixed shape and length, hence this Weyl group and
Γ together act transitively on π −1 (α). Since all elements of these groups
P are bijections that
preserve scalar products and the set π −1 (β), we conclude that (µ,
ν) is the same for
ν∈π −1 (β)
−1
all µ ∈ π (α). Thus, the set Σ of (4.3) is well-defined, and it is clear that Σ ∪ (−Σ) = ΦJ,Γ .
Assume that α1 , α2 ∈ Σ and α1 +α2 ∈ ΦJ,Γ . By [PS09, Lemma 4] for every µ ∈ π −1 (α1 +α2 )
there are µ1 ∈ π −1 (α1 ) and µ2 ∈ π −1 (α2 ) such that µ = µ1 + µ2 . This shows that Σ is
additively closed. By the same argument, if α + β ∈
/ ΦJ,Γ , then (µ, ν) > 0 for all µ ∈ π −1 (α)
−1
and ν ∈ π (β), and hence α ∈ Σ.
It remains to show that Σ is a proper subset of ΦJ,Γ . Let α ∈ π(Π) r {0} be a simple
relative root such that α + β ∈ ΦJ,Γ . Since α − β ∈
/ ΦJ,Γ , we conclude that (µ, ν) 6 0 for all
µ ∈ π −1 (α) and ν ∈ π −1 (β). On the other hand, since α + β ∈ ΦJ,Γ , there are two simple
roots µ ∈ π −1 (α) ∩ Π and ν ∈ π −1 (β) ∩ Π which are connected on the Dynkin diagram by
a chain of roots in Π r J. Then the scalar product of µ and (nu+all roots in this chain) is
negative. Hence α ∈
/ Σ.
5. Relative root subschemes
For any finitely generated projective K-module V , we denote by W (V ) the natural affine
scheme over K associated with V , see [DG70, Exp. I, §4.6] or [Mil12, Ch. IV, § 1, 1.6]. Any
morphism of K-schemes W (V1 ) → W (V2 ) is determined by an element f ∈ Sym∗ (V1∨ ) ⊗K V2 ,
where Sym∗ denotes the symmetric algebra, and V1∨ denotes the dual module of V1 . If
f ∈ Symd (V1∨ ) ⊗K V2 , we say that the corresponding morphism is homogeneous of degree d.
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
8
By abuse of notation, we also write f : V1 → V2 and call it degree d homogeneous polynomial
map from V1 to V2 . In this context, one has
f (rv) = r d f (v)
for any v ∈ V1 and r ∈ K.
Lemma 5.1. [Sta16, Lemma 3.9]. In the setting of Lemma 4.2, for any α ∈ ΦP = Φ(S, G)
there exists a closed S-equivariant embedding of K-schemes
Xα : W Lie(G)α → G,
satisfying the following condition.
(∗) Let R/K be any ring extension such that GR is split with respect to a maximal split
R-torus T ⊆ LR . Let eδ , δ ∈ Φ, be a Chevalley basis of Lie(GR ), adapted to T and P ,
and xδ : Ga → GR , δ ∈ Φ, be the associated system of 1-parameter root subgroups (e.g.
xδ = expδ of [DG70, Exp. XXII, Th. 1.1]). Let
π : Φ = Φ(T, GR ) → ΦP ∪ {0}
P
be the natural projection. Then for any u =
aδ eδ ∈ Lie(GR )α one has
δ∈π −1 (α)
(5.1)
Xα (u) =
Y
Y
xδ (aδ ) ·
i>2
δ∈π −1 (α)
Y
xθ (piθ (u)) ,
θ∈π −1 (iα)
where every piθ : Lie(GR )α → R is a homogeneous polynomial map of degree i, and the
products over δ and θ are taken in any fixed order.
Definition. Closed embeddings Xα , α ∈ ΦP , satisfying the statement of Lemma 5.1, are
called relative root subschemes of G with respect to the parabolic subgroup P .
Relative root subschemes of G with respect to P , actually, depend on the choice of a Levi
subgroup LP , but their essential properties stay the same, so we usually omit LP from the
notation.
Set Vα = Lie(G)α for short. We will use the following properties of relative root subschemes.
Lemma 5.2. [PS09, Theorem 2, Lemma 6, Lemma 9] Let Xα , α ∈ ΦP , be as in Lemma 5.1.
Then
(i) There exist degree i homogeneous polynomial maps qαi : Vα ⊕ Vα → Viα , i > 1, such that
for any K-algebra R and for any v, w ∈ Vα ⊗K R one has
Y
(5.2)
Xα (v)Xα (w) = Xα (v + w)
Xiα qαi (v, w) .
i>1
(ii) For any g ∈ L(K), there exist degree i homogeneous polynomial maps ϕig,α : Vα → Viα ,
i > 1, such that for any K-algebra R and for any v ∈ Vα ⊗R R′ one has
Y
gXα(v)g −1 =
Xiα ϕig,α(v) .
i>1
(iii) (generalized Chevalley commutator formula) For any α, β ∈ ΦP such that mα 6= −kβ
for all m, k > 1, there exist polynomial maps
Nαβij : Vα × Vβ → Viα+jβ , i, j > 0,
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
9
homogeneous of degree i in the first variable and of degree j in the second variable, such that
for any K-algebra R and for any for any u ∈ Vα ⊗K R, v ∈ Vβ ⊗K R one has
Y
(5.3)
[Xα (u), Xβ (v)] =
Xiα+jβ Nαβij (u, v)
i,j>0
∗
(iv) For any subset Ψ ⊆ X (S) r {0} that is closed under addition, the morphism
M
Y
XΨ : W
Vα → UΨ ,
(vα )α 7→
Xα (vα ),
α∈Ψ
α
where the product is taken in any fixed order, is an isomorphism of schemes.
Apart from the above properties of relative root subschemes we will use the following
Lemma, which appeared first as [PS09, Lemma 10] and in a slighly stronger form in [LS12,
Lemma 2]. Note that in both cases the original statements erroneously claim that the image
Im(Nα,β,1,1 ) (respectively, the sum of images in (2) ) equals Vα+β , while in reality the respective
proofs establish only that it generates Vα+γ as an R-module, and hence as an abelian group.
The correct, weaker statement is as follows. One can easily check it is still enough for all the
applications in [PS09, LS12, Sta14b].
Lemma 5.3. Consider α, β ∈ ΦP satisfying α + β ∈ ΦP and mα 6= −kβ for any m, k > 1.
Denote by Φ0 an irreducible component of Φ such that α, β ∈ π(Φ0 ).
(1) In each of the following cases, Im(Nαβ11 ) generates Vα+β as an abelian group:
(a) structure constants of Φ0 are invertible in R (for example, if Φ0 is simply laced);
(b) α 6= β and α − β ∈
/ ΦP ;
0
(c) Φ is of type Bl , Cl , or F4 , and π −1 (α + β) consists of short roots;
(d) Φ0 is of type Bl , Cl , or F4 , and there exist long roots α ∈ π −1 (α), β ∈ π −1 (β) such
that α + β is a root.
(2) If α−β ∈ ΦP and Φ0 6= G2 , then Im(Nαβ11 ), Im(Nα−β,2β,1,1 ), and Im Nα−β,β,1,2 (−, v)
for all v ∈ Vβ together generate Vα+β as an abelian group. Here we assume Im(Nα−β,2β,1,1 ) = 0
if 2β ∈
/ ΦP .
6. Reduction to extraction of unipotents
In this section we establish two important reduction statements. Throughout this section,
we assume that G is a reductive group scheme over a connected commutative ring K, and
R is an arbitrary K-algebra. Let P is a proper parabolic K-subgroup of G, and let ΦP = ΦJ,Γ
be the system of relative roots for P , and Xα (Vα ), α ∈ ΦP , be the relative root subschemes
of G with respect to P that exist by Lemmas 4.2 and 5.1.
Recall that for any ideal q of R we set UP (q) = UP (R) ∩ G(R, q) and denote by EP (R, q)
the normal closure of hUP + (q), UP − (q)i in EP (R).
Lemma 6.1. Assume that for any maximal ideal m of K every semisimple normal subgroup of group GKm contains (Gm ,Km )2 . Then for any K-algebra R the group EP (R, q) is
independent of the choice of a strictly proper parabolic subgroup P .
Proof. Clearly, it is enough to show that for any two strictly proper parabolic subgroups P, Q
one has UP (q) 6 EQ (R, q). Consider the ring of polynomials R[x]. For any x ∈ UP (xR[x])
we have
x ∈ G(R[x], xR[x]) ∩ EP (R[x]) = G(R[x], xR[x]) ∩ EQ (R[x]) = EQ (R[x], xR[x])
by the splitting principle for isotropic groups [Sta14b, Lemma 4.1]. Now let h ∈ UP (q) be
any element. Then, clearly, h is a product of elements Xα (cv), where α ∈ ΦP , c ∈ q and
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
10
v ∈ Vα . By the previous argument we have Xα (xv) ∈ EQ (R[x], xR[x]). Specializing x to c,
we conclude that Xα (cv) ∈ EQ (R, q), as required.
Definition. In the setting of Lemma 6.1 we denote EP (R, q) by E(R, q). We say that the
normal structure of the group G(R) is standard if for each subgroup H 6 G(R) normalized
by E(R), there exists a unique ideal q of R such that
E(R, q) 6 H 6 C(R, q).
We are going to deduce the existence of standard normal structure from the fact that
any H as above contains an elementary root unipotent. Our starting point is the following
theorem established by the first author.
Theorem 6.2 ([Sta14a, Theorem 2]). Assume that the structure constants of the absolute root
system Φ of G are invertible in K, and for any maximal ideal m of K every semisimple normal
subgroup of group GKm contains (Gm,Km )2 . Let P be a strictly proper parabolic K-subgroup
of G, and let E(K) = EP (K) be the elementary subgroup of G(K). Then for any normal
subgroup N 6 E(K) there exists an ideal q = q(N) in R such that N ∩ Xα (Vα ) = Xα (qVα )
for any α ∈ ΦP .
Using this theorem, we prove the following statement. The idea of this reduction goes back
to [Bas64].
Proposition 6.3. In the setting of Theorem 6.2, suppose further that every irreducible component of ΦP contains more than 2 distinct roots, and that for any quotient ring K/q of
K and any noncentral subgroup H 6 G(K/q) normalized by E(K/q) there is α ∈ ΦP and
0 6= u ∈ Vα ⊗K K/q such that Xα (u) ∈ H. Then the normal structure of the group G(K) is
standard.
In order to prove Proposition 6.3, we need the following two statements.
Lemma 6.4. Let α, β ∈ ΦP be two relative roots such that α + β ∈ ΦP and nα 6= −mβ for
all n, m > 1. Assume moreover that α − β ∈
/ ΦP , or the structure constants of the absolute
root system Φ of G are invertible in K. Take 0 6= u ∈ Vβ . Any generating system e1 , . . . , en
of the K-module Vα contains an element ei such that Nαβ11 (ei , u) 6= 0.
`
Proof. Consider an affine fpqc-covering
Spec Rτ → Spec K that splits G. There is a
member Rτ = R of this covering such that the image of Xβ (u) under G(K) → G(R) is
non-trivial. Write
Y
Y Y
Xβ (u) =
xν (aν ) ·
xν (cν ),
π(ν)=β
i>2 π(ν)=iβ
where π : Φ → ΦP ∪ {0} is the canonical projection of the absolute root system of G onto
the relative one, xβ are root subgroups of the split group GR , and aν ∈ R. Since Xβ (u) 6= 0,
the definition of Xβ implies that there exists aν 6= 0. By [PS09, Lemma 4] there
Q exists
Q a root
µ ∈ π −1 (α) such that µ+ν ∈ Φ. Let v ∈ VA ⊗K R be such that Xα (v) = xµ (1)
xη (dη ),
i>2 π(η)=iα
for some dη ∈ R. Then the (usual) Chevalley commutator formula implies that [Xα (v), Xβ (u)]
contains in its decomposition a factor xµ+ν (εaν ), where ε is a structure constant of Φ.
If α−β ∈
/ ΦP , then ε = ±1, otherwise ε is invertible by assumptions. Hence Nαβ11 (v, u) 6= 0.
Since Nαβ11 (v, u) is linear in the first argument, this implies the result.
Lemma 6.5. Under the hypothesis of Proposition 6.3 one has
[G(K, q), E(K)] = EP (K, q).
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
11
Proof. One proves that [G(K, q), E(K)] 6 EP (K, q) exactly as [Ste16, Proposition 5.1], using the splitting principle for isotropic groups [Sta14b, Lemma 4.1]. To prove the inverse
inclusion, by Theorem 6.2 it is enough to show that Xα̃ (Vα̃ ⊗K q) ⊆ [G(K, q), E(K)], where
+
α̃ ∈ Φ+
P is a maximal root. Let β ∈ ΦP be a simple relative root such that α̃ − β ∈ ΦP . Then
by Lemma 5.3 and the generalized Chevalley commutator formula one has Xα̃ (Vα̃ ⊗K q) is
generated by all commutators [Xβ (u), Xα̃−β (v)], where v ∈ Vα̃−β , u ∈ Vβ ⊗K q. This finishes
the proof.
Proof of Proposition 6.3. Let N be a subgroup of G(K) normalized by E(K). If N 6
C(G(K)), there is nothing to prove. Otherwise by our assumption and Theorem 6.2 there
is an ideal q 6= 0 of K such that N ∩ Xα (Vα ) = Xα (qVα ) for all α ∈ ΦP . Then, clearly,
EP (K, q) 6 N. If N is not contained in C(K, q), then ρq (N) is a non-central subgroup of
G(K/q) normalized by E(K/q), and hence by the same token we have N > EP (K/q, q′) for
+
some ideal q′ 6= 0 of K/q. Let α̃ ∈ Φ+
P be a maximal root, and let β ∈ ΦP be a simple
relative root such that α̃ − β ∈ ΦP . Pick 0 6= u ∈ Vα̃−β such that 0 6= ρq (u) ∈ Vα̃−β ⊗K q′ and
Xα̃−β (ρq (u)) ∈ ρq (N). Then there is h ∈ G(K, q) such that Xα̃−β (u)h ∈ N. By Lemma 6.4 for
any generating system e1 , . . . , en of Vβ there is 1 6 i 6 n such that Nα̃−β,β,1,1 (ρq (u), ρq (ei )) 6=
0, and hence Nα̃−β,β,1,1 (u, ei ) 6∈ qVα̃ . By Lemma 6.5 one has
[Xα̃−β (u)h, Xβ (e)] = Xα̃−β (u) [h, Xβ (e)] · Xα̃ (Nα̃−β,β,1,1 (u, ei )) ∈ EP (K, q) · Xα̃(Nα̃−β,β,1,1(u, ei )).
Hence Xα̃ (Nα̃−β,β,1,1 (u, ei )) ∈ N. However, this contradicts the choice of q.
Thus, we have reduced the proof of normal structure theorem to extraction of unipotents
from E(K)-normalized subgroups. The following lemma shows that one can always extract
a unipotent from a parabolic subgroup.
Lemma 6.6. In the setting of Theorem 6.2, let H be a subgroup of G(K) normalized by the
elementary subgroup E(K). Suppose that H ∩ P (K) is not contained in C(G(K)). Then H
contains a root unipotent element Xα (v), α ∈ ΦP , 0 6= v ∈ Vα .
Proof. This is proved exactly as the corresponding statement [VS08, Theorem 1] for Chevalley
groups.
7. Extraction from the main Gauss cell
As usual, assume that G is a reductive group scheme over K, and P is a proper parabolic
K-subgroup of G. Consider the K-subscheme of G
ΩP = UP LP UP − .
This is an open subscheme isomorphic as a K-scheme to the direct product UP × LP × UP − ,
see [DG70, Exp. XXVI, Remarque 4.3.6]. We call ΩP the main Gauss cell associated with
P.
In this section we show how to extract a unipotent from ΩP (K). The following easy lemma
will be used several times.
Lemma 7.1. Let R be a commutative ring, let G be a reductive group scheme over R, let
Ω be an open R-subscheme of G that is finitely presented as an R-scheme. Let g ∈ G(R) be
any element. If for all maximal ideals m of R one has ρm (g) ∈ Ω(R/m), then g ∈ Ω(R).
Proof. First we show that λm (g) ∈ Ω(Rm ) for any fixed maximal ideal m. Since Ω is an
open subscheme of G, it is a union of principal open R-subschemes Gfα = Spec(R[G]fα ) of
G for some fα ∈ R[G]. Since ρm (g) ∈ Ω(R/m), and R/m is a field, there is an index α such
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
12
that ρm (g) ∈ Gfα (R/m). The latter means that the ring homomorphism ρm (g) : R[G] →
R/m satisfies ρm (g)(fα ) ∈ (R/m)× . Since Rm /mRm = R/m and ρm (g) = ρm (λm (g)), we
conclude thatρm (λm (g)) λm (fα ) ∈ (R/m)× . Since Rm is a local ring, this implies that
λm (g) λm (fα ) ∈ (Rm )× . Hence λm (g) ∈ Gfα (Rm ) ⊆ Ω(Rm ), as required.
Since G and Ω are finitely presented R-schemes, Ω(Rm ) is the limit of Ω(Rs ), and G(Rm )
is the limit of G(Rs ), over all s ∈ R \ m. Therefore, for any maximal ideal m of R there
is sm ∈ R \ m such that λsm (g) ∈ Ω(Rsm ). Since all sm together generate the unit ideal of
R, the open subschemes Spec(Rsm ) constitute an open covering of Spec(R) for the Zariski
topology. Consider the system of points λsm (g) ∈ Ω(Rsm ). We would like to show that they
glue in the Zariski topology to a point in Ω(R), and this point is g. Since the functor of
points of Ω is a sheaf for the Zariski topology on Spec(R), it is enough to check that for any
maximal ideals m and n one has λsm (λsn (g)) = λsn (λsm (g)) inside Ω(Rsm sn ). This is true,
since the same equality holds in G(Rsm sn ), and Ω(Rsm sn ) injects into G(Rsm sn ). Hence there
is g ′ ∈ Ω(R) such that λsm (g ′) = λsm (g) for any maximal ideal m of R. It remains to note
that g ′ = g, since Ω(R) injects into G(R).
From now and until the end of this section, assume that K is connected.
Lemma 7.2. Let K be a field and P a parabolic subgroup of G over K. If g ∈ G(K) satisfies
[g, UP (K)] = 1, then g ∈ P (K).
Proof. Let Q 6 P be a minimal parabolic subgroup of G over K, let LQ be its Levi subgroup
contained in LP , and let Q− be the opposite minimal parabolic subgroup contained in P − . Let
S 6 LQ be the maximal split subtorus of C(LQ ). Bruhat decomposition implies that g = uwv,
where u ∈ UQ (K), w ∈ NG (S)(K), v ∈ (UQ )w (K) = {x ∈ UQ (K) | wxw −1 ∈ UQ− (K)}, and
u, v, and the class of w in the Weyl group of ΦQ are unique. We have w ∈ LP (K) if and only
if w is a product of elementary reflections wαi for some simple roots αi ∈ ΦQ belonging to
the root subsystem of ΦQ corresponding to LP [BT65].
Assume first that w ∈
/ LP (K). Then there is a simple root α ∈ ΦQ not belonging to the
root system of LP such that w(α) < 0. Then x = Xα (ξ), ξ ∈ Vα , belongs to UP (K), and since
[g, x] = 1, we have x(uwv) = (uwv)x. The rightmost factor in the Bruhat decomposition of
x(uwv) = (xu)wv equals v. However, since α is a positive root of minimal height, it is clear
that the rightmost factor in the Bruhat decomposition of (uwv)x contains Xα (η + ξ) in its
canonic decomposition, if v contains Xα (η). Therefore, this rightmost factor is distinct from
v, a contradiction.
Therefore, w ∈ LP (K). Then for any x ∈ UP (K) we have wxw −1 ∈ UP (K), hence by the
definition of the Bruhat decomposition v ∈ LP (K) ∩ UQ (K). This means that g = uwv ∈
UQ (K)LP (K) = P (K).
Lemma 7.3. Suppose that the structure constants of the absolute root system of G are
invertible in K. Let Q be a parabolic K-subgroup of G, such that ΦQ is irreducible and
rank(ΦQ ) Q
> 2. Let β ∈ Φ+
Q be a simple relative root. If x ∈ UQ± (K) commutes with Xβ (Vβ ),
then x = Xα (uα ), where each α is a positive (respectively, negative) root of ΦQ satisfying
α+β ∈
/ ΦQ ∪ {0}.
Proof. First, consider the case where x ∈ UQ+ (K). We fix a total ordering on Φ+
Q compatible
k
Q
with the height. Then there is a unique presentation x =
Xαi (ui ), where all roots are
i=1
distinct, each of them lies in Φ+
Q , 0 6= ui ∈ Vαi , and the product is taken in the fixed order.
Let i0 be the smallest index such that αi + β ∈ ΦQ . By Lemma 6.4 there is v ∈ Vβ such
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
13
that Nβ,αi0 ,1,1 (v, ui0 ) = w 6= 0. By the generalized Chevalley commutator formula [Xβ (v), x)]
contains Xαi0 +β (w) in its reduced decomposition, hence it cannot be equal to 1.
k
Q
Second, assume that x =
Xαi (uαi ), where each αi ∈ Φ−
Q \ N · (−β), and 0 6= uαi ∈ Vα .
i=1
We order these roots αi according to the following partial order: first we order these roots
with respect to the sum of coefficients of all simple relative roots distinct from β, with the
largest (negative) sum of coefficients coming first, and then with respect to the coefficient
of β, with the smallest (negative) coefficient coming first. Assume that there is an index
1 6 i 6 k such that αi + β ∈ ΦQ , and let i0 be the smallest such index. By Lemma 6.4 there
is v ∈ Vβ such that Nβ,αi0 ,1,1 (v, uαi0 ) 6= 0. Then
[Xβ (v), x] =
i0Q
−1
i=1
Xαi (uαi )
[Xβ (v), Xαi0 (uαi0 )] ·
i0
Q
i=1
Xαi (uαi )
[Xβ (v),
k
Y
Xαi (uαi )].
i=i0 +1
Generalized Chevalley commutator formula then implies that [Xβ (v), x] contains in its reduced decomposition the non-trivial factor Xβ+αi0 (Nβ,αi0 ,1,1 (v, uαi0 )), and no other factor
corresponding to the root β + αi0 , therefore, [Xβ (v), x] 6= 1, a contradiction.
Finally, let x ∈ UQ− (K) be an arbitrary element. Write x = x1 x2 , where x1 is a product
of negative root elements corresponding to the roots in N(−β), and x2 is a product of other
negative root elements. Then x2 belongs to the unipotent radical U of a parabolic subgroup
+
of G corresponding to the parabolic set Φ−
Q ∪ (Nβ ∩ ΦQ , and Xβ (Vβ ) and x1 are contained in
the Levi subgroup L of this parabolic subgroup corresponding to the set of roots Zβ ∩ ΦQ
(see Lemma 4.1). For any v ∈ Vβ one has
1 = [Xβ (v), x1 x2 ] = [Xβ (v), x1 ]x1 [Xβ (v), x2 ],
where [Xβ (v), x1 ] ∈ L(K) and
x1
[Xβ (v), x2 ] ∈ U(K), hence
1 = [Xβ (v), x1 ] = [Xβ (v), x2 ].
By the previous case we conclude that x2 is a product of root elements corresponding to
α ∈ Φ−
/ ΦQ ; by the definition of x2 , we also have α + β 6= 0, as required.
Q such that α + β ∈
It remains to show that x1 = 1.
k
Q
Assume that x1 6= 1. We can write x1 =
X−iβ (u−iβ ), u−iβ ∈ V−iβ . Let α ∈ Φ+
Q be
i=1
a simple relative root such that α + β ∈ ΦQ . Then α + iβ ∈ ΦQ for any i > 1 such that
iβ ∈ ΦQ by Lemma 4.3. Let i > 1 be the smallest natural number such that u−iβ 6= 0.
Since −iβ − α ∈ ΦQ , by Lemma 6.4 there is w ∈ V−α such that N−iβ,−α,1,1 (u−iβ , w) 6= 0. Set
y = [x1 , X−α (w)]. Then generalized Chevalley commutator formula implies that y ∈ U(K),
and y contains X−iβ−α (N−iβ,−α,1,1 (u−iβ , w)) in its reduced decomposition. On the other hand,
since [Xβ (v), X−α (w)] = 1 for any v ∈ Vβ , we conclude that
[Xβ (v), y] = [Xβ (v), [x1 , X−α (w)]] = 1
for any v ∈ Vβ . By the previous case this implies that y cannot contain X−iβ−α (u) with
non-zero u in its reduced decomposition, a contradiction. This shows that x1 = 1.
Lemma 7.4. Under the hypothesis of Lemma 7.3, let m > 1 be the maximal positive integer
such that mβ ∈ ΦQ . If x ∈ U(β) (K)LQ (K)U(−β) (K) commutes with Xβ (Vβ ), then x ∈
Xmβ (Vmβ )LQ (K).
Proof. Assume that x = ahb for some a ∈ U(β) (K), h ∈ LQ (K), and 1 6= b ∈ U(−β) (K).
m
Q
X−iβ (ui ), where ui ∈ V−iβ , ui0 6= 0. Let α 6= β ∈ ΦQ be a simple relative
Then b =
i=i0
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
14
root adjacent to β. Then [X−α (V−α ), U(β) (K)] = 1 by the generalized Chevalley commutator
formula, and hence [X−α (V−α ), x−1 ], Xβ (Vβ )] = 1. On the other hand, by Lemma 4.3 one
has −i0 β − α ∈ ΦQ , hence by Lemma 6.4 there is v ∈ V−α such that N−i0 β,−α,1,1(ui0 , v) 6= 0.
Then
!−1
m
Y
Y
X−iβ (ui )
[X−α (v), x−1 ] = [X−α (v), h
X−jα−kβ (wj,k ),
]=
i=i0
j>1
k=0∨k>i0
where w1,i0 6= 0. However, this contradicts Lemma 7.3, since (−α − i0 β) + β ∈ ΦQ .
Therefore, b = 1, and x = ah, where a ∈ U(β) (K) and h ∈ LQ (K). Note that by Lemma 5.3
m
Q
Xiβ (ui ),
Xβ (Vβ ) generates U(β) , hence x commutes with all Xiβ (Viβ ), i > 1. Write a =
i=i0
where ui ∈ Viβ , ui0 6= 0. For every j > 1 and v ∈ Vjβ one has
−1
1 = [Xjβ (v), x−1 ] = [Xjβ (v), h−1 ]h [Xjβ (v), a−1 ].
−1
By the generalized Chevalley commutator formula we have h [Xjβ (v), a−1 ] ∈ U{kβ | k>j+1}(K),
and hence [Xjβ (v), h−1 ] ∈ U{kβ | k>j+1}(K). By Lemma 5.2 (ii) this implies that [h, U(β) (K)] =
1. Then [a, Xβ (Vβ )] = 1. By Lemma 7.3 this implies that a ∈ Xmβ (Vmβ ).
Lemma 7.5. Suppose that the absolute root system of G is irreducible and its structure
constants are invertible in K. Let Q be a parabolic K-subgroup of G with a Levi subgroup LQ ,
such that rank(ΦQ ) > 2. For any parabolic subgroup P > Q and any g ∈ ΩP (K) \ C(G(K))
there exists a parabolic subgroup M > LQ of G such that [g E(K) ∩ M(K), E(K)] 6= 1.
Proof. By Theorem 3.2 C(G(K)) = C(G)(K) is contained in every parabolic subgroup, hence
by Lemma 3.4 it is enough to prove the claim under the assumption that G is adjoint.
Write g = ab, where a ∈ UP (K), b ∈ P − (K). Choose any total order on Φ+
Q compatible
E(K)
′
with the height of roots. We show that either g
contains an element g = a′ b′ such that
the minimal root element occuring in a′ is strictly larger than the corresponding element of
a, or there is a parabolic subgroup M as in the claim of the lemma. Since the height of roots
in ΦQ is bounded, the order induction will eventually lead to a′ = 1 and g ′ ∈ P (K).
Assume first that P 6= Q. Let LP > LQ be a Levi subgroup of P . Then there is a simple
relative root β ∈ Φ+
Q that does not belongs to the root system of LP . For any v ∈ Vβ one has
−1
−1
g E(K) ∋ a [Xβ (v), g] = a [Xβ (v), a] · [Xβ (v), b].
−1
Here [Xβ (v), b] ∈ P − (K), while a [Xβ (v), a] ∈ UQ (K) by the generalized Chevalley commutator formula (5.3), and all root elements occuring in the decomposition of the latter
−1
expression are strictly larger than the minimal root element of a. If a [Xβ (v), g] is not
centralized by E(K), then we can apply the induction hypothesis.
−1
−1
Assume that a [Xβ (v), g] is centralized by E(K). Since G is adjoint, we have a [Xβ (v), g] =
−1
1. Then a [Xβ (v), a] ∈ UQ (K) ∩ P − (K) 6 LP (K). However, by the choice of a this implies that [Xβ (v), a] = 1. Then [Xβ (v), b] = 1 as well. Thus, we are in the situation where
[x, a] = [x, b] = 1 for every x ∈ UQ′ (K), where
Q′ = Q ∩ LP
is a proper parabolic subgroup of LP . Write b = hu where h ∈ LP (K), u ∈ UP − (K).
Then 1 = [x, hu] = [x, h] · h [x, u], where [x, h] ∈ LP (K) and h [x, u] ∈ UP − (K), and hence
[x, h] = [x, u] = 1 as well. For any maximal ideal m of K the image of h in LP (K/m)
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
15
commutes with UQ′ (K/m), and hence belongs to Q′ (K/m) by Lemma 7.2. Since ΩQ′ is an
open subscheme of LP , by Lemma 7.1 we have h ∈ ΩQ′ (K). Then
g = ab = ahu ∈ UQ (K) · ΩQ′ (K) · UP − (K) ⊆ ΩQ (K).
Thus, we are reduced to the case where g ∈ ΩQ (K), i.e. to the situation P = Q. Again,
let β ∈ ΦQ be a simple relative root. Let P ′ = P ′+ and P ′− be the “comaximal” parabolic
subgroups of G with the common Levi subgroup LP ′ such that P ′± = UΦ± ∪Zβ and LP ′ = UZβ
Q
in the sense of Lemma 4.1. Clearly,
g ∈ ΩQ (K) ⊆ ΩP ′ (K).
Rerunning, if necessary, the above height induction with respect to the parabolic subgroup
P ′ instead of P , we arrive to a new decomposition g = ahu with a ∈ UQ (K), h ∈ LP ′ (K) =
UZβ (K), u ∈ UP ′− (K), and [a, Xβ (Vβ )] = [h, Xβ (Vβ )] = [u, Xβ (Vβ )] = 1. By Lemmas 7.4
and 7.3 this implies that g is a product of root elements of the form Xα (uα ) where all α
satisfy α + β ∈
/ ΦQ , and of an element of LQ (K). Then by Lemma 4.4 there is a proper
parabolic set of relative roots Σ ⊂ ΦQ such that g ∈ UΣ (K) in the notation of Lemma 4.1.
Then we take M = UΣ .
The following statements follow immediately from Lemmas 7.5 and 6.6.
Corollary 7.6. Under assumptions of Lemma 7.5 the subgroup g E(K) contains a nontrivial
root unipotent element.
Corollary 7.7. Suppose that G is semisimple, K is connected, the absolute root system Φ
of G is irreducible, its structure constants are invertible in K, and G contains (Gm,K )2 . If
H ∩ G(K, Rad K) 6⊆ C(G(K)), then H contains a nontrivial root unipotent element.
Proof. Since G contains (Gm,K )2 , there is a parabolic subgroup Q of G such that rank(ΦQ ) >
2. Take g ∈ (H ∩ G(K, Rad K)) \ C(G(K)). Then ρm (g) ∈ ΩQ (K/m) for any maximal ideal
m of K. Then by Lemma 7.1 g ∈ ΩQ (K), and the result follows from Corollary 7.6.
8. Generic element techniques
Let G be an affine smooth finitely presented group scheme over K. Denote by A = K[G]
the affine algebra of the scheme G. By the definition of an affine scheme, an element h ∈ G(R)
can be identified with a homomorphism h : A → R. We always do this identification, i. e.
we always view elements of the group of points G(R) of the scheme G over a K-algebra R
as homomorphisms from A to R.1 Denote by g ∈ G(A) the generic element of the scheme
G, i. e. the identity map idA : A → A. An element h ∈ G(R) induces the homomorphism
G(h) : G(A) → G(R) by the rule G(h)(a) = h ◦ a for all a ∈ G(A). It follows that the image
of g under the action of G(h) is equal to h.
Recall that for a ring homomorphism ϕ : R → R′ we usually denote the induced group
homomorphism G(ϕ) : G(R) → G(R′ ) again by ϕ. This cannot lead to a confusion as one
always can determine the meaning of ϕ by the argument type of this homomorphism. In
view of this agreement we have h(g) = h ◦ idA = h. If R′ is an R-algebra, then sometimes we
identify elements of G(R) with their canonical images in G(R′ ).
Recall that the fundamental ideal I of A is the kernel of the counit map eK : A → K,
where eK ∈ G(K) is the identity element of this abstract group. The notation A, I, and g
introduced above is kept till the end of the present section.
The following characterization of the principal congruence subgroup was observed in [Ste16].
1Thus,
functions.
for f ∈ A we write h(f ) instead of f (h), as one writes considering the affine algebra as a set of
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
16
Lemma 8.1. An element h of G(R) belongs to the principal congruence subgroup G(R, q) if
and only if h(I) ⊆ q.
Proof. Denote by ρq : G(R) → G(R/q) the natural homomorphism. It is easy to see that
h ∈ G(R, q) iff the following diagram commutes.
h
A −−−→ R
ρq
eK y
y
K −−−→ R/q
And the latter is obviously equivalent to saying that h(I) vanishes modulo q, i. e. h(I) ⊆
q.
Since A is a finitely presented K-algebra, it is a quotient of a polynomial ring in finitely
many variables by a finitely generated ideal. Let K ′ be a Z-subalgebra of K, generated by
all the coeffients of polynomials that generate this ideal. Then A ∼
= A′ ⊗K ′ K, were A′ is
′
a finitely generated algebra over a Noetherian ring K . Thus, there exists an affine smooth
e over a Noetherian ring K ′ such that G = G
eK . In the
finitely presented group scheme G
present paper we prove results about the abstract group G(R) for a K-algebra R. Therefore,
we may often assume without loss of generality that K = K ′ , and hence A is a Noetherian
ring.
An advantage of the Noetherian property of a ring in this context is the following property [Bak91, Lemma 4.10].
Lemma 8.2. Let R be a Noetherian ring and s ∈ R. There exists m ∈ N such that the
restriction of the localization homomorphism λs : R → Rs to the ideal sm A is injective.
The next lemma is a version of clearing denominators.
Lemma 8.3. Let G be an affine smooth finitely presented group scheme over ring K, R a
Noetherian K-algebra, m ∈ N, and s ∈ R. Suppose that the natural map sm A →
As is
injective. Given a ∈ G(Rs ) there exists k ∈ N such that [a, λs (b)] ∈ λs G(R, sm R) for all
b ∈ G(R, sk R).
Proof. Recall that we assume A to be finitely presented. Therefore, there exists a finite set
J ⊆ A that generates A as a K-algebra. Since A = K ⊕ I as a K-module, we may assume
that J ⊆ I. Obviously, under this condition J spans I as an ideal.
Identifying elements of G(A) and G(Rs ) with their canonical images in G(A ⊗K Rs ),
consider the commutator c = [a, λs (g)] ∈ G(A ⊗K Rs , I ⊗K Rs ). By Lemma 8.1 the finite set
c(J) is contained in I ⊗K Rs . Clearly, there exists l ∈ N such that each element of this set can
be written as r/sl for some r ∈ I ⊗K R. Take k = l + m. Take b ∈ G(R, sk R) and consider
the composition d = mult ◦ (b ⊗ id) ◦ c = [a, λs (b)], where mult : R ⊗K Rs → Rs denotes the
multiplication homomorphism. By Lemma 8.1 b(I) ⊆ sk R, hence the set d(J) consists of the
elements of the form mult ◦ (b ⊗ id)(r/sl ) = sk t/sl = sm t for some t ∈ R. Since the restriction
of the localization homomorphism λs to sm R is injective and J generates the K-algebra A,
the homomorphism d factors through d′ : A → R, which means that d = λs (d′ ). Moreover,
the natural choice of d′ provides that d′ (J) ⊆ sm R. Since J generates the fundamental ideal
I, we have d′ (I) ⊆ sm R and by Lemma 8.1 d′ ∈ G(R, sm R).
The next lemma is a key step to the proof of the normal structure theorem.
Lemma 8.4. Let G be a reductive group scheme over a Noetherian connected ring K, and
let Q be a parabolic K-subgroup of G such that ΦQ is irreducible and rank(ΦQ ) > 2. For any
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
17
parabolic subgroup Q P there exist elements cij ∈ g E(A) ∩ ΩP (A), i = 1, . . . , l, j = 1, . . . , n,
satisfying the following property. Let S be a subscheme defined by the formula
S(R) = {h ∈ G(R) | h(cij ) ∈ C(G)(R) for all i = 1, . . . , l and j = 1, . . . , n}
for any K-algebra R. Then the group of points S(F ) does not contain E(F ) for any K-algebra
F that is a field.
Proof. Since the subscheme ΩP is open, hence it is covered by principal open subschemes
SpK Asi , i = 1, . . . , l. Let α, β ∈ Φ−
Q be two relative roots such that −α, −β are simple roots,
α + β ∈ ΦQ , Xα ⊆ LP , Xβ ⊆ UP − . Let {e1 , . . . , en } be a set of generators of the K-module
Vα . Choose i = 1, . . . , l and j = 1, . . . , n and put s = si and v = ej .
Let gs be the image of g in G(As ). In other words, gs = λs is the localization homomorphism. It is a tautology that gs factors through As , which means that gs ∈ SpK As (As ) ⊆
ΩP (As ). Decompose it into a product gs = ab, where a ∈ UP (As ) and b ∈ P − (As ). Since A
is Noetherian, by Lemma 8.2 there exists m ∈ N such that the restriction of λs to sm A is
injective. By Lemma 8.3 there exists a positive integer k > m such that
−1
ds = Xα (sk v)a ∈ λs UP (A, sm A) and fs = dgss = Xα (sk v)b ∈ λs P − (A, sm A) .
m
−1
−
m
Let d = λ−1
s (ds ) ∈ UP (A, s A) and f = λs (fs ) ∈ P (A, s A) (by definition of m these
−1
m
preimages are unique). Put cij = c = [g, d ] ∈ G(A, s A). Then
gs −1
λs (c) = [gs , d−1
= fs d−1
s ] = ds d
s ∈ λs ΩP (A) .
By definition of m we have cij = f d−1 ∈ ΩP (A).
l
S
Let F be a field. Put h = Xβ (u) ∈ ΩP (F ) =
SpK Asi (F ) for some u ∈ Vβ ⊗K F r {0}.
i=1
Choose i such that h ∈ SpK Asi and put s = si . Then h factors through As , i. e. h = e
h ◦ λs
e
e
e
e
for some h : As → F . Since h(g) = h, we have h = h(gs ) = h(a)h(b) = e · Xβ (u). The
uniqueness of representation of h as a product of an element from UP (F ) by an element from
P − (F ) implies that h(a) = e. Thus, we get
h(cij ) = [h(g), h(d−1)] = [Xβ (u), Xα(sk ej )−1 ].
By Lemma 6.4 there exists j such that Nαβ11 (ej , u) 6= 0. By the generalized Chevalley
commutator formula h(cij ) ∈
/ C(G)(F ). Thus, h ∈
/ S(F ), and hence S(F ) does not contain
E(F ), as required.
9. Proof of the normal structure theorem
In this section we prove the main theorem of the present paper.
Proof of Theorem 1.1. Let T 6 G be the subgroup isomorphic to (Gm,K )2 . By Lemma 4.1
there are two parabolic subgroups Q = Q+ and Q− of G with a common Levi subgroup
LQ = CG (T ).
Assume first that K is Noetherian. Then K is a finite product of connected Noetherian
rings, so we can assume without loss of generality that K is Noetherian and connected. The
relative root system ΦQ of Lemma 4.2 is irreducible and rank(ΦQ ) > 2. By Lemma 6.5 we
have EQ (K, q) = [G(K, q), E(K)]. By Lemma 6.1 we have EP (K, q) = EQ (K, q) for any
other parabolic subgroup P . Thus, by Proposition 6.3 it suffices to prove that if H is not
inside C(G(K)), then it contains a nontrivial relative root unipotent element.
Let β ∈ ΦQ be a simple root. Then by Lemma 4.1 P = P + = UΦ+ ∪Zβ is a parabolic
Q
subgroup distinct from Q, with a Levi subgroup LP = UZβ . Let S be a subscheme of G
satisfying conditions of Lemma 8.4. If there exists h ∈ H such that h(cij ) ∈
/ C(G(K)), then
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
18
h(cij ) ∈ hE(K) ∩ ΩP (K) is subject to Corollary 7.6, and hence H contains a nontrivial relative
root unipotent element.
Otherwise H is contained in the set of K-points of the subscheme S. Consider a maximal ideal m of K. The image H of the subgroup H under the canonical homomorphism
G(K) → G(K/m) is contained in S(K/m) and is normalized by E(K/m). Tits’ simplicity theorem [Tit64] then implies that either H contains E(K/m), or H is contained in C(G(K/m)).
Since E(K/m) is not contained in S(K/m), H is contained in C(G(K/m)). It follows that
the image of the subgroup [H, E(K)] vanishes in G(K/m), i. e. [H, E(K)] 6 G(K, m).
Since m is an arbitrary maximal ideal, we have [H, E(K)] ⊆ G(K, J), where J is the
Jacobson radical of K. On the other hand, by Lemma 3.4 [H, E(K)] is not contained in the
center of G(K). Hence, by Corollary 7.7 this subgroup contains a nontrivial relative root
unipotent element. This completes the proof of the Noetherian case.
Now let K be arbitrary. For any finite set of elements Λ ⊆ H there is a finitely generated
e of K and a semisimple reductive group scheme G
e over K
e with a subgroup Te ∼
subring K
=
2
e
e
e
e
(Gm,Ke ) such that G = GK , T = TK , and Λ ⊆ G(K). Clearly, by Lemma 4.1 there is also a
e6G
e such that Q = Q
eK .
parabolic subgroup Q
e there is a maximal ideal m of K such that m
e of K
e ⊆ m,
Since for any maximal ideal m
e
e
e also has irreducible root system whose structure constants are
we conclude that G(K/m)
e
e
e K)
e is subject to the Noetherian case of the theorem.
invertible in K. Thus ΛE(K) 6 G(
e such that
Hence there exists an ideal q(Λ) ⊆ K
e
e q(Λ)) 6 ΛE(K)
e q(Λ)).
E(K,
6 C(K,
e and Λ. Let q be the ideal of K generated by all
Clearly, q(Λ) is uniquely determined by K
e
subsets q(Λ) ⊆ K ⊆ K. Then, clearly, H 6 C(K, q). W show that EQ (K, q) 6 H. In order
to do that, it is enough to check that UQ± ((a)) is contained in H for any finite K-linear
P
e be subring of K corresponding to the
combination a = ci ai of elements ai ∈ q(Λi ). Let K
e and all ci . Then by the
finite set Λ = ∪Λi , and let K ′ be the subring of K generated by K
eK ′ , we conclude that U e± ((a)) = UQ± ((a)) are
Noetherian case of the theorem applied to G
Q
′
E(K )
E(K)
contained in Λ
6Λ
6 H.
The same argument as above also shows that EQ (K, q) = [G(K, q), E(K)] and EP (K, q) =
EQ (K, q) for any other parabolic subgroup P of G, since this equality holds for each finitely
e such that P is defined over K
e and the ideal e
e ∩ q.
generated subring K
q=K
References
[Abe89]
[ABS90]
[AS76]
[Bak69]
[Bak91]
[Bas64]
[Bas73]
[Bix80]
[BLS64]
E. Abe, Normal subgroups of Chevalley groups over commutative rings, Contemp. Math. 83 (1989),
1–17.
H. Azad, M. Barry, and G. Seitz, On the structure of parabolic subgroups, Comm. Algebra 18
(1990), no. 2, 551–562.
E. Abe and K. Suzuki, On normal subgroups of Chevalley groups over commutative rings, Tohoku
Math. J. 28 (1976), no. 2, 185–198.
A. Bak, On modules with quadratic forms, Lecture Notes in Math. 108 (1969), 55–66.
, Nonabelian K-theory: The nilpotent class of K1 and general stability, K-Theory 4 (1991),
363–397.
H. Bass, K-theory and stable algebra, Publ. Math. Inst. Hautes Études Sci 22 (1964), 5–60.
, Unitary algebraic K-theory, Lecture Notes in Math. 343 (1973), 57–265.
Robert Bix, Octonion planes over local rings, Trans. Amer. Math. Soc. 261 (1980), no. 2, 417–438.
MR 580896
H. Bass, M. Lazard, and J.-P. Serre, Sous-groupes d’indice fini dans SL(n, Z), Bull. Amer. Math.
Soc. 70 (1964), 385–392. MR 0161913
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
19
[BMJP67] H. Bass, J. Milnor, and Serre J.-P., Solution of the congruence subgroup problem for SLn (n > 3)
and Sp2n (n > 2), Publ. Math. Inst. Hautes Études Sci 33 (1967), 59–137.
[Bor91]
Armand Borel, Linear algebraic groups, second ed., Graduate Texts in Mathematics, vol. 126,
Springer-Verlag, New York, 1991. MR 1102012
[Bre38]
Joel Brenner, The linear homogeneous group, Ann. of Math. (2) 39 (1938), no. 2, 472–493.
MR 1503419
, The linear homogeneous group. II, Ann. of Math. (2) 45 (1944), 100–109. MR 0009956
[Bre44]
[Bre60]
J. L. Brenner, The linear homogeneous group. III, Ann. of Math. (2) 71 (1960), 210–223.
MR 0110754
[BT65]
A. Borel and J. Tits, Groupes réductifs, Publ. Math. Inst. Hautes Études Sci 27 (1965), 55–150.
[BV85]
Z. I. Borevich and N. A. Vavilov, The distribution of subgroups in the general linear group over a
commutative ring, Proc. Steklov. Inst. Math. 165 (1985), 27–46.
[Che55]
C. Chevalley, Sur certains groupes simples, Tohoku Math. J. 7 (1955), 14–66.
[CK91a] D. L. Costa and G. E. Keller, The E(2,A) sections of SL(2,A), Ann. of Math. (2) 134 (1991),
no. 1, 159–188.
[CK91b]
, Radix redux: normal subgroups of symplectic groups, J. Reine Angew. Math. 427 (1991),
no. 1, 51–105.
, On the normal subgroups of G2 (A), Trans. Amer. Math. Soc. 351 (1999), no. 12, 5051–
[CK99]
5088.
[DG70]
M. Demazure and A. Grothendieck, Schémas en groupes, Lecture Notes in Math. 151–153,
Springer-Verlag, Berlin-Heidelberg-New York, 1970.
[Die48]
J. Dieudonné, Sur les groupes classiques, Hermann, Paris, 1948.
[Gol73]
I. Z. Golubchik, On the general linear group over an associative ring, Uspekhi Math Nauk 28
(1973), no. 3, 179–180 (Russian).
[Kli61a] Wilhelm Klingenberg, Lineare Gruppen über lokalen Ringen, Amer. J. Math. 83 (1961), 137–153.
MR 0124412
, Orthogonale Gruppen über lokalen Ringen, Amer. J. Math. 83 (1961), 281–320.
[Kli61b]
MR 0124414
, Symplectic groups over local rings, Amer. J. Math. 85 (1963), 232–240. MR 0153749
[Kli63]
[KS13]
E. Kulikova and A. Stavrova, centralizer of the elementary subgroup of an isotropic reductive
group, Vestnik St. Petersburg Univ. Math. 46 (2013), no. 1, 22–28.
[LS12]
A. Luzgarev and A. Stavrova, Elementary subgroup of an isotropic reductive group is perfect,
St.Petersburg Math. J. 23 (2012), no. 5, 881–890.
[Mar79] G. A. Margulis, Finiteness of quotient groups of discrete subgroups, Funktsional. Anal. i Prilozhen.
13 (1979), no. 3, 28–39. MR 545365
[Mat69] H. Matsumoto, Sur les sous-groupes arithmétiques des groupes semisimples déployés, Ann. Sci.
Éc. Norm. Supér. (4) 2 (1969), no. 1, 1–62.
[Mil12]
J.
S.
Milne,
Basic
theory
of
affine
group
schemes,
Preprint
http://www.jmilne.org/math/CourseNotes/AGS.pdf, 2012.
[Pla69]
V. P. Platonov, The problem of strong approximation and the Kneser-Tits hypothesis for algebraic
groups, Izv. Akad. Nauk SSSR Ser. Mat. 33 (1969), 1211–1219. MR 0258839
[Pre18]
Raimund Preusser, Sandwich classification for GLn (R), O2n (R) and U2n (R, Λ) revisited, J. Group
Theory 21 (2018), no. 1, 21–44. MR 3739342
[PS09]
V. Petrov and A. Stavrova, Elementary subgroups in isotropic reductive groups, St.Petersburg
Math. J. 20 (2009), no. 4, 625–644.
[Rag76]
M. S. Raghunathan, On the congruence subgroup problem, Inst. Hautes Études Sci. Publ. Math.
(1976), no. 46, 107–161. MR 0507030
[Rag86]
MS Raghunathan, On the congruence subgroup problem, ii, Inventiones mathematicae 85 (1986),
no. 1, 73–117.
[Sta14a] A. Stavrova, On the congruence kernel of isotropic groups over rings, 2014, Preprint:
http://ArXiv.org/abs/1305.0057.
[Sta14b] A. K. Stavrova, Homotopy invariance of non-stable K1 -functors, J. K-Theory 13 (2014), 199–248.
[Sta16]
A. Stavrova, Non-stable K1 -functors of multiloop groups, Canad. J. Math. 68 (2016), 150–178.
[Ste16]
A. V. Stepanov, Structure of Chevalley groups over rings via universal localization, J. Algebra
450 (2016), 522–548.
NORMAL STRUCTURE OF ISOTROPIC REDUCTIVE GROUPS OVER RINGS
[Sus77]
20
A. A. Suslin, On the structure of the special linear group over polynomial rings, Math. USSR.
Izvestija 11 (1977), 221–238.
[Tit64]
J. Tits, Algebraic and abstract simple groups, Ann. of Math. 80 (1964), 313–329.
[Vas81]
L. N. Vaserstein, On the normal subgroups of GLn over a ring, Lecture Notes in Math. 854 (1981),
456–465.
, On normal subgroups of Chevalley groups over commutative rings, Tohoku Math. J. 38
[Vas86]
(1986), 219–230.
, Normal subgroups of orthogonal groups over commutative rings, Amer. J. Math. 110
[Vas88]
(1988), no. 5, 955–973. MR 961501
[VS08]
N. A. Vavilov and A. K. Stavrova, Basic reductions in the description of normal subgroups, J.
Math. Sci. (N. Y.) 151 (2008), no. 3, 2949–2960.
[Wil72]
J. S. Wilson, The normal and subnormal structure of general linear groups, Math. Proc. Cambridge
Philos. Soc. 71 (1972), 163–177.
[Wil09]
Robert A. Wilson, The finite simple groups, Graduate Texts in Mathematics 28, Springer-Verlag,
Berlin, New York, 2009.
[You12]
You Hong, Subgroups of classical groups normalized by relative elementary groups, J. Pure and
Appl. Algebra 216 (2012), 1040–1051.
[Zha10]
Zhang, Zuhong, Subnormal structure of non-stable unitary groups over rings, J. Pure and Appl.
Algebra 214 (2010), no. 5, 622–628.
E-mail address: [email protected]
E-mail address: [email protected]
St. Petersburg State University
| 4math.GR
|
1
On State Observers for Nonlinear Systems: A New
Design and a Unifying Framework
arXiv:1712.08209v1 [cs.SY] 21 Dec 2017
Bowen Yi, Romeo Ortega, and Weidong Zhang
Abstract—In this paper we propose a new observer design
technique for nonlinear systems. It combines the well-known
Kazantzis-Kravaris-Luenberger observer and the recently introduced parameter estimation-based observer, which become
special cases of it—extending the realm of applicability of both
methods. A second contribution of the paper is the proof that
these designs can be recast as particular cases of immersion and
invariance observers—providing in this way a unified framework
for their analysis and design. Simulation results of a physical
system that illustrates the superior performance of the proposed
observer compared to other existing observers are presented.
Index Terms—State observers, nonlinear systems, parameter
estimation.
I. I NTRODUCTION AND P ROBLEM F ORMULATION
In this paper we are interested in the design of state
observers for nonlinear control systems whose dynamics is
described by1
ẋ = f (x, u)
(1)
y = h(x),
where x ∈ Rn is the system state, u ∈ Rm is the control
signal, and y ∈ Rp are the measurable output signals. The
problem is to design a dynamical system
χ̇
= F (y, χ, u)
x̂
= H(y, χ, u)
(2)
with χ ∈ Rnχ , such that
lim |x̂(t) − x(t)| = 0,
t→∞
(3)
where | · | is the Euclidean norm. Following standard practice
in observer theory we assume that u is such that the state
trajectories of (1) are bounded.
Since the publication of the seminal paper [14], which dealt
with linear time-invariant (LTI) systems, this problem has been
extensively studied in the control literature. We refer the reader
to [2], [5] for a review of the literature. In this paper we
are particularly interested in three recently developed observer
design techniques.
This paper is partially supported by the National Natural Science Foundation of China (61473183, U1509211), the Government of Russian Fedration
(074U01, GOSZADANIE 2014/190 (project 2118)), the Ministry of Education
and Science of Russian Federation (14.Z50.31.0031). Corresponding author:
Weidong Zhang.
Bowen Yi and Weidong Zhang are with Department of Automation, Shanghai Jiao Tong University, Shanghai 200240, China. [email protected],
[email protected]
Romeo Ortega is with LSS, CNRS-CentraleSupélec, Plateau du Moulon,
Gif-sur-Yvette 91192, France. [email protected]
1 All mappings in the paper are assumed smooth.
The Kazantzis-Kravaris-Luenberger observer (KKLO)
first presented in [12] as an extension to nonlinear systems of Luenberger’s observer and further developed in
[1].
• Parameter estimation-based observer (PEBO) proposed in
[15], which translates the state observation problem into
an on-line parameter estimation problem.
• Immersion and invariance observer (I&IO), first reported
in [11] and thoroughly elaborated in [2], which proposes a
more general observer framework based on the generation
of attractive and invariant manifolds.
The main contributions of our paper are threefold.
(C1) Propose a new observer design technique, called
[KKL+PEB]O, that combines—in a seamless way—the
KKLO and PEBO designs, yielding a new design applicable to a broader class of systems.
(C2) Prove that KKLO, PEBO and [KKL+PEB]O can be recast
as particular cases of I&IO—providing in this way a
unified framework for their analysis and design.
(C3) Present simulation results of the well-known DC-DC Ćuk
converter that illustrate the superior performance of the
proposed observer compared to other existing observer
designs.
The remainder of the paper is organized as follows. Section
II gives some preliminaries on KKLO and PEBO. In Section
III we present the new [KKL+PEB]O. The unifying framework
based on immersion and invariance is given in Section IV. In
Section V we present two academic examples that illustrate
the interest of the new [KKL+PEB]O and some simulation
results of a physical system that compares the new observer
with other observer designs. The paper is wrapped–up with
concluding remarks and future research directions in Section
VI.
•
II. P RELIMINARIES
In this section we briefly present simple versions of the
KKLO and the PEBO that are motivating to generate the new
[KKL+PEB]O in the next section.
A. KKL Observer
The KKLO design is based on the following proposition,
which is a simplified version of the more general result
reported in [1], [12].
Proposition 1. Consider the system (1) satisfying the following assumption.
2
A1 There exist nξ negative real numbers λi , i = 1, 2, . . . , nξ ,
with nξ ≥ n, and mappings
φ : Rn → Rn ξ
L
φ :R
nξ
p
p
×R →R
Proposition 2. Consider the system (1) satisfying Assumption
A1 of Proposition 1 with Λ = 0 and the dynamic extension
(8).
A2 There exist mappings
n
B : R × Rm → Rnξ
with nξ ≥ n − p, satisfying the following.
(i) The KKLO partial differential equation (PDE)
∇φ> (x)f (x, u) = Λφ(x) + B(h(x), u),
M : Rn ζ × R n ξ × Rp × Rm → Rn ζ
φ (φ(x), h(x)) = x.
such that the dynamical system
ζ̇ = M (ζ, ξ, y, u)
(5)
The KKLO
ξ˙ =
x̂
θ̂ = N (ζ, ξ, y, u),
(12)
defines a stable, consistent parameter estimator for the
regression model (11), that is ζ is bounded and
Λξ + B(y, u)
= φL (ξ, y),
N : R n ζ × Rn ξ × R p × R m → Rn ξ
(4)
∂ >
where ∇ := ( ∂x
) and Λ := diag{λi }.
(ii) φL is a left inverse of φ, that is,
L
where we underscore that y, u, ξ are, of course, measurable.
The main result in [15] may be summarized as follows.
lim |θ̂(t) − θ| = 0.
(6)
ensures (3).
t→∞
(13)
The PEBO (8), (10), (12) verifies (3).
Proof. The proof of this proposition follows immediately
defining the error signal
e := ξ − φ(x)
and noting that ė = Λe.
C. Remarks
R1 Notice that the KKLO (6), together with the dynamics of
the system (1), admits an attractive and invariant manifold
S := {(x, ξ) ∈ Rn × Rnξ | ξ = φ(x)},
B. PEB Observer
The PEBO design proposed in [15], although related with
the KKLO, aims at formulating the state reconstruction problem as a parameter estimation problem. Towards this end, we
are looking for an injection B(h(x), u) and a (left invertible)
mapping φ(x) that transforms the system (1) into2
and the state is (asymptotically) reconstructed, via φL , with
the knowledge of ξ. On the other hand, the PEBO generates
an invariant foliation
φ̇(x) = B(h(x), u).
To reconstruct the state—again via φL —it is necessary to
identify the leaf of the foliation where the system evolves.
These observations are essential to establish the connection
of these observers with the I&IO, which also relies on the
generation of an attractive and invariant manifold, defined by
an invertible mapping.
(7)
In this way, selecting (part of) the observer dynamics as
ξ˙ = B(y, u),
(8)
we establish, via simple integration, the key relationship
φ(x(t)) = ξ(t) + θ,
(9)
where θ is a constant vector defined as θ := φ(x(0)) − ξ(0).
It is clear that, if θ is known, we have that
x = φL (ξ + θ, y).
Hence, the remaining task is to generate an estimate for θ,
denoted θ̂, to obtain the observed state
x̂ = φL (ξ + θ̂, y).
(10)
To achieve this end, we rely on the existence of the regression
model
h(φL (ξ + θ, y))
y
=
,
(11)
ẏ
∇h> (φL (ξ + θ, y))f (φL (ξ + θ, y), u)
2 To avoid cluttering the notation, whenever clear from context, we use
the same symbols to denote mappings playing similar roles in the various
observers. The subindex (·)P or (·)L is later used to identify the PEBO or
KKLO-related mappings in the [KKL+PEB]O.
Sθ := {(x, ξ) ∈ Rn × Rnξ | ξ = φ(x) + θ, θ ∈ Rnξ }.
R2 Besides the additional difficulty of needing to estimate θ,
the main drawback of PEBO is that it relies on the open-loop
integration (8), which might be a problematic operation in
practice. In spite of that, PEBO has proven instrumental in
the solution of numerous physical systems problems [6], [7],
[8], [17]—some of them being unsolvable with other observer
design techniques—with the open integration problem solved
via the addition of “safety nets” similar to the ones used in
PID designs or adaptive control.
R3 We underscore the fact that the PDE that needs to be solved
for PEBO is exactly the one of KKLO with Λ = 0, that is
∇φ> (x)f (x, u) = B(h(x), u).
We refer the reader to [18] where the generation of virtual
outputs via signal injection technique of [9] is proposed to
simplify the solution of this PDE.
3
III. N EW [KKL+PEB] O BSERVERS
In this section we present our first main contribution,
namely, a new observer design technique that combines PEBO
and KKLO. The key idea of the new observer is to split the
states to be estimated in two groups, the first one estimated
with a KKLO and the second one with a PEBO.
A. Main result
The following proposition, whose proof follows verbatim
from Propositions 1 and 2 formalises the discussion above.
For ease of presentation, and without loss of generality, we
assume that the aforementioned groups are arranged one on
top of the other.
Proposition 3. Consider the system (1) satisfying Assumption
A1 of Proposition 1 with
ΛL
0q×(nξ −q)
Λ
=
0(nξ −q)×q 0(nξ −q)×(nξ −q)
(14)
ΛL : =diag{λ1 , . . . , λq },
where 0 ≤ q ≤ nξ , 0k×j is a k × j matrix of zeros, λi <
0, i = 1, . . . , q. Partition the mapping B(y, u) as follows
BL (y, u)
B(y, u) =
, BL ∈ Rq , BP ∈ Rnξ −q .
BP (y, u)
The [KKL+PEB]O
ξ˙L = ΛL ξL + BL (y, u)
ξ˙P = BP (y, u)
ζ̇ = MP (ζ, ξP , y, u)
(15)
θ̂ = NP (ζ, ξP , y, u)
0q
ξL
L
x̂ = φ
+
,y ,
ξP
θ̂
where 0q is a q-dimensional vector of zeros, ensures (3)
provided the mappings
MP : R
NP : R
nζ
nζ
×R
×R
nξ −q
nξ −q
p
m
p
m
×R ×R
×R ×R
nζ
→R ,
→ Rnξ −q ,
case, the dynamics of (ξL , ξP ) in the observer (15) is replaced
by
ξ˙L
ξL
= Λ
+ P B(y, u)
ξP
ξ˙P
0q
ξL
L
>
x̂ = φ (P
+
, y),
ξP
θ̂
where P is a unitary matrix satisfying A = P > ΛP , with Λ
upper triangular.
R6 In the case of input-affine systems, i.e., f (x, u) = F (x) +
g(x)u, it is possible to decompose the PDE (4) into two, that
is,
∇φ> (x)F (x)
R4 It is clear that Proposition 3 contains, as particular cases,
Propositions 1 and 2. Indeed, the former is recovered setting
q = nξ while the latter follows with q = 0.
B(y, u) := BF (y) + Bg (y)u.
Explicit formulas for the solutions of these equations may be
found in [18].
IV. I&I O BSERVERS : A N U NIFYING F RAMEWORK
In this section we show that a mild extension of the I&IO
studied in [11], [2] allows us to capture, as a particular case the
new [KKL+PEB]O proposed in this paper—and, consequently,
it also contains the KKLO and the PEBO.
A. Extension of I&I observers
The main result of the I&IO in [11] is extended in the
following proposition by relaxing a dimension requirement
imposed to some mappings in the original formulation of
I&IO—see R8 in Subsection IV-C.
Proposition 4. Consider the system (1). Assume the existence
of mappings
β : R p × Rn χ → Rn z
φ : R n → Rn z
φ L : R n z × Rp → Rn
with φL (φ(x), y) = x and nχ ≥ nz , such that:
A4 rank ∇χ β > (y, χ) = nz .
A5 The system with state
d˙M =∇y β > ∇h> (x)f (x, u) − ∇h> (x̂)f (x̂, u)
− ∇φ> (x)f (x, u) + ∇φ> (x̂)f (x̂, u)
R5 The result of Proposition 3 can be extended in several
directions. For instance, it is possible to replace the PDE (4)
by
∇φ> (x)f (x, u) = Aφ(x) + B(h(x), u),
where A has an eigenvalue of zero of arbitrary multiplicity,
and all the other eigenvalues are in the open left hand plane.3
Clearly, the degree of freedom provided by the inclusion of
the matrix A enlarges the set of solutions of the PDE. In this
3 With
a slight modification it is also possible to consider the case of A
with purely imaginary eigenvalues.
= Bg (h(x))
and define the observer dynamics via
define a consistent estimator, that is, (13) holds.
B. Remarks
= Aφ(x) + BF (h(x))
∇φ> (x)g(x)
where
x̂
=φL (φ(x) + dM , h(x))
dM
:=β(y, χ) − φ(x),
(16)
(17)
has an asymptotically stable equilibrium dM = 0.
The I&IO
χ̇
x̂
= −[∇χ β > ]† (∇y β > ∇h> (x̂) − ∇φ> (x̂))f (x̂, u) (18)
+(I − [∇χ β > ]† ∇χ β > )Q(y, χ, u)
= φL (β(y, χ), y).
(19)
4
β(y, χ) = ξ + N (ζ, ξP , h(x))
−ΛL φL (x) + ΛL ξL
0nξ −q
Υ(x, χ, u) =
∇ζ β > (−MP (ζ, φP (x), h(x), u) + MP (ζ, φP (x̂), h(x), u))
Q(y, χ, u) = 0.
with [·]† the pseudoinverse and Q : Rp ×Rnχ ×Rm → Rnχ ×nχ
an arbitrary mapping, verifies (3).
Proof. The dynamics of off-the-manifold coordinate dM is
d˙M = ∇y β > ∇h> (x)f (x, u) + ∇χ β > χ̇ − ∇φ> (x)f (x, u).
Replacing the dynamics of χ̇ in (18), we get (16). According
to Assumption A5, we have
lim dM (t) = 0.
= h(col(x̂1 , x2 )) = h(x).
Replacing this limit in (17) and recalling that φ (φ(x), y) = x
ensures limt→∞ |x̂(t) − x(t)| = 0.
d
NP (ζ, φP (x), y) = 0nξ −q ,
dt
thus yielding the following identity.
B. KKL+PEB observers: An I&I interpretation
In this section we will show that if the system admits a
[KKL+PEB]O it also admits an I&IO. To unify the notation
we define
ξL
ξL
ξP
χ := ξP , ξ :=
, υ :=
,
ξP
ζ
ζ
φ(x)
M (ζ, ξP , y, u)
N (ζ, ξP , y, u)
φL (x)
,
φP (x)
0q
:=
,
MP (ζ, ξP , y, u)
0q
:=
,
NP (ζ, ξP , y, u)
:=
(23)
Proof. For the sake of clarity, we assume NP is independent
of u to avoid further complicating the notation. Before the
proof, we present the following two useful facts.
F1 If the output signals are partial states, i.e., x :=
col(x1 , x2 ) and y = x2 without loss of generality, we
have
h(x̂) = h(φL (ξ + N (ζ, ξP , y), y))
L
(22)
F2 When ξP = φP (x), we have
t→∞
introduce the partitions
(21)
>
∇ξP NP> ∇φ>
P (x)f (x, u) + ∇ζ NP MP (ζ, φP (x), y, u)
+ ∇y NP> ∇h> (x)f (x, u) = 0.
(24)
According to (21), equivalently
0q
β(y, χ) = ξ +
,
NP (ζ, ξP , h(x))
Assumption A4 is obviously satisfied. The reminding of the
proof is divided into two parts: 1) the selected mappings yield
the dynamics (18)-(19), having the same structure as (15) in
[KKL+PEB]O; 2) these mappings are solutions of I&IO PDE.
1) The dynamics of χ̇ in (18) has the term
[∇χ β > ]† (∇y β > ∇h> − ∇φ> )f (x̂, u), in which
∇y β > ∇h> (x̂)f (x̂, u) − ∇φ> (x̂)f (x̂, u)
0q − ∇φ>
(x̂)
L
=
f (x̂, u)
∇y NP> ∇h> (x̂) − ∇φ>
P (x̂)
and define the mapping
Υ(x, χ, u) :=∇y β > ∇h> (x)f (x, u) − ∇h> (x̂)f (x̂, u)
− ∇φ> (x)f (x, u) + ∇φ> (x̂)f (x̂, u),
(20)
that, according to (16), defines the dynamics of the off-themanifold coordinate dM .
Proposition 5. Assume the system (1) admits a [KKL+PEB]O
(15) with
rank ([Inξ + ∇ξ N > | ∇ζ N > ]) = nξ ,
and {(x, ξ, ζ)|θ − θ̂ = 0} is invariant. Then, the system admits
an I&IO (18)-(19) with the mappings selected as (21), (22)
and (23). If the measurable output signals are partial states,
the [KKL+PEB]O (15) exactly coincides with the I&IO (18)(19), and
I&IO PDE (20), (22) ⇒ [KKL+PEB]O PDE (4).
(25)
We analyze the above equation in two parts, i.e., −∇φ>
L f and
(∇y NP> ∇h> − ∇φ>
)f
.
For
the
first
part,
the
existence
of a
P
[KKL+PEB] observer yields the PDE
ΛL φL (x) + BL (y, u) = ∇φ>
L (x)f (x, u).
Substituting x by x̂ = φL (ξL ), we have
∇φ>
L (x̂)f (x̂, u) = ΛL ξL + BL (y, u).
(26)
The second partition of (25) verifies the relation below.
(24)
(∇y NP> ∇h> (x̂) − ∇φ>
P (x̂))f (x̂, u)
= − (∇ξP NP> (ζ, φP (x̂), h(x̂)) + I)∇φ>
P (x̂)f (x̂, u)
− ∇ζ NP> (ζ, φP (x̂), h(x̂))MP (ζ, φP (x̂), h(x̂), u)
BP (h(x̂), u)
= − ∇υ β >
.
MP (ζ, ξP , h(x̂), u)
(27)
5
BP (h(x), u)
BP (h(x̂), u)
>
− ∇υ β (h(x), χ)
+ ∇υ β (h(x̂), χ)
MP (ζ, φP (x), h(x), u)
MP (ζ, φP (x̂), h(x̂), u)
>
∇φP (x)f (x, u)
∇φ>
(x̂)f (x̂, u)
>
>
P
= − ∇υ β (h(x), χ)
+ ∇υ β (h(x̂), χ)
MP (ζ, φP (x), h(x), u)
MP (ζ, φP (x̂), h(x̂), u)
>
(29)
⇔∇φ>
P (x)f (x, u) = BP (h(x), u).
Combining (25)-(27), we get the mapping α(·) in I&IO is
ΛL ξL + BL (y, u)
χ̇ = α(y, χ, u) = BP (h(x̂), u) ,
MP (ζ, ξP , h(x̂), u)
showing that if the I&IO PDE has a solution, then the I&IO
asymptotically coincides the [KKL+PEB]O. Furthermore, if
the measurable output signals are partial states, due to fact
F1, the I&IO exactly coincides with the [KKL+PEB]O.
2) We check the solution existence of I&IO PDE. The first
partition of I&IO PDE is verified as follows.
ΛL φL (x) + BL (y, u) = ∇φ>
L (x)f (x, u).
⇒ − ΛL φL (x) − BL (y, u) + ΛL ξL + BL (y, u)
=
−∇φ>
L (x)f (x, u)
⇔ΛL ξL − ΛL φL (x) =
+ ∇φ>
L (x̂)f (x̂, u)
>
∇φL (x̂)f (x̂, u) − ∇φ>
L (x)f (x, u)
(28)
For the second partition of the I&IO PDE, the identity (24)
yields (29). Combining (28)-(29), it shows that the selecting
mappings are solutions of I&IO PDE.
The invariant manifold of [KKL+PEB]O is
n
nχ
ξ + N (ζ, ξP , y) = φ(x) .
M = (x, χ) ∈ R × R
˙
The dynamics of off-the-manifold coordinate is dM
=
Υ(x, χ, u), whose convergence is guaranteed by the consistent
identification Assumption A2 and the fact that the matrix ΛL
is Hurwitz. This completes the proof.
We are in position to present the main result of this paper—
the unified observer framework captured by I&IO.
Corollary 1. For the nonlinear system (1), a [KKL+PEB]O
implies the existence of I&I observers. Moreover, the following “set” relationship holds:
)
PEBO
⊂ [KKL+PEB]O ⊂ I&IO.
(30)
KKLO
C. Remarks
R7 As discussed in R1, an I&IO generates the invariant
manifold
n
nχ
M = (x, χ) ∈ R × R
β(h(x), χ) = φ(x) ,
which is made attractive ensuring—via Assumption A5—that
the zero equilibrium of the dynamics of the off-the-manifold
coordinate (16), which may be written as
d˙M = Υ(x, χ, u)
=: Υ0 (x, dM , u)
has an asymptotically stable equilibrium at the origin.
R8 In the I&IO proposed in [11], [2] we fix nχ = nz ≤ n. In
this case, (18) reduces to
χ̇ = −[∇χ β −1 ]> (∇y β > ∇h> (x̂) − ∇φ> (x̂))f (x̂, u),
and A4 is equivalent to requiring that ∇χ β is a non-singular
square matrix. For PEBO and [KKL+PEB]O, the dimension
of invariant manifold are less than the dimensions of
dynamic extensions. Hence, we generalise I&IO removing
the requirement nξ = nz and using the pseudoinverse.
R9 KKLOs and PEBOs are specific cases of [KKL+PEB]Os,
making them covered by I&IO framework. In detail,
• the KKLO (6) coincides with the I&IO (18)-(19) with
mappings selecting
nχ = nz = q
β(y, χ) = ξ = χ
Υ(x, χ, u) = −Λφ(x) + Λχ,
•
with any mapping Q(y, χ, u). The KKLO PDE sacrifices
the freedom for Υ by fixing Υ = −Λφ(x) + Λχ.
If the measurable output signals are partial states, the
PEBO coincides with the I&IO with mappings selecting4
χ = col(ξ, ζ)
β(y, χ) = ξ + N (ζ, ξ, y)
Υ(x, χ, u) = −∇χ β > ×
0
M (ζ, φ(x), h(x), u) − M (ζ, φ(x̂), h(x̂), u)
Q(y, χ, u) = 0,
where x̂ = φL (β(y, χ), y).
R10 Notice that the condition that “the measurable output
signals are partial states” in Proposition 5 is done without
loss of generality because it is always possible to do a change
of coordinates to verify it.
V. E XAMPLES
In this section we present three examples that illustrate our
developments.
• An academic example for which neither KKLO nor
PEBO are applicable, but it is solvable via our new
[KKL+PEB]O design.
4 Here
we also assume N is independent of u for the sake of clarity.
6
•
•
Identify a class of nonlinear systems, which contain
the aforementioned example, for which [KKL+PEB]O
provides a simple constructive solution.
Consider the widely studied DC-DC Ćuk converter for
which a PEBO and an I&IO were reported in [15] and [2],
respectively. We also design a KKLO, a [KKL+PEB]O
and two high gain observers à la [10]. The purpose of
this example is to compare the performance of all these
observers from the point of view of gain tuning flexibility
and robustness with respect to measurement noise, which
is unavoidable in this application.
Applying this condition to the (1, 3) element of the Jacobian
we conclude that ρ1 (x) should satisfy
ρ1 (x) := ρ03 (x1 )x3 + L(x1 , x2 ),
with L(x1 , x2 ) to be determined and (·)0 the derivative with
respect to its argument. From the (2, 3) element we also
conclude that ρ2 (x) is independent of x3 , that is ρ2 (x) :=
ρ2 (x1 , x2 ).
The terms in the left-hand side of (39) dependent on x3 are
−ρ03 (x1 )x3 x31 + ρ03 (x1 )x3 ex3 + L(x1 , x2 )ex3 , while the one
on the right-hand side is −λ2 ρ3 (x1 )x3 . Thus we conclude
that L(x1 , x2 ) = 0 and
A. Proving the interest of the new observer
−λ2 ρ3 (x1 )x3 = −ρ03 (x1 )x3 x31 + ρ03 (x1 )x3 ex3 .
Proposition 6. Consider the system
Hence
ẋ1 = −x31 + ex3
−λ2 ρ3 (x1 ) = ρ03 (x1 )(−x31 + ex3 ),
x21
ẋ2 = −x2 + + sin x1
1
+ x1 u
ẋ3 = 2
x1 + 1
y = x1 ,
(31)
F3 The system does not admit a KKLO nor a PEBO.
F4 The system admits a [KKL+PEB]O, namely,
ξ˙1 = −ξ1 + y 2 + sin y
ξ˙2 = uy + (y 2 + 1)−1
˙
Θ̂ = γψ(Y − ψ Θ̂)
(32)
x̂2 = ξ1
(35)
x̂3 = ξ2 + ln Θ̂.
(36)
(33)
(34)
where γ > 0 is an adaptation gain and Y, ψ are obtained
via LTI filtering as
α 3
αp
y +
y
Y =
p+α
p+α
α ξ2
ψ =
e , ψ(0) > 0
(37)
p+α
with p :=
d
dt
and α > 0, is a [KKL+PEB]O that ensures
lim |x̂i (t) − xi (t)| = 0, i = 2, 3.
t→∞
Proof. [Proof of F3] KKLO requires φ(x) to be injective. To
guarantee this property at least one of its three components
should depend on x3 . Suppose, without loss of generality, that
φ2 (x) depends on x3 . Define the three-dimensional vector ρ
as
ρ(x) := ∇φ2 (x).
(38)
From the PDE (4) we have
−x31 + ex3
ρ> (x) −x2 + x21 + sin x1 = −λ2 φ2 (x)+B2 (x1 , u) (39)
1
x1 u + 1+x
2
1
Since φ2 (x) dependends of x3 , we have ρ3 (x) 6= 0. The
left hand side term of (39) dependent on u is ρ3 (x)x1 u,
while the one in the right hand side is B2 (x1 , u), from
which we conclude that ρ3 (x) only depends on x1 , that is
ρ3 (x) = ρ3 (x1 ). From Poincare’s lemma we have that (38)
holds if and only if the Jacobian ∇ρ(x) is a symmetric matrix.
whose only solution is ρ3 = 0, which contradicts with the fact
that ρ3 6= 0, due to λ2 6= 0 in the KKLO. Therefore, it shows
that the system (31) does not admit a KKLO.
Since PEBO also requires injectivity of φ(x), some similar
argument proves that the system does not admit a PEBO.
Indeed, let us assume that φ2 (x) depends on x2 and ρ2 (x) 6= 0.
As it follows the argument above, for the PEBO with λ2 = 0
we have ρ03 (x1 ) = 0, yielding ρ3 (x1 ) = c and ρ1 (x) = 0
with a constant c. From the (1,2) and (2,1) elements of the
Jacobian matrix ∇ρ(x) we conclude that ∇x1 ρ2 = 0 and
ρ2 (x1 , x2 ) := ρ2 (x2 ). Then in terms of (39), we have
−x31 + ex3
0 ρ2 (x2 ) c −x2 + x21 + sin x1 = B2 (x1 , u)
1
x1 u + 1+x
2
1
and
−ρ2 (x2 )x2 + ρ2 (x2 )x21 + ρ2 (x2 ) sin x1 = B2 (x1 , u).
Since the first term in the left hand side does not depend on
x1 , it concludes that ρ(x2 ) = 0 leading to a contradiction.
Thus the given system does not have a PEBO.
[Proof of F4] The [KKL+PEB]O PDE (4) with (14), q = 1,
λ1 = −1 and
2
x + sin x1
φ(x) = col(x2 , x3 ), B(h(x), u) = x 1u + 1
1
takes the form
∇φ> (x)f (x, u) =
−1
0
x21 +1
φ(x) + B(h(x), u).
Thus the observer dynamics is given by (32) and (33). From
which we conclude that
lim |ξ1 (t) − x2 (t)| = 0, x3 (t) = ξ2 (t) + θ,
t→∞
with the constant θ to be estimated. Noticing the following
relationship
ẋ1 = −x31 + ex3 ,
we can formulate a linear regression model for the estimation
of θ of the form
Y = ψΘ + t
7
where Θ := exp(θ), Y and ψ are defined in (37) and t
is a (generic) exponentially decaying term that, without loss
of generality, we neglect in the sequel.5 Finally, the choice
of initial condition for ψ ensures that ψ(t) is not square
˙ = −γψ 2 Θ̃ with Θ̃ := Θ̂ − Θ ensures
integrable, thus Θ̃
limt→∞ Θ̂(t) = Θ and consequently guarantees (3).
To verify the above analysis, we did numerical simulation
in Matlab/Simulink. The simulation scenario was as follows:
the control input u is selected as
u = −y + sin(2πt).
The initial states in the plant and observer are selected as
x1 (0) = −1, x2 (0) = 1, x3 (0) = 1, ξ1 (0) = 0, ξ2 (0) =
0, Θ̂(0) = 2 and the initial conditions of the filters set to
zero. The gains are γ = 1 and α = 1. The simulation results
are shown in Fig. 1.
Proposition 7. Consider systems of the form
ẋ1 = f1 (x1 , x2 , x3 , u) + S(x, u)
ẋ2 = A2 x2 + f2 (x1 , u)
ẋ3 = A3 x3 + f3 (x1 , x2 , u)
(40)
ẋ4 = f4 (x1 , x2 , x3 , u)
y = x1 ,
where x := col(x1 , x2 , x3 , x4 ), with xi ∈ Rni , i = 1, . . . , 4,
and u ∈ Rm , verifying the following.
(i) There exists 1 ≤ k ≤ n1 such that the corresponding
element of the vector S satisfies
Sk (x, u) = b> (x1 , x2 , x3 , u)x4
(41)
for some mapping b : Rn1 × Rn2 × Rn3 × Rm → Rn4 .
(ii) The matrices A2 and A3 are Hurwitz.
(iii) The control input guarantees that the trajectories of (40)
are bounded and the following persistency of excitation
condition is verified
Z t+T
b(x1 (s), x2 (s), x3 (s), u(s))×
(42)
t
>
b (x1 (s), x2 (s), x3 (s), u(s))ds ≥ δIn4 ,
for all t ≥ 0 and some δ, T > 0.
The system admits a [KKL+PEB]O of the form
x̂˙ 2 = A2 x̂2 + f2 (y, u)
x̂˙ 3 = A3 x̂3 + f3 (y, x̂2 , u)
ξ˙ = f4 (y, x̂2 , x̂3 , u)
(43)
x̂4 = ξ + θ̂,
(46)
(44)
(45)
with parameter estimator
˙
θ̂ = Γψ̂(Ŷ − ψ̂ > θ̂)
(47)
where
Ŷ
ψ̂
α
αp
yk −
f1,k (y, x̂2 , x̂3 , u)
p+α
p+α
α >
−
b (y, x̂2 , x̂3 , u)ξ
p+α
α
:=
b(y, x̂2 , x̂3 , u) ,
p+α
:=
with f1,k the k-th element of the vector f1 .
Fig. 1. Simulation results of the example of Proposition 6.
B. A class of nonlinear systems for [KKL+PEB]O
We identify below a class of systems whose states can be
reconstructed with the proposed [KKL+PEB]O. The examples
of Subsections V-A and V-C belong to this class.
Proof. We first prove boundedness of x̂2 and x̂3 . Due to the
assumption (iii), we have that f2 (y, u) ∈ L∞ . Hence, from
(43) and the fact that A2 is Hurwitz, we have that x̂2 ∈ L∞ .
Proceeding in the same way with (44) we conclude that x̂3 ∈
L∞ .
Now, we prove that the observation errors x̃i (t) := x̂i (t) −
xi (t) (i = 1, . . . , 3) converge to zero exponentially fast. It is
obvious that
x̃˙ 2 = A2 x̃2 ,
and limt→∞ x̃2 = 0 (exp.). Similarly,
5 See
Remark 3 in [3] where the effect of these term is rigorously analyzed.
x̃˙ 3 = A3 x̃3 + f3 (y, x̂2 , u) − f3 (y, x2 , u).
8
Consider the function V (x̃3 ) := 21 x̃>
3 P x̃3 , with P > 0 the
solution of P A3 + A>
3 P = −`1 I < 0. Its time derivative
satisfies
V̇
= −`1 |x̃3 |2 + 2x̃>
3 P [f3 (y, x̂2 , u) − f3 (y, x2 , u)]
≤
2
−`1 |x̃3 | + `2 |x̃3 ||x̃2 |,
(48)
where we have used the fact that the boundedness of all the
arguments of f3 (·, ·, ·) ensures
|f3 (y, x̂2 , u) − f3 (y, x2 , u)| ≤ `3 |x̃2 |
for some `3 > 0. From (48), the comparison lemma and the
fact that x̃3 ∈ L∞ and x̃2 (t) → 0 (exp.), we conclude that
x̃3 (t) → 0 (exp.).
It only remains to prove that x̃4 := x̂4 − x4 also converges
to zero. Consider (45) and define ξ˜ := ξ − x4 , which satisfies
˙
ξ̃ = f4 (y, x̂2 , x̂3 , u) − f4 (y, x2 , x3 , u).
Hence,
˜
ξ(t)
=
Z
t
[f4 (y(s), x̂2 (s), x̂3 (s), u(s))
0
≤
≤
Z
0
t
˜
−f4 (y(s), x2 (s), x3 (s), u(s))]ds + ξ(0)
Ŷ = ψ̂ > θ
Replacing the equation above in (47) we get the parameter
estimation error equation
˙
θ̃ = Γψ̂ ψ̂ > θ̃,
where θ̃ := θ̂ − θ. The proof of (exponential) convergence of
θ̃(t) to zero is completed invoking standard adaptive control
arguments.
For the sake of clarity we have presented Proposition 7
in a very simple form, being possible to extend it in several
directions.
• Clearly, the number of subsystems of the form ẋi =
Ai xi + fi (y, x1 , · · · , u) can be larger than the two taken
here.
• Invoking the recent results of identification and adaptive
control of nonlinearly parameterised systems—see
[13] and references therein—it is possible to replace
Assumption (i) by:
|[f4 (y(s), x̂2 (s), x̂3 (s), u(s))
(i’) There exists 1 ≤ k ≤ n1 such that the corresponding
element of the vector S satisfies
˜
−f4 (y(s), x2 (s), x3 (s), u(s))]| ds + ξ(0)
Z t
x̃2 (s)
˜
`4
ds + ξ(0),
x̃
3 (s)
0
for some `4 > 0, where we have used the same argument
invoked above to get the second bound. Because of the
exponential convergence to zero of its arguments, the integral
above converges to a constant as t → ∞, consequently, we
can write
x4 (t) = ξ(t) + θ + t ,
(49)
for some constant vector θ—equation (49) corresponds to the
key relationship (9) of PEBO with φ(x) = x4 .
To complete the proof we show now that, under the persistent excitation condition (42), the proposed estimator is
consistent, that is, limt→∞ θ̂(t) = θ that, together with (46)
and (49) establishes the claim that x̃4 (t) → 0. Towards this
end, notice that replacing (41) in the k-th equation of ẋ1 we
get
ẋ1,k
the estimation errors x̃2 (t) and x̃3 (t) converge exponentially
fast to zero, we have that Ŷ (t) = Y (t) +t and ψ̂(t) = ψ +t .
Therefore, neglecting the terms t , we get
Sk (x, u) = b> (x1 , x2 , x3 , u)Φ(x4 ),
•
•
for some monotonic mapping Φ : Rn4 → Rn4 .
Regarding Assumption (i) it is also possible to consider
the existence, not just of one element of S, but several
of them verifying the factorizability condition. This will
give rise to a matrix regressor b for which the persistent
excitation condition (42) would be easier to satisfy.
For simplicity the unknown parameter θ is identified in
Proposition 7 with the classical gradient estimator (47).
However, it is possible to replace this estimator with
the high-performance dynamic regressor extension and
mixing proposed in [3], see also [16]. As shown in these
papers parameter convergence is ensured without the,
often restrictive, persistent excitation condition (42).
= f1,k (x1 , x2 , x3 , u) + b> (x1 , x2 , x3 , u)x4
C. DC-DC Ćuk converter
= f1,k (x1 , x2 , x3 , u) + b> (x1 , x2 , x3 , u)(ξ + θ),
In this section we consider the well-studied DC-DC Ćuk
converter depicted in Fig. 2 and compare, via simulations, the
following observers: PEBO, KKLO, [KKL+PEB]O, I&IO and
the high-gain observer (HGO) of [10].
The averaged model of the system is given as
where we have use (49) to get the second equation. On the
α
we get
other hand, ẏk = ẋ1,k , hence applying the filter p+α
>
the (ideal) regression form Y = ψ θ with
αp
α
Y :=
yk −
f1,k (y, x2 , x3 , u)
p+α
p+α
α >
−
b (y, x2 , x3 , u)ξ
p+α
α
ψ :=
b(y, x2 , x3 , u) ,
p+α
that is, of course, unmeasurable because of the dependence of
f1,k and b on the unknown states. However, due to the fact that
E
1
(1 − u)y1 +
L1
L1
1
G
ẋ2 =
y2 −
x2
C4
C4
1
1
(1 − u)x1 +
uy2
ẏ1 =
C2
C2
1
1
ẏ2 = − uy1 −
x2 ,
L3
L3
ẋ1 = −
(50)
popular boost converter. For the sake of ease of exposition, instead of developing
otationally cumbersome—general theory for a broader class of power converters, we
rred to concentrate on the specific example of the Ćuk power converter, depicted in
2.
L1
E
C2
+ −
i1
L3
v2
i3
+
C4
−
+
−
v4
G
9
is selected as in [15]. The observer parameters were taken as
α = 0.5, γ = 0.001, Γ = diag(0.001, 100), γ1 = 50, γ2 =
1, r1 = 0.05, r2 = 0.005, α1 = α3 = 2,α2 = α4 = 1, to make
the observers have approximate convergence speeds. All the
initial values of dynamic extensions in observers are selected
as 0. The simulation results are given in Fig. 3.
The following remarks are in order.
KKLO and I&IO have 2-D dynamics, clearly, the lowest
order ones. KLLO has the simplest observer structure.
Figure
2: DC–DC
Fig. 2. DC-DC
Ćuk converter
circuitĆuk converter circuit
The parameters in PEBO were the easiest to tune with
guaranteed convergence speed; KKLO and [KKL+PEB]O
he average model of this device is given by
need to resolve PDEs to tune. Besides, for HGO the
where x := col(i1 , v4 ), y := col(v2 , i3 ), and L1 , C1 , L3 , C4 , E
di1
achievable convergence speed is severely limited.
and G are positive
constants.
u
∈
(0,
1)
is
a
duty
cycle.
L1 dt = −(1 − u)v2 + E
• The I&I framework allows to treat in a unified manner
As indicated in the
figure,
i
and
i
are
the
currents
in
the
C2 v̇2 = (11− u)i1 3+ ui3
(41) of state and parameter estimation, see [2]
the problems
3
inductances, whileLv32diand
v4 denote the voltages across the
dt = −uv2 − v4
for the state observation with unknown parameters.
capacitors. More details
converter may be found
C4 v̇4 about
= i3 −the
GvĆuk
4,
• The KKLO has the best performance in the presence
in [2] and references therein. We are interested in estimating x
measurement
noise, probably due to the fact that its
e L1 , C2 , L3 ,with
C4 , E
and G are positive
constants
and udesigns
∈ (0, 1)proposed
is a duty cycle.ofWe
refer
y measurable.
Following
the observer
dynamic extension is a linear system that attenuates the
note and
the ones
reported
eader to [4] in
forthis
further
details
on the
model.in the literature, we obtain
α
effect of the noise. On the other hand, the dynamics in
Table I, in
F (p)
= different
o illustrate the
the observers
generalitygiven
of theinapproach
wewhich
consider
two
p+α andmeasurement sceαp
PEBO, [KKL+PEB]O and I&IO are nonlinear, and seem
W
(p)
=
.
Notice
that
for
the
KKLO,
Λ
is
a
time-varying
s. In the first one we
p+αassume that (v2 , v4 ) are measurable, while in the second one
to have a deleterious impact on the noise.
stable
matrix,
since
1
−
u
∈
/
L
.
3 ) are measurable. Although from the 2practical viewpoint it is “easier” to measure
• The first HGO yields a time-varying error dynamics,
Simulations
were
conduct
with
measurement
noises,
which
ges, we also consider the second one since, as shown in [4], is the one that can be
which is stable because of the physical constraint 1−u >
are generated
by Matlab/Simulink’s
uniform
number
d with immersion
and invariance
(I&I) observers,
withrandom
which we
compare our observer
0. It has oscillations in the transient stage. The second
block
with
sampling
time
of
0.0001s,
and
the
magnitude
mulations below.
HGO has LTI dynamics, where high gain injections are
limitations are [-0.02,0.02] for y1 and [−2 × 10−4 , 2 × 10−4 ]
used to estimate the output derivatives, i.e., ẏ1 and ẏ2 . The
for
y
.
The
parameters
of
the
converter
are
L
=
10
mH,
e I Denoting y :=2 col(v2 , v4 ), x := col(i1 , i3 ) we get from1(41)
operating modes of the converter switch at the moments
C2 =22.0 µF, C4 =22.9 µF, G=0.0447 S and E=12 v. In order
t = 0.2k s (k = 1, . . . , 5), yielding relatively large
1
to give a fair comparison study,
the system Eruns with the ideal
ẋ1 = − (1 − u)y1 +
derivatives of the outputs around these moments.
state-feedback with the stabilizing
controlLlaw
L1
1 given in [2]
•
As expected, the worst performance was systematically
|Vd | ẋ2 = − 1 G|V
|v2 1+yE(x
d−
2 − x1 )
uy
.
1
2
observed for the HGOs because of the high-gain injection
u=
+λ L
,
|Vd | + E
1 +3 (G|Vd |vL23+ E(x2 − x1 ))2
needed to ensure its stability. It is worth pointing out that
the right hand
of the
these
is output
independent
we can
whereside
Vd is
set equations
point for the
voltage of
v4 ,xwhich
wasdirectly select
this (well-known) deleterious effect of high-gain injection
•
φ(x, y) = x.
TABLE I
dynamic extension is given by
State observers for Cúk converters
Type
Mappings
− L11 (1Observer
− u)y1structure
+ LE1
χ̇ =
=: h(y, u),
1
1
− L3 uy1 − L3 y2 −1
−1
Λ = diag(−C −1 G, −L−1 (1 − u))
(6), x̂ = col(L1 ξ2 + L1 C2 y1 , ξ1 )
KKLO
he regression form is
PEBO [15]
ẏ
=
˙
(8),
θ̂ = y,
ΓM
θ̂)
Φ0 (χ,
u)> (Y
+ Φ−1M
(u)θ
−1
x̂ =
+ ξ + col(0, C4
θ := x(0)
−θ̂χ(0)
[KKL+PEB]O
GL3 y2 )
˙
(15), θ̂ = γM (Y − M θ̂)
HGO
(time-varying dynamics)
HGO
(linear dynamics)
1
−1
−1
−1
B = col(L−1
1 (E − (1 − u)y1 ), C4 (y2 + Guy1 )), M = diag(C2 F [1 − u], −L3 )
−1
W [y1 ] − C2 F [ξ1 (1 − u) + uy2 ]
Y=
, α, Γ ∈ R+
−1
−1
W [y2 ] + F [L(42)
3 (uy1 + ξ2 ) + C4 (GL3 y2 )]
−1
Λ = diag(0, −C4−1 G), B = col(L−1
1 E − (1 − u)y1 , C4 y2 ), P = I
13
Y = W [y1 ] − C2−1 F [(1 − u)ξ2 + uy2 ], M = C2−1 F [1 − u], α, γ ∈ R+
x̂ = col(ξ1 + θ̂, ξ2 )
I&IO [2]
4
B = col(C4−1 y2 , (1 + L−1
1 C2 )(−1 + u)y1 + E − uy2 )
ξ̇1 = −γ1 (1 − u)(ξ1 + C2 γ1 y1 ) + γ1 uy2 + L−1
C2 γ1 y1
1 (E − (1 − u)y1 )
, x̂ = ξ +
, γ1 , γ2 ∈ R+
−1
L
γ
y
3 2 2
ξ̇2 = C4 (y2 − G(ξ2 − L3 γ2 y2 )) − γ2 (uy1 + ξ2 − L3 γ2 y2 )
−1
C2 (1 − u)ξ2 + C2−1 (uy2 ) + α1 r1−1 (y1 − ξ1 )
−L−1 (1 − u)y1 + L−1 E + α2 r−2 (y1 − ξ1 )
ξ2
1
1
1
ξ̇ =
,
x̂
=
, r1 ∈ (0, 1], αi > 0
−1
−1
ξ4
−L−1
3 ξ4 − L3 uy1 + α3 r1 (y2 − ξ3 )
−1
−1
−2
C4 y2 − C4 Gξ4 + α4 r1 (y2 − ξ3 )
ξ2 + α1 r1−1 (y1 − ξ1 )
C ξ −uy
−1
−2
−1
2 2
2
−L (1 − u)y1 + L E + α2 r (y1 − ξ1 )
1
1
, x̂ =
1−u
ξ̇ =
, r2 ∈ (0, 1], αi > 0
−1
ξ4 + α3 r (y2 − ξ3 )
−L3 ξ4 − uy1
1
C4−1 y2 − C4−1 Gξ4 + α4 r1−2 (y2 − ξ3 )
10
Fig. 3. The system state variables x1 and x2 with the full-state feedback controller, and observation errors x̃1 = x̂1 − x1 and x̃2 = x̂2 − x2 .
was also observed for mechanical systems in [4].
•
VI. C ONCLUDING R EMARKS
A new observer design technique, called [KKL+PEB]O,
which consists of the combination of KKLO and PEBO was
introduced—providing more degrees of freedom for the solution of the key PDE. Via the suitable selection of the tuning
matrix Λ of the form (14), in the PDE (4), [KKL+PEB]O
reduces to PEBO or KKLO. An example that is not solvable
with KKLO nor PEBO, but it is via [KKL+PEB]O show that
the new observer design extends the applicability of PEBO and
KKLO. Also we identified a class of nonlinear systems, for
which [KKL+PEB]O provides a simple constructive solution.
An additional contribution is the proof that, a slight generalisation of the I&IO, allows us to obtain [KKL+PEB]O,
as well as PEBO and KKLO, as particular cases of I&IO.
This provides a unified framework, based on immersion and
invariance, to treat the three observer designs and establish the
“set” relationship (30).
Further research is underway in the following directions.
• Exploit the constructive approach to find the free mappings in [KKL+PEB]O for some more specific classes of
physical systems.
•
Generalize the coordinate change from φ(x) to φ(x, u)
in order to simplify the solution of the PDEs in these
observers. Along this line of research, one interesting
possibility is extending the theoretical observer existence
results in [1] to control systems with input.
Study [KKL+PEB]O-based output feedback control. In
particular, we are currently investigating if, due to the
presence of the PEBO part, the resulting controller enjoys a “self-tuning” property similar to model reference
adaptive control. That is, if the control objective can be
achieved without requiring that the parameter estimation
error θ̃ converges to zero. Such a property would obviate
the need of excitation conditions for [KKL+PEB]O-based
(or PEBO-based) output feedback control.
ACKNOWLEDGEMENT
The authors would like to thank Stanislav Aranovskiy,
Hassan K. Khalil and Laurent Praly for some suggestions
and comments in Section V, also thank Pauline Bernard and
Vincent Andrieu for clarifications on detectability in KKL
observer design. They are also grateful to the Editor, Associate
Editor and anonymous reviewers for their highly thorough
revisions.
11
R EFERENCES
[1] V. Andrieu and L. Praly, On the existence of a Kazantzis–
Kravaris/Luenberger observer, SIAM J. on Control and Optim., vol. 45,
pp. 432-456, 2006.
[2] A. Astolfi, D. Karagiannis and R. Ortega, Nonlinear and Adaptive
Control with Applications, Springer-Verlag, Berlin, Communications and
Control Engineering, 2008.
[3] S. Aranovskiy, A. Bobtsov, R. Ortega and A. Pyrkin, Performance
enhancement of parameter estimators via dynamic regressor extension
and mixing, IEEE Trans. Automatic Control, vol. 62, No. 7, pp. 35463550, 2017. (See also arXiv preprint arXiv:1509.02763.)
[4] S. Aranovskiy, R. Ortega, J. Romero and D. Sokolov, A globally
exponentially stable speed observer for a class of mechanical systems:
Simulation comparison with high-gain and sliding mode designs, Int. J.
of Control, online, 2017.
[5] G. Besançon (Ed.), Nonlinear Observers and Applications, Lecture
Notes in Control and Information Science, vol. 363, Springer-Verlag,
2007.
[6] A. Bobtsov, A. Pyrkin, R. Ortega, S. Vukosavic, A. M. Stankovic and
E. V. Panteley, A robust globally convergent position observer for the
permanent magnet synchronous motor, Automatica, vol. 61, pp. 47-54,
2015.
[7] A. Bobtsov, A. Pyrkin, R. Ortega and A. Vedyakov, State observers for
sensorless control of magnetic levitation systems, Automatica, (to be
published). (See also arXiv preprint arXiv:1711.02733.)
[8] J. Choi, K. Nam, A. Bobtsov, A. Pyrkin and R. Ortega, Robust adaptive
sensorless control for permanent magnet synchronous motors, IEEE
Transactions on Power Electronics, vol. 32, pp. 3989-3997, 2017.
[9] P. Combes, A. K. Jebai, F. Malrait, P. Martin and P. Rouchon, Adding
virtual measurements by signal injection, 2016 American Control Conference, Boston, USA, July 6-8, pp. 999-1005, 2016.
[10] F. Esfandiari and H. K. Khalil, Output feedback stabilization of fully
linearizable systems, Int. J. of Control, vol. 56, pp. 1007-1037, 1992.
[11] D. Karagiannis, D. Carnevale and A. Astolfi, Invariant manifold based
reduced-order observer design for nonlinear systems, IEEE Trans. on
Automatic Control, vol. 53, No. 11, pp. 2602-2614, 2008.
[12] N. Kazantzis and C. Kravaris, Nonlinear observer design using Lyapunov’s auxiliary theorem, Systems & Control Letters, vol. 34, pp. 241247, 1998.
[13] X. Liu, R. Ortega, H. Su and J. Chu, On adaptive control of nonlinearly
parameterized nonlinear systems: Towards a constructive procedure,
Systems and Control Letters, vol. 10, pp. 36-43, 2011.
[14] D. G. Luenberger, Observers for multivariable systems, IEEE Trans. on
Automatic Control, vol. 11, pp. 190-197, 1966.
[15] R. Ortega, A. Bobtsov, A. Pyrkin and S. Aranovskiy, A parameter
estimation approach to state observation of nonlinear systems, Systems
& Control Letters, vol. 85, pp. 84-94, 2015.
[16] R. Ortega, L. Praly, S. Aranovskiy, B. Yi and W. Zhang, On dynamic
regressor extension and mixing parameter estimators: Two Luenberger
observers interpretations, Automatica, (to appear).
[17] A. Pyrkin, F. Mancilla, R. Ortega, A. Bobtsov and S. Aranovskiy,
Identification of photovoltaic arrays’ maximum power extraction point
via dynamic regressor extension and mixing, Int. J. on Adaptive Control
and Signal Processing,, vol. 39, no. 9, pp. 1337-1349, 2017
[18] B. Yi, R. Ortega and W. Zhang, Relaxing the conditions for parameter
estimation-based observers of nonlinear systems via signal injection,
Systems & Control Letters, vol. 111, pp. 16-28, 2018.
| 3cs.SY
|
1
Securing UAV Communications via Joint
Trajectory and Power Control
Guangchi Zhang, Member, IEEE, Qingqing Wu, Member, IEEE, Miao Cui,
arXiv:1801.06682v1 [cs.IT] 20 Jan 2018
Rui Zhang, Fellow, IEEE
Abstract
Unmanned aerial vehicle (UAV) communication is anticipated to be widely applied in the forthcoming fifth-generation (5G) wireless networks, due to its many advantages such as low cost, high mobility,
and on-demand deployment. However, the broadcast and line-of-sight (LoS) nature of air-to-ground
wireless channels gives rise to a new challenge on how to realize secure UAV communications with
the destined nodes on the ground. This paper aims to tackle this challenge by applying the physical
(PHY) layer security technique. We consider both the downlink and uplink UAV communications with
a ground node, namely UAV-to-ground (U2G) and ground-to-UAV (G2U) communication, respectively,
subject to a potential eavesdropper on the ground. In contrast to the existing literature on wireless
PHY layer security only with ground nodes at fixed or quasi-static locations, we exploit the high
mobility of the UAV to proactively establish favorable and degraded channels for the legitimate and
eavesdropping links, respectively, via its trajectory design. We formulate new problems to maximize
the average secrecy rates of the U2G and G2U transmissions, respectively, by jointly optimizing the
UAV’s trajectory and the transmit power of the legitimate transmitter over a given flight period of
the UAV. Although the formulated problems are non-convex, we propose iterative algorithms to solve
them efficiently by applying the block coordinate descent and successive convex optimization methods.
Specifically, the transmit power and UAV trajectory are each optimized with the other being fixed in
an alternating manner, until they converge. Simulation results show that the proposed algorithms can
improve the secrecy rates for both U2G and G2U communications, as compared to other benchmark
schemes without power control and/or trajectory optimization.
G. Zhang and M. Cui are with the School of Information Engineering, Guangdong University of Technology, Guangzhou,
China (email: {gczhang, cuimiao}@gdut.edu.cn). Q. Wu and R. Zhang are with the Department of Electrical and Computer
Engineering, National University of Singapore (email: {elewuqq, elezhang}@nus.edu.sg). Q. Wu is the corresponding author.
Part of this paper has been presented in IEEE Global Communications Conference (GLOBECOM), Singapore, Dec. 2017 [1].
2
Index Terms
5G and UAV communications, physical layer security, secrecy rate maximization, trajectory design,
power control.
I. I NTRODUCTION
With many advantages such as high mobility, low cost, wide coverage, and on-demand deployment, unmanned aerial vehicles (UAVs) have been extensively used in both military and
civilian applications, such as search and rescue, inspection and surveillance, cargo transportation,
etc. Recently, UAVs have also found increasingly more substantial applications in wireless
communication [2], and are expected to play a significant role in the forthcoming fifth-generation
(5G) wireless networks. To seize this growing opportunity, internationally leading telecommunication companies such as Qualcomm, Ericsson, and China Mobile have already launched
their research projects on integrating UAVs into the 5G networks [3], [4]. Generally speaking,
there are two main paradigms of UAV applications in 5G. In the first one, termed as “UAVassisted wireless communication”, UAVs are utilized as airbone communication platforms such
as mobile base stations (BSs) and/or relays that can be flexibly deployed on demand to assist
the communications in terrestrial networks such as 5G. For example, UAV-mounted BSs can
be used to enable rapid wireless communication service recovery after ground infrastructure
damages, or provide offloading service for terrestrial BSs in extremely crowded areas [5]–[11].
Another example is to use UAVs as mobile relays to provide reliable connectivity between distant
users in remote areas (e.g., an uninhabited desert) that are not covered by any existing wireless
networks [12], [13]. Moreover, in future internet of things (IoT) applications, UAVs can be
dispatched to disseminate/collect data to/from widespread distributed wireless devices efficiently
and with low cost [14], [15]. By contrast, in the other paradigm, known as “cellular-enabled
UAV communication”, UAVs are regarded as new “sky” users in the cellular networks that enable
two-way communications of the UAVs with ground BSs. For example, the future 5G networks
can provide reliable communications for UAVs even beyond the range of their operators’ visual
line-of-sight (LoS) to achieve long-range UAV control in real time [16]. Besides, in UAV-enabled
surveillance applications, the captured pictures and/or videos by the UAVs in real time can be
uploaded timely to the ground data centers via the 5G networks [17].
In the aforementioned UAV communication applications in 5G, due to the broadcast nature
of wireless channels, their security and privacy are of utmost concern [18], [19]. One major
3
advantage of UAV-ground communications is that UAVs usually have LoS channels for the
communications with ground nodes, especially in outdoor environments. However, such LoS
communication links are also more prone to the eavesdropping by illegitimate nodes on the
ground, which gives rise to a new security challenge. Although security was conventionally
viewed as a higher layer communication protocol stack design problem that can be tackled
by using cryptographic methods, physical (PHY) layer security has emerged as a promising
alternative way of defense to realize secrecy in wireless communication.
A key design metric that has been widely adopted in PHY layer security is the so-called secrecy
rate [18]–[25], at which confidential message can be reliably transmitted without having the
eavesdropper infer any information about the message. A non-zero secrecy rate can be achieved
when the strength of the legitimate link is stronger than that of the eavesdropping link. In the
existing literature on PHY layer security, communication nodes are usually assumed to be at fixed
or quasi-static locations. As a result, the average channel quality of the legitimate/eavesdropping
link mainly depends on the path loss and shadowing from the transmitter to receiver, which
are determined if the locations of the legitimate transmitter/receiver and the eavesdropper are
given. Thus, in the case that the average channel gain of the legitimate receiver is smaller
than that of the eavesdropper (e.g., due to longer distance from the legitimate transmitter),
in order to achieve positive secrecy rates, the exploitation of the wireless channel small-scale
fading in time, frequency, and/or space becomes essential, and various techniques such as power
control in time and/or frequency as well as multi-antenna beamforming have been investigated.
In [18], power control with rate adaptation over fading channels is proposed to maximize the
average secrecy rate. This work is also extended to characterize the secrecy rate region of
parallel-fading broadcast channels [19]. In [20], power control over frequency subcarriers is
investigated for secrecy rate maximization in an orthogonal frequency-division multiple access
(OFDMA) system. In [21], joint power control on information signal and artificial noise (AN) is
proposed to maximize the secrecy rate of a simultaneous wireless information and power transfer
(SWIPT) system. In multiple-input multiple-output (MIMO) systems, transmit beamforming can
be jointly employed with AN transmission to effectively enhance the legitimate link capacity and
at the same time degrade that of the eavesdropping link. For example, the legitimate transmitter
can use beamforming to steer a null to the eavesdropper, or send AN in the direction of the
eavesdropper to interferer with it [22], [23]. If one or more relay helpers are available, they
can also cooperatively send AN or jamming signals to interfere with the eavesdroppers to
4
achieve better secrecy communication performance. In [24], optimal cooperative jamming via
relays is studied to maximize the secrecy rate of a single-antenna point-to-point legitimate link.
Besides, transmission scheduling by exploiting multiuser channel diversity is another effective
approach to improve the secrecy communication performance in a system with multiple legitimate
users/eavesdroppers. In [25], a transmission scheduling scheme is proposed to maximize the
secrecy rate of a multiuser cognitive radio network.
However, there are still two major challenges that remain unsolved in the existing PHY layer
security literature. First, the practically achievable secrecy rate can be severely limited if the
distance between the legitimate transmitter and its intended receiver is fixed and significantly
larger than that between it and a potential eavesdropper, even if the various approaches mentioned
above are applied. Second, the channel state information (CSI) of the eavesdropper is usually
required at the legitimate transmitter for the implementation of effective power control and/or
beamforming techniques. This is practically challenging since the eavesdropper is usually a
passive device and thus it is difficult to estimate such CSI. In this paper, we study PHY layer
security in UAV-ground communications, which may potentially overcome the above two critical
issues in conventional studies. First, in contrast to the existing literature with fixed or quasistatic nodes only, the high mobility of UAVs can be exploited to proactively establish stronger
links with the legitimate ground nodes and/or degrade the channels of the eavesdroppers, by
flying closer/farther to/from them, respectively, via proper trajectory design. This approach is
particularly effective in the context of UAV-ground communications (as compared to conventional
terrestrial communications), since the LoS links are usually much more dominant over other
channel impairments such as shadowing and small-scale fading, due to the much larger height
of the UAV than typical ground nodes such as mobile terminals or BSs. Furthermore, since
the LoS channel gain is only determined by the link distance, the UAV can practically obtain
the channel gain to any potential eavesdropper on the ground if its location is known, which
thus resolves the eavesdropper-CSI issue in the existing literature. Note that the location of any
ground node as a potential eavesdropper can be practically detected and tracked by the UAV via
using an optical camera or synthetic aperture radar (SAR) equipped on the UAV [26], [27].
For an initial exposition, in this paper we consider a simplified three-node secrecy UAVground communication system as shown in Fig. 1, where a UAV at fixed altitude intends to
communicate with a ground node, while a potential eavesdropper on the ground may intercept
their communication. The secure communications of both UAV-to-ground (U2G) and ground-
5
z
v
H
(x(t), y(t), H)
UAV
Legitimate
Link (G2U)
Legitimate
Link (U2G)
Eavesdropping
Link (U2G)
y
Eavesdropping
Link (G2U)
Ground Node
Eavesdropper
x
Fig. 1. A UAV wireless communication system consisting of a UAV above ground and a node on the ground. A potential
eavesdropper on the ground intends to intercept the wireless communication between them.
to-UAV (G2U) links are considered. In the U2G case, the UAV and the ground node are the
legitimate transmitter and receiver, respectively, where both the legitimate and eavesdropping
links are modeled as LoS channels. By contrast, in the G2U case, the ground node and the
UAV are the legitimate transmitter and receiver, respectively. Since the legitimate transmitter
and potential eavesdropper are both on the ground in this case, different from the U2G case,
only the legitimate link is modeled as a LoS channel, while the eavesdropping link is practically
modeled as a channel consisting of both distance-dependent path-loss and small-scale Rayleigh
fading. Thus, the problem formulations for the secrecy rate maximization in these two cases
are generally different, which will be detailed later in this paper. Nevertheless, the secrecy rates
of both U2G and G2U transmissions can benefit from the joint design of UAV trajectory and
transmit power control at the legitimate transmitter (i.e., UAV and ground node in the U2G
and G2U cases, respectively), explained as follows. On one hand, the location of the UAV can
be adjusted dynamically to establish stronger channels for the legitimate link than that for the
eavesdropping link. On the other hand, due to practical constraints such as the UAV’s initial and
final locations, the legitimate link may not be always stronger than the eavesdropping link during
the whole flight period of the UAV. In this case, transmit power can be adapted to the channel
variations arising from the UAV’s movement to further improve the secrecy rate. For example,
6
in the U2G case, the UAV should transmit higher power when it flies closer to the ground node
while being more far away from the eavesdropper, and transmit lower or zero power otherwise.
Motivated by this, we aim to design joint UAV trajectory and transmit power optimization
algorithms to secure both U2G and G2U communications. Our goal is to maximize the average
secrecy rate over a finite flight period of the UAV in each case, subject to the practical mobility
constraints on the UAV’s maximum speed and its initial and final locations, as well as the average
and peak transmit power constraints. For the U2G case, the formulated joint trajectory and power
control problem for average secrecy rate maximization is difficult to be solved directly due to
its non-smooth objective function. To tackle this difficulty, we reformulate the problem into an
equivalent problem with a smooth objective function without loss of optimality. Although the nonsmoothness issue is resolved, the reformulated problem is still non-convex due to the coupling
of the transmit power and UAV trajectory optimization variables. We thus propose an efficient
iterative algorithm for solving this problem approximately based on the block coordinate descent
method. Specifically, we divide the optimization variables into two blocks, one for transmit
power control and the other for UAV trajectory optimization. Then the two blocks of variables
are optimized alternately in an iterative manner, i.e., in each iteration one block is optimized
with the other block fixed and vice versa. One corresponding sub-problem that optimizes the
UAV trajectory under given transmit power is still difficult to solve due to its non-convexity.
We thus apply the successive convex optimization method to solve the problem approximately.
Finally, we show that our proposed joint optimization algorithm is guaranteed to converge.
On the other hand, for the G2U case, similar to the U2G case, we also propose an efficient
algorithm to solve the formulated problem by using the block coordinate descent and successive
convex optimization methods, while some modification is made in the problem formulation to
deal with the non-LoS channel of the eavesdropper link in this case. Simulation results show
that the proposed joint trajectory and transmit power designs can improve the average secrecy
rates in both U2G and G2U communications, as compared to other benchmark schemes without
applying the trajectory optimization and/or transmit power control. Furthermore, it is observed
that trajectory optimization and transmit power control are both essential for the U2G case, while
for the G2U case, trajectory optimization is less effective as compared to power control.
It is worth noting that there have been prior works (e.g. [9], [12], [28], [29]) on trajectory
optimization for various UAV communication systems, which consider different system setups
and design objectives. In [9], a UAV mobile BS serving multiple users is considered, where
7
the UAV trajectory and multiuser scheduling are jointly designed to maximize the minimum
throughput of the users. In [12], a UAV-enabled mobile relaying system is investigated, where
the UAV trajectory and transmit power are jointly designed to maximize the throughput. In [28],
the UAV flying heading is optimized to maximize the achievable sum rate from ground nodes to
a UAV by assuming a constant flying speed. In [29], a new design paradigm that jointly considers
both the communication throughput and the UAV’s flying energy consumption is proposed to
maximize the energy efficiency of a point-to-point U2G communication system. Different from
these prior works, in this paper, we apply both trajectory optimization and transmit power control
to maximize the secrecy rates of both U2G and G2U communications.
The remainder of this paper is organized as follows. Section II presents the system model
and problem formulation. Sections III and IV present joint trajectory optimization and transmit
power control algorithms for the U2G and G2U cases, respectively. Section V provides simulation
results to validate the performance of the proposed algorithms as compared to three benchmark
schemes. Finally, Section VI concludes the paper.
II. S YSTEM M ODEL
AND
P ROBLEM F ORMULATION
A. System Model
As shown in Fig. 1, we consider a UAV-enabled wireless communication system where a
UAV above ground and a node on the ground communicate with each other, while a potential
eavesdropper on the ground aims to intercept the communications between them. We assume that
the distance between the ground node and the potential eavesdropper is L meters (m). Without
loss of generality, we consider a three-dimensional (3D) Cartesian coordinate system with the
ground node and the eavesdropper located at (0, 0, 0) and (L, 0, 0) in m, respectively. Their
locations are assumed to be fixed and known to the UAV, where the location of the eavesdropper
can be detected by using an optical camera or SAR equipped on the UAV. On the other hand,
the obtained secrecy rate when the location of the eavesdropper is known serves as an upper
bound for that when the location of the eavesdropper is not known.
We consider a given finite flight period of the UAV, with the duration denoted by T in seconds
(s). It is assumed that the UAV flies at a fixed altitude of H in m above ground, which can be
considered as the minimum altitude required for safety considerations such as terrain or building
avoidance. The coordinate of the UAV over time is denoted as (x(t), y(t), H) in m, 0 ≤ t ≤ T .
For convenience, we divide the period T into N time slots with equal length, i.e., T = Ndt , with
8
dt in s denoting the length of a time slot, which is chosen sufficiently small such that the UAV’s
location can be regarded as unchanged within each time slot from the viewpoint of the ground
node. As a result, the UAV’s coordinate in slot n can be denoted as (x[n], y[n], H), and the UAV’s
horizontal trajectory (x(t), y(t)) over the flight period T can be approximated by the sequence
{x[n], y[n]}N
n=1 . Denote the maximum speed of the UAV as vmax in m/s. Thus, the maximum
flying distance of the UAV in each slot is D = vmax dt . The initial and final locations of the UAV
are assumed to be given, which are denoted by (x0 , y0 , H) and (xF , yF , H) in m, respectively.
For the UAV trajectory to be feasible, we assume that the distance between the initial and final
p
location satisfies that (xF − x0 )2 + (yF − y0 )2 ≤ vmax T . As a result, the mobility constraints
of the UAV can be expressed as
(x[1] − x0 )2 + (y[1] − y0 )2 ≤ D 2 ,
(1a)
(x[n + 1] − x[n])2 + (y[n + 1] − y[n])2 ≤ D 2 , n = 1, . . . , N − 1,
(1b)
(xF − x[N])2 + (yF − y[N])2 ≤ D 2 .
(1c)
We consider both the U2G and G2U communications in the system of interest, which are
specified in detail in the following, respectively.
1) U2G Transmission: In the U2G case, the UAV and the ground node play the roles of
legitimate transmitter and receiver, respectively. The legitimate link from the UAV to the ground
node and the eavesdropping link from the UAV to the eavesdropper are both assumed to be
LoS channels. The measurement results in [30] have shown that the LoS model offers a good
approximation for practical UAV-ground communications. Thus, the LoS channel power gain
from the UAV to the ground node in time slot n follows the free-space path loss model, given
by
gUG [n] = β0 d−2
UG [n] =
x2 [n]
β0
,
+ y 2 [n] + H 2
(2)
where β0 denotes the channel power gain at the reference distance d0 = 1m, which depends
on the carrier frequency and the antenna gains of the transmitter and receiver, and dUG [n] =
p
x2 [n] + y 2 [n] + H 2 is the distance from the UAV to the ground node in time slot n. Similarly,
the LoS channel power gain from the UAV to the eavesdropper in time slot n is given by
gUE [n] =
(x[n] −
L)2
β0
.
+ y 2 [n] + H 2
(3)
9
We denote p[n] as the transmit power of the UAV in time slot n. In practice, p[n]’s are usually
subject to both average and peak limits over time, denoted by P̄ and Ppeak , respectively. Thus,
the transmit power constraints are expressed as
N
1 X
p[n] ≤ P̄ ,
N n=1
0 ≤ p[n] ≤ Ppeak , ∀n.
(4a)
(4b)
To make the constraint in (4a) non-trivial, we assume P̄ < Ppeak in this paper. In the absence
of the eavesdropper, the achievable rate from the UAV to the ground node in bits/second/Hertz
(bps/Hz) in time slot n can be expressed as
p[n]gUG [n]
RUG [n] = log2 1 +
σ2
γ0 p[n]
,
= log2 1 + 2
x [n] + y 2 [n] + H 2
(5)
where σ 2 is the additive white Gaussian noise (AWGN) power at the receiver and γ0 = β0 /σ 2
is the reference signal-to-noise ratio (SNR). Similarly, the achievable rate from the UAV to the
eavesdropper in bps/Hz in time slot n is given by
γ0 p[n]
RUE [n] = log2 1 +
.
(x[n] − L)2 + y 2 [n] + H 2
(6)
With (5) and (6), the average secrecy rate achievable for the U2G link in bps/Hz over the total
N time slots is given by [18]
+
N
γ0 p[n]
γ0 p[n]
1 X
(U2G)
log2 1 + 2
− log2 1 +
,
Rsec =
N n=1
x [n] + y 2 [n] + H 2
(x[n] − L)2 + y 2 [n] + H 2
(7)
where [x]+ , max(x, 0).
2) G2U Transmission: In the G2U case, the ground node and the UAV play the roles of
legitimate transmitter and receiver, respectively. The legitimate channel from the ground node to
the UAV is assumed to be LoS, similar as in the U2G case, whose channel power gain in time
slot n is given by
gGU [n] =
x2 [n]
β0
.
+ y 2[n] + H 2
(8)
Since both the ground node and the eavesdropper are on the ground, the eavesdropping channel
between them is assumed to constitute both distance-dependent path loss with pass-loss exponent
10
κ ≥ 2 and small-scale Rayleigh fading. Thus, the channel power gain from the ground node to
the eavesdropper at any time is given by
gGE =
β0
ζ,
Lκ
(9)
where ζ is an exponentially distributed random variable with unit mean accounting for the
Rayleigh fading.
We denote q[n] as the transmit power of the ground node in time slot n. Similar to the U2G
case, q[n]’s are constrained by average power limit Q̄ and peak power limit Qpeak , i.e.,
N
1 X
q[n] ≤ Q̄,
N n=1
0 ≤ q[n] ≤ Qpeak , ∀n,
(10a)
(10b)
where Q̄ < Qpeak is assumed. Similar to (5), the achievable rate from the ground node to the
UAV in bps/Hz in time slot n can be expressed as
γ0 q[n]
.
RGU [n] = log2 1 + 2
x [n] + y 2[n] + H 2
(11)
The achievable rate from the ground node to the eavesdropper in bps/Hz in time slot n is
expressed as
γ0 q[n]
ζ
RGE [n] = Eζ log2 1 +
Lκ
γ0 q[n]
≤ log2 1 +
Eζ [ζ]
Lκ
γ0 q[n]
= log2 1 +
,
Lκ
(12a)
(12b)
(12c)
where Eζ [·] in (12a) denotes the mathematical expectation with respect to random variable ζ,
and the inequality in (12b) is due to Jensen’s inequality and the fact that log2 (1 + γ0 q[n]ζ/Lκ ) is
concave with respect to ζ. (12c) shows an upper bound of RGE [n]. We consider the worst-case
secrecy rate performance by assuming that the eavesdropper is able to achieve this upper bound.
With (11) and (12c), the following average secrecy rate of the G2U link in bps/Hz over the total
N time slots is thus achievable,
+
N
γ0 q[n]
γ0 q[n]
1 X
(G2U)
log2 1 + 2
− log2 1 +
.
Rsec =
N n=1
x [n] + y 2 [n] + H 2
Lκ
(13)
11
B. Problem Formulation
(U2G)
For the U2G case, our goal is to maximize the average secrecy rate Rsec
in (7) by jointly
†
optimizing the UAV’s transmit power p , [p[1], . . . , p[N]] and the UAV’s trajectory in terms
of its horizontal coordinates x , [x[1], . . . , x[N]]† and y , [y[1], . . . , y[N]]† over all the N
time slots, where the superscript † denotes the transpose operation. The optimization variables
are subject to the UAV’s mobility constraints in (1) and the average and peak transmit power
constraints in (4). We formulate the secrecy rate maximization problem as follows (by dropping
the constant term 1/N in (7))
(P1) : max
x,y,p
N
X
log2
n=1
− log2 1 +
γ0 p[n]
1+ 2
x [n] + y 2 [n] + H 2
γ0 p[n]
(x[n] − L)2 + y 2 [n] + H 2
+
(14)
s.t. (1), (4).
(G2U)
Similarly, for the G2U case, we maximize Rsec
in (13) by jointly optimizing the ground
node’s transmit power q , [q[1], . . . , q[N]]† and the UAV’s horizontal trajectory x and y. The
problem is thus formulated as
N
X
log2 1 +
(P2) : max
x,y,q
n=1
γ0 q[n]
2
x [n] + y 2 [n] + H 2
+
γ0 q[n]
− log2 1 +
Lκ
(15)
s.t. (1), (10).
Note that different from problem (P1), only the first logarithmic function in the objective of
γ0 q[n]
(P2), i.e., log2 1 + x2 [n]+y
2 [n]+H 2 , contains the UAV trajectory variables. This is because the
achievable rate from the ground node to the eavesdropper does not depend on the trajectory of
the UAV.
Problems (P1) and (P2) are both difficult to be solved optimally due to the following two
reasons. First, the operator [·]+ makes the objective functions of (P1) and (P2) non-smooth at
zero value. Second, even without [·]+ , their objective functions are non-concave with respect to
either x, y, or p. In Sections III and IV, we propose efficient algorithms for solving problems
(P1) and (P2) approximately, respectively.
III. P ROPOSED A LGORITHM
FOR
P ROBLEM (P1)
First, we consider problem (P1) for the U2G case. To handle the non-smoothness of the
objective function of (P1), the following lemma is used.
12
Lemma 1. Problem (P1) has the same optimal value as that of the following problem,
N
X
γ0 p[n]
log2 1 + 2
(P3) : max
x,y,p
x [n] + y 2[n] + H 2
n=1
γ0 p[n]
− log2 1 +
(x[n] − L)2 + y 2[n] + H 2
(16)
s.t. (1), (4).
Proof. Denote L1 and L3 as the optimal values of (P1) and (P3), respectively. First, we have
L1 ≥ L3 , since the objective function of (P1) is no smaller than that of (P3), and (P1) and (P3)
have the same constraints.
Next, we show L3 ≥ L1 also holds. Denote (x∗ , y∗ , p∗ ) as the optimal solution to (P1), where
x∗ = [x∗ [1], . . . , x∗ [N]]† , y∗ = [y ∗[1], . . . , y ∗[N]]† , and p∗ = [p∗ [1], . . . , p∗ [N]]† . Define
γ0 p[n]
γ0 p[n]
− log2 1 +
.
f (x[n], y[n], p[n]) = log2 1 + 2
x [n] + y 2[n] + H 2
(x[n] − L)2 + y 2[n] + H 2
We construct a feasible solution to (P3), termed (x̃, ỹ, p̃), such that x̃ = x∗ , ỹ = y∗ , and the
elements of p̃ are obtained as
p̃[n] =
p∗ [n]
0
f (x∗ [n], y ∗ [n], p∗ [n]) ≥ 0,
f (x∗ [n], y ∗ [n], p∗ [n]) < 0.
Denote L̃ as the objective value of (P3) attained at (x̃, ỹ, p̃). The newly constructed solution
(x̃, ỹ, p̃) ensures that L̃ = L1 . Since (x̃, ỹ, p̃) is a feasible solution to (P3), it follows that
L3 ≥ L̃, and thus L3 ≥ L1 . Therefore, L1 = L3 , which completes the proof.
Based on Lemma 1, we only need to focus on solving problem (P3). Although problem (P3)
resolves the non-smoothness issue, it is still non-convex and difficult to solve. However, we
observe that the constraint (1) contains only the variables (x, y) for UAV trajectory optimization
and the constraint (4) contains only the variables p for transmit power control. As such, the
optimization variables of (P3) can be partitioned into two blocks, i.e., p and (x, y), respectively,
which facilitates the development of an iterative algorithm for solving problem (P3) by applying
the block coordinate descent method. Specifically, we solve problem (P3) by solving the following two sub-problems iteratively: one (denoted by sub-problem 1) optimizes the transmit power
p under given UAV trajectory (x, y), while the other (denoted by sub-problem 2) optimizes the
UAV trajectory (x, y) under given transmit power p, as detailed in the next two subsections,
respectively. Then, we present the overall algorithm and show that it is guaranteed to converge.
13
A. Sub-Problem 1: Optimizing Transmit Power Given UAV Trajectory
For given UAV trajectory (x, y), sub-problem 1 can be expressed as
max
p
N
X
log2 (1 + an p[n]) − log2 (1 + bn p[n])
n=1
(17)
s.t. (4),
where
γ0
,
+ y 2 [n] + H 2
γ0
bn =
.
2
(x[n] − L) + y 2 [n] + H 2
an =
(18)
x2 [n]
(19)
Although problem (17) is non-convex, it has been shown in [18] that the optimal solution can
be obtained as
p∗ [n] =
where
s
min ([p̂[n]]+ , Ppeak )
0
1
1
−
2bn 2an
2
an > bn ,
(20)
an ≤ bn ,
1
1
−
bn an
1
1
−
.
(21)
2bn 2an
P
In (21), λ ≥ 0 is a constant that ensures the average power constraint N1 N
n=1 p[n] ≤ P̄ to be
p̂[n] =
1
+
λ ln 2
−
satisfied when the optimal solution of problem (17) is attained, which can be found efficiently
via a one-dimensional bisection search [31].
B. Sub-Problem 2: Optimizing UAV Trajectory Given Transmit Power
For given transmit power p, by letting Pn = γ0 p[n], we can express sub-problem 2 as
N
X
Pn
log2 1 + 2
max
x,y
x [n] + y 2 [n] + H 2
n=1
Pn
− log2 1 +
(x[n] − L)2 + y 2[n] + H 2
(22)
s.t. (1).
Note that the objective function of problem (22) is non-concave with respect to x and y, so it is
a non-convex optimization problem and cannot be solved optimally in general. By introducing
14
slack variables t = [t[1], . . . , t[N]]† and u = [u[1], . . . , u[N]]† , we first consider the following
problem,
max
x,y,t,u
N
X
n=1
Pn
Pn
− log2 1 +
log2 1 +
u[n]
t[n]
s.t. t[n] − x2 [n] + 2Lx[n] − L2 − y 2[n] − H 2 ≤ 0, ∀n,
(23a)
(23b)
x2 [n] + y 2[n] + H 2 − u[n] ≤ 0, ∀n,
(23c)
t[n] ≥ H 2 , ∀n,
(23d)
(1).
At the optimal solution of problem (23), constraints (23b) and (23c) should hold with equalities,
since otherwise t[n] (u[n]) can be increased (decreased) to improve the objective value. Therefore,
problems (22) and (23) have the same optimal solution of (x, y). Next, we focus on solving
problem (23).
The term log2 1 +
Pn
u[n]
in (23a) is convex with respect to u[n], and the terms −x2 [n] and
−y 2 [n] in (23b) are concave with respect to x[n] and y[n], respectively. However, a maximization
problem with a non-concave objective function and/or a non-convex feasible region is in general
non-convex and thus difficult to be solved optimally. Based on the facts that the first-order Taylor
expansion of a convex function is its global under-estimator and that of a concave function is its
global over-estimator [31], we propose an iterative algorithm to solve problem (23) approximately
by applying the successive convex optimization method. The algorithm obtains an approximate
solution to problem (23) by maximizing a concave lower bound of its objective function within
a convex feasible region, which is detailed as follows. First, the algorithm assumes a given
initial point (xfea , yfea , ufea ) which is feasible to (23), where xfea , [xfea [1], . . . , xfea [N]]† , yfea ,
[yfea [1], . . . , yfea [N]]† , and ufea , [ufea [1], . . . , ufea [N]]† . Then, by using the first-order Taylor
Pn
, −x2 [n], and −y 2 [n] at the points ufea [n], xfea [n], and yfea [n],
expansions of log2 1 + u[n]
respectively, i.e.,
Pn
Pn (u[n] − ufea [n])
Pn
≥ log2 1 +
−
,
log2 1 +
u[n]
ufea [n]
ln 2(u2fea [n] + Pn ufea [n])
(24)
− x2 [n] ≤ x2fea [n] − 2xfea [n]x[n],
(25)
2
− y 2 [n] ≤ yfea
[n] − 2yfea [n]y[n],
(26)
15
problem (23) is approximated as
N
X
max
−
x,y,t,u
n=1
Pn
Pn u[n]
− log2 1 +
ln 2(u2fea [n] + Pn ufea [n])
t[n]
(27a)
s.t. t[n] + x2fea [n] − 2xfea [n]x[n] + 2Lx[n] − L2
2
+ yfea
[n] − 2yfea [n]y[n] − H 2 ≤ 0, ∀n,
(27b)
(1), (23c), (23d).
After such approximation, we note that the objective function of problem (27) is concave and
its feasible region is convex. Thus, problem (27) is a convex optimization problem, which can
be optimally solved by the interior-point method [31]. Since the first-order Taylor expansions of
−x2 [n] and −y 2 [n] are their global over-estimators, any solution (x[n], y[n]) satisfying (27b) will
satisfy (23b). As a result, the solution of problem (27) is guaranteed to be a feasible solution of
Pn
problem (23). Moreover, the first-order Taylor expansion of log2 1 + u[n] is its global under-
estimator. As such, problem (27) maximizes a lower bound of the objective function of problem
(23), and the lower bound and the objective function of (23) are equal only at the given point
(xfea , yfea , ufea ); thus, the objective value of problem (23) with the solution obtained by solving
problem (27) is no smaller than that with the given point (xfea , yfea , ufea ).
C. Overall Algorithm
In summary, the overall algorithm applies the block coordinate descent method, and solves the
two sub-problems (17) and (27) alternately in an iterative manner. The details of the proposed
algorithm are summarized in Algorithm 1. As shown in the previous two subsections, the
objective value of (P3) with the solutions obtained by solving the sub-problems (17) and (27)
is non-decreasing over iterations, while the optimal value of (P3) is finite; thus, the solution by
the proposed algorithm is guaranteed to converge.
IV. P ROPOSED A LGORITHM
FOR
P ROBLEM (P2)
In this section, we consider problem (P2) for the G2U case. Similar to (P1), we solve problem
(P2) by considering the following equivalent problem,
N
X
γ0 q[n]
γ0 q[n]
− log2 1 +
(P4) : max
log2 1 + 2
x,y,q
x [n] + y 2[n] + H 2
Lκ
n=1
s.t. (1), (10).
(28)
16
Algorithm 1 Proposed Algorithm for Problem (P1).
1: Initialization: Set an initial feasible solution (p(0) , x(0) , y(0) ), an initial slack variable u(0)
and k = 0.
2:
repeat
3:
Set k ← k + 1;
4:
With given p(k−1) , set the feasible solution xfea = x(k−1) , yfea = y(k−1) and ufea = u(k−1) ,
then update the trajectory variable (x(k) , y(k) ) and the slack variable u(k) by solving
problem (27);
5:
6:
With given (x(k) , y(k) ), update the transmit power control variable p(k) by using (20).
until the fractional increase of the objective value is below a small threshold ǫ.
Although problem (P4) is non-convex, it has a similar structure as problem (P3), which facilitates
us to also apply the block coordinate descent method to find an approximate solution for it. Like
(P3), problem (P4) can also be decomposed into two sub-problems: one (denoted by sub-problem
3) is to optimize transmit power q under given trajectory (x, y); while the other (denoted by
sub-problem 4) is to optimize trajectory (x, y) under given transmit power q. The two subproblems are solved alternately in an iterative manner until convergence. Next, we discuss on
how to solve the two sub-problems, respectively.
A. Sub-Problem 3: Optimizing Transmit Power Given UAV Trajectory
With given (x, y), sub-problem 3 can be expressed as
max
q
N
X
log2 (1 + an q[n]) − log2 (1 + bq[n])
n=1
(29)
s.t. (10),
where an is defined in (18) and
b=
γ0
.
Lκ
(30)
Problem (29) is similar to sub-problem 1 given in (17), thus it can be solved by using (20) and
(21), provided that bn in (20) and (21) is replaced with b in (30).
17
B. Sub-Problem 4: Optimizing UAV Trajectory Given Transmit Power
With given q, by letting Qn = γ0 q[n] and removing the terms in the objective function in (28)
which are irrelevant to x and y, we express sub-problem 4 as
N
X
Qn
max
log2 1 + 2
x,y
x [n] + y 2[n] + H 2
n=1
(31)
s.t. (1).
Unlike sub-problem 2 given in (22) for the U2G case, problem (31) is simplified as maximizing
only the average achievable rate from the ground node to the UAV. This is because the trajectory
of the UAV determines only the channel gain from the ground node to the UAV, but does not
have any effect on the channel from the ground node to the eavesdropper.
Despite the non-convexity of problem (31), we apply the successive convex optimization
to approximately solve it, similar to problem (22). First, we introduce slack variable u =
[u[1], . . . , u[N]]† , and solve the following problem which has the same optimal solution of (x, y)
as problem (31),
N
X
Qn
log2 1 +
max
x,y,u
u[n]
n=1
s.t. x2 [n] + y 2 [n] + H 2 − u[n] ≤ 0, ∀n,
(32a)
(32b)
(1).
With a given initial point ufea , [ufea [1], . . . , ufea [N]]† , which is feasible to (32), and by applying
Qn
the first-order Taylor expansion of log2 1 + u[n]
given in (24) (where Pn is replaced by Qn ),
problem (32) is recast as
max
x,y,u
N
X
n=1
−
Qn u[n]
2
ln 2(ufea [n] + Qn ufea [n])
(33)
s.t. (1), (32b).
In can be shown that problem (33) is a convex quadratically constrained quadratic programming
(QCQP) problem, and thus can be efficiently solved by the interior-point method [31]. The details
of the overall algorithm for solving (P2) are omitted for brevity, given the similarity to that for
(P1).
18
V. S IMULATION R ESULTS
In this section, we provide simulation results to verify the performance of our proposed joint
UAV trajectory optimization and transmit power control algorithm (denoted as “joint traj. opt. &
pow. ctrl.”). For comparison, we also consider the following three benchmark schemes without
optimized trajectory and/or power control:
•
Trajectory optimization without transmit power control (denoted as “traj. opt. w/o pow.
ctrl.”);
•
Best-effort trajectory design with transmit power control (denoted as “best-effort traj. w/
pow. ctrl.”);
•
Best-effort trajectory design without transmit power control (denoted as “best-effort traj.
w/o pow. ctrl.”).
Specifically, the “traj. opt. w/o pow. ctrl.” algorithm designs the UAV trajectories for the U2G
and G2U cases by solving problems (27) and (33) iteratively until convergence, respectively,
with transmit power equally set over time, i.e., p[n] = P̄ and q[n] = Q̄, ∀n. Both the “best-effort
traj. w/ pow. ctrl.” and “best-effort traj. w/o pow. ctrl.” algorithms design the UAV trajectories
in the following heuristic best-effort manner: the UAV first flies straight to the point above the
ground node at its maximum speed, then remains static at that point (if time permits), and finally
flies at its maximum speed to reach its final location by the end of the last time slot. Note that
if the UAV does not have sufficient time to reach the point above the ground node, it will turn
at a certain midway point and then fly to the final location at the maximum speed. Given this
trajectory, the “best-effort traj. w/ pow. ctrl.” algorithm optimizes the transmit power in the U2G
or G2U case by solving problem (17) or (29), respectively; while the “best-effort traj. w/o pow.
ctrl.” algorithm sets transmit power equally over time, i.e., p[n] = P̄ and q[n] = Q̄, ∀n. The
initial feasible solutions for the proposed “joint traj. opt. & pow. ctrl.” and benchmark “traj. opt.
w/o pow. ctrl.” algorithms are generated by the “best-effort traj. w/o pow. ctrl.” algorithm.
The distance between the ground node and the eavesdropper is set as L = 200m, and the flying
altitude of the UAV is set as H = 100m. The maximum speed of the UAV is vmax = 3m/s.
The flight time T is divided into multiple time slots with equal length of dt = 0.5s. The
communication bandwidth is 20MHz with the carrier frequency at 5GHz, and the noise power
spectrum density is −169dBm/Hz. Thus, the reference SNR at the reference distance d0 = 1m
is γ0 = 80dB. The peak transmit power limits for the U2G and G2U links are set as Ppeak = 4P̄
19
600
Initial location
Best-effort traj.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
400
T=400s
200
y (m)
T=403s
Ground node
Eavesdropper
0
-200
T=600s
-400
Final location
-600
-50
0
50
100
150
200
x (m)
Fig. 2. Trajectories of the UAV for the U2G communication in Case 1.
and Qpeak = 4Q̄, respectively. The threshold in Algorithm 1 is set as ǫ = 10−4 . We consider two
cases, denoted as Case 1 and Case 2, in which the UAV has different initial and final locations. In
Case 1, the UAV’s initial and final locations are on the vertical line in the middle of the ground
node and the eavesdropper, and the coordinates of them are set as (x0 , y0) = (100, 600)m and
(xF , yF ) = (100, −600)m, as shown in Fig. 2. In Case 2, the UAV’s initial and final locations are
on the parallel line of that connecting the ground node and the eavesdropper, and the coordinates
of them are set as (x0 , y0 ) = (−500, −150)m and (xF , yF ) = (700, −150)m, as shown in Figs.
5 and 8.
A. U2G Communication
In the U2G case, we first consider Case 1. Fig. 2 shows the trajectories of the UAV by applying
different algorithms with different values of flight time T . The average transmit power is set
as P̄ = −5dBm. The locations of the ground node, eavesdropper, as well as the UAV’s initial
and final locations are marked with
, △, ×, and +, respectively. It is observed that when
20
1.4
Secrecy Rate (bps/Hz)
1.2
1
P̄ =5dBm
0.8
0.6
P̄ =-5dBm
0.4
Best-effort traj. w/o pow. ctrl.
Best-effort traj. w/ pow. ctrl.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
0.2
0
400
450
500
550
600
650
700
T (s)
Fig. 3. Secrecy rate versus flight time T for the U2G communication in Case 1.
T = 400s which is the minimum time required for the UAV to fly from the initial location
to the final location in a straight line at the maximum speed, the trajectories of all algorithms
are identical. As T increases, the trajectories by the proposed “joint traj. opt. & pow. ctrl.” and
the benchmark “traj. opt. w/o pow. ctrl.” are still similar, i.e., the UAV tries to fly as close as
possible to the ground node and at the same time as far away as possible to the eavesdropper,
while they appear different compared to that by the heuristic best-effort trajectory design (same
for with and without power control). When T is sufficiently large, i.e. T = 600s, for the proposed
“joint traj. opt. & pow. ctrl.” or the benchmark “traj. opt. w/o pow. ctrl.”, it is observed that
the UAV first flies at the maximum speed to reach a certain location on the left of the ground
node (not directly above it), then remains stationary at this location as long as possible, and
finally flies to the final location in an arc path at the maximum speed and reaches there by
the end of the last time slot. These stationary locations, which are on the opposite direction of
the eavesdropper, strike an optimal balance between enhancing the legitimate link channel and
degrading the eavesdropping link channel and hence maximize the secrecy rate in each of these
21
1.4
Secrecy Rate (bps/Hz)
1.2
Best-effort traj. w/o pow. ctrl.
Best-effort traj. w/ pow. ctrl.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
1
T=600s
0.8
0.6
0.4
T=403s
0.2
0
-15
-10
-5
0
5
10
15
P̄ (dBm)
Fig. 4. Secrecy rate versus average transmit power P̄ for the U2G communication in Case 1.
two cases. In fact, this is also why the UAV follows an arc path rather than the straight path as
in the heuristic best-effort trajectory.
Fig. 3 shows the corresponding average secrecy rates of different algorithms versus flight time
T when P̄ = −5dBm and 5dBm. It is observed that the secrecy rates of all algorithms increase
significantly with T . This is because for all algorithms the maximum secrecy rate is achieved at
their respective stationary locations (see Fig. 2) for sufficiently large T , and larger T results in
longer hovering time at such locations for the UAV. The proposed “joint traj. opt. & pow. ctrl.”
algorithm always achieves the highest secrecy rate, while the benchmark “best-effort traj. w/o
pow. ctrl.” algorithm has the lowest secrecy rate, as expected. When P̄ = −5dBm, the benchmark
“best-effort traj. w/ pow. ctrl.” algorithm has higher secrecy rate than that of the benchmark “traj.
opt. w/o pow. ctrl.” algorithm. In contrast, when P̄ = 5dBm, the latter algorithm has higher
secrecy rate than the former. Such results suggest that in the low transmit power regime, power
control is more important for improving the secrecy rate; while in the high transmit power regime,
trajectory optimization is more significant. Furthermore, it is worth pointing out that trajectory
22
adaptation with increasing T is essential for the secrecy rate improvement, even for the case with
heuristic best-effort trajectory. Otherwise, if the trajectory is fixed (e.g., following the straight line
p
from the initial to final location with constant speed of (xF − x0 )2 + (yF − y0 )2 /T regardless
of T , then the secrecy rate will remain unchanged as the case with required minimum T = 400s in
Fig. 3, i.e., the UAV cannot exploit its mobility to improve the secrecy rate, even with sufficiently
large T .
Fig. 4 shows the average secrecy rates of different algorithms versus the average transmit
power P̄ when T = 403s and 600s. It is observed that the proposed “joint traj. opt. & pow. ctrl.”
algorithm always achieves the highest secrecy rate, while the benchmark “best-effort traj. w/o
pow. ctrl.” algorithm has the lowest secrecy rate. In addition, when P̄ ≤ −5dB, we note that the
benchmark “best-effort traj. w/ pow. ctrl.” algorithm achieves a secrecy rate performance close
to the proposed “joint traj. opt. & pow. ctrl.” algorithm, and also significantly outperforms the
benchmark “traj. opt. w/o pow. ctrl.” algorithm. However, when P̄ increases, the secrecy rate of
the benchmark “traj. opt. w/o pow. ctrl.” algorithm will eventually exceed that of the benchmark
“best-effort traj. w/ pow. ctrl.” algorithm and even get closer to that of the proposed “joint traj.
opt. & pow. ctrl.” algorithm. Furthermore, the rate gap between the benchmark “traj. opt. w/o
pow. ctrl.” and benchmark “best-effort traj. w/ pow. ctrl.” becomes larger with increasing P̄ .
The above results further demonstrate that transmit power control is more effective in improving
secrecy rate than trajectory optimization when the average transmit power is low, while trajectory
optimization is more effective when the average transmit power is high.
Next, we consider Case 2. Fig. 5 shows the trajectories of the UAV by using different
algorithms when P̄ = −5dBm. It is observed that different from the results in Case 1 shown in
Fig. 2, the trajectories by the proposed “joint traj. opt. & pow. ctrl.” and benchmark “traj. opt.
w/o pow. ctrl.” algorithms with T = 405s or 600s differ significantly, especially when the UAV
flies towards the final location. For the proposed algorithm, the UAV flies along a relatively
direct path towards the ground node and then towards the final location. In contrast, for this
benchmark, the UAV first flies almost directly to the ground node, but then flies along an arc
path to the final location, which inevitably consumes more time on the traveling compared to
the proposed trajectory. The reason for such a difference is that in Case 2, flying from the
ground node towards the final location reduces the distance from the UAV to the eavesdropper
less much as compared to that from it to the ground node, which is undesired. This means that
to improve the secrecy rate, the UAV should reduce transmit power when it gets farther away
23
50
Best-effort traj.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
0
T=600s
Eavesdropper
Ground node
T=405s
y (m)
-50
-100
-150
Final location
Initial location
T=400s
-200
-600
-400
-200
0
200
400
600
800
x (m)
Fig. 5. Trajectories of the UAV for the U2G communication in Case 2.
from the ground node and closer to the final location. Considering this fact, the proposed “joint
traj. opt. & pow. ctrl.” algorithm is able to decrease the UAV transmit power or even turn off
the transmitter to save power and also protect from eavesdropping when the UAV flies directly
towards the final location. However, for the benchmark “traj. opt. w/o pow. ctrl.” algorithm that
employs a constant transmit power, the UAV can only rely on adjusting its trajectory to keep
far away from the eavesdropper to avoid being eavesdropped, which however requires more
traveling time and leads to a longer arc trajectory.
With such a UAV trajectory difference, the secrecy rate performances of different algorithms
versus T and P̄ , which are shown in Figs. 6 and 7 respectively, are also quite different from
those shown in Figs. 3 and 4 for Case 1. Specifically, the secrecy rate gaps between the proposed
“joint traj. opt. & pow. ctrl.” and benchmark “traj. opt. w/o pow. ctrl.” algorithms versus T or
P̄ in Case 2 are significantly larger than those in Case 1. For example, in Fig. 6, the “traj. opt.
w/o pow. ctrl.” algorithm even has lower secrecy rate than the “best-effort traj. w/o pow. ctrl.”
algorithm in the regime of T ≤ 650s when P̄ = −5dBm or in the regime of T ≤ 550s when
24
1.4
Secrecy Rate (bps/Hz)
1.2
1
P̄ =5dBm
0.8
0.6
P̄ =-5dBm
0.4
Best-effort traj. w/o pow. ctrl.
Best-effort traj. w/ pow. ctrl.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
0.2
0
400
450
500
550
600
650
700
T (s)
Fig. 6. Secrecy rate versus flight time T for the U2G communication in Case 2.
P̄ = 5dBm. In addition, in Fig. 7, the “traj. opt. w/o pow. ctrl.” algorithm has a lower secrecy
rate than the “best-effort traj. w/o pow. ctrl.” algorithm over the whole P̄ regime when T = 405s
and in the regime of P̄ ≤ 0dBm when T = 600s. This is mainly because the UAV wastes more
time on travelling along a longer arc trajectory to reach the final location which in turn leads
to the inefficient use of the transmit power. The above results demonstrate the importance and
necessity of the joint UAV trajectory optimization and transmit power control in maximizing the
secrecy rate for U2G communication.
B. G2U Communication
In the G2U case, the channel gain from the ground node to the eavesdropper given in (9)
contains a small-scale Rayleigh fading term ζ. Thus, all secrecy rate results in the following are
averaged over 5000 random independent realizations of ζ, where the path-loss exponent is set
as κ = 3. Since our results obtained for Cases 1 and 2 lead to consistent observations, we only
present the results for Case 2 due to the space limitation.
25
1.4
Secrecy Rate (bps/Hz)
1.2
Best-effort traj. w/o pow. ctrl.
Best-effort traj. w/ pow. ctrl.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
1
T=600s
0.8
0.6
0.4
T=405s
0.2
0
-15
-10
-5
0
5
10
15
P̄ (dBm)
Fig. 7. Secrecy rate versus average transmit power P̄ for the U2G communication in Case 2.
Fig. 8 shows the trajectories of the UAV with different values of T when the average transmit
power is Q̄ = −5dBm. It is observed that the trajectories of the proposed “joint traj. opt. & pow.
ctrl.” and benchmark “traj. opt. w/o pow. ctrl.” algorithms are very similar for different values of
T , i.e., the UAV tries to fly as close as possible to the ground node as T increases. When T is
sufficiently large, i.e., T = 600s, the trajectories of them are the same as the heuristic best-effort
trajectory, i.e., the UAV first flies at the maximum speed to reach the point right above the ground
node, then remains static as long as possible, and finally flies to the final location directly at the
maximum speed in order to reach there by the end of the last time slot. The fundamental reason
of such a result is that in the G2U setup, the channel between the ground node (transmitter)
and the eavesdropper is independent of the UAV’s location. Therefore, the UAV trajectory is
only optimized to maximize achievable rate from the ground node to the UAV. Obviously, the
point right above the ground node is the best location for achieving its largest rate. This explains
why the optimized trajectory also converges to the heuristic best-effort trajectory when T is
sufficiently large.
26
20
Ground node
T=600s
Eavesdropper
0
Best-effort traj.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
-20
-40
T=405s
y (m)
-60
-80
-100
-120
T=400s
-140
Initial location
-160
-600
-400
Final location
-200
0
200
400
600
800
x (m)
Fig. 8. Trajectories of the UAV for the G2U communication in Case 2.
Fig. 9 shows the average secrecy rates of different algorithms versus flight time T when
Q̄ = −5dBm and 5dBm. When T ≥ 410s, the algorithms with transmit power control, i.e. the
proposed “joint traj. opt. & pow. ctrl.” and benchmark “best-effort traj. w/ pow. ctrl.” algorithms,
achieve the same secrecy rate since they have the same trajectory and hence the same transmit
power control, while they both outperform the benchmark algorithms without power control, i.e.,
the “traj. opt. w/o pow. ctrl.” and “best-effort traj. w/o pow. ctrl.” algorithms. These results suggest
that transmit power control is more effective than trajectory optimization in improving secrecy
rate in the G2U case, as shown in Fig. 8, since the optimal trajectory can be easily achieved by
the heuristic best-effort design when T is sufficiently large. Furthermore, the secrecy rate gap
between the algorithms with and without power control when Q̄ = −5dBm is significantly larger
than that when Q̄ = 5dBm, since power control is more effective when the average transmit
power is low.
Fig. 10 shows the average secrecy rates of different algorithms versus average transmit power
Q̄ when T = 405s and 600s. It can be also observed that transmit power control is effective for
27
3.5
Secrecy Rate (bps/Hz)
3
2.5
Q̄=5dBm
2
1.5
1
0
400
Best-effort traj. w/o pow. ctrl.
Best-effort traj. w/ pow. ctrl.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
Q̄=-5dBm
0.5
450
500
550
600
650
700
T (s)
Fig. 9. Secrecy rate versus flight time T for the G2U communication in Case 2.
improving secrecy rate when Q̄ ≤ 0dBm. When T = 405s, the secrecy rate gap between the
proposed “joint traj. opt. & pow. ctrl.” algorithm and the benchmark algorithms with best-effort
trajectory exists due to their trajectory difference. When T = 600s, the secrecy rates of all
algorithms tend to be very similar when Q̄ ≥ 10dBm. This is because their trajectories are the
same and the power control only provides marginal rate gain when transmit power is high.
VI. C ONCLUSION
In this paper, we study the PHY layer security for emerging UAV communications in the
forthcoming 5G wireless networks. Specifically, we propose to enhance the security performance
by proactively controlling channel gains via adjusting the UAV trajectory in addition to applying
the conventional power/rate adaptation, which leads to a new joint optimization framework.
For both the U2G and G2U communications, the transmit power control and UAV trajectory
are jointly designed to maximize the average secrecy rate over a finite horizon, subject to the
average and peak transmit power constraints as well as practical UAV’s mobility constraints.
28
6
Secrecy Rate (bps/Hz)
5
Best-effort traj. w/o pow. ctrl.
Best-effort traj. w/ pow. ctrl.
Traj. opt. w/o pow. ctrl.
Joint traj. opt. & pow. ctrl.
4
3
T=600s
2
1
0
-15
T=405s
-10
-5
0
5
10
15
Q̄ (dBm)
Fig. 10. Secrecy rate versus average transmit power Q̄ for the G2U communication in Case 2.
By applying the block coordinate descent and successive convex optimization methods, efficient
iterative algorithms are proposed to solve the joint design problems. Simulation results show
that joint trajectory optimization and transmit power control improves the PHY layer security
performance, and more significantly in the U2G case compared to the G2U case, as the UAV
trajectory in the U2G case has an effect on both the legitimate and eavesdropping channels,
instead of the legitimate channel only in the G2U case. Furthermore, it is found that both UAV
trajectory optimization and transmit power control are generally necessary in the U2G case; while
in the G2U case, transmit power control is more effective than UAV trajectory optimization for
improving the secrecy rate performance, and the heuristic best-effort trajectory already performs
quite close to the optimized trajectory.
R EFERENCES
[1] G. Zhang, Q. Wu, M. Cui, and R. Zhang, “Securing UAV communications via trajectory optimization,” in Proc. IEEE
GLOBECOM, Dec. 2017, [Online]. Available: https://arxiv.org/abs/1710.04389.
29
[2] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with unmanned aerial vehicles: opportunities and challenges,”
IEEE Commun. Mag., vol. 54, no. 5, pp. 36-42, May 2016.
[3] “Paving the path to 5G: optimizing commercial LTE networks for drone communication,” [Online]. Available:
https://www.qualcomm.com/news/onq/2016/09/06/paving-path-5g-optimizing-commercial-lte-networks-dronecommunication.
[4] “Ericsson and
China
Mobile
conduct
world’s
first 5G
drone
prototype
field trial,” [Online]. Available:
https://www.ericsson.com/en/news/2016/8/ericsson-and-china-mobile-conduct-worlds-first-5g-drone-prototype-field-trial-.
[5] Q. Wu and R. Zhang, “Common throughput maximization in UAV-enabled OFDMA systems with delay consideration,”
submitted to IEEE Trans. Commun., 2017, [Online] Available: https://arxiv.org/abs/1801.00444
[6] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Drone small cells in the clouds: design, deployment and performance
analysis,” in Proc. IEEE GLOBECOM, Dec. 2015.
[7] R. I. Bor-Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-D placement of an aerial base station in next generation
cellular networks,” in Proc. IEEE ICC, May 2016.
[8] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization of UAV-mounted mobile base stations,” IEEE Commun.
Lett., vol. 21, no. 3, pp. 604-607, Mar. 2017.
[9] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication design for UAV-enabled multiple access,” in IEEE
Global Communications Conference (GLOBECOM), Dec. 2017.
[10] Q. Wu, Y. Zeng, and R. Zhang, “Joint trajectory and communication design for multi-UAV enabled wireless networks,”
IEEE Trans. Wireless Commun., 2017, [Online] Available: http://ieeexplore.ieee.org/document/8247211/.
[11] V. Sharma, M. Bennis, and R. Kumar, “UAV-assisted heterogeneous networks for capacity enhancement,” IEEE Commun.
Lett., vol. 20, no. 6, pp. 1207-1210, Jun. 2016.
[12] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for UAV-enabled mobile relaying systems,” IEEE Trans.
Commun., vol. 64, no. 12, pp. 4983-4996, Dec. 2016.
[13] T. A. Johansen, A. Zolich, T. Hansen, and A. J. Sørensen, “Unmanned aerial vehicle as communication relay for autonomous
underwater vehicle – field tests,” in IEEE Globecom Commun. Conf. Workshops, pp. 1469-1474, Dec. 2014.
[14] S. Sotheara, K. Aso, N. Aomi, and S. Shimamoto, “Effective data gathering and energy efficient communication protocol
in wireless sensor networks employing UAV,” in Proc. IEEE WCNC, pp. 2342-2347, 2014.
[15] J. Lyu, Y. Zeng, and R. Zhang, “Cyclical multiple access in UAV-aided communications: a throughput-delay tradeoff,”
IEEE Wireless Commun. Lett., vol. 5, no. 6, pp. 600-603, Dec. 2016.
[16] “Cellular enables safer drone deployments,” [Online]. Available: https://www.qualcomm.com/invention/technologies/lte/advancedpro/cellular-drone-communication.
[17] N. H. Motlagh, M. Bagaa, and T. Taleb, “UAV-based IoT platform: a crowd surveillance use case,” IEEE Commun. Mag.,
vol. 55, no. 2, pp. 128-134, Feb. 2017.
[18] P. K. Gopala, L. Lai, and H. E. Gamal, “On the secrecy capacity of fading channels,” IEEE Trans. Inf. Theory, vol. 54,
no. 10, pp. 4687-4698, Oct. 2008.
[19] Y. Liang, H. V. Poor, and S. Shamai, “Secure communication over fading channels,” IEEE Trans. Inf. Theory, vol. 54, no.
6, pp. 2470-2492, Jun. 2008.
[20] X. Wang, M. Tao, J. Mo, and Y. Xu, “Power and subcarrier allocation for physical-layer security in OFDMA-based
broadband wireless networks,” IEEE Trans. Inf. Forensics Security, vol. 6, no. 3, pp. 693-702, Sep. 2011.
[21] H. Xing, L. Liu, and R. Zhang, “Secrecy wireless information and power transfer in fading wiretap channel,” IEEE Trans.
Veh. Technol., vol. 65, no. 1, pp. 180-190, Jan. 2016.
30
[22] A. Khisti and G. W. Wornell, “Secure transmission with multiple antennas–part II: the MIMOME wiretap channel,” IEEE
Trans. Inf. Theory, vol. 56, no. 11, pp. 5515-5532, Nov. 2010.
[23] Y. Wu, C. Xiao, Z. Ding, X. Gao, and S. Jin, “Linear precoding for finite-alphabet signaling over MIMOME wiretap
channels,” IEEE Trans. Veh. Technol., vol. 61, no. 6, pp. 2599-2612, Jul. 2012.
[24] G. Zheng, L. C. Choo, and K. K. Wong, “Optimal cooperative jamming to enhance physical layer security using relays,”
IEEE Trans. Signal Process., vol. 59, no. 3, pp. 1317-1322, Mar. 2011.
[25] Y. Zou, X. Wang, and W. Shen, “Physical-layer security with multiuser scheduling in cognitive radio networks,” IEEE
Trans. Commun., vol. 61, no. 12, pp. 5103-5113, Dec. 2013.
[26] M. Erdelj, E. Natalizio, K. R. Chowdhury, and I. F. Akyildiz, “Help from the sky: leveraging UAVs for disaster
management,” IEEE Pervasive Comput., vol. 16, no. 1, pp. 24-32, Jan. 2017.
[27] M. Caris, et al., “mm-Wave SAR demonstrator as a test bed for advanced solutions in microwave imaging,” IEEE Aerosp.
Electron. Syst. Mag., vol. 29, no. 7, pp. 8-15, Jul. 2014.
[28] F. Jiang and A. L. Swindlehurst, “Optimization of UAV heading for the ground-to-air uplink,” IEEE J. Sel. Areas Commun.,
vol. 30, no. 5, pp. 993-1005, Jun. 2012.
[29] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with trajectory optimization,” IEEE Trans. Wireless Commun.,
vol. 16, no. 6, pp. 3747-3760, Jun. 2017.
[30] X. Lin, et al., “The sky is not the limit: LTE for unmanned
aerial vehicles,” [Online]. Available:
https://arxiv.org/pdf/1707.07534.
[31] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.
| 7cs.IT
|
Deep Roots:
Improving CNN Efficiency with Hierarchical Filter Groups
Duncan Robertson2
Roberto Cipolla1
2
Antonio Criminisi
1
University of Cambridge, 2 Microsoft Research
arXiv:1605.06489v3 [cs.NE] 30 Nov 2016
Yani Ioannou1
Abstract
be achieved by weight decay or dropout [5]. Furthermore,
a carefully designed sparse network connection structure
can also have a regularizing effect. Convolutional Neural
Networks (CNNs) [6, 7] embody this idea, using a sparse
convolutional connection structure to exploit the locality of
natural image structure. In consequence, they are easier to
train.
We propose a new method for creating computationally efficient and compact convolutional neural networks
(CNNs) using a novel sparse connection structure that resembles a tree root. This allows a significant reduction in
computational cost and number of parameters compared
to state-of-the-art deep CNNs, without compromising accuracy, by exploiting the sparsity of inter-layer filter dependencies. We validate our approach by using it to train
more efficient variants of state-of-the-art CNN architectures, evaluated on the CIFAR10 and ILSVRC datasets. Our
results show similar or higher accuracy than the baseline
architectures with much less computation, as measured by
CPU and GPU timings. For example, for ResNet 50, our
model has 40% fewer parameters, 45% fewer floating point
operations, and is 31% (12%) faster on a CPU (GPU).
For the deeper ResNet 200 our model has 25% fewer floating point operations and 44% fewer parameters, while
maintaining state-of-the-art accuracy. For GoogLeNet, our
model has 7% fewer parameters and is 21% (16%) faster
on a CPU (GPU).
With few exceptions, state-of-the-art CNNs for image
recognition are largely monolithic, with each filter operating on the feature maps of all filters on a previous layer. Interestingly, this is in stark contrast to what we understand of
biological neural networks, where we see “highly evolved
arrangements of smaller, specialized networks which are interconnected in very specific ways” [8].
Recently, learning a low-rank basis for filters was found
to improve generalization while reducing the computational
complexity and model size of a CNN with only full rank filters [9]. However, this work addressed only the spatial extents of the convolutional filters (i.e. h and w in Fig. 1a).
In this work we will show that a similar idea can be applied to the channel extents – i.e. filter inter-connectivity –
by using filter groups [4]. We show that simple alterations
to state-of-the-art CNN architectures can drastically reduce
computational cost and model size without compromising
accuracy.
1. Introduction
This paper describes a new method for creating computationally efficient and compact convolutional neural networks (CNNs) using a novel sparse connection structure
that resembles a tree root. This allows a significant reduction in computational cost and number of parameters compared to state-of-the-art deep CNNs without compromising
accuracy.
It has been shown that a large proportion of the learned
weights in deep networks are redundant [1], a property that
has been widely exploited to make neural networks smaller
and more computationally efficient [2, 3]). It is unsurprising
then that regularization is a critical part of training such networks using large datasets [4]. Without regularization deep
networks are susceptible to over-fitting. Regularization may
2. Related Work
Most previous work on reducing the computational complexity of CNNs has focused on approximating convolutional filters in the spatial (as opposed to the channel) domain, either by using low-rank approximations [9–13], or
Fourier transform based convolution [14, 15]. More general
methods have used reduced precision number representations [16] or compression of previously trained models [17,
18]. Here we explore methods that reduce the computational
impact of the large number of filter channels within state-ofthe art networks. Specifically, we consider decreasing the
number of incoming connections to nodes.
1
stated in [4], is the fact that the AlexNet network has approximately 57% fewer connection weights than the corresponding network without filter groups. This is due to the
reduction in the input channel dimension of the grouped
convolution filters (see Fig. 2). Despite the large difference in the number of parameters between the models,
both achieve comparable accuracy on ILSVRC – in fact the
smaller grouped network gets ≈ 1% lower top-5 validation
error. This paper builds upon these findings and extends
them to state-of-the-art networks.
c2 filters
…
H
W
c1
c1
h1
ReLU
c2
HH2
w1
WW2
(a) Convolution.
c2 filters
H
W
c2ൗ ×
g
c1
h1
c1
w1 ൗg
× c2ൗg
ReLU
Low-dimensional Embeddings. Lin et al. [19] proposed
a method to reduce the dimensionality of convolutional feature maps. By using relatively cheap ‘1×1’ convolutional
layers (i.e. layers comprising d filters of size 1 × 1 × c,
where d < c), they learn to map feature maps into lowerdimensional spaces, i.e. to new feature maps with fewer
channels. Subsequent spatial filters operating on this lower
dimensional input space require significantly less computation. This method is used in most state of the art networks
for image classification to reduce computation [2, 20]. Our
method is complementary.
c2
HH2
WW2
(b) Convolution with filter groups.
Top-5 Val. Error
Figure 1: Filter Groups. (a) Convolutional filters (yellow)
typically have the same channel dimension c1 as the input
feature maps (gray) on which they operate. However, (b)
with filter grouping, g independent groups of c2 /g filters
operate on a fraction c1 /g of the input feature map channels,
reducing filter dimensions from h×w×c1 to h×w×c1 /g.
This change does not affect the dimensions of the input and
output feature maps but significantly reduces computational
complexity and the number of model parameters.
20% 4 groups
GoogLeNet. In contrast to much other work, Szegedy
et al. [2] propose a CNN architecture that is highly optimized for computational efficiency. GoogLeNet uses, as
a basic building block, a mixture of low-dimensional embeddings [19] and heterogeneously sized spatial filters –
collectively an ‘inception’ module. There are two distinct
forms of convolutional layers in the inception module, lowdimensional embeddings (1×1) and spatial (3×3, 5×5).
GoogLeNet keeps large, expensive spatial convolutions (i.e.
5×5) to a minimum by using few of these filters, using more
3×3 convolutions, and even more 1×1 filters. The motivation is that most of the convolutional filters respond to localized patterns in a small receptive field, with few requiring
a larger receptive field. The number of filters in each successive inception module increases slowly with decreasing
feature map size, in order to maintain computational performance. GoogLeNet is by far the most efficient state-of-theart network for ILSVRC, achieving near state-of-the-art accuracy with the lowest computation/model size. However,
we will show that even such an efficient and optimized network architecture benefits from our method.
no groups
19%
2 groups
18%
0.6
0.8
1
Model Parameters (# floats)
1.2
·109
Figure 2: AlexNet Filter Groups. Model Parameters vs.
top-5 error for variants of the AlexNet model on ILSVRC
image classification dataset. Models with moderate numbers of filter groups have far fewer parameters, yet surprisingly maintain comparable error.
AlexNet Filter Groups. Amongst the seminal contributions made by Krizhevsky et al. [4] is the use of ‘filter
groups’ in the convolutional layers of a CNN (see Fig. 1).
While their use of filter groups was necessitated by the practical need to sub-divide the work of training a large network
across multiple GPUs, the side effects are somewhat surprising. Specifically, the authors observe that independent
filter groups learn a separation of responsibility (colour features vs. texture features) that is consistent over different
random initializations. Also surprising, and not explicitly
Low-Rank Approximations. Various authors have suggested approximating learned convolutional filters using
tensor decomposition [11, 13, 18]. For example, Jaderberg
et al. [11] propose approximating the convolutional filters
in a trained network with representations that are low-rank
both in the spatial and the channel domains. This approach
significantly decreases computational complexity, albeit at
the expense of a small amount of accuracy. In this paper
2
m filters
c2 filters
d filters
m
W
w
1
m
H
W
c1
d
c1
h1
…
c
…
*
H
…
1
H
c
1
c
W
…
…
h
*
H
c3 filters
ReLU
c2
HH2
w1
c2
W
c3
WW2
1
W
1
H
ReLU
1
(a) Convolution with d filters of shape h × w × c.
Figure 3: Learning a (Spatial) Basis for Filters. Learning a linear combination of mostly small, heterogeneously
sized spatial filters [9]. Note that all filters operate on all c
channels of the input feature map.
× c2ൗg
c2 filters
W
c1
c2ൗ ×
g
h1
…
H
c3 filters
w1
c1ൗ
g
ReLU
c2
HH2
c2
H
ReLU
W
c3
WW2
1
1
(b) Root-2 Module: Convolution with d filters in g = 2 filter
groups, of shape h × w × c/2.
we are not approximating an existing model’s weights but
creating a new network architecture with explicit structural
sparsity, which is then trained from scratch.
c2 filters
× c2ൗg
c3 filters
× c2ൗg
H
W
c1
…
Learning a Basis for Filters Our approach is connected
with that of Ioannou et al. [9] who showed that replacing 3×3×c filters with linear combinations of filters with
smaller spatial extent (e.g. 1×3×c, 3×1×c filters, see
Fig. 3) could reduce the model size and computational complexity of state-of-the-art CNNs, while maintaining or even
increasing accuracy. However, that work did not address the
channel extent of the filters.
c2ൗ ×
g
c2ൗ ×
g
h1
ReLU
c2
HH2
c2
H
ReLU
W
c3
WW2
w1 c1ൗg
1
1
(c) Root-4 Module: Convolution with d filters in g = 4 filter
groups, of shape h × w × c/4.
Figure 4: Root Modules. Root modules (b), (c) compared
to a typical set of convolutional layers (a) found in ResNet
and other modern architectures. Grey blocks represent the
feature maps over which a layer’s filters operate, while colored blocks represent the filters of each layer.
3. Root Architectures
In this section we present the main contribution of our
work: the use of novel sparsely connected architectures resembling tree roots – to decrease computational complexity
and model size compared to state-of-the-art deep networks
for image recognition.
in the channel extent, since it operates on only a subset of
the channels of the input feature map.
This reduced connectivity also reduces computational
complexity and model size since the size of filters in filter groups are reduced drastically, as is evident in Fig. 4.
Unlike methods for increasing the efficiency of deep networks by approximating pre-trained existing networks (see
§2), our models are trained from random initialization using stochastic gradient descent. This means that our method
can also speed up training and, since we are not merely approximating an existing model’s weights, the accuracy of
the existing model is not an upper bound on accuracy of the
modified model.
Learning a Basis for Filter Dependencies It is unlikely
that every filter (or neuron) in a deep neural network needs
to depend on the output of all the filters in the previous layer.
In fact, reducing filter co-dependence in deep networks has
been shown to benefit generalization. For example, Hinton et al. [5] introduced dropout for regularization of deep
networks. When training a network layer with dropout, a
random subset of neurons is excluded from both the forward and backward pass for each mini-batch. Furthermore,
Cogswell et al. [21] observe a correlation between the covariance of hidden unit activations and overfitting. To explicitly reduce the covariance of hidden activations, they
train networks with a loss function, based on the covariance
matrix of the activations in a hidden layer.
Instead of using a modified loss, regularization penalty,
or randomized network connectivity during training to prevent co-adaption of features, we take a much more direct
approach. We use filter groups (see Fig. 1) to force the network to learn filters with only limited dependence on previous layers. Each of the filters in the filter groups is smaller
Root Module The basic element of our network architecture, a root module, is shown in Fig. 4. A root module has
a given number of filter groups, the more filter groups, the
fewer the number of connections to the previous layer’s outputs. Each spatial convolutional layer is followed by a lowdimensional embedding (1×1 convolution). Like in [9], this
configuration learns a linear combination of the basis filters
(filter groups), implicitly representing a filter of full channel
depth, but with limited filter dependence.
3
Table 1: Network-in-Network. Filter groups in each convolutional layer.
NiN
8.8%
Model
conv1
a
b
conv2
c
a
b
conv3
c
a
b
c
1
1
1
1
1
1
1
1
root-2
root-4
root-8
root-16
1
1
1
1
1
1
1
1
1
1
1
1
2
4
8
16
1
1
1
1
1
1
1
1
1
2
4
8
1
1
1
1
1
1
1
1
Error
1
Tree
Column
16
8.6%
5×5 1×1 1×1 5×5 1×1 1×1 3×3 1×1 1×1
Orig.
Root
·10−2
8.4%
16
4
8.2% 16 8
8
8.0%
42
2
8 4
2
7.8%
0.2
0.4
Table 2: Network-in-Network CIFAR10
0.6
0.8
Model Parameters
1.0
·106
·10−2
GPU
(ms)
Orig.
2.22
9.67
0.9211
39.0
0.623
root-2
root-4
root-8
root-16
1.64
1.23
1.03
0.93
7.37
4.55
3.15
2.45
0.9209
0.9202
0.9215
0.9167
31.2
27.6
24.4
23.0
0.551
0.480
0.482
0.475
8.8%
16
8.6%
Error
Model FLOPS Param. Accuracy CPU
×108
×105
(ms)
8.4%
4
16
8.2% 16 8
8.0%
8
8
4
4
2
2 2
7.8%
1.0
4. Results
Here we present image classification results obtained by
replacing spatial convolutional layers within existing stateof-the-art network architectures with root modules (described in §3) .
1.2 1.4 1.6 1.8 2.0
FLOPS (Multiply-Add)
2.2
·108
Figure 5: Network-in-Network CIFAR10 Results. Spatial filters (3×3, 5×5) are grouped hierarchically. The best
models are closest to the origin. For the standard network,
the mean and standard deviation (error bars) are shown over
5 different random initializations.
4.1. Improving Network in Network on CIFAR-10
Network in Network (NiN) [19] is a near state-of-theart network for CIFAR-10 [22]. It is composed of 3 spatial (5×5, 3×3) convolutional layers with a large number
of filters (192), interspersed with pairs of low-dimensional
embedding (1×1) layers. As a baseline, we replicated the
standard NiN network architecture as described by Lin et al.
[19] but used state-of-the-art training methods. We trained
using random 32×32 cropped and mirrored images from 4pixel zero-padded mean-subtracted images, as in [20, 23].
We also used the initialization of He et al. [24] and batch
normalization [25]. With this configuration, ZCA whitening was not required to reproduce validation accuracies obtained in [19]. We also did not use dropout, having found
it to have little effect, presumably due to our use of batch
normalization, as suggested by Ioffe and Szegedy [25].
To assess the efficacy of our method, we replaced the
spatial convolutional layers of the original NiN network
with root modules (as described in §3). We preserved the
original number of filters per layer but subdivided them into
groups as shown in Table 1. We considered the first of
the pair of existing 1×1 layers to be part of our root mod-
ules. We did not group filters in the first convolutional layer
– since it operates on the three-channel image space, it is
of limited computational impact compared to other layers.
Results are shown in Table 2 and Fig. 5 for various network architectures1 . Compared to the baseline architecture,
the root variants achieve a significant reduction in computation and model size without a significant reduction in accuracy. For example, the root-8 architecture gives equivalent accuracy with only 46% of the floating point operations (FLOPS), 33% of the model parameters of the original
network, and approximately 37% and 23% faster CPU and
GPU timings (see §5 for an explanation of the GPU timing
disparity).
Figure 6 shows the inter-layer correlation between the
adjacent filter layers conv2c and conv3a in the network
1 Here (and subsequently unless stated otherwise) timings are per image
for a forward pass computed on a large batch. Networks were implemented
using Caffe (with CuDNN and MKL) and run on an Nvidia Titan Z GPU
and 2 10-core Intel Xeon E5-2680 v2 CPUs.
4
192
0
0
conv3a
conv3a
192
Table 3: ResNet 50. Filter groups in each conv. layer.
conv2c
192
192
conv2c
Model conv1 res2{a–c} res3{a–d} res4{a–f} res5{a–c}
(b) Root-4: 2 filter groups
conv3a
3×3
1×1
3×3
1
1
1
1
1×1 3×3 1×1
1
1
1
3×3
1
root-2
root-4
root-8
root-16
root-32
root-64
1
1
1
1
1
1
1
1
1
1
1
1
2
4
8
16
32
64
1
1
1
1
1
1
1
2
4
8
16
32
1
1
1
1
1
1
1
1
2
4
8
16
1
1
1
1
1
1
1
1
1
2
4
8
included in §A.1). The results show that the so-called root
topology (illustrated in Fig. 7) gives the best performance,
providing the smallest reduction in accuracy for a given reduction in model size and computational complexity. Similar experiments with deeper network architectures have
delivered similar results and so we have reported results
for root topologies. This aligns with the intuition of deep
networks for image recognition subsuming the deformable
parts model. If we assume that filter responses identify parts
(or more elemental features), then there should be more filter dependence with depth, as more parts (filter responses)
are assembled into complex concepts.
192
192
(c) Root-8: 4 filter groups
1×1
1
192
conv2c
192
conv2c
0
conv3a
0
(a) Standard
7×7
Orig.
(d) Root-32: 16 filter groups
Figure 6: Inter-layer Filter Correlation. The blockdiagonal sparsity learned by a root-module is visible in the
correlation of filters on layers conv3a and conv2c in the
NiN network.
4.3. Improving Residual Networks on ILSVRC
Residual networks (ResNets) [20] are the state-of-the art
network for ILSVRC. ResNets are more computationally
efficient than the VGG architecture [26] on which they are
based, due to the use of low-dimensional embeddings [19].
ResNets are also more accurate and quicker to converge due
to the use of identity mappings.
architectures outlined in Table 1 as evaluated on the CIFAR
test set. The block-diagonalization enforced by the filter
group structure (as illustrated in Fig. 1) is visible, more so
with larger number of filter groups. This shows that the network learns an organization of filters such that the sparsely
distributed strong filter relations, visible in 6a as brighter
pixels, are grouped into a denser block-diagonal structure,
leaving a visibly darker, low-correlated background. See
§A.2 for more images, and an explanation of their derivation.
4.3.1
ResNet 50
As a baseline, we used the ‘ResNet 50’ model [20] (the
largest residual network model to fit onto 8 GPUs with
Caffe). ResNet 50 has 50 convolutional layers, of which
one-third are spatial convolutions (non-1×1). We did not
use any training augmentation aside from random cropping and mirroring. For training, we used the initialization
scheme described by [24] modified for compound layers [9]
and batch normalization [25]. To assess the efficacy of our
method, we replaced the spatial convolutional layers of the
original network with root modules (as described in §3). We
preserved the original number of filters per layer but subdivided them into groups as shown in Table 3. We considered
the first of the existing 1×1 layers subsequent to each spatial convolution to be part of our root modules.
Results are shown in Table 4 and Fig. 8 for various network architectures. Compared to the baseline architecture,
the root variants achieve a significant reduction in compu-
4.2. Grouping Degree with Network Depth
An interesting question concerns how the degree of
grouping in our root modules should be varied as a function of depth in the network. For the NiN-like architectures
described earlier, we might consider having the degree of
grouping: (1) decrease with depth after the first convolutional layer, e.g. 1–8–4 (‘root’); (2) remain constant with
depth after the first convolutional layer, e.g. 1–4–4 (‘column’); or (3) increase with depth, e.g. 1–4–8 (‘tree’).
To determine which approach is best, we created variants
of the NiN architecture with different degrees of grouping
per layer. Results are shown in Fig. 5 (numerical results are
5
ReLU
ReLU
ReLU
…
…
ReLU
…
ReLU
…
…
ReLU
…
W
…
…
…
ReLU
H
ReLU
ReLU
···
3
1
input image
conv1a
1
1
1
1
conv1c
conv1b
conv2a
1
1
1
1
conv2c
conv2b
1
1
conv3a
1
conv3c
conv3b
(a) Standard
ReLU
ReLU
ReLU
…
ReLU
…
ReLU
…
ReLU
…
W
…
…
…
ReLU
H
ReLU
ReLU
3
1
1
1
1
1
1
1
1
1
1
1
1
root-2 module
root-4 module
(b) Root-4 Architecture
Figure 7: Network-in-Network Root Architecture. The Root-4 architecture as compared to the original architecture for all
the convolutional layers. Colored blocks represent the filters of each layer. Here we don’t show the intermediate feature maps
over which a layer’s filters operate, or the final fully connected layer, out of space considerations (see Fig.4). The decreasing
degree of grouping in successive root modules means that our network architectures somewhat resemble plant roots, hence
the name root.
Table 4: ResNet 50 Results.
Model FLOPS Param. Top-1
Acc.
×107
×109
Table 5: ResNet-200 Results
Top-5
Acc.
CPU
(ms)
GPU
(ms)
Orig.
3.86
2.55
0.730
0.916
621
11.6
root-2
root-4
root-8
root-16
root-32
root-64
3.68
3.37
2.86
2.43
2.22
2.11
2.54
2.51
2.32
1.87
1.64
1.53
0.727
0.734
0.734
0.732
0.729
0.732
0.912
0.918
0.918
0.918
0.915
0.915
520
566
519
479
469
426
11.1
11.3
10.7
10.1
10.1
10.2
Model FLOPS ×1012 Param. ×107 Top-1 Err. Top-5 Err.
5.65
6.25
0.2196
0.0623
root-8
root-32
4.84
4.23
4.91
3.51
0.2205
0.2207
0.0626
0.0630
for ILSVRC 2012. To provide a baseline we used code implementing full training augmentation to achieve state-ofthe-art results2 . Table 5 shows the results of these experiments, top-1 and top-5 error are for center cropped images.
The models trained with roots have comparable error to the
baseline network, with fewer parameters and less computation. The root-32 model has 25% fewer FLOPS and 44%
fewer parameters than ResNet 200.
tation and model size without a significant reduction in accuracy. For example, the best result (root-16) exceeds the
baseline accuracy by 0.2% while reducing the model size by
27% and floating-point operations (multiply-add) by 37%.
CPU timings were 23% faster, while GPU timings were
13% faster. With a drop in accuracy of only 0.1% however, the root-64 model reduces the model size by 40%, and
reduces the floating point operations by 45%. CPU timings
were 31% faster, while GPU timings were 12% faster.
4.3.2
Orig.
4.4. Improving GoogLeNet on ILSVRC
We replicated the network as described by Szegedy et al.
[2], with the exception of not using any training augmentation aside from random crops and mirroring (as supported
by Caffe [27]). To train we used the initialization of [24]
modified for compound layers [9] and batch normalization
without the scale and bias [25]. At test time we only evaluate the center crop image.
ResNet 200
To show that the method applies to deeper architectures, we
also applied our method to ResNet 200, the deepest network
2 https://github.com/facebook/fb.resnet.torch
6
···
Top-5 Error
ResNet 50
Table 6: GoogLeNet Results.
All Filters
9%
7%
Model FLOPS Param. Top-1
×109
×107
Acc.
2
8% 64 32
16
1.6
4
8
1.8
2
2.2
2.4
Model Parameters (# Floats)
2.6
·107
Top-5 Error
(a) Model Parameters vs. Top-5 Error.
8% 64 32
GPU
(ms)
Orig.
1.72
1.88
0.694
0.894 315
4.39
root-2
root-4
root-8
root-16
1.54
1.29
0.96
0.76
1.88
1.85
1.75
1.63
0.695
0.693
0.691
0.683
0.893
0.892
0.891
0.886
4.37
4.10
3.72
3.59
285
273
246
207
Table 7: GoogLeNet. Filter groups in each convolutional
layer and Inception module (incp.)
2
16
4
8
Model conv1conv2 incp. 3{a,b} incp. 4{a–e} incp. 5{a,b}
2.5
3
3.5
FLOPS (Multiply-Add)
·109
(b) FLOPS (Multiply-Add) vs. Top-5 Error.
Top-5 Error
CPU
(ms)
9%
7%
9%
2
64
8% 3216
7%
10 10.2 10.4 10.6 10.8 11 11.2 11.4 11.6
GPU Forward (ms)
32
16
7%
450
1×1 3×3 5×5
1
1
1
1
1
1
1
1
1
1
1
1
root-2
root-4
root-8
root-16
1
1
1
1
1
1
1
1
2
4
8
16
1
1
1
1
1
2
4
8
1
2
4
8
1
1
1
1
1
1
2
4
1
1
2
4
1
1
1
1
1
1
1
2
1
1
1
2
ings – and model size, as measured in the number of floating
point parameters. For many of the configurations the top-5
accuracy remains within 0.5% of the baseline model. The
highest accuracy result, is 0.1% off the top-5 accuracy of
the baseline model, but has a 0.1% higher top-1 accuracy
– within the error bounds resulting from training with different random initializations. While maintaining the same
accuracy, this network has 9% faster CPU and GPU timings.
However, a model with only 0.3% lower top-5 accuracy
than the baseline has much higher gains in computational
efficiency – 44% fewer floating point operations (multiplyadd), 7% fewer model parameters, 21% faster CPU and
16% faster GPU timings.
While these results may seem modest compared to the
results for ResNet, GoogLeNet is by far the smallest and
fastest near state-of-the-art model ILSVRC model. We believe that more experimentation in using different cardinalities of filter grouping in the heterogeneously-sized filter
groups within each inception module will improve results
further.
9%
8% 64
7×7 1×1 3×3 1×1 3×3 5×5 1×1 3×3 5×5
Orig.
4
8
(c) GPU Forward Time vs. Top-5 Error.
Top-5 Error
Top-5
Acc.
2
8
4
500
550
CPU Forward (ms)
600
(d) CPU Forward Time vs. Top-5 Error.
Figure 8: ResNet-50 Results. Models with filter groups
have fewer parameters, and less floating point operations,
while maintaining error comparable to the baseline.
While preserving the original number of filters per layer,
we trained networks with various degrees of filter grouping,
as described in Table 7. While the inception architecture is
relatively complex, for simplicity, we always use the same
number of groups within each of the groups of different filter sizes, despite them having different cardinality. For all
of the networks, we only grouped filters within each of the
‘spatial’ convolutions (3×3, 5×5).
As shown in Table 6, and plotted in Fig. 9, our method
shows significant reduction in computational complexity –
as measured in FLOPS (multiply-adds), CPU and GPU tim-
5. GPU Implementation
Our experiments show that our method can achieve a significant reduction in CPU and GPU runtimes for state-ofthe-art CNNs without compromising accuracy. However,
the reductions in GPU runtime were smaller than might
have been expected based on theoretical predictions of computational complexity (FLOPs). We believe this is largely
a consequence of the optimization of Caffe for existing net7
Top-5 Error
12%
GoogLeNet
11% 16
8
10%
1.65
show it is possible to achieve high throughput with large
batches, by implementing a more efficient interface than
that used in the CuBLAS batched calls. We have modified
Caffe to use CuBLAS batched calls, and achieved significant speedups for our root-like network architectures compared to vanilla Caffe without CuDNN, e.g. a 25% speed
up on our root-16 modified version of the GoogleNet architecture. However, our optimized implementation still is not
as fast as Caffe with CuDNN (which was used to generate
the results in this paper), presumably because of other unrelated optimizations in the (proprietary) CuDNN library.
Therefore we suggest that direct integration of CuBLASstyle batching into CuDNN could improve the performance
of filter groups significantly.
All Filters
4
2
1.7
1.75
1.8
1.85
·107
Model Parameters (# Floats)
Top-5 Error
(a) Model Parameters vs. Top-5 Error.
12%
11% 16
8
10%
0.8
4
2
1
1.2
1.4
FLOPS (Multiply-Add)
1.6
·109
6. Future Work
Top-5 Error
(b) FLOPS (Multiply-Add) vs. Top-5 Error.
In this paper we focused on using homogeneous filter
groups (with a uniform division of filters in each group),
however this may not be optimal. Heterogeneous filter
groups may reflect better the filter co-dependencies found in
deep networks. Learning a combined spatial [9] and channel basis, may also improve efficiency further.
12%
11% 16
10%
8
3.6
4
2
3.8
4
4.2
GPU Forward (ms)
4.4
7. Conclusion
Top-5 Error
(c) GPU Forward Time vs. Top-5 Error.
We explored the effect of using complex hierarchical arrangements of filter groups in CNNs and show that imposing a structured decrease in the degree of filter grouping
with depth – a ‘root’ (inverse tree) topology – can allow
us to obtain more efficient variants of state-of-the-art networks without compromising accuracy. Our method appears to be complementary to existing methods, such as
low-dimensional embeddings, and can be used more efficiently to train deep networks than methods that only approximate a pre-trained model’s weights.
We validated our method by using it to create more
efficient variants of state-of-the-art Network-in-network,
GoogLeNet, and ResNet architectures, which were evaluated on the CIFAR10 and ILSVRC datasets. Our results
show similar accuracy with the baseline architecture with
fewer parameters and much less compute (as measured by
CPU and GPU timings). For Network-in-Network on CIFAR10, our model has 33% of the parameters of the original
network, and approximately 37% (23%) faster CPU (GPU)
timings. For ResNet 50, our model has 27% fewer parameters, and was 24% (11%) faster on a CPU (GPU). Even
for the most efficient of the near state-of-the-art ILSVRC
network, GoogLeNet, our model uses 7% fewer parameters
and is 21% (16%) faster on a CPU (GPU).
12%
11% 16
8
10%
220
4
2
240
260
280
CPU Forward (ms)
300
(d) CPU Forward Time vs. Top-5 Error.
Figure 9: GoogLeNet Results. Models with filter groups
have fewer parameters, and less floating point operations,
while maintaining error comparable to the baseline.
work architectures (particularly AlexNet and GoogLeNet)
that do not use a high degree of filter grouping.
Caffe presently parallelizes over filter groups by using
multiple CUDA streams to run multiple CuBLAS matrix
multiplications simultaneously. However, with a large degree of filter grouping, and hence more, smaller matrix multiplications, the overhead associated with calling CuBLAS
from the host can take approximately as long as the matrix
computation itself. To avoid this overhead, CuBLAS provides batched methods (e.g. cublasXgemmBatched),
where many small matrix multiplications can be batched together in one call. Jhurani and Mullowney [28] explore in
depth the problem of using GPUs to accelerate the multiplication of very small matrices (smaller than 16×16), and
References
[1]
8
M. Denil, B. Shakibi, L. Dinh, M. Ranzato, and N. de
Freitas, “Predicting Parameters in Deep Learning,”
in Neural Information Processing Systems (NIPS),
2013, pp. 2148–2156. arXiv:1306 . 0543 (cit. on
p. 1).
[2]
[12]
A. Sironi, B. Tekin, R. Rigamonti, V. Lepetit, and P.
Fua, “Learning separable filters,” IEEE Transactions
on Pattern Analysis and Machine Intelligence, vol.
37, no. 1, pp. 94–106, 2015, ISSN: 01628828 (cit. on
p. 1).
[13]
V. Lebedev, Y. Ganin, M. Rakhuba1, I. Oseledets, and V. Lempitsky, “Speeding-Up Convolutional Neural Networks Using Fine-tuned
CP-Decomposition,” International Conference on
Learning Representations (ICLR), vol. abs/1412.6,
pp. 1–10, 2015. arXiv:arXiv : 1412 . 6553v2
(cit. on pp. 1, 2).
[14]
M. Mathieu, M. Henaff, and Y LeCun, “Fast Training of Convolutional Networks through FFTs,” International Conference on Learning Representations
(ICLR), pp. 1–9, 2014. arXiv:arXiv : 1312 .
5851v5 (cit. on p. 1).
[15]
O. Rippel, J. Snoek, and R. P. Adams, “Spectral Representations for Convolutional Neural Networks,” Advances in Neural Information Processing
Systems 28, pp. 2440–2448, 2015, ISSN: 10495258.
arXiv:1506.03767 (cit. on p. 1).
[16]
S. Gupta, A. Agrawal, K. Gopalakrishnan, and P.
Narayanan, Deep Learning with Limited Numerical
Precision, 2015. arXiv:1502.02551 (cit. on p. 1).
[17]
W. Chen, J. T. Wilson, S. Tyree, K. Q. Weinberger,
and Y. Chen, “Compressing Neural Networks with
the Hashing Trick,” in Proceedings of The 32nd International Conference on Machine Learning, F. R.
Bach and D. M. Blei, Eds., ser. JMLR Proceedings, vol. 37, JMLR.org, 2015, pp. 2285–2294, ISBN:
9781510810587. arXiv:1504.04788 (cit. on p. 1).
[18]
Y.-D. Kim, E. Park, S. Yoo, T. Choi, L. Yang, and D.
Shin, “Compression of Deep Convolutional Neural
Networks for Fast and Low Power Mobile Applications,” in International Conference on Learning Representations (ICLR), 2016, pp. 1–16. arXiv:1511 .
06530 (cit. on pp. 1, 2).
[19]
F. Mamalet and C. Garcia, “Simplifying convnets
for fast learning,” in Artificial Neural Networks and
Machine Learning–ICANN 2012, Springer, 2012,
pp. 58–65 (cit. on p. 1).
M. Lin, Q. Chen, and S. Yan, “Network In Network,” arXiv preprint, vol. abs/1312.4, p. 10, 2013.
arXiv:1312.4400 (cit. on pp. 2, 4, 5).
[20]
M. Jaderberg, A. Vedaldi, and A. Zisserman, “Speeding up Convolutional Neural Networks with Low
Rank Expansions.,” in British Machine Vision Conference, 2014 (cit. on pp. 1, 2).
K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” Arxiv.Org,
vol. 7, no. 3, pp. 171–180, 2015, ISSN: 1664-1078.
arXiv:1512.03385 (cit. on pp. 2, 4, 5).
[21]
M. Cogswell, F. Ahmed, R. B. Girshick, L. Zitnick,
and D. Batra, “Reducing Overfitting in Deep Networks by Decorrelating Representations.,” in International Conference on Learning Representations,
2016 (cit. on p. 3).
C Szegedy, W Liu, Y Jia, and P Sermanet, “Going deeper with convolutions,” in arXiv preprint
arXiv: 1409.4842, 2014, ISBN: 9781467369640.
arXiv:1602.07360 (cit. on pp. 1, 2, 6).
[3]
E. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R.
Fergus, “Exploiting Linear Structure Within Convolutional Networks for Efficient Evaluation,” in arXiv,
2014, pp. 1–11. arXiv:1404.0736 (cit. on p. 1).
[4]
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” in Advances In Neural Information
Processing Systems, P. L. Bartlett, F. C. N. Pereira, C.
J. C. Burges, L. Bottou, and K. Q. Weinberger, Eds.,
2012, pp. 1–9, ISBN: 9781627480031. arXiv:1102.
0183 (cit. on pp. 1, 2).
[5]
[6]
G. E. Hinton, N. Srivastava, A. Krizhevsky, I.
Sutskever, and R. R. Salakhutdinov, Improving neural networks by preventing co-adaptation of feature
detectors, 2012. arXiv:1207 . 0580 (cit. on pp. 1,
3).
K Fukushima, “Neocognitron: A self-organizing
neural network model for a mechanish of pattern
recognition unaffected by shifts in position,” Biological Cybernetics, vol. 36, pp. 193–202, 1980 (cit. on
p. 1).
[7]
Y Lecun, L Bottou, Y Bengio, and P Haffner,
“Gradient-based learning applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11,
pp. 2278–2324, 1998, ISSN: 0018-9219 (cit. on p. 1).
[8]
M. Minsky and S. Papert, Perceptrons. MIT press,
1988 (cit. on p. 1).
[9]
Y. Ioannou, D. P. Robertson, J. Shotton, R. Cipolla,
and A. Criminisi, “Training CNNs with Low-Rank
Filters for Efficient Image Classification,” in International Conference on Learning Representations,
2016 (cit. on pp. 1, 3, 5, 6, 8).
[10]
[11]
9
[22]
A. Krizhevsky, “Learning Multiple Layers of Features from Tiny Images,” Univ. Toronto, Technical Report, 2009, pp. 1–60. arXiv:arXiv : 1011 .
1669v3 (cit. on pp. 4, 11).
[23]
I. J. Goodfellow, D. Warde-Farley, M. Mirza, A.
Courville, and Y. Bengio, “Maxout Networks,” in
Proceedings of the 30th International Conference on
Machine Learning (ICML), vol. 28, 2013, pp. 1319–
1327. arXiv:1302.4389 (cit. on p. 4).
[24]
K. He, X. Zhang, S. Ren, and J. Sun, “Delving
Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” in IEEE Conference on Computer Vision and Patern Recognition
(ICCV), IEEE, 2015, pp. 1026–1034, ISBN: 978-14673-8391-2. arXiv:1502.01852 (cit. on pp. 4–6).
[25]
S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift.,” in Proceedings of the 32
nd International Conference on Machine Learning,
Lille, France, 2015, 2015 (cit. on pp. 4–6).
[26]
K Simonyan and A Zisserman, “Very deep convolutional networks for large-scale image recognition,”
in eprint ar{X}iv:arXiv:1409.1556v5, 1409 (cit. on
p. 5).
[27]
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J.
Long, R. Girshick, S. Guadarrama, and T. Darrell,
“Caffe: Convolutional Architecture for Fast Feature Embedding,” ACM International Conference on
Multimedia, pp. 675–678, 2014, ISSN: 10636919.
arXiv:1408.5093 (cit. on p. 6).
[28]
C. Jhurani and P. Mullowney, “A GEMM interface
and implementation on NVIDIA GPUs for multiple
small matrices,” Journal of Parallel and Distributed
Computing, vol. 75, pp. 133–140, 2015 (cit. on p. 8).
10
192
conv3a
0
0
conv3a
192
A. Appendices
Table 8: Network-in-Network CIFAR10
GPU
(ms)
2.22
9.67
0.9211
39.0
0.623
root-2
root-4
root-8
root-16
1.64
1.23
1.03
0.93
7.37
4.55
3.15
2.45
0.9209
0.9202
0.9215
0.9167
31.2
27.6
24.4
23.0
0.551
0.480
0.482
0.475
tree-2
tree-4
tree-8
tree-16
1.48
1.15
0.99
0.91
4.88
3.31
2.53
2.14
0.9185
0.9147
0.9171
0.9168
31.4
29.1
25.7
20.6
0.541
0.535
0.500
0.512
col-2
col-4
col-8
col-16
1.53
1.18
1.01
0.92
5.71
3.73
2.73
2.24
0.9197
0.9200
0.9192
0.9120
28.8
26.1
23.0
22.8
0.568
0.536
0.475
0.494
192
Orig.
192
Model FLOPS Param. Accuracy CPU
×108
×105
(ms)
conv2c
conv2c
A.1. Full Network-in-Network Results
(a) Non-whitened responses
Figure 10: Covariance for between two layers in the root32 Network-in-Network model with and without whitened
responses
from the data (as shown in Fig. 10a). We can more clearly
show the covariance between layers by first whitening (using ZCA [22]) the samples in X1 and X2 . For a covariance
matrix,
1
XX T ,
N −1
The ZCA whitening transformation is given by,
cov(X, X) =
Table 8 shows the full results for the Network-inNetwork experiments on CIFAR10 on various hierarchical
network topologies.
W =
To show the relationships between filters between adjacent convolutional layers, as illustrated in Fig. 1 , we calculate the covariance of the responses from two adjacent featuremaps, the outputs of convolutional layers with c1 and c2
filters.
Let Xi = [xi,1 ; xi,2 ; . . . ; xi,N ] be the matrix of N samples xi,n from the ci dimensional featuremap for layer i. We
consider each pixel across the two featuremaps to be a sample, and thus each vector xi,n is a single pixel filter response
of dimension ci . If two featuremaps have different spatial dimensions, due to pooling, we up-sample the smaller
featuremap (with nearest neighbor interpolation) such that
there are the same number of pixels (and thus samples) in
each featuremap.
Given two samples X1 , X2 with zero mean (i.e. mean
subtracted) for two adjacent featuremaps, we calculate the
inter-layer covariance,
= E X1 X2T ,
1
X1 X2T .
=
N −1
√
N − 1 XX T
− 12
.
(3)
(4)
Since the covariance matrix is symmetric, it is easily diagonalizable (i.e. PCA),
A.2. Inter-Layer Covariance
cov(X1 , X2 )
(b) Whitened responses
cov(X, X)
=
=
1
XX T ,
N −1
1
P DP T ,
N −1
(5)
(6)
(7)
where P is a orthogonal matrix and D a diagonal matrix.
This diagonalization allows a simplified calculation of the
whitening transformation (see the derivation in Appendix A
of [22]),
W =
√
1
N − 1P D− 2 P T ,
(8)
− 12
where D is simply D with an element-wise power of − 21 .
The covariance between the whitened featuremap responses is then,
cov(W1 X1 , W2 X2 ) = E (W1 X1 ) (W2 X2 )T .
(9)
Figure 11 shows the per-layer (intra-layer) filter correlation. This shows the correlation of filters is more structured in root-networks, filters are learned to be linearly combined into useful filters by the root module, and thus filters
are often grouped together with other filters they correlate
strongly with.
(1)
(2)
While this shows the covariance between layers, it is
conflated with the inherent covariances within X1 and X2
11
0
192 conv3b
0
conv3b 192
0
conv3b 192
conv3a 192
0
192 conv3b
conv3a 192
conv3b 192
192 conv3b
0
0
conv3b 192
192 conv3b
conv3a 192
192 conv3a
0
192 conv3a
conv3a 192
192 conv3a
192 conv3a
conv2c 192
0
0
0
192 conv2c
conv2c 192
192 conv2c
192 conv2a
conv2c 192
192 conv2c
0
conv2b 192
0
(c) Root-8.
conv2a 192
0
192 conv2b
0
conv2b 192
192 conv2b
0
(b) Root-4.
conv2a 192
conv2c 192
192 conv2c
0
0
conv2b 192
192 conv2b
conv1c 96
conv2b 192
192 conv2b
0
192 conv2a
0
192 conv2a
conv2a 192
192 conv2a
96
0
conv1b 160
conv2a 192
(a) Network-in-Network.
0
0
conv1c
0
96
conv1c 96
conv1c
conv1b 160
0
96
conv1c
conv1c 96
conv1c
0
conv1c 96
96
0
160 conv1b
0
conv1b 160
160 conv1b
160 conv1b
conv1a 192
0
0
conv1a 192
192 conv1a
0
12
192 conv1a
conv1b 160
160 conv1b
conv1a 192
192 conv1a
0
192 conv1a
0
conv1a 192
(d) Root-32.
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Figure 11: Network-in-Network Intra-Layer Correlation. Absolute Correlation of filters within each layer of a NiN model variant.
192
conv2c
conv3a
192
conv3a
(c) Root-4
192
0
(b) Root-2
192
0
conv3a
0
0
192
192
192
(a) Standard
0
conv3a
conv2c
192
conv2c
0
conv3a
conv3a
192
conv2c
192
192
192
conv2c
conv2c
13
(d) Root-8
(e) Root-16
(f) Root-32
Figure 12: Filter Inter-layer Covariance conv2c–conv3a. The block-diagonal sparsity learned by a root-unit is visible in the correlation of filters on layers
conv3a and conv2c in the NiN network.
0
conv3b 192 10
0
192
192
conv3a
0
conv3b 192 10
192
conv3a
conv3b 192 10
0
192
192
conv3a 192
0
0
conv2c
conv3a 192
0
192
192
192
conv2c 192
0
0
conv2b 192
0
(b) Root-4.
conv1b 160
conv2c 192
conv2b
0
0
conv2a
conv2a 192
192
conv1c 0
96
0
conv1c
96
conv2b 192
conv3a 192
conv2c
0
0
conv2c 192
conv2b
0
0
conv2a
192
conv1c 0
96
conv2a 192
conv2b 192
(a) Network-in-Network.
160conv1b
192
0
conv1c
96
conv1b 160
conv1a
0
192
160conv1b
conv1a
0
conv1b 160
192
0
conv3b 192 10
192
conv3a
192
0
192
conv3a
conv2c
192
conv2c
192
conv3a 192
conv2b
conv2c 192
192
0
conv2a
conv2a 192
192
conv1c 0
96
0
conv1c
96
0
(c) Root-8.
conv2b 192
0
192
conv2b
conv2a
192
conv1c 0
96
conv2a 192
conv1b 160
160conv1b
192
conv1a
0
192
160conv1b
conv1a
0
14
conv1c
96
(d) Root-32.
Figure 13: Network-in-Network Inter-layer Absolute Covariance. The inter-layer covariance for all layers in variants of the NiN network.
·108
ResNet 50
Root-64
FLOPS
1.0
0.5
0.0
Parameters
·106
2.0
1.0
(a) Standard
(b) Root-2
conv1
res2a-1
res2a-2a
res2a-2b
res2a-2c
res2b-2a
res2b-2b
res2b-2c
res2c-2a
res2c-2b
res2c-2c
res3a-1
res3a-2a
res3a-2b
res3a-2c
res3b-2a
res3b-2b
res3b-2c
res3c-2a
res3c-2b
res3c-2c
res3d-2a
res3d-2b
res3d-2c
res4a-1
res4a-2a
res4a-2b
res4a-2c
res4b-2a
res4b-2b
res4b-2c
res4c-2a
res4c-2b
res4c-2c
res4d-2a
res4d-2b
res4d-2c
res4e-2a
res4e-2b
res4e-2c
res4f-2a
res4f-2b
res4f-2c
res5a-1
res5a-2a
res5a-2b
res5a-2c
res5b-2a
res5b-2b
res5b-2c
res5c-2a
res5c-2b
res5c-2c
fc1000
0.0
Figure 15: ResNet 50 Layer-wise FLOPS/Parameters.
(c) Root-4
(d) Root-8
(e) Root-16
(f) Root-32
these networks are very similar to those learned in the original model, although sometimes inverted or with a different
ordering. This ordering is somewhat consistent in models
with filter groups however, even with different random initializations. This is because filter groups cause filters with
strong mutual information to be grouped adjacent to each
other.
For example, in the root-8 network (Fig. 14d), each row
of filters corresponds to the input of an independent filter
group in conv2. We can see that the first row primarily
is composed of filters giving various directions of the same
color gradient. These filters can be combined in the next
layer to produce color edges easily. Due to the shortcut
layer and the learned combinations of filters however, not
all filter groupings are so obvious.
Figure 14: ResNet 50 conv1 filters. With filter groups
directly after conv1, in conv2, some of the organization
of filters can be directly observed, and give us intuition as
to what is happening in root networks.
A.4. Layer-wise Compute/Parameter Savings
Figure 15 shows the difference in compute and parameters for each layer in a standard ResNet-50 model and a
root-64 variant. The layers in the original networks with
the highest computational complexity are clearly the spatial convolutional layers, i.e. layers with 3×3 spatial filters. When instead a root-module is used, the computational complexity of these layers is reduced dramatically.
While the low dimensional embedding layers (1×1) are not
changed, these have less than half the compute of the spatial convolution layers. The number of parameters in spatial
convolution layers with large numbers of input channels,
which increase towards the end of the network, are similarly reduced.
Figure 12 shows the complete, enlarged version of Fig. 6,
showing the inter-layer filter covariances between layers
conv3a and conv2c. Figure 13 shows the full set of interlayer covariances between all convolutional layers in the
NiN models. Block-diagonal sparsity is visible on the layers with filter groups, conv2a and conv3a. This blockdiagonal is shown for all variants in more detail in Fig. 13.
A.3. The Affect on Image-level Filters of Root Modules
In the ResNet root models, filter groups are used in
conv2, directly after the image level filters of conv1 some
of the organization of filters can be directly observed, and
give us intuition as to what is happening in root networks.
Figure 14 shows the conv0 filters learned for each of the
ResNet 50 models. It is apparent that the filters learned in
15
| 9cs.NE
|
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
1
Attack vulnerability of power systems under an
equal load redistribution model
arXiv:1804.01351v1 [physics.soc-ph] 4 Apr 2018
Talha Cihad Gulcu, Vaggos Chatziafratis, Yingrui Zhang, and Osman Yağan
Abstract—This paper studies the vulnerability of flow networks
against adversarial attacks. In particular, consider a power
system (or, any system carrying a physical flow) consisting of
N transmission lines with initial loads L1, . . . , L N and capacities
C1, . . . , C N , respectively; the capacity Ci defines the maximum
flow allowed on line i. Under an equal load redistribution model,
where load of failed lines is redistributed equally among all
remaining lines, we study the optimization problem of finding
the best k lines to attack so as to minimize the number of
alive lines at the steady-state (i.e., when cascades stop). This
is done to reveal the worst-case attack vulnerability of the
system as well as to reveal its most vulnerable lines. We derive
optimal attack strategies in several special cases of load-capacity
distributions that are practically relevant. We then consider
a modified optimization problem where the adversary is also
constrained by the total load (in addition to the number) of the
initial attack set, and prove that this problem is NP-Hard. Finally,
we develop heuristic algorithms for selecting the attack set for
both the original and modified problems. Through extensive
simulations, we show that these heuristics outperform benchmark
algorithms under a wide range of settings.
Index Terms—Flow networks; cascading failures; vulnerability; optimal attack strategies
I. I NTRODUCTION
Networks that carry or transport physical commodities,
e.g., electricity/water/gas distribution networks and road/public
transportation networks, have been an integral part of our
daily lives for decades. For example, our quality of life
largely depends on the continuous availability of an electrical
power supply. This dependence is expected to be further
amplified in the near future due to the increasing market
share of electric vehicles and increasing integration of major
national infrastructures to the power grid; e.g., water, transport,
communications, etc. All of these point to a future where
the reliability of the flow networks will be paramount, with
the central research question being how we can design these
networks in a robust and reliable manner. In the rest of
the paper, we focus our attention on power systems for
concreteness, but most of the discussion applies generally to
any transport system.
This research was supported by the National Science Foundation through
grant CCF #1422165 and a generous gift from Persistent Systems, Inc.
Preliminary version of some of the material was presented at the Information Theory and Applications Workshop (ITA 2016), San Diego, CA [4].
T. C. Gulcu is with TUBITAK Software Technologies Research Institute,
06100, Cukurambar, Ankara, Turkey. E-mail: [email protected]
V. Chatziafratis is with the Computer Science Department, Stanford University, Palo Alto, CA 94305. Email:[email protected]
Y. Zhang and O. Yağan are with the Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15213. Email:{yingruiz, oyagan}@andrew.cmu.edu
A major problem with existing power systems is the seemingly unexpected large scale failures. Although rare, the sheer
size of such failures has proven to be very costly, at times
affecting hundreds of millions of people [2], [21], [28]; e.g.,
the recent blackout in India [34]. Such events are often
attributed to a small initial shock getting escalated due to
intricate dependencies within a power system [3], [27], [11].
This phenomenon, also known as cascade of failures, has the
potential of collapsing an entire power system as well as other
infrastructures that depend on the power grid [8], [16], [29].
Thus, understanding the dynamics of failures in power systems
and mitigating the potential risks are critical for the successful
development and evolution of many critical infrastructures.
In this paper, we continue our study of the robustness
of power systems under a simple model based on equal
redistribution of flow upon the failure of a power line. Namely,
we consider a power system with N transmission lines with
initial loads L1, . . . , L N and capacities C1, . . . , CN . If a line
fails (for any reason), its load is assumed to be redistributed
equally among all lines that are alive. Thus, the load carried
by a line i may exceed its initial value Li over time due to
load redistribution. The capacity Ci defines the maximum flow
allowed on the line i, meaning that if the load carried by i
exceeds this capacity at any time, the line will be tripped (i.e.,
disconnected) by means of automatic protective equipments
so as to avoid costly damages to the system. Subsequently,
the load that was carried by line i before failure will be
redistributed to remaining lines, which in turn may cause
further failures, possibly leading to a cascade of failures.
The equal load redistribution model gets its appeal from
its ability to capture the long-range [7], [19] nature of the
Kirchhoff’s law, at least in the mean-field sense, as opposed
to the topological models [5], [14], [26] where failed load
is redistributed only locally among neighboring lines. For
example, it was suggested by Pahwa et al. [19] that equal load
redistribution is a reasonable assumption1 especially under the
DC power flow model; the DC model is known [17], [24] to
approximate the AC model well in many cases. With these
in mind, an important goal is to understand the robustness of
systems under the equal load redistribution model described
above against random and targeted attacks. The former case
was studied by Yağan [29] under the assumptions that initial
loads L1, . . . , L N are independent and identically distributed
1 The equal load redistribution model is also similar in spirit to the wellstudied CASCADE model introduced in [8], [9]. There, upon failure of a line,
a fixed amount ∆ of load is redistributed to all functional lines irrespective
of the load being carried by the failed line or the number of remaining lines.
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
with PL (x) = P [L ≤ x] and that capacities are given by
Ci = (1 + α)Li,
i = 1, . . . , N,
where α > 0 denotes the tolerance factor; in [29] all lines assumed to have the same tolerance factor. There, the robustness
of the system against random attacks that target a p-fraction
of the lines was studied; system robustness was quantified by
the final (i.e., steady-state) fraction n∞ (p) of non-failed lines.
Among other results, it was shown that the system robustness
is maximized if all lines are given the same initial load, for a
given fixed mean load E [L].
Recently, Zhang and Yağan [36] extended the results in
[29] to the more general case where lines can have varying
tolerance parameters. Namely, they let
Ci = Li + Si,
i = 1, . . . , N,
with Si denoting the free-space (or, redundancy) available at
line i. The tolerance factor, given by αi = Si /Li , may now vary
from one line to another. Under the assumption that load-‘free
space’ pairs (Li, Si ) are independent and identically distributed
with PLS (x, y) = P [L ≤ x, S ≤ y], they studied the robustness
against random attacks that target a p-fraction of the lines.
Their main result is that, with the mean values E [L] and
E [S] are fixed, n∞ (p) is uniformly maximized for all p values
if all nodes are given the same free space E [S], regardless of
how the loads are distributed. This leads to the counterintuitive
conclusion that lines with higher initial load shall be assigned
smaller tolerance factors to maximize robustness and raises
the possibility that the widely used (both in academy and
industry) setting of equal tolerance factor is not optimal for
system robustness (as far as the metric n∞ (p) is concerned).
With the case of random attacks being well-understood, we
shift our attention in this paper to understanding the vulnerability of power systems under targeted attacks. As before,
the main goal would be to derive design strategies (in the
form of optimal load-‘free space’ distributions) that would lead
to maximum robustness, this time against a knowledgeable
adversary attacking a carefully selected set of lines. However,
for this optimization problem to be well-defined one has to
have a good understanding of the problem from an adversary’s
perspective. With this in mind, this paper aims to develop
good attack strategies that lead to maximal damage to the
system for a given number of lines that can be attacked. The
solution to the optimal attack problem will also help a system
designer by i) revealing the worst-case attack vulnerability of
the system which can help evaluate a given system design; and
ii) revealing the most vulnerable lines in the system that will
potentially be targeted by adversaries; this may then provide
useful design guidelines for improving system robustness.
Formally, we consider the following optimization problem. Given N lines with loads L1, . . . , L N and free spaces
S1, . . . , SN , we seek to find the optimal set A of k lines that
the adversary should attack in order to minimize the final
number n∞ (A) of alive lines. We provide optimal solutions
via greedy algorithms in three special cases: i) when all lines
have the same load; ii) when Si = αLi for each i = 1, . . . , N
(as commonly used in the literature [5], [11], [13], [14], [26],
[29]); and iii) when all lines have the same free space, i.e.,
2
when S1 = · · · = SN . The last case is of particular interest as
it is known to lead to maximum robustness against random
attacks. For general cases of load-‘free-space’ distribution
PLS , we develop heuristic attack algorithms and show via
simulations that these outperform benchmark attack strategies.
In particular, we show that attacking the lines with the largest
Load × Free-space (L ∗ S) is in general a good strategy.
Parametrized attack strategies targeting the lines with the
largest L ∗ S β are also considered where β is in [0, ∞], and
optimum choices for β are provided in several cases.
We also consider a variation of the problem with an additional constraint on the total load of the lines attacked;
i.e.,
Í
when the adversary is further constrained with i ∈ A Li ≤ Q
for some Q. From a practical point of view, this might be
the case if high-load carrying lines are protected better by the
network owner and the cost of attacking them is proportional
to their load. We show that this variation of the optimal attack
problem is in fact NP-Hard, meaning that no polynomialtime algorithm can find the set A that minimizes n∞ (A),
unless P ≡ N P. Our proof is based on a polynomial time
reduction from the k-Subset Sum problem, i.e., the problem
that seeks to find whether a sequence of integers has a subset
of size k whose sum equals Q. For the modified optimization
problem, we develop several heuristic algorithms and evaluate
their performance in comparison with benchmarks through
an extensive simulation study. In particular, we modify the
previously developed heuristics with a switch that, when
actuated during a sequential selection of lines to be attacked,
changes the way algorithm makes the remaining selections;
this idea is inspired from heuristics developed in [12] for
the multi-dimensional 0-1 Knapsack problem. Among other
things, we demonstrate via simulations that the max-L ∗ S
algorithm with a switch performs well in a range of settings.
We believe that our results will shed some light on the
vulnerability of power systems against powerful adversaries
launching targeted attacks. In particular, we expect them to
capture the qualitative properties of a power system well. This
work may also have applications in other fields. In particular,
the model considered here might be relevant for any flow
network that is responsible for carrying (or, transporting) a
physical quantity. A particularly interesting application is the
study of traffic jams in roads [20], where the capacity of a
line is regarded as the traffic flow capacity of a road.
The rest of the paper is organized as follows. We describe
the system model in Section II. In Section III, we survey the
results in [36] concerning robustness against random attacks.
We start our discussion on optimal attack strategies in Section
IV by demonstrating that certain greedy algorithms fail to give
the optimal solution in general. In Section V we consider some
special cases of interest where greedy algorithms are shown
to find optimal attack sets. In Section VI we prove a hardness
result showing that a variation of the optimal attack problem
is NP-Hard. Finally, in Section VII, we develop heuristic
algorithms for both the original and the modified optimization
problems and present a detailed numerical study comparing
the performance of our heuristics with benchmark algorithms
such as max-S, max-L, max-C, and random attacks.
The random variables (rvs) under consideration are defined
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
3
on the probability space (Ω, F , P), where P denotes the probability measure. We denote the expectation operator by E. The
indicator function of an event E is denoted by 1 [E]. We write
| A| for the cardinality of a discrete set A.
II. S YSTEM M ODEL
We consider a power system with N transmission lines
L1, . . . , L N with initial loads (i.e., power flows) L1, . . . , L N .
The capacity Ci of a line Li defines the maximum power flow
that it can sustain, and is given by
Ci = Li + Si,
i = 1, . . . , N,
(1)
where Si denotes the free-space (or, redundancy) available to
line Li . The capacity of a line is typically written [14], [26],
[13], [5] as a factor of the line’s original load, i.e.,
Ci = (1 + αi )Li
with αi > 0 defining the tolerance parameter for line Li .
Put differently, the free space Si is often given in terms of
the initial load Li with Si = αi Li . It is assumed that a line
fails (i.e., outages) if its load exceeds its capacity at any given
time. In that case, the load it was carrying before the failure
is redistributed equally among all remaining lines.
The main question of interest is to characterize the robustness of this power system against i) random attacks that result
initially with a failure of a (randomly selected) p-fraction of
the lines; and ii) targeted attacks that initially fail a specific set
A of lines. The initial set of failures lead to redistribution of
power flows from the failed lines to alive ones (i.e., non-failed
lines), so that the load on each alive line becomes equal to its
initial load plus its equal share of the total load of the failed
lines. This may lead to the failure of some additional lines due
to the updated flow exceeding their capacity. This process may
continue recursively, generating a cascade of failures, with
each failure further increasing the load on the alive lines, and
may eventually result with the collapse of the entire system.
Throughout, let n∞ (p) denote the final fraction of alive
lines when a p-fraction of lines is randomly attacked. The
robustness of the power system can be evaluated [29], [36]
by the behavior of n∞ (p) as the attack size p increases, and
particularly by the critical attack size p? at which n∞ (p) drops
to zero. In the case where a specific set A of lines are attacked,
we define n∞ (A) as the final number of alive lines. The main
goal of this paper is to seek optimal attack strategies, i.e., to
find the set A of lines that minimizes n∞ (A) under certain
constraints; e.g., the size | A| being fixed.
The equal flow redistribution model is inspired by the
democratic fiber bundle model [1], [6], where N parallel fibers
with random failure thresholds C1, . . . , CN (i.e., capacities)
drawn independently from PC (x) share equally an applied total
force of F; see also [20], [22]. This model has been recently
adopted by Pahwa et al. [19] in the context of power systems
with F corresponding to the total load that N power lines
share equally. The formulation considered here, introduced in
[29], [36], differs from the original democratic fiber-bundle
model in that i) we do not assume that the total load of the
system is fixed at F; and ii) we allow for power lines to
carry different initial loads; see also [18]. In addition, [19] is
concerned with failures in the power system that are triggered
by increasing the total force (i.e., load) applied. Instead, our
formulation allows analyzing the robustness of the system
against external attacks or random line failures, which are
known to be the source of system-wide blackouts in many
interdependent systems [3], [32], [35].
III. R ELEVANT W ORK : E VALUATING AND O PTIMIZING
ROBUSTNESS AGAINST R ANDOM ATTACKS
We now survey the results obtained by Zhang and Yağan
[36], [29] on the robustness of power systems under equal load
redistribution of loads. These works consider the problem from
a defender’s perspective and provide means to characterize
and optimize the robustness of the system, assuming that the
adversary will launch a random attack to a certain fraction
of lines. With the randomness involved in the attack model,
as well as load-capacity values, [29], [36] rely on mean-field
analysis to characterize the mean performance of the system
in the asymptotic regime where N approaches infinity.
Throughout this section, assume that the pairs (Li, Si ) are
independently and identically distributed with PLS (x, y) :=
P [L ≤ x, S ≤ y] for each i = 1, . . . , N. The corresponding
∂2
PLS (x, y).
probability density is given by p LS (x, y) = ∂x∂y
We assume that L ≥ Lmin > 0 and S > Smin > 0 with
probability one, and that the marginal densities p L (x) and
pS (y) are continuous on their support.
A. Final system size as a function of the attack size
The main result in [36] characterizes the robustness of
power systems under any distribution of initial load L and
free space S, and any attack size p.
Theorem 3.1 ([36]): Let L and S denote generic random
variables following the same distribution with initial loads
L1, . . . , L N , and free space S1, . . . , SN , respectively. Then, with
x? denoting the smallest solution of
P [S > x] (x + E [L | S > x]) ≥
E [L]
1−p
over x ∈ (0, ∞], the final system size n∞ (p) is given by
n∞ (p) = (1 − p)P S > x? .
(2)
(3)
An intuitive explanation of this result is that x?, defined as
the smallest solution of (2), gives the extra load each alive line
has to carry (in addition to their initial load) at the steady-state;
(3) is then understood easily since a line will still be alive at
the steady-state if i) it is not targeted initially; and ii) it has
enough free-space to handle the extra x? amount of load.
For a graphical solution of n∞ (p), one shall plot
P [S > x] (x + E [L | S > x]) as a function of x, and draw a
horizontal line at E [L] /(1 − p) on the same plot. The leftmost
intersection of these two lines gives the operating point x?,
from which we can compute n∞ (p) = (1 − p)P [S > x?]. When
there is no intersection, we set x? = ∞ and understand that
n∞ (p) = 0; e.g., see Figure 1.
Theorem 3.1 enables computing n∞ (p) as a function of
p, thereby helping understand the response of the system to
attacks of varying magnitude. It also enables computing the
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
4
1
Final system size, n∞ (p)
0.9
0.8
1−p
0.7
Diverging
failure rate
0.6
0.5
0.4
0.3
Abrupt
rupture
0.2
0.1
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Attack size, p
Fig. 1. We see the variation of final system size n∞ (p) as a function of the
(random) attack size p when loads are drawn independently and uniformly from
the range [Lmin, Lmax ] = [10, 50], and free spaces are given by Si = αLi .
Red (lower) curve stands for α = 0.2, whereas blue (upper) curve represents
α = 1.2. For α = 0.2, we observe an abrupt first-order transition of n∞ (p) as
it suddenly jumps to zero at p? = 0.0625, while decaying linearly as 1 − p
up until that point. A slightly different behavior is seen when α = 1.2 in that
n∞ (p) exhibits a continuous divergence from the 1 − p line before complete
system failure, which again takes place through a discontinuous transition.
critical attack size p?, defined as the largest attack the system
can sustain. More precisely, we define
p? = sup n∞ (p) > 0,
p>0
so that for any attack with size p > p?, the system undergoes
a complete breakdown leading to n∞ (p) = 0. The next result
provides a closed form expression for the critical attack size.
Theorem 3.2 ([36]): The critical attack size p? is given by
p? = 1 −
E [L]
.
max{P [S > x] (x + E [L | S > x])}
(4)
x
The operational usefulness of p? derives in part from its ability
to quantify the system robustness by a single scalar, paving
the way to designing the system with optimal robustness by
means of maximizing p?; see Section III-B below.
An interesting question here is whether n∞ (p) decays to
zero continuously (i.e., through a second-order transition),
or discontinuously (i.e., through a first-order transition). The
practical significance of this is that continuous transitions
suggest a more stable and predictable system behavior with
respect to attacks, whereas with discontinuous transitions,
system behavior becomes more difficult to predict based on
past data. In [36], it was shown that the total breakdown of the
system will always be through a first-order (i.e., discontinuous)
transition under random attacks. More precisely, it was shown
that n∞ (p?) > 0, while by definition it holds that n∞ (p? + ε) =
0 for any ε > 0 arbitrarily small. This shows that despite its
simplicity, the equal flow redistribution model can capture realworld phenomena of unexpected large-scale system collapses;
i.e., cases where seemingly identical attacks/failures leading
to entirely different consequences. In addition, depending
on system parameters, the first-order breakdown may have
early indicators at smaller attack sizes such as a diverging
failure rate leading to a non-linear decrease in n∞ (p). In other
cases, the system abruptly transitions to a total collapse while
perfectly resisting attacks until the critical point; see Figure
1 and [36, Fig. 2] for the rich transition behavior that can be
observed under the equal load redistribution model.
B. Achieving Optimal Robustness
An important question from a system designer’s perspective
is concerned with deriving the universally optimum distribution of initial loads L1, . . . L N and free spaces S1, . . . , SN when
the mean values E [L] and E [S], respectively, are fixed. The
results obtained in [36] concerning this are presented next.
Theorem 3.3 ([36]): For any initial load-free space distribution, it holds that
E [S]
E [S]
=
= p?dirac
(5)
p? ≤
E [S] + E [L] E [C]
where p?dirac denotes the critical attack size when the freespace follows a Dirac-delta distribution, i.e., PLS (x, y) =
PL (x)1 [y ≤ E [S]], for any PL (x).
In words, this result states that the critical attack size can
never be larger than the ratio of mean free space to mean
capacity. In addition, this optimal p? value can be achieved
by assigning every line the same free-space, i.e., by setting
S1 = · · · = SN = E [S], regardless of how loads are distributed.
Thus far, we have seen that the equal distribution of free
space leads to the largest possible critical attack size, hereafter
denoted p?opt . It is clear that the final system size after an
attack of size p is at most 1 − p. With this in mind, the next
result establishes the optimality of the Dirac-delta distribution
of free-space in the sense of maximizing the robustness of
power systems uniformly over all attack sizes.
Theorem 3.4 ([36]): With S1 = S2 = · · · = SN we have
1 − p for p < p?opt
n∞ (p) =
0
for p ≥ p?opt
Hence, the Dirac-delta distribution of free-space S maximizes
n∞ (p) over the entire range 0 ≤ p ≤ 1.
These findings suggest that under the equal flowredistribution model considered here, power systems with
homogeneous distribution of redundant space are more robust
against random attacks and failures, as compared to systems
with heterogeneous distribution of redundancy. Interestingly,
this suggests that the optimal robustness is achieved when
the tolerance factor αi = Si /Li decreases with increasing
load, leading to the counter-intuitive conclusion that the lines
carrying the highest load should have the smallest tolerance
factor to achieve maximum robustness; see [36] for a more
detailed discussion on this matter.
IV. T HE OPTIMAL ATTACK PROBLEM AND INSIGHTS
While existing results in [36], [29] shed some light on the
robustness under the equal flow redistribution model, many
real-world threats to power systems would be expected to
come from powerful adversaries targeting specific parts of
the grid to inflict maximum damage on the overall system.
In order to understand the system’s robustness against such
targeted attacks and to reveal its most vulnerable lines, we
now consider the problem from an adversary’s perspective
and seek effective strategies for attacking the system under
given constraints. In particular, we consider a scenario where
the adversary has full information about the system and aims
to find the best set of k of lines to attack so that it fails
maximum number of lines as a result cascading failures. This
optimization problem is formally introduced next.
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
Fig. 2.
In this example we have (load, capacity) values given by (8, 8 +
), (6, 8 + ), (4, 4 + 14/3 + ), (2, 11 + ), (1, 21 + ) where > 0 is arbitrarily
small. The greedy maximum-load attack will need to attack k = 5 containers
to fail all. It will start attacking the leftmost container with load L1 = 8 which
will not lead to any further failures. Then, it will continue with the second one
from the left, again unable to trigger a cascade, and continue until attacking all
containers directly. The optimal solution can be seen to be k = 1 by attacking
the last container, which will trigger a cascading failure destroying the whole
system. We can generalize this counterexample to the case with N containers
with the greedy algorithm’s output being k = N while the optimal solution
being k = 1.
A. The Main Optimization Problem: ER-k
The Equal Redistribution (ER) problem with k attacks is
the optimization problem, denoted ER-k, that aims to find the
set A of k lines such that attacking A leads to the maximum
number of total line failures (as a result of load redistribution
and cascading failures), among all possible attack sets with
size k. Put differently, we seek to find A with | A| = k that
minimizes n∞ (A). Throughout, we find it useful to consider
the decision version of this optimization problem (referred to
as the ER-k-k 0 problem) formally defined as follows.
INPUT: N pairs of non-negative numbers in the form
(Li, Ci ) indicating the load and the capacity of each line, and
integers k and k 0 such that 0 < k < k 0 ≤ N. We assume
Ci > Li so that no line fails initially at its own load.
OUTPUT: The answer to whether or not there is an attack
set A with size k, such that at the end of the cascading failures
the number of failed nodes is at least k 0; i.e., whether there
exists A with | A| = k and n∞ (A) ≤ N − k 0.
B. Heuristic Algorithms that Fail
Here we will present three intuitive greedy algorithms and
give concrete examples demonstrating their poor performance
for the optimization problem described above. In doing so, we
will focus on the special case where k 0 = N meaning that the
goal of the attack is to destroy the whole system, by attacking
a minimum number k of lines.
In what follows, we find it useful to describe the problem
in a simpler way, where we have N water containers with
capacities C1, . . . , CN , and initial water levels L1, . . . , L N . As
in the equal flow-redistribution model, when a container is
"attacked" its content is redistributed equally to the remaining
containers. Also, if the water level in a container exceeds
its capacity, we assume that it has failed and redistribute its
content, again equally, to the remaining containers. With this
formulation, the goal of the attackers is to find the smallest
number k of containers to target so that all containers get
overloaded and fail eventually.
An important observation is that the following intuitive
algorithms can deviate significantly from the optimal solution.
5
Fig. 3. Consider 2n + 1 containers where (load, capacity) values are given by
(, M) for the first n containers and (M −2, M − ) for the last n+1 containers;
here > 0 is arbitrarily small and M > 2(n + 1) . The greedy max-capacity
attack will need to attack k = n + 1 containers to fail the all containers; it will
start attacking the first n containers but cascading failures will not take place.
On the other hand, the optimal solution is k = 1 as it takes to attack only one of
the containers with (M − 2, M − ) to trigger a cascading failure that will fail
all.
a) Greedy max-load attack: This greedy strategy aims
to maximize the load that will be redistributed in each
attack round. Namely, it starts by attacking the container
with the highest load, and proceeds similarly, waiting after
every attack for a steady-state to be reached (meaning that
all load redistribution and potential further failures end). The
algorithm stops when all containers fail. This strategy is not
optimal in general because it fails to recognize the opportunity
to eliminate containers with very large capacities that will
otherwise be difficult to fail by redistributing the load. The
worst-case deviation from the optimal (in terms of the number
of lines needed to be attacked for complete system failure) is
Θ(n); e.g., see Figure 2.
b) Greedy max-capacity attack: This strategy is similar
in spirit with the greedy max-load attack except that this time
the container with the maximum capacity is attacked in each
round. The idea here is that by taking out large containers, the
remaining, supposedly small, containers will be destroyed due
to load redistribution. This strategy is not in general optimal
either, because there may be containers with large capacities
but small (or, even almost zero) loads, rendering an attack to
such containers very ineffective in terms of triggering failures
by means of load redistribution. The worst-case deviation from
the optimal is again Θ(N) as demonstrated in Figure 3.
c) Greedy max-free-space attack: It is clear from the
previous two cases that the optimal attack strategy will be one
that considers both the loads and capacities of the containers
involved. The greedy approach that targets containers with
largest free space (i.e., (capacit y − load) difference) falls
into this category, and is based on the fact that containers
with largest free space will fail the latest in the course of a
cascading failure; e.g., see Section IV-C for a discussion of
this fact. Therefore, it is sensible to eliminate those containers
with a direct attack. On the other hand, containers with small
free space are already on the verge of failing and therefore can
be taken down by means of redistribution of loads. Although
this greedy strategy is intuitive (and in fact optimal in some
special cases), it fails to be the optimal solution in general. The
main reason is that this approach does not take into account the
loads of the containers directly. For example, a container may
have a large free space but its load may be negligible compared
to other containers, rendering a direct attack on this container
ineffective. The worst-case deviation from the optimal is again
Θ(N) as demonstrated in Figure 4.
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
Fig. 4. In this example we have n containers with (load, capacity) values
(, (n + 1) ) for the first n − 1 containers and (M, M + (n − 1) ) for the last
container, where > 0 is arbitrarily small and M satisfies M > (n2 − n) .
The greedy max free-space (C − L) attack will output k = n since it will start
attacking the leftmost containers and no cascading failures will take place. The
optimal solution is obviously k = 1 by attacking the last container.
C. Observations towards Designing a Smart Algorithm
We now present some observations that will be useful in
designing a smart attack algorithm.
a) The order of attack does not matter: In the equal
redistribution model, the order with which we launch an attack
does not affect the final set of failed containers. This is because
the load of the attacked nodes will be distributed to all of the
Í
remaining nodes so at the end an amount of i ∈ A Li will end
up in the remaining containers (leading to new failures or not)
irrespective of the order we chose to attack the containers in
A. We remark that an attack strategy can still be designed in
a greedy fashion, where the set A is determined one member
at a time, waiting for cascades to stop after each attack.
b) Order of failures during the cascading process:
Assume that containers are sorted by increasing free space,
Si = (Ci − Li ). Given that any failed load is redistributed
equally among the remaining containers, it is clear that this
ordering will remain the same throughout the course of cascading failures; the containers that are attacked directly at the
beginning are excluded from this argument. Therefore, in the
process of recursive load redistribution, containers will fail
(due to their free space diminishing to (below) zero) in this
exact same order: the one with smallest free space will fail
first, and so on and so forth.
V. O PTIMAL ATTACK S TRATEGIES UNDER S PECIAL C ASES
We now present three special cases of the ER-k problem
and provide optimal attack strategies for each of them.
a) Same Loads: An interesting situation arises when
initial loads are the same for all containers while capacities
differ. This reflects situations in which all lines in the power
system are given the same initial load, but have different
capacities owing to the physical constraints or material used.
We show that a greedy algorithm finds the optimal solution in
this special case. The ER-k-Same Loads Problem is defined
formally as follows.
INPUT: A non-negative rational number L for the common
load and a list of N non-negative numbers Ci > L, ∀i
indicating the capacity of each line. The integer k represents
the number of attacks we can launch.
OUTPUT: The set A of lines to be attacked that minimizes
n∞ (A) under the constraint | A| = k.
The next result, proved in Supplementary Material, shows
that the max-C-greedy algorithm finds the optimal solution.
6
Theorem 5.5: The max-C-Greedy Algorithm is optimal for
the ER-k-Same Loads Problem.
b) Same Free Spaces: Sometimes it might be the case
that the containers have arbitrary load and capacity but they
have a fixed free space. In [36], this was in fact shown to
be the optimal design that gives maximum robustness against
random attacks. We refer to the corresponding problem as the
ER-k-Same Free Spaces, formally defined as follows.
INPUT: A list of N non-negative rational numbers Li indicating the load of each container and a positive rational number
S indicating the common free space.
OUTPUT: Find the minimum number k of containers needed
to be attacked in order to destroy the whole system.
We changed the output from having a fixed number of lines
to be attacked to inflict the maximum damage, to the case
where we aim to destroy the whole system with the minimum
number of attacks. This is because in the case where every
container has the same free space, there are no intermediate
cascading failures. After an attack, the system will either fail
completely, or no single line will fail other than those attacked
directly. We show in Supplementary Material that the max-LGreedy algorithm that targets lines with the largest loads leads
to the optimal solution for this problem.
Theorem 5.6: The max-L-Greedy Algorithm is optimal for
the ER-k-Same Free Spaces Problem.
c) Capacities Proportional to Loads: In many cases,
the capacities and the loads of power lines are related in a
particular way. Namely, the capacity of a line is often set to
be proportional to its load. For example with α > 0 denoting
the tolerance factor, we have Ci = (1 + α)Li for each line
i = 1, . . . , N. In this variation, we will also show that there is
a greedy algorithm achieving the optimal solution. The ER-k(C ∝ L) Problem is defined formally as follows.
INPUT: A list of N non-negative numbers Li indicating the
load of each container and a positive number α such that
container capacities are set to Ci = (1 + α)Li for each i.
OUTPUT: The set A of lines to be attacked that minimizes
n∞ (A) under the constraint | A| = k.
In this setting, load, free-space, and capacity of the lines
will be ordered in the same way, and as we show in the
Supplementary Material, max-L, C, S-Greedy algorithms that
target lines with the largest load and free-space and capacity
give the optimal solution to this problem.
Theorem 5.7: The max-L, C, S-Greedy Algorithms are optimal for the ER-k-(C ∝ L) Problem.
VI. A MODIFIED OPTIMAL ATTACK PROBLEM WITH TOTAL
LOAD CONSTRAINTS
In this Section, we will prove that a variation of the decision
problem ER-k-k 0 is NP-Complete. In particular, we consider
the ER-k-k 0-Q problem, defined formally as follows.
INPUT: N pairs of non-negative numbers in the form
(Li, Ci ) indicating the load and the capacity of each line,
integers k and k 0 such that 0 < k < k 0 ≤ N, and a positive
number Q. We also assume Ci > Li for each i = 1, . . . , N.
OUTPUT: The answer to whether or not there is an attack
Í
set A with size k, and total sum of loads i ∈ A Li ≤ Q, such
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
that at the end of the cascading failures the number of failed
0
nodes is at least
Í k ; i.e., whether there exists A0 ⊂ {1, . . . , N }
with | A| = k, i ∈ A Li ≤ Q, and n∞ (A) ≤ N − k .
It is clear that the objective is two-fold here and that there
is an inherent trade-off: by attacking lines with larger initial
loads we can shed more load on other lines and have a better
chance to trigger a cascade of failures that would destroy the
whole system. However, the problem enforces a constraint on
the total load of the attacked containers as well. This knapsacklike trade-off is what makes the problem NP-complete as we
now show. Our proof is based on the reduction of the ER-k-k 0Q problem from the k-Subset Sum variant defined as follows:
Given a set of integers and a target sum Q, is there any subset
of size k whose sum is Q?
Theorem 6.8: The ER-k-k 0-Q Problem is NP-Complete.
7
solution with k lines L (1), . . . , L (k) . Then the loads of these
lines constitute a solution to the k-Subset-Sum problem.
The above reduction can be constructed in polynomial time
(more precisely, in linear time), so if there was a polynomial
algorithm that could solve the ER-k-k 0-Q, then the k-Subset
Sum would be in P, which is wrong unless P=NP. Thus, we
conclude that the ER-k-k 0-Q Problem is NP-complete.
An important implication of the above result is that the
optimization version of the ER-k-k 0-Q problem, which seeks
to find the set A of lines that minimizes n∞ (A) under the
Í
constraints | A| = k and i ∈ A Li ≤ Q, is NP-Hard. This means
that under these constraints, the adversary can not launch an
optimal attack in polynomial time unless P=NP.
VII. H EURISTIC ALGORITHMS AND THEIR PERFORMANCE
Proof. First, we show that ER-k-k 0-Q Problem is in NP: The
certificate is a list of the k containers we choose to attack.
We can check in polynomial time (e.g., see the ER-Attack
Projection algorithm in [4]) whether at least k 0 lines in the
system fail or not. Since we have a certificate that can be
checked in polynomial time, ER-k-k 0-Q is in NP!
Given an instance of the k-Subset Sum problem we will
create an instance of the ER-k-k 0-Q problem: Given a set of N
integers a1, a2, ..., a N , the k-Subset Sum problem asks whether
there exists k members Íof the set whose some equals Q. If
N
k = N, we can check if i=1
ai = Q and respond accordingly.
From now on, we suppose k < N and create an equivalent
version of the ER-k-k 0-Q problem in the following manner.
Let lines L1, . . . , L N have loads L1 = a1, L2 = a2, ..., L N = a N
and consider the ER-k-k 0-Q
Í problem; i.e., we seek to find a
set A of k lines such that i ∈ A Li ≤ Q and that attacking A
leads to failure of at least k 0 > k lines in the system. We also
set Ci = Li + Si where the free space is Si = NQ−k for each
i = 1, . . . , N. This last constraint ensures two things. First,
as discussed in Section V, when all lines have the same free
space then attacking k lines can only have two consequences:
either only those k lines that are attacked fail, or all N lines
fail. In either case, there is no cascade of failures and the
system reaches a steady-state immediately. Thus, with equal
free space among all lines, the ER-k-k 0-Q problem with k 0 >
k is equivalent to ER-k-N-Q problem. Secondly, under the
enforced assumptions it is clear that a complete system failure
will take place if and only if the total load failed by the initial
attack A is larger than the sum of the free spaces of those that
are not in the attack set A; i.e., if and only if
Õ
Õ
Q
= Q.
Li ≥
S j = (N − k)
N
−k
i∈A
j ∈ {1,..., N }/A
Here, the first equality follows from the facts that | A| = k
and Si = NQ−k Í
for each i = 1, . . . , N. Recalling
Í further the
constraint that i ∈ A Li ≤ Q, this leads to i ∈ A Li = Q.
Therefore, the created instance of the ER-k-k 0-Q problem
indeed seeks to find a subset A of {a1, . . . , a N } such that
Í
| A| = k and i ∈ A Li = Q, rendering it equivalent to the
k-Subset Sum instance that we have started with. For the
reverse direction, assume that the ER-k-k 0-Q problem has a
Although, it is not known whether the original optimization
problem of finding the best k lines to attack to minimize final
system size is NP-Hard or not, the discussion in the preceding
section indicates that the optimal attack problem is likely to
be computationally challenging; in particular, we know that
the problem is NP-Hard if we are further constrained by the
total load of those we can attack. This prompts us to develop
heuristic algorithms, for both the original and the modified
optimization problems, that work in polynomial time and
have competitive performance under arbitrary load-capacity
distributions. The performance of these heuristics will then
be compared with some benchmark heuristics such as max-L,
max-C, max-S, and random attacks.
In the interest of brevity and concreteness, the discussion is
restricted here to non-greedy algorithms that choose the attack
set A without ever running the attack projection algorithm.
In other words, the algorithms are not allowed to run the
cascading failures initiated by a subset of A, and then continue
making the remaining selections from the lines that survived
the cascade. Of course, all heuristics considered here including
the benchmarks can be modified to operate in a greedy fashion.
One might also be tempted to use a greedy algorithm where,
at each round, the line to be attacked is chosen in an optimal
way; i.e., the line whose failure leads to smallest number
of surviving lines is chosen from among all lines that are
still alive. However, for the problem at hand, one can realize
that unless | A| is relatively large, the final system size equals
n∞ (A) = N − | A| regardless of the set A of attacked lines. For
example, with | A| = k, this will be the case whenever
k≤
Smin
(N − k),
Lmax
meaning that if Smin /Lmax > 0, the greedy heuristic will have
to deal with ties when making its choices for the next line to
be attacked until it makes Ω(n) choices. Since resolving the
ties by randomization for such a large number of selections
is likely to lead to poor performance, one needs heuristic
rules to resolve the ties. Even then, our preliminary simulation
study indicated that the greedy versions of the heuristics
considered here perform only slightly better than their nongreedy counterparts, and the comparison among the greedy
heuristics provided no additional insight to what was already
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
8
observed from non-greedy algorithms; hence the decision to
consider only non-greedy attack strategies here.
A. Heuristics for the original optimization problem
We first consider the original case where there is no constraint on the total load of the lines that can be attacked; i.e.,
we consider the ER-k problem. Let A be the set of lines to
be attacked such that | A| = k. It is clear from the previous
discussions that a good attack should aim to
Õ
maximize
Li, and
(6)
i∈A
minimize S j , j ∈ {1, . . . , N } − A
(7)
In words, the attack should aim to find the lines with the largest
free space while making the total load of the failed lines as
large as possible. Thus, the attack should intuitively look for
lines with large initial load and large free-space. Of course,
most difficult situations arise when the load and free-space
values of the lines are in reverse order; e.g., the highest load
carrying line has the smallest free-space, etc. as in Figure 2.
Our main idea towards handling the trade-off described
above is based on its similarities with the well-studied 0-1
Knapsack problem. In the 0-1 Knapsack problem, we are given
a set of N items, {1, . . . , N }, each with a weight wi and a value
vi , and the goal is to choose items such that their total value
is maximized while the total weight is bounded by W; i.e.,
Õ
maximize
vi xi
(8)
subject to
i
Õ
wi xi ≤ W
and
xi ∈ {0, 1}
a) Maximum Load×Free Space Attack: In this algorithm,
the load free space product, Li ∗ Si is computed for each line
i = 1, . . . , N. After sorting the lines based on this product, the
k lines to be attacked is chosen as the ones having the highest k
load-free space product. As mentioned above, this is inspired
by the 2-approximation heuristic for the Knapsack problem
that orders items according to vi ∗ w1i when the goal is to
choose items with high v and low w. In the ER-k problem, we
wish to choose lines with high L and high S, or equivalently,
with high L and low S1 . Thus, constructing an analogy between
value vi and load Li , and weight (or, cost) wi and 1/Si , our
heuristic chooses lines with the maximum
(9)
i
The 0-1 Knapsack problem is known to be NP-Hard, but
polynomial-time heuristics can still give close-to-optimal solutions. For example, a competitive heuristic is to order the
items based on their "value per weight", i.e., vi /wi , and choose
items according to this order, starting with the one with the
highest vi /wi , until the total weight capacity W is reached.
In fact, with a small modification to handle corner cases, this
heuristic is known to yield at least 50% of the optimum value.
The optimal attack problem we consider, i.e., the ER-k
problem, has some similarities with but is not equivalent to
the 0-1 Knapsack problem. In particular, one can construct an
analogy between the constraints of the 0-1 Knapsack problem
and the ER-k problem by assigning all item weights as wi = 1
and the total weight limit to be W = k. However, the objectives
(6)-(7) of the ER-k problem are much more complex than the
objective (8) of the Knapsack problem. Nevertheless, the two
problems have some similarity in that their main difficulties lie
in the trade-offs involved. In the Knapsack problem the tradeoff is between the value and the weight of the item and it is
desirable to pick items with high value and low weight, while
in the ER-k, the trade-off lies between the possibly conflicting
objectives of choosing lines with high load and high freespace. Inspired with the efficient heuristic for the Knapsack
problem that is based on selecting items with the largest vi /wi
ratio (i.e., items with the biggest bang for the buck), our first
heuristic for the ER-k problem is based choosing lines with
the highest load times free-space, i.e., with the highest Li ∗ Si .
Li ∗
1
= Li ∗ Si .
1/Si
The performance of this heuristic is demonstrated via several
numerical examples in the next subsection along with a
comparison with some benchmark heuristics.
Aside from its connection to a powerful heuristic in a
relevant problem, the maximum L ∗ S heuristic has several
advantages. First of all, this heuristic becomes equivalent to
the optimal attack strategy in the three special cases considered
in Section V; e.g., when all lines have the same load, it
chooses ones with highest free-space (and hence capacity), or
when all lines have the same free-space, it chooses lines with
maximum L, etc. Secondly, considering the product L ∗ S is
an effective way to favor lines with high load and free-space,
while heavily penalizing load or free-space values close to
zero; note that benchmark heuristics including the highestcapacity attack (C = L + S) fail to penalize small L or C
values. In the optimal attack problem, this makes perfect sense
given that a line with almost no load should never be attacked
even if it has very high free-space since its failure will likely
not affect any other line. Similarly, it may not be a good idea
to directly attack a line with almost no free-space even if it
has very high load, since the line will likely fail due to load
redistribution regardless of which other lines are attacked.
b) Maximum L ∗ S β attack: While maximum L ∗ S attack
is intuitive and will be seen to be powerful in many cases,
we observe that its performance can be further improved by a
small modification. To this end, we propose a second heuristic
as a modified version of the max-L ∗ S attack that allows
adjusting the relative importance of load and free-space values
of the lines. In particular, with β in [0, ∞], we consider a
β
heuristic that chooses k lines with the maximum Li ∗ Si .
An added benefit of this heuristic is that it contains several
heuristics as special cases. In particular, the maximum L ∗ S
algorithm described above is obviously a special case of this
algorithm, corresponding to the case β = 1. Also, by setting
β = 0, this heuristic reduces to the max-L attack, while setting
β = ∞ (or, large enough) makes it equal to the max-S attack.
B. Numerical Comparison with Benchmark Heuristics
We now compare the heuristics we developed against some
benchmark heuristic algorithms via numerical experiments.
The benchmark heuristics we will consider are given below:
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
9
5000
β
β
β
β
β
β
β
4000
4000
Final system size
Final system size
5000
3000
2000
1000
highest S
highest L
highest C = L + S
highest L ∗ S
random attack
3000
2000
=0
= 0.3
= 0.6
= 1.0
= 1.5
= 2.0
= 10.0
min. attack to fail all
1700
1600
1000
1500
0
500
1000
1500
2000
Attack size, k
1400
0
0.4
0.8
0
500
Fig. 5. The performance comparison of different heuristic algorithms for
L ∼ U[10, 30], S ∼ U[10, 60], N = 5000.
a) Random attack: This is the most primitive attack
strategy and considered here only for comparison purposes.
The attack picks k lines to be attacked uniformly at random
from amongst all N lines.
b) Highest-L, highest-C, highest-S attacks: These three
attacks are based on sorting lines with respect to their initial
load Li , free-space Si , or capacity Ci = Li + Si , respectively
and attacking the top k lines with the highest value of the
corresponding metric.
We fix the number of lines at N = 5000 for all experiments. First, we consider the case where each line is
independently given an initial load from a uniform distribution,
U(Lmin, Lmax ), where we set Lmin = 10 and Lmax = 30. The
free-space allocated to each line is generated independently
from its load, again from a uniform distribution, U(Smin, Smax )
with Smin = 10, Smax = 60. The capacity of a line Li is
given by the sum Ci = Li + Si . The independence of L and S
leads to some lines having high load but small free-space, or
vice versa, rendering the optimal attack problem non-trivial;
e.g., with these choices, the realized load-capacity values will
almost surely not fall into one of the special cases presented
in Section V where an optimal solution is available.
Under this setup, we compute the final system size as
a function of the number k of lines attacked, where the
set of attacked lines are selected according to the heuristics
considered. The results are given in Figure 5 where each data
point is obtained by averaging over 100 independent runs. We
already see in this simple setting that our attack strategy of
targeting lines with the highest L ∗ S outperforms all other
benchmarks (except a small interval of attack size where maxL attack seems to give the highest damage). In particular, we
see that the highest L ∗S attack is able to fail the whole system
by targeting 90, 180, 210, and 450 fewer lines as compared
to max-C, max-L, max-S, and random attacks, respectively.
Next, we check if this performance can be further improved
by attacking lines with highest L ∗ S β for some β ≥ 0. To
this end, we repeat the previous experiment as β varies from
zero to ten. The results are demonstrated in Figure 6 and as
expected show that with β = 0 or β 1, we obtain the same
performance with max-L and max-S attack, respectively. More
interestingly, we see that the case β = 1 is indeed not the best
one can do. For example, we see that when β = 0.3, the max-
β
1.2
1.6
2
1000
1500
Attack size, k
Fig. 6. The performance comparison of maximum L ∗ S β algorithms for
various β values for L ∼ U[10, 30], S ∼ U[10, 60], N = 5000. Inset:
The minimum number of lines needed to be attacked to fail all lines for
the maximum L ∗ S β attack.
L ∗ S β attack can fail the whole system by attacking 75 fewer
lines the case for β = 1. To demonstrate this better, we plot in
the inset of Figure 6 the minimum number lines needed to be
attacked to fail all N lines. To compute this, we again run 100
independent experiments and pick the minimum attack size
for which all 100 experiments led to entire system failure.
The performance of heuristic algorithms are known to vary
significantly under different parameter settings, and our case is
not expected to constitute an exception. To that end, we have
tested the performance of the max-L∗S β attack with β ∈ [0, 4],
for a wide range of possibilities for the distribution of L and S.
In all cases we considered, we were able to identify a β value
for which the max-L ∗ S β attack was at least as good as all
benchmark attacks (random, highest-C, L, S-attacks) showing
its versatile performance.
As already mentioned, the most challenging cases arise
when the load and free-space values are in reverse order.
To that end, we close this section by demonstrating the
performance of the max-L ∗ S β attack in such cases. In
particular, we start by generating L1, . . . , L N and S1, . . . , SN
independently according to some distribution. Then, the load
values (resp. free-space values) are sorted and re-arranged
in increasing (resp. decreasing) order, leading to highest-load
carrying line having the smallest free-space, and so on. To
make the problem more challenging and interesting, we also
consider Pareto distribution. Namely, a random variable X is
said to follow Pareto distribution, written X ∼ Pareto(Xmin, b)
with Xmin > 0 and b > 0, if its probability density is given by
b
pX (x) = Xmin
bx −b−1 1 [x ≥ Xmin ] .
To ensure that E [X] = bXmin /(b − 1) is finite, one must set
b > 1, while the variance of X is finite only if b > 2.
The results for the case where L and S values are reverse
ordered are depicted in Figure 7. Here, we show a small
number of representative results that correspond to different
behaviors for brevity. As before, all results correspond to the
minimum attack size that led to an entire system collapse in
all 100 experiments. The curves represent the results for the
max-L ∗ S β attack as β varies from zero to two. In each plot,
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
10
3600
L ∼ Pareto(10, 1.1), S ∼ U (10, 200)
L ∼ Pareto(10, 2.5), S ∼ Pareto(8, 1.2)
L ∼ U (0.4, 100), S ∼ U (0.05, 150)
L, S ∼ Pareto(10, 1.2)
highest-C attack
random attack
min. attack to fail all
3000
2400
lines. This is because at the first stage of the cascades, any
line Li that was not directly attacked will fail only if
Í
Si ≤
Li
.
N − | A|
i∈A
1800
1200
600
0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
β
Fig. 7. Minimum number of lines needed to be attacked to fail all lines in the
system is shown when load and free-space values are generated independently
(from the distributions given at the figure legend) and then sorted in reverse
orders; e.g., to ensure L1 ≤ L2 ≤ · · · ≤ L N , while S1 ≥ S2 ≥ · · · ≥ S N .
Curves stand for the results obtained under the max-L ∗ S β attack as a function
of β. Corresponding results for the max-C attack are shown by filled square
symbols, and those for the random attack by filled circles.
we add the corresponding results for the max-C attack (shown
by a filled square) and random attack (shown by a filled circle)
as well; for convenience, the x-axis values for these symbols
are chosen such that they stay on the corresponding curve
showing the results for max-L ∗ S β attack. We note that the
max-L attack is already demonstrated by the case β = 0 while
β = 2 gives a good indication of the performance of the max-S
attack, so these plots provide a comparison of the max-L ∗ S β
attack with all the benchmarks considered here.
The main observations from Figure 7 are as follows. We
see that in all cases there is a particular β value for which the
max-L ∗ S β attack performs the best among all benchmarks;
it is only the case of L, S ∼ Pareto(10, 1.2) where we see that
the best performance of max-L ∗ S β attack is attained when
β = 0, or β ∈ (0.4, 0.7) meaning that max-L attack matches
the performance of the max-L ∗ S β attack. Also, we see that
the β value that leads to the best performance can be equal to,
smaller than, or larger than one in different scenarios showing
the importance of trying different values of β to get the best
performance. Finally, while benchmark attacks (including the
random attack) occasionally give results close to the best maxL ∗ S β attack, we see examples for each benchmark where its
performance is significantly worse than the best max-L ∗ S β
attack; these cases are summarized in Table I.
C. Heuristic attacks for the modified optimization problem
We now consider the modified optimization problem ERk-k 0-Q where the attack set A is further constrained by
Í
i ∈ A Li ≤ Q, in addition to | A| ≤ k. As shown in Theorem
6.8, in this case finding the attack set A that minimizes the
final system size is NP-Hard, prompting us to develop heuristic
strategies. With the additional constraint on the total load of
the lines that we can attack, the trade-offs involved become
more complicated and heuristics developed in the previous
section may not be well-suited for the ER-k-k 0-Q problem.
Ultimately, our strategy should be to choose an attack set that
has k (or, very close to k) lines with total load equal (or, very
close) to Q, and that have the highest free space among all
Í
Thus, to facilitate failures it is desirable to make i ∈ A Li and
| A| as large as possible, while Si as small as possible.
Given the multiple constraints involved, this problems
shows similarity with the 2-dimensional 0-1 Knapsack problem
[12], [23], [25]: Consider a collection of items, where each
item i is given a value vi , has Í
weight wi , and volume
Í qi . The
objective
is
then
to
maximize
v
x
subject
to
i
i
i
i wi xi ≤ W
Í
and i qi xi ≤ Q where xi ∈ {0, 1}; i.e., we want to choose
items with the maximum total value while the total weight is
limited by W and total volume is limited by Q. As can be
inferred from the discussion above, an important difference of
the ER-k-k 0-Q problem is that while it is desirable to choose
lines with high Si (could be thought to be analogous to the
“value” of the item) under given constraints, it is perhaps
equally important to attain or be very close to the limits on
both total load and total number of lines attacked.
With these in mind, our heuristics for the ER-k-k 0-Q
problem are based on incorporating a switch to the previously
developed heuristics that is actuated to ensure that the attack
set A attains or gets close to both constraints on its cardinality
and the total load. This idea is inspired from the greedylike heuristic developed for the multi-dimensional Knapsack
problem in [12]. This algorithm initially starts choosing items
based on a given set of rules until one or more of the
constrained resources become scarce, and then switches to a
different set of rules that favor items that use very little of
the scarce resource. Here, we propose to use a heuristic that
chooses the lines to be attacked one at a time according to the
previously developed max-L ∗ S strategy. After each selection,
we check whether the switch needs to be activated. Namely,
with A0 denoting the set of k 0 lines selected so far, we check
i) Is it still feasible to select all remaining k − k 0 lines
from the smallest load carrying lines available? Namely,
with the remaining lines’ loads sorted in ascending order
0
L (1) ≤ L (2) ≤ · · · ≤ L (N −k ) , we check if
Distribution of
L and S
L ∼ Pareto(10, 1.2)
S ∼ Pareto(10, 1.2)
L ∼ U(0.4, 100)
S ∼ U(0.05, 150)
L ∼ Pareto(10, 2.5)
S ∼ Pareto(8, 1.2)
L ∼ Pareto(10, 1.1)
S ∼ U(10, 200)
L, S from the
UK National Grid∗
Minimum # of lines to attack to fail all
random
max-C
max-L
max-S
best β
981
151
71
2241
71
691
1061
2611
1021
491
1671
1611
1421
2111
1411
791
711
3261
2221
541
2371
1491
1611
2111
1441
TABLE I
Performance comparison of benchmark attacks with the best result of the
max-L ∗ S β attack. The first four rows are obtained from Figure 7, while the
last row is obtained from simulations with UK National Grid data (see Section
VIII for details). Values significantly worse (in the sense of needing to attack
many more lines to fail all) than the best-L ∗ S β attack are made bold.
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
Total load constraint: Ltot ≤ 0.25 ∗ k ∗ E[L]
4000
3000
highest L
highest C
highest L ∗ S
highest LS
highest S
random attack
highest LS with switch
highest L ∗ S with switch
2000
1000
Total load constraint: Ltot ≤ 0.75 ∗ k ∗ E[L]
5000
Final system size
5000
Final system size
11
0
4000
3000
highest L
highest C
highest L ∗ S
highest LS
highest S
random attack
highest LS with switch
highest L ∗ S with switch
2000
1000
0
0
500
1000
1500
2000
2500
3000
0
500
Max. attack size, k
1000
(a)
Final system size
Final system size
3000
highest L
highest C
highest L ∗ S
highest LS
highest S
random attack
highest LS with switch
highest L ∗ S with switch
1000
2500
3000
Total load constraint: Ltot ≤ 1.25 ∗ k ∗ E[L]
5000
4000
2000
2000
(b)
Total load constraint: Ltot ≤ k ∗ E[L]
5000
1500
Max. attack size, k
0
4000
3000
highest L
highest C
highest L ∗ S
highest LS
highest S
random attack
highest LS with switch
highest L ∗ S with switch
2000
1000
0
0
500
1000
1500
2000
2500
Max. attack size, k
(c)
0
500
1000
1500
2000
2500
Max. attack size, k
(d)
Fig. 8. The performance comparison of different heuristic algorithms for L ∼ U[10, 30], S ∼ U[10, 60], N = 5000, when the attack is constrained to k lines
such that their total load satisfies a) Ltot ≤ 0.25 ∗ k ∗ E[L]; b) Ltot ≤ 0.75 ∗ k ∗ E[L]; c) Ltot ≤ 1.0 ∗ k ∗ E[L]; d) Ltot ≤ 1.25 ∗ k ∗ E[L].
k−k
Õ0
L (j) +
j=1
Õ
Li ≤? Q
i ∈ A0
If the answer is YES, we continue with the second
condition for the switch, while if the answer is NO, the
switch is activated and algorithm is finished by appending
A0 with k − k 0 − 1 lines with the smallest load-carrying
0
lines available; i.e., with L (1), . . . L (k−k −1) . Alternatively,
one can release the latest added member of A0 and append
0
it with k − k 0 lines L (1), . . . L (k−k ) ; we found no major
performance difference between these two approaches.
ii) Next, check whether it is feasible to select all remaining
k − k 0 lines from the largest load carrying lines available. Namely, with the remaining lines’ loads sorted in
0
ascending order L (1) ≤ L (2) ≤ · · · ≤ L (N −k ) , we check if
N
−k 0
Õ
N −k+1
L (j) +
Õ
Li ≤? Q
i ∈ A0
If the answer is NO, we continue the algorithm with the
next selection, while if the answer is YES, the switch is
activated and algorithm is finished by appending A0 with
k − k 0 lines with the largest load-carrying lines available;
0
i.e., with L (1), . . . L (k−k −1) .
The two conditions of the switch described above ensure
that while the initial selections are made in line with the
original objectives of picking lines with high load and high
free-space, care is also given so as to be able to pick k (or,
k − 1) lines whose total load is close to the limit Q. Of course,
any algorithm including the benchmarks can be modified using
the switch idea to better accommodate total load constraints. In
particular, when the total load limit Q is extremely stringent,
it would be tempting to pick lines with small load so as to not
exhaust the total load limit quickly, while aiming to choose
lines with high free-space. This prompts us to consider the
max-S/L heuristic as well, including its modification with
the switch idea described above. To keep the discussion brief
we do not present results for the max-L ∗ S β attack (with
or without switch) and consider only the case where β = 1;
this is in part due to the fact that when the switch is added,
the performance of the max-L ∗ S β attack becomes much less
sensitive to variations in β over small ranges.
We now present numerical results to evaluate the performance of the heuristic attacks developed here and compare
them against benchmark algorithms; as before, we will use
max-C, S, L and random attacks as benchmarks. Different from
the experiments conducted in Section VII-A, here we have to
vary not only the maximum number k of lines that can be
attacked, but also the limit Q on the total load of the attacked
lines. In particular, we would expect the performance of the
algorithms to depend heavily on Q. To this end, we find it
meaningful to let Q vary with k and to set it in reference to
the mean total load of k randomly selected lines; i.e., we set
Q = Q(k) = c ∗ k ∗ E[L]
(10)
for some constant c > 0. This choice enables us to tune c
to different levels and check performance in cases where (i)
the total load is extremely limited (i.e., c 1); ii) total load
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
Total load constraint: Ltot ≤ 0.25 ∗ k ∗ E[L]
4000
highest L
highest C
highest L ∗ S
random attack
highest S, highest LS
highest LS with switch
highest L ∗ S with switch
3000
2000
Total load constraint: Ltot ≤ 0.75 ∗ k ∗ E[L]
5000
1000
Final system size
5000
Final system size
12
0
4000
highest L
highest C
highest L ∗ S
random attack
highest S, highest LS
highest LS with switch
highest L ∗ S with switch
3000
2000
1000
0
0
500
1000
1500
2000
2500
3000
0
500
Max. attack size, k
1000
(a)
Final system size
Final system size
3000
2500
Total load constraint: Ltot ≤ 1.25 ∗ k ∗ E[L]
5000
highest L
highest C
highest L ∗ S
random attack
highest S, highest LS
highest LS with switch
highest L ∗ S with switch
4000
2000
(b)
Total load constraint: Ltot ≤ k ∗ E[L]
5000
1500
Max. attack size, k
2000
1000
0
highest L
highest C
highest L ∗ S
random attack
highest S, highest LS
highest LS with switch
highest L ∗ S with switch
4000
3000
2000
1000
0
0
500
1000
1500
2000
2500
Max. attack size, k
(c)
0
500
1000
1500
2000
2500
Max. attack size, k
(d)
Fig. 9. The performance comparison of different heuristic algorithms for L ∼ U[0.4, 100], S ∼ U[0.05, 150], with L and S sorted in reverse order, N = 5000,
when the attack is constrained to k lines such that their total load satisfies a) Ltot ≤ 0.25 ∗ k ∗ E[L]; b) Ltot ≤ 0.75 ∗ k ∗ E[L]; c) Ltot ≤ 1.0 ∗ k ∗ E[L]; d)
Ltot ≤ 1.25 ∗ k ∗ E[L]. Each data point is obtained by averaging over 100 independent runs.
limit is such that heuristics that do not take load into account
(such as max-S or random attacks) will likely be able to select
(close to) k lines with total load very close to Q (i.e., c ≈ 1);
or iii) total load limit is not stringent at all (i.e., c 1) and
the problem is similar to the unconstrained load case.
In the first set of experiments, we set N = 5000 and generate
load-free space values independently from the distributions
L ∼ U[10, 30] and S ∼ U[10, 60]. For brevity we consider four
values of c given at (10): c = 0.25, c = 0.75, c = 1.0, c = 1.25.
The results are presented in Figure 8 from which a number
of interesting observations can be made. When c = 0.25,
i.e., when total load is extremely constrained, we see that all
heuristics without a switch perform poorly and are not able
to fail the whole system even until k = 3000. This can be
attributed to their inability to attack the maximum allowed
number k of lines as they quickly exhaust the total load limit.
On the other hand, we see that both max-L ∗S-with-switch and
max-S/L-with-switch attacks perform much better, and despite
the stringent limit on the total load are able to fail the system
by attacking about 50% more lines than required in the case
where the total load is unlimited. When c is increased to 0.75,
we see that the performance of the benchmarks improve but
still are significantly worse than the two heuristics that use the
switch; in this case we also see that the max-S/L-with-switch
attack slightly outperforms max-L ∗ S-with-switch.
With c = 1, we see that algorithms that ignore the loads
such as max-S and random attacks perform as well as they
do in the unconstrained case; this is expected by virtue of the
law of large numbers. In particular, when c = 1, we would
expect max-S to perform well since it picks the most robust
lines in the system and is likely to reach the limits k and Q
simultaneously given that L is independent from S. Figure 8(c)
confirms this intuition where we also see that both heuristics
with switch match the performance of the max-S attack.
Finally, with c = 1.25, we get close to the unconstrained
load case, and as expected see that the performance of the
max-L ∗ S algorithm becomes the best. What is interesting
here is that the max-L ∗ S-attack-with-switch is able to match
this performance, showing its versatile performance across
very different cases considered here. Overall, these experiments demonstrate that incorporating the switch significantly
improves the performance and the max-L ∗ S-attack performs
well across different ranges of the total load limit.
As in Section VII-A, it is of interest to check the performance of these algorithms in difficult cases where load and
free-space values are sorted in reverse orders. To this end, we
consider one of the settings used in Figure 7, and generate
load and free-space independently from L ∼ U[0.4, 100] and
S ∼ U[0.05, 150], and then sort them in reverse orders so that
the line with maximum load gets the smallest free-space and so
on and so forth. In this setting, max-S and max-S/L heuristics
become equivalent. Also, since large S values are around 150
while L is limited to 100, the lines with maximum-C will
be those with large S (and small L due to reverse ordering).
As seen in Figure 9, this leads to three benchmarks (maxS, C, S/L) performing almost equally in this setting.
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
We see from Figure 9 that the max-L ∗ S-attack-with-switch
once again performs well. It leads to the best performance
among all heuristics considered for c = 0.25, c = 0.75, and
c = 1.0 (equal with max-L ∗ S), while coming second for
c = 1.25 after max-L ∗ S; this is expected since without a
stringent limit on total load, the problem gets closer to the
unconstrained case where a switch is not needed. Also, in this
case we see that the performance of the max-S attack (along
with max-S/L and max-C attacks) is rather unaffected by the
load constraint. We attribute this to the fact that since load and
free-space are reverse ordered, targeting max-S lines is equal
to targeting min-L lines and even when c = 0.25 the total load
limit is not likely to be exhausted easily; i.e., the algorithm is
able to choose k lines without exceeding the total load limit.
We think this is also the reason why the switch is not helping
(and, actually hurting) the max-S/L algorithm in this case.
VIII. S IMULATIONS WITH UK NATIONAL G RID DATA
In this section, we provide simulation results illustrating
how the attack strategies covered here performs when the load
and free space distribution are based on real data. We have
used National Grid Electricity Ten Year Statement 2016 Model
of Great Britain [15] to generate load-free space pairs. To be
more precise, the load distribution is chosen from the winter
peak power flow diagram presented in [15, Appendix C]. For
the free-space distribution, the transmission line ratings given
in [15, Appendix B] has been used. As in the case of previous
examples, the number of lines is taken as N = 5000.
In Figure 10, we show the performance of the heuristic algorithms for the unconstrained case, i.e., for the ER-k problem.
The results are very similar to those obtained under synthetic
load-free space distributions and demonstrate that heuristics
developed here perform well under real-world distributions
as well. In particular, we see that the proposed max-L ∗ S
heuristic performs better than all benchmarks considered, and
its performance can further be improved by the max-L ∗ S β
attack. For the UK National Grid data, our results indicated
that the best performance is obtained when β = 1.5; see also
the last row in Table I.
Next, we consider the ER-k-k 0-Q problem where the total
load of the attack set is bounded by Q. As in Figures 8 and 9,
we set Q according to (10) for several c values, and compare
the performance of our attack strategies with benchmarks;
Final system size
5000
13
this time the load-free space distribution is set according to
the aforementioned UK National Grid data. The results are
depicted in Figure 11. Once again we see that the max-L ∗ S
attack with switch leads to the best overall performance among
all heuristics considered; it leads to the best performance when
c = 0.25 (tied with max-S/L with switch), c = 0.75, and
c = 1.25, while coming as second for c = 1 after maxS/L. This shows that the heuristic attacks proposed here have
versatile performance also under distributions observed in realworld power systems.
IX. C ONCLUSION
In this paper, we consider a flow-redistribution based model
to understand cascading failures in load-carrying networks,
e.g., electrical power systems, transportation systems, etc. We
focus on assessing the vulnerability of such systems against
adversarial attacks. In particular, in a system with N lines with
initial loads L1, . . . , L N and capacities C1, . . . , CN , we study
the constrained optimization problem of finding k initial lines
to be attacked to minimize the final number of alive lines in the
system. We give optimal greedy algorithms in several special
cases, and prove that a variation of the problem (with a bound
on the total load of the initial attack set) is NP-Hard. Several
heuristics are developed and their performance is compared
with benchmark attacks under various settings. Overall, it is
seen that the system is most vulnerable against attacks that
target lines with maximum load free-space product L ∗(C − L).
There are many interesting directions to consider for future
work. First, the complexity of the optimal k-attack problem
(without a bound on the total load) is not known. Also, with
the results of this paper revealing good attack strategies, one
might now seek optimal design strategies (e.g., in the form of
load-capacity distributions) that lead to maximum robustness
against such attacks. It would also be interesting to consider
new cascading failure models for flow networks that combine
local and global redistribution approaches; e.g., a portion of the
failed load is redistributed in the local neighborhood according
to network topology while the rest is redistributed globally. For
this last problem, we provide preliminary simulation results in
Supplementary Material. They suggest that the mean-field (i.e.,
equal) redistribution model captures the qualitative behavior
of system robustness well. Finally, it might be interesting
to study information cascades in social networks [27], [30],
[31], [33] using the models considered here; the optimal
attack problem studied here will then amount to influence
maximization problem [10].
4000
R EFERENCES
3000
2000
1000
highest S
highest L
highest C = L + S
highest L ∗ S
highest L ∗ S 1.5
random attack
0
500
1000
1500
2000
2500
Attack size, k
Fig. 10. The performance comparison of different heuristic algorithms when
L, S follow the UK National Grid data [15].
[1] J. V. Andersen, D. Sornette, and K.-t. Leung. Tricritical behavior in
rupture induced by disorder. Phys. Rev. Lett., 78:2140–2143, Mar 1997.
[2] G. Andersson, P. Donalek, R. Farmer, N. Hatziargyriou, I. Kamwa, et al.
Causes of the 2003 major grid blackouts in north america and europe,
and recommended means to improve system dynamic performance.
Power Systems, IEEE Transactions on, 20(4):1922–1928, 2005.
[3] S. V. Buldyrev, R. Parshani, G. Paul, H. E. Stanley, and S. Havlin.
Catastrophic cascade of failures in interdependent networks. Nature,
464:1025–1028, 2010.
[4] E. Chatziafratis, Y. Zhang, and O. Yağan. On the robustness of power
systems: optimal load-capacity distributions and hardness of attacking.
In Information Theory and Applications Workshop (ITA), 2016.
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
Total load constraint: Ltot ≤ 0.25 ∗ k ∗ E[L]
4000
3000
highest L
highest C
highest L ∗ S
highest S
highest LS
random attack
highest LS with switch
highest L ∗ S with switch
2000
1000
Total load constraint: Ltot ≤ 0.75 ∗ k ∗ E[L]
5000
Final system size
5000
Final system size
14
0
4000
3000
highest L
highest C
highest L ∗ S
highest S
highest LS
random attack
highest LS with switch
highest L ∗ S with switch
2000
1000
0
0
500
1000
1500
2000
2500
3000
3500
4000
0
500
Max. attack size, k
1000
(a)
Final system size
Final system size
3000
highest L
highest C
highest L ∗ S
highest S
highest LS
random attack
highest LS with switch
highest L ∗ S with switch
1000
2500
3000
Total load constraint: Ltot ≤ 1.25 ∗ k ∗ E[L]
5000
4000
2000
2000
(b)
Total load constraint: Ltot ≤ k ∗ E[L]
5000
1500
Max. attack size, k
4000
3000
highest L
highest C
highest L ∗ S
highest S
highest LS
random attack
highest LS with switch
highest L ∗ S with switch
2000
1000
0
0
0
500
1000
1500
2000
2500
3000
Max. attack size, k
(c)
0
500
1000
1500
2000
2500
3000
Max. attack size, k
(d)
Fig. 11. The performance comparison of different heuristic algorithms when L, S pairs are distributed according to the UK National Grid data [15]. The attack
is constrained to k lines such that their total load satisfies a) Ltot ≤ 0.25 ∗ k ∗ E[L]; b) Ltot ≤ 0.75 ∗ k ∗ E[L]; c) Ltot ≤ 1.0 ∗ k ∗ E[L]; d) Ltot ≤ 1.25 ∗ k ∗ E[L].
[5] P. Crucitti, V. Latora, and M. Marchiori. Model for cascading failures
in complex networks. Phys. Rev. E, 69:045104, Apr 2004.
[6] H. Daniels. The statistical theory of the strength of bundles of threads.
i. Proceedings of the Royal Society of London. Series A. Mathematical
and Physical Sciences, 183(995):405–435, 1945.
[7] L. Daqing, J. Yinan, K. Rui, and S. Havlin. Spatial correlation analysis
of cascading failures: congestions and blackouts. Sci. Reps., 4, 2014.
[8] I. Dobson, B. A. Carreras, V. E. Lynch, and D. E. Newman. Complex
systems analysis of series of blackouts: Cascading failure, critical points,
and self-organization. Chaos, 17(2):026103, 2007.
[9] I. Dobson, J. Chen, J. Thorp, B. A. Carreras, and D. E. Newman.
Examining criticality of blackouts in power system models with cascading events. In Proceedings of the 35th Annual Hawaii International
Conference on System Sciences, 2002.
[10] D. Kempe, J. Kleinberg, and É. Tardos. Maximizing the spread
of influence through a social network. In Proc. of ACM SIGKDD
Conference, pages 137–146, 2003.
[11] R. Kinney, P. Crucitti, R. Albert, and V. Latora. Modeling cascading
failures in the north american power grid. The European Physical
Journal B, 46(1):101–107, 2005.
[12] R. Loulou and E. Michaelides. New greedy-like heuristics for the multidimensional 0-1 knapsack problem. Operations Research, 27(6):1101–
1114, 1979.
[13] B. Mirzasoleiman, M. Babaei, M. Jalili, and M. Safari. Cascaded failures
in weighted networks. Physical Review E, 84(4):046114, 2011.
[14] A. E. Motter and Y.-C. Lai. Cascade-based attacks on complex networks.
Phys. Rev. E, 66:065102, Dec 2002.
[15] NationalGrid. http://www2.nationalgrid.com/uk/industry-information/
future-of-energy/electricity-ten-year-statement/. Technical report, 2016.
[16] T. D. O’Rourke.
Critical infrastructure, interdependencies, and
resilience.
BRIDGE-WASHINGTON-NATIONAL ACADEMY OF
ENGINEERING-, 37(1):22, 2007.
[17] T. J. Overbye, X. Cheng, and Y. Sun. A comparison of the ac and dc
power flow models for lmp calculations. In Proceedings, 37th Hawaii
International Conference on System Sciences, 2004.
[18] O. Ozel, B. Sinopoli, and O. Yağan. Robustness of flow networks against
cascading failures under partial load redistribution. arXiv:1802.07664.
[19] S. Pahwa, C. Scoglio, and A. Scala. Abruptness of cascade failures in
power grids. Scientific reports, 4, 2014.
[20] S. Pradhan and B. K. Chakrabarti. Failure properties of fiber bundle
models. Int’l. Journal of Modern Physics B, 17(29):5565–5581, 2003.
[21] M. Rosas-Casals and R. Solé. Analysis of major failures in european
power grid. International Journal of Electrical Power & Energy Systems,
33(3):805–808, 2011.
[22] C. Roy, S. Kundu, and S. Manna. Fiber bundle model with highly
disordered breaking thresholds. Physical Review E, 91(3):032103, 2015.
[23] S. Senju and Y. Toyoda. An approach to linear programming with 0-1
variables. Management Science, pages B196–B207, 1968.
[24] B. Stott, J. Jardim, and O. Alsac. Dc power flow revisited. Power
Systems, IEEE Transactions on, 24(3):1290–1300, Aug 2009.
[25] Y. Toyoda. A simplified algorithm for obtaining approximate solutions
to zero-one programming problems. Management Science, 21(12):1417–
1427, 1975.
[26] W.-X. Wang and G. Chen. Universal robustness characteristic of
weighted networks against cascading failure. Phys. Rev. E, 77, 2008.
[27] D. J. Watts. A simple model of global cascades on random networks.
Proceedings of the National Academy of Sciences, 99:5766–5771, 2002.
[28] Y. Yang, T. Nishikawa, and A. Motter. Small vulnerable sets determine
large network cascades in power grids. Science, 358(6365), 2017.
[29] O. Yağan. Robustness of power systems under a democratic-fiberbundle-like model. Phys. Rev. E, 91:062811, Jun 2015.
[30] O. Yağan and V. Gligor. Analysis of complex contagions in random
multiplex networks. Physical Review E, 86(3):036103, 2012.
[31] O. Yağan, D. Qian, J. Zhang, and D. Cochran. Information diffusion
in overlaying social-physical networks. In Proc. of the 46th Annual
Conference on Information Sciences and Systems (CISS), pages 1–6,
March 2012.
[32] O. Yağan, D. Qian, J. Zhang, and D. Cochran. Optimal allocation
of interconnecting links in cyber-physical systems: Interdependence,
cascading failures, and robustness. IEEE Transactions on Parallel and
Distributed Systems, 23(9):1708–1720, Sept 2012.
[33] O. Yağan, D. Qian, J. Zhang, and D. Cochran. Conjoining speeds
up information diffusion in overlaying social-physical networks. IEEE
Journal on Selected Areas in Communications, 31(6):1038–1048, 2013.
[34] G. Zhang, Z. Li, B. Zhang, and W. A. Halang. Understanding the
cascading failures in indian power grids with complex networks theory.
Physica A, 392(15):3273–3280, 2013.
IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. X, 2018
[35] Y. Zhang, A. Arenas, and O. Yağan. Cascading failures in interdependent
systems under a flow redistribution model. Phys. Rev. E, 97:022307, Feb
2018.
[36] Y. Zhang and O. Yağan. Optimizing the robustness of electrical power
systems against cascading failures. Scientific reports, 6:27625, 2016.
15
| 3cs.SY
|
An efficient method for evaluating BEM singular integrals on curved elements
with application in acoustic analysis
Junjie Rong, Lihua Wen∗, Jinyou Xiao
College of Astronautics, Northwestern Polytechnical University, Xi’an 710072, P. R. China
arXiv:1306.0282v1 [cs.CE] 3 Jun 2013
Abstract
The polar coordinate transformation (PCT) method has been extensively used to treat various singular integrals in
the boundary element method (BEM). However, the resultant integrands of the PCT tend to become nearly singular
when (1) the aspect ratio of the element is large or (2) the field point is closed to the element boundary; thus a large
number of quadrature points are needed to achieve a relatively high accuracy. In this paper, the first problem is
circumvented by using a conformal transformation so that the geometry of the curved physical element is preserved
in the transformed domain. The second problem is alleviated by using a sigmoidal transformation, which makes the
quadrature points more concentrated around the near singularity.
By combining the proposed two transformations with the Guiggiani’s method in [M. Guiggiani, et al. A general algorithm for the numerical solution of hypersingular boundary integral equations. ASME Journal of Applied
Mechanics, 59(1992), 604-614], one obtains an efficient and robust numerical method for computing the weakly-,
strongly- and hyper-singular integrals in high-order BEM with curved elements. Numerical integration results show
that, compared with the original PCT, the present method can reduce the number of quadrature points considerably,
for given accuracy. For further verification, the method is incorporated into a 2-order Nyström BEM code for solving
acoustic Burton-Miller boundary integral equation. It is shown that the method can retain the convergence rate of the
BEM with much less quadrature points than the existing PCT. The method is implemented in C language and freely
available.
Keywords: singular integrals, boundary element method, Nyström method, acoustics
1. Introduction
The boundary element method (BEM) has been a most important numerical method in science and engineering.
Its unique advantages includes the highly accurate solution on the boundary, the reduction of dimensionality, and
the incomparable superior in solving infinite or semi-infinite field problems, etc. Historically, relatively low-order
discretizations have been used with geometries modelled using first order elements and surface variables modelled to
zero or first order on those elements. Recently, however, there has been increasing interest in the use of high order
methods, in order to obtain extra digits of precision with comparatively small additional effort. Successful usage
are seen in acoustics [1], electromagnetics [2, 3], elasticity [4], aerodynamics [5], to name a few. One of the main
difficulties in using high order BEM is the lack of efficient methods for accurately evaluating various singular integrals
over curved elements. For the well-developed low order methods, there is no great difficulty since robust analytical
and numerical integration schemes are generally available for planar elements [6, 7, 9]. When concerned with high
order elements, however, a fully numerical method is required; various techniques have been proposed, for example,
the singularity subtraction [8], special purpose quadrature [10, 3, 11, 12] and the variable transformation methods
[13, 14, 15, 16].
∗ Corresponding
author
Email addresses: [email protected] (Junjie Rong), [email protected] (Lihua Wen), [email protected] (Jinyou Xiao)
Preprint submitted to Elsevier
June 4, 2013
In [8], Guiggiani et al proposed a unified formula (Guiggiani’s method for short) for treating various order of
singular integrals on curved elements. It is a singularity subtraction method, and has found extensive use in BEM. In
this method the singular parts are extracted from the integrand and treated analytically. The remaining parts are regular
so that conventional Gaussian quadrature can be employed, but the number of quadrature points needed would be large,
depending on the regularity of the associated integrands. The special purpose quadrature can be used to substantially
reduce the number of quadrature points, besides it is more robust and highly accurate. Most recently, James Bremer
[3] proposed such a method which can achieve machine precision. The problem with the special purpose quadrature
is that its construction somewhat complicated and time-consuming. Variable transformation method, also known as
singularity cancelation, eliminate the singularity of the integrand by the null Jacobian at the field point through a
proper change of variables. Although simple to implement, it is generally hard to be used in handling hypersingular
integrals.
In most existing integration methods, including those mentioned above, polar coordinates transformation (PCT)
always serves as a common base [8, 3, 12, 13, 16, 21]. It converses the surface integral into a double integral in radial
and angular directions. Many works have been done on dealing with the singularity in the radial direction; numerical
integration on the angular direction, however, still deserves more attention. In fact, after singularity cancelation or
subtraction, although the integrand may behaves very well in the radial direction, its behavior in the angular direction
would be much worst, so that too many quadrature points are needed. Especially when the field point lies close
to the boundary of the element, one can clearly observe near singularity of the integrand in the angular direction.
Unfortunately, this case is frequently encountered in using high order elements, especially for non-conform elements
as will be demonstrated by a Nyström BEM in this paper. Similar problems have been considered in work about
the nearly singular BEM integrals. Effective methods along this line are the subdivision method [17], the Hayami
transformation [18], the sigmoidal transformation [19], etc. For the singular BEM integrals, the angular transformation
has been used for computing weakly singular integrals over planar element [13].
Another problem of the PCT is how to find a proper planar domain to establish the polar coordinates. In usual
practice, the integral is carried out over standard reference domain (i.e., triangle in this paper) in intrinsic coordinates.
However, the reference triangle is independent of the shape of the curved element, and thus the distortion of the
element is brought into the integrand, which can cause near singularity in the angular direction. Consequently, the
performance of quadratures is highly sensitive to the shape of element. The above two problems in the angular
direction were considered and resolved by special purpose quadrature rule in [3]; whereas, as mentioned before, the
algorithm has its own overhead.
In this paper, two strategies are proposed to overcome two problems in the angular direction, respectively. First, a
conformal transformation is carried out to map the curved physical element onto a planar integration domain. Since it
is conformal at the field point, the resultant integration domain perseveres the shape of the curved element. Second, a
new sigmoidal transformation is introduced to alleviate the near singularity caused by the closeness of the field point
to the element boundary. The two proposed techniques, when combined with Guiggiani’s method, lead to a unified,
efficient and robust numerical integration methods for various singular integrals in high order BEM. As a byproduct
of the conformal transformation, the line integral in Guiggiani’s method can be evaluated in close form.
The paper is organized as follows. The Nyström BEM for acoustics and singular integrals encountered are reviewed in section 2. Section 3 describes the Guiggiani’s unified framework for treating various order of singular
integrals, with emphasis on the two reasons which render the poor performance of polar coordinates methods. Two
efficient transformations which represent the main contribution of this paper are proposed in section 4. Numerical
examples are given in section 5 to validate the efficiency and accuracy of the present methods. Section 6 concludes
the paper with some discussions.
2. Model problem: acoustic BIEs and Nyström discretization
The method presented in this paper will be used in solving the acoustic Burton-Miller BIE, which is briefly
recalled here. The BIE is solved by using the Nyström method with the domain boundary being partitioned into
curved quadratic elements. Various singular integrals that will be treated in the following sections are summarized.
2
2.1. Acoustic Burton-Miller formulation
The time harmonic acoustic waves in a homogenous and isotropic acoustic medium Ω is described by the following
Helmholtz equation
∇2 u(x) + k2 u(x) = 0, ∀x ∈ Ω,
(1)
where, ∇2 is the Laplace operator, u(x) is the sound pressure at the point x = (x1 , x2 , x3 ) in the physical coordinate
system, k = ω/c is the wavenumber, with ω being the angular frequency and c being the sound speed. For static case
k = 0, (1) becomes the Laplace equation. By using Green’s second theorem, the solution of Eq. (1) can be expressed
by integral representation
Z
Z
∂G(x, y)
u(x) +
G(x, y)q(y)dS (y) + uI (x), ∀x ∈ Ω,
(2)
u(y)dS (y) =
S ∂n(y)
S
where x is a field point and y is a source point on boundary S ; q(y) = ∂u(y)/∂n(y) is the normal gradient of sound
pressure; n(y) denotes the unit normal vector at the source point y. The incident wave uI (x) will not be presented for
radiation problems. The three dimensional fundamental solution G is given as
G(x, y) =
eikr
,
4πr
(3)
with r = |x − y| denoting the distance between the source and the field points.
Before presenting the BIEs it is convenient to introduce the associated single, double, adjoint and hypersingular
layer operators which are denoted by S, D, M and H, respectively; that is,
Z
G(x, y)q(y)dS (y),
(4a)
S q(y) =
ZS
∂G(x, y)
D u(y) =
u(y)dS (y),
(4b)
S ∂n(y)
Z
∂G(x, y)
q(y)dS (y),
(4c)
M q(y) =
S ∂n(x)
Z
∂2G(x, y)
H u(y) =
u(y)dS (y).
(4d)
S ∂n(x)∂n(y)
The operator S is weakly singular and the integral is well-defined, while the operators D and M are defined in Cauchy
principal value sense (CPV). The operator H, on the other hand, is hypersingular and unbounded as a map from the
space of smooth functions on S to itself. It should be interpreted in the Hadamard finite part sense (HFP). Denoting a
vanishing neighbourhood surrounding x by sε , the CPV and HFP integrals are those after extracting free terms from a
limiting process to make sε tends to zero in deriving BIEs [8].
Letting the field point x approach the boundary S in Eq. (2) leads to the CBIE
C(x)u(x) + Du(x) = Sq(x) + uI (x),
x ∈ S,
(5)
where, C(x) is the free term coefficient which equals to 1/2 on smooth boundary. By taking the normal derivative of
Eq. (2) and letting the field point x go to boundary S , one obtains the hypersingular BIE (HBIE)
C(x)q(x) + Hu(x) = Mq(x) + qI (x),
x ∈ S.
(6)
Both CBIE and HBIE can be applied to calculate the unknown boundary values of interior acoustic problems. For an
exterior problem, they have a different set of fictitious frequencies at which a unique solution can’t be obtained. However, Eqs. (5) and (6) will always have only one solution in common. Given this fact, the Burton-Miller formulation
which is a linear combination of Eqs. (5) and (6) (CHBIE) should yield a unique solution for all frequencies [20]
h
i
C(x)u(x) + (D +α H) u(y) − uI (x) = (S +α M) q(y) − α c(x)q(x) − qI (x) , x ∈ S ,
(7)
where, α is a coupling constant that can be chosen as i/k.
3
2.2. Nyström method and the singular integrals
In this paper, the Nyström method is used to discretize the BIEs. Let K be one of the integral operators in (4)
and K be the associated kernel function. Divide the problem into two regions, a region near and far from the field
point x. If the element locates in far field, Nyström method replaces the integral operator K with a summation under
a quadrature rule
Z
X
K(xi , y)u(y)dS (y)
ω j K(xi , y j )u(y j), ∆S ∈ S \Di ,
(8)
∆S
j
where, xi and y j are quadrature points, Di is the near field of xi , ω j is the jth weight over element ∆S . Such quadrature
rules can be obtained by mapping the Gaussian quadrature rules onto the parameterization of ∆S .
If the element locates in near field, however, the kernels exhibit singularities or even hypersingularities. As a
result, conventional quadratures fail to give correct results. In order to maintain high-order properties, the quadrature
weights are adjusted by a local correction procedure. Thus (8) becomes
Z
X
K(xi , y)u(y)dS (y)
ω̄ j (xi )u(y j), ∆S ∈ Di ,
(9)
∆S
j
where ω̄ j (xi ) represents the modified quadrature weights for specialized rule at the singularity. The local corrected
procedure is performed by approximating the unknown quantities using linear combination of polynomial basis functions which are defined on intrinsic coordinates (Fig.1). Modified weights are obtained by solving the linear system
Z
X
(n)
K(xi , y)φ(n) (y)dS (y), ∆S ∈ Di ,
(10)
ω̄ j φ (y j ) =
∆S
j
where φ(n) are polynomial basis functions. For 2th Nyström method used in this paper, φ(n) employed are of the form
φ(n) (ξ1 , ξ2 ) = ξ1p ξ2q ,
p + q ≤ 2,
(11)
where, p and q are integers, ξ1 and ξ2 denote local intrinsic coordinates.
The right hand side integrals in Eq. (10) are of crucial importance to the accuracy of the Nyström method. They
are referred to as singular integrals when xi lies on the element, and nearly singular integrals when xi is closed to but
not on the element. This paper deals with the integrals in the first case, the nearly singular integrals are treated via an
recursive subdivision quadrature. When concerned with the first three operators in equation (4), the integrals have a
weak singularity of r−1 , while the other operator H is hypersingular of order r−3 , as r → 0.
3. An unified framework for singular integrals
Since various order of singularities appear in Eq. (7) or many other BIEs, it is advantage to find a unified formula
to treat these integrals in the same framework. In such a way, these integrals can be implement in just one program
so that the computational cost will be reduced. By expansion of the singular integrands in polar coordinates, the three
types of singular integrals considered in this paper can be handled in a unified manner by using the formula proposed
by Guiggiani [8]. Despite of this advantage the method can be of low efficiency in practical usage; the very reasons
are then explained.
3.1. Polar coordinates transformation
Following a common practice in the BEM, the curved element is first mapped onto a region ∆ of standard shape
in the parameter plane (Fig. 1). In this case, the integral must be evaluated in Eq. (10) are of the form
Z
K(x, y(ξ))φ(ξ) |J(ξ)| dξ1 dξ2 ,
(12)
I=
∆
where |J| is the transformation Jacobian from global coordinates to the intrinsic coordinates,
#
"
∂y
∂y
∂y
∂y
,
|J(ξ)| =
.
×
J=
∂ξ1 ∂ξ2
∂ξ1 ∂ξ2
4
ξ2
1
2
u2
4x
0.5
5
(0.5, 0.5)
u1
6
1
3
0
0.5
1 ξ1
Figure 1: Description of curved triangle. Left hand image: curved triangle and a field point; right hand image:the reference triangle and typical
field points distribution of 2th order Nyström method in intrinsic coordinate system
Then, polar coordinates (ρ, θ) centered at ξ s (the image of x on intrinsic plane) are defined in the parameter space
(Fig. 2)
(s)
ξ1 = ξ1 + ρ cos θ
ξ2 = ξ(s) + ρ sin θ
2
so that dξ1 dξ2 = ρdρdθ. Due to piecewise smooth property of the boundary of ∆, the triangle is split into three
sub-triangles. The associated integral of Eq. (12) now becomes
I = lim
ρ(ε)→0
3 Z
X
j=1
θj
θ j−1
Z
ρ̂(θ)
ρ(ε)
K(ρ, θ)φ(ρ, θ) |J(ρ, θ)| ρdρdθ,
(13)
where ρ̂(θ) gives a parametrization of the boundary of ∆ in polar coordinates, (θ j−1 , θ j ) are three intervals on which
ρ̂(θ) is smooth. The limiting process means CPV or HFP integral mentioned before, although for well-defined integral
operator S, the limiting is not necessary. According to the geometrical relationship in Fig.2
hj
.
cos θ̄
ρ̂(θ) =
(14)
where, h j is the perpendicular distance from ξ s to jth side of the planar triangle and θ̄ is the angle from the perpendicular to a point ξ (Fig.2). In each sub-triangle, θ̄ equals to θ minus a constant.
ξ2
θ̄
θ̄ ≈ 90◦
θ
ρ̂(θ)
h
ξs
ξs
ξ1
Figure 2: Integration under polar coordinates. Left: Polar coordinates transformation; Right: near singularity as the field point approaches the
boundary
For singularity of order no more than 3, the integrand of (13), denoted by F(ρ, θ), can be expressed as series
expansion under polar coordinate ([23, 8])
F(ρ, θ) = K(ρ, θ)φ(ρ, θ) |J| ρdρ =
∞
X
f−2 (θ) f−1 (θ)
2
+
f
(θ)
+
ρ
f
(θ)
+
ρ
f
(θ)
+
·
·
·
=
ρi fi (θ),
+
0
1
2
ρ
ρ2
i=p
5
(15)
where, fi are just functions of θ, integer p is determined by the order of singularity,
0,
weakly singular;
−1,
strongly singular;
p=
−2, hyper-singular.
First, let’s consider the hypersingular integrals. Due to the appearance of the two terms with (i = −2, −1), the
integration of F(ρ, θ) must be performed in the HFP sense; for more details see Guiggiani’s work [8]. The resultant
formula for hypersingular integrals is given by
3 Z
X
I = I1 + I2 =
θ j−1
j=1
|
+
θj
Z
ρ̂(θ)
0
"
f−2 (θ) f−1 (θ)
F(ρ, θ) −
+
ρ
ρ2
{z
!#
dρdθ
I1
3 Z
X
θj
θ j−1
j=1
|
!
1
dθ .
f−1 (θ) ln ρ̂(θ) − f−2 (θ)
ρ̂(θ)
{z
}
}
(16)
I2
The strongly and weak singular integrals can be treated in a similar manner based on expansion (15). The resultant
computing formulas can also be written in the above form except that for strongly singular integrals f−2 = 0, and for
weak singular integrals f−2 = f−1 = 0 (thus I2 = 0). Therefore, formula (16) actually provide an unified approach to
evaluate singular integrals in BEM.
In (16) the double integral I1 and single integral I2 are all regular, thus one may conclude that it is sufficient to guarantee numerical accuracy in evaluating these two integrals, it would be nevertheless too expensive in practical usage,
especially when used in high order Nyström method considered in this paper. The difficulties and the corresponding
solutions will be presented in the following sections.
3.2. Difficulties in evaluating I1
It is the computation of I1 that accounts for the main cost for evaluating the various singular integrals. The
integrand of I1 can be approximated by polynomials in ρ; that is,
I1 =
3 Z
X
j=1
θj
θ j−1
Z
ρ̂(θ)
0
f0 (θ) + ρ f1 (θ) + ρ2 f2 (θ) + . . . dρdθ.
(17)
The number of terms in this approximation is determined by the kernel function, the order of basis function and the
flatness of the associated element. The relative flatness of element is a basic requirement in BEM in order to guarantee
the accuracy. Consequently, it appears that the integrand of I1 can be well approximated by low order polynomials
in ρ in solving many problems including Laplace, Helmholtz, elasticity and so forth, using quadratic elements. This
implies that low order Gaussian quadratures are sufficient for numerical integration in ρ.
In angular θ direction, however, two difficulties are frequently encountered which severely retard the convergence
rate of Gaussian quadratures. To show this, preforming integration in ρ in (17) one obtains
I1 =
3 Z
X
j=1
θj
θ j−1
!
1 2
ρ̂(θ) f0 (θ) + ρ̂ (θ) f1 (θ) + · · · dθ.
2
(18)
The difficulties in computation of I1 are caused by the near singularities of fi (θ) and ρ̂(θ).
3.2.1. Near singularity in fi (θ)
Functions fi (θ) can be expressed as (Appendix A)
fi (θ) =
6
f˜i (θ)
,
Aα (θ)
(19)
where f˜i are regular trigonometric functions and α is integer determined by the subscript i. Function A(θ) is depend
on the shape of the element and parametric coordinate system (see [8] for a definition of A(θ)). Specifically, let u1 , u2
be the two column vectors of the Jacobian matrix J which spans the space tangent to the element at the point x, i.e.
∂y
∂ξ1
u1 =
and u2 =
y=x
∂y
∂ξ2
.
(20)
y=x
Then A(θ) is given by
q
A(θ) = |u1 |2 cos2 θ + u1 · u2 sin 2θ + |u2 |2 sin2 θ
r
1
1
= u1 · u2 sin 2θ + |u1 |2 − |u2 |2 cos 2θ + |u1 |2 + |u2 |2
2
2
r
1
=
|u1 |2 + |u2 |2 µ sin(2θ + ϕ) + 1 .
2
If we let λ = |u1 | / |u2 | and cos γ =
u1 ·u2
|u1 ||u2 | ,
(21)
then
ϕ = arctan
and
µ=
s
1−
λ2 − 1
,
2λ cos γ
4 sin2 γ
< 1.
(λ + λ−1 )2
(22)
It can be seen from (21) that, if µ → 1, there exist two points θ ∈ [0, 2π] such that A(θ) → 0. Thus function fi
tends to be nearly singular. The circumstance µ → 1 occurs in two cases according to Eq. (22): (1) the aspect ratio
of the element is large, i.e., λ approaches 0 or ∞; (2) peak or big obtuse corner appear in the element which lead to
sin γ → 0. Both of these two cases indicate a distorted shape of the element.
To illustrate the influence of the element aspect ratio on the behavior of function 1/A(θ) which represents the
smoothness of fi , consider the curved element in figure 7; see section 5.1 for more detailed descriptions. The aspect
ratio of the element is controlled by s; larger s implies more distorted shape. Let b be the singular point. The plots of
1/A(θ) in the interval corresponding to the sub-triangle (2-b-3) in figure 7 with various s are exhibited in Fig. 3. It is
clear that 1/A(θ) varies more acutely as the increase of the aspect ratio.
The above analysis shows how element shape affects fi and thus the integrand in (18). In an intuitionistic manner,
the near singularity in fi is induced by the fact that the integration plane ∆ in intrinsic coordinate system is independent
on the shape of element, as a result the distortion is brought into the integrand. One can suppose that if the integration
is performed over another planar triangle which reflects the distortion of the element, the near singularity should be
eliminated. This is the main idea of the our new method in section 4.1.
3.2.2. Near singularity in ρ̂(θ)
In addition to the near singularity in fi , another obstacle retards the convergence of the numerical quadrature for
(18) is the near singularity in ρ̂(θ), which can be clearly seen from (14). When the field point x lies near to the
boundary of the element, the restrict of θ̄ approaches ±π/2, thus the denominator cos θ̄ is close to 0 at the two ends of
the interval (θ j−1 , θ j ) (see Fig.2). The effect of the near singularity in ρ̂(θ) on the total behavior of the integrand F(ρ, θ)
is demonstrated by the left plots of Fig. 6 (a) and (b), where the integrand have large peaks near the two ends of the
interval.
Unfortunately, the situation that causes the near singularity in ρ̂(θ) is ubiquitous in using high order Nyström
method. See Fig. 3.1 for a typical distribution of field points in 2th order Nyström method.
4. Efficient transformation methods
How to effectively resolve the above mentioned two difficulties is crucial to the accurate evaluation of the BEM
singular integrals. More recently, special quadratures are constructed for this purpose in [3]. Although accurate and
7
1.5
s=0.5
s=2.0
s=4.0
s=10
1/A(θ)
1
0.5
0
−0.5
0
0.5
θ (rad)
1
1.5
2
Figure 3: The plot of 1/A(θ) for various value of s. The aspect ratio of the associated element increased with s. The curved element is described in
section 5.1, with field point (b)
robust numerical results are reported, it is noticed that the abscissas and weights of the special quadratures are depend
on the singular (field) point as well as the element on which the integral is defined. The construction of the quadrature
for integral can be rather complicated and time-consuming.
In this section, however, we propose a more simple yet efficient method. First, the intrinsic coordinates (ξ1 , ξ2 )
is transformed onto a new system which results in constant A(θ). An additional benefit of constant A(θ) is that the
line integral I2 can be evaluated in closed form; see section 4.3. Then the sigmoidal transformation is introduced to
alleviate the near singularity caused by ρ̂(θ).
4.1. Conformal transformation
First, the near singularity caused by A(θ) in fi is considered and resolved. The idea is to introduce a new transformation under which A(θ) becomes constant. It can be seen from (21) that if
u1 · u2 = 0
and
|u1 | = |u2 | ,
(23)
then A(θ) will be constant, i.e. A(θ) = |u1 | = |u2 |.
Relations (23) indicate a conformal mapping from the element to the plane at field point x, i.e., both angle and
the shape of the infinitesimal neighborhood at x are preserved. However, this condition generally can not be satisfied
under intrinsic coordinates. In [18] Hayami proposed to connect the three corners of the curved element to establish a
planar triangle which preserves the shape of the element. This operation, nevertheless, can not satisfy condition (23)
exactly.
Here, we propose a transformation in which the curved physical element is mapped to a triangle ∆¯ in plane
(2)
(2)
(η1 , η2 ) as shown in Fig. 4. The coordinates of the three corners of ∆¯ are (0, 0), (η(2)
1 , η2 ) and (1, 0), with η2 > 0.
The transformation from system (ξ1 , ξ2 ) to (η1 , η2 ) can be realized by linear interpolation
η=
3
X
φ̄i (ξ1 , ξ2 )η(i) ,
(24)
i=1
¯ φ̄i are linear interpolating functions
where, η(i) are coordinates of the three corners of ∆,
φ̄1 = 1 − ξ1 − ξ2 ,
φ̄2 = ξ2 ,
φ̄3 = ξ1 .
8
(25)
ξ2
1
0.5
2
x3
5
T
1 ξ1
T−1
u1
6
1
(0.5, 0.5)
0.5
0
J
u2
4x
ξ (s)
(2)
(2)
(η1 , η2 )
η2
3
θ
x2
x1
o
η
(s)
η1
Figure 4: Transformations between coordinate systems. Left: curved element in global coordinates; Top right: intrinsic coordinates and the
reference element; Bottom right: the parametric plane to perform integral
Plugging (25) into (24) yields
η = Tξ,
(26)
where, T is the transformation matrix
"
#
1 η(2)
1
T=
0 η(2)
2
−1
and T
1 − η(2)
1
(2)
.
η2
=
1
0
(2)
η
(27)
2
The Jacobian matrix at x from physical coordinates to (η1 , η2 ) coordinates, denoted by [ū1 ū2 ], can be written as
i
i h
h
−η(2)
u2
1 u1
ū1 ū2 = u1 u2 T−1 = u1
(28)
+
(2)
(2) .
η
η
2
2
In order to obtain constant A(θ), ū1 and ū2 have to satisfy condition (23), thus one has
η(2)
1 =
cos γ
,
λ
η(2)
2 =
sin γ
,
λ
(29)
where, λ and γ are defined in section 3.2.
By using relation (26) the integral (12) can be transformed onto (η1 , η2 ) plane. One can then employ the polar
coordinate transform in Section 3.1. The origin of the polar system is set as the image of x on (η1 , η2 ) plane. Then
integral I1 becomes
!#
3 Z θ j Z ρ̂(θ) "
X
f−2 (θ) f−1 (θ)
T−1 ρdρdθ,
(30)
F(ρ, θ) −
I1 =
+
2
ρ
ρ
j=1 θ j−1 0
in which the near singularity caused by A(θ) has been successfully removed.
Numerical results show that the above transformation can always improve the numerical integration regardless the
shapes of the elements being regular or irregular.
4.2. Sigmoidal transformation
When the field point x approaches the element boundary, function ρ̂(θ) and thus the integral I1 becomes nearly
singular. Here a sigmoidal transformation is introduced to alleviate this problem. The sigmoidal transformation
was first proposed to calculate the singular and nearly singular integrals in two dimensional BEM [22]. Recently,
9
1
0.5
1
Gaussian points
clustered points
Gaussian points
clustered points
0.5
θ̄
0
θ̄
0
−0.5
−0.5
−1
−1
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
w
0.6
0.8
1
w
Figure 5: Sigmoidal transformation with m = 3 to cluster 10 Gaussian quadrature points (x-axis) to the endpoints of the interval. The sub-triangle
(2-b-3) in Fig. 7 is considered, whose θ̄ ∈ [−1.466, 1.005]. Left: A naive use of sigmoidal transformation; Right: Sigmoidal transformation in this
paper.
this approach was adopted as an angular transformation in dealing with nearly singular integrals in 3D BEM [19].
One should notice that the angular transformation proposed by Khayat and Wilton [13] can be alternatively used,
but our numerical experience indicates a better overall performance by using sigmoidal transformation, especially in
hypersingular case.
A sigmoidal transformation can be thought of as a mapping of the interval [0, 1] onto itself whose graph is Sshaped. It has the effect of translating a grid of evenly spaced points on [0, 1] onto a non-uniform grid with the node
points clustered at the endpoints. A typical sigmoidal transformation is given by [22]
σ(w) =
wm
wm
,
+ (1 − w)m
σ, w ∈ [0, 1], m ≥ 1.
(31)
Consider one sub-triangle in figure 2 in which θ̄ ∈ [θ̄ j−1 , θ̄ j ], j = 1, 2, 3. A naive use of the above transformation can
be (see Fig. 5)
θ̄ − θ̄ j−1
= σ(w).
(32)
θ̄ j − θ̄ j−1
However, this would be of low efficiency. Since the integral I1 tends to be nearly singular (due to ρ̂(θ)) only when
θ̄ j−1 or θ̄ j is close to right angle or both, it is thus more reasonable to cluster the quadrature nodes according to the
discrepancy of θ̄ j−1 and θ̄ j to right angles, respectively. In addition there are cases where both θ̄ j−1 and θ̄ j are not very
close to right angles and thus the near singularity is not severe. For these cases, one can use the Gauss quadratures
directly without any transformation. Nevertheless, numerical examples in this paper show that the modification put
forward below can always achieve more accurate results.
The transformation used in this paper is based on the fact that the singularity occurs at θ̄ = ±π/2 and is given by
1
π
θ̄ +
= σ(z), z = (z j − z j−1 )w + z j−1 , w ∈ [0, 1],
(33)
π
2
where, z(θ̄ j−1 ) and z(θ̄ j ) are the values of z corresponding to θ̄ j−1 and θ̄ j which can be easily obtained by (31). Obviously, both transformation (33) and (32) project [θ̄ j−1 , θ̄ j ] onto [0, 1], but (33) can cluster the quadrature points
adaptively according to closeness of θ̄ to right angle, which is shown in Fig. 5. The left hand image demonstrate the
part of sigmoidal transformation used in dealing with the sub-triangle (2-b-3) in Fig. 7. Compared with full sigmoidal
transformation shown left, the distribution of quadrature points are more reasonable since they are more clustered on
θ̄ j−1 which is more closely to right angle. Besides, the points are not too away from the center. A comparison between
the original sigmoidal transformation and the modification is illustrated in table 1. Without any doubt, the modified
transformation is more powerful when dealing with the integrals in this paper.
The degree of cluster is impacted by parameter m, see [22]. Although numerical examples indicates not too much
difference in number of quadrature points when m takes the value between 2 and 3, it should be point out that the
optimum value of m is 3 for weakly singular case and 2 for hypesingular case.
10
Table 1: Comparison of relative errors using different methods in k = 0 case. The curved element is show in Fig. 7, with field point (b). “n” denotes
the number of quadrature points used in angular direction.
n
5
8
10
single
Eq. (32)
3.19e-3
1.16e-5
3.06e-6
Guiggiani
6.24e-4
1.02e-3
3.20e-4
0.8
Present
9.83e-5
6.44e-7
1.26e-8
Present
4.09e-05
1.49e-07
3.71e-09
0.3
integand
8th degree
integrand
8th degree
0.25
0.6
0.2
integrand
integrand
hyper
Eq. (32)
1.57e-04
1.87e-06
5.22e-07
Guiggiani
1.80e-03
1.69e-04
2.39e-05
Norm of residuals = 0.21819
0.4
0.15
Norm of residuals = 0.00738
0.1
0.05
0.2
0
0
-0.5
0
0.5
1
T (rad)
1.5
-0.05
2
-0.5
0
0.5
T (rad)
1
1.5
2
(a) Laplace single kernel, m = 3
4
integrand
8th degree
integrand
8th degree
0.6
0.5
2
integrand
integrand
3
Norm of residuals = 0.83882
1
Norm of residuals = 0.00946
0.4
0.3
0.2
0.1
0
0
-1
-0.5
0
0.5
1
T (rad)
1.5
-0.1
2
-0.5
0
0.5
T (rad)
1
1.5
2
(b) Laplace hypersingular kernel, m = 2
Figure 6: Integrand of I1 in angular direction corresponding to sub-triangle (2-b-3) in Fig. 7;the basis function is selected as ξ22 . 8th order
polynomial is plotted to fit the image. Left: Image of the original Guiggiani’s method; Right: Image after transformations proposed in this paper.
Finally, the integral has the formula
I1 =
3 Z
X
j=1
0
1
Z
ρ̂(θ)
0
"
F(ρ, w) −
f−2 (w) f−1 (w)
+
ρ
ρ2
!#
θ̄′j (w) T−1 ρdρdw,
(34)
where, θ̄ j (w) denotes transformation (33) on interval [θ̄ j−1 , θ̄ j ].
The effects of the two transformations in sub-sections 4.1 and 4.2 are demonstrated in Fig. 6. It can be seen that
after the transformations, the integrands become more regular and can be well approximated by polynomials of order
8, which means that low order Gaussian quadratures can achieve accurate results.
4.3. Analytical integration of the one-dimensional integrals
P
Consider the evaluation of the regular one dimensional integral I2 in Eq. (16), i.e., I2 = 3j=1 I2( j) and
( j)
I2 =
Z
θj
θ j−1
f−1 (θ) ln ρ̂(θ) − f−2 (θ)
!
1
dθ.
ρ̂(θ)
(35)
One notes that the integrand of I2 suffers from the same difficulties as that of the double integral I1 as stated
in section 3.2. The above proposed transformations can also be used to promote the efficiency of the numerical
quadratures.
11
In this section close form expression for I2 is derived. It is attributed to the fact that after the conformal transformation in section 4.1 the denominators of fi , i.e. A(θ), become constant. For Burton-Miller equation, it is readily
to see from (A.7) that f−2 (θ) is constant on each sub-triangle thanks to the constant A(θ), and that f−1 (θ) is typically
made up of combinations of elementary trigonometric functions of θ [8]
f−1 (θ) =c1 cos3 θ + c2 cos2 θ sin θ + c3 cos θ sin2 θ + c4 sin3 θ
+ d1 cos θ + d2 sin θ.
(36)
The coefficients ci and di are given in (A.8). Here we notice that the expressions in (A.8) are also applicable for Laplace
equations. For other problems, such as elasticity, Stokes flow, the expressions of f−2 and f−1 are more complicated
and can be derived analogously as in Appendix A.
To further simplify the derivation, the coordinate system (η1 , η2 ) is rotated so that the axis oη1 is parallel to the
¯ Thus θ = θ̄ over each sub-triangle. By substituting ρ̂(θ), f−2 and f−1 into (35), the
perpendicular line of the side of ∆.
integral with f−2 becomes
Z θj
Z
f ( j) θ j
1
f−2 (θ)
dθ = −2
cos θdθ.
ρ̂(θ)
h j θ j−1
θ j−1
The integral with f−1 consists of terms
ln h j
Z
θj
cos p θ sinq θdθ,
p+q=3
or
p + q = 1,
(37)
θ j−1
and
Z
I¯i =
θj
cos p θ sinq θ ln cos θdt,
p+q=3
or
p + q = 1.
(38)
θ j−1
The explicit expressions of integral (37) can be easily obtained and are omitted. Here the expression of I¯i is derived.
By changing the integral variable
t = sin θ
√
cos θ = 1 − t2
(39)
√
dt = 1 − t2 dθ,
one obtains
I¯1 =
Z
θj
3
cos θ ln cos θdθ =
θ j−1
Z
tj
t j−1
√
√
t 1 − t2 ln 1 − t2 dt = I˜1 − I˜3 ,
√
t − t3
ln 1 − t2 dt = I˜2 − I˜4 ,
√
θ j−1
t j−1
1 − t2
Z tj
Z θj
√
t2 ln 1 − t2 dt = I˜3 ,
cos θ sin2 θ ln cos θdθ =
I¯3 =
I¯2 =
Z
I¯4 =
Z
θj
cos2 θ sin θ ln cos θdθ =
θ j−1
θj
Z
sin3 θ ln cos θdθ =
I¯6 =
θj
cos θ ln cos θdθ =
θj
θ j−1
Z
tj
t j−1
θ j−1
Z
tj
t j−1
θ j−1
I¯5 =
Z
sin θ ln cos θdθ =
Z
Z
tj
t j−1
tj
t j−1
t3
ln
√
1 − t2 dt = I˜4 ,
√
1 − t2
√
ln 1 − t2 dt = I˜1 ,
√
t
ln 1 − t2 dt = I˜2 ,
√
1 − t2
12
(40)
where,
t
√
√
t+1 j
,
1 − t2 dt = t ln 1 − t2 − 1 + ln √
t j−1
1 − t2 t j−1
√
Z tj
√
√
tj
t ln 1 − t2
dt = 1 − t2 1 − ln 1 − t2
=
,
√
t j−1
t j−1
1 − t2
Z tj
1
√
√
1
1 1 + t tj
,
t2 ln 1 − t2 dt = t3 3 ln 1 − t2 − 1 − t + ln
=
9
3
6 1 − t t j−1
t j−1
"
# tj
Z tj 3 √
√
√
1
1
t ln 1 − t2
2
2
2
2
(8 + t ) − (2 + t ) ln 1 − t
dt = 1 − t
.
=
√
9
3
t j−1
1 − t2
t j−1
I˜1 =
I˜2
I˜3
I˜4
Z
tj
ln
(41)
5. Numerical examples
Figure 7: Curved element and the field points in example 1
Two numerical examples are provided to demonstrate the efficiency and robustness of the method in this paper.
The first one is used to illustrate the superior of the method in this paper to the method in 3.1, i.e., the classic polar
coordinates transformation and Guiggiani’s method [8], using a sample curved element. The second one is used to
verify the accuracy and convergence of the method when employed in solving Burton-Miller equation. The method is
implemented in C language. The numerical integration code are freely available.
The focus of this paper is to improve the performance of the quadrature in angular direction, while in radial
direction, as explained in section 3.2, the integrand behaves well so that only a small number of quadrature points
are sufficient. In section 5.1 6 points are used in every test in radial direction. However, when implemented in BEM
programs, fewer points are needed because as the subdivision of the model the diameter of each element will decrease
so that fewer terms are needed in expansion (15) to approximate the integrand. As a result, the rate of convergence of
BEM will not be affected. For the 2th order Nyström BEM in section 5.2, we use 3 points in polar direction.
5.1. Accuracy of the method on a curved element
The accuracy, effectiveness and robustness of the method in this paper are investigated. The integrals is performed
over a curved element extracted from a cylinder surface. The size of the cylinder and the element are given in Fig.7.
Distance s in Fig. 7 is designed to change the aspect ratio of the element. s = 0.5 represents a quite regular element;
the quality of the element becomes bad as the increase of s. Four positions for the field point x are considered, which
correspond to intrinsic coordinates (a) ξ = (0.3, 0.3), (b) ξ = (0.1, 0.8), (c) ξ = (0.45, 0.45), (d) ξ = (0.64, 0.31),
respectively. One should note that points (b) and (c) match approximately with the nodes of the 2th order Nyström
13
Table 2: methods to compare
methods
Guiggiani
Gui+sig
Present
Present+a
description
Guiggiani’s method in section 3.1
Guiggiani’s method with a sigmoidal transformation applied in angular direciton
Method in this paper, i.e. Guiggiani’s method with the two transformations in section 4
Similar to “Present”, except that the line integral is calculated analytically
discretization, while (d) match with that of the 3th order. In addition, for the rather centered point (a) the original
Guiggiani’s method in section 3.1 are believed to work well.
Since no analytical solutions are available, the (relative) error, given by
relative error =
|Icalc − Iref |
|Iref |
is evaluated by comparison with the results Iref of the Guiggiani’s methods using Gauss quadratures with 256 quadrature points in angular direction and 32 points in radial direction. Icalc denotes the result calculated by the various
methods. The single layer operator S and the hypersingular operator H are tested. The parameter m in sigmoidal
transformation takes the value of 3 in weakly singular case and 2 in hypersingular case. The basis function φ is
selected as the second order monomial ξ22 .
A description of the methods tested are listed in table 2. The method in this paper are indicated by “Present” in
which the line integral I2 in (16) is computed by Gauss quadratures; while in “Present+a” the line integral is computed
by using the closed formulations in section 4.3. Note that “Present+a” only validates for hypersingular integrals.
5.1.1. Results for regular element
First the performance of the present method for regular element are verified; for this purpose, we let s = 0.5 and
the wave number k = 0 (Laplace kernels).
Convergence behavior of the original methods and the present method for the four field points are illustrated in
Fig.8. The results of the present method are marked by triangles or circles. The results for single layer kernel are
plotted with dashed line, while real lines are for hypersingular integrals. A comparison between case (a) and the other
three cases shows that the original Guiggiani’s method tends to converge slowly when the field point is closed to the
element boundary; while the present method can achieve almost the same fast convergence. Even in regular case (a)
in which the original method performs better than other three cases, the benefit of the present method is also obvious.
The effect of each transformation scheme in section 4 are studied and exhibited in table 3. Both field points (b)
and (d) are tested. For the weak singular (single layer) integral, the accuracy can be improved by only using the
sigmoidal transformation; whereas the present method, say the combination of the two transformations in section 4,
can further promote the accuracy significantly. One notice that for the hypersingular case the sole use of the sigmoidal
transformation can not improve the accuracy. However, this problem can be completely overcome by using the
parametric coordinates transformation. The analytical formulas for the line integrals can be better than the numerical
quadratures.
Now consider the Helmholtz kernels. The results are similar to that of the Laplace kernels, so only the results for
field point (b) are given in table 4. The wavenumber k is selected as 2.0 so that the wavelength is about 3 times size of
the triangular element. In practical BEM implementation such a mesh is rather coarse. It is obvious that the present
method can greatly improve the accuracy.
5.1.2. Results for distorted element
As have been mentioned in section 4.1, the present method should be robust for elements with high aspect ratio;
here this is validated. The integrals are evaluated over elements with different s and k = 0. Point (a) is taken to be
the field point. Both the original and present methods are used to compute the integrals with relative error 10−8 . The
needed number of Gauss quadrature points in angular direction are listed in table 5. It is seen that this number of the
original method increases drastically as the increase of s, which clearly indicates that the performance of the original
method depend heavily on the shape of the element. On the contrary, the present method behaves more robust with the
14
0
0
10
10
Guiggiani (single)
Present (single)
Guiggiani (hyper)
Present+a (hyper)
−2
10
10
−4
10
relative error
relative error
−4
−6
10
−8
10
−10
−6
10
−8
10
10
−12
2
10
−10
10
10
Guiggiani (single)
Present (single)
Guiggiani (hyper)
Present+a (hyper)
−2
−12
4
6
8
10
12
14
16
18
number of integration points in angular direction
10
20
2
4
6
8
10
12
14
16
18
number of integration points in angular direction
(a) field point (a)
(b) field point (b)
0
0
10
10
Guiggiani (single)
Present (single)
Guiggiani (hyper)
Present+a (hyper)
−2
10
−2
10
−4
relative error
relative error
−4
10
−6
10
−8
10
−10
−6
10
−8
10
10
−12
2
10
−10
10
10
20
Guiggiani (single)
Present (single)
Guiggiani (hyper)
Present+a (hyper)
−12
4
6
8
10
12
14
16
18
number of integration points in angular direction
10
20
2
4
6
8
10
12
14
16
18
number of integration points in angular direction
(c) field point (c)
(d) field point (d)
Figure 8: Convergence comparison for the k = 0 case.
Table 3: Relative errors for different methods. “n” denotes the number of Gauss quadrature points in angular direction.
n
Guiggiani
single
Gui+sig
Present
5
8
10
12
6.24e-04
1.02e-03
3.20e-04
5.65e-05
1.22e-04
2.83e-05
3.08e-07
8.91e-08
9.83e-05
6.44e-07
1.26e-08
2.95e-11
5
8
10
12
4.05e-02
1.89e-02
1.07e-02
5.83e-03
1.18e-03
3.20e-05
5.85e-07
1.50e-08
6.06e-04
6.35e-06
2.23e-07
5.16e-09
Guiggiani
field point (b)
1.81e-03
1.70e-04
2.38e-05
3.01e-06
field point (d)
5.33e-03
2.35e-03
8.11e-04
3.45e-04
hyper
Gui+sig
Present
Present+a
5.51e-03
2.62e-04
2.72e-05
1.12e-05
1.28e-03
5.41e-06
5.64e-08
4.20e-10
4.1e-05
1.49e-07
3.71e-09
6.70e-10
6.28e-03
1.48e-04
2.65e-04
3.15e-05
1.11e-02
7.32e-05
1.93e-06
5.34e-08
1.68e-04
4.08e-06
2.92e-07
1.57e-08
Table 4: Relative errors in dynamic case with k = 2.0, the field point is selected as (b).
n
5
8
10
12
single
Guiggiani Present
2.52e-03 2.00e-04
1.16e-03 2.44e-06
4.09e-04 1.09e-07
1.26e-04 4.66e-09
15
hyper
Guiggiani Present+a
1.92e-03
3.64e-04
2.08e-04
3.22e-07
2.04e-05
6.93e-09
4.36e-06
4.71e-10
20
Table 5: Number of integration points needed in angular direction to make the relative error under 10−8
s
0.5
1.5
2.0
4.0
10.0
single
Guiggiani Present
15
8
19
9
22
9
42
11
82
13
Guiggiani
18
27
31
62
142
hyper
Present
11
12
13
14
15
Present+a
11
11
12
13
16
Figure 9: Geometry model
change of s. For element with aspect ratio larger than 10, high accurate results can still be achieved with a moderate
increase of quadrature points.
5.2. An exterior sound radiation problems
The overall performance of the present integration method is tested by incorporating it into a Nyström BEM code
for solving the Burton-Miller equation. The geometry of the problem consists of three sections of cylinders, as shown
in Fig. 9. The origin of coordinate system lies on the middle point of the symmetry axis of the cylinder. All the
surfaces of the cylinder vibrate with given velocity q(x) = ∂u/∂x (Neumann problem), where u is chosen to be the
fundamental solution u(x) = G(x, y) in (3) with y = (0, 0, 0).
The Nyström BEM with 2th order basis functions is used to discretize Burton-Miller equation. The wave number
k = 2.5. The surface is meshed by quadratic triangular element and refined four times; the finest mesh consists of
1638 elements. In evaluating the singular integrals, both the original Guiggiani’s method and the present method are
tested. A common m = 2.5 is used in the sigmoidal transformation for all the singular integrals, and 3-point Gauss
quadrature are used in radial direction.
Three different quadratures in the angular direction are used and compared, that is, the present method using 4
points and the Guiggiani’s method with 10 and 15 points. The L2 relative errors of the boundary values of u(x) are
demonstrated in Fig. 10. It is shown that the accuracy of BEM is retarded when using the Guiggiani’s method with
10 points. For the Guiggiani’s method with 15 points, the convergence of the BEM tends to slow down with further
refinement of the mesh. The present method with 4 points, however, can achieve the theoretical converge rate.
6. Conclusions
Highly efficient methods with high accuracy and low computational cost are crucial for high-order boundary
element analysis. In [8], Guiggiani proposed an unified framework to treat the singular integrals of various orders
in BEM. It is based on the polar coordinate transformation which has been extensively used in dealing with BEM
singular integrals. However, the performance of the polar coordinate transformation deteriorates when the field point
is close to the element boundary or the aspect ratio of the element becomes large.
16
−2
O(N )
Present(n=4)
Guiggiani(n=10)
Guiggiani(n=15)
−3
2
L error
10
−4
10
−5
10
3
4
10
10
Number of unkowns
Figure 10: L2 -errors of u in example 5.2.
In this paper, first, a conformal transformation is introduced to circumvent the near singularity caused by large
aspect ratio of element. This transformation maps a curved physical element onto a planar triangle. Since it is
conformal at the field point, the resultant integration domain (planar triangle) perseveres the shape of the curved
element. Then, a sigmoidal transformation is applied to alleviate the near singularity due to the closeness of the
field point to the element boundary. The rationale behind this is that the sigmoidal transformation can cluster the
quadrature points to area of near singularity. The combination of the two transformations can effectively alleviate
the two problems of the existing polar coordinate transformation method, and thus leads to considerable reduction of
quadrature points in angular direction.
The efficiency and robustness of the present method are illustrated by various singular integrals on a curved
quadratic element which is typical in high order Nyström BEM. It is shown that highly accurate results with relative
error 10−8 can be achieved with 10 quadrature points in angular direction. Moreover, the method is more stable with
the change of element aspect ratio. For further verification, the method is applied to a 2-order Nyström BEM for
solving acoustic Burton-Miller equation. Theoretical convergence rate of the BEM can be retained with much less
quadrature points than the existing quadrature methods.
Acknowledgements
This work was supported by National Science Foundations of China under Grants 11074201 and 11102154 and
Funds for Doctor Station from the Chinese Ministry of Education under Grants 20106102120009 and 20116102110006.
Appendix A. Coefficients in Eq. (36)
Consider a general integrand in BEM in polar coordinates
F(ρ, θ) =
ρF̄(ρ, θ)
,
rβ
(A.1)
where, β is the order of singularity, F̄(ρ, θ) is a regular function. There holds
∞
X
1
= ρ−β
S ν−β (θ)ρν ,
β
r
ν=0
17
(A.2)
F̄(ρ, θ) =
∞
X
av (θ)ρν ,
(A.3)
ν=0
where, S ν (θ) is given by ([23], Theorem 4)
S ν−β (θ) = A−β−2ν (θ)g3ν (θ),
(A.4)
g3ν (θ) are homogeneous trigonometric polynomials of order 3v.
For hypersingular integrand, β = 3, thus,
f−2 (θ) = S −3 (θ)a0 (θ),
f−1 (θ) = S −2 (θ)a0 (θ) + S −3 (θ)a1 (θ).
(A.5)
See [8] for the expressions of S −2 (θ) and S −3 (θ).
For hypersingular kernel of Burton-Miller equation in this paper, the expressions ai can be obtained similarly with
[8]. Let Jk be the kth component of the vector
∂y
∂y
×
.
∂η1 ∂η2
Then Jk can be expanded as
#
"
∂Jk
∂Jk
cos θ +
sin θ = Jk0 + ρJk1 (θ) + O(ρ2 ).
Jk = Jk0 + ρ
∂η1 η=ηs
∂η2 η=ηs
The basis function φ can be expanded analogously as
"
∂φ
∂φ
φ = φ0 + ρ
cos θ +
∂η1 η=ηs
∂η2
#
sin θ = φ0 + ρφ1 (θ) + O(ρ2 ).
η=ηs
Then, a0 and a1 (θ) in Eq. (A.5) can be expressed as (repeated indicies imply summation)
ni (x)
Ji0 φ0 ,
4π
ni (x)
Ji1 (θ)φ0 + Ji0 φ1 (θ) .
a1 (θ) =
4π
Substituting Eq. (A.6) into Eq. (A.5), yields
a0 =
(A.6)
f−1 (θ) =c1 cos3 θ + c2 cos2 θ sin θ + c3 cos θ sin2 θ + c4 sin3 θ
+ d1 cos θ + d2 sin θ,
ni (x)
f−2 (θ) =
Ji0 φ0 .
4πA3
with
3ni (x)Ji0 φ0 ∂yk ∂2 yk
,
8πA5 ∂η1 ∂η21
3ni (x)Ji0 φ0 ∂yk ∂2 yk
1 ∂yk ∂2 yk
,
c2 = −
+
∂η1 ∂η1 ∂η2 2 ∂η2 ∂η21
4πA5
1 ∂yk ∂2 yk
3ni (x)Ji0 φ0 ∂yk ∂2 yk
,
+
c3 = −
∂η2 ∂η1 ∂η2 2 ∂η1 ∂η22
4πA5
(A.7)
c1 = −
3ni (x)Ji0 φ0 ∂yk ∂2 yk
c4 = −
,
8πA5 ∂η2 ∂η22
!
∂φ
ni (x) ∂Ji
,
φ
+
J
0
i0
∂η1
4πA3 ∂η1
!
∂φ
ni (x) ∂Ji
.
φ
+
J
d2 =
0
i0
∂η2
4πA3 ∂η2
Note that all the derivatives are evaluated at the field point x.
d1 =
18
(A.8)
References
[1] L. F. Canino, J. J. Ottusch, M. A. Stalzer, et al. Numerical solution of the Helmhotz equation in 2D and 3D using a high-order Nyström
discretization. Journal of computational physics, 146 (1998), 627-633.
[2] S. Rjasanow, L. Weggler. ACA accelerated high order BEM for maxwell probelms. Computational mechanics, 51 (2013), 431-441.
[3] J. Bremer, Z. Gimbutas. A Nyström method for weakly singular integral operators on surfaces. Journal of computational physics, 231 (2012),
4885-4903.
[4] X. W. Gao, T. G. Davies. Boundary Element Programming in Mechanics. Cambridge University Press (ISBN: 052177359-8), 2002.
[5] D. J. Willis, J. Peraire, J. K. White. A quadratic basis function, quadratic geometry, high order panel method. In 44th AIAA Aerospace sciences
meeting, number AIAA-2006-1253, 2006.
[6] H. J. Wu, Y. J. Liu, W. K. Jiang. A low frequency fast multiploe boundary element method based on analytical integration of the hypersingular
integral for 3D acoustic problems. Engineering analysis with boundary elements, 37 (2013), 309-318.
[7] S. N. Fata. Explicit expressions for 3D boundary integrals in potential theory. International journal for numerical methods in engineering, 78
(2009), 32-47.
[8] M. Guiggiani, G. Krishnasamy, T. J. Rudolphi, F. J. Rizzo. A general algorithm for the numerical solution of hypersingular boundary integral
equations. ASME Journal of applied mechanics, 59 (1992), 604-614.
[9] S. Järvenpää, M. Taskinen, P. Ylä-Oijala. Singularity extraction technique for integral equation methods with higher order basis functions on
plane triangles and tetrahedra. International journal for numerical methods in engineering, 58 (2003), 1149-1165.
[10] P. Kolm, V. Rokhlin. Numerical quadratures for singular and hypersingular integrals. Computers and Mathematics with Applications, 41
(2001), 327-352.
[11] R. D. Graglia. G. Lombardi. Machine precision evaluation of singular and nearly singular potential integrals by use of Gauss quadrature
formulas for rational functions. IEEE transactions on antennas and propagation, 56 (2008), 981-998.
[12] M. Carley. Numeriacl quadratures for singular and hypersingular integrals in boundary element methods. SIAM journal on scientific computing, 29 (2007), 1207-1216.
[13] M. A. Khayat, D. R. Wilton, P. W. Fink. An improved transformation and optimized sampling scheme for the numberical evaluation of
singular and near-singular potentials. IEEE transactions on antennas and propagation, 7 (2008), 377-380.
[14] M. A. Khayat, D. R. Wilton. Numerical evaluation of singular and near-singular potential integrals. IEEE transactions on antennas and
propagation, 53 (2005), 3180-3190.
[15] M. G. Duffy. Quadrature over a pyramid or cube of integrands with a singularity at a vertex. SIAM Journal on Numerical Analysis, 19(6):
1260-1262, 1982.
[16] B. M. Johnston, P. R. Johnston. A comparsion of transformation methods for evaluating two-dimensional weakly singular integrals. International journal for numerical methods in engineering, 56 (2003), 589-607.
[17] L. Scuderi. On the compuation of nearly singular integrals in 3D BEM collocation. International Journal for numerical methods in engineering, 74 (2008), 1733-1770.
[18] K. Hayami. Variable transformation for nearly singular integrals in the boundary element method. Research institute for mathematical sciences, Kyoto university, 41 (2005), 821-842.
[19] B. M. Johnston, P. R. Johnston, D. Elliott. A new method for the numerical evaluation of nearly singular integrals on triangular elements in
the 3D boundary element method. Journal of Computational and Applied Mathematics, 245 (2013), 148-161.
[20] A. J. Burton, G. F. Miller. The application of integral equation methods to the numerical solution of some exterior boundary-value problems.
Proceedings of the Royal Society of London, Series A, 323 (1971), 201-210.
[21] X. W. Gao. An effective method for numerical evaluation of gneral 2D and 3D high order singular boundary integrals. Computer methods in
applied mechanics and engineering, 199 (2010), 2856-2864.
[22] P. R. Johnston. Application of sigmoidal transformations to weakly singular and near-singular boundary element integrals. International
journal for numerical methods in engineering, 45 (1999), 1333-1348.
[23] C. Schwab, W. L. Wendland. Kernel properties and representations of boundary integral operators. Mathematische Nathrichten, 156 (1992),
187-218.
19
| 5cs.CE
|
T OWARDS LIGHTWEIGHT CONVOLUTIONAL NEURAL
NETWORKS FOR OBJECT DETECTION
arXiv:1707.01395v3 [cs.CV] 5 Oct 2017
Dmitriy Anisimov, Tatiana Khanova
Intel
Nizhny Novgorod, Russia
{dmitry.anisimov,tatiana.khanova}@intel.com
A BSTRACT
We propose model with larger spatial size of feature maps and evaluate it on object
detection task. With the goal to choose the best feature extraction network for our
model we compare several popular lightweight networks. After that we conduct
a set of experiments with channels reduction algorithms in order to accelerate
execution. Our vehicle detection models are accurate, fast and therefore suit for
embedded visual applications. With only 1.5 GFLOPs our best model gives 93.39
AP on validation subset of challenging DETRAC dataset. The smallest of our
models is the first to achieve real-time inference speed on CPU with reasonable
accuracy drop to 91.43 AP.
1
I NTRODUCTION
Object detection in general and vehicle detection specifically has many applications including
surveillance, autonomous driving, etc. Moreover, to be practical, object detector must operate on
embedded processors that have far less compute capabilities than powerful GPUs used for benchmarking on typical computer vision datasets. Hence we focus on design of fast object detection
model, which still retains high quality. We perform our experiments for vehicle detection sub-task
on DETRAC dataset (Wen et al., 2015).
O UR MOTIVATION
To reduce number of operations it is natural to try aggressively decreasing spatial size of feature
maps. But from the other side, it is known that for tasks such as semantic segmentation or object detection it is important to preserve spatial information. So we investigate possibility to train accurate
model with relatively large feature maps and still keep it fast.
Often network weights are initialized from a model, pretrained on classification of a large number
of classes, e.g. ImageNet with 1000 classes (Russakovsky et al., 2014). We hypothesize that for a
lot of practical tasks with few classes a considerable number of channels in convolution layers from
such models are redundant. So in order to achieve high processing speed we focus on reducing the
number of channels.
O UR CONTRIBUTION
• Our findings show that it is possible to get good quality and fast detection model by increasing size of feature maps and decreasing number of channels.
• We demonstrate how to get lightweight model from a large one, by simply picking some
number of channels and after that fine-tuning the resulting model on the given task.
• Straightforward layer-by-layer architecture like ResNet is shown to give good quality on
par with more sophisticated architectures.
• We present the set of accurate models where the smallest one works in real-time on CPU at
34 fps.
1
Table 1: Feature extraction network accuracy on
ImageNet (* - our implementation)
Table 2: Detection models with different feature
extraction networks (on DETRAC validation split)
Model
Top-1
Top-5
GFLOPs
MParams
SqueezeNet1.0
SqueezeNet1.0 bn
ResNet10
MobileNet*
PVANet
57.50
57.67
63.87
69.29
72.34
80.30
81.29
85.16
89.25
91.16
0.859
0.859
0.893
0.579
0.616
1.2
1.2
5.4
4.3
100.9
2
R ELATED
Model
AP
GFLOPs
MParams
MobileNet SSD
PVANet SSD
SqueezeNet1.0 SSD
MobileNet light SSD
ResNet10 SSD
86.92
88.24
88.61
88.93
89.94
1.7
1.8
1.8
0.745
1.8
5.2
5.4
3.5
2.3
4.0
SqueezeNet1.0 bn SSD
91.10
1.8
3.5
WORK
Currently there are two popular approaches to object detection, namely: Faster R-CNN (Ren et al.,
2015) and SSD (Liu et al., 2015). As it was shown in Huang et al. (2016) SSD framework has the
best quality/speed trade-off, so we choose it for our experiments.
Besides detection framework it is important to use lightweight feature extraction network to preserve reasonable number of computations. For this reason we compared several such networks:
ResNet10 (He et al., 2015), SqueezeNet (Iandola et al., 2016), MobileNet (Howard et al., 2017),
PVANet (Kim et al., 2016).
Dilation (i.e. input stride in convolution), which we use in our model, first was proposed for segmentation task in Chen et al. (2016). They also sub-sampled two last fully connected layers in VGG
network (Simonyan & Zisserman, 2014) in order to reduce computations. In SSD framework the
same reduced VGG model was used. We extend this idea to sub-sampling all layers in feature
extraction network.
Concurrent with our work, authors in Yu et al. (2017) applied dilations with similar purpose to classification and segmentation tasks. In comparison with their approach, after removing pooling operations, we additionally sample channels in convolution layers and thus preserve small number of
computations.
In addition to good quality, real-time performance is required for most practical applications. There
are a lot of methods dealing with high computation demand of neural networks. Roughly they can
be divided into two groups, based on dependency on hardware support.
The first group includes such methods as quantization (Gysel et al., 2016; Zhou et al., 2017) and
weight sparsifying (Han et al., 2015). But these approaches require delicate hardware customization
to get practical speedup. For example, quantization relies on hardware support of low-bit operations
and weights pruning leads to sparse computations.
Our target is to get fast model that does not depend on particular hardware, so we focus on the second
hardware independent group. Those methods decrease number of computations using standard DNN
building blocks. Different sorts of decompositions are used, e.g. searching for basis using PCA
decomposition (Wen et al., 2017). Another possible solution is pruning of channels. While weightlevel pruning requires sparse computation support, pruning also can be done on per-channel basis
by eliminating less important channels in convolution filters (Li et al., 2016).
3
3.1
A RCHITECTURE
F EATURE EXTRACTION
NETWORK
We choose four networks for our experiments. Their accuracies on ImageNet classification task are
shown in Table 1. It was shown earlier, for example, for visual odometry (Agrawal et al., 2015),
that features trained on two different tasks with the same network architecture lead to different
quality on third task. The same may hold for fine-tuning from different architectures. So we can’t
choose feature extraction network for detection relying solely on classification accuracies. Below we
demonstrate that the best accuracy on classification task doesn’t necessarily lead to better accuracy
for object detection.
2
Next we study influence of feature extraction network on detection accuracy. We modify chosen
networks to have similar FLOPs and number of parameters for fair comparison. Brief description of
the networks and our modifications is given below.
SquezeNet1.0 Network consists of ”fire” blocks, each block contains two stages: first convolution
layer makes ”bottleneck” by reducing number of channels. Next there are two parallel inception-like
(Szegedy et al., 2014) convolution layers with 1x1 and 3x3 kernels accordingly. We add two ”fire”
blocks at the end of network in the same way as it was done in SqueezeDet paper (Wu et al., 2016).
MobileNet In recently presented lightweight architecture authors replace convolution layer with
depthwise convolution applying a single 3x3 filter to each input channel, followed by conventional
convolution with 1x1 kernels. This network already contains required number of FLOPs and parameters, so we keep it as is.
PVANet PVANet is a relatively sophisticated architecture. It consists of CReLU layers
(Shang et al., 2016) at the beginning of the network, followed by inception blocks. At the end
there are two large fully connected classification layers with 4096 channels each, that’s why it has a
large amount of weights. To reduce number of weights we sample last fully connected layers in the
same way as it was done for VGG model in DeepLab and SSD frameworks:1
• 6x6 kernels are sampled to 3x3 and dilation 6 is added to the following layer in order
to preserve receptive field (to compensate kernel sampling and removing previous max
pooling).
• 4096 channels are sampled to 256. We use 256 instead of 1024 to be able to compare with
other models.
ResNet10 Often only large versions of ResNet architecture are used, e.g. with 50 layers and more,
but we show that small 10-layer model gives good quality as well. We use model with pre-activations
trained in Caffe (Jia et al., 2014) by Computer Vision Group Jena (Simon et al., 2016). We sample
last two layers to 256 channels for comparison with other models.
3.2
SSD
MODIFICATIONS FOR
DETRAC
Since ground truth bounding boxes for DETRAC test dataset are not publicly available, we prepare
our train/validation split. We select 10 videos (totally 15K frames) as validation and the rest 68K as
train trying to preserve ratio of night/day videos and keep similar locations in the same split.
To adapt SSD framework to DETRAC dataset we make the following changes:
• We keep three prior box scale levels based on the bounding box sizes on the dataset. We
use the same parameters, except for the first ”small” scale, for which we decrease range to
15/50.
• Our experiments show that down-sampling images to 320x256 resolution gives good quality/speed trade-off.
Results of this configuration with four feature extraction networks are shown in Table 2. One can
see that ResNet10 provides the best accuracy, so we choose it for further experiments.
In the third version of this paper we fix issue with proper handling ignore regions during evaluation
on DETRAC validation split. All conclusions remain the same, the only change is that metrics
become around 2 AP lower.
In Section 3.6 we get the best result with MobileNet feature extraction network, so we conduct
additional experiment with it. We sample channels to 256 for all layers with large number of channels. This model shows better accuracy comparing to original MobileNet and has smaller number
of parameters, see Table 2.
1
We reduce learning rate for detection model training two times, because it diverges with original value.
3
Table 3: Comparing original ResNet10 feature extraction network with our detection model
name
conv1
pool1
res block
res block
res block
res block
conv16 det
conv32 det
type
7x7/2
3x3
3x3
3x3/2
3x3/2
3x3/2
ResNet10
spatial dim. channels
128x160
64
64x80
64
64x80
64
32x40
128
16x20
256
8x10
512
SSDR
type
spatial dim.
7x7/2
128x160
3x3
64x80
3x3
64x80
3x3/2
32x40
3x3 dil 1,2
32x40
3x3 dil 2,4
32x40
3x3/2
16x20
3x3/2
8x10
1.5
64
64
64
64
128
128
128
128
0.75
channels
49
49
49
49
76
76
76
76
0.47
41
41
41
41
49
49
49
49
After submission to the conference we did the same with SqueezeNet1.0 bn network
(SqueezeNet1.0 with batch normalizations). Since it achieves the best quality we expect that our
approach applied to it would lead to even better results.
3.3
M ODEL
In this section we study possibility of getting accurate and fast detection model by increasing size
of feature maps and decreasing number of channels, we propose the following model. To keep size
of feature maps we remove two last spatial reductions, that reduce image resolution by 16 and 32
times. In order to preserve receptive field of following convolution layers we add dilations 2 and 4 to
them accordingly. We also sample channels in most convolution layers to retain reasonable number
of FLOPs.2 Resulting architecture is shown in Table 3, model SSDR 1.5.
3.4
S TUDYING
EVEN LIGHTER MODELS
In order to accelerate existing model further and achieve real-time performance on CPU, we continue experiments with methods for eliminating channels, see Table 3, models SSDR 0.75 and
SSDR 0.47.
One-shot random sampling First, we take our best SSDR 1.5 model, decrease number of channels following random sampling strategy and fine-tune it. Resulting 473 MFLOPs model gives 91.09
AP on DETRAC validation split.
One-shot pruning Pruning scheme, proposed in Li et al. (2016) appears to be straightforward
improvement of the random sampling approach:
1. For each convolution layer in network, except for the layers in detection part:
(a) For each filter calculate L1 metric to measure usefulness.
(b) Prune specified percentage of filters with the smallest metrics. Percentage of filters to
prune is chosen empirically to be 5% and 10% for the first and the last half respectively.
2. Fine-tune the whole network on the training dataset with smaller learning rate.
Iterative pruning The above algorithm can be repeated until desired quality/speed trade-off is
achieved leading to iterative pruning as an opposite to one-shot approach.
Contrary to Li et al. (2016), for pruning of pre-elementwise sum convolutions we find that using 3x3
filters to calculate metric instead of shortcut convolutions gives better result.
PCA decomposition Another approach we employ was proposed in Wen et al. (2017) and consists
of applying PCA to filters in order to determine basis for each convolution and then decompose layer
into two consecutive convolutions.
2
We experiment with sampling channels by L1 metric (see Section 3.4), but it doesn’t improve results.
4
Table 4: Comparison of detection models (on
DETRAC validation split)
Figure 1: Comparison of channel reduction
techniques on DETRAC validation split. Real-time
performance on CPU is labeled with vertical line.
Model
AP
GFLOPs
MParams
ResNet10 Clustered
ResNet10 SSD
ResNet10 FPN
SSDR 1.5 (ours)
SSDR 0.75 (ours)
SSDR 0.47 (ours)
89.07
89.94
91.50
93.39
92.49
91.43
1.7
1.8
1.4
1.5
0.75
0.47
3.9
4.0
2.8
1.1
0.47
0.24
0.94
SSDR 1.5
Mean AP
SSDR 0.75
SSDR 0.47
0.92
Random sampling
One-shot pruning
Iterative pruning
PCA decomposition
0.9
0.2
0.4
0.6
0.8
1
GFLOPs
1.2
1.4
1.6
Table 5: Comparison of our model with the best models from leader-board (on DETRAC test split)
Model
full set
easy set
medium set
hard set
cloudy set
night set
rainy set
sunny set
RTN
EB
NANO
SSDR 0.75 (ours)
SSDR 1.5 (ours)
FasterRCNN2
YOLO2
SSDR 0.47 (ours)
74.15%
67.96%
63.01%
59.07%
58.68%
58.45%
57.72%
57.07%
91.52%
89.65%
80.33%
77.84%
79.55%
82.75%
83.28%
76.67%
79.16%
73.12%
68.04%
64.41%
63.74%
63.05%
62.25%
62.22%
61.73%
53.64%
50.73%
45.98%
44.93%
44.25%
42.44%
43.89%
77.02%
72.42%
67.00%
62.79%
61.59%
66.29%
57.97%
62.41%
77.20%
73.93%
62.20%
60.88%
62.19%
69.85%
64.53%
58.48%
65.27%
53.40%
55.89%
48.55%
47.47%
45.16%
47.84%
45.26%
84.14%
83.73%
73.89%
74.32%
74.42%
62.34%
69.75%
72.55%
Comparison of aforementioned techniques applied to ours best model is shown in Figure 1. Iterative
pruning strategy gives the best 91.43 AP for 473 MFLOPs model and 92.49 AP for 752 MFLOPs
model. But the results we get with other methods are close to the best one.
3.5
E XPERIMENTS
We compare our models with several recently proposed architectures, namely: base SSD with three
scale levels (Liu et al., 2015), SSD with clustered priors (Erhan et al., 2013; Wu et al., 2016) and
Feature Pyramid Networks (Lin et al., 2016). Our solution gives superior result comparing to all
aforementioned models either on accuracy or processing speed, see Table 4.
Clustered priors Following Erhan et al. (2013); Wu et al. (2016) we cluster DETRAC ground
truth bounding boxes to get dataset specific priors. First, all bounding boxes were clustered into 3
scale groups, and then each of them was clustered into 4 groups to achieve comparable number of
priors with the original SSD. While model with clustered priors does not show the best result (Table
4), clustering approach may provide valuable insight on choice of scale/aspect ratio parameters for
priors generated by other methods.
ACCURACY
AND TIME
Table 5 compares accuracies of our models with other results on different subsets of DETRAC test
set. On time of submission our model is ranked 4th in official leader-board3 by quality and it is at
least an order of magnitude faster than all others. Noticeably, on the test set smaller SSDR 0.75
model gives even better accuracy than the base one.
We measure inference speed of our smallest SSDR 0.47 model on Intel R CoreTMi7-6700K CPU @
4.00GHz x 8 using Intel R MKL library and Caffe and get ∼34 fps.4
3
http://detrac-db.rit.albany.edu/DetRet
Note that we merged batch normalization mean/variance in previous convolution weights to eliminate
redundant computations.
4
5
Table 6: Detection models with different feature
extraction networks (on VOC 2007 test)
Table 7: Comparison of ResNet10 based detection
models (on VOC 2007 test)
Model
AP
GFLOPs
MParams
Model
AP
GFLOPs
MParams
SqueezeNet1.0 SSD
ResNet10 SSD
SqueezeNet1.0 bn SSD
PVANet SSD
MobileNet SSD
SSDM 7.5 (ours)
38.45
64.83
65.61
67.69
70.04
73.08
2.8
2.3
2.8
2.3
2.6
7.5
7.0
6.7
7.0
8.1
8.8
10.1
ResNet10 FPN
ResNet10 SSD
SSDR 5.5 (ours)
63.76
64.83
68.73
2.0
2.3
5.5
5.3
6.7
8.5
3.6
PASCAL VOC
In the previous sections we apply our approach to concrete vehicle detection task. In order to extensively study capabilities of proposed techniques we conduct similar set of experiments on general
object detection PASCAL VOC dataset (Everingham et al., 2015). We don’t experiment with channel reduction for them and use the original SSD 300x300 configuration. The results obtained from
this study are provided in Tables 6 and 7. Since we get the best result for SSD with Mobilenet
feature extraction network, we make use of our modification with replacing two max pooling with
dilations to it as well, see Table 6, SSDM 7.5 model.
Interestingly, performance of detection models with different feature extraction networks on VOC
test set shares tendency in accuracies with Imagenet classification task (Table 1) contrary to what
we get for DETRAC dataset. We argue that this is because of the difference in the number of classes
between these datasets: only one class in DETRAC and 20/1000 in PASCAL VOC and Imagenet
accordingly. Original SqueezeNet1.0 SSD model shows poor result on this dataset, but works quite
well after we add batch normalizations to it, see Table 6, SqueezeNet1.0 bn SSD model.
4
C ONCLUSION
We have presented a set of detection models, which are accurate, fast and therefore suit for realworld applications. We keep high spatial resolution in feature extraction part, rather than losing
spatial structure information caused by progressively reducing the resolution of internal representations. Even though the proposed network is designed for vehicle detection, we believe our design
choices can be applicable to other practical tasks, particularly where detailed understanding of the
scene is important.
Our network design is independent of network compression and quantization, so those methods are
applicable to our network as well to further increase the actual performance in real tasks. We have
shown that relatively simple pruning technique can give real-time performance on CPU with less
than 2% AP loss on test set.
R EFERENCES
Pulkit Agrawal, João Carreira, and Jitendra Malik. Learning to see by moving.
abs/1505.01596, 2015. URL http://arxiv.org/abs/1505.01596.
CoRR,
Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L.
Yuille.
Deeplab:
Semantic image segmentation with deep convolutional nets,
atrous convolution, and fully connected crfs.
CoRR, abs/1606.00915, 2016.
URL
http://arxiv.org/abs/1606.00915.
Dumitru Erhan, Christian Szegedy, Alexander Toshev, and Dragomir Anguelov.
able object detection using deep neural networks. CoRR, abs/1312.2249, 2013.
http://arxiv.org/abs/1312.2249.
ScalURL
M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The
pascal visual object classes challenge: A retrospective. International Journal of Computer Vision,
111(1):98–136, January 2015.
6
Philipp Gysel, Mohammad Motamedi, and Soheil Ghiasi.
Hardware-oriented approximation of convolutional neural networks.
CoRR, abs/1604.03168, 2016.
URL
http://arxiv.org/abs/1604.03168.
Song Han, Jeff Pool, John Tran, and William J. Dally.
Learning both weights and
connections for efficient neural networks.
CoRR, abs/1506.02626, 2015.
URL
http://arxiv.org/abs/1506.02626.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015. URL http://arxiv.org/abs/1512.03385.
Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional
neural networks for mobile vision applications.
CoRR, abs/1704.04861, 2017.
URL
http://arxiv.org/abs/1704.04861.
Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian
Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, and Kevin Murphy. Speed/accuracy
trade-offs for modern convolutional object detectors. CoRR, abs/1611.10012, 2016. URL
http://arxiv.org/abs/1611.10012.
Forrest N. Iandola, Matthew W. Moskewicz, Khalid Ashraf, Song Han, William J. Dally, and Kurt
Keutzer. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and <1mb model size.
CoRR, abs/1602.07360, 2016. URL http://arxiv.org/abs/1602.07360.
Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross B. Girshick,
Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. CoRR, abs/1408.5093, 2014. URL http://arxiv.org/abs/1408.5093.
Kye-Hyeon Kim, Yeongjae Cheon, Sanghoon Hong, Byung-Seok Roh, and Minje Park. PVANET:
deep but lightweight neural networks for real-time object detection. CoRR, abs/1608.08021, 2016.
URL http://arxiv.org/abs/1608.08021.
Hao Li, Asim Kadav, Igor Durdanovic,
Pruning filters for efficient convnets.
http://arxiv.org/abs/1608.08710.
Hanan Samet, and Hans Peter
CoRR, abs/1608.08710, 2016.
Graf.
URL
Tsung-Yi Lin, Piotr Dollár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. CoRR, abs/1612.03144, 2016. URL
http://arxiv.org/abs/1612.03144.
Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott E. Reed, Cheng-Yang Fu,
and Alexander C. Berg. SSD: single shot multibox detector. CoRR, abs/1512.02325, 2015. URL
http://arxiv.org/abs/1512.02325.
Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards realtime object detection with region proposal networks. CoRR, abs/1506.01497, 2015. URL
http://arxiv.org/abs/1506.01497.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and FeiFei Li. Imagenet large scale visual recognition challenge. CoRR, abs/1409.0575, 2014. URL
http://arxiv.org/abs/1409.0575.
Wenling Shang, Kihyuk Sohn, Diogo Almeida, and Honglak Lee. Understanding and improving
convolutional neural networks via concatenated rectified linear units. CoRR, abs/1603.05201,
2016. URL http://arxiv.org/abs/1603.05201.
Marcel Simon, Erik Rodner, and Joachim Denzler. Imagenet pre-trained models with batch normalization. CoRR, abs/1612.01452, 2016. URL http://arxiv.org/abs/1612.01452.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image
recognition. CoRR, abs/1409.1556, 2014. URL http://arxiv.org/abs/1409.1556.
7
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov,
Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions.
CoRR, abs/1409.4842, 2014. URL http://arxiv.org/abs/1409.4842.
Longyin Wen, Dawei Du, Zhaowei Cai, Zhen Lei, Ming-Ching Chang, Honggang Qi, Jongwoo Lim,
Ming-Hsuan Yang, and Siwei Lyu. DETRAC: A new benchmark and protocol for multi-object
tracking. CoRR, abs/1511.04136, 2015. URL http://arxiv.org/abs/1511.04136.
Wei Wen, Cong Xu, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li.
dinating filters for faster deep neural networks. CoRR, abs/1703.09746, 2017.
http://arxiv.org/abs/1703.09746.
CoorURL
Bichen Wu, Forrest N. Iandola, Peter H. Jin, and Kurt Keutzer. Squeezedet: Unified, small, low
power fully convolutional neural networks for real-time object detection for autonomous driving.
CoRR, abs/1612.01051, 2016. URL http://arxiv.org/abs/1612.01051.
Fisher Yu, Vladlen Koltun, and Thomas Funkhouser. Dilated residual networks. ArXiv:1705.09914,
2017. URL https://arxiv.org/abs/1705.09914.
Aojun Zhou, Anbang Yao, Yiwen Guo, Lin Xu, and Yurong Chen. Incremental network quantization: Towards lossless cnns with low-precision weights. CoRR, abs/1702.03044, 2017. URL
http://arxiv.org/abs/1702.03044.
8
| 9cs.NE
|
1
Mutual Information in Frequency and its
Application to Measure Cross-Frequency Coupling
in Epilepsy
arXiv:1711.01629v2 [q-bio.NC] 15 Mar 2018
Rakesh Malladi, Member, IEEE, Don H Johnson, Fellow, IEEE, Giridhar P Kalamangalam,
Nitin Tandon, and Behnaam Aazhang, Fellow, IEEE
Abstract—We define a metric, mutual information in frequency (MI-in-frequency), to detect and quantify the statistical
dependence between different frequency components in the
data, referred to as cross-frequency coupling and apply it to
electrophysiological recordings from the brain to infer crossfrequency coupling. The current metrics used to quantify the
cross-frequency coupling in neuroscience cannot detect if two
frequency components in non-Gaussian brain recordings are
statistically independent or not. Our MI-in-frequency metric,
based on Shannon’s mutual information between the Cramér’s
representation of stochastic processes, overcomes this shortcoming and can detect statistical dependence in frequency between
non-Gaussian signals. We then describe two data-driven estimators of MI-in-frequency: one based on kernel density estimation
and the other based on the nearest neighbor algorithm and
validate their performance on simulated data. We then use
MI-in-frequency to estimate mutual information between two
data streams that are dependent across time, without making
any parametric model assumptions. Finally, we use the MI-infrequency metric to investigate the cross-frequency coupling in
seizure onset zone from electrocorticographic recordings during
seizures. The inferred cross-frequency coupling characteristics
are essential to optimize the spatial and spectral parameters of
electrical stimulation based treatments of epilepsy.
Index Terms—Mutual information in frequency; dependent
data; Cramér’s spectral representation; cross-frequency coupling;
epilepsy; seizure onset zone.
I. I NTRODUCTION
Epilepsy is a very common neurological disorder affecting
nearly 1% of the world’s population. Epilepsy is characterized
by repeated, unprovoked seizures. Nearly a third of all epilepsy
patients have medically refractory epilepsy (medication is not
effective in these patients). For these patients, surgical resection
of the seizure onset zone (SOZ) (the regions of the brain
responsible for generating and sustaining seizure activity [3]) or
electrical stimulation are possible treatment options. However,
the efficacy of these treatments is variable and almost always
never results in a cure [4], [5]. There is tremendous interest
in leveraging the recent advances in electrical stimulation
This work is funded in part by grant 1406447 from National Science
Foundation and Texas Instruments and was done at Rice University. A portion
of this work was presented at Cosyne [1] and Asilomar [2].
Rakesh Malladi is with LinkedIn Corporation, Sunnyvale, CA. Don H Johnson and Behnaam Aazhang are with the Department of Electrical and Computer
Engineering, Rice University, Houston, TX. Giridhar P Kalamangalam is with
Department of Neurology at University of Florida, Gainesville, FL. Nitin
Tandon is with Department of Neurosurgery at University of Texas Health
Center, Houston, TX. E-mail: [email protected], {dhj, aaz}@rice.edu,
[email protected], [email protected].
[6] and optogenetics [7] to develop spatiotemporally specific
approaches to treat epilepsy. A crucial step in this endeavor is
to develop an understanding of the coupling between neuronal
oscillations in different frequency bands during seizures. This
coupling or statistical dependence across frequency components
between signals is referred to as cross-frequency coupling
(CFC) [8], [9]. Our main objective is to learn the dynamics of
cross-frequency coupling during seizures in epilepsy patients
from the electrocorticographic (ECoG) data.
Elaborating the characteristics of epileptic seizures using
cross-frequency coupling between ECoG data has been the
focus of many papers. CFC has been used to predict the onset
of seizure in [10] and detect epileptic seizures in [11]. CFC
has also been used to localize the area for surgical resection
in epilepsy patients [12]–[14]. Variations in CFC from preictal
(before a seizure) to ictal (during a seizure) to postictal (after
a seizures) in epilepsy patients have been analyzed in [15],
[16]. In addition, the CFC in interictal stages is compared with
that around seizures in [11], [17]–[19]. In this paper, we study
CFC within and between various regions inside the seizure
onset zone to determine the dominant frequencies involved in
seizures and to learn the variations in coupling strength between
various spatial regions inside SOZ. The results from this study
are crucial to optimize the spectral and spatial parameters of
next generation epilepsy treatments.
Cross-frequency coupling or dependence across frequencies
in the data could be in a single recording or between recordings,
not necessarily at the same frequency. Coherence can identify
if two frequency components are statistically independent
or not and quantify the dependence for linear, Gaussian
processes [20]. There is no such equivalent metric for nonGaussian signals. Since the time-series data recorded from the
brain are neither linearly related nor Gaussian, neuroscientists
typically use heuristic metrics that cannot identify if two
frequency components are statistically independent or not and
can only capture second-order dependencies. Some of the
popular heuristics estimate the phase-amplitude, amplitudeamplitude, phase-phase coupling between the low and high
frequency components in the electrophysiological recordings
from brain [8], [21]–[23]. In fact, a recent review article on
CFC metrics suggests the use of cross-frequency ‘correlation’
instead of ‘coupling’ to describe these heuristic CFC metrics
[22]. Furthermore, a list of confounds affecting the current
CFC metrics is provided in [22]. A more comprehensive metric
that detects statistical independence and thereby, capture both
2
linear and nonlinear dependencies, would be invaluable in
determining how neuronal oscillations at various frequencies
are involved in the computation, communication, and learning
in the brain. Here we propose a new methodology or metric to
estimate the cross-frequency coupling (CFC) in neuroscience
that overcomes the challenges of the existing approaches and
as a proof-of-concept, we infer CFC characteristics of epileptic
seizures using our metric.
Mutual information in frequency (MI-in-frequency), defined
for linear Gaussian processes using coherence in [24], [25],
can indeed be further developed into a general technique to
estimate CFC. Inspired by prior work [26], we define MIin-frequency between two frequencies in a signal (or two
signals) as the Shannon’s mutual information (MI) between
the Cramér’s spectral representations [27], [28] of the two
signals at the corresponding frequencies. Cramér’s spectral
representation transforms a time-domain stochastic process
into a stochastic process in the frequency domain, the samples
of which can be estimated at each frequency from the timedomain data samples [29]. MI-in-frequency metric is equivalent
to coherence measures for linear, Gaussian signals and can be
thought of as ‘coherence’ for non-Gaussian signals. The MI-infrequency metric is one of the three mutual information based
metrics used in [26] to analyze linear relationships between
seismic data and [26] is not focussed on defining a single
metric to capture the statistical dependence across frequency.
We extend this approach to define a single metric, MI-infrequency, to capture statistical dependencies across frequency
for both linear and nonlinear data and use it measure CFC in the
brain. We then describe two data-driven algorithms – one based
on kernel density estimation (KDMIF) and the other based
on nearest neighbor estimation (NNMIF) – to estimate MI-infrequency without assuming any parametric model of the data.
We considered these two approaches since they outperformed
other approaches in estimating MI from i.i.d. data and there is
no clear winner between them [30], [31]. We also demonstrate
the superiority of MI-in-frequency over existing CFC metrics
by comparing against modulation index [8], [21], a commonly
used CFC metric, on simulated data.
In addition to estimating CFC between ECoG data, we use
MI-in-frequency to develop a data-driven estimator for mutual
information (MI). Note that MI estimation is a solved problem
if the data samples are i.i.d. [32] or are sampled from linear,
Gaussian processes [24], [25], [33], [34]. As mentioned earlier,
real-world data is neither independent across time nor Gaussian
and the underlying model is often unknown. Our data-driven
MI estimation algorithm applies to dependent data, without
making any parametric model assumptions. The key idea is to
make the problem computationally tractable by focussing only
on those frequencies in the two data streams that are statistically
dependent, which are identified by MI-in-frequency metric. Our
MI estimator converges to the true value for Gaussian models
and we validate its performance on nonlinear models.
Finally, we apply the MI-in-frequency estimators to infer the
cross-frequency coupling in the seizure onset zone (SOZ), by
analyzing electrocorticographic (ECoG) data from the SOZ of
9 patients with medial temporal lobe epilepsy in whom a total
of 25 seizures were recorded. We investigate the dynamics of
CFC in preictal, ictal and postictal periods within one SOZ
electrode and between electrodes in different regions in the
SOZ. We observe an increase in coupling in gamma and ripple
high-frequency oscillations during seizures, with the largest
increase within a SOZ electrode and a very small increase
between electrodes in different regions inside SOZ. In addition,
low-frequency coupling and linear interactions between SOZ
electrodes also increase during the postictal state.
II. C RAMÉR ’ S S PECTRAL R EPRESENTATION OF
S TOCHASTIC P ROCESSES
Consider a stochastic processes X (t) , t ∈ R. Let SX (ν) for
ν ∈ R be the spectral distribution function of X and sX (ν),
its power spectral density, if it exists. Two basic spectral
representations are associated with the stochastic process
X (t) - power spectral distribution and Cramér’s representation
[27], [28]. The Cramér’s representation of X (t) and its key
properties are stated in the following theorem.
Theorem 1. (page 380 in [28]) Let X (t) be a second order
stationary, mean-square continuous and zero mean stochastic
process. Then there exists a complex-valued, finite-variance,
e (ν) in the frequency domain
orthogonal increment process X
ν ∈ R, such that
R∞ j2πνt
e (ν) ,
X (t) =
e
dX
−∞
h
i
h
i
e (ν) = 0, and E |dX
e (ν) |2 = dSX (ν) .
with E dX
e (ν) = X
eR (ν) + j X
eI (ν) satisfying the above
The process X
theorem is the spectral process or the Cramér’s representation
e (ν) is the complex random variable representing
of X (t). dX
the amplitude of oscillation in the interval from ν to ν + dν
in X (t). The integral in Theorem 1 is a Fourier-Stieltjes
integral. Intuitively, Theorem 1 decomposes X (t) into an
orthogonal increment complex process in the frequency domain.
e −ν =
Furthermore, if the X (t) is real-valued, then X
e ? ν , E dX
eR ν dX
eI ν = 0, and
X
eR (ν) 2 = E dX
eI (ν) 2 = 1 dSX (ν) . (1)
E dX
2
We have the following theorem for the special case of a realvalued Gaussian process X (t).
Theorem 2. (page 385 in [28]) Let X (t) be a real-valued
stationary, mean-square continuous Gaussian process with zero
mean and power spectral distribution function SX (ν) , ν ∈ R.
eR (ν)
Then the real and imaginary parts of its spectral process X
eI (ν) are zero mean, mutually independent, identically
and X
distributed Gaussian processes satisfying (1).
Example: Consider the zero mean stationary Gaussian
process X (t) = A cos (2πν0 t + Θ), where A is Rayleigh
random variable with parameter σA that is independent of
Θ, which is uniform in [0, 2π). The increments of the spectral
process of X (t) are all zero, except at ν = ±ν0 , where the
increment is A2 exp (±jΘ) [28]. This implies that the sample
e (ν) has two jumps
path of the real part of spectral process X
of same magnitude and direction at frequencies ±ν0 , while
that of the imaginary part has two jumps of same magnitude,
3
but opposite directions at ±ν0 . The magnitude of the jump
at ν0 in the real and imaginary parts is A2 cos Θ and A2 sin Θ
respectively, both of which are Gaussian random variables with
2
mean zero and variance 12 σA
. This spectral process is intuitive
because we know X (t) has all its energy only at frequencies
±ν0 and the variance of the increments of the spectral process
e (ν) is equal to the differential power spectral distribution
dX
of X (t) which is nonzero only at ±ν0 . We therefore expect
e (ν) with non-zero
all sample paths of the random process X
probability to be constant, except for jumps at ±ν0 .
Note that if the process is wide sense-stationary and
Gaussian, then power spectral distribution would have all the
information about the process and its relationship with Cramér’s
representation is given by Theorem 2. Otherwise, power spectral
distribution only captures the second-order dependencies in
the process. Since ECoG signals are not Gaussian, we use
Cramér’s representation to transform a time-domain stochastic
process into a stochastic process in the frequency domain.
2
is zero except at ν = 0, where the spectral increment is A2 ,
2
and at ν = ±2ν0 , where the increment is A4 exp (±j2Θ).
As a result, the frequency components at ±ν0 in X and at
frequencies {0, ±2ν0 } in Y are statistically dependent and
hence the MI-in-frequency obtained from (2) at these frequency
pairs will be positive. In addition, the frequency components
in Y at ν ∈ {0, ±2ν0 } are dependent and hence the MI-infrequency within Y at these frequencies will also be positive.
A. Gaussian Inputs to LTI Filters
Let’s now consider the special case where X (t), a Gaussian
process with power spectral density sX (ν) serves as the input
to a linear, time-invariant (LTI) filter with transfer function
H1 (ν) and Y (t) is output observed in additive colored noise
(white noise W (t) passed through a LTI filter with transfer
function H2 (ν)). The processes X (t) and Y (t) are related
by
y (t) = h1 (t) ∗ x (t) + h2 (t) ∗ w (t) ,
III. M UTUAL I NFORMATION IN F REQUENCY
(4)
where ∗ denotes convolution operation, x(t), y(t) and w(t) are
We first define MI between frequencies within a pro- sample paths of X (t), Y (t) and W (t) respectively. W is
cess and between two processes in continuous time. We a Gaussian process with power spectral density sW (ν) and
then extend this definition to discrete-time stochastic pro- independent of X. h1 (t) and h2 (t) are continuous-time impulse
e (νi ) and dYe (νj ), the increments of
cesses. Consider dX
responses of LTI filters, whose transfer functions are H1 (ν)
spectral processes or the Cramér’s representation of X(t) and H2 (ν) respectively. Let dX
e (ν), dW
f (ν) and dYe (ν) be
and Y (t) at frequencies νi and νj respectively. Let the joint the spectral process increments of the Gaussian processes X,
probability density of the four dimensional random vector W and Y . We have from Theorem 2,
e (νi ) and dYe (νj )
of the real and imaginary parts of dX
eR (ν) , dX
eI (ν) ∼N 0, 1 sX (ν) I ,
dX
eR (νi ) , dX
eI (νi ) , dYeR (νj ) , dYeI (νj ) .
be denoted by P dX
2
fR (ν) , dW
fI (ν) ∼N 0, 1 sW (ν) I ,
The corresponding two-dimensional
marginal densities are
dW
(5)
2
eR (νi ) , dX
eI (νi ) , P dYeR (νj ) , dYeI (νj ) . The MI-inP dX
where N (µ, Σ) represents Gaussian distribution with mean µ
frequency between X (t) at νi and Y (t) at νj is defined as
and covariance Σ, 0 is the two element zero vector and I is
MIXY (νi , νj )
the 2 × 2 identity matrix. In addition, we can show for the
e
e
e
e
model in (4) that
= I dXR (νi ) , dXI (νi ) ; dYR (νj ) , dYI (νj ) ,
eR (νi ),dX
eI (νi ),dY
e (ν ),dY
eI (νj )
e (ν) + H2 (ν) dW
f (ν) .
P dX
dYe (ν) = H1 (ν) dX
(6)
R j
,
= E log
(2)
eR (νi ),dX
eI (νi ) P dY
eR (νj ),dY
eI (νj )
P dX
The proof of (6) is in the appendix. The MI-in-frequency
where I ({·, ·} ; {·, ·}) is the standard mutual information defined in (2) is further simplified for the model in (4) using
between two pairs of two dimensional real-valued random (5), (6) and stated in the following theorem.
vectors [34]. The MI between two different frequencies νi , νj
Theorem 3. For the model given in (4), the MI between X (t)
in the same process Y (t) is similarly defined as
at frequency νi and Y (t) at frequency νj is zero, when νi 6= νj
MIY Y νi , νj =I dYeR νi ,dYeI νi ; dYeR νj ,dYeI νj . (3) and the MI between X (t) and Y (t) at frequency νi = νj =
ν 6= 0 is
The MI between the components of Y at frequencies νi =
eR (ν) , dX
eI (ν) ; dYeR (ν)
νj =
MIXY (ν, ν) = 2 × I dX
ν, MIY Y (ν, ν),
is ∞, a consequence of the fact
that dYeR (ν) , dYeI (ν) is a continuous-valued random vector
|H1 (ν)|2 sX (ν)
= log 1 + |H
.
(7)
2
2 (ν)| sW (ν)
whose conditional differential entropy is not lower bounded.
MI-in-frequency defined in (2), (3) is a non-negative number. If
The proof of the above theorem is in the appendix. Note
MI-in-frequency between two frequencies is zero, then they are that at ν =
X and Y is
0, the MI-in-frequency between
eR (ν) , dX
eI (ν) ; dYeR (ν) , which is just half
independent and if not, MI-in-frequency is a measure of the sta- equal to I dX
tistical dependence between the two frequency components. MI- of the right hand side of (7). We intuitively expect different
in-frequency between two processes is not symmetric in general, frequency components in the Gaussian input and its output
i.e., MIXY (νi , νj ) 6= MIXY (νj , νi ). However, it is symmetric from a linear system to be independent and Theorem 3 confirms
within a process, i.e., MIY Y (νi , νj ) = MIY Y (νj , νi ).
that the proposed definition of MI-in-frequency agrees with
Example: Continuing with our example in section II, let this intuition. In addition, the MI between X and Y is ∞
2
X (t) = A cos (2πν0 t + Θ) and Y (t) = X (t) . Then dYe (ν) when |H2 (ν) | = 0, since the components of X and Y at
4
such ν are linearly related. The MI between two different
frequencies in Y (t), generated from (4), is zero due to
the linearity of the filters and Gaussian inputs. Furthermore,
we can also show for the Gaussian processes X and Y
related by (4) that MI-in-frequency is related to coherence
CXY (ν) ∈ [0, 1], by MIXY (ν, ν) = − log (1 − CXY (ν)).
The proof is in the appendix. This result implies MI-infrequency between Gaussian processes related by (4) can
be estimated with the coherence. In addition, Theorem 3
also shows that MI-in-frequency between Gaussian processes
related by (4) can be estimated by estimating
the mutual
eR (νi ) , dX
eI (νi ) and dYeR (νj ), a
information between dX
three dimensional estimate as opposed to a four dimensional
estimate in general.
B. Discrete-time Stochastic Processes
A. Kernel Density Based MI-in-frequency (KDMIF) Estimator
1) Estimation of Samples of Spectral Process Increments:
The first step of the algorithm is estimating the samples of
spectral process increments of X and Y from N dependent
data samples. We assume there is a finite memory in both
these processes and choose a value for a parameter Nf , which
is much larger than the length of dependence or memory in
the data and determines the frequency resolution of our MI-infrequency estimates. We assume data in different windows are
independent of each other. Ideally, consecutive windows should
be separated to ensure no dependence across windows and
avoid the dependence across the window boundaries, but our
simulation results demonstrate that not separating the windows
doesn’t affect performance significantly. N samples of X are
split into Ns non-overlapping windows with Nf = NNs data
points in each window. Let us denote the samples in lth window
of X and Y respectively by two Nf element one-dimensional
vectors, xl and yl , for l = 1, 2, · · · , Ns .
Let us now focus on estimating
samples of the random
e (λi ). Let F xl (α) denote the discrete-time
variable dX
Fourier transform (DTFT) of xl at normalized frequency α.
For λi = Nif ∈ [0, 1] and i ∈ [0, Nf − 1], let us define de
xl (λi )
l
and integrated Fourier spectrum, x
e (λi ), by
We now extend the definition of MI-in-frequency between
continuous-time stochastic processes in (2), (3) to discretetime stochastic processes. In practice, we only have access
to data samples from a real-valued, discrete-time stochastic
process, sampled at a given Nyquist sampling frequency Fs .
Sampled signals have periodic spectra, with a period equalling
Fs . In addition, components in the process with frequencies in
i
the range [Fs /2, Fs ] correspond to negative frequencies [35].
P
xl (λi ) = F xl (λi ) and x
el (λi ) =
F xl (λm ) . (8)
Therefore, the actual frequency content in the signal is confined de
m=0
to [0, Fs /2]. We use normalized frequency λ = Fνs ∈ [0, 0.5] to
el (λi )
describe the frequency axis in case of discrete-time stochastic It is stated in [29] that the random variable for which x
processes, instead of ν which was used for continuous-time is just one realization, tends to the spectral process of X
stochastic processes. The MI-in-frequency between discrete- at λi in mean of order γ, for any γ > 0, as the number
time processes is therefore obtained by replacing νi , νj by of samples goes to infinity and assuming the underlying
the normalized frequencies λ1 , λ2 ∈ [0, 0.5] in (2), (3). distribution is stationary and satisfies a mixing assumption.
xl (λi ), which is the increment in x
el (λi ) between λi
Multivariate autoregressive models, commonly used to model Also, de
electro-physiological signals recorded from brain [20], [36], and λi + dλ, is just the DTFT of the samples in window
are a special case of the discrete-time equivalent of (4). The l. Calculating the DTFT with the FFT for each of the Ns
an Nf × Ns matrix, whose
ith
analytic expression for MI at frequency λ for such discrete- windows separately
yields
1
2
Ns
x (λi ) = de
x (λi ) , de
x (λi ) , · · · , de
x (λi ) is the
time Gaussian processes is therefore similarly obtained by row, de
e (λi ),
complex-valued
vector
containing
N
samples
of dX
replacing the frequencies ν by λ in (7), which is also equal
s
i
the
spectral
process
increments
of
X
at
λ
=
.
The
lth
to − log (1 − CXY (λ)). This shows that for the special case
i
Nf
of discrete-time Gaussian processes, MI-in-frequency metric is element of de
x (λi ), de
xl (λi ) = de
xlR (λi ) + ide
xlI (λi ), is a
e
equivalent to coherence and the definitions in [24], [25].
particular realization of dX (λi ). A similar procedure is used
to obtain the Ns samples of the spectral process increments of
Y at λj = Njf , j ∈ [0, Nf − 1] and the resulting samples are
1
IV. DATA -D RIVEN E STIMATION OF MI- IN - FREQUENCY
denoted by de
y (λj ) = de
y (λj ) , de
y 2 (λj ) , · · · , de
y Ns (λj ) .
We describe two data-driven estimators–a kernel density
2) Estimating MI-in-frequency: The MI-in-frequency
based (KDMIF) and a nearest neighbor based (NNMIF) estimate is now
the
obtained from
l Ns samples,
l
c XY (λi , λj ), between
estimator to estimate MI-in-frequency, MI
de
xlR λi , de
xlI λi
and
de
yR
λj , de
yI λj ,
for
λi component of X and λj component of Y . The input to l = 1, 2, · · · , Ns , using a kernel density based plug-in
both these algorithms are the N samples of X and Y . The nonparametric estimator [32]. The Ns data samples are split
first step in both KDMIF and NNMIF estimators involves into Ntr training and Nts test samples. The training data is
e (λi ) used to estimate the four-dimensional joint probability density
estimating the samples of spectral process increments dX
e
eR (λi ) , dX
eI (λi ) , dYeR (λj ) , dYeI (λj ) . The density is
and dY (λj ), of X at λi and of Y at λj respectively. In the P dX
second step, the KDMIF estimator uses the kernel density based estimated using a kernel density estimator with Gaussian
MI estimator [32], [37], whereas NNMIF estimator uses the kernels, the optimal bandwidth matrix selected using smoothed
k-nearest neighbor based MI estimator [32], [38] to estimate cross-validation criterion [37] and implemented using ‘ks’
e (λi ) package in R [39]. The joint density is marginalized to estimate
MI from the samples of spectral process increments, dX
e
eR (λi ) , dX
eI (λi ) and
and dY (λj ).
the two-dimensional densities, P dX
5
P dYeR (λj ) , dYeI (λj ) , by recognizing that the bandwidth
matrix for the two-dimensional marginal is the appropriate
2 × 2 sub-matrix from the 4 × 4 bandwidth matrix of the joint
density. The estimates of the joint and the marginal densities
at the Nts test samples are plugged into the following equation
(9) to estimate MI-in-frequency.
c XY (λi , λj )
MI
b
P
P
= N1ts
log b
l
l
de
xlR (λi ),de
xlI (λi ),de
yR
(λj ),de
yIl (λj )
.
l (λ ),de
b de
P de
xlI (λi ) P
yR
yIl (λj )
xlR (λi ),de
j
(9)
B. Nearest Neighbor Based MI-in-frequency (NNMIF) Estimator
1) Estimation of Samples of Spectral Process Increments:
The first step in the nearest neighbor based MI-in-frequency estimator is exactly same as that of KDMIF estimator. Following
the steps described in section IV-A1, we estimate de
xl (λi ) and
l
de
y (λj ), for l = 1, 2, · · · , Ns , the Ns samples of the spectral
process increments of X at λi and Y at λj respectively.
2) Estimating MI-in-frequency: MIXY (λi , λj ) is now
estimated from de
xl (λi ) ∈ R2 and de
y l (λj ) ∈ R2 , for
l = 1, 2, · · · , Ns using nearest neighbor based MI estimator
[38]. We apply the first version of the algorithm in
e (λi ) and
[38] to two-dimensional random variables dX
e
c
dY (λj ) to compute MIXY (λi , λj ). Consider
the joint four
e (λi ) , dYe (λj ) ∈ R4 . The distance
dimensional space dX
between two data points with indices l1 , l2 ∈ [1, Ns ]
is calculated
using the infinity norm, according to
max kde
xl1 (λi ) − de
xl2 (λi ) k, kde
y l1 (λj ) − de
y l2 (λj ) k .
Let l denote the
distance between the data sample
de
xl (λi ) , de
y i (λj ) and its K th nearest neighbor, for
l = 1, 2, · · · , Ns . We used K = 3 in this paper [30]. Let
e (λi ) and
nlx and nly denote the number of samples of dX
e
dY (λj ) within an infinity norm ball of radius less than l
centered at de
xl (λi ) and de
y i (λj ) respectively. From [38], the
MI-in-frequency between X and Y at normalized frequencies
λi and λj is given by
c XY (λi , λj ) = ψ (K) + ψ (Ns )
MI
Ns
P
− N1s
ψ nlx + 1 + ψ nly + 1 , (10)
l=1
where ψ (·) is the Digamma function.
of independence. The permuted MI estimates will be almost
zero since the permutations make the spectral processes almost
c XY (λi , λj ), is
independent. If the actual MI estimate, MI
judged larger than all the permuted Np estimates, then there
is a statistically significant dependence between the processes
at these frequencies.
V. MI BETWEEN DATA WITH T EMPORAL D EPENDENCIES
We now use MI-in-frequency to estimate mutual information
between dependent data. The data-driven MI estimator, summarized in Algorithm 1, takes in N samples of X and Y as input
and outputs the mutual information between X and Y , Î (X; Y ),
c XY (λi , λj ), where λi = i , λj = j ,
by estimating MI
Nf
Nf
∀ (i, j) such that i, j ∈ [0, Nf − 1].
Algorithm 1: Mutual Information Estimator
Data: (x [n] , y [n]), for x [n] , y [n] ∈ R, n ∈ [0, N − 1].
Result: Î (X; Y )
Algorithm:
c XY (λi , λj ) at all possible pairs (λi ,λj ),
A) Estimate MI
using either the KDMIF or the NNMIF estimator.
Identify the sets Λx , Λy , such that for each λip ∈ Λx
c XY λi , λj is
there exists a λjq ∈ Λy such that MI
p
q
statistically significant and vice-versa. Let P, Q
respectively denote the cardinality of Λx , Λy .
e (Λx ) = dX
e (λi ) , · · · , dX
e (λi ) ∈ R2P ,
B) Let dX
1
P 2Q
dYe (Λy ) = dYe (λj1 ) , · · · , dYe λjQ ∈ R . The
mutual information between X and Y is given by
1
e (Λx ) ; dYe (Λy ) ,
Î dX
Î (X; Y ) = max(P,Q)
where the right hand side is estimated from Ns i.i.d.
samples using any nonparametric MI estimator [32].
A. Identifying Coupled Frequencies
The first step in our MI estimator involves estimating the
c XY (λi , λj ), between λi = i frequency
MI-in-frequency, MI
Nf
C. Significance Testing
component in X and λj = Njf component in Y , for all
The statistical significance of the MI-in-frequency estimates (i, j) such that i, j ∈ [0, Nf − 1] using either the KDMIF
obtained from both KDMIF and NNMIF estimators is now (section IV-A) or the NNMIF (section IV-B) algorithms.
tested using the following procedure. We permute the samples Statistical significance of the resulting estimates is assessed
in the vector de
x (λi ) randomly and estimate the MI-in- using the procedure described in section IV-C. The resultant
frequency between the permuted vector and the Ns samples MI-in-frequency estimates across all frequency pairs can be
of dYe (λj ). Instead of adding random phase or permuting graphically visualized by plotting the statistically significant
the phase time series, which are typically used to test the MI-in-frequency estimates on a two-dimensional image grid,
statistical significance of phase-amplitude coupling metrics whose rows and columns correspond to frequencies of X and
[40], we permute the samples of spectral process increments Y respectively. Let Λx and Λy respectively denote the set
since our metric can detect coupling across phase and amplitude of frequency components of X and Y , such that for each
exists at least one λjq ∈ Λy for which
jointly. This process is repeated Np times to obtain Np λip ∈ Λx , there
c XY λi , λj is statistically significant and vice-versa.
permuted MI-in-frequency estimates, under the null hypothesis MI
p
q
6
B. Estimating Mutual Information
The final step in our algorithm estimates MI between the
spectral process increments of X and Y at frequencies in Λx
and Λy respectively. With P, Q denoting the cardinality of
e (Λx ) and dYe (Λy ) denote the 2P
Λx , Λy respectively, let dX
and 2Q-dimensional random vector comprising the spectral
process increments of X, Y at all frequencies in Λx and Λy
respectively. We already computed Ns i.i.d. samples of these
two random vectors to estimate MI-in-frequency estimates in
the previous step of this algorithm. The desired MI estimate
e (Λx )
is computed from the mutual information between dX
and dYe (Λy ), which is estimated using the k-nearest neighbor
based estimator developed in [38], according to
1
e (Λx ) ; dYe (Λy ) .
Î (X; Y ) =
Î dX
(11)
max (P, Q)
The MI estimator in (11) can be further simplified for
discrete-time Gaussian processes. Without loss of generality,
consider two Gaussian processes X and Y , related by
y[n] = h1 [n] ∗ x[n] + h2 [n] ∗ w[n],
(12)
Finally, as we mentioned earlier, MI estimation between
Gaussian processes is a solved problem in the sense that we
can analytically compute it if the covariance of the Gaussian
processes is known [3] and there are several estimators whose
performance is thoroughly analyzed [4]. MI in frequency for
Gaussian processes is analyzed by Brillinger [24]. In this paper,
we extended Brillinger’s work to define MI-in-frequency for
any process. In the following section, we use simulated data to
validate that the extensions we proposed to any process in this
paper are still in agreement with the prior work on Gaussian
processes and also work for non-Gaussian processes.
VI. P ERFORMANCE E VALUATION ON S IMULATED DATA
The performance of the data-driven MI-in-frequency and
mutual information estimators described in section IV and
section V respectively is validated on simulated data. The
statistical significance of the estimates was assessed using the
procedure described in section IV-C. In addition, we compare
the performance of the MI-in-frequency estimators against
modulation index [8], [9], [22], a commonly used phaseamplitude coupling metric in neuroscience.
where h1 [n], h2 [n] are linear time-invariant (LTI) filters and
W is white Gaussian noise independent of X. For the model A. Comparing the KDMIF and NNMIF Estimators
Consider two stochastic processes X and Y , where X is a
in (12), which is the discrete-time equivalent of (4), the datawhite Gaussian process with standard deviation σx and Y is
driven estimation in (11) can be further simplified to
obtained by
NP
f /2
1
i
c
Î (X; Y ) = Nf
MIXY (λi ; λi ) , where λi = Nf . (13)
y[n] = h[n] ∗ x[n] + w[n],
(15)
i=0
This result is obtained because linear models do not introduce where W is a white Gaussian process with standard deviation
cross-frequency dependencies and because negative frequencies σw that is independent of X and h[n] is a linear time-invariant
do not carry any extra information. Furthermore, the relation- filter. We compared the performance of the kernel density
ship between the MI and the MI-in-frequency for two processes based and nearest neighbor based estimators by benchmarking
the estimates against the true value of MI-in-frequency and
related by (12) is stated in the following theorem.
the mutual information between X and Y for the model in
Theorem 4. Consider two discrete-time Gaussian stochastic (15). We used two different filers: a two-tap low pass filter,
processes X and Y related by (12). The mutual information h[n] = [β, 1 − β] , for β ∈ [0, 1] and a 33-tap bandpass filter
between these processes, a scalar, is given by
with passband in [0.15, 0.35] normalized frequency range. We
0.5
observed that modulation index, a popular CFC metric, was
R
I (X; Y ) = MIXY (λ, λ) dλ.
(14) unable to correctly detect and quantify the strength of cross0
frequency coupling for both these models.
The proof of the above theorem is in the appendix. This
1) Lowpass Filter: The samples of X and Y are generated
theorem means that MI between two Gaussian processes over from (15) with σx = σw = 1 and a lowpass filter with unitthe entire time can be obtained by integrating the contribution impulse response [β, 1 − β], for various values of β ∈ [0, 1].
from each frequency component. It is easy to see that the right The true value of MI-in-frequency at normalized frequency
hand side of (13) is just the Riemann sum of the integral on the λ ∈ [0, 0.5] is obtained substituting the parameters of this
right hand side of (14), which converges to the true value as model in (7) and is plotted in Fig. 1a for β = 0.5. In addition,
Nf tends to infinity. This implies our MI estimator converges the MI-in-frequency estimated by the KDMIF and NNMIF
to the true value for discrete-time Gaussian processes.
algorithms from N = 64 × 104 data samples, with Nf =
Note that the MI estimation algorithm does not make any 64, Ns = 104 is also plotted in Fig. 1a. It is seen that the
parametric assumptions on the underlying model between estimates from both algorithms follow the true value closely,
X and Y . The computation of MI via (11) can be greatly without the knowledge of the underlying model. In addition,
simplified by clustering the frequencies in Λx and Λy into we evaluate the bias and the rate of convergence of both these
groups such that there are no significant dependencies across algorithms as a function of Ns , with Nf = 64 in Fig. 1b.
groups and using the chain rule of mutual information. In The bias is defined as the average value of the ratio of MI-inaddition, if we observe after the first step that significant
MI
frequency estimate and its true value in the passband of the
in-frequency estimates occur only at (λi , λi ) ,∀i ∈ 0, Nf − 1 , lowpass filter. We observe that the NNMIF algorithm converges
then the MI can be estimated using (13).
faster and has lower bias than the KDMIF algorithm. We now
1
0.8
KDMIF
NNMIF
0
0
0.25
0.6
102
0.5
103
Mutual Information
True Value
KDMIF
NNMIF
0.35
0.36
1.1
0.7
Bias
MI between X and Y at 6
7
True Value
Kernel Density Based
Nearest Neighbor Based
0.26
0.16
0
104
0.5
Number of Samples, Ns
6
(a) MI-in-frequency
1
-
(b) Bias of the Estimators
(c) Mutual Information
KDMIF
NNMIF
True Value
KDMIF
NNMIF
0
0.25
6
(a) MI-in-frequency
0.8
0.5
0.6
102
103
Number of Samples, Ns
(b) Bias of the Estimators
104
Mutual Information
0.4
0
0.4
1
0.8
Bias
MI between X and Y at 6
Fig. 1. Comparing the performance of the kernel density based and nearest neighbor based estimators, KDMIF and NNMIF respectively, on simulated
generated from (15) using a two-tap lowpass filter. In Fig. 1a, the MI-in-frequency estimates obtained from KDMIF and NNMIF estimators along with the true
value of MI-in-frequency are plotted against the normalized frequency λ for β = 0.5. Fig. 1b plots the bias (mean of the ratio of the estimate and the true
value in the filter passband) against the number of data samples used for estimation for β = 0.5. Fig. 1c plots the MI estimate between X and Y obtained
from kernel density and nearest neighbor algorithms along with the true value of MI for β ∈ [0, 1].
True Value
Kernel Density Based
Nearest Neighbor Based
0.2
0
0.5
1
1.5
2
<w
(c) Mutual Information
Fig. 2. Comparing the performance of the kernel density based and nearest neighbor based estimators, KDMIF and NNMIF respectively, on simulated generated
from (15) using a 33-tap bandpass filter with passband in [0.15, 0.35] normalized frequency. In Fig. 2a, the MI-in-frequency estimates obtained from KDMIF
and NNMIF estimators along with the true value of MI-in-frequency are plotted against the normalized frequency λ for σw = 1. Fig. 2b plots the bias (mean
of the ratio of the estimate and the true value in the filter passband) against the number of data samples used for estimation for σw = 1. Fig. 2c plots the plots
the MI estimate between X and Y from kernel density and nearest neighbor algorithms along with the true value of MI for different values of σw ∈ [0.5, 2].
use both these algorithms to estimate the mutual information and Y , as evident from Fig. 2a, Fig 2c respectively. In addition,
between X and Y for β ∈ [0, 1]. The analytical expression for nearest neighbor based MI-in-frequency algorithm runs faster
the true value of MI1 for this model is derived in [41]. It is than kernel density based algorithm. We, therefore, conclude
evident from Fig. 1c that the MI estimates obtained from the that the nearest neighbor based MI-in-frequency algorithm
nearest neighbor based estimator is closer to the true value outperforms kernel density based algorithms and only depict
than those from the kernel density based estimator.
the results obtained from nearest neighbor based algorithm in
2) Bandpass Filter: The samples of X are generated from the remainder of the paper.
a standard white Gaussian random process with σx = 1 and
those of Y are generated from (15) using a 33-tap finite- B. Comparison with Modulation Index
impulse-response bandpass filter with passband in [0.15, 0.35]
We now compare the effectiveness of MI-in-frequency
normalized frequency range for different values of noise
against modulation index in detecting cross-frequency coupling,
standard deviation, σw ∈ [0.5, 2]. We used the kernel density
using the simulated model commonly used to validate CFC
and the nearest neighbor based algorithms to estimate the MImetrics [21], [22], [42]. Modulation index quantifies the relain-frequency and the mutual information between X and Y .
tionship between the phase and amplitude envelopes extracted
The true value of MI-in-frequency is obtained from (7) and
by the Hilbert transform [8]. Consider two random cosine
of mutual information is numerically calculated using power
waves, sl [n] and sh [n], at frequencies fl and fh respectively.
spectral density (chapter 10 in [34]). It is clear from Fig. 2b
Let fs denote the sampling frequency. The samples of timethat the nearest neighbor based algorithm converges to the true
series X and Y are generated from the following model:
value faster than the kernel density based algorithm. The nearest
neighbor based algorithm also provides more accurate estimates sl [n] = A cos 2π fl n + θ , sh [n] = A cos 2π fh n + θ
fs
fs
of both MI-in-frequency and mutual information between X
x[n] = sl [n] + w1 [n], y[n] = (1 + sl [n]) sh [n] + w2 [n], (16)
1 Note that for this particular model, mutual information is equal to the
directed information from X to Y and the analytical expression is given in
equation (18) in [41].
where A is a Rayleigh random variable with parameter 1 and
θ is a uniformly distributed random variable between 0 and
45
1.2
60
75
90
5 15
30
45
60
75
90
30
45
0.24
60
75
90
0
5 15
Frequency of X (Hz)
30
45
60
75
90
2.4
5
15
30
45
1.2
60
75
90
0
5 15
Frequency of X (Hz)
(a) MI-in-frequency
30
45
60
75
90
0
Frequency of X (Hz)
(b) Modulation Index
Frequency of Y (Hz)
30
0.48
5
15
Frequency of Y (Hz)
2.4
5
15
Frequency of Y (Hz)
Frequency of Y (Hz)
8
0.048
5
15
30
45
0.024
60
75
90
5 15
30
45
60
75
90
0
Frequency of X (Hz)
(c) MI-in-frequency
(d) Modulation Index
4
8
1.1
12
16
0
4
8
12
16
6j (#2!5 cycles=sample)
c Y Y (λi , λj )
(a) MI
0
4.4
0
4
8
2.2
12
16
0
4
8
12
16
6j (#2!5 cycles=sample)
c XY (λi , λj )
(b) MI
0
Mutual Information
2.2
0
6i (#2!5 cycles=sample)
6i (#2!5 cycles=sample)
Fig. 3. Comparing the performance of MI-in-frequency against modulation index in detecting cross-frequency coupling in data generated from (16). In Fig. 3a
and Fig. 3b, MI-in-frequency estimates obtained from nearest neighbor algorithm and modulation index are plotted respectively, when fl = 5 Hz and fh = 60
Hz in (16). Fig. 3c and Fig. 3d respectively plot the MI-in-frequency estimates and modulation index estimates, when fl = 15 Hz and fh = 60 Hz in (16).
2.8
1c
MIXY
2
(60 ; 260 )
^I (X; Y )
1.6
0.4
0
5
10
<w
(c) MI between X and Y
Fig. 4. (a) MI-in-frequency estimates from the nearest neighbor based algorithm between the frequency components within the random processes Y , obtained
from the single cosine data-generation model, (18) with σw = 1. Note that the MI-in-frequency estimates along the principal diagonal are not plotted, since
they are equal to ∞. (b) MI-in-frequency estimates between random processes X and Y related by the single cosine data-generation model with σw = 1.
It is clear that MI-in-frequency estimator correctly identifies the pairwise frequency dependencies. (c) MI-in-frequency between X at λ0 and Y at 2λ0 ,
c XY (λ0 , 2λ0 ), obtained from (10) along with the MI estimate between X and Y , Î (X; Y ), obtained from Algorithm 1 for various values of the noise
MI
standard deviation, σw .
2π that is independent of A. w1 [n], w2 [n] are samples of i.i.d estimated CFC (note the smearing in Fig. 3d, when compared to
white Gaussian noise process with standard deviation 1. We Fig. 3b) [22], [42]. In addition, we tested modulation index on
generated samples from this model with fl = 5 Hz, fh = 60 data generated from (15) and (17) and found that modulation
Hz and fs = 200 Hz. MI-in-frequency between X and Y is index is unable to detect the cross-frequency coupling for these
estimated using the nearest neighbor based algorithm from relationships. This is not surprising since the modulation index
N = 40 × 104 samples with Ns = 104 and plotted in Fig. 3a. like metrics are tuned to detect CFC when the underlying
Modulation index between X and Y estimated by using the coupling is of the form in (16), whereas the MI-in-frequency
Matlab toolbox [21], with the amplitude envelope estimated defined in this paper overcomes this shortcoming, as evident
by the Hilbert transform and is plotted in Fig. 3b. It is clear from its performance on various simulated models.
that both MI-in-frequency and modulation index successfully
detect the cross-frequency coupling between 5 Hz component C. Nonlinear Models
of X and {55, 60, 65} Hz components of Y for these parameter
We now consider square nonlinearity, where the random
values. We then generated X and Y from (16) with fl = 15 processes X and Y are related by
Hz and all other parameter values unchanged. Fig. 3c plots the
y[n] = x[n]2 + w[n],
(17)
MI-in-frequency estimates obtained via NNMIF algorithm and
as expected, we detect the CFC between 15 Hz component of
where w[n] is white Gaussian noise with standard deviation
X and {45, 60, 75} Hz components of Y . However, modulation
σw . Modulation index was not able to detect and quantify
index, depicted in Fig. 3d, was not able to correctly detect the
the cross-frequency coupling for this model. We estimated
CFC between X and Y for these parameter values. In addition,
the MI-in-frequency between frequency components within Y ,
the strength of the modulation index decreased from around c
MIY Y (λi , λj ), between the frequency components of X and
0.5 when fl = 5 Hz in Fig. 3b to 0.05 when fl = 15 Hz in
c XY (λi , λj ), and the mutual information between X and
Y , MI
Fig. 3d. This is because metrics like modulation index can only
Y , Î (X; Y ), from N = 32 × 104 samples of X and Y with
detect the CFC correctly with good frequency resolution only
Ns = 104 , for different values of noise standard deviation,
when one of the frequencies involved is very small compared
σw ∈ [0, 10]. Computing the true value of MI-in-frequency and
to the other frequency. Otherwise, the bandwidth of the filter
mutual information is nontrivial because of the nonlinearity. The
used to extract the phase and the amplitude envelope should
performance of the algorithms is assessed by checking if they
be larger, which will reduce the frequency resolution in the
detect the cross-frequency coupling at expected frequency pairs
4
8
0.9
12
16
0
4
8
12
!5
16
0
1
4.4
0
4
8
2.2
12
16
0
4
8
12
!5
16
6j (#2 cycles=sample)
6j (#2 cycles=sample)
c Y Y (λi , λj )
(a) MI
c XY (λi , λj )
(b) MI
0
Mutual Information
1.8
0
6i (#2!5 cycles=sample)
6i (#2!5 cycles=sample)
9
^I (X; Y )
0.6
0.2
0
5
10
<w
(c) MI between X and Y
Fig. 5. (a) MI-in-frequency estimates from the nearest neighbor based algorithm between the frequency components within the random processes Y ,
obtained from the two cosine data-generation model, (19). The MI-in-frequency estimates are not plotted along the diagonal, since they are equal to ∞. (b)
MI-in-frequency estimates between random processes X and Y related by the two cosine data-generation model. It is clear that MI-in-frequency estimator
correctly identifies the pairwise frequency dependencies between X and Y . (c) Î (X; Y ), the MI estimate between X and Y obtained from Algorithm 1 for
various values of the noise standard deviation, σw .
4
6
A1 , A2 , and λ1 = 32
. As before, the frequency
, λ2 = 32
components of X are statistically independent. However, after
some basic algebra, it is easy to see that the all possible pairs of
frequency components of Y in {0, λ2 − λ1 , 2λ1 , λ2 + λ1 , 2λ2 }
1) Random Cosine with Squared Nonlinearity: The samples are statistically dependent, except for (2λ1 , 2λ2 ) frequency pair,
and we expect to see statistically significant MI-in-frequency
of X are generated from a random cosine wave,
estimates between these frequency components. This is conx[n] = A cos (2πλ0 n + θ) ,
(18) firmed by Fig. 5a, which plots the MI-in-frequency estimates
within Y , generated with σw = 1 and obtained by the NNMIF
where A is a Rayleigh random variable with parameter 1,
algorithm. The pairwise frequency dependencies between X
θ is a uniform random variable between 0 and 2π that is
and Y occur at (λ1 , 0), (λ1 , λ2 − λ1 ), (λ1 , 2λ1 ), (λ1 , λ2 + λ1 ),
4
independent of A and λ0 = 32 . It is easy to see that frequency
(λ2 , 0), (λ2 , λ2 − λ1 ), (λ2 , λ2 + λ1 ) and (λ2 , 2λ2 ). Fig. 5b
components of X are statistically independent and this is
plots the estimates of pairwise MI-in-frequency between X
confirmed by the NNMIF estimator. However, because of
and Y generated with σw = 1 and obtained by the datathe square nonlinearity in (17), the DC component of Y and
driven NNMIF algorithm using (10). The algorithm correctly
the 2λ0 component of Y will be statistically dependent and
identifies all the dependent frequency pairs and P = 2, Q = 5.
this is confirmed by Fig. 4a, which plots the MI-in-frequency
We then apply the algorithm described in section V and plot
between components of Y generated with σw = 1 using
the estimates the MI for different values of noise standard
the NNMIF algorithm. The common information between
deviation σw in Fig. 5c. Again, the MI decreases with increasing
these two processes will be present between λ0 component of
noise power, as expected. These different models validate
X and the {0, 2λ0 } components of Y . This cross-frequency
the superiority of MI-in-frequency over other existing metrics
dependence is confirmed by Fig. 4b, which plots the estimates
to detect cross-frequency coupling and also demonstrate the
of MI-in-frequency between X and Y obtained by the NNMIF
performance and accuracy of the data-driven MI-in-frequency
algorithm from (10): we observe that significant dependencies
and MI estimators.
occur only at (λ0 , 0) and (λ0 , 2λ0 ) frequency pairs. As a
result, P = 1, Q
MI estimate from
= 2. The
Algorithm 1,
VII. CFC IN S EIZURE O NSET Z ONE
1
e
e
e
Î (X; Y ) = 2 Î dX(λ0 ); dY (0), dY (2λ0 )
is plotted in
Epilepsy is a common neurological disorder characterized
Fig. 4c. The MI estimate decreases with increasing σw as
by
repeated, unprovoked seizures. The seizure onset zone
expected. In addition, we note for this model that the DC
(SOZ)
comprises regions of the brain that are responsible for
component of Y does not contain any extra information about
generating
and sustaining seizures [3]. Surgical resection of the
X,given the 2λ0 component of
Y . Therefore, we expect
seizure
onset
zone is the prescribed treatment for a large portion
1
1
e
e
e
c XY (λ0 ; 2λ0 ), a re= 2 MI
2 Î dX(λ0 ); dY (0), dY (2λ0 )
of medically refractory epilepsy patients with focal epilepsy.
sult verified in Fig. 4c, since the two curves are very close.
However, surgical resection risks damage to critical functional
2) Two Random Cosines with Squared Nonlinearity: The zones that are frequently adjacent or even overlapping with
samples of random process X are generated according to
the seizure focus, depending on location of the focus [45]. An
ideal solution might be a closed-loop neuromodulation strategy
x[n] = A1 cos (2πλ1 n + θ1 ) + A2 cos (2πλ2 n + θ2 ) , (19)
that stimulates the epileptic [41], [46] and other networks [47]
where A1 , A2 are independent Rayleigh random variables at the optimal frequency with spatial and temporal specificity
with parameter 1, θ1 , θ2 are independent uniformly distributed [6], [7]. In this paper, we focus on learning more about the
random variables between 0 and 2π that are independent of characteristic frequencies and the spatial specificity of epileptic
and by checking if the mutual information estimates decrease
with increasing noise power as expected. We considered two
different models for the stochastic process X, such that its
samples are dependent across time.
10
TABLE I
C LINICAL D ETAILS OF THE PATIENTS A NALYZED .
Patient
Number of Seizures
Analyzed
Age/Sex
P1
3
22/M
P2
3
61/M
P3
2
29/M
P4
3
21/F
P5
3
24/M
P6
3
35/M
P7
3
26/M
P8
3
41/M
P9
2
18/F
Seizure Onset Zone
RAH 1-2, RPH 4,
RAMY 2-3
LAH 2-4, LPH 2
PD 4, 5
LF 28, LP 4
MST 1, TP 1, HD 1
LPH 5, 6, LPSM 8, LMH 5,
RMH 4, 5, RPSM 7
AH 3-5, PH 4
AMY 2, 3
AH 1, 2, 5, PH 5
TOP 3, 4
LAH 5, LAMY 3
RAH 3-5, LPH 6,
RPH 5-7
Outcome
of Surgery
Class IA
Class IIIA
Class IA
Class IA
Class IB
Class IA
Class IIB
N/A
Class IB
The full forms of the electrodes in seizure onset zone column in Table I: RAH - right anterior hippocampus, RPH - right posterior hippocampus,
RAMY - right amygdala, PD - posterior hippocampal depth, MST - mid-subtemporal lobe, TP - temporopolar, HD - hippocampal depth and
AST - anterior sub-temporal lobe, LMH - left mid hippocampus, AH - anterior hippocampus, PH - posterior hippocampus, AMY - amygdala,
TOP - temporo-occipito-parietal. The outcomes are in Engel epilepsy surgery outcome scale [43], [44]: “Class IA - completely seizure free
since surgery, class IB - non disabling simple partial seizures only since surgery, class IIB - rare disabling seizures since surgery (‘almost
seizure-free’), class IIIA - worthwhile seizure reduction, class IV - no worthwhile improvement".
networks. Specifically, we investigate cross-frequency coupling
between various regions in the seizure onset zone during the
evolution of seizures and identify the frequencies with strong
coupling. We estimate the cross-frequency coupling (CFC)
from ECoG data recorded from the SOZ electrodes using our
nearest neighbor based MI-in-frequency estimator. We infer
the characteristics of CFC within and between various regions
inside the seizure onset zone.
We analyzed ECoG data, sampled at Fs = 1 kHz, from a
total of 25 seizures recorded from nine patients with medial
temporal lobe epilepsy. Clinical details of the patients, along
with the seizure onset zone identified from ECoG data [41],
are summarized in Table I. The seizure start and end time were
marked by the neurologist. We analyzed ECoG recordings
from SOZ electrodes during preictal (window spanning up to
3 minutes immediately before the seizure starts), ictal (during
seizures) and postictal (window spanning up to 3 minutes
immediately after the seizure ends) periods. We only focussed
on the oscillations in alpha (7.5-12.5 Hz), beta (12.5 - 30
Hz), gamma (30-80 Hz) and ripples (80-200 Hz), excluding
60 Hz line noise and its harmonics. The CFC oscillations are
analyzed at spectral resolution of 10 Hz by choosing Nf = 100,
and the exact frequencies considered are {10, 20, · · · , 200} Hz,
excluding {60, 120, 180} Hz. The resulting 17×17 CFC matrix
from each ECoG electrode and between all pairs of ECoG
electrodes in the SOZ is estimated using nearest neighbor based
estimator (section IV-B) during preictal, ictal and postictal
periods during all the twenty five seizures.
1
5
7
11
13
17
19
20
1
5 7
1113
171920
Fig. 6. Binary mask plotting the frequency pairs with statistically significant
differences across ahypotheses tested after applying false discovery rate
correction. White and black colored regions represent frequency pairs with
and without statistically significant variation respectively.
CFC within a SOZ electrode, between two electrodes in the
same anatomical region and between electrodes in different
anatomical regions. For instance, consider patient P1. ECoG
electrodes in the SOZ of patient P1 are grouped into three
different anatomical regions–RAH, RPH, and RAMY (Table I).
We estimated 5 CFC matrices, one per SOZ electrode, to infer
the average CFC within an electrode in SOZ in this patient.
We estimated 20 CFC matrices between all pairs of electrodes
in the SOZ. Of these, 4 CFC matrices (2 to learn the CFC
between the 2 SOZ electrodes in RAH and 2 to learn the CFC
between the 2 SOZ electrodes in RAMY regions) are grouped
to learn the average CFC between electrodes in the same
We then grouped the ECoG electrodes into distinct anatom- anatomical region in SOZ. The remaining 16 CFC matrices are
ical regions based on their label and analyzed the average grouped to learn the CFC between different regions inside the
11
1.9
1
5
7
0.7
1
5
7
1
11
13
17
19
20
1
57
1113
171920
0.2
(a) Preictal period
0.4
1
5
7
0.3
11
13
-0.2
11
13
17
19
20
1
57
1113
171920
-0.2
(b) Difference between ictal and preictal period
17
19
20
1
57
1113
171920
-0.8
(c) Difference between postictal and ictal period
Fig. 7. Cross-frequency coupling within an electrode inside the seizure onset zone. In Fig. 7a, MI-in-frequency estimates over the frequencies {10, 20, · · · , 200}
Hz excluding {60, 120, 180} Hz are obtained from each electrode in SOZ during preictal period and the median of the resulting CFC estimates from all the
SOZ electrodes in the twenty five seizures from the nine temporal lobe epilepsy patients analyzed is plotted. In Fig. 7b, MI-in-frequency estimates are obtained
from each electrode in SOZ in the ictal period and the difference between the median CFC estimate in ictal and preictal period is plotted. Similarly, Fig. 7c
plots the difference in the median CFC between postictal and ictal periods.
0.1
1
5
7
0.2
1
5
7
0.05
11
13
17
19
20
1
57
1113
171920
(a) Preictal period
0
0.26
1
5
7
0.1
11
13
17
19
20
1
57
1113
171920
0
(b) Difference between ictal and preictal period
0.13
11
13
17
19
20
1
57
1113
171920
0
(c) Difference between postictal and ictal period
Fig. 8. Cross-frequency coupling between electrodes in different regions inside the seizure onset zone. In Fig. 8a, MI-in-frequency estimates over the frequencies
{10, 20, · · · , 200} Hz excluding {60, 120, 180} Hz are obtained between electrodes in different SOZ regions during the preictal period and the median of the
resulting CFC estimates from the twenty five seizures in the nine temporal lobe epilepsy patients analyzed is plotted. In Fig. 8b, MI-in-frequency estimates are
obtained between electrodes in different SOZ regions from the ictal period and the difference between the median CFC estimate from the ictal and preictal
period is plotted. Similarly, Fig. 8c plots the difference in the median CFC between postictal and ictal periods.
SOZ. The estimated CFC matrices are grouped into these three
spatial categories for all the nine patients during preictal, ictal
and postictal periods. We only presented the results for CFC
within a SOZ electrode and between electrodes in different
SOZ regions during preictal, ictal and postictal periods.
level of 0.01 to account for multiple comparisons [49]. The
frequency pairs with significant statistical variation across
all the hypotheses considered are depicted using a binary
mask in Fig. 6, in which black and white colored regions
respectively represent frequency pairs without statistically
significant variation and with statistically significant variation.
Lack of statistical significance at the black regions in Fig. 6
could be because of insufficient data or could be due to a
neuronal transition mechanism as the brain moves from preictal
to ictal to postictal state. It is important to note that if we
tested only a subset of the thirteen hypotheses, then some of
the frequency pairs in black colored regions in Fig. 6 could
become statistically significant.
We used the permutation procedure outlined in section IV-C
to estimate the CFC under the null hypothesis and assess
the significance of the estimated CFC values across the six
conditions considered (CFC during preictal, ictal, postictal
periods within a SOZ electrode and between electrodes in
different SOZ regions) using Wilcoxon signed-rank test [48].
We also used the Wilcoxon signed-rank test to identify the
frequency pairs with significant variation in CFC between
preictal and ictal periods and between ictal and postictal periods,
The median CFC within an electrode in SOZ during preictal,
both within a SOZ electrode and between electrodes in different ictal and postictal periods grouped across all twenty five
SOZ regions (four hypotheses in total). In addition, we used seizures in nine patients analyzed is plotted in Fig. 7. In
the Mann-Whitney U-test [48] to identify frequency pairs Fig. 7a, median CFC in the preictal period is plotted, while the
with significant changes in CFC within a SOZ electrode and difference between median CFC in the ictal and preictal period,
between electrodes in different SOZ regions across preictal, and between postictal and ictal period is plotted in Fig. 7b
ictal and postictal periods (three hypotheses in total). We and Fig. 7c respectively. We need to multiply the binary mask
estimated 3621 p-values in total (13 × 17 × 16 + 5 × 17) in Fig. 6 with the plots in Fig.7 to obtain frequency pairs
and applied false discovery rate correction at a significance with significant statistical variation. The (i, j)th element in
12
the matrix in Fig. 7a is the median MI-in-frequency between
the 10i and 10j Hz frequency components during preictal
period across all SOZ electrodes in the twenty five seizures
analyzed. The principal diagonal in the three CFC matrices is
not plotted since MI-in-frequency between same frequencies
in a signal is infinity. It is clear from this figure that ripple
frequencies are heavily synchronized during preictal stage
within an electrode in SOZ. The synchronization between
all frequency pairs, particularly in gamma and ripples, seemed
to increase during the seizure when compared to just before
the seizure. And finally, the synchronization between highfrequency bands decreased, and low frequencies become more
synchronized amongst themselves and with high-frequencies
in the postictal period compared to the ictal period within an
electrode in SOZ.
The median CFC between electrodes in different SOZ regions
grouped across all twenty five seizures in nine patients analyzed
is plotted in Fig. 8. We need to multiply the binary mask in
Fig. 6 with the plots in Fig. 8 to obtain frequency pairs with
significant statistical variation. The median CFC during the
preictal period is plotted in Fig. 8a. It is clear from the principal
diagonal that neighboring regions in SOZ have weak linear
interactions (possibly due to their spatial proximity) just before
a seizure starts. From Fig. 7a and Fig. 8a, it is clear that the
CFC strength is much lower between regions when compared
to within an electrode. From Fig. 8b, we observe a small
increase in CFC between regions as the brain transitions to
seizure state. However, the increase is much smaller between
regions when compared to the increase observed in Fig. 7b,
which suggests that different SOZ regions potentially drive
the rest of the brain into a seizure state independently, which
implies any non-surgical treatment should target these different
regions simultaneously to disrupt the epileptic network. As
the brain transitions to postictal state, we observe a sharp
increase in linear coupling between electrodes in different
SOZ regions, which suggests that postictal periods, unlike ictal
periods, are characterized by an increase in linear interactions.
These results highlight the role of gamma and ripple highfrequency oscillations (HFOs) during seizures and the dynamic
reorganization of synchronization between neuronal oscillations
inside the seizure onset zone during the course of a seizure.
These results also suggest that multiple regions inside the
seizure onset zone might have to be targeted simultaneously
using neuromodulation techniques to control seizure activity.
VIII. D ISCUSSION AND C ONCLUSIONS
Detecting and quantifying relationships between multiple
data streams recorded from a physical system is of interest in
many science and engineering disciplines. However, since the
underlying model is often unknown and nonlinear, detecting
and quantifying the relationships in data is very challenging
in most real-world applications. Brownian distance covariance
[50], maximal information coefficient [51] are some of the
recent works that attempt to overcome this challenge in the
most general case. Furthermore, in neuroscience, we are
also interested in decomposing the relationships in frequency
domain and estimating cross-frequency coupling (CFC) from
electrophysiological recordings. Motivated to understand nonlinear frequency coupling in electrophysiological recordings
from the brain and inspired by [26], we defined MI-in-frequency
between stochastic processes that are not necessarily Gaussian
and estimated it using data-driven estimators. We found that the
nearest neighbor based MI-in-frequency estimator outperforms
the kernel-based MI-in-frequency estimator. MI-in-frequency
can be thought of as ‘coherence’ for non-Gaussian signals. At
a first glance, CFC could be estimated by first filtering the
data into appropriate frequency bands and then applying the
techniques in [21], [50], [51]. However, [22] summarizes all the
caveats and confounds in estimating CFC using this approach.
In contrast, the MI-in-frequency metric estimates CFC without
explicitly band-pass filtering the data into appropriate frequency
bands.
We then compared the performance of MI-in-frequency
against modulation index [8], [21], a popular CFC metric used
to measure phase-amplitude coupling that involves bandpass
filtering, on simulated data and observed that MI-in-frequency
outperforms the existing metrics used to estimate CFC. The
main advantages of the MI-in-frequency approach over existing
methods to estimate CFC are that it detects statistical independence, detects dependencies across phase and amplitude
jointly, applies to linear and nonlinear dependencies, and is not
dependent on parameters like the filter bandwidth. Our approach
will need more data when compared with coherence since MIin-frequency detects both linear and nonlinear dependencies
in frequency. From the simulation results on linear models,
we need about 103 samples to be within 10% of the true
value. For the ECoG data sampled at 1 kHz and a desired
spectral resolution of 10 Hz, this implies the total number of
data samples is of the order of 100 seconds or a couple of
minutes, which is roughly the size of preictal, ictal and postictal
windows used in section VII. In summary, we developed a
metric to detect statistical independence in frequency which
outperforms existing CFC metrics and for the first time, utilized
frequency domain to estimate mutual information over time
between dependent data.
The MI-in-frequency metric can be further extended along
several directions and some of them are outlined here. We
can move to wavelet based analysis to improve the fixed
time-frequency resolution of our Fourier-based approach in
future work. The assumption of data stationarity in observation
window (also assumed by most CFC metrics) can be potentially
relaxed by utilizing time-frequency distributions and developing
heuristics to measure the dependencies across frequency.
However, the inherent trade-off involved is that we are not
guaranteed to detect statistical independence. It is also possible
to define and estimate conditional MI-in-frequency to eliminate
indirect coupling estimated between two signals because of a
third signal which is coupled to both.
We then apply the MI-in-frequency estimators to infer the
coupling between neuronal oscillations before, during and
after seizures in the seizure onset zone. Spatially, we used
the electrode labels to identify the different regions in the SOZ.
This is just one possible way to analyze the spatial variation
in CFC. Some of the other possible options include using the
distance between electrodes or using the underlying neuronal
13
1.6
1
5
7
0.16
1
5
7
0.8
11
13
17
19
20
1
57
1113
171920
(a) Within a SOZ electrode
0.1
seizure activity and will serve as the first step towards the
development of a patient-specific, closed-loop, non-surgical
treatment for epilepsy.
0.08
11
13
IX. ACKNOWLEDGMENTS
17
19
20
1
57
1113
171920
0
(b) Between electrodes in different SOZ
regions
Fig. 9. Cross-frequency coupling during interictal periods. In Fig. 9a, MIin-frequency estimates over the frequencies {10, 20, · · · , 200} Hz excluding
{60, 120, 180} Hz are obtained from each electrode in SOZ during interictal
period and the median of the resulting CFC estimates from all the SOZ
electrodes in patients P1 and P2 is plotted. In Fig. 9b, MI-in-frequency estimates
over the frequencies {10, 20, · · · , 200} Hz excluding {60, 120, 180} Hz are
obtained between electrodes in different SOZ regions during the interictal
period and the median of the resulting CFC estimates in patients P1 and P2
is plotted.
cell types to split the electrodes into different regions in SOZ.
Our MI-in-frequency metric provides a framework that can
be utilized to learn the CFC characteristics for any desired
spatial grouping. In addition, the frequency resolution of our
estimated CFC was constant and wavelet transform, instead of
Fourier transform, can be utilized to provide greater resolution
at lower frequencies.
We observed that the high-frequency synchronization within
an ECoG electrode in SOZ increases during seizures and
decreases immediately after the seizure, which is accompanied
by an increase in low-frequency coupling. However, the
coupling between different anatomical regions in SOZ does
not increase noticeably during seizures and is also followed
by a large increase in linear interactions immediately after
a seizure. These observations suggest that seizure activity is
characterized by nonlinear interactions and is potentially due
to the independent efforts by various regions within SOZ,
which implies that all these regions are potential spatial targets
for electrical stimulation. Furthermore, we did a preliminary
investigation to learn if there are the differences in CFC between
interictal periods and seizure periods. Fig. 9 plots the CFC
within an ECoG electrode and between ECoG electrodes in
different regions in SOZ during interictal period in two patients
(P1 and P2). Comparing Fig. 9 with Fig. 7a and Fig. 8a, it looks
like the CFC within a SOZ electrodes at higher frequencies
slightly increases, while CFC between electrodes in different
regions across the diagonal (or equivalently, linear interactions)
slightly decreases as the brain transitions from interictal to
preitctal periods. We plan to extend this analysis to a larger
patient cohort. Building a real-time seizure prediction system
utilizing the variations in CFC between interictal and seizure
periods is the focus of our current [52] and future work. In
addition, the CFC characteristics were patient-specific and we
presented the median CFC across all the patients considered.
Going forward, the MI-in-frequency metric should be applied
to infer the CFC between channels in SOZ and outside SOZ to
learn how SOZ drives the rest of the brain into a seizure state
in each epilepsy patient. The results from such an analysis will
improve our understanding of the CFC mechanisms underlying
The authors wish to thank Suganya Karunakaran for the
helpful discussions on statistical hypothesis testing and proofreading the manuscript.
R EFERENCES
[1] R. Malladi, D. H. Johnson, G. Kalamangalam, N. Tandon, and
B. Aazhang, “Measuring cross-frequency coupling using mutual information and its application to epilepsy,” in Cosyne Abstracts, Salt Lake
City, USA, 2017.
[2] ——, “Data-driven estimation of mutual information using frequency
domain and its application to epilepsy,” in Asilomar Conference on
Signals, Systems and Computers, 2017.
[3] H. O. Lüders, I. Najm, D. Nair, P. Widdess-Walsh, and W. Bingman,
“The epileptogenic zone: general principles,” Epileptic Disorders, vol. 8,
no. 2, pp. 1–9, 2006.
[4] F. Rosenow and H. Lüders, “Presurgical evaluation of epilepsy,” Brain,
vol. 124, no. 9, pp. 1683–1700, 2001.
[5] G. K. Bergey, M. J. Morrell, E. M. Mizrahi, A. Goldman, D. KingStephens, D. Nair, S. Srinivasan, B. Jobst, R. E. Gross, D. C. Shields
et al., “Long-term treatment with responsive brain stimulation in adults
with refractory partial seizures,” Neurology, vol. 84, no. 8, pp. 810–817,
2015.
[6] S. Sunderam, B. Gluckman, D. Reato, and M. Bikson, “Toward rational
design of electrical stimulation strategies for epilepsy control,” Epilepsy
& Behavior, vol. 17, no. 1, pp. 6–22, 2010.
[7] E. Krook-Magnuson and I. Soltesz, “Beyond the hammer and the scalpel:
selective circuit control for the epilepsies,” Nature neuroscience, vol. 18,
no. 3, pp. 331–338, 2015.
[8] R. T. Canolty, E. Edwards, S. S. Dalal, M. Soltani, S. S. Nagarajan, H. E.
Kirsch, M. S. Berger, N. M. Barbaro, and R. T. Knight, “High gamma
power is phase-locked to theta oscillations in human neocortex,” Science,
vol. 313, no. 5793, pp. 1626–1628, 2006.
[9] R. T. Canolty and R. T. Knight, “The functional role of cross-frequency
coupling,” Trends in Cognitive Sciences, vol. 14, no. 11, pp. 506–515,
2010.
[10] C. Alvarado-Rojas, M. Valderrama, A. Fouad-Ahmed, H. FeldwischDrentrup, M. Ihle, C. Teixeira, F. Sales, A. Schulze-Bonhage, C. Adam,
A. Dourado et al., “Slow modulations of high-frequency activity (40–140
Hz) discriminate preictal changes in human focal epilepsy,” Scientific
reports, vol. 4, 2014.
[11] K. Edakawa, T. Yanagisawa, H. Kishima, R. Fukuma, S. Oshino,
H. M. Khoo, M. Kobayashi, M. Tanaka, and T. Yoshimine, “Detection
of epileptic seizures using phase–amplitude coupling in intracranial
electroencephalography,” Scientific reports, vol. 6, 2016.
[12] M. Guirgis, Y. Chinvarun, M. del Campo, P. L. Carlen, and B. L.
Bardakjian, “Defining regions of interest using cross-frequency coupling
in extratemporal lobe epilepsy patients,” Journal of Neural Engineering,
vol. 12, no. 2, p. 026011, 2015.
[13] S. A. Weiss, A. Lemesiou, R. Connors, G. P. Banks, G. M. McKhann,
R. R. Goodman, B. Zhao, C. G. Filippi, M. Nowell, R. Rodionov et al.,
“Seizure localization using ictal phase-locked high gamma a retrospective
surgical outcome study,” Neurology, vol. 84, no. 23, pp. 2320–2328,
2015.
[14] S. Liu, Z. Sha, A. Sencer, A. Aydoseli, N. Bebek, A. Abosch, T. Henry,
C. Gurses, and N. F. Ince, “Exploring the time–frequency content of high
frequency oscillations for automated identification of seizure onset zone
in epilepsy,” Journal of Neural Engineering, vol. 13, no. 2, p. 026026,
2016.
[15] H. Zhou, Y. Li, Y.-L. Hsin, and W. Liu, “Phase-amplitude coupling
analysis for seizure evolvement using Hilbert Huang transform,” in IEEE
38th Annual International Conference of the Engineering in Medicine
and Biology Society (EMBC). IEEE, 2016, pp. 1022–1025.
[16] R. Zhang, Y. Ren, C. Liu, N. Xu, X. Li, F. Cong, T. Ristaniemi, and
Y. Wang, “Temporal-spatial characteristics of phase-amplitude coupling
in electrocorticogram for human temporal lobe epilepsy,” Clinical
Neurophysiology, vol. 128, no. 9, pp. 1707–1718, 2017.
14
[17] M. Amiri, B. Frauscher, and J. Gotman, “Phase-amplitude coupling is
elevated in deep sleep and in the onset zone of focal epileptic seizures,”
Frontiers in human neuroscience, vol. 10, 2016.
[18] M. Cotic, Y. Chinvarun, M. del Campo, P. L. Carlen, and B. L. Bardakjian,
“Spatial coherence profiles of ictal high-frequency oscillations correspond
to those of interictal low-frequency oscillations in the ecog of epileptic
patients,” IEEE Transactions on Biomedical Engineering, vol. 63, no. 1,
pp. 76–85, 2016.
[19] B. Frauscher, N. von Ellenrieder, F. Dubeau, and J. Gotman, “Different
seizure-onset patterns in mesiotemporal lobe epilepsy have a distinct
interictal signature,” Clinical Neurophysiology, vol. 128, no. 7, pp. 1282–
1289, 2017.
[20] L. Faes and G. Nollo, Multivariate frequency domain analysis of causal
interactions in physiological time series.
INTECH Open Access
Publisher, 2011.
[21] A. C. Onslow, R. Bogacz, and M. W. Jones, “Quantifying phase–
amplitude coupling in neuronal network oscillations,” Progress in
Biophysics and Molecular Biology, vol. 105, no. 1, pp. 49–57, 2011.
[22] J. Aru, J. Aru, V. Priesemann, M. Wibral, L. Lana, G. Pipa, W. Singer,
and R. Vicente, “Untangling cross-frequency coupling in neuroscience,”
Current Opinion in Neurobiology, vol. 31, pp. 51–61, 2015.
[23] R. Pascual-Marqui, P. Faber, T. Kinoshita, Y. Kitaura, K. Kochi, P. Milz,
K. Nishida, and M. Yoshimura, “The dual frequency RV-coupling
coefficient: a novel measure for quantifying cross-frequency information
transactions in the brain,” arXiv preprint arXiv:1603.05343, 2016.
[24] D. R. Brillinger, “Second-order moments and mutual information in the
analysis of time series,” Recent Advances in Statistical Methods, pp.
64–76, 2002.
[25] R. Salvador, A. Martinez, E. Pomarol-Clotet, J. Gomar, F. Vila, S. Sarro,
A. Capdevila, and E. Bullmore, “A simple view of the brain through a
frequency-specific functional connectivity measure,” Neuroimage, vol. 39,
no. 1, pp. 279–289, 2008.
[26] D. R. Brillinger and A. Guha, “Mutual information in the frequency
domain,” Journal of Statistical Planning and Inference, vol. 137, no. 3,
pp. 1076–1084, 2007.
[27] H. Cramér and M. Leadbetter, Stationary and related stochastic processes:
sample function properties and their applications, ser. Wiley series in
probability and mathematical statistics. Tracts on probability and statistics.
Wiley, 1967.
[28] H. J. Larson and B. O. Shubert, Probabilistic models in engineering
sciences. Wiley, 1979, vol. 2.
[29] D. R. Brillinger, Time Series: Data Analysis and Theory. Philadelphia,
PA, USA: Society for Industrial and Applied Mathematics, 2001.
[30] S. Khan, S. Bandyopadhyay, A. R. Ganguly, S. Saigal, D. J. Erickson III,
V. Protopopescu, and G. Ostrouchov, “Relative performance of mutual
information estimation methods for quantifying the dependence among
short and noisy data,” Physical Review E, vol. 76, no. 2, p. 026209,
2007.
[31] E. Schaffernicht, R. Kaltenhaeuser, S. S. Verma, and H.-M. Gross, “On
estimating mutual information for feature selection,” in International
Conference on Artificial Neural Networks. Springer, 2010, pp. 362–367.
[32] Q. Wang, S. R. Kulkarni, and S. Verdú, “Universal estimation of
information measures for analog sources,” Foundations and Trends in
Communications and Information Theory, vol. 5, no. 3, pp. 265–353,
2009.
[33] M. Pinsker, Information and information stability of random variables
and processes, ser. Holden-Day series in time series analysis. HoldenDay, 1964.
[34] T. M. Cover and J. A. Thomas, Elements of information theory. John
Wiley & Sons, 2012.
[35] A. V. Oppenheim, R. W. Schafer, J. R. Buck et al., Discrete-time signal
processing. Prentice Hall Englewood Cliffs, NJ, 1989, vol. 2.
[36] E. Pereda, R. Q. Quiroga, and J. Bhattacharya, “Nonlinear multivariate
analysis of neurophysiological signals,” Progress in Neurobiology, vol. 77,
no. 1, pp. 1–37, 2005.
[37] D. W. Scott, Multivariate density estimation: theory, practice, and
visualization. John Wiley & Sons, 2015.
[38] A. Kraskov, H. Stögbauer, and P. Grassberger, “Estimating mutual
information,” Physical Review E, vol. 69, no. 6, p. 066138, 2004.
[39] T. Duong et al., “ks: Kernel density estimation and kernel discriminant
analysis for multivariate data in R,” Journal of Statistical Software,
vol. 21, no. 7, pp. 1–16, 2007.
[40] D. Dvorak and A. A. Fenton, “Toward a proper estimation of phase–
amplitude coupling in neural oscillations,” Journal of Neuroscience
Methods, vol. 225, pp. 42–56, 2014.
[41] R. Malladi, G. Kalamangalam, N. Tandon, and B. Aazhang, “Identifying
seizure onset zone from the causal connectivity inferred using directed
information,” IEEE Journal of Selected Topics in Signal Processing,
vol. 10, no. 7, pp. 1267–1283, Oct 2016.
[42] J. I. Berman, J. McDaniel, S. Liu, L. Cornew, W. Gaetz, T. P. Roberts,
and J. C. Edgar, “Variable bandwidth filtering for improved sensitivity
of cross-frequency coupling metrics,” Brain Connectivity, vol. 2, no. 3,
pp. 155–163, 2012.
[43] J. Engel, “Update on surgical treatment of the epilepsies summary of the
second international palm desert conference on the surgical treatment of
the epilepsies (1992),” Neurology, vol. 43, no. 8, pp. 1612–1612, 1993.
[44] C. Tonini, E. Beghi, A. T. Berg, G. Bogliun, L. Giordano, R. W. Newton,
A. Tetto, E. Vitelli, D. Vitezic, and S. Wiebe, “Predictors of epilepsy
surgery outcome: a meta-analysis,” Epilepsy Research, vol. 62, no. 1, pp.
75–87, 2004.
[45] U. Gleissner, R. Sassen, M. Lendt, H. Clusmann, C. Elger, and
C. Helmstaedter, “Pre-and postoperative verbal memory in pediatric
patients with temporal lobe epilepsy,” Epilepsy research, vol. 51, no. 3,
pp. 287–296, 2002.
[46] S. Karunakaran, M. J. Rollo, K. Kim, J. A. Johnson, G. P. Kalamangalam,
B. Aazhang, and N. Tandon, “The interictal mesial temporal lobe epilepsy
network,” Epilepsia, 2017.
[47] K. Kim, A. D. Ekstrom, and N. Tandon, “A network approach for
modulating memory processes via direct and indirect brain stimulation:
toward a causal approach for the neural basis of memory,” Neurobiology
of learning and memory, vol. 134, pp. 162–177, 2016.
[48] G. W. Corder and D. I. Foreman, Nonparametric statistics: A step-by-step
approach. John Wiley & Sons, 2014.
[49] Y. Benjamini and Y. Hochberg, “Controlling the false discovery rate:
a practical and powerful approach to multiple testing,” Journal of the
Royal Statistical Society. Series B (Methodological), pp. 289–300, 1995.
[50] G. J. Székely, M. L. Rizzo et al., “Brownian distance covariance,” The
annals of applied statistics, vol. 3, no. 4, pp. 1236–1265, 2009.
[51] D. N. Reshef, Y. A. Reshef, H. K. Finucane, S. R. Grossman, G. McVean,
P. J. Turnbaugh, E. S. Lander, M. Mitzenmacher, and P. C. Sabeti,
“Detecting novel associations in large data sets,” Science, vol. 334, no.
6062, pp. 1518–1524, 2011.
[52] S. Hooper, E. Biegert, M. Levy, J. Pensock, L. V. D. Spoel, X. Zhang,
T. Zhang, N. Tandon, and B. Aazhang, “On developing an FPGA based
system for real time seizure prediction,” in Asilomar Conference on
Signals, Systems and Computers, 2017.
X. A PPENDIX
a) Proof of (6): We have from (4),
R∞
R∞
y(t) =
h1 (t − τ )x(τ )dτ+ h2 (t − τ )w(τ )dτ (20)
−∞
⇒
R∞
−∞
ej2πνt dYe (ν) =
−∞
+
R∞
h2 (t − τ )
−∞
=
R∞
ej2πνt
−∞
R∞
−∞
=
e
−∞
R∞ j2πντ
e
R∞
e (ν) dτ
ej2πντ dX
−∞
f (ν) dτ (from Theorem 1)
dW
e (ν)+
h1 (t − τ )e−j2πν(t−τ ) dτ dX
−∞
R∞ j2πνt R∞
j2πνt
h1 (t − τ )
−∞
e
R∞
R∞
f (ν) (21)
h2 (t − τ )e−j2πν(t−τ ) dτ dW
−∞
n
o
e (ν)+H2 (ν) dW
f (ν) .
H1 (ν) dX
(22)
−∞
e (ν) + H2 (ν) dW
f (ν) .
=⇒ dYe (ν) = H1 (ν) dX
b) Proof of Theorem 3: We will first prove that
MIXY (ν1 , ν2 ) is zero, when X and Y are related by (4) and
ν1 6= ν2 . Since the processes X (t) and W (t) are independent,
their spectral processes are also independent. In addition, we
also know from Theorem 2 that the spectral increments of
the Gaussian process X (t) are independent.
It is clear from
(6) that given H1 (ν) and H2 (ν), dYeR (ν2 ) , dYeI (ν2 ) is
15
completely
determined by the two-dimensional
random
veceR (ν2 ) , dX
eI (ν2 ) and dW
fR (ν2 ) , dW
fI (ν2 ) , both
tors dX
of which
are independent of the two-dimensional random
eR (ν1 ) , dX
eI (ν1 ) when ν1 6= ν2 . This implies the
vector dX
e
eR (ν1 ) , dX
eI (ν1 ) ,
MI between dYR (ν2 ) , dYeI (ν2 ) and dX
which is defined as MIXY (ν1 , ν2 ), is zero.
We will now derive the analytical expression for MIXY (ν,ν),
for ν 6= 0. Let H1 (ν)=H1R (ν)+jH1I (ν) and H2 (ν)=H2R (ν)+
jH2I (ν). We can see from (5), (6) that
dYeR (ν) , dYeI (ν) ∼N 0, 12 sX (ν) |H1 (ν) |2 +
1
2
I ,
(23)
2 sW (ν) |H2 (ν) |
where N represents Gaussian distribution, 0 is a two element
zero vector and I is the 2 × 2 identity matrix. In addition,
eR (ν),dX
eI (ν),dYeR (ν),dYeI (ν) ∼N 0, Σ11 Σ12 , (24)
dX
Σ21 Σ22
where Σ11 = 12 sX (ν) I, Σ22 = 12 σY2e (ν) I, σY2e (ν) =
sX (ν) |H1 (ν) |2 + sW (ν) |H2 (ν) |2 , I is the 2 × 2 identity
matrix and 0 is a four element zero vector. In addition,
1
1
sX (ν) H1R (ν)
sX (ν) H1I (ν)
T
2
2
Σ12 = Σ21 =
.
− 12 sX (ν) H1I (ν) 21 sX (ν) H1R (ν)
Now, the MI between X and Y at frequency ν is given by
eR (ν) , dX
eI (ν) ; dYeR (ν) , dYeI (ν)
MIXY (ν, ν) = I dX
eR (ν) , dX
eI (ν) ; dYeR (ν) +
= I dX
eR (ν) , dX
eI (ν) ; dYeI (ν) |dYeR (ν)
I dX
(25)
=I
eR (ν) ,dX
eI (ν) ; dYeR (ν) +
dX
eR (ν) , dX
eI (ν) ; dYeI (ν) ,
I dX
(26)
where (25) follows from the chain rule of mutual information
[34] and (26) follows because the real and imaginary parts
of the spectral process of a Gaussian
process are independent
eR (ν) , dX
eI (ν) , dYeR (ν)
from Theorem 2. In addition, dX
is a Gaussian distributed random vector with zero mean and
covariance matrix Σ0 , which is easily obtained from (24). Since
the mutual information between components of a Gaussian
random vector depends only on the determinants of the joint
distribution’s covariance matrices and that of marginals [34],
we can easily show that
1 2
eR (ν) , dX
eI (ν) ; dYeR (ν) = 1 log |Σ11 |( 02 σYe )
I dX
2
|Σ |
|H1 (ν)|2 sX (ν)
= 12 log 1 + |H
. (27)
2
2 (ν)| sW (ν)
Similarly, we can also show that
eR (ν) , dX
eI (ν) ;dYeI (ν) =
I dX
1
2
log 1 +
|H1 (ν)|2 sX (ν)
|H2 (ν)|2 sW (ν) .
(28)
From (26), (27) and (28), we have
eR (ν) , dX
eI (ν) ; dYeR (ν)
MIXY (ν, ν) = 2 × I dX
|H1 (ν)|2 sX (ν)
.
(29)
= log 1 + |H
2
2 (ν)| sW (ν)
At ν = 0, MI-in-frequency between X and Y is equal to
eR (ν) , dX
eI (ν) ; dYeR (ν) , since the imaginary part of
I dX
Y is zero.
c) Relationship between MI in frequency and coherence:
The coherence CXY (ν) ∈ [0, 1] between two processes X and
Y related by (4) is given by
CXY (ν) =
|sXY (ν)|2
sX (ν)sY (ν)
=
|H1 (ν)|2 sX (ν)
sX (ν)|H1 (ν)|2 +sW (ν)|H2 (ν)|2 .
⇒ − log (1 − CXY (ν)) = log 1 +
|H1 (ν)|2 sX (ν)
|H2 (ν)|2 sW (ν)
= MIXY (ν, ν) .
(30)
d) Proof of Theorem 4: Now we consider two discretetime Gaussian stochastic processes X [n] and Y [n] that are
related by
y[n] = h1 [n] ∗ x[n] + h2 [n] ∗ w[n],
(31)
where h1 [n] and h2 [n] are the impulse responses of two
discrete-time linear, time-invariant filters. (31) is the discretetime equivalent of (4). It was shown in chapter 10 in [33]
that mutual information between the discrete-time Gaussian
stochastic processes X [n] and Y [n] is related to coherence
according to
I (X; Y ) = −
0.5
R
log (1 − CXY (λ)) dλ.
(32)
0
From (30) and (32), we have
I (X; Y ) =
0.5
R
0
MIXY (λ, λ) dλ.
(33)
| 7cs.IT
|
Three Colors Suffice: Conflict-Free Coloring of Planar Graphs
Zachary Abel∗ Victor Alvarez† Erik D. Demaine‡ Sándor P. Fekete†
arXiv:1701.05999v1 [cs.DM] 21 Jan 2017
Aman Gour§ Adam Hesterberg∗
Phillip Keldenich†
Abstract
A conflict-free k-coloring of a graph assigns one of k
different colors to some of the vertices such that, for
every vertex v, there is a color that is assigned to exactly
one vertex among v and v’s neighbors. Such colorings
have applications in wireless networking, robotics, and
geometry, and are well-studied in graph theory. Here we
study the natural problem of the conflict-free chromatic
number χCF (G) (the smallest k for which conflict-free
k-colorings exist), with a focus on planar graphs.
For general graphs, we prove the conflict-free variant of the famous Hadwiger Conjecture: If G does not
contain Kk+1 as a minor, then χCF (G) ≤ k. For planar
graphs, we obtain a tight worst-case bound: three colors
are sometimes necessary and always sufficient. In addition, we give a complete characterization of the algorithmic/computational complexity of conflict-free coloring.
It is NP-complete to decide whether a planar graph has
a conflict-free coloring with one color, while for outerplanar graphs, this can be decided in polynomial time.
Furthermore, it is NP-complete to decide whether a planar graph has a conflict-free coloring with two colors,
while for outerplanar graphs, two colors always suffice.
For the bicriteria problem of minimizing the number of
colored vertices subject to a given bound k on the number of colors, we give a full algorithmic characterization
in terms of complexity and approximation for outerplanar and planar graphs.
1
Introduction
Coloring the vertices of a graph is one of the fundamental problems in graph theory, both scientifically and historically. Proving that four colors always suffice to color
∗ Mathematics
Department,
MIT,
Cambridge,
Massachusetts, USA
† Algorithms
Group,
TU
Braunschweig,
Braunschweig, Germany
‡ Computer Science and Artificial Intelligence Laboratory,
MIT, Massachusetts, USA
§ Computer Science and Engineering, IIT Bombay, Mumbai, India
Christian Scheffer†
a planar graph [5, 6, 26] was a tantalizing open problem
for more than 100 years; the quest for solving this challenge contributed to the development of graph theory,
but also to computers in theorem proving [29]. A generalization that is still unsolved is the Hadwiger Conjecture [18]: A graph is k-colorable if it has no Kk+1 minor.
Over the years, there have been many variations
on coloring, often motivated by particular applications.
One such context is wireless communication, where
“colors” correspond to different frequencies. This also
plays a role in robot navigation, where different beacons
are used for providing direction. To this end, it is vital
that in any given location, a robot is adjacent to a
beacon with a frequency that is unique among the ones
that can be received. This notion has been introduced
as conflict-free coloring, formalized as follows. For any
vertex v ∈ V of a simple graph G = (V, E), the closed
neighborhood N [v] consists of all vertices adjacent to v
and v itself. A conflict-free k-coloring of G assigns one
of k different colors to a (possibly proper) subset S ⊆ V
of vertices, such that for every vertex v ∈ V , there
is a vertex y ∈ N [v], called the conflict-free neighbor
of v, such that the color of y is unique in the closed
neighborhood of v. The conflict-free chromatic number
χCF (G) of G is the smallest k for which a conflict-free
coloring exists. Observe that χCF (G) is bounded from
above by the proper chromatic number χ(G) because
in a proper coloring, every vertex is its own conflictfree neighbor.
Conflict-free coloring has received an increasing
amount of attention. Because of the relationship to
classic coloring, it is natural to investigate the conflictfree coloring of planar graphs. In addition, previous
work has considered either general graphs and hypergraphs (e.g., see [25]) or geometric scenarios (e.g., see
[20]); we give a more detailed overview further down.
This adds to the relevance of conflict-free coloring of
planar graphs, which constitute the intersection of general graphs and geometry. In addition, the subclass of
outerplanar graphs is of interest, as it corresponds to
subdividing simple polygons by chords.
There is a spectrum of different scientific challenges
when studying conflict-free coloring. What are worstcase bounds on the necessary number of colors? When
is it NP-hard to determine the existence of a conflictfree k-coloring, when polynomially solvable? What
can be said about approximation? Are there sufficient
conditions for more general graphs? And what can be
said about the bicriteria problem, in which also the
number of colored vertices is considered? We provide
extensive answers for all of these aspects, basically
providing a complete characterization for planar and
outerplanar graphs.
1.1 Our contribution. We present the following
results.
1. For general graphs, we provide the conflict-free
variant of the Hadwiger Conjecture: If G does not
contain Kk+1 as a minor, then χCF (G) ≤ k.
2. It is NP-complete to decide whether a planar graph
has a conflict-free coloring with one color. For
outerplanar graphs, this question can be decided
in polynomial time.
3. It is NP-complete to decide whether a planar graph
has a conflict-free coloring with two colors. For
outerplanar graphs, two colors always suffice.
4. Three colors are sometimes necessary and always
sufficient for conflict-free coloring of a planar graph.
5. For the bicriteria problem of minimizing the number of colored vertices subject to a given bound
χCF (G) ≤ k with k ∈ {1, 2}, we prove that the
problem is NP-hard for planar and polynomially
solvable in outerplanar graphs.
6. For planar graphs and k = 3 colors, minimizing
the number of colored vertices does not have a
constant-factor approximation, unless P = NP.
7. For planar graphs and k ≥ 4 colors, it is NPcomplete to minimize the number of colored vertices. The problem is fixed-parameter tractable
(FPT) and allows a PTAS.
1.2 Related work. In a geometric context, the
study of conflict-free coloring was started by Even,
Lotker, Ron, and Smorodinsky [15] and Smorodinsky [27], who motivate the problem by frequency assignment in cellular networks: There, a set of n base
stations is given, each covering some geometric region
in the plane. The base stations service mobile clients
that can be at any point in the total covered area. To
avoid interference, there must be at least one base station in range using a unique frequency for every point
in the entire covered area. The task is to assign a frequency to each base station minimizing the number of
frequencies. On an abstract level, this induces a coloring problem on a hypergraph where the base stations
correspond to the vertices and there is an hyperedge
between some vertices if the range of the corresponding
base stations has a non-empty common intersection.
If the hypergraph is induced by disks, Even et
al. [15] prove that O(log n) colors are always sufficient.
Alon and Smorodinsky [4] extend this by showing that
each family of disks, where each disk intersects at most
k others, can be colored using O(log3 k) colors. Furthermore, for unit disks, Lev-Tov and Peleg [23] present an
O(1)-approximation algorithm for the number of colors.
Horev et al. [21] extend this by showing that any set of
n disks can be colored with O(k log n) colors, even if
every point must see k distinct unique colors. Abam et
al. [1] discuss the problem in the context of cellular networks where the network has to be reliable even if some
number of base stations fault, giving worst-case bounds
for the number of colors required.
For the dual problem of coloring a set of points
such that each region from some family of regions contains at least one uniquely colored point, Har-Peled and
Smorodinsky [19] prove that with respect to every family of pseudo-disks, every set of points can be colored
using O(log n) colors. For rectangle ranges, Elbassioni
and Mustafa [14] show that it is possible to add a sublinear number of points such that a conflict-free coloring with O(n3/8·(1+ε) ) colors becomes possible. Ajwani
et al. [2] complement this by showing that coloring a
set of points with respect to rectangle ranges is always
possible using O(n0.382 ) colors. For coloring points on
a line with respect to intervals, Cheilaris et al. [10]
present a 2-approximation algorithm, and a 5 − k2 approximation algorithm when every interval must see
k uniquely colored vertices. Hoffman et al. [20] give
tight bounds for the conflict-free chromatic art gallery
problem under rectangular visibility in orthogonal polygons: Θ(log log n) are sometimes necessary and always
sufficient. Chen et al. [13] consider the online version
of the conflict-free coloring of a set of points on the
line, where each newly inserted point must be assigned
a color upon insertion, and at all times the coloring has
to be conflict-free. Also in the online scenario, Bar-Nov
et al. [9] consider a certain class of k-degenerate hypergraphs which sometimes arise as intersection graphs of
geometric objects, presenting an online algorithm using
O(k log n) colors.
On the combinatorial side, some authors consider
the variant in which all vertices need to be colored; note
thus χCF (G) ≤ χP (G). For some k, we define the
k
conflict-free domination number γCF
(G) of G to be the
minimum number of vertices that have to be colored in a
k
conflict-free k-coloring of G. We set γCF
(G) = ∞ if G is
not conflict-free k-colorable. Because the set of colored
vertices is a dominating set, the conflict-free domination
k
number satisfies γCF
(G) ≥ γ(G) for all k, where γ(G),
the domination number of G, is the size of a minimum
dominating set of G. Moreover, for any graph, there is
k
a k ≤ γ(G) such that γCF
(G) = γ(G).
We denote the complete graph on n vertices
by Kn := ({1, . . . , n}, {{u, v} | u, v ∈ {1, . . . , n}, u 6= v}),
and the complete bipartite graph on n and m vertices
as Kn,m . We define the graph Kn−3 := (V (Kn ), E(Kn )\
E(K3 )), which is obtained by removing any three edges
forming a single triangle from a Kn .
We also provide a number of results for outerplanar graphs. An outerplanar graph is a graph that has
a planar embedding for which all vertices belong to the
outer face of the embedding. An outerplanar graph is
called maximal iff no edges can be added to the graph
without losing outerplanarity [8]. Maximal outerplanar
graphs can also be characterized as the graphs having
an embedding corresponding to a polygon triangulation,
which illustrated the particular relevance in a geometric context. In addition, maximal outerplanar graphs
exhibit a number of interesting graph-theoretic properties. Every maximal outerplanar graph is chordal, a
2-tree and a series-parallel graph. Also, every maximal outerplanar graph is the visibility graph of a simple
polygon.
For some of our NP-hardness proofs, we use a
variant of the planar 3-SAT problem, called Positive
2 Preliminaries
Planar 1-in-3-SAT. This problem was introduced and
For every vertex v ∈ V , the closed neighborhood is shown to be NP-complete by Mulzer and Rote [24], and
denoted by NG [v] := NG (v) ∪ {v}. A partial k-coloring consists of deciding whether a given positive planar 3of G is an assignment χ : V 0 → {1, . . . , k} of colors to a CNF formula allows a truth assignment such that in
subset V 0 ⊆ V (G) of the vertices. χ is called conflict- each clause, exactly one literal is true.
free k-coloring of G iff, for each vertex v ∈ V , there is
a vertex w ∈ NG [v] ∩ V 0 such that χ(w) is unique in Definition 2.1. (Positive planar formulas)
NG [v], i.e., for all other w0 ∈ NG [v] ∩ V 0 , χ(w0 ) 6= χ(w). A formula φ in 3-CNF is called positive planar iff it
We call w the conflict-free neighbor of v.
is both positive and backbone planar. A formula φ
In order to avoid confusion with proper k-colorings, is called positive iff it does not contain any negation,
i.e., colorings that color all vertices such that no adja- i.e. iff all occurring literals are positive. A formula
cent vertices receive the same color, we use the term φ, with clause set C = {c1 , . . . , cl } and variable set
proper coloring when referring to this kind of coloring. X = {x1 , . . . , xn }, is called backbone planar iff its
The minimum number of colors needed for a proper col- associated graph G(φ) := (X ∪C, E(φ)) is planar, where
oring of G, also known as the chromatic number of G, E(φ) is defined as follows:
is denoted by χP (G), whereas the minimum number
• xi cj ∈ E(φ) for a clause cj ∈ C and a variable
of colors required for a conflict-free coloring of G (G’s
xi ∈ X iff xi occurs in cj ,
conflict-free chromatic number ) is written as χCF (G).
Note that, because every vertex sees itself, every proper
coloring of G is also a conflict-free coloring of G, and
• xi xi+1 ∈ E(φ) for all 1 ≤ i < n.
that this does not change asymptotic results for general
graphs and hypergraphs: it suffices to introduce one additional color for vertices that are left uncolored in our
constructions. Regarding general hypergraphs, Ashok
et al. [7] prove that maximizing the number of conflictfreely colored edges in a hypergraph is FPT when parameterized by the number of conflict-free edges in the
solution. Cheilaris et al. [11] consider the case of hypergraphs induced by a set of planar Jordan regions and
prove an asymptotically tight upper bound of O(log n)
for the conflict-free list chromatic number of such hypergraphs. They also consider hypergraphs induced by
the simple paths√of a planar graph and prove an upper bound of O( n) for the conflict-free list chromatic
number. For hypergraphs induced by the paths of a
simple graph G, Cheilaris and Tóth [12] prove that it
is coNP-complete to decide whether a given coloring is
conflict-free if the input is G. Regarding the case in
which the hypergraph is induced by the neighborhoods
of a simple graph G, which resembles our scenario, Pach
and Tárdos [25] prove that the conflict-free chromatic
number of an n-vertex graph is in O(log2 n). Glebov
et al. [17] extend this from an extremal and probabilistic point of view by proving that almost all G(n, p)graphs have conflict-free chromatic number O(log n) for
p ∈ ω(1/n), and by giving a randomized construction for
graphs having conflict-free chromatic number Θ(log2 n).
In more recent work, Gargano and Rescigno [16] show
that finding the conflict-free chromatic number for general graphs is NP-complete, and prove that the problem is FPT w.r.t. vertex cover or neighborhood diversity number.
The path formed by the latter edges is also called the 1 on the vertices of J can be removed. Therefore, we
backbone of the formula graph G(φ).
can assume each of the two copies of J to be colored
in a conflict-free manner using the colors {2, . . . , k}.
3 Conflict-Free Coloring of General Graphs
Observe that, due to χCF (J) = k − 1, each of these
In this section we consider the Conflict-Free k- colors must be used at least once in each copy. This
Coloring problem on general simple graphs. In implies that both u and v see each color at least twice:
Section 3.1, we prove that this problem is NP-complete The two copies of J enforce two occurrences of the colors
for any k ≥ 1. In Section 3.2, we provide a sufficient {2, . . . , k}, and color 1 is assigned to both u and v, which
criterion that guarantees conflict-free k-colorability. In are connected by an edge. This is a contradiction, and
Section 3.3, we consider the conflict-free domination therefore, both u and v must be colored with distinct
number and prove that, for any k ≥ 3, there is no colors.
k
constant-factor approximation algorithm for γCF .
Next, we give an inductive construction of graphs, Gk ,
with χCF (Gk ) = k. The proof of NP-hardness relies on
3.1 Complexity
this hierarchy.
Theorem 3.1. Conflict-Free k-Coloring is NP1. The first graph G1 of the hierarchy consists of a
complete for any fixed k ≥ 1.
single isolated vertex. G2 is a K1,3 with one edge
subdivided by another vertex, or, equivalently, a
Membership in NP is clear. For k ≥ 3, we prove NPpath of length 3 with a leaf vertex attached to one
hardness using a reduction from proper k-Coloring.
of the inner vertices.
For k ∈ {1, 2}, refer to Section 4, where we prove
Conflict-Free k-Coloring of planar graphs to be
2. Given Gk and Gk−1 , Gk+1 is constructed as follows
NP-complete for k ∈ {1, 2}.
for k ≥ 2:
Central to the proof is the following lemma that
enables us to enforce certain vertices to be colored, and
• Take a complete graph G = Kk+1 on
both ends of an edge to be colored using distinct colors.
k + 1 vertices.
• To each vertex v ∈ V (Kk+1 ), attach two
Lemma 3.2. Let G be any graph, u, v ∈ V (G) and
disjoint and independent copies of Gk , adding
vu = e ∈ E(G). If N (v) contains two disjoint and
an edge from v to every vertex of both copies
independent copies of a graph H with χCF (H) = k, not
of Gk .
adjacent to any other vertex w ∈ G, every conflict-free
• For each edge e = vw ∈ E(Kk+1 ), add
k-coloring of G colors v. If the same holds for u and
two disjoint and independent copies of Gk−1 ,
in addition, NG (u) ∩ NG (v) contains two disjoint and
adding an edge from v and w to every vertex
independent copies of a graph J with χCF (J) = k − 1,
of both copies.
not adjacent to any other vertex w ∈ G, every conflictfree k-coloring of G colors u and v with different colors. The number of vertices of the graphs Gk obtained by
Proof. Assume towards a contradiction that there was a the above construction satisfies the recursive formula
conflict-free k-coloring χ that avoids coloring v. Then,
due to the copies of H being independent, disjoint
and not connected to any other vertex, the restriction
of χ to the vertices of each of the two copies must
induce a conflict-free coloring on H. As χCF (H) =
k, this implies that χ uses k colors on each copy.
Therefore, in the open neighborhood of v, there are at
least two vertices colored with each color. This leads
to a contradiction, because v cannot have a conflictfree neighbor.
For the second proposition, suppose there was a
conflict-free coloring assigning the same color to u and v.
Without loss of generality, let this color be 1. As every
vertex of the two copies of J now sees two occurrences
of color 1, color 1 can not be the color of the unique
neighbor of any vertex of J, and any occurrence of color
|G1 | = 1, |G2 | = 5, |Gk+1 | = (k+1)·(2|Gk |+k|Gk−1 |+1),
which is in Ω 2k and O 2k log k . Figure 1 depicts the
graph G3 , which in addition to being planar is a seriesparallel graph.
Lemma 3.3. For Gk constructed in this manner,
χCF (Gk ) = k.
Proof. The proof uses induction over k. Application
of Lemma 3.2 implies that all vertices of the Kk+1
underlying Gk+1 have to be colored using different
colors. Therefore, χCF (Gk+1 ) ≥ k + 1. By coloring all
k + 1 vertices of the underlying Kk+1 with a different
color, we obtain a conflict-free (k + 1)-coloring of Gk+1 ,
implying χCF (Gk+1 ) ≤ k + 1.
G2
G2
G1
G1
Figure 1: The graph G3 .
Lemma 3.4. For k ≥ 2, k-Coloring 4 ConflictFree k-Coloring. Therefore, for k ≥ 3, ConflictFree k-Coloring is NP-complete.
Proof. Given a graph G for which to decide proper kcolorability for a fixed k. We construct a graph G0 that
is conflict-free k-colorable iff G is k-colorable. G0 is
constructed from G by attaching two copies of Gk to
each vertex v ∈ V (G), by adding an edge from v to each
vertex of the copies of Gk . For each edge uv ∈ E(G),
we attach two copies of Gk−1 to both endpoints of uv
by adding an edge from u and v to all vertices of both
copies. As k is fixed, |Gk | and |Gk−1 | are constant,
implying that G0 can be constructed in polynomial time.
A proper k-coloring of G induces a conflict-free kcoloring of G0 by leaving all other vertices uncolored. On
the other hand, by Lemma 3.2, a conflict-free k-coloring
χ of G0 colors all vertices v ∈ V (G) and for every edge,
the colors of both endpoints are distinct. Therefore, the
restriction of χ to V (G) is a proper k-coloring of G.
3.2 A Sufficient Criterion for k-Colorability. In
this section we present a sufficient criterion for conflictfree k-colorability together with an efficient heuristic
that can be used to color graphs satisfying this criterion
with k colors in a conflict-free manner. This heuristic
is called iterated elimination of distance-3-sets and is
detailed in Algorithm 1. The main idea of this heuristic
is to iteratively compute maximal sets of vertices at
pairwise (link) distance at least 3, coloring all vertices
in one of these sets using one color, and then removing
these vertices and their neighbors until all that remains
is a collection of disconnected paths, which can then be
colored using one color.
conflict-free k-coloring that can be found in polynomial
time using iterated elimination of distance-3 sets.
Proof. For k = 1, a graph G with neither a K3 nor a
K4−3 = K1,3 minor consists of a collection of isolated
paths. A path on 3n vertices can be colored with
one color by coloring the middle vertex of every three
vertices. This does not color the vertices at either end,
so up to two vertices can be removed from the path to
get colorings for paths on 3n − 1 and 3n − 2 vertices.
For k ≥ 2, we use induction as follows: First,
we color an inclusion-wise maximal subset D ⊆ V of
vertices at pairwise distance at least 3 to each other
using color 1. This provides a conflict-free neighbor
of color 1 to every vertex in N [D]. Therefore, the
vertices in N [D] are covered and can be removed from
the graph. The remaining graph consists of vertices at
distance 2 to some vertex in D; we call these vertices
unseen in the remainder of the proof. We show that
−3
the remaining graph has no Kk+1 and no Kk+2
as a
minor. By induction, iterated elimination of distance 3
sets requires k − 1 colors to color the remaining graph,
and thus k colors suffice for G.
If the graph is disconnected, iterated elimination of
distance 3 sets works on all components separately, so
we can assume G to be connected. We claim that there
is no set U of unseen vertices that is a cutset of G.
Suppose there were such a cutset U and let H be any
component of G \ U not containing v, the first selected
vertex during the construction of D. At least one vertex
of H is colored: every vertex in U is at distance at
least two from every colored vertex not in H, therefore,
every vertex in H is at distance at least three from every
colored vertex not in H. Consider the iteration where
the first vertex w of H is added to the set of colored
vertices D. At this point, w is at distance exactly 3 from
some colored vertex not in H. However, this implies w
is adjacent to some vertex from U , contradicting the
fact that all vertices in U are unseen.
Now, suppose for the sake of contradiction that the
−3
set W of unseen vertices contains a Kk+1 or Kk+2
minor.
W is not the whole graph, because at least one vertex is
colored, so there must be a vertex v not in the Kk+1 or
−3
Kk+2
minor. For every vertex w ∈ W , there is a path
from v to w that intersects W only at w. Otherwise,
W \ {w} would be a cutset separating v from w. So, if
−3
the graph induced by W had a Kk+1 or Kk+2
minor, we
could contract G \ W to a single vertex, which would be
−3
adjacent to all vertices in W , yielding a Kk+2 or Kk+3
minor of G, which does not exist.
−3
Observe that Gk+1 contains a Kk+3
as a minor, but not
Theorem 3.5. Let G be a graph and k ≥ 1. If G a Kk+2 , proving that just excluding Kk+2 as a minor
−3
has neither Kk+2 nor Kk+3
as a minor, G admits a does not suffice to guarantee k-colorability. Moreover,
Algorithm 1 Iterated elimination of distance-3-sets
1: i ← 1, χ ← ∅
2: Remove all isolated paths from G
3: while G is not empty do
4:
D←∅
5:
For each component of G, select some vertex v and add it to D
6:
while there is a vertex w at distance ≥ 3 from all vertices in D do
7:
Choose w at distance exactly 3 from some vertex in D
8:
D ← D ∪ {w}
9:
∀u ∈ D : χ(u) ← i
10:
i←i+1
11:
Remove N [D] from G
12:
Remove all isolated paths from G
13: Color all removed isolated paths using color i
−3
note that Kk+1 is a minor of Kk+2 and Kk+3
. This G, or adjacent to just one vertex of G. Conversely, let χ
yields the following corollary, which is the conflict-free be a conflict-free coloring of G0 , coloring just n vertices.
If χ did not assign a color to some vertex v of G, it would
variant of the Hadwiger Conjecture.
have to color all M ≥ n + 1 neighbors of v. If χ assigned
Corollary 3.6. All graphs that do not have Kk+1 as the same color to any pair v, w of vertices adjacent in G,
a minor are conflict-free k-colorable.
it would have to color all M vertices adjacent only to v
and w. Therefore, χ is a proper coloring of G. Running
a c-approximation algorithm on G0 results in a conflict3.3 Conflict-Free Domination Number. In this
free coloring using at most c · n < M vertices if G is
section we consider the problem of minimizing the
k-colorable, and using at least M vertices if G is not;
number of colored vertices in a conflict-free k-coloring
thus we could decide k-colorability in polynomial time.
k
for a fixed k, which is equivalent to computing γCF . We
call the corresponding decision problem k-ConflictFree Dominating Set. We show that approximating
4 Planar Conflict-Free Coloring
the conflict-free domination number in general graphs
is hard for any fixed k. In Section 5 we discuss the k- This section deals with the Planar Conflict-Free kConflict-Free Dominating Set problem for planar Coloring problem which consists of deciding conflictfree k-colorability for fixed k on planar graphs. Due to
graphs.
the 4-color theorem, we immediately know that every
Theorem 3.7. Unless P = NP, for any k ≥ 3, planar graph is conflict-free 4-colorable. This naturally
there is no polynomial-time approximation algorithm for leads to the question of whether there are planar graphs
k
(G) with constant approximation factor.
γCF
requiring 4 colors or whether fewer colors might already
suffice for a conflict-free coloring, which we address in
Proof. We use a reduction from proper k-Coloring the following two sections.
for the proof. Assume towards a contradiction that
there was a polynomial-time approximation algorithm 4.1 Complexity. For k ∈ {1, 2} colors, we show that
k
for γCF
(G) with approximation factor c ≥ 1. Let the problem of deciding conflict-free k-colorability on
G be a graph on n vertices for which we want to planar graphs is NP-complete. This implies that 2 colors
decide k-colorability. For each vertex v of G, add are not sufficient.
M := (n + 1)(c + 1) vertices uv to G and connect them
to v. For each edge vw of G, add M vertices uvw to Theorem 4.1. Deciding planar conflict-free 1-colorG and connect them to both v and w. Let G0 be the ability is NP-complete.
resulting graph. Clearly, the size of G0 is polynomial
in the size of G. Additionally, G0 is planar if G is, Proof. Membership in NP is obvious. The proof of
and G0 has a conflict-free k-coloring of size n iff G is NP-hardness is done by reduction from the problem
properly k-colorable: Any proper k-coloring of G is a Positive Planar 1-in-3-SAT. From a positive planar
conflict-free k-coloring of G0 , as every vertex added to 3-CNF formula φ with clauses C = {c1 , . . . , cl } and
G is either adjacent to two distinctly colored vertices of variables X = {x1 , . . . , xn } we construct in polynomial
time a graph G1 (φ) such that φ is 1-in-3-satisfiable iff
G1 (φ) admits a conflict-free 1-coloring.
First, find and fix a planar embedding d of G(φ).
G1 (φ) is constructed from G(φ) and d as follows: For
every variable xi , there is a cycle Zi = (zi,1 , . . . , zi,12 ) of
length 12. The vertices zi,1 , zi,4 , zi,7 , zi,10 are referred to
as true vertices of Zi , all other vertices are false vertices.
Moreover, vertices zi,1 , zi,2 , zi,3 are called upper vertices
of Zi , and vertices zi,7 , zi,8 , zi,9 are called lower vertices
of Zi . Additionally, vertices zi,4 , zi,5 , zi,6 are called right
vertices of Zi and zi,10 , zi,11 , zi,12 are called left vertices
of Zi .
For each clause cj , there is a cycle (cj,1 , . . . , cj,4 )
of length 4 in G1 (φ).
To each variable xi for
i ∈ {2, . . . , n − 1}, we associate two disjoint se|Ui |
|Li |
quences Ui = uj j=1 and Li = lj j=1 of clauses
xi appears in. The sequences are constructed using a clockwise (with respect to d) enumeration
of the edges of xi in G(φ), starting with xi−1 xi .
Let (xi−1 xi , xi cj1 , . . . , xi cjλ , xi xi+1 , xi cjλ+1 , . . . , xi cjµ )
be the sequence of edges encountered in this manner
and set Ui := (cj1 , . . . , cjλ ) and Li := (cjλ+1 , . . . , cjµ ).
For i ∈ {1, n}, Li is empty and Ui contains all clauses
xi appears in, again in clockwise order. In G1 (φ), the
clauses and variables are connected such that for each
clause cj that xi occurs in, either the upper or the lower
true vertex of xi is adjacent to cj,1 . More precisely, for
variable xi , if cj = um , we add the edge cj,1 zi,1 to connect the upper true vertex to the clause. If cj = lm ,
we add cj,1 zi,7 to connect the lower true vertex to the
clause. Because the order of edges around each vertex
is preserved by the construction, the graph G1 (φ) obtained in this way can be embedded in the plane by a
suitable adaptation of d. See Figure 2 for an example
of the construction.
Now we prove that G1 (φ) is conflict-free 1-colorable
iff φ is 1-in-3-satisfiable. Regarding necessity, a valid
truth assignment b : X → B yields a valid conflictfree coloring by coloring the vertex cj,3 of every clause,
coloring all true vertices of variables with b(xi ) = 1 and
coloring the false vertices zi,3 , zi,6 , zi,9 , zi,12 of all other
variables. Thus, in every cycle Zi , every third vertex
is colored, providing a conflict-free neighbor to every
vertex of Zi . Moreover, in each clause, by virtue of cj,3
being colored, vertices cj,2 , cj,3 , cj,4 have a conflict-free
neighbor. Because b is a valid truth assignment, for each
clause, the vertex cj,1 is adjacent to exactly one colored
true vertex. Therefore, the coloring constructed in this
way is conflict-free.
Regarding sufficiency, we first argue that the vertices cj,1 , cj,2 , cj,4 can never be colored: If cj,1 receives
a color, then cj,3 still enforces that one of cj,2 , cj,3 , cj,4
is colored, leading to a contradiction in either case. If
c1 c1,3
c2
c1,1
z1,1
z1,3
x1
c3
x2
x3
x4
x5
c4
Figure 2: A formula graph G(φ) (dashed) and the
corresponding G1 (φ) (solid).
cj,2 receives a color, then cj,4 cannot have a conflict-free
neighbor and vice versa. Therefore, no clause vertex
can be the conflict-free neighbor of any vertex of Zi .
Thus, the conflict-free neighbor of every vertex of Zi
must itself be a vertex of Zi . Moreover, the conflictfree neighbor of every vertex cj,1 must be a true vertex.
Thus, there are exactly three ways to color each cycle
Zi : either by coloring the true vertices (one possibility),
or by coloring every other false vertex (two possibilities). A valid conflict-free 1-coloring of G1 (φ) satisfies
the property that for each clause cj , exactly one of the
true vertices adjacent to cj,1 is colored. Hence, a valid
conflict-free 1-coloring of G1 (φ) induces a valid truth
assignment b by setting b(xi ) = 1 iff all true vertices of
xi are colored.
Theorem 4.2. It is NP-complete to decide whether a
planar graph admits a conflict-free 2-coloring.
The proof requires the gadget G≤1 depicted in Figure 3.
G≤1 consists of three vertices v, w1 , w2 forming a triangle. Each edge ux of the triangle has two corresponding
1
2
vertices yux
, yux
, each connected to u and x. Furthermore, both w1 and w2 are attached to two copies of a
cycle on 4 vertices, where every vertex of both cycles is
adjacent to the corresponding wi . G≤1 can be used to
enforce that the vertices connected to its central vertex
v are colored using at most one distinct color:
Lemma 4.3. Let G = (V, E) be any graph, let v ∈ V
and let G0 be the graph resulting from adding a copy of
G≤1 to G by identifying v in G with v in G≤1 . Then
(1) G0 is planar if G is, and (2) every conflict-free 2coloring of G0 leaves v uncolored and uses at most one
color on NG [v].
Proof. The planarity of G0 follows from the planarity
of G by the observation that G≤1 is planar and can
be embedded in any face incident to v in a planar
embedding of G. Now consider a conflict-free 2-coloring
χ of G0 . χ must color both w1 and w2 . Otherwise,
χ restricted to each of the two 4-cycles adjacent to
wi must be a valid conflict-free 2-coloring. However,
as C4 requires at least 2 different colors, wi then sees
two occurrences of both colors, and thus cannot have a
conflict-free neighbor anymore. Furthermore, χ(w1 ) 6=
1
2
χ(w2 ), as otherwise, yw
and yw
must both be
1 w2
1 w2
colored with the other color; but then, w1 and w2 again
see two occurrences of both colors. By an analogous
argument, χ must not color v. Moreover, χ cannot
use more than one color on NG [v], because v already
sees one occurrence of each color, so adding another
occurrence of both colors would yield a conflict at v.
1
yw
1 w2
w1
w2
≤1
v
...
...
NG [v]
Figure 3: Gadget G≤1
cj,3
cj,1
cj
≤1
upper
t f
f
right
left
Zi−1
≤1
≤1
≤1
Zi+1
cycle and between the right and left vertices of two
adjacent variable cycles Zi and Zi+1 . This is depicted
in Figure 4. More precisely, for every cycle Zi , we add
one copy of gadget G≤1 , and connect its central vertex
v to all vertices of the cycle. In a planar embedding of
G2 (φ), these gadgets can be embedded within the face
defined by the cycles Zi and thus do not harm planarity.
By Lemma 4.3, this enforces that on every cycle, only
one color can be used. Moreover, for every edge xi xi+1
in G(φ), we add one copy of G≤1 that we connect to the
right vertices of xi and the left vertices of xi+1 . This
preserves planarity because these gadgets and the added
edges can be embedded in the face crossed by xi xi+1 in
some fixed embedding d of G(φ). As one of the right
vertices of xi and one of the left vertices of xi+1 must
be colored, this enforces that the same single color must
be used to color all cycles Zi . Finally, we add a copy of
G≤1 to every clause cj and connect it to cj,1 , . . . , cj,4 .
Again, this preserves planarity because the gadget may
be embedded in the face defined by (cj,1 , . . . , cj,4 ).
We now argue that G2 (φ) is conflict-free 2-colorable
iff G1 (φ) is conflict-free 1-colorable. A 1-coloring of
G1 (φ) induces a 2-coloring of G2 (φ) by copying the
color assignment and coloring the internal vertices of the
added gadgets as described in the proof of Lemma 4.3.
Now, let G2 (φ) be conflict-free 2-colorable and fix a
valid 2-coloring χ. In each clause, χ must color cj,3 and
neither of cj,1 , cj,2 nor cj,4 can be colored. Therefore,
no clause vertex can be the conflict-free neighbor of any
vertex of Zi . Thus, the conflict-free neighbor of every
vertex of Zi must itself be a vertex of Zi . Moreover, the
conflict-free neighbor of every vertex cj,1 must be a true
vertex. As there is only one color available to color all
cycle vertices of all variables, the restriction of χ to the
vertices of G1 (φ) yields a valid 1-coloring except for the
fact that some cj,3 might use a different color than the
one used for the variables. However, this can be fixed
by simply replacing all occurring colors with one single
color. Hence, G2 (φ) is conflict-free 2-colorable iff G1 (φ)
is conflict-free 1-colorable.
Zi
lower
Figure 4: Clause and variable gadget for k = 2
Proof. [Proof of Theorem 4.2] NP-hardness is proven by
constructing, in polynomial time, a planar graph G2 (φ)
from the graph G1 (φ) used in the hardness proof for
k = 1, such that G2 (φ) is conflict-free 2-colorable iff
G1 (φ) is conflict-free 1-colorable.
The construction is carried out by adding a gadget
G≤1 to every variable cycle Zi of G1 (φ), to every clause
4.2 Sufficient Number of Colors. As shown
above, it is NP-complete to decide whether a planar
graph has a conflict-free k-coloring for k ∈ {1, 2}. On
the positive side, we can establish the following result,
which follows from the more general results discussed in
Section 3.2.
Corollary 4.4. (of Theorem 3.5) Every outerplanar graph is conflict-free 2-colorable and every planar
graph is conflict-free 3-colorable. Moreover, such colorings can be computed in polynomial time.
Outerplanar graphs are not the only interesting
graph class for which one might suspect two colors to
be sufficient. Two other interesting subclasses of planar
graphs are series-parallel graphs and pseudo-maximal
planar graphs. However, each of these classes contains
graphs that do not admit a conflict-free 2-coloring: The
graph G3 as defined in Section 3 is an example of a
series-parallel graph requiring three colors. Figure 5
depicts a maximal outerplanar graph O9 satisfying
χCF (O9 ) = 2. This graph can be used to obtain a
pseudomaximal planar graph M with χCF (M ) = 3 by
adding two copies of O9 to the neighborhood of every
vertex of a triangle, similar to the construction of G3 ,
and adding gadgets on the inside of the triangle as
depicted in Figure 6.
5
Minimizing the Number of Colored Vertices
in Planar Graphs
In this section we consider the decision problem kConflict-Free Dominating Set for planar graphs.
In Section 5.1, we deal with the cases when k ∈ {1, 2} for
planar and outerplanar graphs, and we give a polynomial time algorithm to compute an optimal conflict-free
coloring of outerplanar graphs with k ∈ {1, 2} colors.
Section 5.2 discusses the problem for k ≥ 3.
5.1 At Most Two Colors. We start by pointing
out that, for every conflict-free 1-colorable graph G,
1
γCF
(G) = γ(G) holds. Moreover, Corollary 5.1 discusses the complexity of k-Conflict-Free Dominating Set and Theorem 5.2 states positive results for outerplanar graphs.
Corollary 5.1. (of Theorems 4.1 and 4.2)
k-Conflict-Free Dominating Set is NP-complete
for k ∈ {1, 2} for planar graphs.
Theorem 5.2. Let k ∈ {1, 2} and let G be an outerplanar graph. We can decide in polynomial time whether
χCF (G) ≤ k. Moreover, we can compute a conflict-free
k-coloring of G that minimizes the number of colored
vertices in O(n4k+1 ) time.
The proof of Theorem 5.2 relies on a polynomialtime algorithm that computes a k-coloring of the input
outerplanar graph G if and only if such a coloring
exists (which thus solves the decision problem). In the
following, we describe our algorithm.
Let G = (V, E) be an outerplanar graph. Let
χ : V 0 ⊆ V (G) → {0, 1, . . . , k} be a partial coloring
of the vertices of G and let v ∈ V . Observe that χ
defined like this differs from the definition given earlier
in the introduction. We call a pair Cv = [χ(v), Sv ] a
configuration of v, where χ(v) ∈ {0, 1, . . . , k} denotes
the color of v. If χ(v) = 0, we regard v as uncolored.
The set Sv ⊆ N [v] is the set of conflict-free neighbors
of v, along with their colors. That is, every w ∈ Sv is
a conflict-free neighbor of v under χ. For e = uv ∈ E
we call a pair Ce = [Cu , Cv ] a configuration of e. By Cew
we denote the configuration of an endpoint w ∈ {u, v}
of e. Observe that if χ was conflict-free, then Sv 6= ∅,
Figure 6: The pseudomaximal planar graph M , without and Cu and Cv do not conflict with each other. For the
the O9 gadgets.
latter property we say that Cu and Cv are compatible
and we denote this by Cu ↔ Cv . If Cev = Cev0 for a
pair e = uv, e0 = vw of incident edges, then we say
Furthermore, observe that Theorem 4.4 does not C 0 is compatible with C . The following observation
e
e
hold if every vertex must be colored. In this case, there is straightforward:
are outerplanar graphs requiring 3 colors for a conflictfree coloring. One can obtain an example of such a Observation 5.3. Let G be an outerplanar graph. Let
graph by adding a chord to a cycle of length 5.
C = {C1 , . . . , C|E| } be a set of configurations over the
Figure 5: The maximal outerplanar graph O9 .
edges of G using k colors. If for every pair e = uv,
e0 = vw of incident edges, Cu ↔ Cv and Cv ↔ Cw holds
and Ce0 is compatible with Ce , then a conflict-free kcoloring can be obtained from C.
Now let v ∈ V (G). Observe that the number
of different configurations Cv = [χ(v), Sv ] is upperbounded by O(nk ), as there cannot be more than
|N [v]|
· k! different sets Sv . Thus the following
k
observation is straightforward.
f2
f1
r
f1∗
Ge
f2∗
e
b
Observation 5.4. Let G = (V, E) be an outerplanar Figure 7: Graph construction of faces, windows, and
graph and let e = uv ∈ E. The number of different the corresponding dual (sub)graphs. The shaded are
configurations Ce = [Cu , Cv ] is upper-bounded by O(n2k ). corresponds to already processed faces of G (the past).
The face f1 is the face to be processed next (the
We can now describe our algorithm, which is based present). Edge b is the window between f1 and f2 . The
on non-serial dynamic programming. For the sake of rest of the graph corresponds to faces to be processed
simplicity, let us assume that the weak dual G∗ = in the future.
(V ∗ , E ∗ ) of the outerplanar graph G is connected. This
means that G∗ is a tree. It is well-know that, in general,
the weak dual graph of an outerplanar graph G is a consist of the edges he1 = (u1 , v1 ), . . . , e` = (u` , v` )i
forest [28]. We discuss later how to convert this forest where, without loss of generality, w = e1 if w is an
edge. Otherwise w = u1 if w is a vertex. Also, let
into a tree as long as G is connected.
Let us root G∗ at an arbitrary dual vertex r ∈ V ∗ . L(ei ) be the setS of all possible configurations of the
Thus, each dual vertex has a unique parent vertex on edge ei . By Cu1 we denote the number of conflictconfiguration Cu1 , i.e., if
the path from the vertex to r. For an edge e = vw ∈ E ∗ , free neighbors of u1 given the
S
where v is the parent of w, we consider the subtree Te Cu1 = (χ(u1 ), Su1 ), then Cu1 = |Su1 |. The algorithm
rooted at w. Let Ge be the primal subgraph of G whose populates a family {Pi } of sets containing pairs of
compatible configurations and their scores. In the
dual graph is Te .
We define a window b as the edge or vertex in the algorithm GenerateScore, δ(Cei , Cei−1 ) is the number
primal graph G separating two faces f1 , f2 . Observe of newly-colored vertices resulting from combining the
that b corresponds to an edge e in the dual graph G∗ . two compatible configurations Cei and Cei−1 .
If f1∗ and f2∗ are two (dual) vertices in the dual graph,
then the corresponding faces f1 and f2 only have b
in common, see Figure 7. Assume that f2 has been
conflict-free k-colored. Then, to color f1 in a conflictfree manner, we would need all the possible configurations of the window b allowed by the conflict-free coloring of the face f2 . The algorithm performs dynamic
programming starting by computing all possible configurations of the leaves of G∗ and propagating them towards the root in a compatible manner (conflict-freely).
Let f be a face of G and f ∗ be the corresponding
dual vertex in G∗ . Let b be the window of f and
let e = b∗ be the dual edge of b connecting f ∗ to its
parent p = p(f ∗ ). For any configuration Cb , we compute
the score S(Cb ), which is the number of colored vertices
corresponding to Cb in the conflict-free k-coloring
of the
subgraph Ge . We store the pairs Cw , S(Cw ) which
are then combined with the other children of p to
compute the compatible configurations of p. Given a
window w of a face fl , the algorithm GenerateScore
computes S(Cw ) for a given configuration Cw . Let fl
Lemma 5.5. For a fixed k ≥ 1, we can compute the
scores S(Cb ) for all configurations Cb of all windows b
in O(n4k+1 ) time.
Proof. We process the dual graph G∗ starting from the
leaves. Let b be the window between the two faces f1
and f2 . The window corresponds to an edge between
the a dual vertex and its parent in the dual graph. Let
f1 = he1 = (u1 , v1 ), . . . , e` = (u` , v` )i such that e1 = b,
v` = u1 , and vi = ui+1 for i ∈ {1, . . . , ` − 1}. We
compute S(Cb ) by applying Algorithm 2. Inductively,
we can compute the score for all configurations of all
windows going up in the dual graph in this manner.
For each window there are at most O(n2k ) configurations. This implies that for each pair of edges, there
are at most O(n4k ) pairs of configurations. As Algorithm 2 considers O(n) pairs of edges overall, we obtain
a running time of O(n4k+1 ) for the algorithm.
Proof. [Proof of Theorem 5.2] By applying the approach
of Algorithm 2 we can compute the scores of all windows
Algorithm 2 Processing a configuration of a window
1: function GenerateScore(C
e1 , f = he1 = (u1 , v1 ), . . . , e` = (u` , v` )i)
2:
P1 ← Ce1 , CuS1
3:
for i = 2, . . . , ` do
4:
Pi ← ∅
5:
for (Cei−1 , h) ∈ Pi−1 do
6:
for Cei ∈ L(ei ) do
7:
if Cei−1 is compatible with Cei then
8:
Pi ← Pi ∪ {(Cei , h + δ(Cei , Cei−1 ))}
9:
10:
11:
12:
S(Ce1 ) ← ∞
for (Ce` , h) ∈ P` do
if Ce` is compatible with Ce1 then
S(Ce1 ) ← min{S(Ce1 ), h}
of the graph G. At the root node we have a set
of configuration for each window that results in the
minimum number of colored vertices in the whole graph.
Such a set can be obtained by backtracking. Combining
this with Observation 5.3, we get a conflict-free coloring
with a minimal number of colored vertices for the graph
G, if and only if χCF (G) ≤ k.
What remains to be discussed is how we treat the
case in which G∗ is not a tree but a forest (assuming G
is connected). The dual G∗ becomes disconnected if G
has cut edges or cut vertices. In such a case, we use the
following construction depicted in Figure 8 to connect
the components of G∗ to obtain a tree.
(1) For a cut vertex v, let hf1 , . . . , ft i be the t faces
containing v. Let hf1∗ , . . . , ft∗ i be the corresponding
vertices in G∗ . We make one of fi∗ a parent to
all the others by adding an edge between them.
Note that this does not create a cycle because G
is outerplanar.
5.2 Approximability for Three or More Colors.
In Section 4.2 we stated that every planar graph is
conflict-free 3-colorable. In this section we deal with
conflict-free 3-colorings of planar graphs that, additionally, minimize the number of colored vertices.
Theorem 5.6. Let k ≥ 3 and let G be a planar graph.
The following holds:
(1) Unless P = NP, there is no polynomial-time
approximation algorithm providing a constant3
factor approximation of γCF
(G) for planar graphs.
3-Conflict-Free Dominating Set is NPcomplete for planar graphs.
(2) For k ≥ 4, k-Conflict-Free Dominating Set
k
is NP-complete. Also, γCF
(G) = γ(G), and the
problem is fixed-parameter tractable with paramek
(G). Furthermore, there is a PTAS for
ter γCF
k
γCF (G).
k
(3) If G is outerplanar, then γCF
(G) = γ(G) and there
k
is a linear-time algorithm to compute γCF
(G).
(2) If we have a cut edge, we consider the cut edge as
a face. In this way, for a cut edge, we have a vertex The proof of Theorem 5.6 is based on the following
in the dual graph.
polynomial-time algorithm, which transforms a dominating set D of a planar graph G into a conflict-free
k-coloring of G, coloring only the vertices of D: Let D
...
...
be a dominating set of a planar graph G. Every verfi
tex v ∈ V (G) \ D is adjacent to at least one vertex in
D. Pick any such vertex u ∈ D and contract the edge
f1 v ft
f1
f2
uv ∈ E(G) towards u. Repeat this until only the verv
w
tices from D remain. Because G is planar, the graph
G0 = (D, E 0 ) obtained in this way is planar, as G0 is a
(1)
(2)
minor of G. By the 4-coloring theorem, we can compute
a proper 4-coloring of G0 .
Figure 8: Two cases leading to a forest: (1) a cut vertex,
(2) a cut edge.
Lemma 5.7. The 4-coloring generated by this procedure
induces a conflict-free 4-coloring of G.
Proof. Every vertex u ∈ D is a conflict-free neighbor
to itself as its color does not appear in NG (u). Let
v ∈ V (G) \ D be some uncolored vertex, and let u ∈ D
be the vertex that v was contracted towards by the
algorithm. In G0 , this contraction made u adjacent to all
other vertices in NG (v) ∩ D, which guarantees that the
color of u is unique in NG (v) ∩ D. As V (G) \ D remains
uncolored, the color of u is thus unique in NG [v].
Proof. [Proof of Theorem 5.6] Proposition (1) follows
from Theorem 3.7 of Section 3.3: The reduction used
there preserves planarity and proper planar 3-coloring
k
is NP-complete. For (2), γCF
(G) = γ(G) implies NPhardness in planar graphs because planar minimum
dominating set is NP-hard. Moreover, the coloring algorithm lets us apply any approximation scheme for planar dominating set to conflict-free k-coloring. We obtain a PTAS for the conflict-free domination number by
applying our coloring algorithm to the dominating set
produced by the PTAS of Baker and Hill [8]. Additionally, Alber et al. [3] proved that planar dominating set is
FPT with parameter γ(G), implying that computing the
planar conflict-free domination number for k ≥ 4 is FPT
k
with parameter γCF
(G). For (3), the class of outerplanar graphs is properly 3-colorable in linear time and
closed under taking minors. Kikuno et al. [22] present a
linear time algorithm for finding a minimum dominating
set in a series-parallel graph, which includes outerplanar graphs. The result follows by combining this linear
time algorithm with the coloring algorithm mentioned
above, but using just three colors instead of four.
6
Conclusion
A spectrum of open questions remain. Many of them
are related to general graphs, in particular with our
sufficient condition for general graphs. For every k ≥
2, Gk+1 provides an example that excluding Kk+2 as
a minor is not sufficient to guarantee k-colorability.
However, for k ≥ 2 we have no example where excluding
−3
Kk+3
as a minor does not suffice.
Variants of our problems arise from modifying the
considered neighborhoods. In our definition of the
neighborhood N [v] of a vertex v ∈ V , we allow the
vertex itself to serve as the one that is uniquely colored.
In some settings (e.g., for guiding a robot to other locations), it is also interesting to require that another vertex must be uniquely colored. This distinction has been
dubbed “closed” neighborhood (N [v], including v) and
“open” neighborhood (N (v), excluding v) by Gargano
and Rescigno [16]. It would be interesting to expand our
positive results to the case of open neighborhoods; another proof of NP-completeness seems straightforward.
Another distinction arises from requiring that all ver-
tices must be colored. It is clear that one extra color suffices for this purpose; however, it is not always clear that
this is also necessary, in particular, for planar graphs.
Adapting our argument to this situation does not seem
straightforward, especially since there are outerplanar
graphs requiring three colors in this setting.
In addition, there is a large set of questions related
to geometric versions of the problem. What is the
worst-case number of colors for straight-line visibility
graphs within simple polygons? It is conceivable that
Θ(log log n) is the right answer, just like for rectangular
visibility, but this is still an open problem, just like
complexity and approximation. Other questions arise
from considering geometric intersection graphs, such as
unit-disk intersection graphs, for which necessary and
sufficient conditions, just like upper and lower bounds,
would be quite interesting.
Acknowledgments
We thank Bruno Crepaldi, Pedro de Rezende, Cid
de Souza, Stephan Friedrichs, Michael Hemmer and
Frank Quedenfeld for helpful discussions. Work on this
paper was partially supported by the DFG Research
Unit “Controlling Concurrent Change”, funding number
FOR 1800, project FE407/17-2, “Conflict Resolution
and Optimization”.
References
[1] M. A. Abam, M. de Berg, and S.-H. Poon. Faulttolerant conflict-free colorings. In Proc. 20th Canadian
Conference on Computational Geometry (CCCG’08),
pages 13–16, 2008.
[2] D. Ajwani, K. Elbassioni, S. Govindarajan, and
S. Ray. Conflict-free coloring for rectangle ranges using
O(n.382 ) colors. In SPAA ’07: Proc. 19th ACM Symposium on Parallelism in Algorithms and Architectures,
pages 181–187, 2007.
[3] J. Alber, M. R. Fellows, and R. Niedermeier.
Polynomial-time data reduction for dominating set.
Journal of the ACM, 51(3):363–384, 2004.
[4] N. Alon and S. Smorodinsky. Conflict-free colorings
of shallow discs. In Proc. 22nd Annual Symposium on
Computational Geometry, pages 41–43. ACM, 2006.
[5] K. Appel and W. Haken. Every planar map is four colorable. Part I. Discharging. Illinois J. Math., 21:429–
490, 1977.
[6] K. Appel and W. Haken. Every planar map is four
colorable. Part II. Reducibility. Illinois J. Math.,
21:491–567, 1977.
[7] P. Ashok, A. Dudeja, and S. Kolay. Exact and FPT algorithms for max-conflict free coloring in hypergraphs.
In Proc. 26th International Symposium on Algorithms
and Computation, pages 271–282, 2015.
[8] B. S. Baker. and M. Hill. Approximation algorithms
for NP-complete problems on planar graphs. Journal
of the ACM, 41(1):153–180, 1994.
[9] A. Bar-Noy, P. Cheilaris, S. Olonetsky, and
S. Smorodinsky. Online conflict-free colouring for hypergraphs. Combinatorics, Probability and Computing,
19(04):493–516, 2010.
[10] P. Cheilaris, L. Gargano, A. A. Rescigno, and
S. Smorodinsky. Strong conflict-free coloring for intervals. Algorithmica, 70(4):732–749, 2014.
[11] P. Cheilaris, S. Smorodinsky, and M. Sulovsky. The
potential to improve the choice: list conflict-free coloring for geometric hypergraphs. In Proc. 27th Annual
Symposium on Computational Geometry, pages 424–
432. ACM, 2011.
[12] P. Cheilaris and G. Tóth. Graph unique-maximum and
conflict-free colorings. Journal of Discrete Algorithms,
9(3):241–251, 2011.
[13] K. Chen, A. Fiat, H. Kaplan, M. Levy, J. Matousek,
E. Mossel, J. Pach, M. Sharir, S. Smorodinsky, U. Wagner, and E. Welzl. Online conflict-free coloring for intervals. SIAM J. Computing, 36:1342–1359, 2007.
[14] K. Elbassioni and N. H. Mustafa. Conflict-free colorings of rectangles ranges. In Annual Symposium on
Theoretical Aspects of Computer Science, pages 254–
263. Springer, 2006.
[15] G. Even, Z. Lotker, D. Ron, and S. Smorodinsky.
Conflict-free colorings of simple geometric regions with
applications to frequency assignment in cellular networks. SIAM Journal on Computing, 33(1):94–136,
2003.
[16] L. Gargano and A. A. Rescigno. Complexity of
conflict-free colorings of graphs. Theoretical Computer
Science, 566:39–49, 2015.
[17] R. Glebov, T. Szabó, and G. Tardos. Conflict-free
coloring of graphs. Combinatorics, Probability and
Computing, 23:434–448, 2014.
[18] H. Hadwiger. Über eine Klassifikation der Streckenkomplexe. Vierteljschr. Naturforsch. Ges. Zürich,
88:133–143, 1943.
[19] S. Har-Peled and S. Smorodinsky. Conflict-free coloring of points and simple regions in the plane. Discrete
& Computational Geometry, 34(1):47–70, 2005.
[20] F. Hoffmann, K. Kriegel, S. Suri, K. Verbeek, and
M. Willert. Tight bounds for conflict-free chromatic
guarding of orthogonal art galleries. In 31st International Symposium on Computational Geometry, volume 34 of LIPIcs, pages 421–435, 2015.
[21] E. Horev, R. Krakovski, and S. Smorodinsky. Conflictfree coloring made stronger. In Proc. 12th Scandinavian Symposium and Workshop on Algorithm Theory,
volume 6139, pages 105–117, 2010.
[22] T. Kikuno, N. Yoshida, and Y. Kakuda. A linear
algorithm for the domination number of a seriesparallel graph. Discrete Applied Mathematics, 1983.
[23] N. Lev-Tov and D. Peleg. Conflict-free coloring of unit
disks. Discrete Applied Mathematics, 157(7):1521–
1532, 2009.
[24] W. Mulzer and G. Rote. Minimum-weight triangulation is NP-hard. Journal of the ACM, 55(2):11, 2008.
[25] J. Pach and G. Tárdos. Conflict-free colourings of
graphs and hypergraphs. Combinatorics, Probability
and Computing, 18(05):819–834, Sept. 2009.
[26] N. Robertson, D. Sanders, P. Seymour, and
R. Thomas. The four-colour theorem. J. Combinatorial Theory Series B, 70:2–44, 1997.
[27] S. Smorodinsky. Combinatorial Problems in Computational Geometry. PhD thesis, School of Computer
Science, Tel-Aviv University, 2003.
[28] M. M. Sysło. Characterizations of outerplanar graphs.
Discrete Mathematics, 26:47–53, 1979.
[29] R. Wilson. Four colours suffice: How the map problem
was solved. Princeton University Press, 2013.
| 8cs.DS
|
Gayaz Khakimzyanov
Institute of Computational Technologies, Novosibirsk, Russia
Denys Dutykh
arXiv:1706.08790v1 [physics.flu-dyn] 27 Jun 2017
CNRS, Université Savoie Mont Blanc, France
Numerical modelling of surface
water wave interaction with a
moving wall
arXiv.org / hal
Last modified: June 28, 2017
Numerical modelling of surface water wave
interaction with a moving wall
Gayaz Khakimzyanov and Denys Dutykh∗
Abstract. In the present manuscript we consider the practical problem of wave interaction with a vertical wall. However, the novelty here consists in the fact that the wall
can move horizontally due to a system of springs. The water wave evolution is described
with the free surface potential flow model. Then, a semi-analytical numerical method
is presented. It is based on a mapping technique and a finite difference scheme in the
transformed domain. The idea is to pose the equations on a fixed domain. This method
is thoroughly tested and validated in our study. By choosing specific values of spring
parameters, this system can be used to damp (or in other words to extract the energy of)
incident water waves.
Key words and phrases: wave/body interaction; full Euler equations; movable wall;
wave run-up; wave damping; wave energy extraction
MSC: [2010] 76B15 (primary), 76B07, 76M20 (secondary)
PACS: [2010] 47.35.Bb (primary), 47.35.Fg (secondary)
∗
Corresponding author.
3 / 72
Wave/moving wall interaction
Contents
1
Introduction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2
Problem formulation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Potential flow model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Initial conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Piston motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Dimensionless problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3
Equations on a fixed domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1 On some properties of elliptic operators in curvilinear coordinates . . . . . . . . . . . . . . . 15
Example 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Example 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4
Finite difference scheme in curvilinear coordinates
. . . . . . . . . . . . . . 18
4.1 Approximation in interior nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Approximation of boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Difference operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4 Mathematical study of the finite difference problem for the velocity potential . . . . . . 27
Some discrete identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.5 The main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Final remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5
Numerical algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.1 Approximation of the dynamic boundary condition (1) . . . . . . . . . . . . . . . . . . . . . . . . 38
5.2 Computation of the velocity potential (2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3 Free surface motion (3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.4 Moving wall displacement (4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.5 Elliptic mapping construction (5) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
The re-computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.6 Stability of the scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Practical choice of the time step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6
Numerical results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.1 Solitary wave run-up on a fixed wall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.2 Wave generation by a numerical wave maker . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.3 Surface wave run-up on a movable wall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Single pulse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Solitary wave . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3 Side effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
A
Transformation of coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
B
Transformation of the Laplace operator . . . . . . . . . . . . . . . . . . . . . 65
C
Boundary conditions transformation . . . . . . . . . . . . . . . . . . . . . . . 66
D
Piston motion modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
References
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Wave/moving wall interaction
5 / 72
1. Introduction
The mathematical and numerical high fidelity modeling of water waves is a central topic
in coastal and naval engineering. The incident waves come and interact with various
coastal features. Nowadays the interaction of water waves with bathymetric features is
relatively well understood (at least with some special features [49]). A more challenging
situation is the interaction of waves with various coastal structures [53]. At this level the
most studied situation is the wave/wall interaction and the wall is assumed to be fixed
and impermeable. Violent wave impacts have to be modelled in general using more CFDlike methods in the Eulerian mesh-based [9, 20] or Lagrangian particle-based [21]
approaches. In the present study we apply the free surface approximation by neglecting all
the processes happening in the air above the free surface [10]. In this line of thinking, the
interaction of periodic waves with a fixed vertical wall was recently studied in [5] in the
framework of a fully nonlinear weakly dispersive wave model [24]. The conditions under
which an extreme wave run-up on a vertical wall may happen were describe in [5] as well.
In this study, we focus on wave interactions with a movable vertical wall. The wall
motion can be prescribed. In this case we model the wave generation process with a
piston-type wave maker [38]. For instance, this problem was considered in the framework
of Boussinesq-type equations in [63]. Otherwise, the wall can move under the action of
incident waves. We can even assume that a system of horizontal tension/extension springs
(with tunable rigidities) is attached to the wall. Thus, the wall may present a certain
resistance to the action of waves. This problem can be regarded also as a piston’s free
motion under the forcing of water waves. The piston mechanical energy conversion and
recuperation is a different technological problem which is out of scope of the present study.
The extraction of ocean wave energy on industrial scales is not yet a very common practice
[33]. However, very active researches in this field are conducted since at least forty years
[31, 32]. Consequently, the numerical methods developed below can be used to design and
to optimize such Wave Energy Conversion (WEC) devices. Movable walls have been used,
for example, in a triplate system proposed back in 1977 by Dr. Francis Farley. The
rigidity of strings is chosen to minimize the reflected wave amplitude so that most of the
wave energy be converted into the mechanical energy of the device. Mathematical and
numerical modeling of this type of WEC devices is considered below. The main point is
that (ordinary and partial) differential equations posed in time-varying domains are known
to pose notorious theoretical and numerical difficulties [47].
In the present work we consider a two-dimensional non-stationary problem of surface
wave motion in a domain with one moving wall. We assume that the wall remains vertical
during its motion. Moreover, at least in a vicinity of the moving wall the bathymetry has to
be flat to allow its free motion under the action of waves or to follow a prescribed trajectory.
The fluid flow is assumed to be potential and we address the complete (i.e. fully nonlinear
and fully dispersive) water wave problem [73]. We propose first a reformulation of this
problem on a fixed and domain (a square) by introducing a new curvilinear coordinate
G. Khakimzyanov & D. Dutykh
6 / 72
system. Then, we propose a robust finite difference discretization of this problem with
mathematically proven good qualities. The performance of this algorithm is illustrated
on several examples of practical interest. In particular, we study the influence of springs
rigidity on the oscillatory motion of the wave/wall system.
The results presented in this study can be viewed under a different angle and, thus, they
can be applied to another problem of coastal structures protection from wave loads and
impacts. In particular, the catastrophic consequences of 2011 Tohoku earthquake and
tsunami are widely known nowadays [61]. The protecting structures are made very solid
by implementing various security coefficients. However, more economic protections can be
designed if we allow them to move under the action of waves. The moving parts can be
related to fixed ones by a system of flexible strings and by tuning their rigidity one can
reduce significantly the wave run-up on moving parts as well as wave loads on fixed solid
structural elements.
We would like to mention also related works. The generation of water waves by an
accelerating moving vertical plate in a channel of constant depth was studied in [77]. The
Authors measured the free surface elevation and pressure distributions on the wall for three
different values of plate’s accelerations. In another recent work [76] a nonlinear model for
incompressible free surface flows was proposed. Then, this model was used to study wave
run-up on a sloping beach and on a moving vertical wall. The two-dimensional unsteady
problem of wave interaction with a moving vertical wall was studied analytically in [48].
Moreover, the comparisons with numerical solutions obtained with the complex boundary
element method was presented as well. Another numerical model was presented in [40] as
well. In last two works the waves were generated by an initial disturbance located close to
the moving wall.
The present manuscript is organized as follows. The physical and mathematical problems are formulated in Section 2 and then it is reformulated on a fixed domain in Section 3.
The proposed finite difference approximation is described in Section 4 and studied mathematically in Section 4.4. The construction of the mapping in the fixed domain is discussed
in Section 5.5 and the numerical algorithm in general is described in Section 5. Some numerical results are presented in Section 6. Finally, the main conclusions and perspectives
are outlined in Section 7.
2. Problem formulation
Consider a two-dimensional motion of an incompressible, homogeneous and ideal fluid
with free surface. The sketch of the fluid domain is depicted in Figure 1. The Cartesian
coordinate system O x y , x = (x, y) is chosen such that the horizontal axis O x is
directed along the undisturbed free surface and the axis O y points vertically upwards.
The fluid domain Ω (t) is assumed to be simply connected and bounded from below by a
def
horizontal bottom Γb := {x ∈ Ω (t) | y = −h0 } and from above by the free surface
def
— Γs := {x ∈ Ω (t) | y = η(x, t)} . On the sides the domain Ω (t) is bounded by two
7 / 72
Wave/moving wall interaction
y
s(t)
y=η(x,t)
O
x
h0
Figure 1. Sketch of the physical problem of wave interaction with a moving wall.
def
vertical walls. The right wall Γr := {x ∈ Ω (t) | x = ℓ} is assumed to be fixed∗, while
def
the left one Γℓ := {x ∈ Ω (t) | x = s (t)} is connected to a system of springs and
can move horizontally under the wave action. By s (t) we denote the displacement of the
moving wall† with respect to its undisturbed position x = 0 . The closed domain Ω (t)
together with its boundaries is denoted by Ω̄ (t) .
Remark 1. In order to be able to simulate water waves with a shape, which is not necessarily a graph, we can adopt a parametric representation of the free surface Γs :
x = ξ (q, t) ,
x = η (q, t) ,
(2.1)
where q ∈ R is a real parameter and t is the time evolution variable. Functions ξ (·, ·) :
R×R + 7→ R and η (·, ·) : R×R + 7→ R are assumed to be sufficiently smooth. Without
any restriction we can assume that q ∈ [ 0, 1 ] . Below in the text (see Section 3) this
parametrization will appear more naturally when we map the unknown fluid domain Ω (t)
to a fixed reference domain.
The fluid flow is described by Euler equations [52]:
∇ ·u = 0,
1
∇p = g ,
u t + (u · ∇) u +
ρ
∗Thus,
(2.2)
(2.3)
the length of the undisturbed domain is equal to ℓ > 0 .
assume additionally that the wall is non-deformable and remains vertical during the interaction
with waves.
†We
8 / 72
G. Khakimzyanov & D. Dutykh
where u = (u, v) is the velocity vector, ρ = const > 0 is the fluid density, p is the
total pressure and g = (0, − g) is the gravity acceleration. The subscripts such as u t
def
denote partial derivatives, i.e. u t := ∂u
. These equations have to be supplemented by ap∂t
propriate boundary conditions in order to have a well-posed problem. The impermeability
condition on fixed solid surfaces (the right vertical wall Γr and the bottom Γb ) reads
u·n = 0,
x ∈ Γb ∪ Γr ,
t > 0,
where n is the vector of unitary exterior normal to the corresponding boundary of the
domain Ω (t) . The left wall is impermeable as well and the boundary condition reads
u · n = ṡ (t) ,
x ∈ Γℓ ,
t > 0.
Taking into account the horizontal character of the wall motion, this condition can be
further simplified to give
u = ṡ (t) ,
x ∈ Γℓ ,
t > 0.
Traditionally on the free surface we prescribe the kinematic
ηt + u · ηx = v ,
x ∈ Γs ,
t > 0
(2.4)
p = pa = const ,
x ∈ Γs ,
t > 0
(2.5)
and dynamic
boundary conditions [73]. Here pa is the constant atmospheric pressure. The boundary conditions mentioned above are enough to obtain a closed system of equations which describe
the motion of an ideal fluid. One has to prescribe also the initial conditions, but their form
is highly problem-dependent. Consequently, it will not be discussed at the current level of
generality.
Remark 2. Let us obtain also kinematic free surface boundary conditions when the free
surface Γs is given in a parametric form (2.1). The free surface Γs is characterized by
the fact that its velocity is determined by the velocity of fluid particles constituting this
boundary. In other words, the point x ∈ Γs moves with the velocity of the fluid particle
located in this point, i.e.
dx
= u x(t), t .
(2.6)
dt
By taking the material derivative of both sides of the parametric form (2.1), we have
dx
∂ξ
∂ξ dq
≡
+
·
,
dt
∂t
∂q dt
∂η
∂η dq
dy
≡
+
·
.
dt
∂t
∂q dt
9 / 72
Wave/moving wall interaction
Thus, by taking into account the physical information (2.6) we obtain the following kinematic boundary conditions for ∀ x ∈ Γs and for ∀ t > 0 :
∂ξ
∂ξ dq
+
·
− u = 0,
∂t
∂q dt
∂η dq
∂η
+
·
− v = 0.
∂t
∂q dt
(2.7)
(2.8)
In the conditions above the quantity dq
is the rate of change of the parameter q for fluid
dt
particles located at the free surface. If one uses the Lagrangian description of a flow, fluid
particles keep their labels and, thus, dq
≡ 0 . However, we use an arbitrary parametrization
dt
dq
and in our case dt 6= 0 in general.
2.1. Potential flow model
The formulation presented above is still too general. We shall simplify it further by
assuming the flow to be irrotational, i.e.
rot u = −
∂v
∂u
+
= 0.
∂y
∂x
It implies the existence of a function ϕ : R2 × R+ 7→ R which is called the velocity
potential such that the velocity field is given by
(2.9)
u = ∇ϕ.
Substituting this form into the incompressibility condition (2.2) yields that the velocity
potential has to be a harmonic function
∆ϕ = 0,
(x, t) ∈ Ω (t) × R+ .
The kinematic boundary condition is obtained straightforwardly by substituting into (2.4)
the velocity components by their representations in terms of the velocity potential:
ηt + ϕx · ηx = ϕy ,
y = η (x, t) .
We note also that velocity components u and v in the kinematic boundary conditions (2.7),
(2.8) are expressed in terms of the velocity potential ϕ according to (2.9).
The momentum equation (2.3) can be integrated and combined with the dynamic boundary condition (2.5) to give the so-called Euler–Lagrange integral∗ evaluated at the free
surface:
ϕ t + 21 | ∇ϕ | 2 + g η = 0 ,
y = η (x, t) .
On solid stationary walls the velocity potential satisfies the homogeneous Neumann condition
∇ϕ · n = 0 ,
x ∈ Γb ∪ Γr ,
∗The
Euler–Lagrange integral becomes the Bernoulli integral for steady flows. We implicitly
chose the gauge where the Bernoulli ‘constant’ is identically zero.
10 / 72
G. Khakimzyanov & D. Dutykh
(with n defined above) while on the moving wall Γℓ it satisfies the following non-homogeneous
Neumann condition
∂ϕ
∇ϕ · n ≡
= ṡ (t) ,
x ∈ Γℓ .
∂x
So, instead of seeking for two components of the velocity field u(x, t) = u(x, t), v(x, t)
and the pressure function p(x, t) , the formulation was simplified to one unknown function,
i.e. the velocity potential ϕ(x, t) , thanks to the irrotationality assumption. The potential
flow formulation with free surface is known as the full water wave problem and it was shown
in numerous studies to be an excellent model for water waves [41, 73] (see [12] for water
wave problem history review).
2.1.1 Initial conditions
In order to obtain a well-posed problem, we have to specify also the appropriate initial
conditions. The free surface elevation in the parametric form is initially given by two
functions:
ξ (q, 0) = ξ 0 (q) ,
η (q, 0) = η 0 (q) ,
q ∈ [ 0, 1 ] .
Let us assume that the initial distribution of fluid particles is known as well:
u (x, 0) = u 0 (x) ,
x ∈ Ω (0) .
Then, we can construct an initial condition for the velocity potential ϕ in the following
way:
ˆ
ϕ (x, 0) = ϕ 0 (x) ≡ ϕ 0 (x 0 ) +
u 0 dx + v 0 dy ,
γx 0 ;x
where ϕ 0 (x 0 ) is the value of the velocity potential in an arbitrary point x 0 ∈ Ω̄ (0)
and γx 0 ;x ⊆ Ω̄ (0) is an arbitrary piecewise smooth curve connecting points x 0 with
x . We assume that the initial distribution of the velocities u 0 (x) is potential. Thus, the
curvilinear integral above does not depend on the path γx 0 ;x .
2.2. Piston motion
In order to describe the horizontal motion of the piston we adopt the following very
simple model based on the second law of Newton [50, 62]:
m s̈ + k s = − F (t) − F (0) ,
(2.10)
where m is the wall mass and k is the stiffness coefficient of springs. Finally, F (t) is the
force acting on the left moving wall. It can be computed by integrating the contributions
of the whole water column
ˆ ηs (t)
def
F (t) =
p s(t), y, t dy ,
η s (t) := η s (t), t .
−h0
11 / 72
Wave/moving wall interaction
If initially the fluid was at rest, the force F (0) consists only of the hydrostatic loading on
the wall, i.e.
h 20
F (0) = g
.
2
(2.11)
The pressure p (x, t) ≡ p (x, y, t) can be computed at any point inside the fluid thanks
to the Euler–Lagrange integral:
p (x, y, t) = − ϕ t −
1
2
| ∇ϕ | 2 − g y .
A more detailed derivation of equation (2.10) can be found in Appendix D.
The second order nonlinear and non-autonomous Ordinary Differential Equation (ODE)
(2.10) has to be completed with two initial conditions. If nothing is indicated, we start the
integration from the rest state, i.e.
s (0) = ṡ (0) = 0 .
2.3. Dimensionless problem
Equations given in previous Section can be further simplified by choosing appropriate
scaled variables. Namely, we introduce the following scaled independent
x
,
x =
h0
∗
y
∗
y
=
,
h0
∗
t
= t
r
g
h0
and dependent
s∗ =
s
,
h0
η∗ =
η
,
h0
ϕ∗ =
ϕ
√
,
h0 g h0
p∗ =
p
,
ρ g h0
F∗ =
F
ρ g h 30
variables. The dimensional coefficients and parameters appearing in governing equations
scale as follows:
ℓ∗ =
ℓ
,
h0
m∗ =
m
,
ρ h 03
k∗ =
k
.
ρ g h 20
12 / 72
G. Khakimzyanov & D. Dutykh
Finally, dimensionless governing equations read (where for simplicity we drop out all asterisk ∗ symbol from superscripts):
∇2 ϕ = 0 ,
ϕt +
1
2
ηt + ϕx · ηx = ϕy ,
|∇ϕ|
2
+ η = 0,
ϕy = 0 ,
ϕ x = ṡ ,
(x, t) ∈ Ω (t) × R+ ,
(2.12)
y = η (x, t) ,
(2.14)
y = −1 ,
(2.15)
y = η (x, t) ,
x = s (t) ,
ϕ x = 0,
x = ℓ,
−1 6 y 6 η (s(t), t) ,
−1 6 y 6 η (ℓ, t) ,
m s̈ + k s = − F (t) − F (0) ,
ˆ η (s(t), t)
F (t) =
p s (t), y, t dy ,
(2.13)
(2.16)
(2.17)
(2.18)
(2.19)
−1
p (x, t) = − ϕt −
1
2
| ∇ϕ |2 − y .
(2.20)
In numerical simulations below we shall solve this dimensionless system of equations. It is
equivalent to setting simply h0 = 1 and g = 1 in the computer code.
3. Equations on a fixed domain
The main difficulty of the problem described above is that the computational domain
Ω (t) is time dependent. First of all, it is due to the motion of the free surface, but also due
to the motion of the left wall. Consequently, the strategy adopted in this study consists in
transforming the problem to a fixed domain Q 0 . Obviously, this transformation will be time
dependent due to the unsteady character of the problem. In the past; the idea of using
conformal mappings from Complex Analysis has become popular in 2D Hydrodynamics
[52]. For the 2D water wave problem it was proposed by L. Ovsyannikov (1974) [64]
and implemented much later numerically by A. Dyachenko et al. (1996) [29]. Strictly
speaking, in our developments we do not need the conformal property of the underlying
mapping. Consequently we shall relax this assumption. Moreover, it will allow us to have
a fixed and finite transformed domain Q 0 . For instance, we can choose Q 0 as a unit square
def
Q 0 := (q 1 , q 2 ) | 0 6 q 1 , q 2 6 1 ≡ [0, 1]2
and we consider a smooth bijective mapping (i.e. a diffeomorphism) x : Q 0 7→ Ω (t) :
(
x = x (q 1 , q 2 , t) ,
x = x (q, t)
⇐⇒
(3.1)
y = y (q 1 , q 2 , t) .
x−1 (q, t)
Q0
Ω (t)
x (q, t)
(3.2)
13 / 72
Wave/moving wall interaction
Figure 2. Schematic illustration of the mapping (3.1). Dashed lines show the
correspondence between two boundaries of domains Q 0 and Ω (t). The same
property holds for two other lateral boundaries, but it is not shown for the sake of
picture clarity.
Additionally we assume that the Jacobian matrix J (q, t) of transformation (3.1) is nonsingular, i.e.
def
J (q, t) :=
xq1 xq2
yq1 yq2
≡ x q 1 y q 2 − x q 2 y q 1 6= 0 .
(3.3)
This situation is schematically depicted in diagram (3.2) and in Figure 2. For more details
on the computation of partial derivatives we refer to Appendix A.
It is assumed that left Γℓ and right Γr walls are images of left and right sides of the
square Q 0 , i.e. q 1 = 0 and q 1 = 1 correspondingly (0 6 q 2 6 1). Similarly, the
upper (q 2 = 1) and lower (q 2 = 0) sides of the square are respectively mapped on the
free surface Γs and bottom Γb . The boundary of the square Q 0 will be denoted by γ , i.e.
def
γ := ∂ Q 0 = γ s ∪ γ b ∪ γ r ∪ γ ℓ .
The practical construction of the mapping x (q, t) is described in Section 5.5. At the
present stage of the exposition we shall assume that mapping x (q, t) is simply known.
In this Section we shall skip the intermediate computations, which are rather standard
in the Differential Geometry, for example. We detail them in Appendices A and B. The
14 / 72
G. Khakimzyanov & D. Dutykh
system of governing equations (2.12) – (2.20) posed on the fixed domain Q 0 reads:
∂
∂ϕ
K
= 0,
q ∈ Q 0 , α, β = 1, 2,
α
β
∂q α
∂q β
q2 = 1,
ηt + v1 · ηq1 = v ,
q2 = 1,
ϕ t − (u · xt + v · yt ) + 12 |∇x ϕ|2 + η = 0 ,
∂ϕ
∂ϕ
K21 1 + K22 2 = 0 ,
q2 = 0,
∂q
∂q
∂ϕ
∂ϕ
K 1 1 1 + K 1 2 2 = yq 2 ṡ ,
q1 = 0,
∂q
∂q
(3.4)
(3.5)
(3.6)
(3.7)
0 6 q2 6 1,
(3.8)
K11
∂ϕ
∂ϕ
+ K 1 2 2 = 0,
1
∂q
∂q
ˆ
F (t) =
q 1 = 1,
1
0
0 6 q2 6 1,
p 0, q 2 , t dq 2 ,
p = − ϕ t + (u · xt + v · yt ) −
(3.9)
(3.10)
1
2
|∇x ϕ|2 − y ,
(3.11)
where the implicit convention about the summation over repeated indices is adopted. The
transformation of the Laplace equation is detailed in Appendix B.
Remark 3. If the free surface is given in the parametric form (2.1), then we have to add
one more kinematic free surface boundary condition:
ξt + v1 · ξq1 = u ,
q ∈ γs .
(3.12)
See also Appendix C for some hints on the derivation of boundary conditions in the transformed domain.
Notice that equation (2.18) is not affected by transformation (3.1). That is why we
do not repeat it here. Above we introduced the following notations. The coefficients in
transformed Laplacian are defined as
g12
def g 2 2
def
def g 1 1
K 1 1 :=
,
K 1 2 ≡ K 2 1 := −
,
K 2 2 :=
,
(3.13)
J
J
J
where G = (gαβ )α,β = 1, 2 is the familiar metric tensor
def
g 1 1 := (x q 1 )2 + (y q 1 )2 ,
def
g 1 2 ≡ g 2 1 := x q 1 x q 2 + y q 1 y q 2 ,
def
g 2 2 := (x q 2 )2 + (y q 2 )2 . (3.14)
In fact, tensor K = (K αβ )α, β = 1,2 can be expressed in a compact matrix form through the
metric tensor G as
K = J G−1 .
15 / 72
Wave/moving wall interaction
The Cartesian (u, v) and the contravariant v 1 components of the velocity field are defined
as
def 1
u :=
(3.15)
ϕq 1 · y q 2 − ϕq 2 · y q 1 ,
J
def 1
v :=
(3.16)
−ϕq 1 · xq 2 + ϕq 2 · xq 1 ,
J
1
∂q 1
∂q 1
1 def ∂q
v :=
+ u·
+ v·
.
(3.17)
∂t
∂x
∂y
In order to compute the derivatives of the inverse mapping, we use the following expressions:
∂q 1
1 h ∂y ∂x
∂x ∂y i
1 ∂y
∂q 1
1 ∂x
∂q 1
,
=
−
=
,
= −
. (3.18)
2
2
2
∂t
J ∂t ∂q
∂t ∂q
∂x
J ∂q
∂y
J ∂q 2
1 h ∂x ∂y
∂y ∂x i
∂q 2
1 ∂y
∂q 2
1 ∂x
∂q 2
=
−
,
=
−
,
=
. (3.19)
1
1
1
∂t
J ∂t ∂q
∂t ∂q
∂x
J ∂q
∂y
J ∂q 1
The mathematical problem of a potential flow formulation with free surface in curvilinear
coordinates consists in determining the velocity potential ϕ (q 1 , q 2 , t) and the free surface
profile η (q 1 , t) . The velocity potential satisfies in the domain Q 0 an elliptic equation (3.4)
with non-constant coefficients along with boundary conditions (3.5) – (3.9) on ∂Q 0 . In
order to obtain a well-posed problem we have to complete the formulation with two initial
conditions at t = 0 directly in the fixed domain:
ϕ (q 1 , 1, 0) = ϕ 0 (q 1 ) ,
η (q 1 , 0) = η 0 (q 1 ) ,
q = (q 1 , 1) ∈ Γs ,
q = (q 1 , 1) ∈ Γs .
Earlier the initial conditions were discussed in Section 2.1.1. They can be transposed on
the fixed domain Q 0 using the inverse mapping x −1 (q, t) .
3.1. On some properties of elliptic operators in curvilinear coordinates
It can be noticed that the Laplace equation in curvilinear coordinates (3.4) has a
more complex form comparing to the initial Cartesian coordinates O x y . In particular,
constant coefficients become variable in space and time. Moreover, the mixed derivatives
appear as well. However, some properties are important to construct qualitatively correct
numerical discretizations. For example, below we shall construct a finite difference scheme
for equation (3.4) with a positive definite difference operator. The proof of this fact relies
heavily on the uniform ellipticity property of operator (3.4).
Lemma 1. Partial differential equation (3.4) is uniformly elliptic.
Proof. By our assumptions the mapping (3.1) is differentiable with bounded derivatives in
Q 0 and non-degenerate with a positive Jacobian J > 0 , ∀q ∈ Q 0 , ∀t > 0 . Thanks
to the identity
g 1 1 · g 2 2 ≡ J 2 + g 12 2 ,
16 / 72
G. Khakimzyanov & D. Dutykh
the metric components g 1 1 and g 2 2 take strictly positive values. Thanks to definitions
(3.13), it is straightforward to conclude that K 1 1 , K 2 2 are also positive. The matrix
!
K11 K12
K =
K21 K22
is symmetric and its determinant is equal to det K = 1 due to (3.13). Consequently, it
possesses two orthonormal eigenvectors corresponding to eigenvalues
√
g
+
g
−
D
def
1
1
2
2
(−)
0 6 λK :=
6 1,
2J
√
D
(+) def g 1 1 + g 2 2 +
1 6 λK :=
,
2J
where the discriminant is defined as
2
def
D := g 1 1 + g 2 2 − 4 J 2 =
g11 − g22
2
+ 4 g 12 2 > 0 .
Then, for any real numbers ∀ξ , ζ ∈ R and for any point ∀q ∈ Q 0 we have the following
inequalities:
c1 ξ 2 + ζ 2) 6 Q (ξ, ζ) 6 c2 ξ 2 + ζ 2 ) ,
(3.20)
where Q is the quadratic form defined as
def
Q (ξ, ζ) := K 1 1 ξ 2 + 2 K 1 2 ξ ζ + K 2 2 ζ 2 .
(3.21)
The constants c 1, 2 are defined as
def
inf
g11 + g22 −
2J
sup
g11 + g22 +
2J
√
D
> 0,
(3.22)
< ∞.
(3.23)
The positivity of c 1 > 0 implies the uniform ellipticity property.
c1 :=
def
c2 :=
q ∈ Q0
q ∈ Q0
√
D
The constants c 1, 2 defined in the proof above can be used to estimate the convergence
speed of iterative methods, which depends directly on the conditioning number [19] of the
matrix corresponding to the difference operator [69]. The conditioning number depends on
c2
. High ratio implies a poorly conditioned difference operator and, thus, more
the ratio
c1
iterations are needed to converge to the solution within prescribed accuracy. We underline
also that constants c 1, 2 depend only on the mapping (3.1). Below we provide two examples
c2
which illustrate situations where the ratio
≫ 1.
c1
17 / 72
Wave/moving wall interaction
3.1.1 Example 1
Let us take a physical domain in the form of a rectangle:
def
Ω := x = (x, y) | 0 6 x 6 ℓ 1 , 0 6 y 6 ℓ 2
with substantially different sizes of the walls, e.g. we can assume 0 < ℓ 1 ≪ ℓ 2 . The
mapping (3.1) from Q 0 7→ Ω is given explicitly by formulas
x = q1 · ℓ1 ,
y = q2 · ℓ2 .
(3.24)
Then, we can compute explicitly the metric coefficients
g 1 1 = ℓ 21 ,
g 2 2 = ℓ 22 ,
g12 = 0 ,
the Jacobian
J = ℓ1 · ℓ2 ,
and eigenvalues of the matrix K :
ℓ1
ℓ2
,
c2 =
.
ℓ2
ℓ1
Thus, the conditioning of the finite difference operator will scale with
ℓ 2
c2
2
≡
≫ 1.
c1
ℓ1
c2
becomes large. If on the reference domain
Thus, for highly distorted domains the ratio
c1
Q 0 we use a uniform square grid, than mapping (3.24) generates a uniform grid with
distorted cells. The convergence of iterative methods on such grids will be slowed down.
Thus, in real computations we should avoid highly distorted cells since they will penalize
the convergence of linear solvers.
c1 =
3.1.2 Example 2
Let us take another physical domain Ω having the shape of a parallelogram. It can be
obtained as an image of the reference domain Q 0 under the following mapping
x = q 1 + q 2 · cos ψ ,
y = q 2 · sin ψ ,
(3.25)
π
where 0 < ψ ≪
is a small angle. Then, the metric coefficients of this mapping are
2
g11 = 1 ,
g 1 2 = cos ψ ,
g22 = 1 .
The Jacobian J = sin ψ and eigenvalues of the matrix K are
c 1 = tan
and henceforth
ψ
,
2
c 2 = cot
ψ
c2
≡ cot 2
≫ 1.
c1
2
ψ
,
2
18 / 72
G. Khakimzyanov & D. Dutykh
c2
becomes large for domains Ω featuring a small angle.
c1
The problem is that the grid generated by mapping (3.25) is substantially non-orthogonal,
since it consists of parallelograms with an acute angle ψ . On such grids we should expect
some reduction of iterative methods convergence speed. Thus, in our computations we
should avoid grids featuring small angles.
So, we showed that the ratio
4. Finite difference scheme in curvilinear coordinates
In the numerical simulation of free surface potential flows of an ideal fluid, an elliptic
equation to determine the velocity potential ϕ has to be solved at every time step. According to our numerical algorithm (see Section 5 for more details), we determine first
the velocity potential value on the free surface Γs , then using this value we reconstruct
the velocity potential ϕ in the whole fluid domain (by taking into account other boundary
conditions on Γb , Γr and Γℓ ). So, in this Section we assume that we know the values of
the velocity potential in nodes which constitute the pre-image of the free surface Γs . The
goal is to determine the values of the velocity potential in all remaining grid nodes. These
values will be determined by solving a system of difference equations constructed below.
Moreover, we assume that the curvilinear grid Ω nh on the nth time layer is already constructed. The nodes x nj are images under the mapping (3.1) of fixed nodes q j , which
def
constitute the uniformly distributed grid Q 0h ⊆ Q 0 . Here j := (j 1 , j 2 ) is a multi-index
def
and q j := q j11 , q j22 . The uniform grid is traditionally defined as
def
q jαα := j α · h α ,
j α = 0, . . . , N α ,
def
h α :=
1
,
Nα
α = 1, 2 .
We make an additional geometrical (and not very restrictive) assumption on the mapping
(3.1): boundary components γ ℓ , γ b , γ r and γ s are mapped on corresponding boundary
components Γℓ , Γb , Γr and Γs of the fluid domain Ω(t) (see Figure 2 for an illustration).
The boundary of the reference domain Q 0 after the discretization is denoted as
def
γ h := ∂ Q h0 ≡ γ ℓh ∪ γ bh ∪ γ rh ∪ γ sh .
The sketch of the discretized reference domain Q 0h is depicted in Figure 3.
In order to construct a finite difference approximation we employ the integro-interpolation
method using the terminology of Tikhonov & Samarskii [75]. In the western literature
this method is closer to the finite volume/conservative finite difference methods [35, 36].
The choice for finite differences seems to be natural since we would like to solve an elliptic
equation on a simple Cartesian domain [68]. For this purpose we replace equation (3.4)
by the following integral relation
“
̥ 1 dq 2 − ̥ 2 dq 1 = 0 ,
(4.1)
∂C
19 / 72
Wave/moving wall interaction
Figure 3. Sketch of the discretized reference domain Q 0h .
where C ⊆ Q 0 is an arbitrary sub-domain with a piece-wise smooth boundary. We
introduced above the notation
∂ϕ
∂ϕ
∂ϕ
∂ϕ
def
def
̥ 2 := K 2 1 · 1 + K 2 2 · 2 .
̥ 1 := K 1 1 · 1 + K 1 2 · 2 ,
∂q
∂q
∂q
∂q
Below we construct discrete approximations based on this integral formulation.
Remark 4. Please, notice that the transformed Laplace equation (3.4) can be compactly
rewritten using the new notation:
∂̥ 1
∂̥ 2
+
= 0,
∂q 1
∂q 2
q =
q 1, q 2) ∈ Q 0 .
4.1. Approximation in interior nodes
Let us consider an interior grid Q 0h node marked with symbol 0 in Figure 4(a) along
with the integration contour A B C D . For this contour the integral equation (4.1) takes
the form
ˆ
ˆ
ˆ
ˆ
1
2
1
2
2
1
− ̥ dq − − ̥ dq + − ̥ dq − − ̥ 2 dq 1 = 0 .
(4.2)
BC
AD
DC
AB
Now, by applying a quadrature rule to all integrals above, one can obtain a discrete analogue of this integral relation. In the present study we employ the quadrature formula of
20 / 72
G. Khakimzyanov & D. Dutykh
4
8
1
j2
7
D
N
C
W
0
E
A
S
2
B
5
3
6
j1
(a)
4
D
C
0
j2
7
A
3
E
B
6
2
j1=0
(b)
Figure 4. Finite difference stencil and the integration contour (shown with a
dotted line) in an interior (a) and a left boundary (b) node.
21 / 72
Wave/moving wall interaction
rectangles. For the integral over the segment B C we have
K 1 1 (0) + K 1 1 (3) ϕ 3 − ϕ 0
+
·
2
h1
BC
1h
ϕ7 − ϕ6
ϕ4 − ϕ2 i
K 1 2 (3)
h2 =
+ K 1 2 (0)
2
2 h2
2 h2
1n
K 1 1 ϕ ♮q 1 j 1 , j 2 + K 1 1 ϕ q♭ 1 j 1 +1, j 2 +
2
io
1h
h2 ,
K 1 2 ϕ q♮ 2 j 1 +1, j 2 + K 1 2 ϕ ♭q 2 j 1+1, j 2 + K 1 2 ϕ ♮q 2 j 1 , j 2 + K 1 2 ϕ q♭ 2 j 1 , j 2
2
(4.3)
ˆ
−
1
̥ dq
2
≈
where we used the grid numbering shown in Figure 4(a) and we introduced the following
notation
ϕj 1 +1, j 2 − ϕj 1 , j 2
,
h1
def ϕj 1 , j 2 +1 − ϕj 1 , j 2
,
:=
h2
def
ϕj 1 , j 2 − ϕj 1−1, j 2
,
h1
def ϕj 1 , j 2 − ϕj 1 , j 2 −1
:=
.
h2
def
ϕ q♮ 1 , j :=
ϕ ♭q 1 , j :=
ϕ q♮ 2 , j
ϕ ♭q 2 , j
for left- and right-sided finite difference operators. By using similar approximations to
discretize other integrals over sides A D , D C and A B we obtain a fully discrete analogue
of the integral equation (4.1):
h
i
1
K 1 1 ϕ ♮q 1 j 1, j 2 + K 1 1 ϕ ♭q 1 j 1+1, j 2 +
2
i
h
1
♮
♮
♭
♭
h2
K 1 2 ϕ q 2 j 1+1, j 2 + K 1 2 ϕ q 2 j 1 +1, j 2 + K 1 2 ϕ q 2 j 1 , j 2 + K 1 2 ϕ q 2 j 1 , j 2
4
h
i
1
K 1 1 ϕ ♮q 1 j 1 −1, j 2 + K 1 1 ϕ ♭q 1 j 1 , j 2 +
−
2
i
h
1
♮
♮
♭
♭
h2
K 1 2 ϕ q 2 j 1, j 2 + K 1 2 ϕ q 2 j 1, j 2 + K 1 2 ϕ q 2 j 1−1, j 2 + K 1 2 ϕ q 2 j 1−1, j 2
4
h
i
1
K 2 2 ϕ ♮q 2 j 1, j 2 + K 2 2 ϕ ♭q 2 j 1, j 2+1 +
+
2
i
h
1
♮
♮
♭
♭
h1
K 2 1 ϕ q 1 j 1, j 2+1 + K 2 1 ϕ q 1 j 1 , j 2 +1 + K 2 1 ϕ q 1 j 1 , j 2 + K 2 1 ϕ q 1 j 1 , j 2
4
h
i
1
♮
♭
K 2 2 ϕ q 2 j 1 , j 2 −1 + K 2 2 ϕ q 2 j 1 , j 2 +
−
2
i
1h
♮
♮
♭
♭
K 2 1 ϕ q 1 j 1 , j 2 + K 2 1 ϕ q 1 j 1 , j 2 + K 2 1 ϕ q 1 j 1 , j 2 −1 + K 2 1 ϕ q 1 j 1 , j 2 −1
h1 = 0 .
4
22 / 72
G. Khakimzyanov & D. Dutykh
By dividing the both parts of the last equation by the surface h 1 · h 2 of the rectangle
A B C D , we obtain more compact difference equations in any interior node q j ∈ Q 0h :
♮ o
♭
1n
+
K 1 1 ϕ ♮q 1 q 1 + K 1 1 ϕ q♭ 1 q 1
j
2
♮
♭ o
1n
♮
♮
♭
♭
K12 ϕq2 + K12 ϕq2 q1 + K12 ϕq2 + K12 ϕq2 q1
+
4
j
♮ o
♭
1n
K 2 2 ϕ q♮ 2 q 2 + K 2 2 ϕ ♭q 2 q 2
+
2
j
♭ o
♮
1n
♮
♮
♭
♭
= 0.
K21 ϕq1 + K21 ϕq1 q2 + K21 ϕq1 + K21 ϕq1 q2
4
j
2
These
equations approximate the original differential equation (3.4) to the order O h 1 +
2
h 2 provided that solutions are sufficiently smooth. This is achieved by using the 9−point
stencil shown in Figure 4(a).
4.2. Approximation of boundary conditions
For the sake of manuscript conciseness we explain the boundary conditions treatment on
def
the example of condition (3.8) imposed on the fixed∗ boundary γ ℓ . Let γ ℓh := q 0, j 2 ∈
γ h | j 2 = 1, 2, . . . , N2 −1 be the set of ‘interior’ grid nodes belonging to the left boundary
γ ℓ except two angular nodes q 0, N2 and q 0, 0 , which deserve a special consideration. For a
boundary node q j ∈ γ ℓh we choose the integration contour as it is shown in Figure 4(b).
The quadrature rule for the integral over side B C is given in equation (4.3) (one has just
to substitute j 1 ←֓ 0). On the segment A D we have to use the boundary condition (3.8)
that we can rewrite in a compact form:
̥1
def
q ∈γℓ
= y q 2 ṡ =: μ 0 .
(4.4)
Using this information, we have directly the following approximation:
ˆ
− ̥ 1 dq 2 ≅ μ 0j 2 h 2 .
AD
For the side D C the formula of rectangles yield the following approximation of the integral:
ˆ
K 2 2 (0) + K 2 2 (4) ϕ 4 − ϕ 0
2
1
·
+
− ̥ dq ≅
2
h2
DC
ϕ7 − ϕ4
ϕ3 − ϕ0 i h1
1h
K 2 1 (4) ·
+ K 2 1 (0) ·
=
2
h1
h1
2
h
i h
i 1h
1
1
♮
♮
♮
♭
K 2 2 ϕ q 2 j 1, j 2 + K 2 2 ϕ q 2 j 1, j 2+1 +
K 2 1 ϕ q 1 j 1 , j 2+1 + K 2 1 ϕ q 1 j 1, j 2
.
2
2
2
∗The left wall might be moving in the physical space. However, it becomes fixed in the transformed
domain.
23 / 72
Wave/moving wall interaction
For the segment A B we give only the final result without intermediate calculations:
h
ˆ
i
1
2
1
− ̥ dq ≅
K 2 2 ϕ q♮ 2 j 1 , j 2 −1 + K 2 2 ϕ ♭q 2 j 1 , j 2 +
2
AB
i h
1h
1
♮
♮
K 2 1 ϕ q 1 j 1 , j 2 + K 2 1 ϕ q 1 j 1 , j 2 −1
.
2
2
Now we substitute all these approximations into the integral equation (4.2) and after
dividing its both parts by h 2 , we obtain the following difference relation in any node γ ℓh :
i
1h
K 1 1 j 1 + 1 , j 2 ϕ ♮q 1 j +
K 1 2 ϕ ♮q 2 + K 1 2 ϕ ♭q 2 j 1+1, j 2 + K 1 2 ϕ q♮ 2 + K 1 2 ϕ ♭q 2 j +
2
4
h
♮ i 1 h
♭ i
♮
h1
1
♭
♮
♮
♮
♭
K22 ϕq2 q2 + K22 ϕq2 q2 +
K21 ϕq1 q2 + K21 ϕq1 q2
= μ 0j 2 .
2
2
2
j
(4.5)
Here we introduced the following notation:
def
K11
j 1 +1, j 2
+
K11
j ,j
1 2
.
2
This difference equation is inhomogeneous because of the wall motion. If the (left) wall is
fixed, then μ 0 ≡ 0 and this relation becomes homogeneous.
K11
j 1 + 21 , j 2
:=
Remark 5. Please, notice that in all formulas in this Section j 1 ≡ 0 .
4.2.1 Consistency
In this Section we show that equation (4.5) approximates the boundary condition (3.8)
to the order O(h 21 + h 22 ) . Indeed, let us substitute a smooth solution ϕ (q 1 , q 2 , t) into
equation (4.5) and perform local Taylor expansions. Then, the consistency error ψ can
be computed:
1
∂ϕ
∂ϕ
∂ϕ
2
2
K12
+
(E) + O h 1 +
(3) + K 1 2
(0) + O h 2
ψ 0, j 2 = K 1 1
∂q 1
2
∂q 2
∂q 2
h1
∂
∂ϕ
2
K22
(0) + O h 2
+
∂q 2
∂q 2
2
h ∂ϕ
h ∂ϕ
i
i
∂ 2ϕ
∂ 2ϕ
h1
2
2
K21
+ h1
+ O h 1 (4) − K 2 1
+ h1
+ O h 1 (2)
− μ 0j 2
1
1
2
1
1
2
∂q
∂(q )
∂q
∂(q )
4 h2
∂ϕ
∂ϕ
∂ϕ
h1 ∂
h1 ∂
∂ϕ
K11 1 + K12
K12
+
+
+
= K11
1
1
2
1
∂q
2 ∂q
∂q
∂q
2 ∂q
∂q 2
∂ϕ
h1 ∂
∂ϕ
h1 ∂
K22
+
K21
− μ 0j 2 =
2
2
2
1
2 ∂q
∂q
2 ∂q
∂q
1
∂̥ 2 h 1
∂̥
2
2
1
0
+
+
O
h
+
h
̥ − μ +
1
2 .
∂q 1
∂q 2
2
24 / 72
G. Khakimzyanov & D. Dutykh
In the very last expression all quantities are evaluated in the same node q 0, j 2 , j 2 =
1, 2, . . . , N2 −1 . Moreover, taking into account the boundary condition (4.4) and assuming
that Laplace equation (3.4) (see also Remark 4) is fulfilled up to the boundary γ ℓ , we
obtain that
ψ 0, j 2 = O h 21 + h 22 .
Thus, the boundary condition is approximated to the second order in space. In a similar
def
way we can construct finite difference approximations in boundary nodes γ bh := q j 1 , 0 ∈
def
γ h | j 1 = 1, 2, . . . , N1 − 1 and γ rh := q N1 , j 2 ∈ γ h | j 2 = 1, 2, . . . , N2 − 1 . In
the derivation of these boundary conditions one has to use the following impermeability
conditions (compare with conditions (3.7) and (3.9)):
̥2
q ∈γb
= 0,
̥1
q ∈γr
= 0.
All these difference equations have six point stencils to achieve the second order accuracy:
three nodes lie in the interior of the domain Q h0 and three on the boundary γ h . Finally,
there are also two angular nodes q 0, 0 and q N1 , 0 . The stencils contain four nodes: one
interior and three on the boundaries.
4.3. Difference operator
In order to determine the discrete values of the velocity potential ϕ j , we constructed
above a finite difference problem with the uniform second order accuracy. In this difference
problem one has to find the values of the velocity potential ϕ j in all nodes except those
on the free surface, where the Dirichlet-type condition is imposed:
ϕ j 1, N2 = μ sj 1 ,
j 1 = 0, 1, . . . , N1 .
In this way we have to determine only N1 + 1 · N2 the values ϕ j j in the nodes
x j 1, j 2 with j 1 = 0, 1, .. . , N1 and j 2 = 0, 1, . . . , N2 − 1 . In the resulting problem
we have precisely N1 + 1 · N2 equations.
In order to study theoretically the finite difference problem, it will be convenient to make
a change of variables in order to get homogeneous Dirichlet’s boundary conditions in
nodes q j 1, N2 , j 1 = 0, 1, . . . , N1 . For this purpose we introduce a new grid function
( s
μj1 ,
j 1 = 0, 1, . . . , N1 , j 2 = N2 ,
def
ϕ̃ j 1, j 2 :=
0,
j 1 = 0, 1, . . . , N1 , j 2 < N2 ,
and a new dependent variable
def
φ := ϕ − ϕ̃ .
h
Then, in the nodes q j ∈ γ s on the free surface the function φ j will take zero values. The
difference problem for φ will differ from the one for ϕ by inhomogeneous right hand sides
in the nodes close to the free surface pre-image γ sh . For the new problem we introduce the
operator notation
Λφ = ν,
(4.6)
25 / 72
Wave/moving wall interaction
where ν is the vector of the right hand sides. The difference operator Λ can be decomposed
in sub-operators for our convenience:
Λ ≡ Λ1 + Λ2 ,
and on one more level:
Λ1 ≡ Λ11 + Λ12 ,
Operators Λ α β
Λ11 φj
1 6 α, β 6 2
Λ2 ≡ Λ21 + Λ22 .
are defined below as follows.
h
♮ i
1
♮ ♭
♭
, 0 < j 1 < N1 , 0 6 j 2 < N2 ,
+
K
φ
K
φ
1
1
11 q
11 q
q1 j
q1
2
2
def
j 1 = 0 , 0 6 j 2 < N2 ,
K 1 1 1 , j 2 φ ♮q 1 0, j 2 ,
:=
2
h1
2
−
K 1 1 N1 − 1 , j 2 φ ♭q 1 N1 , j 2 ,
j 1 = N1 , 0 6 j 2 < N2 .
2
h1
(4.7)
h
♮
♭ i
1
♮
♮
♭
♭
+
K
φ
K
φ
+
K
φ
+
K
φ
,
2
2
2
2
1
2
1
2
1
2
1
2
q
q
q
q
q1
q1 j
4
0 < j 1 < N1 , 0 < j 2 < N2 ,
i
h
1
♮
♮
♭
♭
,
+
K
φ
+
K
φ
K
φ
+
K
φ
12 q2
1 2 q 2 0, j
12 q2
1 2 q 2 1, j
2
2
2 h1
j 1 = 0 , 0 < j 2 < N2 ,
i
h
1
def
−
K 1 2 φ ♮q 2 + K 1 2 φ ♭q 2 N1 −1, j 2 + K 1 2 φ ♮q 2 + K 1 2 φ ♭q 2 N1 , j 2 ,
Λ 1 2 φ j :=
2 h1
j 1 = N1 , 0 < j 2 < N2 ,
h
♭ i
♮
1
♮
♮
, 0 < j 1 < N1 , j 2 = 0 ,
K12 φq2 q1 + K12 φq2 q1
2
j 1, 0
i
1 h
♮
♮
,
j1 = 0 , j2 = 0 ,
K
φ
+
K
φ
2
2
12 q
12 q
1, 0
0, 0
h
1
h
i
♮
♮
−1
j 1 = N1 , j 2 = 0 ,
K 1 2 φ q 2 N1 −1, 0 + K 1 2 φ q 2 N1 , 0 ,
h1
(4.8)
26 / 72
G. Khakimzyanov & D. Dutykh
def
Λ 2 1 φ j :=
Λ22 φj
♭ i
♮
1h
K 2 1 φ ♮q 1 + K 2 1 φ ♭q 1 q 2 + K 2 1 φ ♮q 1 + K 2 1 φ ♭q 1 q 2 ,
j
4
0 < j 1 < N1 , 0 < j 2 < N2 ,
h
♭ i
♮
1
, j 1 = 0 , 0 < j 2 < N2 ,
K 2 1 φ ♮q 1 q 2 + K 2 1 φ ♮q 1 q 2
0, j 2
2
♭ i
♮
1h
♭
♭
, j 1 = N1 , 0 < j 2 < N2 ,
K21 φq1 q2 + K21 φq1 q2
N1 , j 2
2
i
1 h
K 2 1 φ ♮q 1 + K 2 1 φ ♭q 1 j 1 , 0 + K 2 1 φ ♮q 1 + K 2 1 φ ♭q 1 j 1 , 1 ,
2 h2
0 < j 1 < N1 , j 2 = 0 ,
h
i
1
♮
♮
j1 = 0 , j2 = 0 ,
K 2 1 φ q 1 0, 1 + K 2 1 φ q 1 0, 0 ,
h2
i
1 h
j 1 = N1 , j 2 = 0 ,
K 2 1 φ ♭q 1 N1 , 0 + K 2 1 φ ♭q 1 N1 , 1 ,
h2
(4.9)
h
♮ i
♭
1
K 2 2 φ ♮q 2 q 2 + K 2 2 φ ♭q 2 q 2 , 0 6 j 1 6 N1 ,
def
2
j
:=
2
K 2 2 j 1 , 1 ϕ ♮q 2 j 1, 0 ,
0 6 j 1 6 N1 ,
2
h2
0 < j 2 < N2 ,
j2 = 0 .
(4.10)
Above we used the notation
Kαα
Kββ
Kαα
Kββ
j 1 + 21 , j 2
j 1 , j 2 + 21
j 1 − 12 , j 2
j 1 , j 2 − 21
def
:=
def
:=
def
:=
def
:=
Kαα
Kββ
j 1, j 2
Kαα
Kββ
j 1, j 2
j 1, j 2
j 1, j 2
+
Kαα
j 1 +1, j 2
2
+ K β β j 1 , j 2 +1
2
+ K α α j 1 −1, j 2
2
+ K β β j 1 , j 2 −1
2
,
(4.11)
,
(4.12)
≡ Kαα
E
≡ Kββ
N
≡ Kαα
W
≡ Kββ
S
,
,
(4.13)
(4.14)
with α, β = 1, 2 . Moreover, all the operators vanish in the nodes of the upper boundary
γ sh , i.e.
Λαβ φj ≡ 0 ,
∀q j ∈ γ sh ,
α, β = 1, 2 .
Below we study the properties of the operator Λ to show that the problem (4.6) is wellposed.
27 / 72
Wave/moving wall interaction
4.4. Mathematical study of the finite difference problem for the velocity potential
Let us define by H the set of functions defined on the grid Q 0h and which take zero values
on γ sh , i.e.
def
H := φ : Q h0 7→ R | φ q ∈ γ h ≡ 0 .
s
j
We introduce also on space H the scalar product
def
h u, v i H :=
NX
1−1 N
2 −1
X
u j 1, j 2 v j 1 , j 2 h1 h2 +
j1 =1 j2 =1
NX
1−1
NX
2−1
u 0, j 2 v 0, j 2
j2 =1
u j 1, 0 v j 1 , 0
j1=1
h1 h2
+
2
NX
2−1
h1 h2
h1 h2
+
u N1 , j 2 v N1 , j 2
+
2
2
j =1
2
u 0, 0 v 0, 0
h1 h2
h1 h2
+ u N1 , 0 v N 1 , 0
.
4
4
This scalar product induces a natural norm
def
2
k u kH
:= h u, u i H .
def
Let us define a new operator A := −Λ and study its properties. The operator Λ was
introduced in (4.6) and rigorously defined earlier.
For any fixed value of index j 2 , a function u ∈ H can be considered as a function being
defined on a 1D grid:
def
Q 1h := q j11 ∈ [ 0, 1 ] | j 1 = 0, 1, . . . , N1 .
The class of all grid functions defined on Q 1h will be denoted by H 1 . For two arbitrary
functions u, v ∈ H we define the scalar product on the layer j 2 as
def
h u, v i H 1, j 2 :=
NX
1 −1
u j 1, j 2 v j 1 , j 2 h 1 + u 0, j 2 v 0, j 2
j1=1
h1
h1
+ u N 1, j 2 v N 1, j 2
.
2
2
(4.15)
Then, it is not difficult to see that
h u, v i H ≡
NX
2−1
j2 =1
h u, v i H 1, j 2 h 2 + h u, v i H 1 , 0
h2
.
2
(4.16)
However, we can proceed in a different way by considering instead ‘vertical’ slices. Let
H 2 be the set of 1D functions defined on the grid
def
Q 2h := q j22 ∈ [ 0, 1 ] | j 2 = 0, 1, . . . , N2 .
and taking zero value for j 2 = N 2 , i.e.
def
H 2 := φ : Q h2 7→ R | φ
j2 =N2
≡ 0 .
28 / 72
G. Khakimzyanov & D. Dutykh
Similarly, in H we can introduce a scalar product on the layer j 1 as follows:
def
h u, v i H 2, j 1 :=
NX
2 −1
u j 1, j 2 v j 1 , j 2 h 2 + u j 1, 0 v j 1 , 0
j2 =1
h2
,
2
(4.17)
and the scalar product in H can be similarly expressed as
h u, v i H ≡
NX
1−1
j1=1
h u, v i H 2 , j 1 h 1 + h u, v i H 2, 0
h1
h1
+ h u, v i H 2, N 1
.
2
2
(4.18)
It is not difficult to show that formulas (4.16) and (4.18) provide the same value for h u, v i H
and, thus, they are nothing else than two different ways to express the scalar product in H .
Intuitively, we can understand it in the following way. Imagine that one wants to compute
the sum of all elements of a matrix. To do it, one can move along the lines and then
along the columns, which corresponds to (4.18), or vice versa (4.16). Since the addition is
commutative, the resulting sum will be the same.
4.4.1 Some discrete identities
N
N
N
Let us remind that for any discrete functions α j j = 0 , u j j = 0 and v j j = 0 defined
N
on a 1D grid Q 1h = q j ∈ [ 0, 1 ] j = 0 , the first discrete Green identity can be shown:
E
D
i
D
♭
♭
♭ ♮
(4.19)
α u q q , v ≡ − α u q , v q + α u q♭ v N − α 1 u q♮ v 0 ,
and also
D
α u q♮
♭
,v
q
E
h
E
≡ − α u ♮q , v q♮ + α N −1 u ♭q v
N
−
♮
α uq v
0
,
(4.20)
where for the sake of simplicity we introduced the following notations
D
α, β
E
def
:= h
N
−1
X
j=1
αβ
def
j
αj βj ,
:= α j β j ,
D
α, β
i
= h
αβγ
N
X
j=1
def
j
:= α j β j γ j ,
αj βj ,
h
α, β
E
Moreover, the following summation by parts formulas hold as well
u ♮q , v ≡ − u, v q♭ + u v N − u 1 v 0 ,
u ♭q , v ≡ − u, v q♮ + u N −1 v N − u v 0 .
= h
N
−1
X
αj βj .
j=0
(4.21)
(4.22)
All these properties mentioned above show that the discretization proposed in the present
study belongs to the class of operational finite difference schemes [70], which became later
known as mimetic methods [54].
29 / 72
Wave/moving wall interaction
4.5. The main result
Theorem 1. The operator A ≡ −Λ is self-adjoint in the space H .
Proof. For two arbitrary grid functions u, v ∈ H we consider their scalar product on a
layer 0 6 j 2 < N2 :
Λ 1 1 u, v
= h1
H 1, j 2
N
1 −1
X
j1 =1
v Λ11 u
j 1, j 2
+
v Λ11 u
h1
+ v Λ11 u
2
0, j 2
N 1, j 2
h1
.
2
Taking into account definition (4.7) of the operator Λ 1 1 and Green’s identity (4.19) we
obtain
Λ 1 1 u, v
H 1, j 2
N1 − 1
1 X
= −
K 1 1 u q♮ 1 v q♮ 1
2 j =0
j 1, j 2
h1 +
1
1
−
K 1 1 u q♮ 1 v
2
−
0, j 2
N1
1 X
−
K 1 1 u ♭q 1 v q♭ 1
2 j =1
1
K 1 1 1, j 2 u q♮ 1 v
2
H 1, j 2
h +
j 1, j 2 1
1
0, j 2
1
K 1 1 u q♭ 1 v
2
2
h1
K 1 1 1/2, j 2 u ♮q 1 v 0, j 2
h1
2
2
K 1 1 N1 − 1/2, j 2 u q♭ 1 v
−
h1
+
or simply
Λ 1 1 u, v
1
K 1 1 N1 −1, j 2 u q♭ 1 v
2
N1 − 1
1 X
K 1 1 u q♮ 1 v q♮ 1
= −
2 j =0
j 1, j 2
h1
1
N1
1 X
−
K 1 1 u q♭ 1 v q♭ 1
2 j =1
N1 , j 2
N1 , j 2
N1 , j 2
j 1, j 2
h1
,
2
h1 .
1
(4.23)
By substituting the last expression into formula (4.16), we obtain the equality
h Λ 1 1 u, v i H
N2 − 1 NX
N1
1−1
X
1 X
♮
♮
♭
♭
K11 uq1 vq1 +
K11 uq1 vq1
h1 h2
= −
2 j =1 j =0
j
,
j
1
2
j =1
2
1
−
2
1
NX
1−1
j1 =0
1
K 1 1 u q♮ 1 v q♮ 1
j 1, 0
N1
h1 h2
1 X
K 1 1 u q♭ 1 v q♭ 1
−
2
2 j =1
1
j 1, j 2
h1 h2
.
2
The last identity is absolutely symmetric with respect to the functions u and v , i.e.
h Λ 1 1 u, v i H ≡ h u, Λ 1 1 v i H .
Consequently, the operator Λ 1 1 is self-adjoint in space H, i.e.
Λ 1∗ 1 ≡ Λ 1 1 .
Using the summation by parts formulas (4.21), (4.22) along with the definition (4.8) of
the operator Λ 1 2 in interior and boundary nodes, we obtain the following expression for
30 / 72
G. Khakimzyanov & D. Dutykh
the scalar product
N2 − 1 X
N1
NX
1−1
♭
♮
1 X
♮
♮
♭
♭
= −
K12 uq2 + uq2 vq1 +
K12 uq2 + uq2 vq1
h1 h2
4 j =1 j =1
j
,
j
1
2
j =0
h Λ 1 2 u, v i H
1
2
−
2
N1
X
1
1
K 1 2 u q♮ 2 v q♭ 1
j 1, 0
j1 =1
It is obvious that
1
h1 h2
−
2
2
NX
1 −1
j1=0
K 1 2 u ♮q 2 v q♮ 1
j 1, 0
h1 h2
. (4.24)
2
h Λ 1 2 u, v i H 6= h u, Λ 1 2 v i H .
Thus, the operator Λ 1 2 is not self-adjoint. Nevertheless, we continue the proof without any
deception. Consider right now the operator Λ 2 1 defined in (4.9). Using similar techniques
we obtain the following expression for the scalar product of Λ 2 1 u with v :
NX
N1 − 1 X
N2
2−1
♮
♭
1 X
♮
♮
♭
♭
h Λ 2 1 u, v i H = −
h1 h2
K21 uq1 + uq1 vq2
K21 uq2 + uq1 vq2 +
4 j =1 j =1
j 1, j 2
j =0
1
1
2
−
−
N2
X
2
2
K 2 1 u q♮ 1 v q♭ 2
0, j 2
j2=1
N2
1 X
K 2 1 u q♭ 1 v q♭ 2
2 j =1
N1 , j 2
2
h1 h2
1
−
2
2
NX
2−1
j2 =0
K 2 1 u q♮ 1 v q♮ 2
N2 − 1
1 X
h1 h2
−
K 2 1 u q♭ 1 v q♮ 2
2
2 j =0
0, j 2
N1 , j 2
2
h1 h2
2
h1 h2
. (4.25)
2
From the last formula it can be clearly seen that
h Λ 2 1 u, v i H 6= h u, Λ 2 1 v i H .
Thus, the operator Λ 2 1 as well as Λ 1 2 is not self-adjoint. However, we can study the sum
Λ 2 1 + Λ 1 2 . By regrouping the terms in (4.24) and (4.25) we obtain
h Λ 1 2 u + Λ 2 1 u, v i H =
N1 − 1 NX
2 − 1h
i
h1 h2 X
−
K 1 2 u q♮ 1 + u q♭ 1 · v q♮ 2 + v q♭ 2 + K 1 2 v q♮ 1 + v q♭ 1 · u ♮q 2 + u q♭ 2
4
j 1, j 2
j =1 j =1
1
2
+
NX
2 − 1h
K 1 2 u q♮ 1
v q♮ 2
+
j2=1
+
NX
2 − 1h
K 1 2 u ♭q 1 v q♮ 2 + v q♭ 2
j2=1
+
NX
1 − 1h
+
K 1 2 u q♮ 1
v q♮ 2
+
+
+ K 1 2 v q♭ 1 u q♮ 2 + u ♭q 2
K 1 2 v q♮ 2 u ♮q 1 + u q♭ 1
j1=1
h
v q♭ 2
K 1 2 u ♮q 2
v q♮ 1
i
0, 0
+
h
K 1 2 v q♮ 1
u ♮q 2
+
u q♭ 2
i
0, j 2
i
N1 , j 2
+ K 1 2 u ♮q 2 v q♮ 1 + v q♭ 1
i
K 1 2 u q♮ 2
v q♭ 1
K 1 2 u q♭ 1
v q♮ 2
+
j 1, 0
i
N1 , 0
. (4.26)
31 / 72
Wave/moving wall interaction
Above we used also the fact that K 1 2 ≡ K 2 1 . From the last formula (4.26) we readily
obtain that
h Λ 1 2 u + Λ 2 1 u, v i H ≡ h u, Λ 1 2 v + Λ 2 1 v i H .
Thus, the operator Λ 2 1 + Λ 1 2 is self-adjoint in the space H , i.e.
∗
Λ12 + Λ21
≡ Λ12 + Λ21 ,
even if constitutive operators Λ 1 2 and Λ 2 1 are not self-adjoint if taken separately.
Finally, consider the operator Λ 2 2 , which was defined in (4.10). We employ first
the
discrete Green identities (4.19), (4.20) to evaluate scalar products on layers j 1 ∈
0, 1, . . . , N1 :
Λ 2 2 u, v
N2 − 1
1 X
= −
K 2 2 u ♮q 2 v q♮ 2
2 j =0
H 2, j 1
j 1, j 2
h2
2
N2
1 X
−
K 2 2 u q♭ 2 v q♭ 2
2 j =1
j 1, j 2
h2 .
2
Using this partial result we can now compute the scalar product in the H space as well:
h Λ 2 2 u, v i H
−
−
1
2
N1 − 1 NX
N2
2−1
X
1 X
♮
♮
♭
♭
h1 h2
K22 uq2 vq2 +
K22 uq2 vq2
= −
2 j =1 j =0
j 1, j 2
j =1
1
2
2
N2 − 1
1 X
K 2 2 u q♮ 2 v q♮ 2
2 j =0
0, j 2
2
NX
2−1
j2 =0
K 2 2 u q♮ 2 v q♮ 2
N1 , j 2
N2
h1 h2
1 X
K 2 2 u q♭ 2 v q♭ 2
−
2
2 j =1
0, j 2
2
h1 h2
1
−
2
2
N2
X
K 2 2 u q♭ 2 v q♭ 2
j2=1
N1 , j 2
h1 h2
2
h1 h2
. (4.27)
2
By changing symmetrically grid functions u and v , we obtain
h Λ 2 2 u, v i H ≡ h u, Λ 2 2 v i H .
Consequently, the operator Λ 2 2 is self-adjoint in the space H, i.e.
Λ 2∗ 2 ≡ Λ 2 2 .
Since Λ ≡ Λ 1 1 + Λ 1 2 + Λ 2 1 + Λ 2 2 and each term is an self-adjoint operator, the
operator Λ is self-adjoint as well. Hence, the same holds for A ≡ −Λ .
Below we are going to show also that the operator A is positive definite. The proof
will be based on the uniform ellipticity property of equation (3.4), which was established
in Lemma 1. Another key ingredient consists in the properties of the finite difference
approximation to the following mixed Boundary Value Problem (BVP) for the Laplace
equation posed on Q 0 :
∆ϕ = 0,
q = q 1, q 2 ∈ Q 0 ,
(4.28)
32 / 72
G. Khakimzyanov & D. Dutykh
with boundary conditions
∂ϕ
∂q 1
q1 =1
ϕ
q2
=1
∂ϕ
2
=
μ
q
,
ℓ
∂q 1 q 1 = 0
∂ϕ
= 0.
∂q 1 q 2 = 0
= 0,
= μ̃ q 1 ,
(4.29)
(4.30)
This problem is a particular case of the BVP considered earlier for elliptic equation (3.4)
with the following coefficients:
K11 = K22 ≡ 1 ,
K12 = K21 ≡ 0 .
In this case the finite difference operator Λ admits a much simpler form since finite difference operators corresponding to mixed derivatives vanish, i.e.
Λ12 = Λ21 ≡ 0 .
The finite difference scheme for the BVP (4.28) – (4.30) can be written as
∆h ϕ = ν ,
def
(1)
∆ h := ∆ h
(2)
+ ∆h ,
(1, 2)
with operators ∆ h
defined as
♭, ♮
ϕ q 1 q 1 j 1, j 2 ,
2 ♮
ϕq1 0j2 ,
def
(1)
h1
∆ h ϕ j :=
2 ♭
ϕ q 1 N1 j 2 ,
−
h1
0,
(2)
∆h ϕj
♭, ♮
ϕ q 2 q 2 j 1, j 2 ,
def
2 ♮
:=
ϕ q 2 j 1, 0 ,
h2
0,
0 < j 1 < N1 − 1 ,
j1 = 0 ,
0 6 j 2 < N2 ,
j 1 = N1 ,
0 6 j 2 < N2 ,
0 6 j 1 6 N1 ,
def
0 6 j 2 < N2 ,
j 2 = N2 ,
0 6 j 1 6 N1 ,
0 < j 2 < N2 ,
0 6 j 1 6 N1 ,
j2 = 0 ,
0 6 j 1 6 N1 ,
j 2 = N2 .
(1)
def
(2)
def
Below we introduce also the operators B 1 := −∆ h , B 2 := −∆ h and B := B 1 + B 2 ,
def
i.e. B := − ∆ h . By applying Theorem 1 we conclude that the operator B : H 7→ H
is self-adjoint in the space H .
It is not difficult to see that the family of grid functions
(k, l) def
ψ j 1, j 2 := α k ·cos π k q j11 ·cos π (l + 12 ) q j22 , k = 0, 1, . . . , N1 , l = 0, 1, . . . N2 − 1
(4.31)
satisfies the following identities:
B · ψ (k, l) ≡ λ (k, l) · ψ (k, l) .
(4.32)
33 / 72
Wave/moving wall interaction
Henceforth, the functions {ψ (k, l) } k, l are eigenfunctions∗ of the operator B and numbers
{λ (k, l) } k, l ∈ R are its eigenvalues:
1
4
4
(k, l) def
2 π k h1
2 (l + 2 ) π h 2
λ
+ 2 sin
.
:= 2 sin
h1
2
h2
2
If we choose the normalization factor α k as
(√
2,
k = 0 ∧ N1 ,
αk =
2,
k = 1, 2, . . . , N1 − 1 ,
then the eigenfunctions {ψ (k, l) } k, l of the operator B form an orthonormal basis in H , i.e.
(
1,
k = r ∧ l = s,
h ψ (k, l) , ψ (r, s) i H = δ k r · δ l s =
0,
k =
6 r ∨ l 6= s .
In other words, an arbitrary function u ∈ H can be represented as a sum of a finite
Fourier series:
N1 NX
2−1
X
u =
c k l ψ (k, l) ,
(4.33)
k=0
l=0
where the numbers c k l ≡ h u, ψ (k, l) i H k, l are called the Fourier coefficients. The
Parseval theorem holds in the finite dimensional setting as well:
2
k u kH
≡ h u, u i H =
N1 NX
2−1
X
k=0
c k2 l .
(4.34)
l=0
Since all eigenvalues of the operator B are positive, the following Lemma holds:
Lemma 2. The finite difference operator B is positive definite in space H and for any
function u ∈ H we have the following estimations
2
2
λ min k u k H
6 h B u, u i H 6 λ max k u k H
,
(4.35)
with
λ min
4
2 π h2
,
:= 2 sin
h2
4
def
λ max
4
4
2 π h2
:= 2 + 2 cos
.
h1
h2
4
def
(4.36)
Proof. Let us take an arbitrary function u ∈ H and we expand it (4.33) on the basis of
orthonormal eigenfunctions of the operator B . Then we have
X
N1 NX
N1 NX
2−1
2−1
X
(r, s)
(k, l) (k, l)
,
h B u, u i H =
ckl λ
ψ
,
crs ψ
k=0
∗Please,
l=0
r=0 s=0
H
notice that each eigenfunction ψ (k, l) is a grid function taking in each node j =
(k, l)
value ψ j 1 , j 2 given in equation (4.31).
j 1 , j 2 the
34 / 72
G. Khakimzyanov & D. Dutykh
where we used the fact that ψ (k, l) is an eigenfunction of B (4.32). Then, by taking into
account the orthonormality property we can greatly simplify the scalar product:
h B u, u i H =
N1 NX
2−1
X
k=0
c 2k l λ (k, l) .
l=0
Obviously, among a finite number of eigenvalues {λ (k, l) } k, l we can always choose the
minimal and maximal ones. Thus,
λ min 6 λ (k, l) 6 λ max .
∀ (k, l) :
Taking into account the last estimation along with the Parseval identity (4.34), we obtain
immediately the estimations (4.35).
Thanks to the property of self-adjointness and positive definiteness of the operator B , we
def
can define another scalar product h u, v i B := h B u, v i H , which implies the corresponding
energy norm:
def p
h u, u i B ,
k u k B :=
generated by the operator B . Below we use an equivalent representation of the energy
norm, which is more suitable for our purposes:
2
k u kB
≡ h B 1 u, u i H + h B 2 u, u i H .
Lemma 3. For any function u ∈ H the following identities hold:
N1 NX
2 − 1n
X
2o
1 n ♭ 2o
♭
h B 1 u, u i H =
uq1
+
uq1
h1 h2 ,
j 1, j 2
2
j 1, 0
j =1 j =1
1
h B 2 u, u i H =
N2 NX
1 − 1n
X
j2=1
j1 =1
(4.37)
2
2
u q♭ 2
o
j 1, j 2
1 n ♭ 2o
1 n ♭ 2o
h1 h2 .
+
uq2
uq2
+
2
2
0, j 2
N1 , j 2
(4.38)
Proof. Taking into account equation (4.15) and discrete identities shown in Section 4.4.1,
we obtain the equality
B 1 u, u
H 1, j 2
=
N1 n
X
u q♭ 1
j1=1
2o
j 1, j 2
2o
j 1, j 2
h1 .
Then, from the last result and relation (4.16) follows immediately the requested identity
(4.37). Similarly, the discrete identities along with definition (4.17) yield
B 2 u, u
H 2, j 1
=
N2 n
X
j2=1
u q♭ 2
h2 .
Then, using relation (4.18) we obtain the second requested identity (4.38).
Now we can come back to the original finite difference operator A and prove the following
35 / 72
Wave/moving wall interaction
Theorem 2. The operator A ≡ − Λ is positive definite in the space H , and we have the
following estimations:
2
2
c 1 λ min k u k H
6 h A u, u i H 6 c 2 λ max k u k H
,
(4.39)
where the constants c 1, 2 were defined in (3.22), (3.23) and λ min, max — in (4.36).
Proof. As in the first step of the proof we take v ≡ u ∈ H and substitute it into
equations (4.23). After regrouping interior and boundary nodes, we obtain:
h Λ 1 1 u, u i H
h1 h2
= −
4
NX
2 − 1h
NX
1−1 N
2 − 1h
X
j1 =1 j2 =1
2
u q♮ 1
2 K11
j2 =1
NX
1 − 1h
K11
j1=1
2
u q♮ 1
2 K 1 1 u ♮q 1
2
u q♭ 1
+ K11
i
+
0, j 2
i
NX
2 − 1h
2
+ 2 K 1 1 u q♭ 1
2 K 1 1 u ♭q 1
j2=1
+
j 1, 0
h
i
2
u q♮ 1
K11
2 i
0, 0
2 i
j 1, j 2
N1 , j 2
+
h
+
+
2
u q♭ 1
K11
i
N1 , 0
.
We perform the same operation with equation (4.27) as well:
h Λ 2 2 u, u i H
NX
2 − 1h
K22
j2=1
h1 h2
= −
4
NX
1−1 N
2 − 1h
X
2
u q♮ 2
+ K22
2 K22
u q♮ 2
NX
1 − 1h
j1 =1
NX
1 − 1h
j1=1
2 i
2 K 2 2 u ♮q 2
j1 =1 j2 =1
2
u q♭ 2
j 1, 0
2
2 K 2 2 u ♭q 2
+
i
i
0, j 2
h
NX
2 − 1h
+
+ 2 K 2 2 u q♭ 2
K 2 2 u q♮ 2
j2=1
2
K 2 2 u ♮q 2
j 1 , N2
2
+
h
i
0, 0
+
2
K 2 2 u ♭q 2
i
h
2
2 i
+ K 2 2 u q♭ 2
2
K 2 2 u ♮q 2
0, N2
j 1, j 2
+
h
i
K22
+
2 i
N1 , 0
N1 , j 2
+
+
2
u q♭ 2
i
N1 , N2
.
36 / 72
G. Khakimzyanov & D. Dutykh
Similarly, we set v ≡ u in equation (4.26) and using the last two formulas we have
h A u, u i H =
NX
1 −1 N
2 − 1h
X
i
h1 h2
♮
♮
♮
♮
♭
♭
♭
♭
Q u q 1, u q 2 + Q u q 1, u q 2 + Q u q 1, u q 2 + Q u q 1, u q 2
4
j 1, j 2
j =1 j =1
1
+
NX
2 − 1h
j2 =1
+
NX
1 − 1h
j1 =1
2
i
Q u q♮ 1 , u q♮ 2 + Q u q♮ 1 , u q♭ 2
i
Q u q♮ 1 , u ♮q 2 + Q u q♭ 1 , u q♮ 2
+
NX
1 − 1h
2 Q 0,
u ♭q 2
j1=1
i
0, j 2
NX
2 − 1h
j2=1
j 1, 0
h
+
j 1 , N2
+
i
Q u q♭ 1 , u q♮ 2 + Q u q♭ 1 , u q♭ 2
h
i
+ Q u q♮ 1 , u q♮ 2
u ♭q 2
Q 0,
i
0, 0
+
0, N2
h
N1 , j 2
h
i
+ Q u q♭ 1 , u q♮ 2
Q 0,
u q♭ 2
i
N1 , N 2
N1 , 0
.
where we used the quadratic form Q(·, ·) defined earlier in (3.21). By taking into account
the uniform ellipticity property (3.20), we can estimate the quantity h A u, u i H from below:
h A u, u i H
NX
2 − 1h
2
j2 =1
NX
1 − 1h
j1 =1
h1 h2
·
> c1
4
2
u q♮ 1
2
u q♮ 1
+
+
2
u q♮ 2
2
u ♭q 1
+ 2
NX
1 −1 N
2 −1
X
h
2
j1=1 j2=1
2
u ♭q 2
+
+2
2
u q♮ 2
NX
1−1
u q♭ 2
j1 =1
i
i
0, j 2
j 1, 0
+
2
j 1 , N2
u ♮q 1
+
2
NX
2 − 1h
2 u q♭ 1
j2 =1
h
+
2
u q♮ 1
+
2
u q♭ 2 0, N2
2
+ u q♭ 2
2
+ u q♭ 2
2
+ u q♮ 2
2
+ u q♮ 2
+ u q♭ 1
2
u ♮q 2
+
i
0, 0
+
h
2
u ♭q 2 N1 , N2
u q♭ 1
2
2 i
2 i
+
+
j 1, j 2
+
N1 , j 2
2
u q♮ 2
i
N1 , 0
def
=: B low . (4.40)
Then, in the expression above for B low we replace∗ all right derivatives u q♮ α by their left
counterparts u q♭ α :
B low = c 1
X
N1 h NX
2 −1
j1=1
j2=1
u q♭ 1
2
N2 h NX
1 −1
X
j2 =1
∗We
j1=1
j 1, j 2
+
2 i
1
u q♭ 1 j 1 , 0 +
2
2
u q♭ 2 j 1, j 2
i
2
2
1
1
♭
♭
u q 2 0, j 2 +
u q 2 N1 , j 2
+
2
2
h1 h2 .
can do it since the right finite difference can be considered as the left one for the subsequent value
of the index. Thus, this operation can be seen as a change of index by one in all summations.
37 / 72
Wave/moving wall interaction
Now we can make use of Lemma 3, where we showed identities (4.37) and (4.38), which
yield:
n
o
B low ≡ c 1 h B 1 u, u i H + h B 2 u, u i H .
In other words, we just showed the following estimation from below:
h A u, u i H > c 1 h B u, u i H .
Finally, by applying Lemma 2 we obtain that
2
h B u, u i H > λ min k u k H
.
The two last inequalities correspond precisely to the required lower bound in (4.39). By
departing again from equation (4.40) and using the same techniques we can show the upper
bound in (4.39). This completes the proof of this Theorem.
Remark 6. During the proof of Theorem 2 we established also the following estimates:
c 1 h B u, u i H 6 h A u, u i H 6 c 2 h B u, u i H ,
∀u ∈ H ,
or equivalently
c1 B 6 A 6 c2 B .
Such operators A and B satisfying the inequalities above are usually referred to as energetically equivalent [69].
4.5.1 Final remarks
From the positive definiteness property of the operator A follows the unique solvability
property of the finite difference problem (4.6). To solve this problem numerically one can
use almost any iterative technique. It is natural to use the conjugate gradient method
whose convergence is guaranteed if the operator is positive definite and self-adjoint [67].
The speed of convergence depends essentially on the conditioning number κ of the difference
operator A , which can be estimated as
4
4
2 π h2
+
cos
c 2 h 21
c 2 λ max
h 22
4
·
=
·
.
κ 6
4
c 1 λ min
c1
2 π h2
sin
h 22
4
|
{z
}
(⋆)
For larger values of κ the convergence will be slower. As we showed earlier in Section 3.1, the
c2
is completely determined by the properties of the geometric mapping (3.1).
first factor
c1
This ratio becomes large for the meshes with highly elongated or highly distorted elements.
The second factor (⋆) depends on the number of nodes N1, 2 chosen in the directions O q 1, 2
correspondingly and (⋆) becomes larger when we refine the grid (i.e. when N1, 2 increase).
G. Khakimzyanov & D. Dutykh
38 / 72
5. Numerical algorithm
In this Section we describe briefly how the numerical code is organized and we provide
the details on the treatment of kinematic and dynamic boundary conditions on the free
surface and on the lateral fixed and moving walls.
Assume that the position of the wall x = s n on the time layer t = t n is known along
with all other fields. The grid Ω nh with nodes x nj , ordered using a multi-index j = j 1 , j 2
in
On this grid we know the values of grid functions
n space isconstructed.
nthe physical
n
ξ j 1 j 1 , η j 1 j 1 and ϕ j j . The time marching requires to compute these quantities
on the following time layer t = t n+1 . It is done in several stages:
(1) First, we compute the values of the velocity potential ϕ n+1
j 1 , N2 j 1 at the free surface
node q j 1, N2 .
(2) These values are used as the
data in the linear problem (4.6) to deter Dirichlet
n+1
mine the velocity potential ϕ j 1 , j 2 j 1 , j 2 in all other nodes q j with 0 6 j 1 6 N1
and 0 6 j 2 < N2 .
n+1
(3) Then, we find new positions ξ jn+1
,
η j 1 j 1 of the free surface by using the fi1
j1
nite difference approximation of the kinematic boundary conditions (3.5) and (3.12).
(4) The new position x = s n+1 of the moving wall is found from the discretized version
of the boundary condition (3.8).
(5) Finally, the new grid Ω̄ n+1
is constructed on the following time layer t = t n+1 .
h
Every step above involves grid functions on various layers and to achieve a better understanding of all the stages of our algorithm, below we provide a detailed description of every
item.
5.1. Approximation of the dynamic boundary condition (1)
The finite difference approximation to the dynamic boundary condition (3.6) takes the
following form:
ϕ jn+1 − ϕ nj
(5.1)
− u nj · x t, j + v jn · y t, j + 12 | u nj | 2 + η jn1 = 0 ,
τn
where j = j 1 , j 2 is a multi-index with j 1 = 1, 2, . . . , N1 − 1 and j 2 = N2 . In order
to compute the speeds of grid nodes we employ finite differences applied to two last grids
Ω̄ nh and Ω̄ n−1
:
h
x nj − x jn−1
x t, j =
,
(5.2)
τ n−1
2
2
where τ n−1 ≡ t n − t n−1 is the local time step, | u nj | 2 ≡ u nj + v jn . In formula (5.2)
we use only the first order approximation in time. This is done for the sake of algorithm
memory efficiency. Otherwise, to achieve at least the second order accuracy in τ , we would
39 / 72
Wave/moving wall interaction
have to keep in memory the grids for three time layers. In the current implementation we
keep track of the preceding grid only.
The Cartesian components u jn and v jn of the velocity vector u nj are computed after
approximating formulas (3.15) and (3.16):
n
n
ϕq 1 · y q 2 − ϕq 2 · y q 1
−ϕq 1 · xq 2 + ϕq 2 · xq 1
n
n
uj =
,
vj =
.
(5.3)
J
J
j
j
Above one has to approximate also the partial derivatives ϕ q 1 , x q 1 over the independent
variable q 1 along the upper side of the square Q 0 . These derivatives are computed with
standard central finite differences up to the second order accuracy. The derivatives ϕ q 2 ,
x q 2 are computed with one-sided finite differences of the second order as well (in order to
have the uniform second order accuracy in space).
We use equation (5.1) in order to compute the updated values of the velocity potential
only in interior nodes of the grid γ sh . For two nodes located in the upper corners q 0, N2 =
(0, 1) and q N1 , N2 = (1, 1) we use modified approximation formulas which take into
account the lateral walls γ ℓ, r impermeability. In the computational domain the left wall
has the fixed coordinate q 1 = 0 . The intersection point of the free surface with the left
boundary is a triple point and it has the coordinate x 0, N2 . This point is permanently
moving up and down the left wall.
dx
In the coordinate system O q 1 q 2 the Cartesian components u =
become the
dt
dq
with
so-called contravariant components of the velocity v =
dt
vα =
∂q α
∂q α
∂q α
dq α
=
+ u
+ v
,
dt
∂t
∂x
∂y
α = 1, 2 .
(5.4)
Consequently, if a fluid particle moves such that its coordinate q α does not change, then
the corresponding contravariant velocity component v α has to vanish. For instance, the
fluid particles, which constitute the free surface, have the coordinate q 2 = 1 , consequently
v2
q2=1
= 0.
Similarly, for fluid particles sliding along the left vertical wall we always have q 1 = 0 ,
which yields
v1
The point x 0, N2 ∈ γ sh ∩ γ ℓh , thus
v 1 q 0, N2
q1=0
= 0.
= v 2 q 0, N2
≡ 0.
(5.5)
From formula (5.4) and using also equations (3.18), (3.19) we obtain the following expressions for contravariant components of the velocity:
− u − xt · yq1 + v − yt · xq1
u − xt · yq2 − v − yt · xq2
2
1
,
v =
.
v =
J
J
40 / 72
G. Khakimzyanov & D. Dutykh
From these expressions and using boundary condition∗ (5.5) for the triple point x 0, N2 we
obtain that
xt q
= u,
yt q
= v.
0, N2
2
2
0, N2
2
Henceforth, | u | ≡ u + v = u x t + v y t and the free surface dynamic boundary
condition (3.6) in the corner q 0, N2 becomes:
ϕ t − 12 u · x t + v · y t + η = 0 ,
x = x 0, N2 .
The difference equation (5.1) correspondingly takes the form:
ϕ jn+1 − ϕ nj
j = j 1 , j 2 ≡ (0, N2 ) . (5.6)
− 21 u nj ·x t, j + v jn ·y t, j + η jn1 = 0 ,
τn
In a similar way one can derive the finite difference discretization of the boundary condition
in the right triple point. We give only the final result. It coincides with equation (5.6)
provided that we make a substitution j 1 = N1 .
5.2. Computation of the velocity potential (2)
On the second stage of the numerical algorithm wesolve the finite difference problem (4.6)
in order to find the values of the velocity potential ϕ jn+1 j in the nodes q j ∈ Q̄ 0h \ γ sh .
In other words, in the nodes j = j 1 , j 2 ) with 0 6 j 1 6 N1 and 0 6 j 2 < N2 .
In order to invert the linear system (4.6) we tested several iterative methods and no
method clearly outperformed the others. In the final implementation of the code we used
the Successive Over Relaxation (SOR) method [78] that we remind briefly here. The choice
of the SOR method can be explained by two main reasons:
(1) We would like to have a sufficiently simple to implement and sufficiently efficient
method to find approximate solutions to relatively large linear systems
(2) The method does not have to rely heavily on the matrix structure (such as Thomas’s
algorithm, for example). For instance, if we include obstacles in the fluid domain,
its topology might change, which implies some drastic changes in the pattern of
non-zero elements of the matrix.
Thus, to our opinion SOR method represents the best trade-off among the efficiency, generality and ease of implementation.
We use a nine-point stencil in the finite difference scheme. Thus, equation j has the
following general form:
8
X
k=0
A k, j ϕ k, j = B j ,
q j ∈ Q̄ 0h \ γ sh .
(5.7)
The last condensed
form is obtained from (4.6) by regrouping all the terms in front of
n+1
. The corresponding coefficient is denoted by A k, j . In equation (5.7)
unknowns ϕ j
j
∗We
implicitly use also the fact that the transformation (3.1) is non-degenerate.
41 / 72
Wave/moving wall interaction
we used a local numeration of indices depicted in Figure 4(a). The right hand side is given
by
q j ∈ γ ℓh ,
μ 0, j 2 h 2 ,
def
B j :=
1
2
(5.8)
q j ≡ q 0, 0 ,
μ 0, 0 h 2 ,
0,
otherwise .
Then, one step of the SOR method reads:
X
X
1
def
(m+1)
(m)
ϕ̃ 0, j := −
A k, j ϕ k, j
+
A k, j ϕ k, j − B 0 ,
A 0, j
k ∈ {1, 2, 5, 6}
k ∈ {3, 4, 7, 8}
(m+1) def
ϕ 0, j
(m)
:= θ ϕ̃ 0, j + (1 − θ) ϕ 0, j ,
where (m) is the iteration number and θ is the relaxation parameter. The value of this
parameter weakly influences the iterative process speed of convergence. The question of
finding the optimal value of the relaxation parameter θ is the key to the efficiency of the
SOR method. This question can be studied theoretically for the Dirichlet problem of
the Poisson equation in a square domain [ 0, ℓ ] 2 with uniform grids. We know also that
the optimal value θ ⋆ always belongs to the interval:
θ ⋆ ∈ (1, 2) .
For the Poisson equation on a square domain with the uniform discretization h 1 = h 2 ≡
h , one can show
2
θ⋆ =
π h .
1 + sin
ℓ
1
In the last formula for ℓ = 1 and h = 30
we obtain the value θ ⋆ ≈ 1.81 . In our discrete
problem for the velocity potential ϕ j we cannot determine theoretically the optimal value
θ ⋆ . This value was determined experimentally for each problem under consideration. In
numerical computations below we always took θ ⋆ ∈ [ 1.85, 1.95 ] , depending on the
discretization parameters h 1, 2 as well.
In boundary nodes the stencil contains six points and in corner points only four. However,
it does not change anything for the SOR scheme since extra coefficients A k, j can be set to
zero for our convenience. The expressions of coefficients A k, j are given in Table 1. In this
Table we use the following notation:
def h 2
def h 1
def h 2
def h 1
ρ 1 :=
ρ 2 :=
ρ 3 :=
ρ 4 :=
K11 W ,
K22 S ,
K11 E ,
K22 N ,
h1
h2
h1
h2
def
K12
1
def
K12
3
ρ 5 :=
ρ 7 :=
def
σ α :=
K12
0
− K12
4
α
,
+ K12
4
+ K12
4
def
K12
2
def
K12
4
2
,
ρ 6 := −
4
,
ρ 8 := −
α = 1, 2 ,
def
σ β :=
K12
β
+ K12
4
+ K12
4
− K12
4
3
,
1
,
0
,
β = 3, 4 .
42 / 72
G. Khakimzyanov & D. Dutykh
Node type
qj ∈
qj ∈
qj ∈
qj ∈
q 0, 0
q N1 , 0
Q h0
γ ℓh
γ bh
γ rh
A 1, j
ρ1
0
1
ρ − σ1
2 1
ρ1
0
1
ρ − σ1
2 1
A 2, j
A 3, j
ρ2
ρ2 − σ2
0
1
ρ + σ2
2 2
0
0
ρ3
σ3
1
ρ − σ3
2 3
0
1
ρ − σ3
2 3
0
1
2
A 4, j
ρ4
ρ4 −
ρ4
1
ρ +
2 4
1
ρ −
2 4
1
ρ +
2 4
1
2
A 5, j A 6, j A 7, j A 8, j
σ4
σ4
σ4
σ4
ρ5
0
0
ρ5
0
0
ρ6
ρ6
0
0
0
0
ρ7
ρ7
ρ7
0
ρ7
0
ρ8
0
ρ8
ρ8
0
ρ8
Table 1. Matrix elements of the difference equation (5.7).
In the definition of quantities ρ k we used the notation introduced in equations (4.11) –
(4.14). Finally, the coefficient A 0, j is determined as
def
A 0, j := −
4
X
A k, j .
k=1
This concludes the description of our velocity potential solver.
5.3. Free surface motion (3)
The free surface position on the following time layer t = t n+1 is found by integrating
in time kinematic conditions (2.7), (2.8). Here we consider the simplest approximation of
these equations using an explicit upwind scheme:
η n+1
− η nj1
j1
τn
v j1, n + | v j1, n | ♭, n
v j1, n − | v j1, n | ♮, n
+
η q 1, j 1 +
η q 1 , j 1 − u nj = 0 ,
2
2
def
j = j 1 , N2 ,
(5.9)
where we introduced the vector η := η, ξ and index j 1 = 1, 2, . . . , N1 − 1 . The
contravariant velocity component v j1 is approximated as
u nj − x nt, j · y q♭,2n, j − v jn − y t,n j · x ♭,q 2n, j
1, n
.
(5.10)
vj =
J nj
The components of the (Cartesian) velocity vector u nj are computed using formulas (5.3)
on the subsequent
except the fact that we use the values of the velocity potential ϕ n+1
j
j
n+1
time layer t = t
.
We employ formula (5.9) in order to compute the free surface position in ‘interior’ nodes
q j 1 , N1 with j 1 = 1, 2, . . . , N1 − 1 , which lie on the upper side of the square Q h0 . Thanks
to impermeability conditions (5.5), kinematic conditions (2.7), (2.8) take a very simple
form:
ηt − u = 0 ,
q ∈ q 0, N2 , q N1 , N2 .
43 / 72
Wave/moving wall interaction
Taking into account this information, the finite difference formula (5.9) takes a much
simpler form as well:
η n+1
− η nj1
j1
τn
= u nj1 , N2 ,
j1 ∈
0, N1 .
The fully discrete scheme to compute the free surface transport is described.
5.4. Moving wall displacement (4)
The movable wall position is given by the function x = s (t) , which is a solution to
equation (2.10). The fluid pressure p is determined from the Cauchy–Lagrange integral
given in dimensionless variables in equation (2.20). The force acting on the vertical wall is
given in dimensionless variables in equation (2.19). In curvilinear coordinates the pressure
is given in equation (3.11) and it can be used to compute the force in curvilinear coordinates
as well (3.10). The advantage of this formulation is that the wall has a fixed position
q 1 = 0 in the transformed domain. The integral in equation (3.10) is discretized using
the trapezoidal quadrature formula:
F (t) ≈
NX
2−1
j2=0
p 0, j 2 + p 0, j 2 + 1
h2 .
2
In order to compute the time derivative of the velocity potential ϕ t appearing in the
N
pressure term p 0, j 2 j 2 2= 0 , we use the values of the velocity potential on three time lay
ers ϕ jn−1 j , ϕ nj j and ϕ n+1
. So, the time derivative of the velocity potential is
j
j
approximated as:
n−1
n
3 ϕ n+1
0, j 2 − 4 ϕ 0, j 2 + ϕ 0, j 2
.
ϕ t 0, j 2 ≈
2 ∆t
However, due to the change of variables in (3.10) one has to compute also the vertical
velocity y q 2 of grid nodes. This is done as specified in (5.2) with Cartesian velocities
computed as specified in (5.3) with the only difference — we use the values of velocity
potential at the new time level ϕ jn+1 j .
In order to integrate numerically the nonlinear differential equation (2.18), we rewrite
this equation as a system of two first order differential equations:
ṡ = υ ,
m υ̇ + k s = − F (t) − F (0) ,
together with initial conditions:
s (0) = 0 ,
υ (0) = 0 .
The last system of equations can be rewritten in the vectorial form for our convenience:
υ̇ = Ξ · υ + Σ (t) ,
υ (0) = 0 ,
(5.11)
44 / 72
G. Khakimzyanov & D. Dutykh
where we introduced the following notations:
!
!
def
def
s (t)
0 1
υ (t) :=
, Ξ :=
,
k
υ (t)
−m
0
def
Σ (t) :=
− m1
!
0
.
F (t) − F (0)
The Cauchy problem (5.11) is solved using the so-called modified Euler method∗:
υ⋆ − υn
= Ξ · υn + Σn ,
∆t
1
υ n+1 − υ n
Ξ · υn + Ξ · υ⋆ + Σn .
=
∆t
2
If we exclude the intermediate variable υ ⋆ from last equations and rewrite the obtained
equations in the component-wise form, we obtain the following scheme:
s n+1 − s n
∆t k n
∆
F (t) − F (0) ,
= υn −
s +
∆t
2 m
2m
n+1
n
∆t k n
k n
1
υ
− υ
F (t) − F (0) −
= −
s +
υ .
∆t
m
m
2 m
The last scheme is implemented in our numerical code.
5.5. Elliptic mapping construction (5)
In previous sections we assumed that mapping (3.1) was known. In order to complete
our numerical method description we have to describe how to construct this mapping in
practice. Namely, we shall use the so-called equidistribution method [45]. A typical grid
generated using this method is shown in Figure 5. Recently this method was successfully
applied in 1D to the simulation of conservation laws [44]. Below we provide a detailed
description of this method to two spatial dimensions.
⊆
According to the equidistribution method, the coordinates of grid nodes x n+1
j
j
n+1
are determined by
Ω (t) in the physical space on the subsequent time layer t = t
solving the following vectorial parabolic equation:
∂x
∂x
∂
∂x
∂
ς
g22 ̟ 1 +
g11 ̟ 2 ,
(5.12)
=
∂t
∂q 1
∂q
∂q 2
∂q
where ς > 0 is a positive smoothing parameter whose value is chosen to minimize the
oscillations in grid nodes trajectories. Above ̟ = ̟ (x, t) is the so-called monitoring
2
function, which determines the local density of the grid nodes and g α α α = 1 are metric
tensor components defined in (3.14).
The equidistribution method based on the solution of parabolic equations for grid nodes
coordinates was proposed in [71]. A few years later this method was applied to the simulation of 1D (compressible) gas dynamics problems in [1, 72]. This method has been
especially designed for non-stationary problems and its application allows to avoid (or simply reduce) abrupt changes in the nodes positions when we move from one time layer to
∗This
scheme is also called the second order explicit Runge–Kutta scheme [4].
45 / 72
Wave/moving wall interaction
y
O
x
Figure 5. A typical grid generated using the equidistribution method. On this
picture we show a zoom on the area near the moving wall. We underline that the
free surface profile is chosen to illustrate the adaptive features of the grid and it
does not come from any physical simulation.
the following. In this method one constructs the new mesh by taking into account the
position of nodes at the last time layer. The new coordinates of grid nodes are determined
by solving a discretized version of the parabolic problem (5.12). This modification of the
classical equidistribution method (proposed in [17]) of the construction of dynamically
adapted moving grids has been used to solve numerous 1D problems [13–15, 18, 22, 58, 80],
this list is not being exhaustive.
The system of nonlinear equations (5.12) is solved using an iterative method of alternating directions. The monitoring function ̟ was computed at the known solution on the
n
time
n layer t = t . As the initial guess for this iterative process we take the known mesh
x j j from the previous time layer as well. We mention that at the initial moment of time
t = 0 we employ the traditional (steady) equidistribution method as it was described in
[46]. It is equivalent to solve the system (5.12) with ς ≡ 0 .
46 / 72
G. Khakimzyanov & D. Dutykh
Since the position of boundary nodes varies∗ from
one time step to another, before solving
n+1
equations (5.12) we have to determine the grid x j
along the boundary ∂ Q h0 . Let us
j
assume that a portion Γ̊ ∈ Γ ℓ , Γ r , Γ b , Γ s ≡ ∂ Q 0 of the boundary Γ n+1 ≡ Γ(t n+1 )
is given in the following parametric form:
x = x (ζ) ,
y = y (ζ) ,
where ζ is a real parameter. Then, the coordinates of grid nodes x n+1
j
boundary can be determined by formulas:
x n+1
= x (ζ n+1
),
j
j
j
⊆ Γ̊ h on the
y jn+1 = y (ζ n+1
).
j
We assume that the nodes on the boundary under consideration are re-ordered with a single
scalar index j for the sake
of notation compactness. So, the problem is to determine the
n+1
in the parameter space. To do it, we solve numerically
discrete scalar function ζ j
j
the following scalar nonlinear difference equation:
n+1
n+1
ζ n+1
ζ n+1
ζ n+1
− ζ n+1
n n+1
1 n n+1
j+1 − ζ j
j+1 − ζ j
j
j−1
ς
=
̟ J̃
,
− ̟ J̃
j+ 21
j− 21
τn
h
h
h
(5.13)
where h ∈ h 1 , h 2 is the discrete step in space, which depends on the boundary under
o n+1
nq
def
2
2
is the trace of the transformation
:=
+
y
x
consideration. The function J̃ n+1
j
ζ
ζ
j
Jacobian at this boundary. Please, notice that we compute the monitoring function ̟ n on
the known position of the nodes. The difference equation (5.13) is solved on four sides (left
and right walls, bottom and free surface) of the discretized rectangle Q h0 . The distribution
of boundary nodes obtained in this way is used as the Dirichlet-type boundary condition
for the 2D grid generation, which is achieved by solving numerically
n+1 equation (5.12). By
.
the end of this step we know the grid Ω̄ n+1
nodes
coordinates
xj
h
j
5.5.1 The re-computation
to take into account the new
At the next step we recompute quantities η n+1 and ϕ n+1
j
j
mesh. For this, we solve again equation (5.1) (or an analogue of (5.6) on the boundaries) in
which the Cartesian components of the velocity (5.3) are still computed on the previous
grid Ω̄ nh and using the velocity potential values from the previous time layer t = t n .
However, in formulas (5.3) we take the new velocity of grid nodes. In other words, formula
(5.2) is replaced by
x n+1
− x nj
j
x t, j =
.
τn
During the second solution of system (5.7) the right hand side member (5.8) is modified as
well to take into account the new velocity of the left wall and, obviously, the new location
of mesh nodes on this boundary. Another modification concerns the velocity components
∗Indeed,
we are dealing with free surface flows in the presence of a moving wall.
47 / 72
Wave/moving wall interaction
u , v when they are used in the computation of the contravariant component v j1, n whose
discretized formula is given in (5.10). The quantity v j1, n is needed to determine the new
position of the free surface elevation η n+1 . During the present step, the components u , v
are computed as
n+1
n+1
−ϕq 1 · xq 2 + ϕq 2 · xq 1
ϕq 1 · y q 2 − ϕq 2 · y q 1
n+1
n+1
,
vj
=
.
uj
=
J
J
j
j
The last formulas replace equations (5.3) during this stage.
5.6. Stability of the scheme
In order to study the stability of the proposed scheme, we consider the linearized governing equations. Moreover, we consider only the Initial Value Problem∗ (IVP) in order to
remove the complexity of (lateral) boundary conditions treatment. The linear water wave
problem is known as the Cauchy–Poisson problem [27, 42, 59, 73] since the pioneering
works of Augustin Louis Cauchy [6] and Siméon Denis Poisson [66]. Let us formulate
this problem in precise mathematical terms. Consider a fluid layer of infinite horizontal
extent (−∞ < x < +∞) over a solid bottom of uniform depth y = −h 0 const . The
fluid domain is bounded from above by the free surface, whose location is assumed to be
at y = 0 after the linearization. The Cauchy–Poisson problem consists in finding the
free surface elevation η and the velocity potential ϕ by solving the Laplace equation† in
a strip:
∆ϕ = 0,
−∞ < x < +∞ ,
− h0 6 y 6 0 .
The last equation has to be completed by the following free surface boundary conditions:
ηt − ϕy = 0 ,
ϕt + g η = 0 ,
y = 0,
(5.14)
y = 0,
(5.15)
and by one bottom impermeability condition:
ϕy = 0 ,
y = −h 0 .
It is not difficult to see that free surface boundary conditions given above are nothing else
but linearized versions of equations (2.13) and (2.14). A particular solution to the Cauchy–
Poisson problem can be easily obtained using some elementary Fourier analysis [27, 42,
59, 73]:
η (x, t) = η 0 ei (k x − ω t) ,
ϕ (x, y, t) = ϕ 0 ei (k x − ω t) cosh k (y + h 0 ) , (5.16)
∗In fact, we have the so-called Cauchy–Poisson problem, which is of mixed type between IVP and
BVP. We say that we have an IVP since there are no boundary conditions on lateral boundaries, i.e. the
domain is unbounded in the horizontal directions. However, there are still boundary conditions to be
satisfied on the impermeable bottom and on the (linearized) free surface.
†This equation is to be compared with equation (2.12).
48 / 72
G. Khakimzyanov & D. Dutykh
def 2 π
def 2 π
is the wave number, λ is the wave length, ω :=
is the wave frequency
where k :=
λ
T
and T is its period. Wave amplitudes η 0 ∈ R and ϕ 0 ∈ R are some real numbers. The
wave frequency ω is related to the wave number k through the so-called dispersion relation
of gravity waves:
p
ω (k) = ± g k tanh(k h 0 ) .
(5.17)
For the sake of simplicity we consider only waves moving rightwards. It fixes the branch +
in the relation above. We reiterate the fact that the dispersion relation (5.17) is a necessary
condition for the existence of solutions (5.16).
The numerical scheme considered above can be applied to the Cauchy–Poisson problem as well. The semi-discretization in time of free surface boundary conditions (5.14),
(5.15) reads:
η n+1 (x) − η n (x)
− ϕ yn+1 = 0 ,
(5.18)
τn
ϕ n+1 (x, 0) − ϕ n (x, 0)
+ g η n (x) = 0 .
(5.19)
τn
From now on, for the sake of simplicity we take the time step τ n ≡ τ > 0 to be constant.
The elementary solution (5.16) can be semi-discretized as well:
η n (x) = η 0 ρ n ei k x ,
ϕ (x, y, t) = ϕ 0 ρ n ei k x cosh k (y + h 0 ) ,
def
where we introduced the notation ρ := e− i ω τ . By substituting this semi-discrete solution
ansatz into relations (5.18), (5.19) we obtain the following linear system of equations with
respect to wave amplitudes η 0 and ϕ 0 :
ρ − 1
g η0 + ϕ0
cosh k h 0 = 0 ,
τ
ρ − 1
η 0 − ϕ 0 ρ k sinh k h 0 = 0 .
τ
In order to have non-trivial solutions, the determinant of this system has to vanish. It
gives us the following quadratic equation with respect to the transfer coefficient ρ :
τ 2 ω2
2
ρ − 2 1 −
ρ + 1 = 0.
(5.20)
2
To have the linear stability property of our scheme, it is necessary that both roots of this
equation verify the inequality | ρ ± | 6 1 . To meet this requirement, it is sufficient to ask
that the discriminant of the quadratic equation (5.20) is not positive, i.e.
τ 2 ω 2
τ 2 ω2
− 1 6 0,
4
or equivalently
2
,
(5.21)
τ 6
ω
where the wave frequency was defined in equation (5.17) (with the sign + by our convention).
49 / 72
Wave/moving wall interaction
Let ∆ x be the discretization step size in the horizontal direction. The minimal wave
length λ inf , which can be represented on the grid with spacing ∆ x is λ inf ≡ 2 ∆ x . All
other waves satisfy the inequality λ > λ inf . These considerations on the wave length can
be translated into the language of wave numbers, i.e.
π
def
k 6 k sup :=
.
∆x
Henceforth, we can derive the following estimation for the wave frequency using the dispersion relation (5.17):
s
r
p
p
p
tanh k h 0
π
k
tanh k h 0 = k g h 0 ·
6
.
g h0 ·
g h0
ω (k) =
h0
k h0
∆x
Consequently, in order to satisfy the stability condition (5.21), it is sufficient to impose the
following restriction on the time step τ magnitude:
τ 6
2∆x
√
.
π g h0
(5.22)
An analogue of this stability condition will be used below during the simulation of fully
nonlinear problems. The most important conclusion of this Section is that we have a
hyperbolic-type stability Courant–Friedrichs–Lewy condition [11] — the time step is
a linear function of the mesh spacing.
5.6.1 Practical choice of the time step
In the previous Section we studied the scheme stability in the linear case. However, in
numerical simulations presented below, we solve the nonlinear problem. Henceforth, one
may ask the question how to choose the time step in practical nonlinear simulations. Below
we explain our approach to this problem.
In order to cover the whole family of problems, we choose to work in scaled variables.
The CFL condition (5.22) can be rewritten in dimensionless variables as
τ∗ 6
2 ∗
h ≈ 0.64 h 1∗ ,
π 1
h1
. As grid spacing h 1 , we take the smallest horizontal spacing along the
h0
free surface (since most important variations take place there):
def
where h ∗1 :=
def
h 1 ∼ ∆x nmin :=
min
0 6 j1 < N1
| x j 1 +1, N 2 − x j 1 , N 2 | .
Consequently, the dimensionless CFL condition can be rewritten as
n, ∗
τ n∗ 6 0.64 ∆x min
.
However, our numerical simulations show that this estimation is too pessimistic. Consequently, in all simulations presented below we took the time step according to the following
50 / 72
G. Khakimzyanov & D. Dutykh
less restrictive formula:
∗
τ n∗ 6 κ ∆x n,
min ,
with κ = 0.95 . Most probably, the last condition is pessimistic as well. However, it
guaranteed the stability of our nonlinear computations.
6. Numerical results
Above we described the proposed finite difference scheme and our resolution algorithm
on a fixed reference domain. Below we present several validation tests and numerical
experiments which show the performance and abilities of our numerical approach.
6.1. Solitary wave run-up on a fixed wall
In order to illustrate the applicability of our numerical algorithm, we consider the classical problem of the solitary wave/fixed wall interaction. Due to symmetry considerations,
this set-up is equivalent to the head-on collision of two equal solitary waves. This problem
is well-studied in the literature [7, 10, 57, 60, 74] and it can serve as the first validation
test.
Consider a solitary wave of amplitude α moving in the leftward direction. The channel
has a constant depth and the wall is assumed to be fixed in order to be able to perform
comparisons with previous investigations. On the right the channel is also bounded by a
fixed vertical wall. The total length of the channel is equal to ℓ = 20 . In this Section
we provide all values in dimensionless variables as it was explained earlier in Section 2.3.
Moreover, we assume in these computations that waves do not overturn. In other words,
the free surface is traditionally given as the graph of a function y = η (x, t) . The initial
condition for the free surface elevation η 0 (x) ≡ η (x, 0) and velocity field u 0 (x) ≡
u (x, 0) are given by the following approximate formulas [65]:
r
h
i
3α
def
2 κ
η 0 (x) = α sech
(x − x 0 ) ,
,
(6.1)
κ :=
α + 1
|2
{z
}
def
=: Θ
η 0 (x)
1
α2
3 y + 1 2
u 0 (x, y) = − 1 + α
·
+ √
−
1 + η 0 (x)
4 η 0 (x) + 1
1 + α 4
3 − η 0 (x)
η 0 (x) − 1
2
4
sech Θ +
sech Θ , (6.2)
· 2
η 0 (x) + 1
η 0 (x) + 1
√
v 0 (x, y) = −
√
3 α 3 (1 + y)
where x 0 is the wave crest initial position.
cosh Θ sinh Θ
2 ,
α + cosh 2 Θ
(6.3)
51 / 72
Wave/moving wall interaction
The interaction of a solitary wave with amplitude α = 0.4 with a vertical (fixed)
wall is shown in Figure 6 at different moments of time. Under a wave we show also the
adapted grid. On this Figure we can see that after the reflection the solitary wave does
not recover completely its initial shape. In particular, behind the main wave we observe
a slight dispersive tail. The numerical simulations on refined grids show that it is not
a numerical effect. This dispersive tail appears on all grids and it reflects an intrinsic
property of the full Euler equations — their non-integrability [28, 30]. In this example
we use the following monitoring function in order to adapt the grid to the solution:
̟ (x, t) = 1 + ϑ | η (x, t) | ,
(6.4)
where ϑ > 0 is a positive ad-hoc parameter. In computations presented in the Section we
use ϑ = 10 . The main rationale behind this choice of the monitoring function ̟ (x, t) is
to put more nodes in areas where the waves are large.
In Figure 6 we can see that the refined area follows somehow the solitary wave crest
during its motion towards and fromwards the vertical wall. From equation (6.4) it can be
seen that the monitoring function does not depend on the vertical variable y . As a result,
we obtain the grids with almost vertical lines. Consequently, in accordance with stability
condition (5.22) to determine an admissible local time step τ n we can use the following
formula:
n
τ n = ν ∆ x min
,
where 0 < ν 6 1 is the security factor and ∆ x nmin is the minimal mesh spacing on the
free surface at time t = t n , i.e.
n
def
∆ x nmin :=
min
| x j 1+1, N2 − x nj1, N2 | .
0 6 j 1 < N1
One of the main characteristics of the wave/wall interaction is the maximal amplitude
of the wave on the wall. This quantity is called the maximal run-up and will be denoted in
our study as R max . Obviously, in our experimental conditions this quantity depends on the
amplitude of the incident solitary wave, i.e. R max = R max (α) . In Figure 7 we represent
this dependence according to our numerical simulations (solid black line), experimental
data (filled markers), other computations (empty markers) and the following asymptotic
analytical prediction [74] (dashed line):
h
3 2i
1
+ o(α 3) ,
as
α → 0.
α +
α
R max (α) = 2 α 1 +
4
8
In Figure 7 we can see that there is an overall good agreement among all presented data up
to the amplitude α . 0.4 . For higher waves some divergences start to appear. However,
the agreement between our numerical model with other potential flow solvers [10, 34]
continues up to α . 0.6 . The experimental points go rather below our predictions. It
can be easily explained by the neglection of viscous and friction effects in our numerical
(and mathematical) model. Nevertheless, we would like to mention that our numerical
results agree particularly well with experimental data reported in [16].
52 / 72
G. Khakimzyanov & D. Dutykh
1.0
y
0.5
0.0
-0.5
-1.0
0
5
10
15
20
x
15
20
x
15
20
x
15
20
x
(a) t = 0
1.0
y
0.5
0.0
-0.5
-1.0
0
5
10
(b) t = 7
1.0
y
0.5
0.0
-0.5
-1.0
0
5
10
(c) t = 9
1.0
y
0.5
0.0
-0.5
-1.0
0
5
10
(d) t = 18
Figure 6. Solitary Wave (SW) interaction with a vertical fixed wall at different
instances of time. SW’s amplitude is α = 0.4 . The propagation direction is
shown with a horizontal arrow.
53 / 72
Wave/moving wall interaction
2.4
R
1
2
3
4
5
6
7
8
9
2.0
1.6
1.2
0.8
0.4
0.0
0
0.1
0.2
0.3
0.4
0.5
0.6
a
0.7
Figure 7. Dependence of the maximal run-up R max at the vertical wall on the
incident solitary wave amplitude α . Experimental data: (1) [79], (2) [57], (3)
[16], (4) [56]. Numerical data: (5) [34], (6) [8], (7) [10]. Analytical prediction
[74] (8) (dashed line). The present study (9) is given by the solid line.
6.2. Wave generation by a numerical wave maker
The study of water wave interaction with movable partially submerged bodies and objects
is a very important problem of the modern computational hydrodynamics. In the previous
Section 6.1 the wall was considered to be fixed in order to compare our numerical predictions
with available data. Starting from this Section we allow the left wall as a movable object.
As the first step towards the freely moving solid boundary, we consider first the situation
where left wall motion is prescribed by a given law. Physically, it corresponds to the
vertical piston motion used in many experimental facilities. Traditionally, the generated
waves are understood using linear or weakly nonlinear theories [55]. Recently this situation
was modelled numerically with Boussinesq-type equations in [63]. The results presented
below are fully nonlinear and fully dispersive.
54 / 72
G. Khakimzyanov & D. Dutykh
Consider a numerical wave tank with horizontal bottom and of uniform depth d = 1 m
(when the water is unperturbed). The left wall is initially located in the point x = 0 . For
times t > 0 the wall moves according to the following law:
s (t) = α 1 − e− β t sin(ω t) .
(6.5)
Thus, during the initial times the wall moves to the right. In this way, the wall motion
is completely determined by three parameters α > 0 , β > 0 and ω . The first one
(α) specifies the maximal wall oscillation amplitude (in the horizontal extent), the second
parameter β controls the speed of the relaxation∗ towards the stationary periodic regime
and ω is the frequency of wall oscillations (or equivalently it controls the oscillation period).
In Figure 8 we depict a few simulations results, which were obtained by varying the
parameter ω for fixed values of two other parameters:
α = 0.4 m ,
β = 0.5 Hz .
Figure 8 shows clearly that the generated wave amplitude (measured at the moving wall)
depends essentially on the wall oscillation period, which is determined by parameter ω .
This
effect is illustrated in Figure 8 for three different values of the parameter ω ∈
0.5 Hz, 1.0 Hz, 3.0 Hz . For each wall trajectory we show also the corresponding free
surface excursion on the moving vertical wall. The increase in ω results in the reduction
of the oscillation period. We can also witness the relaxation effect on the free surface
oscillation. Indeed, as the time t goes on, the wave amplitude increases and becomes
practically stationary. The speed of relaxation depends on the parameter β . We underline
the fact that for all three (considered) values of the wall oscillation frequency ω , the
amplitude of oscillations α was the same (α ≡ 0.4 m). Nevertheless, the wave amplitudes
registered on the wall were significantly different. This is the first practical conclusion
that we can draw from our numerical simulations: the piston oscillation period has a much
bigger influence on the generated wave amplitude, than the amplitude of these oscillations.
We can explain this observation with a simple physical argument: during ‘fast’ piston
oscillations, the energy is pumped into the fluid close to the piston faster than waves can
evacuate this energy by propagating into the wave tank.
The second important observation that we can make based on our simulations is that
the maximal wave run-up does not take place at the moments where the wall displacement
is maximal. A more detailed investigation showed that the wave amplitude on the wall
is maximal when the wall acceleration is maximal. It happens when the wall passes by
its initial (or mean, or unperturbed) position x = 0 . After passing this point, the wall
decelerates and the wave has time to flow away from the piston, which reduces the value
of run-up.
Finally, we can also observe that the wave run-up value on the wall is asymmetric with
respect to the still (unperturbed) water level. This effect is much better visible for fast wall
∗It
is easy to see that the wall oscillation amplitude tends to its maximum value α as t → +∞ . This
justifies the term ‘relaxation’ associated to this parameter β . The speed of this convergence depends on
the value of parameter β , i.e. bigger is faster.
55 / 72
Wave/moving wall interaction
η
s, cm
40
η
s, cm
40
40
40
20
20
20
0
0
0
5
4
20
3
0
1
2
-20
-20
-20
-40
-40
-20
6
-40
0
5
10
15
20
t, s
0
5
(a)
10
15
20
-40
t, s
(b)
Figure 8. Free surface excursion on the moving wall η s(t), t (thick lines 1 , 3
and 5) along with the wall trajectory s (t) (thin lines 2 , 4 and 6) for (a)
ω = 0.5 Hz (lines 1 and 2); ω = 1.0 Hz (lines 3 and 4); (b) ω = 3.0 Hz (lines
5 and 6).
motions. Indeed, for ω = 3.0 Hz the maximal run-up is ≈ 43 cm , while the maximal
run-down is ≈ 34 cm .
6.3. Surface wave run-up on a movable wall
In this Section we consider the wave/wall interaction problem, where the wall motion
is not prescribed, but it is determined as a part of the problem solution. The left wall is
attached to a spring system, which can deform under the wave action.
6.3.1 Single pulse
To begin, we consider first the case of a single localized wave impulse interacting with
a moving wall. The channel depth is taken to be d = 1 m and the channel length is
ℓ = 20 m . At the initial moment of time t = 0 , the velocity field is taken to be
quiescent and the free surface shape is given by the following formula:
h
i
α 1 + cos π x − x , | x − x | 6 Υ ,
0
0
2
Υ
η 0 (x) =
(6.6)
0,
| x − x0 | > Υ .
For subsequent times t > 0 this initial condition is separated in two waves of the amplitude
≈ α2 , which move in opposite directions. The wave moving leftwards interacts with the
56 / 72
G. Khakimzyanov & D. Dutykh
η/a
s/a
1.0
1
2
0.025
0.5
0.000
-0.025
0.0
1
2
-0.050
-0.5
0
5
10
15
20
25
30
t(g/h
0)
(a)
1/2
0
5
10
15
20
25
30
t(g/h
0)
1/2
(b)
Figure 9. Free surface excursion as a function of time at point x = x 0 = 0.7
(a) and wall oscillations (b) for the initial condition given in Equation (6.6). The
parameter α = 1 cm (1) and α = 10 cm (2).
moving wall. We would like to underline the fact that the initial left wall position (x = 0)
does not coincide with the wall equilibrium position in the absence of water. Even if the
free surface coincides with the still water level, the springs are deformed under the action
of the hydrostatic pressure. Thus, the equilibrium wall position x = x e with undeformed
springs is located somewhere to the right from its initial position, i.e. x e > 0 . We
performed a series of numerical simulations by varying the parameters α , x 0 and Υ of the
initial condition (6.6). Additionally we varied also the wall mass m and the springs rigidity
coefficient k .
First, we conducted a series of numerical experiments for small initial wave amplitudes
α in view of comparisons against the numerical simulations reported in [40]. The Authors
of the preceding study employed the Boundary Integral Equations Method (BIEM) [2,
3]. Unfortunately, the Authors of [40] did not publish any tabulated data to perform
quantitative comparisons. However, we performed qualitative comparisons of numerical
results. We choose the same values as in [40], i.e. x 0 = 0.7 m , Υ = 0.5 m and
dimensionless spring rigidity coefficient k = 10 . The wall dimensionless mass is m = 1.5 .
In Figure 9(a) we show the free surface oscillations recorded at the initial pulse location∗
x = x 0 = 0.7 m . We show the results for two initial wave amplitudes. The line (1)
corresponds to α = 1 cm , while α = 10 cm is depicted by line (2). Please, notice
that curves presented in Figure 9 are shown relatively to the initial wave amplitude α .
These curves compare quantitatively and qualitatively well to those presented in [40]. To
the graphical resolution the oscillation decay rates and oscillation periods are the same.
∗Here
we mean the center or pulse crest.
57 / 72
Wave/moving wall interaction
In Figure
9(b) we show moving wall trajectories in time for two initial wave amplitudes
α ∈ 1 cm, 10 cm . These curves also compare well with previous computations reported in [40]. We can also see that for small times the solid and dashed curves shown in
Figure 9(a,b) almost coincide. It means that for small initial wave amplitudes α the results
of computations depend almost linearly on α . This observation supports the applicability
of the linear theory [48].
Let us study now the influence of the springs rigidity parameter k on the resulting wave
field. We took the following parameters in this computation:
x0 = 2 m ,
Υ = 1 m,
α = 10 cm ,
m = 1.5 .
The parameter k is variable. In Figure 10(a) we can see that the first run-up on the wall
is large. Then, the following run-ups are generally decaying in amplitude since the wave
energy spreads over the whole wave tank during the dispersive (and possibly nonlinear)
effects. Moreover, for k = 5 (curve 2) the decay rate is the strongest since a larger part
of the wave energy is converted into the wall motion. Thus, we have the wave energy
transformation into the elastic energy of the spring system. When the springs rigidity
is high, the wall performs small later motions, as it can be seen in Figure 10(b). The
increase in k results also in the increase of the wall oscillation frequency around the initial
equilibrium position x = 0 . In the same time, the wave run-up records on the wall are
practically identical with the fixed wall case (cf. curves 3 and 4 in Figure 10(a)). In other
words, for high rigidities k ≫ 1 , the ‘moving’ wall interacts in a way very similar to
the fixed wall. On the other hand, when the rigidity decreases, the wall oscillation period
and amplitude both increase. Wave run-up on the wall decay rate also increases with
the parameter k . However, in contrast to large values of k , a flexible spring system can
allow the maximal wave run-up value during the second (or even further subsequent) wave
interactions with the wall. In Figure 10(a) we can see that the value of the maximal wave
run-up on the wall is generally lower for weak springs comparing to quasi-fixed wall results
(k ≫ 1). From these simulations we can conjecture that flexible spring systems should
be used for more efficient wave run-up reduction.
Let us consider now the case of the weak springs system (k = 1) and a moving wall
with mass m = 1 . The question we would like to investigate is how the wave run-up
and wall trajectory depend on the incident wave amplitude α ? In Figure 11(a) we show
the wave run-up value on the moving wall for a moderate amplitude α = 20 cm (curve
1) and a fairly strong amplitude α = 60 cm (curve 2). The initial condition is given
by the single pulse formula (6.6). The initial impulse center is x 0 = 12 m and its halflength Υ = 5 m . In Figure 11(b) we show the wall displacement s (t) . As expected,
the wall displacement becomes larger as the incident wave amplitude α increases. It can
be seen that for weak springs the wall first retracts under the incident wave action. Then,
it comes back towards its initial position and slightly oscillates around this point. It is
very interesting to observe the high amplitude pulse (curve 2) interaction with the moving
wall depicted in Figure 11(a). Here the wave run-up value grows until the wall attains its
left-most position. Then, on the way back the wall motion provokes the second maximal
run-up value. It explains the non-monotonic behaviour of the curve 2 around its maximum.
58 / 72
G. Khakimzyanov & D. Dutykh
0.08
η/h
0
s/h0
0.02
0.04
0.00
0.00
-0.04
-0.08
-0.02
4
3
2
1
0
5
10
1
2
3
-0.04
15
t(g/h
0)
1/2
0
5
10
(a)
15
t(g/h
0)
1/2
(b)
Figure 10. Free surface excursion on the vertical moving wall (a) and the wall
position (b) as functions of time for the spring rigidity parameter k = 1 (1),
k = 5 (2) and k = 25 (3). For the sake of comparison the curve (4) shows the
wave run-up on the fixed wall, i.e. k = ∞ .
η
50
0
1
2
40
30
-20
1
2
20
-40
10
0
0
2
4
6
(a)
8
t, s
10
-60
0
2
4
6
8
t, s
10
(b)
Figure 11. Free surface excursion on the vertical moving wall (a) and the wall
position (b) as functions of time for α = 20 cm (1) and α = 60 cm (2). The
initial impulse center is x 0 = 12 m and its half-length Υ = 5 m .
This behaviour is possible only for high amplitude waves and weak (i.e. easily deformable)
springs.
59 / 72
Wave/moving wall interaction
η, cm
η, cm
21
21
18
18
1
2
3
15
0
5
10
15
m
1
2
3
20
15
(a)
0
2
4
6
8
k
10
(b)
Figure 12. Maximal run-up on the vertical wall (during the whole simulation
time) as a function of the moving wall mass (a) for fixed springs rigidity k = 1
(1), k = 3 (2) and k = 5 (3). On the right panel we show the same quantity as
a function of the springs rigidity (b) for fixed wall masses m = 1 (1), m = 3
(2) and m = 5 (3). This computation corresponds to the incident initial pulse
(6.6).
We naturally come to the question of the wall mass m and springs rigidity k on the
maximal wave run-up value of the impulse (6.6) on the movable wall. We use the following
parameters of the incident pulse:
x 0 = 12 m ,
Υ = 5 m,
α = 20 cm .
In Figure 12(a) we depict the maximal run-up value as the function of the wall mass
m . It can be easily seen that for a fixed rigidity parameter k , the maximal run-up is a
non-monotonic function of m . Even more, for every value of k , one can find the value
m ⋆ for which the maximal wave run-up will be minimal. It implies that the wave runup can always be reduced by choosing the appropriate wall parameters k ⋆ and m ⋆ . In
Figure 12(b) we show
the maximal wave run-up dependence on k for fixed dimensionless
wall masses m ∈ 1, 3, 5 . So, it can be seen that the run-up depends monotonically
on k . Namely, it increases with wall rigidity k . As we increase k → +∞ , the maximal
run-up tends to its limiting value, which depends on the incident wave parameters and
which is independent on the dimensionless wall mass m .
6.3.2 Solitary wave
Finally, we show some numerical results on the solitary wave interaction with a moving
wall. This test case is a natural extension of the case considered earlier in Section 6.1.
60 / 72
G. Khakimzyanov & D. Dutykh
η
η
40
40
36
36
1
2
3
1
2
3
32
32
0
2
4
6
(a)
8
m
10
0
2
4
6
8
k
10
(b)
Figure 13. Maximal run-up on the vertical wall (during the whole simulation
time) as a function of the moving wall mass (a) for fixed springs rigidity k = 1
(1), k = 3 (2) and k = 5 (3). On the right panel we show the same quantity as
a function of the springs rigidity (b) for fixed wall masses m = 1 (1), m = 3
(2) and m = 5 (3). This computation corresponds to the incident solitary wave.
The dashed line on both panels shows the maximal run-up value on a fixed wall.
The initial condition is provided by the same formulas (6.1) – (6.3). In order to avoid any
interactions of reflected waves from the right wall (during the simulation time), we increase
the computational domain length to ℓ = 40 m (from ℓ = 20 m). The computations
were performed for a solitary wave with moderate amplitude α = 20 cm . We would like
to study the influence of moving wall parameters on the maximal solitary wave run-up.
In Figure 13 we show the maximal run-up as a function of the wall mass m (a) and as a
function of springs rigidity k (b) (the other parameter being fixed). With a dashed line
we show the maximal run-up value on a fixed wall. The comparison of this result with
Figure 12 indicates that there are qualitative similarities in the behaviour of a single pulse
with a solitary wave (cf. Figure 13). So, for solitary waves the run-up can be significantly
reduced as well. All differences observed in Figures 12 and 13 can be solely explained by
the incident wave amplitude and its shape.
In Figure 14(a) we show the wave run-up record under the action of a solitary wave on
the moving wall as a function of time. This computation was performed for the wall mass
m = 5 and springs rigidity is k = 1 . In Figure 14(b) we report the wall displacement
during the wave/wall interaction process. In Figure 14(a) we depict also the wave run-up
record for a fixed wall as well. It can be clearly seen that the presence of a moving wall
reduces significantly wave run-up height on it. We can also notice that for a movable wall
the wave run-up happens twice in contrast to the fixed wall case. Indeed, the first (and
usually maximal) run-up takes place when the wall is retracting under the incident wave
61 / 72
Wave/moving wall interaction
η
s, cm
20
40
1
2
30
10
0
20
-10
10
-20
-30
0
-40
-10
0
3
6
9
(a)
12
t, s
0
3
6
9
12
t, s
(b)
Figure 14. Solitary wave run-up on the wall (a): free surface excursion on the
fixed (1) and moving (2) walls. Wall position as a function of time (b). The wall
mass m = 5 and springs rigidity is k = 1 .
action. Then, at some moment the springs accumulate enough elastic energy in order to
push back the wall towards its equilibrium position. Right after this turning point the
second run-up takes place. However, its value is usually lower than the first one.
In Figure 15 we show also the free surface evolution in space and in time. The left
panel 15(a) shows the classical fixed wall case for the sake of comparison. In this case the
wave field essentially consists of one incident and one reflected waves. Eventual inelastic
collision effects are negligible for a solitary wave of amplitude α = 20 cm . A much more
interesting wave field is generated by the interaction with a moving wall. It is depicted in
Figure 15(b). The wall motion generates multiple reflected wave, which travel with lower
speed (in agreement with their lower amplitude).
7. Discussion
The present study was devoted to the simulation of free surface waves in a two-dimensional
tank of variable size. Namely, a vertical wall is attached to a system of springs and it can
move under the action of incident water waves. Modelling and numerical simulation of
this problem was discussed in this study. Below we outline the main conclusions and
perspectives of the present study.
62 / 72
G. Khakimzyanov & D. Dutykh
(a)
(b)
Figure 15. Free surface dynamics represented as a surface in space-time
y = η (x, t) during a solitary wave run-up on a fixed (a) and moving (b) vertical
wall. The solitary wave amplitude α = 20 cm and x 0 = 10 m . The axis O x
shows only a part of the numerical wave tank. The whole length is ℓ = 40 m .
7.1. Conclusions
In the present study we considered the problem of surface water wave modelling in
domains with variable geometry. In general, problems on time-varying are known to be
notoriously difficult [47] (both theoretically and numerically). Here we considered a special
case of a numerical wave tank with a moving wall. This wall can move according to the
prescribed law. In this case we model the wave generation process by a piston-type wave
maker [38]. In our formulation the wall can also freely move under the wave action. It is
realized by connecting the wall to a system of elastic springs. In this way, the wall mass
m and springs rigidity k are taken into account in our model by writing an additional nonlinear second order differential equation for the wall position. Surface waves are described
mathematically using the full water wave problem, i.e. the irrotational incompressible
perfect fluid flow described by Euler equations [73]. No approximation was made and,
thus, the presented results are fully nonlinear and fully dispersive. Moreover, we proposed
a numerical algorithm on adaptive moving grids. The adaptation was achieved using the
so-called equidistribution principle [43, 44]. The discrete problem was rigorously studied
and the proposed discrete operator was shown to be self-adjoint on adaptive grids as well
as the continuous problem. Thus, the numerical method preserves qualitatively some properties of the continuous operator, which is referred to as the ‘structure preserving’ property
[39] in modern numerical analysis.
Wave/moving wall interaction
63 / 72
The proposed algorithm was first validated on the well-known case of a fixed rigid vertical
wall. There is enough analytical, numerical [10, 34] and experimental [16] data available
to validate the solver. Then, this numerical tool was applied to generate periodic waves
in silico by a moving piston-type wave maker. Finally, the wall motion and incident wave
run-up on this moving wall was studied using our numerical means. Our results compared
quite well with some published data [40]. The influence and effect of various parameters
on the wave run-up was investigated.
7.2. Perspectives
In the present article the geometry, mathematical formulation and the numerical method
have been presented for the two-dimensional configuration. In future works we plan to
extend this methodology for 3D flows. A priori, this generalization is going to be tedious,
but straightforward.
As another possible direction for future research, we would like to underline that the
moving (solid non-deformable) part in our study was a wall. This object is very important
in coastal engineering. However, in future works we would like to incorporate moving
objects of more general geometrical shape. As a minor point of improvement, one can
notice that in all simulations presented earlier the bottom was taken to be flat. It was
done on purpose to isolate the effect of the moving wall on incident waves. However, it is
not a limitation of the method. The formulation presented earlier works for general bottoms
and the simultaneous effect of the moving boundary and local bathymetric features on the
wave run-up on this wall has to be investigated.
Finally, the formulation we used was fully conservative. In other words, no dissipation
mechanism was included in our study except for the motion of the wall under the action
of water waves (however, the whole wave/wall system was conservative). Of course, this
model is an idealization of the reality. In real flows some dissipative effects are also present
due to the molecular viscosity, friction, turbulence or other mechanisms. For this purpose
a suitable visco-potential formulation has been proposed [23, 25, 26], which occupies the
intermediate level between the potential flows and Navier–Stokes equations.
7.3. Side effects
Our work has also at least one nice side effect. Namely, if we simplify the problem
considered in the present study by fixing the wall position, e.g. s(t) ≡ 0 , we obtain a
robust numerical wave tank for 2D surface waves similar to [37]. However, the technique
presented here is fully adaptive. By the way, we are not aware of any recent study using
similar adaptive redistribution methods for the full Euler equations with free surface. So,
this aspect might be new.
G. Khakimzyanov & D. Dutykh
64 / 72
Acknowledgments
This research was supported by RSCF project No 14-17-00219. D. Dutykh acknowledges the support of the CNRS under the PEPS InPhyNiTi project FARA and project
N◦ EDC26179 — “Wave interaction with an obstacle” as well as the hospitality of the Institute of Computational Technologies SB RAS during his visit in October 2015. D. Dutykh
would like to acknowledge the help of his colleague Tom Hirschowitz who timely provided
tea leaves in cases of emergency.
A. Transformation of coordinates
In this Appendix we explain some calculation details of coordinate transformations which
allow to pass from the system (2.12) – (2.20) posed on a variable domain Ω (t) to the system
(3.4) – (3.11) on the fixed domain Q 0 . Consider a smooth bijective map (3.1) (see also
Figure 2 for an illustration):
(
(
1
2
x = x (q , q , t) ,
q 1 = q 1 (x, y, t) ,
⇐⇒
(A.1)
y = y (q 1 , q 2 , t) .
q 2 = q 2 (x, y, t) .
We consider that the time variable is the same in both coordinate systems. Two maps
written above are mutually inverse. So, let us write this condition
x ≡ x q 1 (x, y, t), q 2 (x, y, t), t ,
(A.2)
1
y ≡ y q (x, y, t), q 2 (x, y, t), t .
(A.3)
Then, we differentiate (A.2) and (A.3) with respect to x . We thus obtain two relations
∂q 2
∂q 1
xq1
+ xq2
= 1,
∂x
∂x
∂q 1
∂q 2
yq1
+ yq2
= 0.
∂x
∂x
The last two relations can be considered as a system of two linear equations with respect
∂q 2
∂q 1
and
as unknowns. By trivially applying the Cramer rule we obtain
to
∂x
∂x
yq 2
yq 1
∂q 2
∂q 1
=
,
= −
,
(A.4)
∂x
J
∂x
J
def
where J := xq 1 yq 2 − xq 2 yq 1 is the Jacobian defined in (3.3). The Jacobian can be
also seen as the determinant of a matrix:
!
1
0
0
xq1 xq2
≡ xq 1 yq 2 − xq 2 yq 1 .
J = det x t x q 1 x q 2 ≡ det
yq1 yq2
yt yq1 yq2
Wave/moving wall interaction
65 / 72
Similarly, by differentiating (A.2) and (A.3) with respect to y and using Cramer’s rule,
one can show that
xq 2
xq 1
∂q 1
∂q 2
= −
,
=
.
(A.5)
∂y
J
∂y
J
Now, by differentiating (A.2) and (A.3) with respect to t we obtain the following two
equations
∂q 2
∂q 1
xq1
+ xq2
= −x t ,
∂t
∂t
1
2
y q 1 ∂q + yq 2 ∂q = − y t .
∂t
∂t
∂q 1
∂q 2
Solving the last linear system with respect to
and
yields
∂t
∂t
yt · xq2 − xt · yq2
xt · yq1 − yt · xq1
∂q 2
∂q 1
=
,
=
.
(A.6)
∂t
J
∂t
J
This concludes the computation of partial derivatives of the inverse mapping q = q (x, t) .
B. Transformation of the Laplace operator
Now we can rewrite the Laplace operator (2.12) in curvilinear coordinates q, t . The
idea consists in viewing the velocity potential as
ϕ (x, t) ≡ ϕ q (x, t), t .
The first derivatives can be easily computed:
∂ϕ
∂ϕ ∂q 1
∂ϕ ∂q 2
=
·
+
·
,
∂x
∂q 1 ∂x
∂q 2 ∂x
and using the just derived relation (A.4) we can express everything in terms of derivatives
solely with respect to q1 , q2 :
∂ϕ
∂ϕ i
1h
∂ϕ
yq2 · 1 − yq1 · 2 .
(q, t) =
∂x
J
∂q
∂q
Similarly, we can compute the other partial derivative of the velocity potential ϕ with
respect to y in new coordinates:
∂ϕ
∂ϕ
∂ϕ i
1h
− xq2 ·
+ xq1 · 2 .
(q, t) =
∂y
J
∂q1
∂q
Now we can apply recursively the just obtained results for ∂ϕ
(q, t) and ∂ϕ
(q, t) to compute
∂x
∂y
the second derivatives as well:
∂2ϕ
∂ ∂ϕ ∂q 1
∂ ∂ϕ
∂ ∂ϕ ∂q 2
=
·
·
=
+
=
∂x 2
∂x ∂x
∂q 1 ∂x
∂x
∂q 2 ∂x
∂x
∂ϕ ∂q 2 ∂q 1
∂ ∂ϕ ∂q 1
∂ϕ ∂q 2 ∂q 2
∂ ∂ϕ ∂q 1
·
·
·
+
·
+
·
+
·
.
∂q 1 ∂q 1 ∂x
∂q 2 ∂x
∂x
∂q 2 ∂q 1 ∂x
∂q 2 ∂x
∂x
66 / 72
G. Khakimzyanov & D. Dutykh
By taking into account the relations derived earlier, we obtain that
2
2
∂2ϕ
∂ h − ϕq1 yq1 yq2 + ϕq2 yq1 i
∂ h ϕq1 yq2 − ϕq2 yq1 yq2 i
1
+ 2
.
(q, t) =
∂x 2
J ∂q 1
J
∂q
J
2
Using similar methods we can obtain the following expression for ∂∂y ϕ2 (q, t) :
2
2
∂2ϕ
∂ h ϕq1 xq2 − ϕq2 xq1 xq2 i
∂ h − ϕq1 xq1 xq2 + ϕq2 xq1 i
1
+ 2
.
(q, t) =
∂y 2
J ∂q 1
J
∂q
J
The Laplace operator can be particularly compactly written in new variables if we introduce the metric tensor components:
def
g 1 1 := x 2q 1 + y q2 1 ,
def
g 1 2 ≡ g 2 1 := x q 1 · x q 2 + y q 1 · y q 2 ,
def
g 2 2 := x 2q 2 + y q2 2 ,
along with the implied anisotropic diffusion coefficients:
g12
def g 2 2
def
def g 1 1
K 1 1 :=
,
K 1 2 ≡ K 2 1 := −
,
K 2 2 :=
.
J
J
J
Using this notation we can finally write the Laplace operator in transformed coordinates:
∂2ϕ
∂
∂
∂2ϕ
1
.
+
≡
K11 ϕq1 + K12 ϕq2 +
K21 ϕq1 + K22 ϕq2
∂x 2
∂y 2
J ∂q 1
∂q 2
As a result, the Laplace equation reads:
∂ϕ
∂
Kαβ β = 0 ,
α, β = 1, 2 ,
∂q α
∂q
where we employed an implicit summation over repeated indices∗. This result coincides
exactly with equation (3.4) presented earlier in the manuscript.
C. Boundary conditions transformation
Finally, let us derive some relations which turn out to be useful in the transformation of
boundary conditions. In particular, in this Appendix we focus on the kinematic condition
(2.4). First of all, during the derivation of boundary conditions one needs time derivatives
of the inverse mapping in (A.1):
∂y
∂x
∂q 2
∂x
∂y
1
1
∂q 1
xq2
,
yq1
.
=
− yq2
=
− xq1
∂t
J
∂t
∂t
∂t
J
∂t
∂t
These formulas were given also earlier in (A.6) with more details on their derivation. Let
us compute Cartesian components of the velocity vector:
∂ϕ
(q, t) = ϕ q 1
∂x
∂ϕ
v (q, t) =
(q, t) = ϕ q 1
∂y
u (q, t) =
∗In
∂q 1
+ ϕq2
∂x
∂q 1
+ ϕq2
∂y
∂q 2
∂x
∂q 2
∂y
(A.4)
=
(A.5)
=
ϕq1 yq2 − ϕq2 yq1
,
J
− ϕq1 xq2 + ϕq2 xq1
.
J
the literature this rule is referred to as the Einstein summation convention.
67 / 72
Wave/moving wall interaction
The free surface elevation η on the transformed domain becomes a function of the variable
q 1 and time t :
η = η̃ (q 1 , t) ≡ η x (q 1 , q 2 ≡ 1, t), t .
Now, we can compute partial derivatives of the function η , which appear in the boundary
conditions:
∂q 1
∂q 2
η x = η̃ q 1
+ ηq2
,
∂x
∂x
∂q 1
η t = η̃ t + η̃ q 1
.
∂t
By substituting all these elements into equation (2.4), we obtain the following condition:
η̃ t + v 1 η̃ q 1 − v = 0 ,
(C.1)
where v 1 is a contravariant component of the velocity, which can be computed according
to this formula:
∂q 1
∂q 1
∂q 1
v1 =
+ u
+ v
.
∂t
∂x
∂y
The last equation (C.1) completes the derivation of the free surface kinematic boundary
conditions. Other boundary conditions can be derived in a similar manner.
Remark 7. Above, in the main text of our manuscript we do not use the tilde notation
η̃ in order to avoid overloading of the text with multiple notations. It is assumed that the
reader can make the difference between η (x, t) and η̃ (q 1 , t) depending on the context.
D. Piston motion modelling
Let us introduce a Cartesian coordinate system with the horizontal axis O χ looking
rightwards such that the left moving wall is located at the point χ = 0 in the absence of
any kind of loading (only the hydrostatic atmospheric pressure that we neglect here). In
this case the wave tank is dry. Now we fill it with still water of the uniform depth h 0 > 0 .
The hydrostatic force F (0) > 0 (defined in equations (2.11)) appears on the left wall
and the springs accumulate some elastic deformation. It results in the wall displacement
of magnitude Θ (0) > 0 in the negative direction of the axis O χ . The displacement Θ (0)
can be found by applying the Hook law [51]:
k Θ (0) = F (0) ,
where k is the springs rigidity. For t > 0 we have another force F (t) acting on the left
wall due to the wave motion. This force acts against the wall. Thus, in the second law of
Newton we take it with the opposite sign [50, 62]:
m χ̈ + k χ = − F (t) ,
where m is the wall mass. The initial conditions for this ODE are:
χ (0) = − Θ (0) ,
χ̇ (0) = 0 .
(D.1)
68 / 72
G. Khakimzyanov & D. Dutykh
Now, we make a change of coordinates 0 χ
O x so that the wall is located initially at
the point x = 0 , i.e.
x = χ + Θ (0) .
Then, the initial conditions become by construction:
x (0) = 0 ,
ẋ (0) = 0 .
Equation of motion (D.1) transforms to:
m ẍ + k x −
The last equation can be rewritten as
k Θ (0) = − F (t) .
| {z }
≡ F (0)
m ẍ + k x = − F (t) − F (0) .
def
Finally, by denoting the wall displacement by s (t) := x (t) we recover equation (2.10).
References
[1] J. B. Bell and G. R. Shubin. An adaptive grid finite difference method for conservation laws.
J. Comp. Phys., 52(3):569–591, dec 1983. 44
[2] C. A. Brebbia. The Boundary Element Method for Engineers. Pentech Press Ltd, London,
2 edition, 1980. 56
[3] C. A. Brebbia, J. C. F. Telles, and L. C. Wrobel. Boundary element techniques: Theory and
applications in engineering. Springer-Verlag, Berlin and New York, 1984. 56
[4] J. C. Butcher. Numerical Methods for Ordinary Differential Equations. Wiley, Chichester,
UK, 3 edition, 2016. 44
[5] F. Carbone, D. Dutykh, J. M. Dudley, and F. Dias. Extreme wave run-up on a vertical cliff.
Geophys. Res. Lett., 40(12):3138–3143, 2013. 5
[6] A.-L. Cauchy. Mémoire sur la théorie de la propagation des ondes à la surface d’un fluide
pesant d’une profondeur indéfinie. Mém. Présentés Divers Savans Acad. R. Sci. Inst. France,
1:3–312, 1827. 47
[7] J. Chambarel, C. Kharif, and J. Touboul. Head-on collision of two solitary waves and residual
falling jet formation. Nonlin. Processes Geophys., 16:111–122, 2009. 50
[8] R. K.-C. Chan and R. L. Street. A computer study of finite-amplitude water waves. J. Comp.
Phys., 6(1):68–94, aug 1970. 53
[9] L. F. Chen, J. Zang, A. J. Hillis, G. C. J. Morgan, and A. R. Plummer. Numerical investigation of wave-structure interaction using OpenFOAM. Ocean Engineering, 88:91–109, sep
2014. 5
[10] M. J. Cooker, P. D. Weidman, and D. S. Bale. Reflection of a high-amplitude solitary wave
at a vertical wall. J. Fluid Mech., 342:141–158, 1997. 5, 50, 51, 53, 63
[11] R. Courant, K. Friedrichs, and H. Lewy. Über die partiellen Differenzengleichungen der
mathematischen Physik. Mathematische Annalen, 100(1):32–74, 1928. 49
[12] A. D. D. Craik. The origins of water wave theory. Ann. Rev. Fluid Mech., 36:1–28, 2004. 10
Wave/moving wall interaction
69 / 72
[13] N. A. Dar’in and V. I. Mazhukin. An approach to adaptive grid construction for
non-stationary problems. USSR Computational Mathematics and Mathematical Physics,
28(2):99–103, jan 1988. (in Russian) 45
[14] N. A. Dar’in, V. I. Mazhukin, and A. A. Samarskii. A finite-difference method for solving
the equations of gas dynamics using adaptive nets which are dynamically associated with
the solution. USSR Computational Mathematics and Mathematical Physics, 28(4):164–174,
jan 1988. (in Russian)
[15] N. A. Dar’in, V. I. Mazhukin, and A. A. Samarskii. A finite difference method for the solution
of 1D gas dynamics on adaptive grids. Doklady Mathematics, 302(5):1078–1081, 1988. (in
Russian) 45
[16] V. H. Davletshin. Force action of solitary waves on vertical structures. In Tsunami meeting,
pages 41–43, Gorky, 1984. Institute of Applied Physics. (in Russian) 51, 53, 63
[17] C. de Boor. Good approximation by splines with variable knots. II. In G. A. Watson, editor,
Conference on the Numerical Solution of Differential Equations, pages 12–20. Springer-Verlag,
Berlin, Heidelberg, 1974. 45
[18] L. M. Degtyarev and T. S. Ivanova. The adaptive-grid method in one-dimensional nonstationary convection-diffusion problems. Differ. Uravn., 29(7):1179–1192, 1993. (in Russian)
45
[19] J. W. Demmel. Applied Numerical Linear Algebra. SIAM, Philadelphia, 1997. 16
[20] F. Dias, D. Dutykh, and J.-M. Ghidaglia. A two-fluid model for violent aerated flows. Comput.
& Fluids, 39(2):283–293, 2010. 5
[21] E. Didier, D. R. C. B. Neves, R. Martins, and M. G. Neves. Wave interaction with a vertical
wall: SPH numerical and experimental modeling. Ocean Engineering, 88:330–341, sep 2014.
5
[22] E. A. Dorfi and L. O. Drury. Simple adaptive grids for 1−D initial value problems. J. Comp.
Phys., 69(1):175–195, mar 1987. 45
[23] D. Dutykh. Visco-potential free-surface flows and long wave modelling. Eur. J. Mech.
B/Fluids, 28(3):430–443, 2009. 63
[24] D. Dutykh, D. Clamond, P. Milewski, and D. Mitsotakis. Finite volume and pseudo-spectral
schemes for the fully nonlinear 1D Serre equations. Eur. J. Appl. Math., 24(05):761–787,
2013. 5
[25] D. Dutykh and F. Dias. Dissipative Boussinesq equations. C. R. Mecanique, 335:559–583,
2007. 63
[26] D. Dutykh and F. Dias. Viscous potential free-surface flows in a fluid layer of finite depth.
Comptes Rendus Mathematique, 345(2):113–118, 2007. 63
[27] D. Dutykh, F. Dias, and Y. Kervella. Linear theory of wave generation by a moving bottom.
C. R. Acad. Sci. Paris, Ser. I, 343:499–504, 2006. 47
[28] A. I. Dyachenko, D. I. Kachulin, and V. E. Zakharov. On the nonintegrability of the free
surface hydrodynamics. JETP Letters, 98(1):43–47, sep 2013. 51
[29] A. I. Dyachenko, E. A. Kuznetsov, M. D. Spector, and V. E. Zakharov. Analytical description
of the free surface dynamics of an ideal fluid (canonical formalism and conformal mapping).
Phys. Lett. A, 221(1-2):73–79, sep 1996. 12
[30] A. I. Dyachenko and V. E. Zakharov. Is free-surface hydrodynamics an integrable system?
Phys. Lett. A, 190(2):144–148, jul 1994. 51
[31] D. V. Evans. Power From Water Waves. Ann. Rev. Fluid Mech., 13(1):157–187, jan 1981. 5
G. Khakimzyanov & D. Dutykh
70 / 72
[32] J. Falnes. Ocean Waves and Oscillating Systems: Linear Interactions Including Wave-Energy
Extraction. Cambridge University Press, Cambridge, 2002. 5
[33] J. Falnes. A review of wave-energy extraction. Marine Structures, 20(4):185–201, oct 2007.
5
[34] J. D. Fenton and M. M. Rienecker. A Fourier method for solving nonlinear water-wave
problems: application to solitary-wave interactions. J. Fluid Mech., 118:411–443, apr 1982.
51, 53, 63
[35] S. K. Godunov. Reminiscences about Difference Schemes. J. Comput. Phys., 153:6–25, 1999.
18
[36] S. K. Godunov and V. S. Ryabenkii. Difference Schemes. North-Holland, Amsterdam, 1987.
18
[37] S. T. Grilli, J. Skourup, and I. A. Svendsen. An efficient boundary element method for
nonlinear water waves. Engng. Anal. Bound. Elem., 6(2):97–107, jun 1989. 63
[38] K. Guizien and E. Barthélemy. Accuracy of solitary wave generation by a piston wave maker.
J. Hydr. Res., 40(3):321–331, may 2002. 5, 62
[39] E. Hairer, C. Lubich, and G. Wanner. Geometric Numerical Integration, volume 31 of Spring
Series in Computational Mathematics. Springer-Verlag, Berlin, Heidelberg, second edition,
2006. 62
[40] G. He and M. Kashiwagi. Nonlinear Solution For Vibration of Vertical Elastic Plate By
Initial Elevation of Free Surface. Int. J. Offshore Polar Eng., 19(3):189–197, 2009. 6, 56, 57,
63
[41] R. S. Johnson. A Modern Introduction to the Mathematical Theory of Water Waves. Cambridge University Press, 2004. 10
[42] Y. Kervella, D. Dutykh, and F. Dias. Comparison between three-dimensional linear and
nonlinear tsunami generation models. Theor. Comput. Fluid Dyn., 21:245–269, 2007. 47
[43] G. Khakimzyanov and D. Dutykh. On supraconvergence phenomenon for second order centered finite differences on non-uniform grids. J. Comp. Appl. Math., 326:1–14, dec 2017.
62
[44] G. S. Khakimzyanov, D. Dutykh, D. E. Mitsotakis, and N. Y. Shokina. Numerical solution
of conservation laws on moving grids. Submitted, pages 1–28, 2016. 44, 62
[45] G. S. Khakimzyanov, Y. I. Shokin, V. B. Barakhnin, and N. Y. Shokina. Numerical Simulation of Fluid Flows with Surface Waves. Sib. Branch, Russ. Acad. Sci., Novosibirsk, 2001.
(in Russian) 44
[46] G. S. Khakimzyanov and N. Y. Shokina. Equidistribution method for the construction of
adaptive grids. Russ. J. Numer. Anal. Math. Modelling, 14(4):339–358, 1999. 45
[47] E. Knobloch and R. Krechetnikov. Problems on Time-Varying Domains: Formulation, Dynamics, and Challenges. Acta Appl. Math., 137(1):123–157, jun 2015. 5, 62
[48] A. A. Korobkin, S. V. Stukolov, and I. V. Sturova. Motion of a vertical wall fixed on springs
under the action of surface waves. J. Appl. Mech. Tech. Phys., 50(5):841–849, sep 2009. 6,
57
[49] A. A. Kurkin, S. V. Semin, and Y. A. Stepanyants. Transformation of surface waves over a
bottom step. Izvestiya, Atmospheric and Oceanic Physics, 51(2):214–223, mar 2015. 5
[50] L. D. Landau and E. M. Lifshitz. Mechanics. Elsevier Butterworth–Heinemann, Amsterdam,
1976. 10, 67
[51] L. D. Landau and E. M. Lifshitz. Theory of Elasticity. Elsevier Sciences, Oxford, 1986. 67
Wave/moving wall interaction
71 / 72
[52] M. A. Lavrent’ev and B. V. Shabat. Hydrodynamics problems and their mathematical models.
Nauka, Moscow, 2 edition, 1977. (in Russian) 7, 12
[53] C. M. Linton and P. McIver. Mathematical Techniques for Wave/Structure Interactions.
Chapman & Hall/CRC, 2001. 5
[54] K. Lipnikov, G. Manzini, and M. Shashkov. Mimetic finite difference method. J. Comp.
Phys., 257:1163–1227, jan 2014. 28
[55] O. S. Madsen. On the generation of long waves. J. Geophys. Res., 76(36):8672–8683, dec
1971. 53
[56] S. V. Manoylin. Some experimental and theoretical methods of estimation of tsunami wave
action on hydro-technical structures and seaports. Technical report, Siberian Branch of
Computing Center, Krasnoyarsk, 1989. (in Russian) 53
[57] T. Maxworthy. Experiments on collisions between solitary waves. J. Fluid Mech, 76:177–185,
1976. 50, 53
[58] V. I. Mazhukin, A. A. Samarskii, O. Castellanos, and A. V. Shapranov. The method of
dynamic adaptation for non-stationary problems with large gradients. Math. Model., 5(4):32–
56, 1993. (in Russian) 45
[59] C. C. Mei. The applied dynamics of ocean surface waves. World Scientific, 1994. 47
[60] S. M. Mirie and C. H. Su. Collision between two solitary waves. Part 2. A numerical study.
J. Fluid Mech., 115:475–492, 1982. 50
[61] N. Mori, T. Takahashi, T. Yasuda, and H. Yanagisawa. Survey of 2011 Tohoku earthquake
tsunami inundation and run-up. Geophys. Res. Lett., 38(7), apr 2011. 6
[62] I. Newton. Philosophiae Naturalis Principia Mathematica. Imprimatur. s. Pepys Reg. Soc.
Praeses, London, 1687. 10, 67
[63] J. Orszaghova, A. G. L. Borthwick, and P. H. Taylor. From the paddle to the beach - A
Boussinesq shallow water numerical wave tank based on Madsen and Sorensen’s equations.
J. Comp. Phys., 231(2):328–344, jan 2012. 5, 53
[64] L. V. Ovsyannikov. To the shallow water theory foundation. Arch. Mech., 26:407–422, 1974.
12
[65] L. V. Ovsyannikov. Parameters of cnoidal waves. In Problems of Mathematics and Mechanics,
M. A. Lavrentiev Memorial Volume, pages 150–166. Nauka, Novosibirsk, 1983. (in Russian)
50
[66] S. D. Poisson. Mémoire sur la théorie des ondes. Mém. Acad. R. Sci. Inst. France, 1816(1):70–
186, 1818. 47
[67] B. T. Polyak. The conjugate gradient method in extremal problems. USSR Computational
Mathematics and Mathematical Physics, 9(4):94–112, jan 1969. (in Russian) 37
[68] R. D. Richtmyer and K. W. Morton. Difference Methods for Initial-Value Problems. Interscience Publishers, New York, 1967. 18
[69] A. A. Samarskii. The Theory of Difference Schemes. CRC Press, New York, 2001. 16, 37
[70] A. A. Samarskii, V. F. Tishkin, A. P. Favorskii, and M. Y. Shashkov. Operational finitedifference schemes. Differential Equations, 17(7):854–862, 1981. 28
[71] Y. I. Shokin and A. I. Urusov. On the construction of adaptive algorithms for unsteady
problems of gas dynamics in arbitrary coordinate systems. In Eighth International Conference
on Numerical Methods in Fluid Dynamics, pages 481–486. Springer, Berlin, Heidelberg, 1982.
44
[72] Y. I. Shokin and N. N. Yanenko. Method of Differential Approximation. Application to Gas
Dynamics. Nauka, Novosibirsk, 1985. (in Russian) 44
G. Khakimzyanov & D. Dutykh
72 / 72
[73] J. J. Stoker. Water Waves: The mathematical theory with applications. Interscience, New
York, 1957. 5, 8, 10, 47, 62
[74] C. H. Su and R. M. Mirie. On head-on collisions between two solitary waves. J. Fluid Mech.,
98:509–525, 1980. 50, 51, 53
[75] A. N. Tikhonov and A. A. Samarskii. Equations of Mathematical Physics. Dover Publications,
Inc., New York, 1963. 18
[76] K.-H. Wang, Z. Dai, and H. S. Lee. Modeling Wave Run-up along a Sloping or a Moving
Wall Boundary. J. Coast. Res., 27(6):1159–1169, nov 2011. 6
[77] S. A. Yang and A. T. Chwang. An experimental study of nonlinear waves produced by an
accelerating plate. Phys. Fluids A, 4(11):2456, 1992. 6
[78] D. M. Young. Iterative Methods for Solving Partial Difference Equations of Elliptic Type.
PhD, Harvard University, 1950. 40
[79] N. N. Zagryadskaya, S. V. Ivanova, L. S. Nudner, and A. I. Shoshin. Action of long waves
on a vertical obstacle. Bulletin of VNIIG, 138:94–101, 1980. (in Russian) 53
[80] P. A. Zegeling and J. G. Blom. An evaluation of the gradient-weighted moving-finite-element
method in one space dimension. J. Comp. Phys., 103(2):422–441, dec 1992. 45
G. Khakimzyanov: Institute of Computational Technologies, Siberian Branch of the
Russian Academy of Sciences, Novosibirsk 630090, Russia
E-mail address: [email protected]
URL: http://www.ict.nsc.ru/ru/structure/Persons/ict-KhakimzyanovGS
D. Dutykh: LAMA, UMR 5127 CNRS, Université Savoie Mont Blanc, Campus Scientifique, F-73376 Le Bourget-du-Lac Cedex, France
E-mail address: [email protected]
URL: http://www.denys-dutykh.com/
| 5cs.CE
|
1
STOCHASTIC VOLATILITY MODELS INCLUDING OPEN, CLOSE, HIGH AND LOW
PRICES
ENRIQUE TER HORST, ABEL RODRIGUEZ, HENRYK GZYL, AND GERMAN MOLINA
A BSTRACT. Mounting empirical evidence suggests that the observed extreme prices within a trading
period can provide valuable information about the volatility of the process within that period. In this
arXiv:0901.1315v1 [q-fin.ST] 9 Jan 2009
paper we define a class of stochastic volatility models that uses opening and closing prices along with the
minimum and maximum prices within a trading period to infer the dynamics underlying the volatility
process of asset prices and compares it with similar models that have been previously presented in the
literature. The paper also discusses sequential Monte Carlo algorithms to fit this class of models and
illustrates its features using both a simulation study and data form the S&P500 index.
1. I NTRODUCTION
Understanding the volatility of financial assets is a key problem in econometrics and finance. Over
the last two decades, the literature that deals with the modelling of financial asset volatility has expanded significantly. Volatility, a latent variable as such, has been shown to be highly persistent,
and all kinds of models have been proposed to extract that persistence and incorporate it into the investment decision-making process, and is of especial interest in the derivatives world, as one of the
key inputs for pricing. From the early Autoregressive Conditional Heteroskedastic (ARCH) model of
Engle (1982) and the generalized version (GARCH), offered by Bollerslev (1986), to the stochastic
volatility approaches (Shephard, 2005), the literature has expanded exponentially, bringing all kinds
1
Enrique ter Horst is Assistant Professor, Instituto de Estudios Superiores de Administración, Caracas 1010, Venezuela,
[email protected]. Abel Rodriguez is Assistant Professor, Department of Applied Mathematics and
Statistics, University of California, Santa Cruz, [email protected]. Henryk Gzyl is Professor, Instituto de Estudios
Superiores de Administración, Caracas 1010, Venezuela, [email protected]. German Molina is Quantitative
Trading Researcher at a hedge fund, London, UK, [email protected].
Key words and phrases. Stochastic Volatility; Functionals of Brownian motion; Maximum; Minimum; High; Low; Open;
Close; Particle Filter; Bayesian Statistics, Sequential Monte Carlo.
1
of adaptations to particular cases (fractional integration, volatility-in-mean equations, etc.), or expansions into the multivariate space and applications in asset management, derivatives pricing and risk
management. Most of them share the common feature of taking closing prices over the frequency
considered to do model fitting and forecasting.
More recently some structural variations have been proposed. Some of them allow for more flexible
uses of the data available. (Russell & Engle, 1998), use non-equally spaced observations to extract
volatility features. (Andersen et al., 2001a; Barndorff-Nielsen & Shephard, 2002) use data at higher
frequencies to provide better estimates of the volatility and its persistence at the lower ones. These
original papers, and the expansions and advances that followed them, tend to focus on the use of larger
amounts of the information available to achieve the common goal of better fitting and forecasting, but
still tend to focus their efforts on closing prices at the highest frequency considered, ignoring the paths
between those points, even if this information is available.
When it comes to data availability, there have been major changes in the last two decades. From
having only (if that) access to daily or weekly closing prices, to having tick-by-tick/bid-ask data
available for many series. The increased availability of information must also induce new classes of
models and algorithms that tackle that, and that combine the best possible use of that information
with an efficient way to process it, so that they remain practical and useful for practitioners. This
manuscript tries to achieve these purposes by adding three commonly available series to the statistical
process. We continue using closes over the frequency chosen, but add to them the open, high and low
data as well, over that period and frequency considered. This data is available for most series through
the common data providers (Datastream, Reuters, Bloomberg), and it provides several advantages.
First, the use different sources of information (order statistics together with start/end points) about
the movement of the underlying, which help to better understand and make inference on its volatility
levels and dynamics. The size of the intra-period swings is very informative, and the information
content can potentially be very different from that available from the close-to-close data only. Second,
2
adding that information does not have to be a computational burden. We can still provide a quick
model that meets the requirements of speed of fitting, while using a larger set of information. Third,
because it opens the door to alternative models that could take advantage of the information embedded
in those added data features. For example, modifications in the elicitation of the leverage effect in
the likelihood, which could potentially be modeled as a relationship between not only return and
volatility, but downside range and volatility. Although we do not explore these improvements in this
paper, it is worth mentioning that the flexibility provided by the use of this data is there.
The relevance of Close, High, Low and Open (CHLO) data, that is, the relevance of the path
followed by the series rather than the start and end points, is well-known. Financial newspapers include it for the frequency reported (daily, weekly, monthly or yearly), data providers add it to their
data series, and practitioners in technical analysis have been using it in the past to model volatility
by using estimators like ranges or average true ranges Wilder (1978). Chartists have also been using CHLO data as a key source of information from a graphical/visual standpoint to identify price
patterns, trends or reversals. By assuming that log-prices follow a log-normal diffusion, Parkinson
(1980) showed that high/low prices provide a highly efficient estimator of the variance compared to
estimators based solely on open/close prices. By incorporating information from CHLO prices, Garman & Klass (1980) extended the estimator of Parkinson and gained a significant amount of efficiency
compared to only including open/close prices. Ball & Tourus (1984) derived a maximum likelihood
analogue of the estimator of Parkinson. In the context of time series modeling of asset return volatility,
Gallant et al. (1999) and Alizadeh et al. (2002) take one very relevant step forward by using the range
of observed prices (rather than the maximum and minimum directly) to estimate stochastic volatility
models. Brandt & Jones (2005) extend the work in Alizadeh et al. (2002) by including the range and
the closing price in the estimation of a stochastic volatility model, but again ignore the actual levels
of the maximum and minimum prices.
3
Full CHLO prices have been used by a number of authors, including Rogers & Satchell (1991),
Rogers et al. (1994) and Rogers (1998). In particular, Magdon & Atiya (2003) derives a maximum
likelihood estimator assuming constant volatility, obtaining better performance than existing previous
methods on simulated data. Their method, however, is not integrated into a (G)ARCH or stochastic
volatility framework, something done by Lildholdt (2002) from a maximum likelihood perspective.
We introduce a Bayesian stochastic volatility model that uses full CHLO prices and develop a particle
filter approach for infrerence.
Closing prices, especially at the lower frequencies (daily or lower), have been less and less reliable
to ascertain realized volatilities based on them, given the patterns recently seen in volumes. For
example, the S&P shows unusually large volumes of trades in the last 15 minutes of the sessions.
They have been linked to several factors, like high frequency funds unwinding positions accumulated
during the day, or exchange-traded funds and hedgers operating in the last minutes to adjust their
positions. All of those flows push prices towards outside the extremes observed during the day.
However, the actual behavior of the underlying for volatility (risk) purposes is more extreme. Stoploss/profit-taking levels, as well as positions sizings and their dynamics, are usually defined based
on both technical levels and volatility measurements. If those volatilities are based on extremes and
paths as well as close-to-close levels, they can potentially lead to very different levels than those based
only on close-to-close levels, this applying to any frequency of investment. Intra-day ranges offer, a
cleaner, more liquidity-adjusted picture of volatility. Volatility can be very large and still end up with
closing prices not far from where they started, while big swings have happened throughout the day.
This is especially the case during periods of low liquidity, where the average holding period of any
position diminishes significantly, and people react more violently to moves. Stop-loss orders can
exacerbate those intra-day swings, while squaring of daily positions for intra-day funds can produce
the opposite effect. This is reflective of higher realized volatility and lower liquidity, but it will not be
captured by models that only use close-to-close returns.
4
Our work combines the use of CHLO data with a Bayesian approach, all in a stochastic volatility framework. However, we use a particle filter algorithm to do the filtered estimation, which is
computationally quick and can be applied to any frequency of data; which is of special relevance to
practitioners handling large amounts of data at the higher frequencies or for processing large numbers
of series (or when a quick decision is needed, even if operating with lower frequency data). Section 2
provides a quick introduction to the theory behind stochastic volatility models and an introduction to
the notation used throughout the paper. Sections 3 and 4 present the analytical framework for the joint
density of the CHLO prices, the elicitation of prior densities and a description of the particle filtering
algorithm used. Sections 5 and 6 show how to deal with problems of missing data, and applies our
model to weekly CHLO data of the S&P. We also provide a comparison with the stochastic volatility
models using only closing prices data and show how different the results can be once added the information contained in the observed extremes. Section 7 concludes with a summary and a description of
potential applications and extensions.
In summary, our net contribution is threefold. First, we provide a coherent model that links the traditional stochastic volatility model with the CHLO data without the need of assumptions beyond those
used in that traditional stochastic volatility model. Second, we provide a quick and simple algorithm
that allows for fast estimation of this model, which should be very appealing to practitioners. Third,
by having changed only the observation equation, we show that these changes can be embedded in
any model that uses any other types of evolution equations.
2. S TOCHASTIC VOLATILITY MODELS
Because of its simplicity and tractability, geometric Brownian motion is by far the most popular
model to explain the evolution of the price of financial assets, and has a history dating back to Louis
Bachelier (Courtault et al., 2000). A stochastic process {St : t ∈ R+ } is said to follow a Geometric
5
Brownian motion (GBM) if it is the solution of the stochastic differential equation
(1)
dSt = µSt dt + St σdBt
where Bt is a standard Wiener process and µ and σ are, respectively, the instantaneous drift and
instantaneous volatility of process. GBM implies that the increments of yt = log St over intervals of
the same length are independent, stationary and identically distributed, i.e., yt+∆ − yt = log St+∆ −
log St ∼ N(∆µ, ∆σ 2 ), or equivalently,
St+∆ = St exp {∆µ + σ(B(t + ∆) − B(t))}
By construction, GMB models assume that the volatility of returns is constant. However, empirical
evidence going back at least to Mandelbrot (1963), Fama (1965) and Officer (1973) demonstrates that
the price volatility of financial assets tends to change over time and therefore the simple model in (1)
is generally too restrictive. The GBM model can be generalized by assuming that the both the price
and volatility processes follow general diffusions,
dSt = µ(St , σt )dt + ν(St , σt )dWt
dσt = α(St , σt )dt + β(St , σt )dDt
One particularly simple (and popular) version of this approach assumes that, just as before, the price
follows a GBM with time varying drift, and that the log-volatility follows an Ornstein-Uhlenbeck
(OU) processes (Uhlenbeck & Ornstein, 1930),
(2)
(3)
dYt = µdt + σt dBt
d log(σt ) = κ(ψ − log(σt )) + τ dDt
6
Practical implementation of these models typically relies on a discretization of the continuos time
model in (2) and (3). For the remainder of the paper, we assume that the drift and volatility are scaled
so that ∆ = 1 corresponds to one trading period (e.g., day or week) and we focus attention on the
state-space model
(4)
(5)
yt = µ + yt−1 + 1t
1t ∼ N(0, σt2 )
log(σt ) = α + φ[log(σt−1 ) − α] + 2t
2t ∼ N(0, τ 2 ),
where α = κψ and φ = 1 − κ. It is common to assume that 0 ≤ φ < 1 and log(σ0 ) ∼
N(α, τ 2 /(1 − φ2 )), so that volatilities are positively correlated and the volatility process is stationary
(which ensures that the process for the prices is a martingale) with α determining the median of the
long-term volatility, ν = exp{α}. Therefore, we expect the volatility of returns to take values greater
than ν half of the time, and viceversa.
Unlike ARCH and GARCH models (Engle, 1982; Bollerslev, 1986), where a single stochastic process controls both the evolution of the volatility and the observed returns, stochastic volatility models
use two coupled processes to explain the variability of the returns. By incorporating dependence
between 1t and 2t , the model can accommodate leverage effects, while additional flexibility can be
obtained by considering more general processes for the volatility (for example, higher order autoregressive process, jump process, and linear or nonlinear regression terms).
Although theoretical work on stochastic volatility models goes back at least to the early 80s, practical application was limited by computational issues. Bayesian fitting of stochastic volatility models
has been discussed by different authors, including Jacquier et al. (1994) and Kim et al. (1998). Popular
approaches include Gibbs sampling schemes that directly sample from the full conditional distribution of each parameter in the model, algorithms based on offset mixture representations that allow for
joint sampling of the sequence of volatility parameters, and particle filter algorithms. In the sequel,
7
we concentrate on sequential Monte Carlo algorithms for the implementation of stochastic volatility
models.
3. I NCORPORATING EXTREME VALUES IN STOCHASTIC VOLATILITY MODELS
3.1. Joint density for the closing price and the observed extremes of geometric Brownian motion. The goal of this section is to extend the stochastic volatility model described in the previous
Section to incorporate information on the full CHLO prices. To do so, note that the Euler approximation in (4) implies that, conditionally on the volatility σt , the distribution of the increments of the asset
price follows a Geometric Brownian motion with constant volatility σt during the t-th trading period.
Therefore, the discretization allows us to interpret the process generating the asset prices as a a sequence of conditionally independent processes defined over disjoint and adjacent time periods; within
each of these periods the price process behaves like a GBM with constant volatility, but the volatility
is allowed to change from period to period. This interpretation of the discretized process is extremely
useful, as it allows us to derive the joint distribution for the closing, maximum and minimum price of
the asset within a trading period using standard stochastic calculus tools.
Theorem 1. Let Yt be a Brownian motion with drift and consider the evolution of the process over
a time interval of unit length where Yt−1 and Yt are the values of the process at the beginning and
end of the period, and let Mt = supt−1≤s≤t {Ys } and mt = inf t−1≤s≤t {Ys } be, respectively, the
supremum and the infimum values of the process over the period. If we denote by µ and σt the drift
and volatility of the process between t − 1 and t (which are assumed to be fixed within this period),
the joint distribution of Mt , mt and Yt conditional on Yt−1 = yt−1 is given by
Z
ct
Pr(mt ≥ at , Mt ≤ bt , Yt ≤ ct |Yt−1 = yt−1 ) =
q(yt , at , bt |yt−1 )dyt
−∞
8
for mt ≤ min{ct , yt−1 }, Mt ≥ max{ct , yt−1 } and mt ≤ Mt , where
(6)
∞
1
[µ2 − 2µ(yt − yt−1 )] X
q(yt , at , bt |yt−1 ) = √
exp −
(exp {−d1 (n)} − exp {−d2 (n)})
2σt2
2πσt
n=−∞
and
d1 (n) =
[yt − yt−1 − 2n(bt − at )]2
2σt2
d2 (n) =
[yt + yt−1 − 2at − 2n(bt − at )]2
2σt2
The proof of this theorem, which is a simple extension of results in Feller (1971), Dym et al. (1985)
and Klebaner (2005), can be seen in Appendix A. An equivalent but more involved expression was
obtained by Lildholdt (2002), who used it to construct GARCH models that incorporate information
on CHLO prices. Note that if we take both at → −∞ and bt → ∞, the cumulative distribution in (6)
reduces to the integral of a Gaussian density with mean yt−1 + µ and variance σt2 , which agrees with
(1).
Corollary 1. The joint density for mt , Mt , Yt |Yt−1 = yt−1 is given by
(7)
∂ 2q
q(yt , at , bt |yt−1 )
∂at ∂bt
2
µ − 2µ(yt − yt−1 )
1
exp −
=√
×
2σt2
2πσt3
p(at , bt , yt |yt−1 ) = −
∞
X
4n2 (2d1 (n) − 1) exp {−d1 (n)} − 4n(n − 1) (2d2 (n) − 1)) exp {−d2 (n)}
n=−∞
for mt ≤ min{yt , yt−1 }, Mt ≥ max{yt , yt−1 } and mt ≤ Mt , and zero otherwise.
Equation (7) provides the likelihood function for the closing, maximum and minimum prices given
the volatility and drift, under the first order Euler approximation to the system of stochastic differential
equations in (2) and (3). Therefore, the basic underlying assumptions about the behavior of asset
prices are the same as in standard volatility models; however, by employing (7) instead of (4) we are
9
able to coherently incorporate information about the observed price extremes in the inference of the
price process. After a simple transformation rt = bt − at and wt = at and marginalization over wt ,
we recover the likelihood function described in Brandt & Jones (2005),
(yt − yt−1 − µ)2
1
exp −
p(rt , yt |yt−1 ) = √
2σt2
2πσt
∞
X
4n2
(2nrt − |yt − yt−1 |)2
(2nrt − |yt − yt−1 |)2
√
− 1 exp
2
2
σ
2σt2
2πσ
t
t
n=−∞
(8)
(2nrt − |yt − yt−1 |)2
2n(n − 1)
(2nrt − |yt − yt−1 |) exp
+ √
2σt2
2πσt2
2n(n − 1)
(2(n − 1)rt + |yt − yt−1 |)2
+ √
(2(n − 1)rt + |yt − yt−1 |) exp
2σt2
2πσt2
for rt > |yt − yt−1 |, while a further marginalization over the closing price yt yields the (exact)
likelihood underlying the range-based model in Alizadeh et al. (2002), which is independent of the
opening price yt−1 ,
(9)
2 2
∞
2
X
nr
n−1 n rt
exp − 2t
p(rt ) = 8
(−1) √
2σt
2πσt
n=1
Using either (8) or (9) as likelihoods entails a loss of information with respect to the full joint
likelihood in (7). In the case of (9), the effect is clear as the range is an ancillary statistic for the drift
of the diffusion, and therefore provides no information about it. This leads Alizadeh et al. (2002) to
assume that the drift is zero, which has little impact in the estimation of the volatility of the process,
but might have important consequences for other applications of the model, such as option pricing.
In the case of (8), although there is information about the drift contained in the opening and closing
prices, the model ignores the additional information about the drift contained in the actual levels of
the extremes.
In order to emphasize the importance of the information provided by the minimum and maximum
returns, we present in Figure 1 plots of the likelihood function for the first observation of the S&P500
10
−5
−12
−11
−10
−9
log(σ)
−8
−7
−6
−5
−6
−7
−8
−9
−12
−11
−10
log(σ)
−0.02
−0.01
0.00
0.01
0.02
−0.02
µ
−0.01
0.00
0.01
0.02
µ
F IGURE 1. Likelihood functions for a single observation. The left panel corresponds
to the Gaussian likelihood obtained solely from the closing price (equation (4)), while
the right panel corresponds to the joint likelihood (equation (7)).
dataset discussed in Section 6.1 (the week ending on April 21, 1997). When only the closing price is
available, the likelihood provides information about the drift of the process, but not about the volatility
(note that in this case the likelihood is unbounded in a neighborhood of σt = 0, and therefore the
maximum likelihood estimator does not exist). Therefore, information about the volatility in this type
of model is obtained solely through the evolution of prices, and is therefore strongly influenced by the
underlying smoothing process. In other words, the volatility parameters are only weakly identifiable,
with the identifiability begin provided by the autoregressive prior in (5). In practice, this means that
formally comparing alternative models for the volatility is extremely difficult. However, when the
maximum and minimum are included in the analysis, the likelihood for a single time period does
provide information about the volatility in that period, which can greatly enhance our ability to infer
and test volatility models.
11
Although dealing with an infinite sum can seem troublesome, our experience suggests that a small
number of terms (less than 20) suffice to provide an accurate approximation. In addition, we note that
since the general term of the sum is strictly decreasing for all the likelihoods discussed above, it is
easy to implement an adaptive scheme that stops adding terms once the change in the value is smaller
than a given tolerance.
Note that equation (6) assumes that the log closing price at period t − 1, yt−1 is the same as the
log opening price at period t (call this xt ). However, in some cases these two prices can differ;
for example, wars or unexpected news can happen during holidays whenever markets are closed
provoking jumps that are not governed by the diffusion processes. As with other stochastic volatility
models, we can easily include this type of “weekend effect” by realizing that equations (6) and (7)
use yt−1 simply as the initial value of the GBM; therefore, they remain valid if we substitute yt−1 by
xt .
3.2. Variance evolution and prior specification. The previous discussion focused on the characteristics of the likelihood function for the discretized process conditional on its volatility. The full
specification of the model also requires that we define the evolution of the volatility in time. For
illustrative purposes, this paper focuses on the simple autoregressive process for the log volatility
described in Section 2,
(10)
log(σt )| log(σt−1 ) ∼ N α + φ[log(σt−1 ) − α], τ
2
log(σ0 ) ∼ N α,
τ2
1 − φ2
However, we would like to stress that more complex stochastic processes can easily be incorporated in this models without significantly increasing the computational complexity. For example,
stationary AR(p) processes for the log volatility (which can potentially provide information on quasiperiodicities in the volatility process) can be introduced using the prior specification discussed by
12
Huerta & West (1999). Similarly, including volatility jumps is straightforward using Markov switching models (Carvalho & Lopes, 2007), as well as estimating the parameters of stochastic volatility
models with several factors varying at different time scales (Molina et al., 2009).
The model is completed by introducing priors for the unknown structural parameters in the models.
Following standard practice, we assume that
(11)
µ ∼ N(dµ , Dµ )
α ∼ N(dα , Dα )
φ ∼ Beta(qφ , rφ )
τ 2 ∼ IG(uτ , vτ ).
This choice of priors ensures that all parameters have the right support; in particular, the beta prior
for φ ensures that 0 ≤ φ ≤ 1 (remember our discussion in Section 2). The eight hyperparametes dµ ,
Dµ , dα , Dα , qφ , rφ , uτ and vτ have to be chosen according to the available prior information about
the problem at hand. For example, it is natural to choose dµ to be close to the market risk-free rate,
while choosing dα close the logarithm of the long-term average volatility for the asset. In this paper,
we avoid noninformative priors because the sequential Monte Carlo algorithms we employ for model
fitting (see Section 4 below) require proper priors in order to be implemented.
4. I NFERENCE USING PARTICLE FILTERS
The use of simulation algorithms to explore the posterior distribution of complex Bayesian hierarchical models has become popular in the last 20 years. In particular, Markov chain Monte Carlo
(MCMC) algorithms, which generate a sequence of dependent samples from the posterior distribution
of interest, have become ubiquitous. For inference in non-linear state space models, sequential Monte
Carlo algorithms, and in particular particle filters, have become a standard tool. Particle filters use
a finite discrete approximations to represent the distribution of the state parameters at time t given
the observations up to t, which in our case reduces to p(σt |{yl , bl , al }tl=1 )), and sequentially updates
it to obtain p(σt+1 |{yl , bl , al }t+1
l=1 ). As with other Monte Carlo approaches to inference, the resulting
samples can be used to obtain point and interval estimates, as well as to test hypothesis of interest.
13
However, unlike MCMC algorithms, there is no need to check for convergence of the algorithm, study
its mixing properties, or devising proposal distributions. Doucet et al. (2001) provides an excellent
introduction to sequential Monte Carlo Methods.
Most sequential Monte Carlo algorithms are unable to handle structural parameters that do not
evolve in time, and assume them fixed and known in advance. However, in our stochastic volatility model structural parameters such as µ, α, φ and τ 2 are unknown and need to be estimated from
the data. Therefore, in the sequel we concentrate on a version of the auxiliary particle filter (Pitt &
Shephard, 1999) developed by Liu & West (2001) for exactly this purpose. Their algorithm introduces an artificial Gaussian perturbation in the structural parameters and applies an auxiliary particle
filter to the modified problem. In order to correct for the information loss generated by the artificial
perturbation, the authors introduce a shrunk kernel approximation constructed in such a way as to
preserve the mean and covariance of the distribution of the structural parameters. This kernel density
approximation usually works well in practice, producing accurate reconstructions of the posterior distribution of both structural and state parameters while avoiding loss of information that plagues the
self-organizing state space models in Kitagawa (1998).
Since the artificial evolution is assumed to follow a Gaussian distribution, it is typically necessary to transform the structural parameters so that the support of their distribution is the whole
real line. Therefore, we describe our algorithm in terms of the transformed structural parameter
η = (η1 , η2 , η3 , η4 ) = (µ, α, logit(φ), log(τ 2 )) ∈ R4 . After choosing a discount factor 0.5 < < 1
(controlling both the size of the perturbation and the level of shrinkage in the density estimator) and
generating a sample of N particle from the prior distributions in (11), the algorithm proceeds by
repeating the following steps for t = 1, . . . , T
14
(j)
(1) For each particle j = 1, . . . , N , identify prior point estimates (zt+1 , mjt+1 ) for the joint vector
(j)
(j)
of state and structural parameters (σt+1 , ηt+1 ) such that,
(j)
(j)
(j)
(j)
(j)
zt+1 = exp{η2,t + expit(η3,t )[log(σt ) − η2,t ]}
(j)
(j)
mt+1 = aηt + (1 − a)η̄t
where a = (3 − 1)/(2) and η̄t =
(1)
(j)
PN
ηt .
j=1
(N )
(2) Sample auxiliary indicators ξt , . . . , ξt
, each one also taking values in the set {1, . . . , N },
so that
(j)
k
Pr(ξt = k) = p(yt+1 , bt+1 , at+1 |yt , µ = mk1,t+1 , σ = zt+1
)
where p(yt+1 , bt+1 , at+1 |yt , µ, σt2 ) is given in (7).
j
(3) Generate a set of new structural parameters ηt+1
by sampling
ξ
(j)
(j)
t
ηt+1 ∼ N(mt+1
, (1 − a2 )Vt )
where Vt =
(j)
j=1 (ηt
PN
(j)
− η̄t )0 (ηt − η̄t )/N .
(4) Sample a value of the current state vector
(j)
(j)
(j)
(ξt )
(j)
log(σt+1 ) ∼ N(η2,t+1 + expit(η3,t+1 )[log(σt
(j)
(j)
) − η2,t+1 ], exp(η4,t+1 ))
(5) Resample the particles according to probabilities
(j)
ωt+1
∝
j
k
p(yt+1 , bt+1 , at+1 |yt , µ = η1,t+1
, σ = σt+1
)
j
p(yt+1 , bt+1 , at+1 |yt , µ = mj1,t+1 , σ = zt+1
)
Although particle filters are not iterative algorithms and therefore mixing and convergence are
not issues, sequential Monte Carlo algorithms might suffer from particle impoverishment. Particle
impoverishment happens when the particle approximation at time t differs significantly from the
approximation at time t + 1, leading to a small number of particles receiving most of the posterior
15
weight; in the worst case, a single particle receives all the posterior weight. A similar issue arises
in importance sampling algorithms, where efficiency decreases as the distribution of the importance
weights becomes less uniform. In the sequential Monte Carlo literature it is common to use the
effective sample size (ESS) to monitor particle impoverishment,
ESSt =
N
1+
V(ωt )
[E(ωt )]2
Values of the ESS close to N point to well behaved samplers with little particle impoverishment,
while small values of N usually indicate uneven weights and particle representations that might be
missing relevant regions of the parameter space.
5. M ISSING DATA
In some instances (for example, non-exchange traded assets), data on the observed extremes might
be unavailable or unreliable for some trading periods. When MCMC algorithms are used for inference, their structure can be easily exploited to deal with this type of situation by adding an additional
sampling step in which the missing or unreliable data is imputed conditionally on the current value of
the parameters. However, this type of iterative procedure is not available in particle filter algorithms
such as the one described in Section 4. This means that implementation under missing data requires
that we compute the corresponding marginal distributions from the joint density derived from (6),
which are to be used in steps (2) and (5) of the particle filter instead of (6).
As we discussed in Section 3, the marginal density of the log closing price is simply a normal
distribution with mean µ and variance σt2 . Therefore, if both the maximum and minimum are missing,
the likelihood for the period simply becomes the likelihood for a standard stochastic volatility model
in (4). When the minimum is not available, the marginal density for the maximum and the closing
16
price is simply given by
2(2bt − (yt + yt−1 ))
(2bt − (yt + yt−1 ))2 µ(yt − yt−1 ) µ2 t
p
q(yt , bt |yt−1 ) =
+
− 2 ,
exp −
2σt2 t
σt2
2σt
2πσt6 t3
while if the maximum is missing, the marginal density for the minimum and the closing is given by,
2(yt + yt−1 − 2at )
(yt + yt−1 − 2at )2 µ(yt − yt−1 ) µ2 t
p
+
− 2
q(yt , at |yt−1 ) =
exp −
2σt2 t
σt2
2σt
2πσt6 t3
The density of the maximum and minimum can be computed as well by integrating equation (7) from
a to b with respect to the variable yt . Proof of these results can be seen Dana et al. (2007), or can be
directly obtained by computing the corresponding limits on expression (7).
6. I LLUSTRATIONS
6.1. Simulation study. In this section we use a simulation study to compare the performance of four
stochastic volatility models: STSV, which uses the Gaussian likelihood in (1) and therefore employs
solely the information contained in the opening and closing prices; RASV, which uses the likelihood
(9) and represents a Bayesian version of the range-only model described by Alizadeh et al. (2002);
RCSV, which uses the likelihood in (8) and therefore corresponds to the model based on the range and
the opening/closing prices described in Brandt & Jones (2005); and EXSV, our proposed model using
the full likelihood (7) and employing all the information contained in the opening/closing/high/low
prices.
Our simulation study uses 100 random samples, each comprising 156 periods of returns generated under the stochastic volatility model described in equations (4) and (5). First, we generate the
sequence of volatilities using the Ornstein-Uhlenbeck model in (5), with parameters α = −3.75,
φ = 0.9 and τ = 0.11 (note that these values correspond to the means of the prior distributions used
for the analysis of the weekly S&P500 data in Section 6.2 below). Then, for each period, we generate
17
Models
RMSD
MAD
STSV/RASV 1.43 (1.21, 1.64) 1.46 (1.05, 1.82)
RASV/RCSV 1.02 (0.90, 1.13) 0.99 (0.85, 1.16)
RCSV/EXSV 1.06 (0.92, 1.22) 1.11 (0.96, 1.26)
RASV/EXSV 1.07 (0.97, 1.18) 1.10 (0.93, 1.27)
TABLE 1. Results from our simulation example. We show the median along with the
5% and 95% quantiles (in parenthesis) of the ratio between deviation measures for
three pairs of models, computed over a total of 100 simulated data sets.
a sample path for the geometric Brownian motion in (4) over a grid with 1000 nodes using the corresponding value for the volatility, assuming that µ = 0.000961 (for weekly returns, this corresponds
approximately to an average 5% annual return). The maximum, minimum, opening and closing prices
over the period were computed from this sample path. We assume that the value of the assets at the
beginning of each simulation S0 is $100 and compute both the root mean squared deviation (RMSD)
of the estimated volatility with respect to their true value, and their median absolute deviation (MAD).
The particle filter algorithm described in Section 4 was used to fit all four models, and was implemented in MATLAB using 30,000 particles. Unless noted otherwise, the results we report below
correspond to prior hyperparameters dµ = 0, Dµ = 0.0001, qφ = 9 and rφ = 1, dα = −3.75,
Dα = 0.025 and uτ = 6 vτ = 0.06, so that the prior distributions are centered around the true
parameter values for all models.
Table 1 compares the performance of the different stochastic volatility models in this simulation
scenario. The first number in each cell corresponds to the median ratio of the pertinent performance
measure across 100 simulations, while the numbers in parenthesis correspond to the 5% and 95%
quantiles, respectively. Ratios close to one indicate that the two models in question have similar
performance, while ratios much larger than one indicate that the first model has a much larger RMSD
(or MAD) than the second one, and therefore performs worse.
In agreement with the results reported by Alizadeh et al. (2002), the first row of Table 1 shows
that using the range alone as a volatility proxy produces consistently more accurate volatility estimates than using opening and closing prices alone. Similarly, the second row suggests that, although
18
including the opening prices tends to improve the volatility estimates over those obtained from the
range-based model most of the time, it can sometimes decrease accuracy (specially when the MAD is
used to measure the accuracy of the reconstruction). A similar situation, although much less severe,
happens with our model based on the full likelihood; our model tends to improve over RASV and
RCSV most of the time (with an average efficiency gain in the range of 5-10%), but can underperform in some data sets.
Although the impact of the full information on the estimation of the volatility is moderate, including
the full information contained in CHLO data can greatly improve the estimation of the drift of the
model. For example, the ratio of the absolute error in the estimate of the drift µ based on the 156 time
points between RASV and EXSV has a median of 1.12, with the 5% and 95% quantiles being 0.93 and
1.46. This indicates an average gain in efficiency of about 12%. We also note that EXSV tends to be
more robust to prior misspecification. Sensitivity analysis performed using different hyperparameters
(results not shown) showed that estimates tend to be less affected by prior choice under our EXSV
model.
Further insight into the behavior of these four models is provided by Figure 2, which shows the true
and reconstructed volatility paths for one simulation in our study. Note that paths generated by STSV
tend to underestimate the volatility of volatility. Also, the paths from RASV, RCSV and EXSV tend
to be quite similar to each other, specially RCSV and EXSV. These results suggest that much of the
information about the volatility contained in the High and Low prices is provided by the range, but
also that the actual levels of the minimum and maximum can provide additional helpful information
in most cases.
6.2. Estimating the volatility in the S&P500 index. In this section we consider the series of the
weekly S&P500 prices covering the ten-year period between April 21, 1997 and April 9, 2007, for a
total of 520 observations. Figure 3 shows the evolution of the log returns at closing, as well as the
observed ranges in log returns (computed as the maximum observed log return minus the minimum
19
0.30
0.30
0.25
0.20
Annualized weekly volatility
0.20
145
133
0.20
Annualized weekly volatility
145
133
121
109
97
85
73
61
49
37
25
1
13
145
133
121
109
97
85
73
61
49
37
25
1
13
0.10
0.15
0.20
0.15
0.10
0.25
True
EXSV
0.25
True
RCSV
Annualized weekly volatility
121
97
109
85
73
61
49
37
25
1
13
0.30
145
133
121
109
97
85
73
61
49
37
25
1
0.30
13
0.10
0.15
0.15
0.25
True
RASV
0.10
Annualized weekly volatility
True
STSV
F IGURE 2. True and reconstructed volatility paths for one simulation in our study for
STSV (top left panel), RASV (top right), RCSV (bottom left) and EXSV (bottom
right).
observed log return over the week). Note that both plots provide complementary but distinct information about the volatility in prices. The series of closing returns does not exhibit any long term
20
0.15
0.10
0.05
Observed Range of Returns
0.05
0.00
−0.05
4/2/07
7/17/06
10/31/05
6/1/04
2/14/05
9/15/03
4/15/02
12/30/02
7/23/01
11/6/00
6/7/99
2/22/00
1/5/98
9/21/98
4/21/97
4/2/07
7/17/06
10/31/05
6/1/04
2/14/05
9/15/03
4/15/02
12/30/02
7/23/01
11/6/00
6/7/99
2/22/00
1/5/98
9/21/98
4/21/97
0.00
−0.10
Closing Returns
F IGURE 3. Observed weekly closing returns (left panel) and return ranges (right
panel) in the S&P500 data.
trend, but different levels of volatility can be clearly seen from both plots, with the period 1997-2002
presenting a higher average volatility than the period 2003-2007.
The data was analyzed using three of the models considered in the previous section: STSV, RASV
and EXSV. As in the previous section, prior hyperparameters were chosen so that dµ = 0, Dµ =
0.0001 (therefore, we expect the average weekly returns to be between -0.03 and 0.03 with high
probability), qφ = 9 and rφ = 1 (so that we expect the autoregressive coefficient to be around 0.9),
dα = −3.75, Dα = 0.025 and uτ = 6 vτ = 0.06 (so that the median of the annualized longterm volatility is a priori around 20%). The models were fitted using the particle filter algorithm
described in Section 4. We used a total of 100,000 particles for each model, and monitored particle
impoverishment by computing the effective sample sizes at each point in time. Although not a serious
issue in any of the three models, particle impoverishment is more pronounced for EXSV. This is
probably due to the additional information provided by the extremes, which makes the likelihood
tighter and decreases the viability of individual particles.
21
22
0.25
0.8
0.30
0.9
1.0
4/15/02
7/23/01
11/6/00
2/22/00
6/7/99
9/21/98
1/5/98
4/21/97
4/2/07
7/17/06
10/31/05
2/14/05
6/1/04
9/15/03
12/30/02
4/15/02
7/23/01
11/6/00
2/22/00
6/7/99
9/21/98
1/5/98
4/21/97
6/1/04
2/14/05
10/31/05
7/17/06
4/2/07
6/1/04
2/14/05
10/31/05
7/17/06
4/2/07
9/15/03
0.7
9/15/03
0.20
0.6
12/30/02
0.15
0.5
τ
12/30/02
0.10
0.4
φ
4/15/02
7/23/01
11/6/00
2/22/00
6/7/99
9/21/98
1/5/98
4/21/97
4/2/07
7/17/06
10/31/05
2/14/05
6/1/04
9/15/03
12/30/02
4/15/02
7/23/01
11/6/00
2/22/00
6/7/99
9/21/98
1/5/98
4/21/97
0.3
−4.1
−4.0
−0.01
−3.9
0.00
−3.8
−3.7
0.01
−3.6
−3.5
0.02
µ
α
F IGURE 4. Filtered means and 5% and 95% posterior quantiles for the structural parameters under EXSV.
Point and interval estimates for the structural parameters under EXSV are shown in Figure 4.
These are filtered estimates, which means that they incorporate information available only until the
time they were computed. We note that there is substantial learning about the structural parameters. After T = 520 observations, the posterior mean for the median of the stationary distribution of
volatility, ν = exp{α}, is 0.1546, with a symmetric 90% credible interval (0.1510, 0.1670), while
the autoregressive coefficient for the volatility has a posterior mean of 0.8935 with credible interval
(0.8832, 0.9134). The results from STSV (not shown) are similar, but tend to produce a larger value of
the autocorrelation coefficient (posterior mean 0.9690, 90% credible interval (0.9245, 0.9873)). One
surprising feature of these estimates is the pronounced drop in the autocorrelation coefficient φ on the
week of January 24, 2000, which is accompanied by an increase in the volatility of the volatility, τ .
This can be explained by the large negative return observed for this week (around -11% in the week).
This observation suggests that a model that includes volatility jumps would be more appropriate for
this data, however, the development of such a model is beyond the scope of this paper and will be
discussed elsewhere.
Figure 5 shows point estimates for the volatility of returns under all three models. Again, these are
filtered estimates. Although the overall level of the volatility series seems to be similar in all three
instances, and the estimates obtained from RASV and EXSV are similar, there are striking differences in the behavior of STSV with respect to RASV and EXSV. For example, note the peaks during
the high volatility period of 1997-2002 are much more pronounced under RASV and EXSV, while
the average volatility between 2003 and 2007 seems to be lower under both RASV and EXSV than
under STSV. Therefore, including information on extreme values seems to help to correct not only
for underestimation, but also for overestimation of the volatility. To complement the information in
Figure 5, we present in Figure 6 the posterior distribution for the volatility of returns on December 6,
2000 and April 2, 2007 under all three models. All distributions are right skewed, but the distributions under RASV and EXSV are shifted to the left, which suggest that STSV overestimates the true
volatility of model in both cases. In addition, for December 6, 2000 the distributions for RASV and
EXSV are almost identical; however, for April 2, 2007, they are noticeable different. More important,
23
0.4
0.3
0.2
4/2/07
7/17/06
10/31/05
2/14/05
6/1/04
9/15/03
12/30/02
4/15/02
7/23/01
11/6/00
2/22/00
6/7/99
9/21/98
1/5/98
4/21/97
0.1
Annualized weekly volatility
STSV
EXSV
RASV
F IGURE 5. Estimated volatilities of S&P500 data using both stochastic volatility models.
the posterior under both EXSV and RASV has a lower variability than that under STSV, reflecting the
additional information contained in the extreme values.
In order to better understand the effect of the extreme values on the estimation of volatility, we
compare our volatility estimates with the value of VIX index for the dates under consideration, which
provides an independent source of information about the volatility on the S&P500 index. The VIX is
constructed as a weighted average from the implied volatilities obtained from options whose underlying asset are included in the S&P500, and are therefore not directly linked to the S&P500 extremes.
Of course, the VIX is the implied forward looking volatility priced in the market, with 1-2 months
24
4/2/07
15
6/12/00
20
STSV
EXSV
RASV
10
Density
0
0
5
5
Density
10
15
STSV
EXSV
RASV
0.1
0.2
0.3
0.4
0.05
Volatility
0.10
0.15
0.20
0.25
0.30
Volatility
F IGURE 6. Posterior distribution for the volatility of returns for December 6, 2000
(left panel) and April 2, 2007 (right panel)
average tenors and which includes the price of risk, while our estimates refer to current filtered estimates, but we make this comparison under the assumption that the VIX will reflect current market
expectations of where the volatility of the underlying is, and any bias will affect equally all models
under consideration. Figure 7 presents scatter plots of the observed VIX prices at closing, which show
strong association with the volatilities estimated under all models. Correlation between the closing
VIX price and the filtered volatilities is 0.801 for the STSV model 0.853 for RASV and 0.861 for
EXSV. Very similar results are obtained if the average between the opening and closing VIX price are
used instead (correlations are 0.816 for STSV, 0.854 for RASV and 0.863 for EXSV). The stronger
correlation with EXSV volatilities provides additional evidence that the information contained in the
realized extremes can indeed improve the performance of the model, although the difference between
RASV and EXSV is small and might not be relevant.
25
●
●
20
25
30
0.20
0.15
●
●
●
●
●
●
●
●
●
●
35
40
45
●
●
●
●
●
●●
0.30
0.25
●
●
● ● ● ●●
●
●
●● ● ●
● ●
● ● ●●
●● ●
●
● ● ● ●
●
●
●
● ●
●
●●
● ●
● ● ●●●
●
● ● ●●
● ●
●
●●●
●
●
●
●
●
● ●● ●● ●●●●
●
● ●
● ● ●●
●●● ●
●
●
●
●
●
●
●
●
●
●● ● ●
● ●
●●● ●
●●
● ●
● ●
● ●●
●
●● ●
●● ●
●●●
●● ●●●
●●
●
●● ● ●● ●●
● ●●
●●
●
● ●●
● ●
●● ●
●
● ● ●
●
●●
●
●
●●
●
●
●
●●
●
●● ●●●
●●
●●●
●
●
●
●●
●●
● ●●●
●
●
●● ●●
●
●●
● ●● ●
●
●
●●
● ●●●●●
● ●
●
●
●
● ●
●
●
●●●●
●
●
●
●
●
●●
● ●●
●
●
● ●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ●
● ● ●
●
●
●● ●
●
●
●
●
●
●
●
●●
●●
●● ● ●
●
●
●
●
●
●
●
●
●
●
●
● ● ● ●
● ●
●
●
●●
● ●●
●● ●
●
●
● ● ● ●●
●●
●● ●
●●
●
●● ●●
●
●●
●●
●
●
●●●
●●
●
●●
●
●
●
●● ●●●
●
●●
●
●●●●●
●
●
●
●
●
●●
●
●●●●●
●● ●●●●
●
●
● ●●
●
●●
●● ●
●
●
●
●
●
●
●
●
●●● ●
●
●
●●●● ●● ●
●●
●
●
●● ● ●
●
●
●
●●
●●
●●
●
●●
●
●●● ● ●
●
●● ● ●●●
●●
●
●●●●
●
●
●
●
●● ●
15
Estimated volatility − RASV
●
0.10
●
●
● ●
10
●
●
●
●
●
●
●
●
●
●
●
●
●
0.05
0.25
●
● ●
● ●●
●
0.10
Estimated volatility − STSV
●
0.20
●
●
●
●
●
●
●
0.15
0.30
●
●
●
0.35
●
●
●
●
● ●
●●
●
●●
●●
●
●
●
●●●●
●●
●
●
●
● ●
●●●● ● ● ●
●
● ●
●
●● ●
●●●
●
● ●
● ●
●
●
●
●
● ●● ●
●
● ●●
●
● ● ● ●●
●
●● ● ●
●
●
●
●
●●● ●●●
●
●
●
●
●
●
●●●● ●● ●●
●● ●● ●●●
●
●
● ●
● ●
●●
●
●●●●
● ●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●● ●
●●
● ●●●●
●
●●
●
●●
● ● ●
●●
●●
●
●
●●
●
●●
●●
●
●
● ●
●
●●
● ●●● ●
● ●●●●
●
● ●
●
●
●
●●
●
●
●
●● ●
●●● ●●
●●●●●
●
●●
●●
● ●
●
●
●●
●
●●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●●
●● ●
●
●
●●
●
●
● ●●
●
●
●
●
● ● ●
●
●
●●
●
●
●●●
●
●●
●●
●●●
●
● ●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●●
●●●
●●
●
●●
● ●●● ● ●
●●
●
● ●
● ●
●
●●
●●
●
●
● ●●
●
●
● ●
●● ●
●●
● ●●
●●●●●
●
●
●
●●●●●
●
●●
● ●
●
●
●●
●
●●●
●
●●● ●
●●
●●
●
●
●
●
●
●●
●●●
● ● ●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●
● ●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●● ●
●
●
●●
●
●
●●
●
●
10
15
20
25
35
40
45
VIX (closing)
0.4
VIX (closing)
30
●
●
●
●
●
0.3
0.2
●
●
● ●●
●
●●
● ●
●
●
●
●
●
●●
● ●
●●
●
●
●●
●
● ● ● ● ●●
●
● ● ● ●
● ●
●
●
●
●
● ● ● ●● ● ● ●
●● ●
●
●
●
●●
●●
●
● ●
●
● ● ●●● ●●
●● ●
● ●● ● ● ●
●
● ●●● ●● ●● ●
● ●
●
●● ●●● ● ●●●
●
●
●
● ●●
●● ●
●
● ●● ●
●
● ●● ●●
●
●●●●●
●●
●●
●
● ● ● ●● ●●
● ●● ●
●●
●
● ●● ●
●
●
●
●
●● ●
●
●
●
●
●
●●●
● ● ●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ●
● ●
● ●
●●
●●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ● ●
●● ●
●
●
●
●
●●
●●
●●
●
● ● ●●●
●
●●●
●
●
●
●
●●●●
●
●●
●●
●
●
●●
●●●
●●●
● ● ●● ●
●
●
●
● ●●●●●●
●
●
● ●
●
●
●
●● ●
●● ●●● ●●
●●●
●●
● ●
●●
●●
●
●
●● ●
●
●●
●
●●
●
●●
●
●●●●●
●●●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ●
●
●●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●● ●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
● ●
●●●
●● ● ●●
● ●
●
●
●
●
●
●
●
●
● ●
●
●
●
0.1
Estimated volatility − EXSV
●
●
●●
●
●
●●
●
●
●
10
15
20
25
30
35
40
45
VIX (closing)
F IGURE 7. Closing VIX prices versus filtered volatilities for all three models under consideration.
7. D ISCUSSION AND FUTURE WORK
This paper adds to the growing body of literature suggesting that the information contained in
the observed extremes of asset prices can greatly contribute to increase the accuracy and efficiency
volatility estimates. In addition, our results also suggest that the information contained in the level of
26
the extreme returns (which is lost when using the observed ranges for inferences) can also contribute
to more efficient estimation of the volatility, and almost as important in certain applications, to the
estimation of the drift of the process. This paper also provides a unifying framework in which to
understand several CHLO models available in the literature, allowing for a direct comparison of
different assumptions.
Some extensions of the simple stochastic volatility model discussed in this paper are immediate
and have already been hinted at. For example, it is well known that leverage effects can be incorporated by introducing a correlation between the innovations in the price and volatility processes. In
addition, non-equally spaced observation can be easily accommodated by slightly rewriting the likelihood equations in terms of non-unit length intervals. The strong identifiability of model parameters
provided by using the full information in the model allows us to reliably compare different evolution
dynamics (e.g., Ornstein-Uhlenbeck, Markov-switching, jump and squared root processes). Indeed,
model comparisons in this setting is straightforward thanks to our reliance on particle filters for computation. Finally, we plan to investigate how the additional information provided by the range can
help when reconstructing option prices. In this regard, note that, conditionally on the parameters of
the underlying stochastic process, the pricing formulas in Heston (1993) can be used almost directly
in our problem. Therefore, our model can be extended to generate a posterior distribution for the price
of any option of interest, which can be compared to the prices observed in the market.
Another avenue that deserves future exploration is the use of extreme prices for the estimation
of covariance across multiple asset prices. In that regard, we can start by writing a model for the
multivariate log-asset prices yt as a multivariate GBM, with time variance covariance matrix, and
deriving the joint density for closing, high and low prices conditional on the opening prices, in a
manner similar to what we did in this paper. This joint likelihood contains all relevant information in
the extremes and avoid the calculation of cross ranges (Rogers & Zhou, 2008), which scale badly to
higher dimensions.
27
A PPENDIX A. P ROOF OF THEOREM 1
Consider Xt , a driftless Brownian motion with unit variance starting at zero. In Chapter 3, p. 75 of
Klebaner (2005), the joint cumulative distribution for the maximum Mt = max0≤s≤t {Xs }, minimum
mt = min0≤s≤t {Xs } and final value of Xt is obtained, yielding
Z
P (mt ≥ at , Mt ≤ bt , Xt ≤ xt ) =
xt
∞
X
[p(z − 2n(bt − at )) − p(z + (2n(bt − at ) − 2at )]dz
−∞ n=−∞
where p(z) corresponds to the density of a standard gaussian density. To include the drift and nonunit volatility, we can use Girsanov’s theorem. Finally, information about the opening price can be
incorporated by translating the Brownian motion, i.e, letting a∗t = at − yt−1 , b∗t = bt − yt−1 and
xt = yt − yt−1 , which yields (7). An alternative, more convoluted proof, can also rely on a more
probabilistic version of Kelvin’s method of images combined with basic Fourier analysis in section 8
of chapter 2 of Port & Stone (1978).
R EFERENCES
A LIZADEH , S., B RANDT, M. & D IEBOLD , F. (2002). Range-based estimation of stochastic volatility
models. Journal of Finance 57.
A NDERSEN , T., B OLLERSLEV, T., D IEBOLD , F. & L ABYS , P. (2001a). The distribution of exchange
rate volatility. Journal of the American Statistical Association 96.
BALL , C. & T OURUS , W. (1984). The maximum likelihood estimation of security price volatility:
Theory, evidence, and application to option pricing. Journal of Business 57(1).
BARNDORFF -N IELSEN , O. & S HEPHARD , N. (2002). Econometric analysis of realised volatility and
its use in estimating stochastic volatility models. Journal of the Royal Statistical Society, Series B
63.
B OLLERSLEV, T. (1986). Generalized autoregressive conditional heteroscedasticity. Journal of
Econometrics 31, 307–327.
28
B RANDT, M. & J ONES , C. (2005). Estimating the volatility of stock prices: a comparison of methods
that use high and low prices. Finance Research Letters 2, 201–209.
C ARVALHO , C. & L OPES , H. F. (2007). Simulation-based sequential analysis of markov switching
stochastic volatility models. Computational Statistics and Data Analysis 51, 4526–4542.
C OURTAULT, J., K ABANOV, Y., B RU , B., C REPEL , P., L EBON , I. & L E M ARCHAND , A. (2000).
Louis Bachelier on the centenary of theorie de la speculation. Mathematical Finance 10, 341–353.
DANA , R., J EANBLANC , M. & K ENNEDY, A. (2007). Financial Markets in Continuous Time. New
York: Springer Verlag, 1st edition.
D OUCET, A.,
DE
F REITAS , N. & G ORDON , N. (Eds.) (2001). Sequential Monte Carlo Methods in
Practice. Springer.
DYM , H., M C K EAN , H. P., A LDOUS , D. & T ONG , Y. L. (1985). Fourier Series and Integrals. San
Francisco: Academic Press, 1st edition.
E NGLE , R. F. (1982). Autoregressive conditional heteroscedasticity with estimates of variance of
united kingdom inflation. Econometrica 50, 987–1008.
FAMA , E. F. (1965). The behavior of stock market prices. Journal of Business pp. 34–105.
F ELLER , W. (1971). An Introduction to Probability Theory and its Applications. New York: Wiley,
2nd edition.
G ALLANT, A., H SU , C. & TAUCHEN , G. (1999). Using daily range data to calibrate volatility
diffusions and extract the forward integrated variance. The Review of Economic and Statistics
81(4).
G ARMAN , M. & K LASS , M. (1980). On the estimation of security price volatilities from historical
data. Journal of Business 53, 67–78.
H ESTON , S. L. (1993). A closed-form solution for options with stochastic volatility with applications
to bond and currency options. Review of Financial Studies 6, 327–343.
29
H UERTA , G. & W EST, M. (1999). Priors and component structures in autoregressive time series
models. Journal of the Royal Statistical Society, Series B 61, 881–899.
JACQUIER , E., P OLSON , N. G. & ROSSI , P. E. (1994). Bayesian analysis of stochastic volatility
models. Journal of business and Economic Statistics 12, 371–389.
K IM , S., S HEPHARD , N. & C HIB , S. (1998). Stochastic volatility: Likelihood inference and comparison with ARCH models. Review of Economic Studies 65.
K ITAGAWA , G. (1998). A self-organizing state-space model. Journal of the American Statistical
Association 93, 1203–1215.
K LEBANER , F. C. (2005). Introduction to stochastic calculus with applications. Inperial College
Press, second edition edition.
L ILDHOLDT, P. M. (2002). Estimation of GARCH models based on open, close, high, and low prices.
Centre for Analytical Finance Working Papers Series 128, Department of Economics, University
of Aarhus, and Center for Analytical Finance. Available at http://www.cls.dk/caf/wp/wp-128.pdf.
L IU , J. & W EST, M. (2001). Combined parameter and state estimation in simulation-based filtering.
In Sequential Monte Carlo Methods in Practice, Eds. A. Doucet, N. de Freitas & N. Gordon.
Springer.
M AGDON , M. & ATIYA , A. (2003). A maximum likelihood approach to volatility estimation for a
Brownian motion using high, low and close price data. Quantitative Finance 3, 376–384.
M ANDELBROT, B. (1963). The variation of certain speculative prices. Journal of Business pp. 394–
419.
M OLINA , G., H AN , C. & F OUQUE , J. (2009). MCMC estimation of multiscale stochastic volatility
models. In Handbook of Quantitative Finance and Risk Management, Eds. C. Lee & A. Lee.
Springer.
O FFICER , R. R. (1973). The variability of the market factor of the New York stock exchang. Journal
of Business 46, 434–453.
30
PARKINSON , M. (1980). The extreme value method for estimating the variance of the rate of return.
Journal of Business 53, 61–65.
P ITT, M. K. & S HEPHARD , N. (1999). Filtering via simulation: Auxiliary prticle filters. Journal of
the American Statistical Association 94, 590–599.
P ORT, S. & S TONE , C. (1978). Brownian Motion and Classical Potential Theory. San Diego:
Academic Press, 1st edition.
ROGERS , L. (1998). Volatility estimation with price quanta. Mathematical Finance 8, 61–65.
ROGERS , L. & S ATCHELL , S. (1991). Estimating variance from high, low and closing prices. Annals
of Applied Probability 1, 504–512.
ROGERS , L., S ATCHELL , S. & YOON , Y. (1994). Estimating the volatility of stock prices: a comparison of methods that use high and low prices. Applied Financial Economics 4, 241–247.
ROGERS , L. & Z HOU , F. (2008). Estimating correlation from high, low, opening and closing prices.
Annals of Applied Probability 18, 813–823.
RUSSELL , J. & E NGLE , R. (1998). Econometric analysis of discrete-valued irregularly-spaced financial transactions data using a new autoregressive conditional multinomial model. CRSP Working
Paper 470, University of California at San Diego Working Paper No. 98-10. Available at SSRN:
http://ssrn.com/abstract=106528 or DOI: 10.2139/ssrn.106528.
S HEPHARD , N. (Ed.) (2005). Stochastic Volatility: Selected Readings. Oxford University Prss.
U HLENBECK , G. & O RNSTEIN , L. (1930). On the theory of Brownian motion. Phys.Rev. 36, 823–
841.
W ILDER , J. (1978). New Concepts in Technical Trading Systems. Trend Research, 1st edition.
31
| 5cs.CE
|
arXiv:1608.03912v1 [cs.PL] 12 Aug 2016
A Hiking Trip Through the Orders of
Magnitude: Deriving Efficient Generators for
Closed Simply-Typed Lambda Terms and
Normal Forms
Paul Tarau
Department of Computer Science and Engineering
University of North Texas
[email protected]
Abstract. Contrary to several other families of lambda terms, no closed
formula or generating function is known and none of the sophisticated
techniques devised in analytic combinatorics can currently help with
counting or generating the set of simply-typed closed lambda terms of
a given size.
Moreover, their asymptotic scarcity among the set of closed lambda
terms makes counting them via brute force generation and type inference quickly intractable, with previous published work showing counts
for them only up to size 10.
By taking advantage of the synergy between logic variables, unification
with occurs check and efficient backtracking in today’s Prolog systems,
we climb 4 orders of magnitude above previously known counts by deriving progressively faster Horn Clause programs that generate and/or
count the set of closed simply-typed lambda terms of sizes up to 14. A
similar count for closed simply-typed normal forms is also derived up to
size 14.
Keywords: logic programming transformations, type inference, combinatorics of lambda terms, simply-typed lambda calculus, simply-typed normal forms.
1
Introduction
Generation of lambda terms [1] has practical applications to testing compilers
that rely on lambda calculus as an intermediate language, as well as in generation
of random tests for user-level programs and data types. At the same time, several
instances of lambda calculus are of significant theoretical interest given their
correspondence with logic and proofs.
Simply-typed lambda terms [2,3] enjoy a number of nice properties, among
which strong normalization (termination for all evaluation-orders), a cartesian
closed category mapping and a set-theoretical semantics. More importantly, via
the Curry-Howard correspondence lambda terms that are inhabitants of simple
types can be seen as proofs for tautologies in minimal logic which, in turn, correspond to the types. Extended with a fix-point operator, simply-typed lambda
terms can be used as the intermediate language for compiling Turing-complete
functional languages. Random generation of simply-typed lambda terms can also
help with automation of debugging compilers for functional programming languages [4].
Recent work on the combinatorics of lambda terms [5,6,7,8], relying on recursion equations, generating functions and techniques from analytic combinatorics
[9] has provided counts for several families of lambda terms and clarified important properties like their asymptotic density. With the techniques provided by
generating functions [9], it was possible to separate the counting of the terms
of a given size for several families of lambda terms from their more computation intensive generation, resulting in several additions (e.g., A220894, A224345,
A114851) to The On-Line Encyclopedia of Integer Sequences, [10].
On the other hand, the combinatorics of simply-typed lambda terms, given
the absence of closed formulas, recurrence equations or grammar-based generators, due to the intricate interaction between type inference and the applicative
structure of lambda terms, has left important problems open, including the very
basic one of counting the number of closed simply-typed lambda terms of a given
size. At this point, obtaining counts for simply-typed lambda terms requires going through the more computation-intensive generation process.
As a fortunate synergy, Prolog’s sound unification of logic variables, backtracking and definite clause grammars have been shown to provide compact combinatorial generation algorithms for various families of lambda terms [11,12,13,14].
For the case of simply-typed lambda terms, we have pushed (in the unpublished draft [15]) the counts in sequence A220471 of [10] to cover sizes 11 and
12, each requiring about one magnitude of extra computation effort, simply by
writing the generators in Prolog. In this paper we focus on going two more magnitudes higher, while also integrating the results described in [15]. Using similar
techniques, we achieve the same, for the special case of simply-typed normal
forms.
The paper is organized as follows. Section 2 describes our representation of
lambda terms and derives a generator for closed lambda terms. Section 3 defines
generators for well-formed type formulas. Section 4 introduces a type inference
algorithm and then derives, step by step, efficient generators for simply-typed
lambda terms and simple types inhabited by terms of a given size. Section 5
defines generators for closed lambda terms in normal form and then replicates
the derivation of an efficient generator for simply-typed closed normal forms.
Section 6 aggregates our experimental performance data and section 7 discusses
possible extensions and future improvements. Section 8 overviews related work
and section 9 concludes the paper.
The paper is structured as a literate Prolog program. The code has been
tested with SWI-Prolog 7.3.8 and YAP 6.3.4. It is also available as a separate
file at http://www.cse.unt.edu/~tarau/research/2016/lgen.pro.
2
2
Deriving a generator for lambda terms
Lambda terms can be seen as Motzkin trees [16], also called unary-binary trees,
labeled with lambda binders at their unary nodes and corresponding variables
at the leaves. We will thus derive a generator for them from a generator for
Motzkin trees.
2.1
A canonical representation with logic variables
We can represent lambda terms [1] in Prolog using the constructors a/2 for applications, l/2 for lambda abstractions and v/1 for variable occurrences. Variables
bound by the lambdas and their occurrences are represented as logic variables. As
an example, the lambda term λa.(λb.(a (b b)) λc.(a (c c))) will be represented as
l(A,a(l(B,a(v(A),a(v(B),v(B)))),l(C,a(v(A),a(v(C),v(C)))))). As variables share a unique scope (the clause containing them), this representation assumes that distinct variables are used for distinct scopes induced by the lambda
binders in terms occurring in a given Prolog clause.
Lambda terms might contain free variables not associated to any binders.
Such terms are called open. A closed term is such that each variable occurrence
is associated to a binder.
2.2
Generating Motzkin trees
Motzkin-trees (also called binary-unary trees) have internal nodes of arities 1 or
2. Thus they can be seen as a skeleton of lambda terms that ignores binders and
variables and their leaves.
The predicate motzkin/2 generates Motzkin trees with S internal and leaf
nodes.
motzkin(S,X):-motzkin(X,S,0).
motzkin(v)-->[].
motzkin(l(X))-->down,motzkin(X).
motzkin(a(X,Y))-->down,motzkin(X),motzkin(Y).
down(s(X),X).
Motzkin-trees, with leaves assumed of size 1 are counted by the sequence
A001006 in [10]. Alternatively, as in our case, when leaves are assumed of size
0, we obtain binary-unary trees with S internal nodes, counted by the entry
A006318 (Large Schröder Numbers) of [10].
Note the use of the predicate down/2, that assumes natural numbers in unary
notation, with n s/1 symbols wrapped around 0 to denote n ∈ N. As our combinatorial generation algorithms will usually be tractable for values of n below 15,
the use of unary notation is comparable (and often slightly faster) than the call
to arithmetic built-ins. Note also that this leads, after the DCG translation, to
“pure” Prolog programs made exclusively of Horn Clauses, as the DCG notation
3
can be eliminated by threading two extra arguments controlling the size of the
terms.
To more conveniently call these generators with the usual natural numbers
we define the converter n2s as follows.
n2s(0,0).
n2s(N,s(X)):-N>0,N1 is N-1,n2s(N1,X).
Example 1 Motzkin trees with 2 internal nodes.
?- n2s(1,S),motzkin(S,T).
S = s(0), T = l(v) ;
S = s(0), T = a(v, v) .
2.3
Generating closed lambda terms
We derive a generator for closed lambda terms by adding logic variables as labels
to their binder and variable nodes, while ensuring that the terms are closed, i.e.,
that the function mapping variables to their binders is total.
The predicate lambda/2 builds a list of logic variables as it generates binders.
When generating a leaf variable, it picks “nondeterministically” one of the binders
among the list of binders available, Vs. As in the case of Motzkin trees, the predicate down/2 controls the number of internal nodes.
lambda(S,X):-lambda(X,[],S,0).
lambda(v(V),Vs)-->{member(V,Vs)}.
lambda(l(V,X),Vs)-->down,lambda(X,[V|Vs]).
lambda(a(X,Y),Vs)-->down,lambda(X,Vs),lambda(Y,Vs).
The sequence A220471 in [10] contains counts for lambda terms of increasing
sizes, with size defined as the number of internal nodes.
Example 2 Closed lambda terms with 2 internal nodes.
?- lambda(s(s(0)),Term).
Term = l(A, l(B, v(B))) ;
Term = l(A, l(B, v(A))) ;
Term = l(A, a(v(A), v(A))) .
3
A visit to the other side: the language of types
As a result of the Curry-Howard correspondence, the language of types is isomorphic with that of minimal logic, with binary trees having variables at leaf
positions and the implication operator (“->”) at internal nodes. We will rely
on the right associativity of this operator in Prolog, that matches the standard
notation in type theory.
The predicate type skel/3 generates all binary trees with given number of
internal nodes and labels their leaves with unique logic variables. It also collects
the variables to a list returned as its third argument.
4
type_skel(S,T,Vs):-type_skel(T,Vs,[],S,0).
type_skel(V,[V|Vs],Vs)-->[].
type_skel((X->Y),Vs1,Vs3)-->down,
type_skel(X,Vs1,Vs2),
type_skel(Y,Vs2,Vs3).
Type skeletons are counted by the Catalan numbers (sequence A000108 in [10]).
Example 3 All type skeletons for N=3.
?- type_skel(s(s(s(0))),T,_).
T = (A->B->C->D) ;
T = (A-> (B->C)->D) ;
T = ((A->B)->C->D) ;
T = ((A->B->C)->D) ;
T = (((A->B)->C)->D) .
The next step toward generating the set of all type formulas is observing that
logic variables define equivalence classes that can be used to generate partitions
of the set of variables, simply by selectively unifying them.
The predicate mpart of/2 takes a list of distinct logic variables and generates
partitions-as-equivalence-relations by unifying them “nondeterministically”. It
also collects the unique variables, defining the equivalence classes as a list given
by its second argument.
mpart_of([],[]).
mpart_of([U|Xs],[U|Us]):mcomplement_of(U,Xs,Rs),
mpart_of(Rs,Us).
To implement a set-partition generator, we will split a set repeatedly in
subset+complement pairs with help from the predicate mcomplement of/2.
mcomplement_of(_,[],[]).
mcomplement_of(U,[X|Xs],NewZs):mcomplement_of(U,Xs,Zs),
mplace_element(U,X,Zs,NewZs).
mplace_element(U,U,Zs,Zs).
mplace_element(_,X,Zs,[X|Zs]).
To generate set partitions of a set of variables of a given size, we build a list
of fresh variables with the equivalent of Prolog’s length predicate working in
unary notation, len/2.
partitions(S,Ps):-len(Ps,S),mpart_of(Ps,_).
len([],0).
len([_|Vs],s(L)):-len(Vs,L).
5
The count of the resulting set-partitions (Bell numbers) corresponds to the
entry A000110 in [10].
Example 4 Set partitions of size 3 expressed as variable equalities.
?- partitions(s(s(s(0))),P).
P = [A, A, A] ;
P = [A, B, A] ;
P = [A, A, B] ;
P = [A, B, B] ;
P = [A, B, C].
We can then define the language of formulas in minimal logic, among which
tautologies will correspond to simple types, as being generated by the predicate
maybe type/3.
maybe_type(L,T,Us):-type_skel(L,T,Vs),mpart_of(Vs,Us).
Example 5 Well-formed formulas of minimal logic (possibly types) of size 2.
?- maybe_type(s(s(0)),T,_).
T = (A->A->A) ;
T = (A->B->A) ;
T = (A->A->B) ;
T = (A->B->B) ;
T = (A->B->C) ;
T = ((A->A)->A) ;
T = ((A->B)->A) ;
T = ((A->A)->B) ;
T = ((A->B)->B) ;
T = ((A->B)->C) .
The sequence 2,10,75,728,8526,115764,1776060,30240210 counting these formulas corresponds to the product of Catalan and Bell numbers.
4
Merging the two worlds: generating simply-typable
lambda terms
One can observe that per-size counts of both the sets of lambda terms and their
potential types are very fast growing. There is an important difference, though,
between computing the type of a given lambda term (if it exists) and computing
an inhabitant of a type (if it exists). The first operation, called type inference is
an efficient operation (linear in practice) while the second operation, called the
inhabitation problem is P-space complete [17].
This brings us to design a type inference algorithm that takes advantage of
operations on logic variables.
6
4.1
A type inference algorithm
While in a functional language inferring types requires implementing unification
with occurs-check, as shown for instance in [5], this operation is available in
Prolog as a built-in predicate, optimized, for instance, in SWI-Prolog [18], to
proceed incrementally, only checking that no new cycles are introduced during
the unification step as such.
The predicate infer type/3 works by using logic variables as dictionaries
associating terms to their types. Each logic variable is then bound to a term of
the form X:T where X will be a component of a fresh copy of the term and T will
be its type. Note that we create this new term as the original term’s variables
end up loaded with chunks of the partial types created during the type inference
process.
As logic variable bindings propagate between binders and occurrences, this
ensures that types are consistently inferred.
infer_type((v(XT)),v(X),T):-unify_with_occurs_check(XT,X:T).
infer_type(l((X:TX),A),l(X,NewA),(TX->TA)):-infer_type(A,NewA,TA).
infer_type(a(A,B),a(X,Y),TY):-infer_type(A,X,(TX->TY)),infer_type(B,Y,TX).
Example 6 illustrates typability of the term corresponding to the S combinator
λx0 .λx1 .λx2 .((x0 x2 ) (x1 x2 ))
and untypabilty of the term corresponding to the Y combinator
λx0 .(λx1 .(x0 (x1 x1 )) λx2 .(x0 (x2 x2 ))).
?- infer_type(l(A,l(B,l(C,a(a(v(A),v(C)),a(v(B),v(C)))))),X,T),
portray_clause((T:-X)),fail.
(A->B->C)-> (A->B)->A->C :l(D,l(F,l(E, a(a(v(D), v(E)), a(v(F), v(E)))))).
?- infer_type(
l(A,a(l(B,a(v(A),a(v(B),v(B)))),l(C,a(v(A),a(v(C),v(C)))))), X, T).
false.
By combining generation of lambda terms with type inference we have our
first cut to an already surprisingly fast generator for simply-typable lambda
terms, able to generate in a few hours counts for sizes 11 and 12 for sequence
A220471 in [10].
lamb_with_type(S,X,T):-lambda(S,XT),infer_type(XT,X,T).
Example 7 Lambda terms of size up to 3 and their types.
?- lamb_with_type(s(s(s(0))),Term,Type).
Term = l(A, l(B, l(C, v(C)))), Type = (D->E->F->F) ;
Term = l(A, l(B, l(C, v(B)))), Type = (D->E->F->E) ;
Term = l(A, l(B, l(C, v(A)))), Type = (D->E->F->D) ;
Term = l(A, l(B, a(v(B), v(A)))), Type = (C-> (C->D)->D) ;
Term = l(A, l(B, a(v(A), v(B)))), Type = ((C->D)->C->D) ;
7
Term
Term
Term
Term
=
=
=
=
l(A, a(v(A), l(B, v(B)))),
l(A, a(l(B, v(B)), v(A))),
l(A, a(l(B, v(A)), v(A))),
a(l(A, v(A)), l(B, v(B))),
Type
Type
Type
Type
=
=
=
=
(((C->C)->D)->D) ;
(C->C) ;
(C->C) ;
(C->C).
Note that, for instance, when one wants to select only terms having a given
type, this is quite inefficient. Next, we will show how to combine size-bound term
generation, testing for closed terms and type inference into a single predicate.
This will enable more efficient querying for terms inhabiting a given type, as one
would expect from Prolog’s multi-directional execution model, and more importantly for our purposes, to climb two orders of magnitude higher for counting
simply-typed terms of size 13 and 14.
4.2
Combining term generation and type inference
We need two changes to infer type to turn it into an efficient generator for
simply-typed lambda terms. First, we need to add an argument to control the
size of the terms and ensure termination, by calling down/2 for internal nodes.
Second, we need to generate the mapping between binders and variables. We
ensure this by borrowing the member/2-based mechanism used in the predicate
lambda/4 generating closed lambda terms in subsection 2.3.
The predicate typed lambda/3 does just that, with helper from DCG-expanded
typed lambda/5.
typed_lambda(S,X,T):-typed_lambda(_XT,X,T,[],S,0).
typed_lambda(v(V:T),v(V),T,Vs)--> {
member(V:T0,Vs),
unify_with_occurs_check(T0,T)
}.
typed_lambda(l(X:TX,A),l(X,NewA),(TX->TY),Vs)-->down,
typed_lambda(A,NewA,TY,[X:TX|Vs]).
typed_lambda(a(A,B),a(NewA,NewB),TY,Vs)-->down,
typed_lambda(A,NewA,(TX->TY),Vs),
typed_lambda(B,NewB,TX,Vs).
Like lambda/4, the predicate typed lambda/5 relies on Prolog’s DCG notation to thread together the steps controlled by the predicate down. Note also
the nondeterministic use of the built-in member/2 that enumerates values for
variable:type pairs ranging over the list of available pairs Vs, as well as the
use of unify with occurs check to ensure that unification of candidate types
does not create cycles.
Example 8 A term of size 15 and its type.
l(A,l(B,l(C,l(D,l(E,l(F,l(G,l(H,l(I,l(J,l(K,
a(v(I),l(L,a(a(v(E),v(J)),v(J)))))))))))))))
M->N->O->P-> (Q->Q->R)->S->T->U-> ((V->R)->W)->Q->X->W
8
We will discuss exact performance data later, but let’s note here that this
operation brings down by an order of magnitude the computational effort to
generate simply-typed terms. As expected, the number of solutions is computed
as the sequence A220471 in [10]. Interestingly, by interleaving generation of closed
terms and type inference in the predicate typed lambda, the time to generate
all the closed simply-typed terms is actually shorter than the time to generate
all closed terms of the same size, e.g., 17.123 vs. 31.442 seconds for size 10 with
SWI-Prolog. As, via the Curry-Howard isomorphism, closed simply typed terms
correspond to proofs of tautologies in minimal logic, co-generation of terms and
types corresponds to co-generation of tautologies and their proofs for proofs of
given length.
4.3
One more trim: generating inhabited types
Let’s first observe that the actual lambda term does not need to be built, provided that we mimic exactly the type inference operations that one would need
to perform to ensure it is simply-typed. It is thus safe to remove the first argument of typed lambda/5 as well as the building of the fresh copy performed
in the second argument. To further simplify the code, we can also make the
DCG-processing of the size computations explicit, in the last two arguments.
This gives the predicate inhabited type/4 and then inhabited type/2,
that generates all types having inhabitants of a given size, but omits the inhabitants as such.
inhabited_type(X,Vs,N,N):member(V,Vs),
unify_with_occurs_check(X,V).
inhabited_type((X->Xs),Vs,s(N1),N2):inhabited_type(Xs,[X|Vs],N1,N2).
inhabited_type(Xs,Vs,s(N1),N3):inhabited_type((X->Xs),Vs,N1,N2),
inhabited_type(X,Vs,N2,N3).
Clearly the multiset of generated types has the same count as the set of their
inhabitants and this brings us an additional 1.5x speed-up.
inhabited_type(S,T):-inhabited_type(T,[],S,0).
One more (easy) step, giving a 3x speed-up, makes reaching counts for sizes 13
and 14 achievable: using a faster Prolog, with a similar unify with occurs check
built-in, like YAP [19], with the last value computed in less than a day.
Example 9 The sequence A220471 completed up to N=14
first 10: 1,2,9,40,238,1564,11807,98529,904318,9006364
11:
12:
96,709,332
1,110,858,977
13: 13,581,942,434
14: 175,844,515,544
9
5
Doing it once more: generating closed simply-typed
normal forms
We will devise similar methods for an important subclass of simply-typed lambda
terms.
5.1
Generating normal forms
Normal forms are lambda terms that cannot be further reduced. A normal form
should not be an application with a lambda as its left branch and, recursively,
its subterms should also be normal forms. The predicate normal form/2 uses
normal form/4 to define them inductively and generates all normal forms with
S internal nodes.
normal_form(S,T):-normal_form(T,[],S,0).
normal_form(v(X),Vs)-->{member(X,Vs)}.
normal_form(l(X,A),Vs)-->down,normal_form(A,[X|Vs]).
normal_form(a(v(X),B),Vs)-->down,normal_form(v(X),Vs),normal_form(B,Vs).
normal_form(a(a(X,Y),B),Vs)-->down,normal_form(a(X,Y),Vs),normal_form(B,Vs).
Example 10 illustrates closed normal forms with 2 internal nodes.
?NF
NF
NF
normal_form(s(s(0)),NF).
= l(A, l(B, v(B))) ;
= l(A, l(B, v(A))) ;
= l(A, a(v(A), v(A))) .
The number of solutions of our generator replicates entry A224345 in [10] that
counts closed normal forms of various sizes.
The predicate nf with type applies the type inference algorithm to the generated normal forms of size S.
nf_with_type(S,X,T):-normal_form(S,XT),infer_type(XT,X,T).
5.2
Merging in type inference
Like in the case of the set of simply-typed lambda terms, we can define the more
efficient combined generator and type inferrer predicate typed nf/2.
typed_nf(S,X,T):-typed_nf(_XT,X,T,[],S,0).
It works by calling the DCG-expended typed nf/4 predicate, with the last
two arguments enforcing the size constraints.
typed_nf(v(V:T),v(V),T,Vs)--> {
member(V:T0,Vs),
unify_with_occurs_check(T0,T)
}.
10
typed_nf(l(X:TX,A),l(X,NewA),(TX->TY),Vs)-->down,
typed_nf(A,NewA,TY,[X:TX|Vs]).
typed_nf(a(v(A),B),a(NewA,NewB),TY,Vs)-->down,
typed_nf(v(A),NewA,(TX->TY),Vs),
typed_nf(B,NewB,TX,Vs).
typed_nf(a(a(A1,A2),B),a(NewA,NewB),TY,Vs)-->down,
typed_nf(a(A1,A2),NewA,(TX->TY),Vs),
typed_nf(B,NewB,TX,Vs).
Example 11 Simply-typed normal forms up to size 3.
?- typed_nf(s(s(s(0))),Term,Type).
Term = l(A, l(B, l(C, v(C)))),
Type = (D->E->F->F) ;
...
Term = l(A, a(v(A), l(B, v(B)))),
Type = (((C->C)->D)->D) .
We are now able to efficiently generate counts for simply-typed normal forms
of a given size.
Example 12 Counts for closed simply-typed normal forms up to N=14.
first 10: 1,2,6,23,108,618,4092,30413,252590,2297954
11:
22,640,259
12:
240,084,189
13: 2,721,455,329
14: 32,783,910,297
Note that if we would want to just collect the set of types having inhabitants of
a given size, the preservation of typability under β-reduction property [3] would
allow us to work with the (smaller) set of simply-typed terms in normal form.
6
Experimental data
Figure 1 gives the number of logical inferences as counted by SWI-Prolog. This
is a good measure of computational effort except for counting operations like
unify with occurs check as a single step, while its actual complexity depends
on the size of the terms involved. Therefore, figure 2 gives actual timings for the
same operations above N=5, where they start to be meaningful.
The “closed λ-terms” column gives logical inferences and timing for generating all closed lambda terms of size given in column 1. The column “gen, then
infer” covers the algorithm that first generates lambda terms and then infers
their types. The column “gen + infer” gives performance data for the significantly faster algorithm that merges generation and type inference in the same
predicate. The column “inhabitants” gives data for the case when actual inhabitants are omitted in the merged generation and type inference process. The
11
column “typed normal form” shows results for the fast, merged generation and
type inference for terms in normal form.
As moving from a size to the next typically adds one order of magnitude of
computational effort, computing values for N=15 and N=16 is reachable with our
best algorithms for both simply typed terms and their normal form subset.
Size closed λ-terms gen, then infer gen + infer inhabitants typed normal form
1
15
19
16
9
19
2
44
59
50
28
47
3
166
261
188
113
127
4
810
1,517
864
553
429
5
4,905
10,930
4,652
3,112
1,814
6
35,372
92,661
28,878
19,955
9,247
7
294,697
895,154
202,526
143,431
55,219
8
2,776,174
9,647,495
1,586,880
1,146,116
377,745
9
29,103,799
114,273,833 13,722,618 10,073,400
2,896,982
10
335,379,436 1,471,373,474 129,817,948 96,626,916
24,556,921
Fig. 1. Number of logical inferences used by our generators, as counted by SWI-Prolog
Size closed λ-terms gen, then infer gen + infer inhabitants typed normal form
5
0.001
0.001
0.001
0.000
0.001
6
0.005
0.011
0.004
0.002
0.004
7
0.028
0.114
0.029
0.018
0.011
8
0.257
1.253
0.242
0.149
0.050
9
2.763
15.256
2.080
1.298
0.379
10
32.239
199.188
19.888
12.664
3.329
Fig. 2. Timings (in seconds) for our generators up to size 10 (on a 2015 MacBook,
with 1.3 GHz Intel Core M processor)
7
Discussion
An interesting open problem is if this can be pushed significantly farther. We
have looked into term hash based indexing and tabling-based dynamic programming algorithms, using de Bruijn terms. Unfortunately as subterms of closed
terms are not necessarily closed, even if de Bruijn terms can be used as ground
keys, their associated types are incomplete and dependent on the context in
which they are inferred.
12
While it only offers a constant factor speed-up, parallel execution is a more
promising possibility. However, given the small granularity of the generation
and type inference process, the most useful parallel execution mechanism would
simply split the task of combined generation and inference process into a number
of disjoint sets, corresponding to the number of available processors. A way to
do this, is by using unranking functions (bijections originating in N) to the sets
of combinatorial objects involved, and then, for k processors, assign work on
successive numbers belonging to the same equivalence class modulo k. Another
way is to first generate Motzkin trees and then launch threads to “flesh them
up” with logic variables, run the type inference steps and collect success counts
atomically.
We have not seen any obvious way to improve these results using constraint
programing systems, partly because the “inner loop” computation is unification
with occurs check with computations ranging over Prolog terms rather than
being objects of a constraint domain. On the other hand, for a given size, exploring grounding to propositional formulas or answer-set programming seems
worth exploring as a way to take advantage of today’s fast SAT-solvers.
Our techniques can be easily adapted to a different size definition like the
ones in [20,21] where variables in de Bruijn notation have a size proportional
to the distance to their binder. We have not discussed here the use of similar
techniques to improve the Boltzmann samplers described to [22], but clearly interleaving type checking with the probability-driven building of the terms would
improve their performance, by excluding terms with ill-typed subterms as early
as possible, during the large number of retries needed to overcome the asymptotically 0-density of simply-typed terms in the set of closed terms [6].
8
Related work
The classic reference for lambda calculus is [1]. Various instances of typed lambda
calculi are overviewed in [3].
The combinatorics and asymptotic behavior of various classes of lambda
terms are extensively studied in [5,8]. Distribution and density properties of
random lambda terms are described in [6].
Generation of random simply-typed lambda terms and its applications to
generating functional programs from type definitions is covered in [23].
Several concepts of size have been used in the literature, partly to facilitate
convergence of formal series in analytic combinatorics [20,21].
Asymptotic density properties of simple types (corresponding to tautologies
in minimal logic) have been studied in [24] with the surprising result that “almost
all” classical tautologies are also intuitionistic ones.
In [4] a “type-directed” mechanism for the generation of random terms is
introduced, resulting in more realistic (while not uniformly random) terms, used
successfully in discovering some bugs in the Glasgow Haskell Compiler (GHC).
A statistical exploration of the structure of the simple types of lambda terms of
13
a given size in [14] gives indications that some types frequent in human-written
programs are among the most frequently inferred ones.
Generators for closed simply-typed lambda terms, as well as their normal
forms, expressed as functional programming algorithms, are given in [5], derived
from combinatorial recurrences. However, they are significantly more complex
than the ones described here in Prolog, and limited to terms up to size 10.
In the unpublished draft [15] we have collected several lambda term generation algorithms written in Prolog and covering mostly de Bruijn terms and
a compressed de Bruijn representation. Among them, we have covered linear,
affine linear terms as well as terms of bounded unary height and in binary
lambda calculus encoding. In [15] type inference algorithms are also given for
SK and Rosser’s X-combinator expressions. A similar (but slower) program for
type inference using de Bruijn notation is also given in the unpublished draft
[15], without however describing the step-by-step derivation steps leading to it,
as done in this paper.
In [25] a general constraint logic programming framework is defined for sizeconstrained generation of data structures as well as a program-transformation
mechanism. While our fine-tuned interleaving of term generation and type inference directly provides the benefits of a CLP-based scheme, the program transformation techniques described in [25] are worth exploring for possible performance
improvements.
9
Conclusion
We have derived several logic programs that have helped solve the fairly hard
combinatorial counting and generation problem for simply-typed lambda terms,
4 orders of magnitude higher than previously published results.
This has put at test two simple but effective program transformation techniques naturally available in logic programming languages: 1) interleaving generators and testers by integrating them in the same predicate and 2) dropping
arguments used in generators when used simply as counters of solutions, as in
this case their role can be kept implicit in the recursive structure of the program. Both have turned out to be effective for speeding up computations without changing the semantics of their intended application. We have also managed
(after a simple DCG translation) to work within in the minimalist framework
of Horn Clauses with sound unification, showing that non-trivial combinatorics
problems can be handled without any of Prolog’s impure features.
Our techniques, combining unification of logic variables with Prolog’s backtracking mechanism and DCG grammar notation, recommend logic programming as a convenient meta-language for the manipulation of various families of
lambda terms and the study of their combinatorial and computational properties.
14
Acknowledgement
This research has been supported by NSF grant 1423324.
We thank the anonymous reviewers of LOPSTR’16 for their constructive
suggestions and the participants of the 9th Workshop Computational Logic and
Applications (https://cla.tcs.uj.edu.pl/) for enlightening discussions and
for sharing various techniques clarifying the challenges one faces when having a
fresh look at the emerging, interdisciplinary field of the combinatorics of lambda
terms and their applications.
References
1. Barendregt, H.P.: The Lambda Calculus Its Syntax and Semantics. Revised edn.
Volume 103. North Holland (1984)
2. Hindley, J.R., Seldin, J.P.: Lambda-calculus and combinators: an introduction.
Volume 13. Cambridge University Press Cambridge (2008)
3. Barendregt, H.P.: Lambda calculi with types. In: Handbook of Logic in Computer
Science. Volume 2. Oxford University Press (1991)
4. Palka, M.H., Claessen, K., Russo, A., Hughes, J.: Testing an optimising compiler
by generating random lambda terms. In: Proceedings of the 6th International
Workshop on Automation of Software Test. AST’11, New York, NY, USA, ACM
(2011) 91–97
5. Grygiel, K., Lescanne, P.: Counting and generating lambda terms. J. Funct.
Program. 23(5) (2013) 594–628
6. David, R., Raffalli, C., Theyssier, G., Grygiel, K., Kozik, J., Zaionc, M.: Some
properties of random lambda terms. Logical Methods in Computer Science 9(1)
(2009)
7. Bodini, O., Gardy, D., Gittenberger, B.: Lambda-terms of bounded unary height.
In: ANALCO, SIAM (2011) 23–32
8. David, R., Grygiel, K., Kozik, J., Raffalli, C., Theyssier, G., Zaionc, M.: Asymptotically almost all λ-terms are strongly normalizing. Preprint: arXiv: math.
LO/0903.5505 v3 (2010)
9. Flajolet, P., Sedgewick, R.: Analytic Combinatorics, year = 2009, isbn =
0521898064, 9780521898065, edition = 1, publisher = Cambridge University Press,
address = New York, NY, USA,
10. Sloane, N.J.A.: The On-Line Encyclopedia of Integer Sequences. (2014) Published
electronically at https://oeis.org/.
11. Tarau, P.: On Logic Programming Representations of Lambda Terms: de Bruijn Indices, Compression, Type Inference, Combinatorial Generation, Normalization. In
Pontelli, E., Son, T.C., eds.: Proceedings of the Seventeenth International Symposium on Practical Aspects of Declarative Languages PADL’15, Portland, Oregon,
USA, Springer, LNCS 8131 (June 2015) 115–131
12. Tarau, P.: Ranking/Unranking of Lambda Terms with Compressed de Bruijn
Indices. In Kerber, M., Carette, J., Kaliszyk, C., Rabe, F., Sorge, V., eds.: Proceedings of the 8th Conference on Intelligent Computer Mathematics, Washington,
D.C., USA, Springer, LNAI 9150 (July 2015) 118–133
13. Tarau, P.: On a Uniform Representation of Combinators, Arithmetic, Lambda
Terms and Types. In Albert, E., ed.: PPDP’15: Proceedings of the 17th international ACM SIGPLAN Symposium on Principles and Practice of Declarative
Programming, New York, NY, USA, ACM (July 2015) 244–255
15
14. Tarau, P.: On Type-directed Generation of Lambda Terms. In De Vos, M., Eiter,
T., Lierler, Y., Toni, F., eds.: 31st International Conference on Logic Programming
(ICLP 2015), Technical Communications, Cork, Ireland, CEUR (September 2015)
available online at http://ceur-ws.org/Vol-1433/.
15. Tarau, P.: A logic programming playground for lambda terms, combinators, types
and tree-based arithmetic computations. CoRR abs/1507.06944 (2015)
16. Stanley, R.P.: Enumerative Combinatorics. Wadsworth Publ. Co., Belmont, CA,
USA (1986)
17. Statman, R.: Intuitionistic Propositional Logic is Polynomial-Space Complete.
Theor. Comput. Sci. 9 (1979) 67–72
18. Wielemaker, J., Schrijvers, T., Triska, M., Lager, T.: SWI-Prolog. Theory and
Practice of Logic Programming 12 (1 2012) 67–96
19. Costa, V.S., Rocha, R., Damas, L.: The YAP Prolog system. Theory and Practice
of Logic Programming 12 (1 2012) 5–34
20. Grygiel, K., Lescanne, P.: Counting and Generating Terms in the Binary Lambda
Calculus (Extended version). CoRR abs/1511.05334 (2015)
21. Bendkowski, M., Grygiel, K., Lescanne, P., Zaionc, M.: A Natural Counting of
Lambda Terms. In Freivalds, R.M., Engels, G., Catania, B., eds.: SOFSEM 2016:
Theory and Practice of Computer Science - 42nd International Conference on Current Trends in Theory and Practice of Computer Science, Harrachov, Czech Republic, January 23-28, 2016, Proceedings. Volume 9587 of Lecture Notes in Computer
Science., Springer (2016) 183–194
22. Lescanne, P.: Boltzmann samplers for random generation of lambda terms. CoRR
abs/1404.3875 (2014)
23. Fetscher, B., Claessen, K., Palka, M.H., Hughes, J., Findler, R.B.: Making random judgments: Automatically generating well-typed terms from the definition of
a type-system. In: Programming Languages and Systems - 24th European Symposium on Programming, ESOP 2015, Held as Part of the European Joint Conferences
on Theory and Practice of Software, ETAPS 2015, London, UK, April 11-18, 2015.
Proceedings. (2015) 383–405
24. Genitrini, A., Kozik, J., Zaionc, M.: Intuitionistic vs. Classical Tautologies, Quantitative Comparison. In Miculan, M., Scagnetto, I., Honsell, F., eds.: Types for
Proofs and Programs, International Conference, TYPES 2007, Cividale del Friuli,
Italy, May 2-5, 2007, Revised Selected Papers. Volume 4941 of Lecture Notes in
Computer Science., Springer (2007) 100–109
25. Fioravanti, F., Proietti, M., Senni, V.: Efficient generation of test data structures
using constraint logic programming and program transformation. Journal of Logic
and Computation 25(6) (2015) 1263–1283
16
| 6cs.PL
|
arXiv:1710.02673v2 [math.AC] 24 Oct 2017
SYZYGIES OF COHEN–MACAULAY MODULES OVER ONE
DIMENSIONAL COHEN–MACAULAY LOCAL RINGS
TOSHINORI KOBAYASHI
Abstract. We study syzygies of (maximal) Cohen–Macaulay modules over one dimensional Cohen–Macaulay local rings. We compare these modules to Cohen–Macaulay
modules over the endomorphism ring of the maximal ideal. After this comparison, we
give several characterizations of almost Gorenstein rings in terms of syzygies of Cohen–
Macaulay modules.
1. Introduction
Throughout this paper, let (R, m) be a 1-dimensional singular local Cohen–Macaulay
ring which is generically Gorenstein, and set E = EndR (m). All subcategories are assumed
to be full. First of all, we give some notations.
Notation 1.1. Let A be a commutative Noetherian ring. We denote:
(1) by mod(A) the category of finitely generated A-modules,
(2) by CM(A) the subcategory of Cohen–Macaulay A-modules,
(3) by CM′ (A) the subcategory of Cohen–Macaulay A-modules without free summands,
(4) by ΩCM(A) the subcategory of first syzygies of Cohen–Macaulay A-modules, and
(5) by ΩCM′ (A) the subcategory ΩCM(A) ∩ CM′ (A).
Here we say that a finitely generated A-module M is Cohen–Macaulay if the localization
Mp satisfies depth Mp ≥ ht Ap for all prime ideals p of A.
The category ΩCM(A) has been used to study surface singularities. For example, the
relationship between ΩCM(A) and the category of special Cohen–Macaulay modules is
established in [11]. Furthermore, the following theorem is shown in [6, Corollary 3.3].
Theorem 1.2 (Dao–Iyama–Takahashi–Vial). Let A be an excellent henselian local normal
domain of dimension two with algebraically closed residue field. Then A has a rational
singularity if and only if ΩCM(A) is of finite type.
Here we say that a subcategory X of mod(A) is of finite type if there is a module
M ∈ mod(A) with add(M) = X , where add(M) is the additive closure of M in mod(A).
It is natural to ask what kind of one-dimensional local ring R is such that ΩCM(R) is
of finite type. To answer this question, we investigate the structure of ΩCM(R) to get
the following result.
Theorem 1.3. The endomorphism ring E is a commutative Cohen–Macaulay semilocal
ring, and the following hold true.
2010 Mathematics Subject Classification. 13C14, 13D02, 13H10.
Key words and phrases. Cohen–Macaulay ring/module, syzygy, almost Gorenstein ring.
1
2
TOSHINORI KOBAYASHI
(1) One has ΩCM(E) ⊆ ΩCM′ (R) ⊆ CM(E) ⊆ CM′ (R) as subcategories of mod(R).
(2) Assume that the residue field R/m is infinite. Then ΩCM(E) = ΩCM′ (R) if and only
if R has minimal multiplicity.
b of R is generically Gorenstein. Then CM(E) =
(3) Assume that the completion R
′
ΩCM (R) if and only if R is almost Gorenstein in the sense of [7].
The purpose of this paper is to give a proof of this theorem. As an application, we
construct some examples of rings R such that ΩCM(R) is of finite type. We also deduce
the following results.
Corollary 1.4. Assume that R is almost Gorenstein. Then ΩCM(R) is of finite type if
and only if CM(E) is of finite type.
We remark that there is an example of a ring R which is not almost Gorenstein such that
ΩCM(R) is of finite type (see Example 2.16). As another application, we also obtain a
result on the dimension of a triangulated category, which has been introduced by Rouquier
[16]. It turns out that each integer is realized as the dimension of the stable category of
Cohen–Macaulay modules.
Corollary 1.5. Let n > 0 be any integer. Let R = k[[H]] be the numerical semigroup
ring with H = h2n+1 , {2n+1 + 2i }i=1,...,n , 2n+1 + 3i. Then R is a complete intersection with
dim CM(R) = n,
where CM(R) stands for the stable category of CM(R).
2. The endomorphism ring of the maximal ideal
In this section, we give a proof of Theorem 1.3. Let A be a commutative Noetherian
ring. Recall the definition of reflexive modules. A finitely generated A-module M is
called reflexive if the natural map M → HomA (HomA (M, A), A) is an isomorphism. We
denote by Ref(A) the subcategory of reflexive A-modules. For any M ∈ mod(A), we
denote by TrM the Auslander transpose of M. For an integer n ≥ 1, we define Fn (A) =
{M | ExtiR (TrM, R) = 0 for i = 1, . . . , n} as a full subcategory of mod(A). A module
M ∈ mod(A) is called n-torsionfree if it is in Fn .
We have the following characterization of modules in CM(A), ΩCM(A), or Ref(A).
Lemma 2.1. (1) One has Ref(A) = F2 (A).
(2) Assume that A is a Cohen–Macaulay ring of dimension d such that Ap is Gorenstein
for every prime ideal p of A with ht p being at most d − 1. Then Fd (A) = CM(A) and
Fd+1 (A) = ΩCM(A).
Proof. (1) See [14, Proposition 12.5] for instance.
(2) See [5, Theorems 3.6 and 3.8] to prove the equality Fd (A) = CM(A) holds. The
proof of [11, Proposition 2.4] shows that Fd+1 (A) = ΩCM(A) holds.
We denote by Q(R) the total quotient ring of R and by R the integral closure of R in
Q(R). Then we can identify the ring E(= EndR (m)) with m :Q(R) m in the natural way.
Note that E is a one-dimensional Cohen–Macaulay semilocal ring contained in R because
SYZYGIES OF CM MODULES OVER ONE DIMENSIONAL CM LOCAL RINGS
3
E is a Cohen–Macaulay R-module. Using the propositions in previous section, we have
the following identifications.
Lemma 2.2. ΩCM(R) = F2 (R) = Ref(R) and ΩCM(E) = F2 (E) = Ref(E).
Proof. Since R is 1-dimensional and generically Gorenstein, the assumption of Lemma 2.1
(2) is satisfied. Thus ΩCM(R) = F2 (R). The ring E satisfies the same condition, because
the total quotient ring of E coincides with Q(R). Therefore ΩCM(E) = F2 (E). Finally,
we have F2 (R) = Ref(R) and F2 (E) = Ref(E) from Lemma 2.1 (1).
To prove Theorem 1.3, we will need the following lemma which is observed in Bass’s
“ubiquity” paper [2]. For the proof, see [14, Lemma 4.9].
Lemma 2.3. Let M ∈ ΩCM′ (R). Then M has an E-module structure which is compatible
with the action of R on M.
The module-finite ring extension R ⊆ E induces an inclusion CM(E) ⊆ CM(R) since
E is Cohen-Macaulay over R. Via this inclusion and the following lemma, we can view
CM(E) as a subcategory of CM(R).
Lemma 2.4. Let N, M ∈ CM(E). Then HomR (M, N) = HomE (M, N).
We prepare the following three lemmas about ΩCM(R). For a finitely generated Rmodule M, we denote by µ(M) the minimal number of generators of M and by ΩM
the (first) syzygy of M in the minimal free resolution. The first lemma follows from [12,
Lemma 2.1].
Lemma 2.5. Let M be a Cohen–Macaulay R-module. Then ΩM has no free summand.
α
Lemma 2.6. Let 0 → L → M −
→ N → 0 be an exact sequence of R-modules in CM(R).
If M is in ΩCM′ (R), then so is L.
β
γ
→ M′ → 0
Proof. Since M is in ΩCM′ (R), there is an exact sequence 0 → M −
→ R⊕a −
with a Cohen–Macaulay R-module M ′ . As M has no free summand, γ is a minimal free
cover. In particular, µ(M ′ ) = a. Taking the pushout of homomorphisms β and γ, we
have the following diagram.
0
0
0
L
/
/
M
α
N
/
/
0
/
0
β
0
/
L
/
R⊕a
δ
/
P
γ
M′
0
M′
0
4
TOSHINORI KOBAYASHI
The surjections R⊕a → P and P → M ′ show that a = µ(M ′ ) ≤ µ(P ) ≤ a. Thus δ is a
minimal free cover and L = ΩP . Lemma 2.5 implies that L has no free summand and
hence L is in ΩCM′ (R).
Lemma 2.7. The R-modules m and E belong to ΩCM′ (R).
Proof. Take a non zerodivisor x of R. Then R/(x) has depth zero. Therefore there is
a short exact sequence 0 → k → R/(x) → N → 0 with some R-module N. Applying
horseshoe lemma to this sequence, we have an exact sequence 0 → Ωk → R⊕n → ΩN → 0
for some n ≥ 1. Since ΩN is Cohen–Macaulay and Ωk = m, it follows that m ∈ ΩCM(R).
This implies that m is reflexive over R by Lemma 2.2. By [1, Corollary 5.7], m has no free
summand. It yields that E = HomR (m, m) = HomR (m, R). Thus E = Ω2 Trm = Ω(ΩTrm).
In particular, E is in ΩCM(R). If E has a nonzero free summand, then, since m is reflexive
and E = HomR (m, R), m must have a nonzero free summand. This is a contradiction. As
a consequence, m and E ∈ ΩCM′ (R).
Now we can give a proof of Theorem 1.3 (1).
Proof of Theorem 1.3 (1). To show the inclusion CM(E) ⊆ CM′ (R), we only need to
prove that all modules in CM(E) have no R-free summand. Assume that M ∈ CM(E)
has a nonzero R-free summand. Then E also has a nonzero R-free summand, since there
is a surjection E ⊕n → M. This contradicts Lemma 2.7.
The inclusion ΩCM′ (R) ⊆ CM(E) follows by Lemma 2.3.
Finally, we show ΩCM(E) ⊆ ΩCM′ (R). Let M be an E-module in ΩCM(E). Then we
have an exact sequence 0 → M → E ⊕n → N → 0 with some E-module N in CM(E). In
view of Lemmas 2.5 and 2.6, this sequence yields that M is in ΩCM′ (R).
The R-module m plays an important role to prove the theorem. Next lemma says that
m can be regarded as a “cogenerator” of ΩCM′ (R).
Lemma 2.8. Let M be an R-module in ΩCM′ (R). Then there is an exact sequence
0 → M → m⊕n → N → 0
of modules in CM(E).
α
Proof. As M is in ΩCM′ (R), we have an exact sequence 0 → M −
→ R⊕n → N ′ → 0
with a Cohen-Macaulay R-module N. Then, since M has no free summand, there is
a homomorphism β : M → m such that α = i ◦ β, where i is the natural inculusion
m⊕n → R⊕n . Let N be the cokernel of β. We have the following commutative diagram
with exact rows and columns.
0
/
M
α
/
R⊕n
O
/
NO ′
/
0
NO
/
0
i
0
/
M
β
/
m⊕n
O
0
/
0
Since β ∈ HomR (M, m) = HomE (M, m), N is a E-module. The exactness of 0 → N → N ′
implies that N is Cohen-Macaulay over R. Thus N is in CM(E).
SYZYGIES OF CM MODULES OVER ONE DIMENSIONAL CM LOCAL RINGS
5
To prove Theorem 1.3 (2), we will use the following lemma (cf. [4, Corollary 3,2]).
Lemma 2.9. Let A be a 1-dimensional Cohen–Macaulay generically Gorenstein ring and
I be an ideal of A containing a non zerodivisor such that HomA (I, I) is naturally isomorphic to A. If I is reflexive, then it is isomorphic to A.
Proof. Denote by σ the natural homomorphism I → HomA (HomA (I, A), A), which maps
a to [φ 7→ φ(a)] for a ∈ I and φ ∈ HomA (I, A). Then σ is isomorphism by the reflexivity
of I. Let ev : I ⊗A HomA (I, A) → A be the evaluation homomorphism, which maps a ⊗ φ
to φ(a). Let K be the kernel of ev, C be the cokernel of ev, and T be the image of ev.
we have the following exact sequences.
(2.9.1)
0 → K → I ⊗A HomA (I, A) → T → 0
(2.9.2)
0→T →A→C→0
Let Q(A) be the total quotient ring of A. Then the ideal I ⊗A Q(A) of Q(A) is isomorphic
to Q(A). Thus ev ⊗ Q(A) is an isomorphism. Then K ⊗A Q(A) = 0 = C ⊗A Q(A), and it
follows from Cohen-Macaulayness of A that HomA (K, A) = 0 = HomA (C, A). Applying
HomA (−, A) to the above sequences, we have HomA (T, A) ∼
= HomA (I ⊗A HomA (I, A), A)
and the following exact sequence
f
0 → A = HomA (A, A) −
→ HomA (T, A) → Ext1A (C, A) → 0.
Via the isomorphism HomA (T, A) ∼
= HomA (I ⊗A HomA (I, A), A), the map f corresponds
to the map HomA (ev, A) : A → HomA (I ⊗A HomA (I, A), A), which maps 1A to ev. By the
hom-tensor adjointness and composition with HomA (I, σ −1 ), we get an isomorphism
Ψ : HomA (I ⊗A HomA (I, A), A) → HomA (I, HomA (HomA (I, A), A) → HomA (I, I).
Consider the composition Ψ ◦ HomA (ev, A) : A → HomA (I, I). Then it maps 1A to
idI . Therefore Ψ ◦ HomA (ev, A) is an isomorphism by the assumption. It yields that
HomA (ev, A) and f are isomorphisms and hence Ext1A (C, A) = 0. Since A is 1-dimensional,
we obtain that C = 0. In this case, ev is a surjection and we can see that there is a
surjective map I → A. This yields that I is isomorphic to A.
Now we show the Theorem 1.3 (2).
Proof of Theorem 1.3 (2). Assume the equality ΩCM(E) = ΩCM′ (R) holds. Then m ∈
ΩCM′ (R) = Ref(E). Since HomE (m, m) = EndR (m) = E, we can apply Lemma 2.9 to
see that m is isomorphic to E. By [15, Proposition 2.5], it follows that R has minimal
multiplicity.
Conversely, assume that R has minimal multiplicity. Then m is isomorphic to E by
using [15, Proposition 2.5] again. Let M ∈ ΩCM′ (R). Lemma 2.8 yields that there is an
exact sequence 0 → M → m⊕n → N → 0 with N ∈ CM(E). Since m⊕n ∼
= E ⊕n , we have
′
M ∈ ΩCM(E). Thus ΩCM(E) = ΩCM (R).
We denote by ω a canonical module of R and set (−)† = HomR (−, ω). If ω exists,
then we can give an equivalent condition to the equality CM(S) = ΩCM′ (R) by using the
canonical dual (−)† .
6
TOSHINORI KOBAYASHI
Lemma 2.10. Assume that R has a canonical module ω. Then the equality CM(E) =
ΩCM′ (R) holds if and only if E † ∈ ΩCM′ (R).
Proof. The “only if” part is clear. Now we assume E † ∈ ΩCM′ (R). Let M be in CM(E).
Taking a free cover of M † over E, we get an exact sequence 0 → N → E ⊕n → M † → 0
with some E-module N. Since M † , E ∈ CM(R), N is also in CM(R). Applying (−)† to
this sequence, we have an exact sequence 0 → M → (E † )⊕n → N † → 0. Using Lemma
2.6, M is in ΩCM′ (R). This shows that CM(E) = ΩCM′ (R).
b of R is generically Gorenstein, then R has a canonical module
If the completion R
by [7, Proposition 2.7]. In this situation, we see in the next lemma that the condition
CM(E) = ΩCM′ (R) is stable under flat local extension.
Corollary 2.11. Let ϕ : (R, m) → (R′ , m′ ) be a flat local homomorphism such that
b of R is generically Gorenstein. Then CM(E) =
mR′ = m′ . Assume that the completion R
ΩCM′ (R) if and only if CM(EndR′ (m′ )) = ΩCM′ (R′ ).
b′ is also generically Gorenstein by [7, Proposition
Proof. Let E ′ = EndR′ (m′ ). Note that R
2.12]. In addition, ω ⊗R R′ is a canonical module of R′ . Therefore, by Lemma 2.10,
CM(E) = ΩCM′ (R) if and only if E † ∈ ΩCM′ (R), and CM(E ′ ) = ΩCM′ (R′ ) if and only
if HomR′ (E ′ , ω ⊗R R′ ) ∈ ΩCM′ (R). Here E ′ = E ⊗R R′ and hence HomR′ (E ′ , ω ⊗R R′ ) =
(E † ) ⊗R R′ . Lemma 2.2 implies that E † ∈ ΩCM′ (R) if and only if Ext1R (Tr(E † ), R) =
0 = Ext2R (Tr(E † ), R). The later condition is equivalent to the equations Ext1R′ (Tr((E † ) ⊗R
R′ ), R′ ) = 0 = Ext2R′ (Tr(E † ) ⊗R R′ ), R′ ) since R → R′ is faithfully flat and the Auslander
transpose is preserved under a base change.
b We have one more
Using the above lemma, we can replace R with the completion R.
′
equivalent condition to being CM(E) = ΩCM (R).
b of R is generically Gorenstein. Then E † ∈
Lemma 2.12. Assume that the completion R
ΩCM′ (R) if and only if E † ∼
= m.
Proof. Thanks to Corollary 2.11, we can assume that R is complete. If E † ∼
= m, then we
′
′
†
†
have E ∈ ΩCM (R). Conversely, we assume E ∈ ΩCM (R). Using Lemma 2.8, we get
an exact sequence
(2.12.1)
α
0 → E† −
→ m⊕m → N → 0
of modules in CM(E). By the Krull-Schmidt theorem on R, we have a unique decomposition m = m1 ⊕ · · · ⊕ mn , where mi are indecomposable R-modules. Then we obtain
E = EndR (m1 ) × · · · × EndR (mn ) as an R-algebra. The components Ei = EndR (mi ) of E
are local rings because of the indecomposablity of mi . Set ni the maximal ideal of E corresponding to the maximal ideal of Ei . Note that the localization (E † )ni = (HomR (E, ω))ni
is the canonical module of Ei and the localization mni is equal to (mi )ni . Thus, after
localizing at ni , the sequence (2.12.1) becomes split exact and (E † )ni is a direct summand
†
of (mi )⊕m
ni . The modules (E )ni and (mi )ni are both indecomposable. Hence we obtain an
isomorphism (E † )ni ∼
= (mi )ni by the Krull-Schmidt theorem. The homomorphism α is a
split injection, since it becomes a split injection after localizing at ni for all i = 1, . . . , n.
SYZYGIES OF CM MODULES OVER ONE DIMENSIONAL CM LOCAL RINGS
7
1
n
Therefore E † is isomorphic a direct summand of m⊕m . Set E † ∼
⊕ · · · ⊕ m⊕a
= m⊕a
1
n . Then
† ∼
the localization at ni shows that ai = 1. Consequencely, E = m1 ⊕ · · · ⊕ mn = m.
The following lemma will be used to prove Theorem 1.3 (3).
Lemma 2.13. Let A be a ring with total quotient ring T , A be the integral closure of A in
T , and X be an A-submodule of A containing A. If there is an isomorphism φ : A → X
of A-modules, then X = A.
Proof. Let i : A → X be the inclution homomorphism. Then φ−1 ◦ i : A → A is
an endmorphism of A. Hence it is a multiplication map by r for some r ∈ A. Since
r = φ−1 ◦ i : A → A is injective, 1/r is in T . We have 1 = i(1) = φ(r) = rφ(1) in A
and hence 1/r = φ(1) ∈ A. It means that 1/r is integral over A. Therefore we have an
equation of integral dependence
(1/r)n + a1 (1/r)n−1 + · · · + an = 0,
where ai ∈ A for all i = 1, . . . , n. Multiplying r n , we get 1 + r(a1 + · · · + an r n−1) = 0.
This equation yields that r is a unit of A. Thus the endmorphism r = φ−1 ◦ i : A → A is
an automorphism, i is an isomorphism, and A = X.
Assume that R is complete and has a inifite residue field. Then there is an R-submodule
K of Q(R) such that R ⊂ K ⊂ R, and as an R-module, K is a canonical module of R; see
[7, Corollary 2.9]. Using this module K, we have the following theorem, which essensially
proves Theorem 1.3 (3).
Theorem 2.14. Assume that R is complete and the residue field R/m is infinite. Let K
be an R-submodule of Q(R) such that R ⊂ K ⊂ R, and as an R-module, K is a canonical
module of R. Then the following conditions are equivalent.
(1) R is almost Gorenstein.
(2) K : m = E.
(3) m : K = m.
(4) K : E is isomorphic to m.
(5) K : E is in ΩCM′ (R).
(6) ΩCM′ (R) = CM(E).
Proof. (1) =⇒ (2): If R is Gorenstein, then there is nothing to prove. Thus we assume
that R is almost Gorenstein but not Gorenstein. In this case, [7, Theorem 3.16] says that
(E =)m : m = K : m.
(2) =⇒ (3): We can easely see m = K : (K : m) = K : E. Thus m : K = (K : E) :
K = (K : K) : E = R : E = m.
(3) =⇒ (1): We have mK ⊂ m ⊂ R. This implies that R is almost Gorenstein by [7,
Theorem 3.11].
(2) =⇒ (4): We have already got m = K : E. In particular, K : m is isomorphic to m.
(4) =⇒ (2): We assume that R is not Gorenstein. Then K : m ⊂ R[K] ⊂ R by
[7, Corollary 3.8]. On the other hand, K : m is isomorphic to (K : K) : E = E by the
assumption. Applying Lemma 2.13 to X = K : m and A = E, we obtain K : m = E.
(4) ⇐⇒ (5) and (5) ⇐⇒ (6): These follow by Lemmas 2.12 and 2.10.
8
TOSHINORI KOBAYASHI
Proof of Theorem 1.3 (3). By Corollary 2.11, we may assume that R is complete and has
infinite residue field. Thus the assersion follows by Theorem 2.14.
We end the section by giving some examples of rings R with ΩCM(R) is of finite type.
Example 2.15. Let R = k[[t3 , t7 , t11 ]]. Then E = k[[t3 , t4 ]], which is Gorenstein. Therefore ΩCM(E) = CM(E) and hence ΩCM(E) = ΩCM′ (R) = CM(E) by Theorem 1.3 (1).
It follows that R is almost Gorenstein and has minimal multiplicity by Theorem 1.3 (2)
and (3). Note that CM(E) is of finite type. This implies that ΩCM′ (R) has finitely many
indecomposable objects and thus ΩCM(R) is also of finite type.
Example 2.16. Let R = k[[t3 , t7 , t8 ]]. Then E = k[[t3 , t4 , t5 ]] and CM(E) is of finite type.
This yields ΩCM(R) is also of finite type. On the other hand, R is not almost Gorenstein.
3. The dimension of the stable category of CM(R)
In this section, as an application of our results obtained in the preceding sections, we
consider the dimension of the stable category of Cohen–Macaulay R-modules in the case
where R is Gorenstein. We shall construct for each integer n an example of a Gorenstein
ring R such that dim CM(R) = n.
Assume that R is a complete domain. Then E is a finite product of local rings because
R is henselian and E is module-finite over R. Since E is a subring of R, it is also a
domain. Thus E is a complete local domain. Using this argument inductively, we obtain
a family {(Rn , mn )}n of local rings with
E
if n = 1,
Rn =
EndRn−1 (mn−1 ) otherwise.
If R is Gorenstein, then as is well known, CM(R) is a Frobenius exact category. Thus
its stable category CM(R) is a triangulated category; see [8, Chapter 1] for details. In
this case, we can consider the dimension of the triangulated category CM(R) in the sense
of Rouquier [16].
Proposition 3.1. Assume that R is a complete Gorenstein domain. If CM(Rn ) is of
finite type, then dim CM(R) ≤ n − 1.
Proof. By the assumption, there is an Rn -module G′ such that CM(Rn ) = add(G′ ). Let
G = G′ ⊕ Rn−1 ⊕ · · · ⊕ R1 .
We claim that CM(R) = hGin . In fact, let M1 ∈ CM(R). To show that M1 ∈ hGin , we
may assume that M1 is in CM′ (R). Since R is Gorenstein, CM′ (R) = ΩCM′ (R) and hence
CM′ (R) = CM(R1 ) by Theorem 1.3 (1). Therefore M1 ∈ CM(R1 ). Taking a minimal free
resolution of M1 over R1 , we have an exact sequence 0 → M2 → R1⊕ → M → 0. Here M2
is in ΩCM′ (R1 ) by Lemma 2.5 and thus in CM(R2 ) by Theorem 1.3 (1). Applying this
argument inductively, we obtain R-modules Mi ∈ CM(Ri ) with exact sequences
(3.1.1)
⊕
0 → Mi → Ri−1
→ Mi−1 → 0
for each integer i ≥ 2. Since Mn ∈ CM(Rn ), it is obvious that Mn ∈ hGi1 . Then induction
shows that Mi ∈ hGin−i+1 by considering the exact triangle in CM(R) induced by the
sequence (3.1.1) for each i. In particular, M1 ∈ hGin .
This claim yields that dim CM(R) ≤ n − 1.
SYZYGIES OF CM MODULES OVER ONE DIMENSIONAL CM LOCAL RINGS
9
Example 3.2. Let n ≥ 1 be an integer and R = k[[H]] be the numerical semigroup
ring with H = h2n+1 , {2n+1 + 2i }i=1,...,n , 2n+1 + 3i. Then R is a complete intersection of
codimension n + 1. Thus dim CM(R) ≥ n by [3, Corollary 5.10]. On the other hand, we
can see that Rn+1 is isomorphic to k[[t2 , t3 ]]. Note that CM(k[[t2 , t3 ]]) is of finite type.
Therefore dim CM(R) ≤ n by Proposition 3.1. Consequently, one has dim CM(R) = n.
Remark 3.3. Let R be as in the above example. Since Rn+1 = k[[t2 , t3 ]], Rn+2 is equal to
k[[t]], which is the integral closure of R in its quotient field. Therefore, the endmorphism
ring Γ := EndR (R ⊕ R1 ⊕ · · · Rn+2 ) has global dimension less than or equal to n + 2; see
[10, Example 2.2.3(2)]and [13, Theorem 4]. It is known that dim CM(R) + 2 ≤ gl.dim Γ.
This shows that dim CM(R) ≤ n.
Acknowlegments. The author is grateful to his supervisor Ryo Takahashi for giving him
helpful advice throughout this paper. The author also thanks to Osamu Iyama for providing helpful suggestions on Theorem 1.3 and Remark 3.3.
References
[1] M. Auslander; S. Ding; Ø. Solberg, Liftings and weak liftings of modules, J. Algebra 156 (1993),
273–317.
[2] H. Bass, On the ubiquity of Gorenstein rings, Math. Zeitschrift 82 (1963), 8–28.
[3] P. A. Bergh; S. B. Iyengar; H. Krause; S. Oppermann, Dimensions of triangulated categories
via Koszul objects, Math. Z. 265 (2010), no. 4, 849–864.
[4] J. P. Brennan; W. V. Vasconcelos, On the structure of closed ideals, Math. Scand. 88 (2001),
no. 1, 3–16.
[5] E. G. Evans; P. Griffith, Syzygies, London Mathematical Society Lecture Note Series, 106,
Cambridge University Press, Cambridge, 1985.
[6] H. Dao; O. Iyama; R. Takahashi; C. Vial, Non-commutative resolutions and Grothendieck
groups, J. Noncommut. Geom. 9 (2015), no. 1, 21–34.
[7] S. Goto; N. Matsuoka; T. T. Phuong, Almost Gorenstein rings, J. Algebra, 379 (2013), 355–381.
[8] D. Happel, Triangulated categories in the representation theory of finite-dimensional algebras, London Mathematical Society Lecture Note Series, 119, Cambridge University Press, Cambridge, 1988.
[9] J. Herzog; E. Kunz, Der kanonische Modul eines. Cohen-Macaulay-Rings, Lecture Notes in Mathematics 238, Springer–Verlag, 1971.
[10] O.Iyama, Rejective subcategories of artin algebras and orders, arXiv:031128.
[11] O. Iyama; M. Wemyss, The classification of special Cohen–Macaulay modules, Math. Z. 265
(2010), no. 1, 41–83.
[12] T. Kobayashi, On delta invariants and indices of ideals, arXiv:1705.05042.
[13] G. J. Leuschke, Endomorphism rings of finite global dimension. Canad. J. Math. 59 (2007), no.
2, 332-342.
[14] G. J. Leuschke; R. Wiegand, Cohen–Macaulay Representations, Mathematical Surveys and
Monographs, vol. 181, American Mathematical Society, Providence, RI, 2012.
[15] A. Ooishi, On the self-dual maximal Cohen–Macaulay modules, J. Pure Appl. Algebra, 106, (1996),
93–102.
[16] R. Rouquier, Dimensions of triangulated categories, J. K-Theory 1 (2008), no. 2, 193–256.
Graduate School of Mathematics, Nagoya University, Furocho, Chikusaku, Nagoya,
Aichi 464-8602, Japan
E-mail address: [email protected]
| 0math.AC
|
arXiv:1610.05471v1 [math.OC] 18 Oct 2016
Optimal Communication Scheduling and Remote
Estimation over an Additive Noise Channel ?
Xiaobin Gao, Emrah Akyol, Tamer Başar
University of Illinois at Urbana-Champaign
Abstract
This paper considers a sequential sensor scheduling and remote estimation problem with one sensor and one estimator. The
sensor makes sequential observations about the state of an underlying memoryless stochastic process, and makes a decision
as to whether or not to send this measurement to the estimator. The sensor and the estimator have the common objective of
minimizing expected distortion in the estimation of the state of the process, over a finite time horizon. The sensor is either
charged a cost for each transmission or constrained on transmission times. As opposed to the prior work where communication
between the sensor and the estimator was assumed to be perfect (noiseless), in this work an additive noise channel with
fixed power constraint is considered; hence, the sensor has to encode its message before transmission. Under some technical
assumptions, we obtain the optimal encoding and estimation policies in conjunction with the optimal transmission schedule.
The impact of the presence of a noisy channel is analyzed numerically based on dynamic programming. This analysis yields
some rather surprising results such as a phase-transition phenomenon in the number of used transmission opportunities, which
was not encountered in the noiseless communication setting.
Key words: Stochastic Control; Sensor networks
1
1.1
Introduction
Background
The communication scheduling and remote state estimation problem arises in the applications of wireless sensor
networks, such as environmental monitoring and networked control systems. As an example of environmental monitoring, researchers at the National Aeronautics and Space Administration (NASA) Earth Science group are interested
in monitoring the evolution of the soil moisture, which is used in weather forecast, ecosystem process simulation and
the like [1]. In order to achieve that goal, the sensor networks are built over an area of interest. The sensors collect
data on the soil moisture and send them to the decision unit at NASA via wireless communication. The decision
unit at NASA forms estimates on the evolution of the soil moisture based on the messages received from the sensors.
Similarly, in networked control systems, where the objective is to control some remote plants, sensor networks are
built to measure the states of the remote plants. Sensors transmit their measurements to the controller via a wireless
communication network, and the controller estimates the state of the remote plant and generates a control signal
based on that estimate [2]. In both scenarios, the quality of the remote state estimation strongly a↵ects the quality
of decision making at the remote site, that is, weather prediction or control signal generation. The networked sensors
are usually constrained by limits on power [3]. They are not able to communicate with the estimator at every time
? This paper was not presented at any IFAC meeting. This research was supported in part by NSF under grant CCF 11-11342,
and in part by the U.S. Air Force Office of Scientific Research (AFOSR) MURI grants FA9550-10-1-0573 and N00014-16-12710.
Email addresses: [email protected] (Xiaobin Gao), [email protected] (Emrah Akyol), [email protected]
(Tamer Başar).
Preprint submitted to Automatica
16 October 2016
step and thus, the estimator has to produce its best estimate based on the partial information received from the sensors. Therefore, the communication between the sensors and the estimator should be scheduled judiciously, and the
estimator should be designed properly, so that the state estimation error is minimized subject to the communication
constraints.
1.2
Literature Review
Research on the general sensor scheduling problem dates back to the 1970s. In one of the earliest works [4], the
problem formulation is such that only one out of several sensors can be selected at each instant of time to observe
the output of a linear stochastic system. Using the measurements over a finite time interval, the goal is to form
prediction on some future state of the system. Furthermore, each sensor is associated with a certain measurement
cost. The author proposed an o↵-line deterministic sensor scheduling strategy that minimizes the sum of measurement
cost over the time interval and prediction error. Gupta et al [5] studied the sensor scheduling problem over infinite
time horizon. Similar to the problem in [4], only one sensor can be selected at each instant of time. However, there
is no measurement cost associated with each sensor. The authors proposed an o↵-line stochastic sensor scheduling
strategy such that the expected steady state estimation error is minimized. Yang and Shi [6] studied the o↵-line
sensor scheduling problem where there is only one sensor observing the state of a linear stochastic system. The
sensor can communicate with the remote estimator only a limited number of times. The objective was to minimize
the cumulative estimation error over a finite time horizon. It was shown that the optimal sensor scheduling strategy
is to distribute the limited communication opportunities uniformly over the time horizon. The authors of the papers
discussed above considered o↵-line sensor scheduling problems. “O↵-line sensor scheduling” means the sensor is
scheduled to take observation or conduct communication based on some a priori information about the system (e.g.
statistics of random variables, system matrices). The on-line information (e.g. sensor’s observation, battery’s energy
level) is not taken into account when making schedules. Some other selected work on o↵-line sensor scheduling
problems can be found in [7–9].
With the advances in hardware devices, sensors are endowed with stronger computational capabilities. Consequently,
the sensors are able to make schedules based on all the information they have (a priori information as well as on-line
information), which motivates the formulation of on-line sensor scheduling problems. Åström and Bernhardsson [10]
considered a state estimation problem with a first-order stochastic system. They compared the estimation error over
infinite time horizon obtained by periodic sampling and threshold event-triggered sampling. The periodic sampling
is one of the o↵-line sensor scheduling strategies while the threshold event-triggered sampling is one of the online sensor scheduling strategies. They showed that the threshold event-triggered sampling, which is also called
“threshold-based communication strategy”, leads to better performance in state estimation compared with periodic
sampling. The global optimality of threshold-based communication strategy in this contexts is proved later by Nar
and Başar [11]. Imer and Başar [12] considered the on-line sensor scheduling and remote state estimation problem
over a finite time horizon. In the formulation, the sensor is restricted to communicate only a limited number of
times. By considering the communication strategies within the class of threshold-based strategies, the paper has
shown that there exists a unique threshold-based communication strategy achieving the best performance on remote
state estimation. Furthermore, the optimal threshold can be computed by solving a dynamic programming equation.
Bommannavar and Başar [13] later extended the result of [12] to multi-dimensional systems. The continuous-time
version of the problem in [12] has been studied by Rabi et al [14]. Xu and Hespanha [15] considered the networked
control problem involving state estimation and communication scheduling, which can be viewed as a sensor scheduling
and remote estimation problem. They fixed the estimator to be Kalman-like and designed an event-triggered sensor
that minimizes the time average of the sum of the communication cost and estimation error over infinite time horizon.
They showed that the optimal communication strategy is deterministic and stationary, and is a function of the
estimation error. Wu et al [16] considered the sensor scheduling and estimation problem subject to constraints on the
average communication rate over infinite time horizon. The authors assumed that the sensor has noisy observations
on the system state. By restricting the sensor scheduling strategies to the threshold event-triggered class, they derived
the exact minimum mean square error (MMSE) estimator. However, the exact MMSE estimator is nonlinear and
thus computationally intractable. Under a Gaussian assumption on the a priori distribution, the authors derived
an approximate MMSE estimator, which is Kalman-like. Based on the approximated MMSE estimator, the authors
derived conditions on the thresholds so that the average sensor communication rate will not exceed its upper bound.
You and Xie [17] extended the work in [16] by deriving conditions on the thresholds so that the estimator is stable.
Han et al. [18] showed that if the sensor is fixed to apply some stochastic event-triggered strategy, then the exact
MMSE estimator is Kalman-like. Other selected work on remote estimation with event-based sensor operations can
be found in [19, 20]. The work in [16–18] can also be viewed as Kalman-filtering with scheduled observations, which
is related to Kalman-filtering with intermittent observations studied in [21, 22].
2
The approaches of [15, 16] involved fixing the communication strategies or estimation strategies to be of a certain
type and then deriving the corresponding optimal estimation strategies and communication strategies, respectively.
The approach of [12], on the other hand, is to derive the jointly optimal communication strategies and estimation
strategies. Similarly, Lipsa and Martins [23] considered the sensor scheduling and remote estimation problem where
the sensor is not constrained by communication times but is charged a communication cost. They proposed a
threshold event-triggered sensor and a Kalman-like estimator and proved that the proposed sensor and estimator are
jointly optimal, minimizing the sum of communication cost and estimation error over a finite time horizon. Nayyar
et al [24] considered a similar problem where the sensor is equipped with an energy harvesting sensor. In the work
of [24], the problem formulation is such that the sensor is constrained by the energy level of the battery and is also
charged a communication cost. It is shown in [24] that an energy dependent threshold event-triggered sensor and a
Kalman-like estimator are jointly optimal. Hence, the result of [24] can be viewed as generalization of the results
of [12,23]. In both [23] and [24], majorization theory was used to prove the optimality of the respective results, which
is closely related to the approach in [25].
It is worth drawing attention to the two di↵erent types of constraints that arise in the works mentioned above–hard
and soft constraints–as featured in the problem setups of [12] and [23]. In the problem of [12], the sensor can only
communicate for a pre-specified number of times. Such a communication constraint is called hard constraint. In
the work of [23], however, the sensor is charged a communication cost. This kind of communication constraint is
called soft constraint. In the problem with hard constraint, the communication strategy must take the remaining
communication opportunities as a variable and schedule no communication if there is no remaining opportunity.
Such communication strategies guarantee that the number of transmissions made over the time horizon of interest
will not exceed the given constraint. In the problem with soft constraint, however, the sensor is not constrained
by the number of transmissions, therefore the communication strategy need not take the remaining communication
opportunities (which is always equal to the remaining time steps) as a variable. Therefore, the results obtained in
one problem cannot be applied directly to the other. For example, if we apply the communication strategy obtained
in [23] to the problem of [12], then there exists a positive probability that the sensor decides to communicate at every
instant of time, which certainly would violate the hard constraint on communication times. A detailed discussion on
the di↵erence between optimization problems with soft constraints and hard constants can be found in [26, 27].
1.3
Contributions
In this paper, we extend the lines of research in [12], [23], and [24]. In previous works, the communication between
the sensor and the estimator was assumed to be perfect (no channel noise), which may not be realistic, even though
it was an important first step. This paper investigates the e↵ect of communication channel noise on the design
of optimal sensor scheduling and remote estimation strategies. The paper consists of two parts: In the first part,
we consider a discrete time sensor scheduling and remote estimation problem with a soft constraint. At each time
step, the sensor makes a perfect observation on the state of an independent identically distributed (i.i.d.) source.
Next, the sensor decides whether to transmit its observation to the remote estimator or not. The sensor has a soft
communication constraint (i.e., the sensor is charged a cost for each transmission). Since the communication channel
is noisy, the sensor encodes the message before transmitting it to the estimator. The remote estimator generates
a real-time estimate on the state of the source based on the noise corrupted messages received from the sensor.
The estimator is charged for estimation error. Our goal is to design the communication scheduling strategy and
encoding strategy for the sensor, and the estimation strategy (decoding strategy) for the estimator, to minimize the
expected value of the sum of communication cost and estimation cost over the time horizon. Our solution consists of
a threshold-based communication scheduling strategy, and a pair of piecewise linear encoding/decoding strategies.
We prove optimality under some technical assumptions. Then, we extend the result to the problem with a hard
constraint. Using a dynamic programming approach, we obtain the optimal communication scheduling, encoding
and decoding strategies. Beyond the qualititively expected results, we notice some rather surprising e↵ects of the
noisy communication considerations in this class of remote estimation problems. For example, over a time horizon T
and with a hard transmission limit, N T , if the state realizations were so that at time step K, the sensor has used
only N T + K transmissions out of N , the intuitively appealing solution to the noiseless variation of the problem
was to transmit from that time on all the observed state realizations without any thresholding, i.e., the threshold is
e↵ectively set to zero for samples at time steps K + 1, . . . , T . However, in the noisy setting, we have noticed that this
is not the case, the sensor may not use all the transmission opportunities left. This is due to the fact that threshold
information–that is whether or not the state sample belongs to an interval– may be more valuable than a “noisy”
observation of the state. In fact, depending on the signal-to-noise ratio (SNR) of the channel, there is a fixed number
of useful (in average) number of transmissions, and allowing transmissions more than this number, on the average,
does not help decrease the expected mean square error (MSE).
3
1.4
Organization
The rest of the paper is organized as follows: in Section 2, we formulate the optimization problems with soft/hard
constraints. In Sections 3 and 4, we present the main results for problems with soft/hard constraints. In Section 5,
we present some numerical results for the problem with hard constraint. Finally, in Section 6, we draw concluding
remarks and discuss future work.
The authors have five conference papers on the general topic of this paper, listed as [31]-[35]. The specific topics and
results of the last three, that is [33]-[35], are beyond the scope of this paper, as explained in Section VI (Conclusions).
The first two, that is [31] and [32], have some overlap as far as the problem formulations go, but the current paper
substantially improves upon the results in these two papers, as explained in Remark 9 in Section IV.
2
2.1
Problem Formulation
System Model
Fig. 1. System model
Consider a discrete time communication scheduling and remote estimation problem over a finite-time horizon, that
is, t = 1, 2, . . . , T . There is one sensor, one encoder and one remote estimator (which is also called “decoder”). A
source process {Xt } is a one-dimensional, independent, and identically distributed (i.i.d.) stochastic process, which
has density pX . At time t, the sensor observes Xt . Since the sensor is assumed to have communication constraint
(which will be introduced later), it needs to decide whether or not to transmit its observation. Let Ut 2 {0, 1} be
the sensor’s decision at time t, where Ut = 1 stands for transmission and Ut = 0 stands for no transmission. The
communication channel is assumed to be noisy. Hence, if the sensor decides to transmit its observation, it sends Xt
to the encoder. If the sensor decides not to transmit, it does not send anything to the encoder but a free symbol ✏
stands for its decision. Denote by X̃t the message received by the encoder; then
X̃t =
(
Xt ,
if Ut = 1
✏,
if Ut = 0
If the encoder receives Xt from the sensor, it sends an encoded message Yt to the communication channel. The
encoder operates under the average power constraint:
E[Yt2 |Ut = 1] PT ,
where the expectation is taken over Yt . Furthermore, PT is known and is invariant of time. The encoded message
Yt is corrupted by an additive channel noise Vt . The noise process {Vt } is a one-dimensional i.i.d. stochastic process
with density pV . When sending Yt to the communication channel, the encoder is able to transmit the sign of Xt to
the decoder via a side channel, which is assumed to be noise-free. If the encoder receives ✏ from the sensor, it does
not send anything to the communication channel, while it sends ✏ to the side channel. Consequently, the decoder
can deduce the sensor’s decision from the message conveyed via the side-channel . We use Ỹt and St to denote the
messages received by the decoder from the communication channel and the side channel, respectively, that is
Ỹt =
⇢
Y t + Vt ,
Vt ,
if Ut = 1
,
if Ut = 0
St =
⇢
sgn(Xt ),
✏,
if Ut = 1
if Ut = 0
After receiving Ỹt and St , the decoder produces an estimate on Xt , denoted by X̂t . The estimator will be charged
for distortion in estimation. We assume that the distortion function ⇢(Xt , X̂t ) is the squared error, (Xt X̂t )2 .
4
2.2
Communication Constraint
The sensor is said to have a soft constraint if there is a positive cost function associated with Ut , denoted by C(Ut ).
Here, the cost function is assumed to have the form of
C(Ut ) = cUt =
(
0,
if Ut = 0
c,
if Ut = 1
,
where c is called the communication cost (c > 0), which is known and is invariant of time. The sensor is said to have
a hard constraint if it is restricted to use the noisy channel for no more than N times, i.e.,
T
X
t=1
1{Ut =1} N,
where 1{·} is the indicator function, and 0 < N < T .
2.3
Decision Strategies
Assume that at time t, the sensor has memory on all its observations up to t, denoted by X1:t , and all the decisions
it has made up to t 1, denoted by U1:t 1 . The sensor determines whether or not to transmit its observation at time
t, based on its current information (X1:t , U1:t 1 ), namely
Ut = ft (X1:t , U1:t
1 ),
where ft is the communication scheduling policy at time t, and f = {f1 , f2 , . . . , fT } is the communication scheduling
strategy.
Similarly, at time t, the encoder has memory on all the messages received from the sensor up to t, denoted by X̃1:t ,
and all the messages it has sent to the communication channel and the side channel up to t 1, denoted by Y1:t 1
and S1:t 1 , respectively. The encoder generates the encoded message at time t, based on its current information
(X̃1:t , Y1:t 1 , S1:t 1 ), namely
Yt = gt (X̃1:t , Y1:t 1 , S1:t 1 ),
where gt is the encoding policy at time t, and g = {g1 , g2 , . . . , gT } is the encoding strategy.
Finally, we assume that at time t, the decoder has memory on all the messages received from the communication
channel up to t, denoted by Ỹ1:t , and all the messages received from the side channel up to t, which are S1:t . The
decoder generates the estimate at time t, based on its current information (Ỹ1:t , S1:t ), namely
X̂t = ht (Ỹ1:t , S1:t ),
where ht is the decoding policy at time t, and h = {h1 , h2 , . . . , hT } is the decoding strategy.
Remark 1. Although we do not assume that the decoder has memory on its previous estimates up to t, yet it can
deduce them from (Ỹ1:t 1 , S1:t 1 ) and h1 , h2 , . . . , ht 1 .
For simplicity, we call the sensor, the encoder, and the decoder as “decision makers”. Correspondingly, we call the
communication scheduling policy, encoding policy, and decoding policy as “decision policies”.
2.4
Optimization Problem
Consider the settings described above, with the time horizon T , the probability density functions pX and pV , and
the power constraint PT as given.
5
Optimization problem with soft constraint: Given the communication cost c, determine (f, g, h) minimizing the function
( T
)
X
2
J(f, g, h) = E
cUt + (Xt X̂t ) .
t=1
Optimization problem with hard constraint: Given the number of transmission opportunities N < T , determine
(f, g, h) minimizing, under the hard constraints, the function
J(f, g, h) = E
3
(
T
X
(Xt
X̂t )
t=1
2
)
.
Optimization Problem with Soft Constraint
To begin with, we show that the optimization problem with soft constraint can be simplified to a “one-stage”
problem, as described in the theorem below.
Theorem 1. Consider the optimization problem in Section 2-D with the soft constraint.
(1) Without loss of optimality, one can restrict all the decision makers to apply the decision policies (ft , gt , ht ) in
the forms of:
Ut = ft (Xt ), Yt = gt (X̃t ), X̂t = ht (Ỹt , St ).
(1)
(2) Without loss of optimality, one can restrict all the decision makers to apply stationary decision strategies (f, g, h),
i.e.,
f = {f, f, . . . , f }, g = {g, g, . . . , g}, h = {h, h, . . . , h}.
Before proving Theorem 1, we first introduce the following notations. For any 1 a b T , let fa:b , ga:b , ha:b
denote the subsets of f, g, h such that
fa:b = {fa , fa+1 , . . . , fb },
ga:b = {ga , ga+1 , . . . , gb },
ha:b = {ha , ha+1 , . . . , hb }.
Furthermore, let Ist , Iet , Idt denote the information about the past system states available to the sensor, the encoder,
and the decoder, respectively, at time t (t > 1), i.e.,
Ist = {X1:t
1 , U1:t 1 },
Iet = {X̃1:t
1 , Y1:t 1 , S1:t 1 },
Idt = {Ỹ1:t
1 , S1:t 1 }.
Furthermore, let It be union of Ist , Iet , and Idt , i.e.,
It = {X1:t
1 , U1:t 1 , X̃1:t 1 , Y1:t 1 , S1:t 1 , Ỹ1:t 1 }.
Proof of Theorem 1: It is easy to see the validity of the following sequence of equalities:
min J(f, g, h)
f,g,h
= min E
f,g,h
=
=
nP
T
t=1
cUt + (Xt
⇢
min E cU1 + (X1
f1 ,g1 ,h1
⇢
min E cU1 + (X1
f1 ,g1 ,h1
X̂t )2
X̂1 )2 +
o
min
f2:T ,g2:T ,h2:T
E
nP
T
t=2
n
X̂1 )2 + min E cU2 + (X2
f2 ,g2 ,h2
cUt + (Xt
X̂t )2
X̂2 )2 + . . . +
o
min
fT ,gT ,hT
E cUT + (XT
Then, at time t = T , the optimization problem is to design (fT , gT , hT ) minimizing
n
o
2
JT1 (fT , gT , hT ) := E cUT + (XT X̂T ) ,
6
X̂T )2 . . .
o
call it Problem T1. Recall that the decisions at time T are generated by
UT = fT (XT , IsT ),
YT = gT (X̃T , IeT ),
X̂T = hT (ỸT , ST , IdT ).
We will now show that using information about the past (IsT , IeT , IdT ) when making decisions cannot help improve
the performance (that is, reduce the expected cost). Consider another problem, call it Problem T2, where IT is
available to all the decision makers, and one needs to design (fT0 , gT0 , h0T ) minimizing
n
JT2 (fT0 , gT0 , h0T ) := E cUT + (XT
X̂T )
2
o
,
where
UT = fT0 (XT , IT ), YT = gT0 (X̃T , IT ), X̂T = h0T (ỸT , ST , IT ).
Since the sensor, the encoder, and the decoder can always ignore the redundant information and behave as if they
only know IsT , IeT , IdT , respectively, the performance of the system in Problem T2 is no worse than that in Problem
T1, i.e.,
min
JT2 (fT0 , gT0 , h0T ) min JT1 (fT , gT , hT ).
0
0
0
(fT ,gT ,hT )
(fT ,gT ,hT )
Similarly, consider a third problem, call it Problem T3, where IsT , IeT , IdT are not available to the sensor, the
encoder, and the decoder, respectively. One needs to design (fT00 , gT00 , h00T ) to minimize
n
JT3 (fT00 , gT00 , h00T ) = E cUT + (XT
o
X̂T )2 ,
where
UT = fT00 (XT ), YT = gT00 (X̃T ), X̂T = h00T (ỸT , ST )
By a similar argument as above, the system in Problem T1 cannot perform worse than the system in Problem T3.
Hence,
min JT1 (fT , gT , hT ) 00min
JT3 (fT00 , gT00 , h00T ).
00
00
(fT ,gT ,hT )
(fT ,gT ,hT )
Let us come back to Problem T2. One can observe that the communication cost c, the distortion function ⇢(·, ·),
and the power constraint of the encoder do not depend on IT . Furthermore, since {Xt } and {Vt } are i.i.d. stochastic
processes, XT and VT are also independent of IT . Therefore, there is no loss of optimality in restricting
UT = fT0 (XT ), YT = gT0 (X̃T ), X̂T = h0T (ỸT , ST ),
and thus
min
0 ,h0 )
(fT0 ,gT
T
JT2 (fT0 , gT0 , h0T ) =
min
00 ,h00 )
(fT00 ,gT
T
JT3 (fT00 , gT00 , h00T )
The equality above indicates that in Problem T1 the sensor, the encoder, and the decoder can ignore their information
about the past, namely IsT , IeT , and IdT , respectively, and there is no loss of optimality in restricting
UT = fT (XT ), YT = gT (X̃T ), X̂T = hT (ỸT , ST ).
Since (fT , gT , hT ) do not take IT as a parameter, the design of (fT , gT , hT ) is independent of the design of
(f1:T 1 , g1:T 1 , h1:T 1 ). Consequently, the problem can be viewed as a T 1 stage problem and a one stage problem. By induction, we can show that (f1 , g1 , h1 ), (f2 , g2 , h2 ), . . . , (fT , gT , hT ) can be designed independently, and
(ft , gt , ht ) is designed to minimize the stage-wise cost E{cUt + (Xt X̂t )2 }. Hence, the optimal decision policies
(ft , gt , ht ) are in the form of (1). Furthermore, since {Xt } and {Vt } are i.i.d. stochastic processes, the optimal
decision policies (ft , gt , ht ) should be the same for all t = 1, 2, . . . , T .
By Theorem 1, the optimization problem with soft constraint can be reduced to a “one-stage” problem. Therefore,
for simplicity we suppress the subscript for time in all the expressions for the rest of this section. To present our
main results for the one-stage problem, we need the following four assumptions:
Assumption 1. The source density pX is nonatomic, even, and log-concave with support R. Furthermore, pX is
continuously di↵erentiable on (0, 1) (and on ( 1, 0) by symmetry).
7
Remark 2. There are several probability density functions satisfying Assumption 1, e.g., Gaussian distribution with
zero mean, Laplace distribution and a few others. For simplicity, we assume here that pX has support R. However,
the results also hold for the source density with support ( a, a), a > 0, e.g., uniform distribution. In this case, we
require that pX is continuously di↵erentiable on (0, a).
f
Given any communication scheduling policy f , let T0f , T1+
, and T1f be the non-transmission region, the positive
transmission region and the negative transmission region, respectively, where
f
T0f := {x 2 R|f (x) = 0}, T1+
:= {x > 0|f (x) = 1}, T1f := {x < 0|f (x) = 1}
Note that T0f , T1f , T2f may not be connected regions. Then, we make the following assumption on the communication
scheduling policy.
Assumption 2. The sensor is restricted to apply the communication scheduling policy f satisfying
f
E[X|X 2 T1f ] < E[X|X 2 T0f ] < E[X|X 2 T1+
]
(2)
Remark 3. There is a wide class of communication scheduling policies satisfying inequality (2). For example, given
any even communication scheduling policy f , i.e.,
f (x) = f ( x) 2 {0, 1},
and any even source density function pX , we have
E[X|X 2 T1f ] < 0,
E[X|X 2 T0f ] = 0,
f
E[X|X 2 T1+
] > 0,
Then, Assumption 2 is satisfied.
Assumption 3. The communication channel noise V has zero mean, and finite variance, denoted by
more, V and X are independent conditioned on sgn(X).
Assumption 4. The encoder and the decoder are restricted to apply piecewise affine policies:
g(X̃) =
h(Ỹ , S) =
where
= PT /
2
V
, ↵(S) =
p
(
S · ↵(S) · (X
0,
8
>
<S ·
E [X|U = 1, S]) ,
2
V
. Further-
if U = 1
if U = 0
1
↵(S)
+1
Ỹ + E [X|U = 1, S] ,
>
:E[X|U = 0],
if U = 1
if U = 0
PT /Var(X|U = 1, S), and Var(X|U = 1, S) is the conditional variance.
One can see that, when U = 1, the encoder can deduce (X, S) from X̃, namely X = X̃ and S = sgn(X̃). Furthermore,
when applying the encoding policy described above, the power consumption of the encoder can be computed as
⇥
⇤
P
E g 2 (X̃)|U = 1 = i2{
P
= i2{
=
P
i2{
E[g 2 (X̃)|U = 1, S = i] P(S = i)
h
i
2
2
E[X|U = 1, S = i] |U = 1, S = i P(S = i)
1,1} E ↵ (S) X
h
i
PT
2
E
X
E[X|U
=
1,
S
=
i]
|U
=
1,
S
=
i
P(S = i)
1,1}
Var(X|U = 1, S = i)
1,1}
= PT ,
which satisfies the average power constraint. Moreover, the events U = 0, (U = 1, S = 1), and (U = 1, S = 1) are
f
equivalent to the events X 2 T0f , X 2 T f , and X 2 T1+
, respectively. Therefore, the encoding and decoding policies
(g, h) are determined by the source density pX and the communication scheduling policy f . For simplicity, we use
J(f ) instead of J(f, g, h) to denote the cost function in the rest of this section.
8
Remark 4. Note that the assumption of piece-wise affine encoding policies originates from a prior work [28], which
analyzed a memoryless zero-sum jamming game between a pair of transmitter and receiver and an adversary that
generates an additive channel noise subject to second order (power) statistical constraints. It was shown in [28] that
the saddle-point equilibrium associated with this zero-sum game is achieved by linear encoding/decoding policies for
the transmitter & receiver pair. Here, we utilize such piece-wise linear policies, not only because they facilitate a
tractable analysis but also because they posess such mini-max robustness properties (see [28] for more details).
Remark 5. Since the communication channel noise V depends on the sign of X, the encoder and the decoder will
apply di↵erent encoding/decoding policies for positive/negative realizations of the source.
Theorem 2. Consider the one-stage problem under Assumptions 1-4. Then, the optimal communication scheduling
policy is of the symmetric threshold type:
f (x) =
(
0,
if |x| <
1,
otherwise
,
where > 0 is the threshold. Furthermore, there exists a unique value
all such thresholds.
⇤
minimizing the cost function J(f ) among
To prove Theorem 2, we need the following definitions and lemmas. We first introduce a quantization problem.
Quantization Problem: The problem is one of quantizing the realizations (denoted by x) of a real-valued random
variable (denoted by X) to N codepoints (N is finite and is known) according to some quantization rule (or quantizer)
Q, i.e,
Q(x) = qi , if x 2 Si , i 2 {1, 2, . . . , N },
where S1 , S2 , . . . , SN are called quantization regions and q1 , q2 , . . . , qN are the corresponding codepoints. Note that
S1 , S2 , . . . , SN are mutually disjoint sets and their union equals R. The distortion error between a realization x
and the its quantized value Q(x) is ⇢(|x Q(x)|), where ⇢ : [0, 1) ! [0, 1) is called the distortion function. The
performance of the quantizer Q is evaluated by its mean distortion error, denoted by D(Q), i.e.,
h
i
D(Q) := E ⇢ |X Q(X)|
Then, given the probability distribution of X, the optimization problem is to design a quantizer Q = Q⇤ (i.e., design
{S1 , S2 , . . . , SN } and {q1 , q2 , . . . , qN }) that minimizes D(Q).
We recall here a result on the regularity of the optimal quantizer, which we will use shortly.
Lemma 1 ([29], Theorem 1 and Corollary 1). Assume that the source X has nonatomic distribution pX , and
⇢ : [0, 1) ! [0, 1) is convex and nondecreasing. Then, for any N -level quantizer Q with quantization regions
{S1 , S2 , . . . , SN } and the corresponding codepoints {q1 , q2 , . . . , qN }, there exists a quantizer Q̂ with quantization
regions {Ŝ1 , Ŝ2 , . . . , ŜN } and the corresponding codepoints {q̂1 , q̂2 , . . . , q̂N } such that
(1)
(2)
(3)
(4)
Ŝi is convex, and P(X 2 Ŝi ) = P(X 2 Si ), for all i = 1, . . . , N .
If qi < qj , then Ŝi < Ŝj , i.e., x < y for any x 2 Ŝi and y 2 Ŝj .
q̂i 2 Ŝi , for all i = 1, . . . , N .
D(Q̂) D(Q).
Now returning to our problem, for any communication scheduling policy f , we can construct a three level quantizer,
f
denoted by Qf , with quantizing regions (T0f , T1+
, T1f ) and the corresponding codepoints (E[X|X 2 T0f ], E[X|X 2
f
f
f
T1+ ], E[X|X 2 T1 ]). Let D(Q ) be the mean squared distortion of Qf , i.e.,
D(Qf ) = E
h
X
Qf (X)
2
i
=
X
i2{0,1+,1 }
=
X
i2{0,1+,1 }
h
E X
i
2
E[X|X 2 Tif ] |X 2 Tif P(X 2 Tif )
Var(X|X 2 Tif ) P(X 2 Tif )
By Lemma 1, we have the following result.
9
Lemma 2. Suppose the source density pX is nonatomic and even. Then, for any communication scheduling policy
f satisfying Assumption 2, we can construct a threshold-based communication scheduling policy f (1) such that
(1) T0f
(1)
=(
2,
f (1)
1 ),T1+
(1)
Tif )
(2) P(X 2
= P(X 2
(1)
(3) D(Qf ) D(Qf ).
f (1)
1 , 1),T1
=(
Tif ),
= ( 1,
2 ),
where
> 0,
1
2
> 0 are thresholds.
for all i 2 {0, 1+, 1 }.
f
Proof. By Lemma 1, given a 3-level quantizer Qf with quantization regions (T0f , T1+
, T1f ) and the corresponding
f
codepoints (E[X|X 2 T0f ], E[X|X 2 T1+
], E[X|X 2 T1f ]), there exists a 3-level quantizer Q̂ with quantization
regions (T0f
Tif
(1)
(1)
(1)
f
, T1+
, T1f
(1)
) and corresponding codepoints (ĉ0 , ĉ1+ , ĉ1 ) such that D(Q̂) D(Qf ). Furthermore,
is convex and P(X 2 Tif
f
E[X|X 2 T1+
], we have T1f
pX is even, we have
(1)
(1)
) = P(X 2 Tif ), for all i 2 {0, 1+, 1 }. Since E[X|X 2 T1f ] < E[X|X 2 T0f ] <
< T0f
(1)
(1)
f
f
< T1+
. Note that since T1+
✓ (0, 1), T1f ✓ ( 1, 0), and the source density
1
,
2
(1)
f
f
P(X 2 T1+
) = P(X 2 T1+
)
(1)
P(X 2 T1f
(1)
(1)
) = P(X 2 T1f )
1
,
2
(1)
f
which implies that T0f = ( 2 , 1 ),T1+
= ( 1 , 1), and T1f = ( 1,
0. Also note that
2 ), for some 1 , 2
when the distortion function is the squared error, the optimal codepoints corresponding to quantization regions
(1)
(1)
(1)
(1)
f (1)
f (1)
(T0f , T1+
, T1f ) are (E[X|X 2 T0f ], E[X|X 2 T1+
], E[X|X 2 T1f ]). Hence, we have D(Q(1) ) D(Q̂)
D(Qf ).
Based on Lemma 2, we further propose a proposition, and we will prove it later.
Proposition 1. Suppose the source density pX satisfies Assumption 1. Then, for any communication scheduling
policy f satisfying Assumption 2, we can construct a symmetric threshold-based communication scheduling policy
f (2) such that
(1) T0f
(2)
=(
(2)
f
, ),T1+
= ( , 1),T1f
(2)
T0f )
(2) P(X 2
= P(X 2
(2)
(3) D(Qf ) D(Qf ).
(2)
T0f ).
= ( 1,
), where
> 0.
To prove Proposition 1, we need to apply results from majorization theory, introduced below. Given a Borel measurable set A, we use A to denote its symmetric rearrangement, i.e., A = [ a, a], and L(A ) = L(A). Given a
nonnegative integrable function p : R ! R, we use p to denote its symmetric rearrangement, which is described as
follows,
Z
1
p (x) :=
0
1{z2R|p(z)
⇢}
1{z2R|p(z)
⇢}
(x)d⇢,
x 2 R.
(x) is the indicator function on whether x is an element of {z 2 R|p(z)
1{z2R|p(z)
⇢}
(x) =
(
1,
if x 2 {z 2 R|p(z)
0,
otherwise
⇢} or not, i.e.,
⇢}
Definition 1. Given two probability densities p and q defined on R, we say p majorizes q, denoted by p
Z
Z
q (x)dx
p (x)dx, for all t 0
|x|<t
q, if
|x|<t
Lemma 3 ([23], Lemma 4). Let pX and pX 0 be probability density functions defined on R. Assume that pX is even
and logconcave, and pX pX 0 . Then,
Z 1
Z 1
2
x pX (x)dx
(x y)2 pX 0 (x)dx, for all y 2 R,
1
1
10
or equivalently,
Var(X) Var(X 0 )
Remark 6. Lemma 4 in [23] assumes that pX is even, quasi-concave, and there exists b 2 R such that pX is nondecreasing on ( 1, b] and non-increasing on (b, 1). Note that a positive logconcave function is also quasi-concave.
Moreover, it can be easily shown that if pX is even and logconcave. In view of this, pX is non-decreasing on ( 1, 0]
and non-increasing on (0, 1), and hence as stated above Lemma 3 is a valid slightly modified version of Lemma 4
of [23].
Lemma 4 ([23], Lemma 2). Let pX and pX 0 be probability density functions defined on R. Assume
that pX is even
R
and logconcave, and pX
pX 0 . Let A =R[ ⌧, ⌧ ] be any symmetric
closed
interval
such
that
p
(x)dx
> 0 and let
X
A
R
h : R ! [0, 1] be any function such that R h(x)pX 0 (x)dx = A pX (x)dx. Then,
R
pX|X2A
where
pX|X2A
and
R
h · pX 0
,
h(x)p
X 0 (x)dx
R
8
p (x)
>
<R X
,
p (x)dx
=
A X
>
:
0,
if x 2 A
,
otherwise
h · pX 0
h(x) · pX 0 (x)
(x) = R
h(x)pX 0 (x)dx
h(x)pX 0 (x)dx
R
R
Combining Lemmas 3 and 4, we have the following lemma.
Lemma
5. Let pX be an even and logconcave density. Let
R
R A = [ ⌧, ⌧ ]R be any symmetric closed interval such that
p
(x)dx
>
0,
and
let
B
be
any
subset
of
R
such
that
p (x)dx = A pX (x)dx. Then,
X
A
B X
Var(X|X 2 A) Var(X|X 2 B)
Proof. One can see that pX majorizes itself. Furthermore,
we choose Rh(x) to be the Rindicator function on whether
R
x belongs to B or not, i.e., h(x) = 1{x2B} . Then, R h(x)pX (x)dx = B pX (x)dx = A pX (x)dx. By Lemma 4, the
conditional density pX|X2A majorizes the conditional density pX|X2B . Since A is symmetric about zero, pX|X2A is
even and logconcave. By Lemma 3, we conclude that Var(X|X 2 A) Var(X|X 2 B).
To prove Proposition 1, we also need to apply properties of logconcave probability density functions, which is
introduced below.
Lemma 6 ([30], Theorem 6). Let pX be a continuously di↵erentiable and logconcave probability density function
defined on (a, b). Let be a variable belonging to interval (a, b). Then, the function GX ( ), defined below, is monotone
decreasing in :
GX ( ) := E[X|X > ]
.
(3)
We will frequently refer to this function GX ( ) in the rest of the paper. We next provide an extension of Lemma 6
as follows.
Lemma 7. Let pX be an even and logconcave probability density function defined on R. Furthermore, let pX be
continuously di↵erentiable on (0, 1), and taking values in (0, 1). Then, GX ( ) as defined by (3) is monotone
decreasing in .
Proof. Let Y be a random variable such that Y = |X|. Denote by pY be the probability function of Y . Since the
probability density of X, pX is even, we have
pY (y) =
(
2pX (y),
if y > 0
0,
otherwise
11
Since pX is continuously di↵erentiable on (0, 1), so is pY . Furthermore, for any
E[Y |Y > ] = R 1
1
pY (y)dy
Z
1
2 (0, 1), we have
ypY (y)dy
Z 1
1
= R1
y · 2pX (y)dy
2pX (y)dy
Z 1
1
= R1
xpX (x)dx
pX (x)dx
= E[X|X > ].
Hence GY ( ) = GX ( ). By Lemma 6, GY ( ) is monotone decreasing in . Hence, we conclude that GX ( ) is also
monotone decreasing in when > 0.
We are now in a position to prove Proposition 1.
Proof of Proposition 1: By Lemma 2, given any communication scheduling policy f (0) satisfying Assumption 2, we
can construct a threshold-based policy f (1) such that
(1) T0f
(1)
=(
2,
f (1)
1 ),T1+
(1)
Tif )
(2) P(X 2
= P(X 2
(0)
f (1)
(3) D(Q ) D(Qf ).
=(
f (1)
1 , 1),T1
(0)
Tif ),
= ( 1,
2 ).
for all i 2 {0, 1+, 1 }.
Based on policy f (1) , we can construct a symmetric threshold-based policy f (2) such that
(1) T0f
(2)
(2)
=(
(2) P(X 2 T0f
(2)
f
, ),T1+
= ( , 1),T1f
) = P(X 2 T0f
(1)
(2)
).
Then, we only need to show that D(Qf
D(Qf
D(Qf
By Lemma 5, Var(X|X 2 T0f
(2)
(1)
(2)
(2)
) =
) =
= ( 1,
) D(Qf
P
P
(2)
(1)
). Note that D(Qf
i2{0,1+,1 }
i2{0,1+,1 }
) Var(X|X 2 T0f
Var(X|X 2 T0f
).
(1)
Var(X|X 2 Tif
(1)
Var(X|X 2 Tif
(2)
). Since P(X 2 T0f
)P(X 2 T0f
(2)
) P(X 2 Tif
(2)
(1)
) and D(Qf
) P(X 2 Tif
(1)
) P(X 2 Tif
(2)
(2)
(2)
)
)
) = P(X 2 T0f
) Var(X|X 2 T0f
(1)
) can be expressed as
)P(X 2 T0f
(1)
(1)
), we have
).
Hence, we will be done if we show that
X
i2{1+,1 }
Var(X|X 2 Tif
(2)
)
X
i2{1+,1 }
Var(X|X 2 Tif
(1)
) P(X 2 Tif
(1)
).
Consider the class of threshold-based communication scheduling policies, denoted by F, whose generic element f is
in the form of
f
T0f = ( 2 , 1 ), T1+
= ( 1 , 1), T1f = ( 1, 2 ),
0,
1, 2
and
P(X 2 T0f ) = P(X 2 T0f
12
(0)
)
It is clear that f (1) and f (2) are elements of F. Let P D(Qf ) be the sum of the mean squared distortions of Qf over
f
regions T1+
and T1f , i.e.,
P
P D(Qf ) :=
i2{1+,1 }
Var(X|X 2 Tif ) P(X 2 Tif )
= Var X|X <
P X<
2
= Var X|X >
P X>
2
2
+ Var X|X >
+ Var X|X >
2
1
1
P X>
P X>
1
,
1
where the last equality holds since pX is even. We now show that f (2) is the global minimizer of P D(Qf ) among all
f
elements in F. Since P(X 2 T0f ) is a constant, so is P(X 2 T1+
) + P(X 2 T1f ). Then,
Z
2
pX (x)dx +
1
Taking the derivatives of both sides with respect to
d
d
2
1
d
d 2
·
Z
Z
1,
1
pX (x)dx = constant
1
we have
d
d 1
Z
pX ( 1 ) = 0 )
d
d
2
pX (x)dx +
a
a
pX (x)dx = 0,
1
which implies that
d
d
2
1
pX (
2)
2
pX ( 1 )
.
pX ( 2 )
=
1
(4)
The equality above holds because pX is even. Now taking the derivative of P D(Qf ) with respect to
d
d
P D(Qf ) =
d 1
d
2
1
·
d
Var X|X >
d 2
2
P X>
2
+
d
Var X|X >
d 1
1
P X>
1
1,
.
we have
(5)
The second term in (5) can be computed as follows:
d
Var(X|X > 1 ) P(X > 1 )
d 1
✓
◆
Z 1
d
=
Var(X|X > 1 )
pX (x)dx
d 1
1
✓Z 1
◆
Z 1
d
2
=
x2 pX (x)dx
E[X|X > 1 ]
pX (x)dx
d 1
1
1
0
1
R1
2
Z 1
xpX (x)dx
d @
A
R11
=
x2 pX (x)dx
d 1
pX (x)dx
1
1
=
=
2
2
1 pX ( 1 )
pX ( 1 ) ·
⇣
1
R1
1
xpX (x)dx ·
E[X|X >
1]
⌘2
1 pX ( 1 )
.
(6)
R1
R1
2
· 1 pX (x)dx +
xpX (x)dx pX ( 1 )
1
R1
2
pX (x)dx
1
Similarly, we have the first term in (5):
d
Var X|X >
d 2
2
P X>
2
=
13
pX ( 2 ) ·
⇣
2
E[X|X >
2]
⌘2
(7)
Plugging (4), (6), and (7) into (5), we have
⇣
⌘2
⇣
⌘2
pX ( 1 )
· pX ( 2 ) · 2 E[X|X > 2 ]
pX ( 1 ) · 1 E[X|X > 1 ]
pX ( 2 )
h
i
⇣
⌘
2
2
= pX ( 1 ) 2 E[X|X > 2 ]
E[X|X > 1 ]
= pX ( 1 ) G2X ( 2 ) G2X ( 1 )
1
d
P D(Qf ) =
d 1
By Lemma 7, GX ( ) is a non-negative and monotone decreasing function. Therefore, G2X ( ) is monotone decreasing,
and
d
d
d
P D(Qf ) 0, if 1 > 2 ,
P D(Qf ) = 0, if 1 = 2 ,
P D(Qf ) 0, if 1 < 2 .
d 1
d 1
d 1
which implies that 1 = 2 (corresponding to f (2) ) is the global minimizer.
Now we are in a position to prove Theorem 2 by applying Proposition 1.
Proof of Theorem 2: Consider any communication scheduling policy f . The expected cost corresponding to f can be
computed as follows:
h
i
h
i
X
J(f ) = E cU + (X X̂)2 =
E cU + (X X̂)2 |X 2 Tif P(X 2 Tif )
i2{0,1+,1 }
When X 2 T0f , we have U = 0 and X̂ = E[X|X 2 T0f ]. Hence,
h
E cU + (X
i
h
X̂)2 |X 2 T0f = E X
⇥
f ⇤
E X|X 2 T1+
f
When X 2 T1+
, we have U = 1, Y = ↵(1) X
1
↵(1)
1
=
↵(1)
X̂ =
=
+1
+1
+1
X+
where
↵(1) =
Hence,
⇥
E cU + (X
2
X̂) |X 2
f ⇤
T1+
i
⇥
⇤ 2
E X|X 2 T0f |X 2 T0f = Var(X|X 2 T0f )
⇥
f ⇤
Ỹ + E X|X 2 T1+
⇥
f ⇤
(Y + V ) + E X|X 2 T1+
1
↵(1)
s
+1
V +
⇥
1
f ⇤
E X|X 2 T1+
+1
PT
Var(X|X 2
f
T1+
)
,
=
PT
2
V
⇣
(8)
⌘2
⇥
1
1
f ⇤
f
X E X|X 2 T1+
V |X 2 T1+
+1
↵(1) + 1
h
i
2
⇥
⇥
1
1
f ⇤ 2
f
f ⇤
= c+
E
X
E
X|X
2
T
|X
2
T
+
E V 2 |X 2 T1+
1+
1+
2
2
2
( + 1)
↵(1) ( + 1)
h
i
⇥
⇤
2
f
f
E X E X|X 2 T1+
V |X 2 T1+
↵(1) ( + 1)2
= c+E
2
1
1
f
2
Var(X|X 2 T1+
)+
2
2
( + 1)
↵(1) ( + 1)2 V
1
f
f
= c+
Var(X|X 2 T1+
)+
Var(X|X 2 T1+
)
( + 1)2
( + 1)2
1
f
= c+
Var(X|X 2 T1+
)
+1
= c+
14
The third equality holds since conditioning on the event that X > 0, V is zero-mean and independent of X
(Assumption 3). The fourth equality holds due to the definitions of ↵(1) and , which are described by (8). Similarly,
one can compute that
⇥
⇤
1
E cU + (X X̂)2 |X 2 T1f = c +
Var(X|X 2 T1f ).
+1
Hence, J(f ) can be further expressed as
J(f ) = Var(X|X 2 T0f )P(X 2 T0f ) +
1
Var(X|X 2 T1f )P(X 2 T1f ) + c P(X 2 T1f )
+1
+
1
D(Qf ) +
+1
=
1
f
f
f
Var(X|X 2 T1+
)P(X 2 T1+
) + c P(X 2 T1+
)
+1
+1
(9)
f
Var(X|X 2 T0f )P(X 2 T0f ) + c P(X 2 T1+
) + c P(X 2 T1f )
Given any communication scheduling policy f , we can construct a symmetric threshold-based communication
scheduling policy f 0 such that
0
(1) T0f = (
(2) P(X 2
P(X 2
0
0
f
, ),T1+
= ( , 1),T1f = ( 1,
0
T0f )
f0
T1+
)
= P(X
+ P(X
2 T0f ),
0
2 T1f )
).
or equivalently,
f
= P(X 2 T1+
) + P(X 2 T1f )
0
0
By Proposition 1 and Lemma 5, we have D(Qf ) D(Qf ) and Var(X|X 2 T0f ) Var(X|X 2 T0f ). Furthermore,
0
0
f0
f
we have P(X 2 T0f ) = P(X 2 T0f ) and c P(X 2 T1+
) + c P(X 2 T1f ) = c P(X 2 T1+
) + c P(X 2 T1f ). Hence,
0
we conclude that J(f ) J(f ), which implies that without loss of optimality, we can restrict the search of the
optimal communication scheduling policy to the class of symmetric threshold-based policies. Denote by J( ) the
expected cost corresponding to the symmetric-threshold based communication scheduling policy with threshold ,
where
0. By (9), J( ) can be computed as
J( ) =
Z
x2 pX (x)dx +
+
1
Var(X|X > )P(X > ) + c
+1
= 2
Z
1
Var(X|X <
+1
x2 pX (x)dx + 2
0
)P(X <
Z
1
)+c
Z
pX (x)dx
1
pX (x)dx
1
Var(X|X > )P(X > ) + 2c
+1
Z
1
pX (x)dx,
where the second equality holds since pX is even. Taking the derivative of J( ) with respect to , and by (6), we
have
⇣
⌘
d
1
2
J( ) = 2pX ( ) 2
E[X|X > ]
c
d
+1
⇣
⌘
1
= 2pX ( ) 2
G2X ( ) c
+1
Since c > 0 and GX ( ) is monotone decreasing, there exists a unique ⇤ in [0, 1) such that
⇤2
=
1
G2 (
+1 X
⇤
)+c
Furthermore, dJ( )/d < 0 when < ⇤ and dJ( )/d > 0 when > ⇤ . Hence, ⇤ is the unique global minimizer
among all > 0.
Remark 7. If the density function pX has support ( a, a) and 0 < a < ⇤ , then dJ( )/d is always negative, which
implies that the minimizing is just a. This means the optimal communication scheduling policy is to always choose
not to transmit regardless of sensor’s observation. Such a case can occur when the communication cost is very high.
15
4
Optimization Problem with hard constraints
To present our main results for the problem with hard constraints, we introduce a number of terms. First, we let
Et denote
Pt 1 the number of communication opportunities left at the beginning of the t-th time interval, i.e., Et =
N
i=1 Ui . Then, evolution of Et is described by
E1 = N
(10)
E t = Et
Ut
1
1,
t
2
Furthermore, the communication constraint is
U t Et ,
for all t = 1, 2, . . . , T
(11)
Recall that U1:t 1 is the common information shared by the sensor, the encoder, and the decoder, and hence Et is
also known by all the decision makers.
Second, we let J ⇤ (t, Et ) be the optimal cost to go if the system is initialized at time t (or equivalently, at the beginning
of the t-th time interval) with Et number of communication opportunities. Specifically, we define J ⇤ (T + 1, ·) = 0
for any number of communication opportunities.
Third, for any Et > 0, we let c(t, Et ) denote the di↵erence between two optimal costs to go, i.e.,
c(t, Et ) = J ⇤ (t + 1, Et
1)
J ⇤ (t + 1, Et ).
Remark 8. c(t, Et ) can be interpreted as the opportunity cost for choosing to communicate with the estimator rather
than not to communicate.
The following theorem ensures that without loss of optimality we can restrict all the decision makers to consider
only their current inputs and Et when making decisions at time t. Furthermore, the optimal decision policies can be
obtained via solving the dynamic programming equation.
Theorem 3. Consider the optimization problem with hard constraint as formulated in section 2-D. Without loss of
optimality, we can restrict the communication scheduling, encoding and decoding policies to the forms:
Ut = ft (Xt , Et ), Yt = gt (X̃t , Et ), X̂t = ht (Ỹt , St , Et )
Furthermore, the optimal decision policies (ft⇤ , gt⇤ , h⇤t ) together with the optimal cost to go J ⇤ (t, Et ) can be obtained
by solving the dynamic programming (DP) equation:
J ⇤ (T + 1, ·) = 0
n
o
J ⇤ (t, Et ) = min E (Xt X̂t )2 + J ⇤ (t + 1, Et+1 )
(12)
ft ,gt ,ht
The proof of Theorem 3 is similar to that of Theorem 1, and hence we keep using notations (fa:b , ga:b , ha:b ) and
(Ist , Iet , Idt ), and It .
Proof. It is easy to see validity of the following equality:
min J(f, g, h)
⇢
= min E cU1 + (X1
f,g,h
f1 ,g1 ,h1
+
fT
min
1 ,gT
1 ,hT
1
n
X̂1 )2 + min E cU2 + (X2
n
E cUT
f2 ,g2 ,h2
1
+ (XT
1
X̂T
16
1)
2
+
X̂2 )2 + . . .
min
fT ,gT ,hT
E cUT + (XT
(13)
X̂T )2
o
...
Then, at time t = T , the optimization problem, call it Problem T1, is to design (fT , gT , hT ) minimizing
n
JT (fT , gT , hT ) := E (XT
X̂T )2
o
where
UT = fT (XT , IsT ), YT = gT (X̃T , IeT ), X̂T = hT (ỸT , ST , IdT ).
Consider another problem, call it Problem T2, where IT is available to the sensor, the encoder, and the decoder, and
one needs to design (fT0 , gT0 , h0T ) minimizing
n
JT2 (fT0 , gT0 , h0T ) := E cUT + (XT
X̂T )
2
o
,
where
UT = fT0 (XT , IT ),
YT = gT0 (X̃T , IT ),
X̂T = h0T (ỸT , ST , IT ).
Since the sensor, the encoder, and the decoder can always ignore the redundant information and behave as if they
only know IsT , IeT , IdT , respectively, the performance of the system in Problem T2 is no worse than that in Problem
T1, i.e.,
JT2 (fT0 , gT0 , h0T ) min JT1 (fT , gT , hT ).
min
0
0
0
(fT ,gT ,hT )
(fT ,gT ,hT )
Similarly, consider a third problem, call it Problem T3, where IsT , IeT , IdT are not available to the sensor, the
encoder, and the decoder, respectively. Instead, all decision makers share the information about ET . One needs to
design (fT00 , gT00 , h00T ) minimizing
n
o
JT3 (fT00 , gT00 , h00T ) = E cUT + (XT X̂T )2 ,
where
UT = fT00 (XT , ET ),
YT = gT00 (X̃T , ET ),
X̂T = h00T (ỸT , ST , ET )
Since the sensor, the encoder, and the decoder can deduce ET from IsT , IeT , IdT , respectively, the system in Problem
T1 cannot perform worse than the system in Problem T3. Hence,
min
(fT ,gT ,hT )
JT1 (fT , gT , hT )
min
00 ,h00 )
(fT00 ,gT
T
JT3 (fT00 , gT00 , h00T ).
We now consider system 2. Note that the random variables (XT , VT ), the distortion function ⇢(·, ·), and the power
constraint of the encoder do not depend on IT . Furthermore, the communication constraint as described by (11)
depends on IT only via ET . Therefore, there is no loss of optimality in restricting
UT = fT0 (XT , ET ), YT = gT0 (X̃T , ET ), X̂T = h0T (ỸT , ST , ET ),
and thus
min
0 ,h0 )
(fT0 ,gT
T
JT2 (fT0 , gT0 , h0T ) =
min
00 ,h00 )
(fT00 ,gT
T
JT3 (fT00 , gT00 , h00T )
The equality above indicates that in Problem T1, the sensor, the encoder, and the decoder can ignore their information
about the past, namely IsT , IeT , and IdT , respectively, and there is no loss of optimality in restricting
UT = fT (XT , ET ), YT = gT (X̃T , ET ), X̂T = hT (ỸT , ST , ET ).
Equivalently, all the decision makers in system 1 can behave as if the system is initialized at time T with ET number
of communication opportunities. Hence,
min
fT ,gT ,hT
JT (fT , gT , hT ) = J ⇤ (T, ET ).
17
Now consider the optimization problem at stage T
1. By (13), the problem is to design (fT 1 , gT
minimizing
⇢
nP
o
2
T
JT 1 (fT 1 , gT 1 , hT 1 ) := E (XT 1 X̂T 1 )2 + min E
X̂t )
t=1 cUt + (Xt
fT ,gT ,hT
n
= E (XT
n
= E (XT
1
X̂T
2
⇤
1 ) + J T, ET
1
X̂T
1)
2
+ J ⇤ T, ET
o
UT
1
1
1,
hT
1)
(14)
o
Depending on whether ET 1 is zero or not, we have the following: when ET 1 = 0, by the communication constraint,
UT = UT 1 = 0. Then, J ⇤ (T, 0) is a constant, which is independent of (fT 1 , gT 1 , hT 1 ). Hence, minimizing
JT 1 (fT 1 , gT 1 , hT 1 ) is equivalent to minimizing E{(XT 1 X̂T 1 )2 }. When ET 1 > 0, (14) can be further
expressed as
n
o
JT 1 (fT 1 , gT 1 , hT 1 ) = J ⇤ T, ET 1 + E (XT 1 X̂T 1 )2 + c(T, ET 1 ) · UT 1
Note that J ⇤ (T, ET 1 ) is independent of (fT 1 , gT 1 , hT 1 ). Hence, minimizing JT 1 (fT 1 , gT 1 , hT 1 ) is equivalent to minimizing E{(XT 1 X̂T 1 )2 + c(T, ET 1 )UT 1 }. In both cases, the random variables (XT 1 , VT 1 ), the
distortion function ⇢(·, ·), and the power constraint of the encoder do not depend on IT 1 . Furthermore, (in both
cases) the communication constraints and (in the second case) the opportunity cost c(T 1, ET 1 ) depend on IT 1
only via ET 1 . By a similar argument as above, it is sufficient for all the decision makers to consider only ET 1
instead of Is(T 1) , Ie(T 1) , Id(T 1) . Equivalently, all the decision makers can behave as if the system is initialized at
time T 1 with ET 1 number of communication opportunities. Hence,
fT
min
1 ,gT
1 ,hT
JT
1 (fT
1 , gT
1 , hT
1)
= J ⇤ (T
1, ET
1 ).
1
By induction one can show that there is no loss of generality by restricting Ut = ft (Xt , Et ), Yt = gt (X̃t , Et ), X̂t =
ht (Ỹt , St , Et ), and the optimal cost to go satisfies the DP equation (12).
Let us take a closer look at the DP equation (12). When Et = 0, by the communication constraint Ut = ft (Xt , Et ) = 0
regardless of the realization of Xt . Consequently, we have Et+1 = 0. Then, the DP equation can be easily solved as
follows:
n
o
J ⇤ (t, 0) = min E (Xt X̂t )2 + J ⇤ (t + 1, 0) = Var(Xt ) + J ⇤ (t + 1, 0)
ft ,gt ,ht
The last equality holds since without any information about Xt , the optimal estimator is E[Xt ] and the mean squared
error is Var(Xt ). When Et > 0, the DP equation can be written as
n
o
n
o
J ⇤ (t, Et ) = min E (Xt X̂t )2 + J ⇤ (t + 1, Et+1 ) = J ⇤ (t, Et ) + min E c(t, Et )Ut + (Xt X̂t )2
(15)
ft ,gt ,ht
ft ,gt ,ht
Note that the minimization in the second line of (15) is just the one-stage problem discussed in section 3 with
communication cost c(t, Et ). This now motivates us to make the following assumptions.
Assumption 5. The sensor is restricted to apply the communication scheduling policies ft such that for any 1
t T and Et > 0,
E[Xt |ft (Xt , Et ) = 1, Xt < 0] < E[Xt |ft (Xt , Et ) = 0] < E[Xt |ft (Xt , Et ) = 1, Xt > 0]
Assumption 6. The encoder and the decoder are restricted to apply piecewise affine policies:
(
St · ↵(St ) · (Xt E [Xt |Ut = 1, St ]) , if Ut = 1
gt (X̃t , Et ) =
0,
if Ut = 0
8
1
>
< St ·
Y˜t + E [Xt |Ut = 1, St ] , if Ut = 1
↵(St ) + 1
ht (Ỹt , St , Et ) =
>
:E[X |U = 0],
if U = 0
t
t
t
18
where
= PT /
2
V
, ↵(St ) =
p
PT /Var(Xt |Ut = 1, St ), and Var(Xt |Ut = 1, St ) is the conditional variance.
Then, we have the following theorem on the optimality of symmetric threshold-based communication scheduling
strategy.
Theorem 4. Consider the problem with hard constraints under Assumptions 1, 3, 5 and 6
ft⇤ (Xt , Et )
where
⇤
t (Et )
=
(
1,
if Et > 0 and |Xt | >
⇤
t (Et )
0,
if Et = 0 or |Xt |
⇤
t (Et )
(16)
is non-negative and is the unique solution to the fixed-point equation:
2
=
1
G2 ( ) + c(t, Et ),
+1 X
⇥
⇤
:= E Xt |Xt >
GX
,
0
(17)
Remark 9. The major di↵erences between the problem considered in this paper and the problems considered in [31]
and [32] are as follows: In [31] and [32], we had restricted the sensor to apply symmetric threshold-based policies and
shown that the optimal encoding and decoding policies are piecewise affine. Furthermore, the results only hold for
specific source and noise densities (e.g., Laplace source and Gamma noise with specific parameters). In this paper,
however, we restrict the encoder and the decoder to apply piecewise affine encoding and decoding policies, and we
show that under some weak technical assumptions (Assumptions 2 and 5), the optimal communication scheduling
policy is symmetric and threshold-based. Moreover, the results hold for a large class of source densities (e.g., general
even and logconcave densities).
Remark 10. Consider the case where Et > T t, that is, the sensor is always allowed to communicate with the
estimator for the remaining time. First, we note that the opportunity cost c(t, Et ) is zero. Since GX (0) = E[X|X >
0] > 0, the solution to (17) is not zero, which means it is strictly positive. Then, even though the sensor can always
communicate with the estimator, the optimal communication policy is still the threshold based policy with threshold
⇤
t (Et ) > 0, which might seem counter-intuitive: why would the sensor not transmit its observation although it is
allowed to do so? This surprising result is due to the fact that threshold information, i.e., whether or not the state
sample belongs to a fixed, known interval, might be more informative than a noisy observation of the state at the
output of the noisy channel. Hence, it might be better not to communicate explicitly over the noisy channel but
rely on the side channel which signals where the sample lies. For example, at the extreme case of a very noisy
channel ( ! 0) the output of the communication channel, Ỹt , is e↵ectively useless, irrespective of the realization
Xt . However, depending on the threshold and the realization Xt , thresholding information could be significantly more
informative.
5
Numerical Results
In this section, we present the numerical results for the problem with hard constraint. We select the source density
to be Laplace density with parameter , i.e.,
pX (x) =
(1
x
2
e
1
2
e x,
,
if x
0
if x < 0
Then, it is easy to see that
GX ( ) = E[Xt |Xt > ]
=
1
,
for all
0.
Hence, the solution to (17) is
⇤
t (Et )
=
r
p
1 1
+ c(t, Et ) = m + c(t, Et ),
2
+1
where
m :=
1 1
.
+1 2
19
Then, the optimal communication scheduling policy is described by (16). Furthermore, the optimal encoding/decoding
policies (gt⇤ , h⇤t ) are as follows
gt (X̃t , Et ) =
(
⇤
t (Et )
↵ · |Xt |
0,
8
⇣1
>
< St ·
↵
ht (Ỹt , St , Et ) =
>
:0,
+1
1
,
if Ut = 1
if Ut = 0
⌘
Y˜t + t⇤ (Et ) + 1 , if Ut = 1
if Ut = 0
p
where = PT / V2 , and ↵ = PT / 2 . By plugging the optimal communication scheduling, encoding, and decoding
policies (ft⇤ , gt⇤ , h⇤t ) into the DP equation (12), we obtain the explicit update rule for the optimal cost to go J ⇤ (t, Et ),
as shown below
J ⇤ (t, Et ) = J ⇤ (t + 1, Et ) + 2
2
J ⇤ (t, Et ) = J ⇤ (t + 1, Et ) + 2
2
,
if Et = 0
2
⇤
t (Et )
1
2
+
e
⇤
t (Et )
(18)
, if Et > 0
We choose the parameters as follows: T = 100, = 1, and the signal-to-noise ratio (SNR) = 0.1, 1, 10. We solve
the optimal cost to go J ⇤ (t, Et ) by applying the update rule (18). We have plotted the optimal 100-stage estimation
error J ⇤ (1, N ) versus the number of communication opportunities N under di↵erent SNRs, as shown in Fig. 2.
200
SNR = 0.1
SNR = 1
SNR = 10
Optimal 100-stage estimation cost
180
160
140
120
100
80
60
40
20
0
0
10
20
30
40
50
60
70
80
90
100
Number of communication opportunities
Fig. 2. 100-stages estimation error vs. number of communication opportunities
One can see that, as to be expected, for each fixed SNR, the optimal 100-stage estimation error is non-increasing in
terms of the number of communication opportunities. To be more specific, there exists a threshold on the number
of communication opportunities (call it opportunity threshold ) such that the optimal 100-stage estimation error
decreases when the number of communication opportunities is below the threshold, and it stays constant above
the threshold. We call minimal error as the optimal 100-stage estimation error with the number of communication
opportunities above the opportunity threshold. One can also see from Fig. 2 that when the SNR increases, the
opportunity threshold increases and the minimal error decreases. The existence of opportunity threshold can be
interpreted
as follows: since we restrict the sensor to apply the threshold based policy with threshold t⇤ (Et ) =
p
p
c(t, Et ) + m
m, thep expected number of communication opportunities that will be used is upper
bounded by
p
p
m
m
T ·P(|Xt |
m) = T e
. Therefore, when the communication opportunities is greater than T e
, the optimal
expected estimation error will not decrease even though the sensor can have more communication opportunities. It
can also
be checked from Fig. 2 that the opportunity
thresholds
under di↵erent signal to noise ratios are roughly
p
p
p
1 1
m
m
Te
. Moreover, since m = +1
= T e 1/ +1 , which is an increasing function of the SNR . Hence,
2, Te
the opportunity threshold increases with the SNR.
20
Fig. 3 depicts a sample path of the number of communication opportunities left when the sensor applies the symmetric
threshold-based scheduling policies described in Theorem 4. When generating the plot, we have chosen T = 100,
= 1, = 0.1, and the number of communication opportunities N = 50. One can see that the communication
opportunities are not used up by the end of the time horizon.
Number of communication opportunities left
55
50
45
40
35
30
25
20
15
10
5
0
0
10
20
30
40
50
60
70
80
90
100
Time
Fig. 3. A sample path of the number of communication opportunities left vs. time
When the number of communication opportunities is larger than the opportunity threshold, the optimal estimation
error does not change with respect to the number of communication opportunities. Without loss of generality, we
can assume that the sensor is allowed
step, that is, N = T . Then, the opportunity cost is
p to communicate at each
1 1
c(t, Et ) = 0. Recall that t⇤ (Et ) = c(t, Et ) + m and m = +1
2 . Hence, the update rule for the cost function can
be simplified as follows:
p
⇣2
p ⌘
2 m
2
m
J ⇤ (t, T ) = J ⇤ (t + 1, T ) +
+ 2 ·e
2
with J ⇤ (T + 1, T ) = 0, which implies that
J ⇤ (1, T ) = T ·
⇣2
2
p
2 m
+
2
2
·e
p
m
⌘
=T ·2
2
· 1
p
1
1+
+1 ·e
p1
1+
.
It is straightforward to check that J ⇤ (1, T ) is a decreasing function of the SNR . Hence, the minimal error decreases
as the SNR increases.
Plotting the opportunity threshold T e m versus minimal error J ⇤ (1, T ) under di↵erent SNRs (dash line) in Fig.
2, we arrive at Figure 4. One can see that the intersection between the dash line and each solid line is roughly the
turning point of the solid line. Therefore, the plot of opportunity threshold versus minimal error under di↵erent
SNRs is an important one. In fact, the plot suggests the lowest capacity of the battery that one should choose when
building a physical system so that the expected estimation error is minimized. In addition, the plot predicts the
minimal expected estimation error.
1 1
Consider the asymptotic case where the SNR ! 1, and thus m = +1
2 ! 0. Then the opportunity threshold
m
⇤
Te
! T , and the minimal error J (1, T ) ! 0. Hence, the optimal 100-stage estimation error will be strictly
decreasing in terms of the number of communication opportunities in the asymptotic case, as also noted in the prior
work [12]. Moreover, the estimation error will reach zero when the number of communication opportunities is equal
to the time horizon.
6
Conclusions
In this paper, we have considered a communication scheduling and remote estimation problem with a noisy communication channel. Under some technical conditions, we have obtained optimal solutions for both soft-constrained and
hard-constrained problems, which consist of a symmetric threshold-based communication scheduling strategy and
21
200
SNR = 0.1
SNR = 0.2
SNR = 0.5
SNR = 1
SNR = 2
Optimal 100-stage estimation cost
180
160
140
120
100
80
60
40
20
0
10
20
30
40
50
60
70
80
90
100
Number of communication opportunities
Fig. 4. Opportunity threshold vs. minimal error under di↵erent signal to noise ratios
a pair of piecewise affine encoding/decoding strategies. Moreover, we have generated numerical results to illustrate
the e↵ect of the presence of channel noise.
There are several directions for future work: (1) Here, we assumed that the sensor is restricted to apply the communication scheduling policy satisfying Assumption 2. Under this assumption, we showed that without loss of optimality,
one can restrict the search of optimal communication scheduling policy to the symmetric class. However, it is plausible that this assumption can be relaxed or even removed, that is, for any communication scheduling policy (which
may or may not satisfy Assumption 2), there exists a symmetric policy achieving no greater costs. This possible
extension is not immediate and requires more e↵ort. (2) Here, we considered the setting with a noisy channel. It
will be interesting to consider a more general setting where there are two channels. One is noisy but not costly, and
the other one is perfect (has high communication quality) but is costly. Then, at each time step, the sensor needs to
choose whether to transmit its observation or not. If the sensor chooses to transmit its oberservation, it also needs to
choose which channel it will use. More details on this problem can be found in [33]. (3) Here, the encoding power was
taken to be time invariant. What if the encoder can distribute its total encoding power over the time horizon? More
details on problems with power allocation can be found in [34, 35]. (4) Finally, we considered here a one-dimensional
system. It would be interesting to consider extensions to multi-dimensional systems, that is, when the input and the
communication channel are chosen from multi-dimensional spaces.
References
[1] D. Shuman, A. Nayyar, A. Mahajan, Y. Goykhman, K. Li, M. Liu, D. Teneketzis, M. Moghaddam, and D. Entekhabi. Measurement
scheduling for soil moisture sensing: From physical models to optimal control. Proceedings of the IEEE, 98(11):1918–1933, 2010.
[2] J. P. Hespanha, P. Naghshtabrizi, and Y. Xu. A survey of recent results in networked control systems. Proceedings of the IEEE,
95(1):138–162, 2007.
[3] I. Akyildiz, W. Su, S. Sankarasubramaniam, and E. Cayirci. A survey on sensor networks. IEEE Communications Magazine,
40(8):102–114, 2002.
[4] M. Athans. On the determination of optimal costly measurement strategies for linear stochastic systems. Automatica, 8:397–412,
1972.
[5] V. Gupta, T. H. Chung, B.Hassibi, and R. M. Murray. On a stochastic sensor selection algorithm with applications in sensor
scheduling and sensor coverage. Automatica, 42:251–260, 2006.
[6] C. Yang and L. Shi. Deterministic sensor data scheduling under limited communication resource. IEEE Transactions on Signal
Processing, 59(10):5050–5056, 2011.
[7] Y. Mo, E. Garone, A. Casavola, and B. Sinopoli. Stochastic sensor scheduling for energy constrained estimation in multi-hop wireless
sensor networks. IEEE Transactions on Automatic Control, 56(10):2489–2495, 2011.
[8] L. Shi and H. Zhang. Scheduling two gauss-markov systems: An optimal solution for remote state estimation under bandwidth
constraint. IEEE Transactions on Signal Processing, 60(4):2038–2042, 2012.
[9] Z. Ren, P. Cheng, J. Chen, L. Shi, and Y. Sun. Optimal periodic sensor schedule for steady-state estimation under average
transmission energy constraint. IEEE Transactions on Automatic Control, 58(12):3265–3271, 2013.
[10] K. J. Åström and B. M. Bernhardsson. Comparison of riemann and lebesgue sampling for first order stochastic systems. In
Proceedings of the 41st IEEE Conference on Decision and Control, pages 2011–2016, 2002.
22
[11] K. Nar and T. Başar. Sampling multidimensional wiener processes. In Proceedings of the 53rd IEEE Conference on Decision and
Control, 2013, to appear.
[12] O. C. Imer and T. Başar. Optimal estimation with limited measurements.
Communications, 2(1-3):5–29, 2010.
International Journal of Systems Control and
[13] P. Bommannavar and T. Başar. Optimal estimation over channels with limits on usage. In 17th IFAC World Congress, pages
6632–6637, 2008.
[14] M. Rabi, G. V. Moustakides, and J. S. Baras. Multiple sampling for real-time estimation on a finite horizon. In Proceedings of the
45th IEEE Conference on Decision and Control, pages 1351–1357, 2006.
[15] Y. Xu and J. Hespanha. Optimal communication logics for networked control systems. In Proceedings of the 43rd IEEE Conference
on Decision and Control, pages 3527–3532, 2004.
[16] J. Wu, Q. Jia, K. H. Johansson, and L. Shi. Event-based sensor data scheduling: Trade-o↵ between communication rate and
estimation quality. IEEE Transactions on Automatic Control, 58(4):1041–1046, 2013.
[17] K. You and L. Xie. Kalman filtering with scheduled measurements. IEEE Transactions on Signal Processing, 61(6):1520–1530,
2013.
[18] D. Han, Y. Mo, J. Wu, S. Weerakkody, B. Sinopoli, and L. Shi. Stochastic event-triggered sensor schedule for remote state estimation.
IEEE Transactions on Automatic Control, 60(10):2661–2675, 2015.
[19] D. Shi, R. J. Elliott, and T. Chen. Event-based state estimation of discrete-state hidden markov models. Automatica, 65:12–26,
2016.
[20] S. Weerakkody, Y. Mo, B. Sinopoli, D. Han, and L. Shi. Multi-sensor scheduling for state estimation with event-based stochastic
triggers. IFAC Proceedings Volumes, 46(27):15–22, 2013.
[21] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M. I. Jordan, and S. S. Sastry. Kalman filtering with intermittent observations.
IEEE Transactions on Automatic Control, 49(9):1453–1464, 2004.
[22] K. You, M. Fu, and L. Xie. Mean square stability for Kalman filtering with Markovian packet losses. Automatica, 47(12):2647–2657,
2011.
[23] G. M. Lipsa and N. C. Martins. Remote state estimation with communication costs for first-order LTI systems. IEEE Transactions
on Automatic Control, 56(9):2013–2025, 2011.
[24] A. Nayyar, T. Başar, D. Teneketzis, and V. V. Veeravalli. Optimal stategies for communication and remote estimation with an
energy harvesting sensor. IEEE Transactions on Automatic Control, 58(9):2246–2260, 2013.
[25] B. Hajek, K. Mitzel, and S. Yang. Paging and registration in cellular networks: Jointly optimal policies and an iterative algorithm.
IEEE Transactions on Information Theory, 54(2):608–622, 2008.
[26] A. Gupta, C. Langbort, and T. Başar. Optimal control in the presence of an intelligent jammer with limited actions. In Proceedings
of the 49th IEEE Conference on Decision and Control, pages 1096–1101, 2010.
[27] A. Gupta, A. Nayyar, C. Langbort, and T. Başar. A dynamic transmitter-jammer game with asymmetric information. In Proceedings
of the 51st IEEE Conference on Decision and Control, pages 6477 – 6482, 2012.
[28] E. Akyol, K. Rose, and T. Başar. Optimal zero-delay jamming over an additive noise channel. IEEE Transactions on Information
Theory, 61(8):4331–4344, 2015.
[29] A. György and T. Linder. On the structure of optimal entropy-constrained scalar quantizers. IEEE Transactions on Information
Theory, 48(2):416–427, 2002.
[30] M. Bagnoli and T. Bergstrom. Log-concave probability and its applications. Economic theory, 26(2):445–469, 2005.
[31] X. Gao, E. Akyol, and T. Başar. Optimal sensor scheduling and remote estimation over an additive noise channel. In Proceedings
of 2015 American Control Conference, pages 2723–2728, 2015.
[32] X. Gao, E. Akyol, and T. Başar. Optimal estimation with limited measurements and noisy communication. In Proceedings of the
54th IEEE Conference on Decision and Control, pages 1775 – 1780, 2015.
[33] X. Gao, E. Akyol, and T. Başar. On remote estimation with multiple communication channels. In Proceedings of 2016 American
Control Conference, pages 5425–5430, 2016.
[34] X. Gao, E. Akyol, and T. Başar. On remote estimation with communication scheduling and power allocation. In Proceedings of the
55th IEEE Conference on Decision and Control, to appear.
[35] X. Gao, E. Akyol, and T. Başar. Joint optimization of communication scheduling and online power allocation in remote estimation.
In 2016 Asilomar Conference on Signals, Systems, and Computers, to appear.
23
| 3cs.SY
|
Ensuring convergence in total-variation-based
reconstruction for accurate microcalcification
imaging in breast X-ray CT
arXiv:1111.2616v1 [physics.med-ph] 10 Nov 2011
Jakob H. Jørgensen, Student Member, IEEE, Emil Y. Sidky, Member, IEEE, and Xiaochuan Pan, Fellow, IEEE
Abstract—Breast X-ray CT imaging is being considered in
screening as an extension to mammography. As a large fraction
of the population will be exposed to radiation, low-dose imaging
is essential. Iterative image reconstruction based on solving
an optimization problem, such as Total-Variation minimization,
shows potential for reconstruction from sparse-view data. For
iterative methods it is important to ensure convergence to
an accurate solution, since important image features, such as
presence of microcalcifications indicating breast cancer, may not
be visible in a non-converged reconstruction, and this can have
clinical significance. To prevent excessively long computational
times, which is a practical concern for the large image arrays in
CT, it is desirable to keep the number of iterations low, while
still ensuring a sufficiently accurate reconstruction for the specific
imaging task. This motivates the study of accurate convergence
criteria for iterative image reconstruction. In simulation studies
with a realistic breast phantom with microcalcifications we
compare different convergence criteria for reliable reconstruction. Our results show that it can be challenging to ensure
a sufficiently accurate microcalcification reconstruction, when
using standard convergence criteria. In particular, the gray level
of the small microcalcifications may not have converged long after
the background tissue is reconstructed uniformly. We propose
the use of the individual objective function gradient components
to better monitor possible regions of non-converged variables.
For microcalcifications we find empirically a large correlation
between nonzero gradient components and non-converged variables, which occur precisely within the microcalcifications. This
supports our claim that gradient components can be used to
ensure convergence to a sufficiently accurate reconstruction.
a particularly challenging problem as the total exposure is
restricted to the equivalence of two digital mammograms. Such
a low X-ray dose can be achieved either by drastically reducing
the intensity compared to a diagnostic-quality CT scan, or by
reconstruction from sparse-view data.
Total-Variation (TV)-regularized image reconstruction exploits approximate sparsity of the spatial gradient of cross
sections of the human body to compensate for reduction
in data. TV-reconstructions have been shown to compare
favorably with standard Filtered Back Projection from sparseview data [2], [3]. We are investigating the optimal trade-off
between low intensity views and sparse-view data for breast
CT by means of TV-reconstruction [4].
The TV-reconstruction is obtained by solving a nonlinear
optimization problem. A practical concern is that the extremely
large systems in CT, where image arrays of 109 voxels are
standard, are challenging to solve accurately in acceptable
time. Complicating this issue is the fact that clinically relevant
features are often very small—occupying only a few voxels.
As result both global and pointwise convergence of an iterative
reconstruction algorithm may have clinical impact. We demonstrate this issue in the present preliminary investigation, where
we examine a realistic simulation of CT for breast cancer
screening, and compare strategies for ensuring convergence
to a sufficiently accurate TV-reconstruction.
Index Terms—X-ray CT, breast CT, algorithm convergence,
total variation, compressed sensing
II. I MAGE RECONSTRUCTION BY TV- MINIMIZATION
I. I NTRODUCTION
D
OSE reduction has gained considerable interest in diagnostic computed tomography (CT) in recent years [1].
The potential to employ CT for screening, where a large
population fraction will be exposed to radiation dose and
the majority of subjects will be asymptomatic, also motivates
the interest in low intensity X-ray CT. Breast CT poses
This work is part of the project CSI: Computational Science in Imaging,
supported by grant 274-07-0065 from the Danish Research Council for
Technology and Production Sciences. This work was supported in part by
NIH R01 grants CA120540, EB000225. The contents of this article are solely
the responsibility of the authors and do not necessarily represent the official
views of the National Institutes of Health.
J. H. Jørgensen is with the Department of Informatics and Mathematical
Modelling, Technical University of Denmark, Richard Petersens Plads, Bygning 321, 2800 Kongens Lyngby, Denmark (e-mail: [email protected]).
E. Y. Sidky and X. Pan are with the Department of Radiology, University
of Chicago, 5841 S. Maryland Ave., Chicago IL, 60637, USA (e-mail:
{sidky,xpan}@uchicago.edu).
We consider TV-regularized image reconstruction in order
to exploit gradient sparsity to compensate for the few-view
projection data. The present study works with the discrete-todiscrete imaging model, Au = b, see [5]. For reconstruction
we consider the minimization problem
uTV = argmin f (u),
(1)
u
where
f (u) = kAu − bk1 + λkukTV
(2)
and
kukTV =
X
kDj uk2
(3)
j
and Dj is a forward difference approximation to the image
gradient at pixel j.
Instead of the more commonly used `2 norm for measuring
data fidelity we use the `1 norm. TV-regularized `2 norm
Fig. 1. Left: Original full breast phantom, 20482 pixels. Right: 1202 pixel
ROI around simulated microcalcifications. Gray level window: [0.9, 1.2].
minimization is known to be contrast-reducing, in particular
for objects of small scale [6], such as microcalcifications.
`1 minimization does not remove this problem, but tends to
reduce it [7].
Both terms in (2) are non-differentiable, and in order
to apply standard gradient-based optimization algorithms we
apply the standard smoothing trick of the replacements:
Xq
kDj uk22 +
replaces
kukTV .
(4)
X pi
|(Au)i − bi |2 +
replaces
kAu − bk1 .
(5)
i
In our simulations we use = 10−4 , which we found
sufficiently small to prevent any change in visual appearance
of the reconstructed image compared to using = 0.
An important question is how well a TV reconstrution
is capable of reproducing the salient image features, such
as microcalcifications in the present case. Numerous studies
demonstrate that of TV-reconstruction can produce clinically
useful reconstructions, see e.g. [2], [3].
Our main question of interest in the present work arises
when using an iterative algorithm to solve the TV minimization problem: When can we reliably stop iterating and accept
the computed solution as a good approximation of the true
minimizer to (2)? In other words, what is a good termination
criterion?
Note that, in general, uTV is biased compared to the original
underlying image, and the size of this bias is parameter
dependent, in particular the bias depends on λ. It is not our
goal to select a well-suited λ here; we only consider the
question of, given a choice of λ, how do the iterates approach
the solution, i.e., the minimizer of (2)? We will consider
two different choices for termination criterion for the iterative
algorihtm used for solving (1):
1) k∇f (u)k2 < τ
2) 1 + cos α < τ ,
where α is the angle between the gradients of each of the
two terms in (2), see the original reference [3] for details, and
τ is a user-specified tolerance, where a smaller τ leads to a
more accurate solution. Both criteria correspond to theoretical
optimality conditions [8] in the limit of τ = 0.
For solving (1) we use a convergent, gradient-based optimization algorithm, which is optimal in a certain sense, see
Fig. 2. Reconstructions of full image and ROIs. Top: λ = 2 · 10−2 , middle:
λ = 2 · 10−3 , bottom: λ = 2 · 10−4 . Gray level window: [0.9, 1.2].
[9]. The algorithm was developed for TV-regularized `2 data
fidelity, but is applicable to any smooth objective function, and
we have found that it works well for solving (the smoothed
version of) the problem in (1).
III. B REAST CT MODEL
Breast CT imaging is being considered as a potential
addition to mammography in screening for breast cancer.
One particular indicator of breast cancer is formation of
microcalcifications—very small, highly attenuating calcium
deposits. For screening, low-dose imaging is pertinent to
minimize accumulated X-ray dose, while accurate and reliable
microcalcification shape and attenuation reconstruction may be
important for detecting malignancy.
In the present work we use the breast phantom from [10]
discretized on a 20482 pixel grid, as shown in Fig. 1 along
with a 1202 pixel region of interest (ROI) around a simulated
cluster of microcalcifications, also discretized. Gray values
in units of water attenuation are given in Table I. Note the
fairly complex phantom structure, which makes the phantom
semi-realistic, and at the same time poses a challenge for TVbased reconstruction, which tends to favor piecewise constant,
“cartoon-like” images.
1.6
209
1.5
408
674
1.4
917
1.3
1163
1.2
1461
1.1
1.00
20
40
60
80
100
1010
10
10-1-2
10-3
10-4
10-5
10-6
10-7
||∇f(u)||2
10-8
1 + cosα
10-9
10 200 400 600 800 1000 1200 1400 1600
1032
(f−f )/f
101
100
10
10-1-2
10-3
10-4
10-5
10-6
10 200 400 600 800 1000 1200 1400 1600
1.6
1.5
1.4
1.3
1.2
1.1
1.00
1010
10
10-1-2
10-3
10-4
10-5
10-6
10-7
10-8
10-9
10 0
1032
101
100
10
10-1-2
10-3
10-4
10-5
10-6
10 0
1.6
1.5
1.4
1.3
1.2
1.1
20
40
60
80
100 1.00
1010
10
10-1-2
10-3
10-4
10-5
10-6
10-7
||∇f(u)||2
10-8
1 + cosα
10-9
200 400 600 800 1000 1200 10 0
1032
(f−f )/f
101
100
10
10-1-2
10-3
10-4
10-5
10-6
200 400 600 800 1000 1200 10 0
16
72
208
497
765
1159
13
97
311
907
1641
3064
20
40
60
80
100
||∇f(u)||2
1 + cosα
500 1000 1500 2000 2500 3000 3500
(f−f )/f
500 1000 1500 2000 2500 3000 3500
Fig. 3. Top row: Vertical profile through microcalcifications for iteration number corresponding to terminating iterations at τ = 101 , 100 , . . . , 10−4 using
criterion 1. Middle row: Values of the two convergence criteria vs. number of iterations. Bottom row: Convergence in objective function relative to the reference
solution vs. number of iterations. Left column: λ = 2 · 10−2 , center column: λ = 2 · 10−3 , right column: λ = 2 · 10−4 .
Tissue
Fat
Fibro-glandular tissue
Skin
Microcalcifications
Value
1.00
1.10
1.15
1.80 − 2.10
TABLE I
G RAY VALUES FOR BREAST PHANTOM , IN UNITS OF WATER ATTENUATION .
IV. N UMERICAL RESULTS
Different choices for the regularization parameter λ lead to
very different solutions, and the question of how to choose a
well-suited λ is important. However, our goal here is merely
to demonstrate that very different convergence is observed for
different choices of λ; not to propose a certain λ over others.
For that purpose we make three choices: α = 2 · 10−2 , α =
2 · 10−3 , and α = 2 · 10−4 . We generate noise-free 64-view,
1024-detector-bin fan-beam data by forward projection (using
a line intersection-based ray-driven projector) of the original
discrete 20482 pixelized phantom with microcalcifications. We
solve (1) with termination criterion 1 for τ = 10−4 to obtain
accurate solutions. The obtained reconstructions are shown in
Fig. 2.
As expected, with increasing λ the reconstructed images becomes smoother, and the microcalcifications gradually become
invisible. Only at λ = 2·10−4 is the smallest microcalcification
visible, so it is clear that we need to use a λ smaller than or
equal to 2 · 10−4 .
We rerun the three reconstructions and store this time
iterates along the way, at thresholds τ = 101 , 100 , . . . , 10−4
for termination criterion 1. We use the most accurate iterate, at
each λ as a reference solution for comparing the convergence
of the earlier iterates. We denote the reference solution by u?
and its value for the objective function by f ? .
In the top row in Fig. 3 we show reconstruction profiles
through two of the microcalcifications (the two that are on the
same vertical line) for each of the stored iterates, including
the reference solutions. For the largest λ we see that the
iterates converge to the reference solution very quickly: after
438 iterations i.e. at τ = 100 the solution is indistinguishable
from the reference solution. For the middle λ we see a
different behavior to which we will refer as non-uniform
convergence: For the most part, the iterates converge to the
reference solution rapidly, but precisely within the larger of
the two microcalcifications, the iterates converge very slowly.
For the smallest λ the non-uniform convergence is even more
pronounced and only at the reconstruction stored before the
reference solution we see no further improvement of the
iterates. It seems natural that for even smaller values of λ
we would see even more severe non-uniform convergence.
Our concern about non-uniform convergence arises from
two facts: First, detecting non-uniform convergence can be
very challenging as we will demonstrate. Second, if we are
not aware of non-uniform convergence, we risk accepting
a solution which is not yet converged everywhere. Such a
reconstruction has much lower contrast than the true TVsolution, which will make it difficult to spot the microcalcifications. This can lead us to the, incorrect, conclusion that the
TV-solution is not capable of reproducing microcalcifications
faithfully, but in fact the lack of contrast in the reconstruction
was a result of accepting a too early iterate returned by
the iterative solver and not because of the TV-minimization
problem itself.
Difference images to pseudo-solution
τ
λ = 2 · 10−2
λ = 2 · 10−3
λ = 2 · 10−4
Gradient components
λ = 2 · 10−2
λ = 2 · 10−3
λ = 2 · 10−4
10−1
10−2
10−3
Fig. 4. Left: Difference images u − u? . Gray level windows: Top: [−1 · 10−1 , 1 · 10−1 ]. Middle: [−2 · 10−2 , 2 · 10−2 ]. [−2 · 10−3 , 2 · 10−3 ]. Right:
Gradient components ∇f (u). Gray level windows: Top: [−1 · 10−4 , 1 · 10−4 ]. Middle: [−1 · 10−5 , 1 · 10−5 ]. [−1 · 10−6 , 1 · 10−6 ].
To illustrate that detecting non-uniform convergence is
challenging, we compare the use of termination criteria 1
and 2. In the middle row of plots in Fig. 3 we have plotted
the values of the two criteria vs. the number of iterations
used for the considered iterates . Furthermore, in the bottom
row of plots we show the convergence in terms of objective
function value relative to the reference solution vs. the number
of iterations. The difference in function value relative to the
reference solution acts as an indicator of the accuracy of
the iterate. Now imagine that we use τ = 10−3 as our
convergence criterion. With criterion 1 and the largest λ we
find an (f − f ? )/f ? of approx. 10−6 , which we consider to
be very accurate. However, for the two smaller values λ we
find (f − f ? )/f ? of approx. 10−4 and 10−3 , indicating much
less accurate reconstruction. A similar trend can be observed
for termination criterion 2. The different accuracies obtained
confirm our observations from the profile plots in the top row.
We note that we are able to detect non-uniform by comparing
the final function value differences in the plots in the bottom
row. However, in practice, we do not have access to the true
solution or a reference solution, as we would like keep the
number of iterations low. We do have access to the values
of the termination criterion functions, but as can be seen by
inspecting the middle row of plots, we cannot trust that a using
a fixed τ will provide a uniformly converged reconstruction.
V. G RADIENT COMPONENTS
As a first step towards a more reliable convergence criterion
we wish to point out a connection that can possibly exploited.
The two considered convergence criteria both involve the
gradient of the objective function f . However, as we saw,
they do not clearly show that a few pixels have not yet
reached convergence. We believe this is due to computing a
single number from the full gradient for comparing with a τ ,
thereby “averaging out” the differences between the individual
components of the gradient. Many small gradient components
will tend to hide the presence of a few larger ones. We propose
instead to monitor the full objective function gradient ∇f (u)
during the iterations.
In the right half of Fig. 4 we display as an image the
ROI gradient components of the objective function, for the
iterates obtained with τ = 10−1 , 10−2 , 10−3 for each of
the three choices of λ. In the left half of Fig. 4 we show
the corresponding ROI difference images u − u? between
the iterates and the reference solution image. Note that with
decreasing τ we are making the gray level windows narrower
to emphasize small components.
For the largest λ both difference images and gradient components appear to be zero (in the chosen gray level window)
for all three choices of τ . This agress well with our observatoin
that the iterates converged rapidly to the reference solution, so
we should precisely expect very small gradient components
everywhere here.
For the smaller choices of λ we observe a highly nonuniform nonzero gradient component pattern for both the
difference images and the gradient components, with large
(negative) components exactly at the microcalcifications. The
gradient components are negative, which agress with the variables still growing as seen in the profiles in the top row of Fig.
3. For the more accurate reconstructions the microcalcification
pixel values in the difference image and gradient components
remain distinct while their magnitude approach zero.
There is a clear correlation between the difference images
and the gradient components, indicating a close connection.
This suggests the possibility for ensuring local convergence in
the microcalcifications by means of monitoring the gradient
components.
At the most accurate solution, while the microcalcification
pixels are still visible in difference images and gradient
components, the intensity is of the level of the background.
This leads us to the conclusion that at this point the iterate has
converged, and we can relibly accept it as an accurate solution.
VI. D ISCUSSION
Note that we are only able to monitor the convergence
using the difference images, since we computed the reference
solution. In practice, we wish to monitor convergence at
any given iteration without a much more accurate reference
solution. The gradient components are readily available during
the iterations, and as our simulation shows, they can be used
to monitor non-converged pixels.
We are investigating strategies other than visual inspection
of the gradient components for a quantitative convergence
criterion. For instance by forcing maxj |(∇f (u))j | below a
appropriately chosen threshold , all gradient components will
be smaller than , thereby ensuring global convergence. When
applying a single number based convergence criterion such
criteria 1 and 2, the fact that the majority of the variables are
at optimum can conceal by averaging out the contributions
from the few variables that are not. The rationale in forcing
all gradient components below is that small areas of nonconvergent varibles will prevent termination of the algorithm.
A different approach would be to exploit the spatial structure
in the nonzero gradient components, e.g. by not terminating
iterations until no spatial correlation is present.
VII. C ONCLUSION
We have conducted a preliminary comparative investigation
of convergence criteria for ensuring accurate reconstruction
of microcalcifications in breast CT. We have demonstrated
that the nonzero gradient components can be used to monitor
the regions of non-converged variables and thereby preventing termination of the optimization algorithm before global
convergence is reached.
Accepting a reconstruction which is not globally converged
may have clinical significance, for instance, as in the example
given, by providing insufficient contrast for detecting the
microcalcifications.
The use of the objective gradient in a convergence criterion
is well-known, at least the use of the norm of the gradient. Explicit use of the individual gradient components for monitoring
local convergence for small objects such as microcalcifications
has not, to the best of our knowledge, been studied before. An
interesting direction for future work is to apply the approach
to other optimization based reconstruction techniques.
R EFERENCES
[1] C. H. McCollough, A. N. Primak, N. Braun, J. Kofler, L. Yu, and
J. Christner, “Strategies for reducing radiation dose in ct,” RADIOLOGIC
CLINICS OF NORTH AMERICA, vol. 47, pp. 27–40, 2009.
[2] E. Y. Sidky, C.-M. Kao, and X. Pan, “Accurate image reconstruction
from few-views and limited-angle data in divergent-beam CT,” JOURNAL OF X-RAY SCIENCE AND TECHNOLOGY, vol. 14, no. 2, pp.
119–139, 2006.
[3] E. Y. Sidky and X. Pan, “Image reconstruction in circular cone-beam
computed tomography by constrained, total-variation minimization,”
PHYSICS IN MEDICINE AND BIOLOGY, vol. 53, no. 17, pp. 4777–
4807, 2008.
[4] J. H. Jørgensen, P. C. Hansen, E. Y. Sidky, I. S. Reiser, and X. Pan,
“Toward optimal X-ray flux utilization in breast CT,” in Proceedings
of the 11th International Meeting on Fully Three-Dimensional Image
Reconstruction in Radiology and Nuclear Medicine, Potsdam, Germany,
2011, available from http://arxiv.org/abs/1104.1588.
[5] H. H. Barrett and K. J. Myers, Foundations of Image Science. Hoboken,
NJ: John Wiley & Sons, 2004.
[6] D. Strong and T. Chan, “Edge-preserving and scale-dependent properties
of total variation regularization,” INVERSE PROBLEMS, vol. 19, no. 6,
pp. S165–S187, 2003.
[7] T. F. Chan and S. Esedoglu, “Aspects of total variation regularized L-1
function approximation,” SIAM JOURNAL ON APPLIED MATHEMATICS, vol. 65, no. 5, pp. 1817–1837, 2005.
[8] J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed. New
York: Springer, 2006.
[9] T. L. Jensen, J. H. Jørgensen, P. C. Hansen, and S. H. Jensen, “Implementation of an optimal first-order method for strongly convex total
variation regularization,” BIT Numerical Mathematics, to appear, 2011,
preprint available from: http://arxiv.org/abs/1105.3723.
[10] I. Reiser and R. M. Nishikawa, “Task-based assessment of breast
tomosynthesis: Effect of acquisition parameters and quantum noise,”
Med. Phys., vol. 37, pp. 1591–1600, 2010.
| 5cs.CE
|
ALGEBRAIC GEOMETRY OVER HYPERRINGS
arXiv:1512.04837v1 [math.AG] 15 Dec 2015
JAIUNG JUN
Abstract. We develop basic notions and methods of algebraic geometry over
the algebraic objects called hyperrings. Roughly speaking, hyperrings generalize rings in such a way that an addition is ‘multi-valued’. This paper largely
consisits of two parts; algebraic aspects and geometric aspects of hyperrings.
We first investigate several technical algebraic properties of a hyperring. In
the second part, we begin by giving another interpretation of a tropical variety as an algebraic set over the hyperfield which canonically arises from a
totally ordered semifield. Then we define a notion of an integral hyperring
scheme (X, OX ) and prove that Γ(X, OX ) ≃ R for any integral affine hyperring scheme X = Spec R.
Contents
1. Introduction
2. Review: Hyperrings
3. Algebraic aspects of Hyperrings
3.1. Quotients of hyperrings
3.2. Congruence relations
3.3. Analogues of classical lemmas
4. Geometric aspects of Hyperrings
4.1. A tropical variety as an algebraic set over a hyperfield
4.2. Construction of hyperring schemes
4.3. From hyperring schemes to the classical schemes
Appendix A. From semistructures to hyperstructures
A.1. Basic definitions of semiring theory
A.2. The symmetrization process
References
1.
1
4
7
8
10
13
16
16
23
31
33
33
34
36
Introduction
An idea of algebraic objects with a multi-valued operation has been first considered by
F.Marty in 1934 (see [33]). Marty introduced the notion of hypergroups which generalizes groups by stating that an addition is multi-valued. Later, in 1956, M.Krasner introduced
the concept of hyperrings in [26] to use them as a technical tool on the approximation of
valued fields. Since its first appearance, very few algebro-geometric perspective of hyperrings
has been studied (cf. [12], [36]). However, lately there has been considerable attention drawn
to the theory of hyperrings under various motivations.
In [6], A.Connes and C.Consani noticed that Connes’ adèle class space of a global field has
2010 Mathematics Subject Classification. 14A99(primary), 14T99(secondary).
Key words and phrases. hyperring scheme, hyperring, hypergroup, tropical variety, F1 -geometry.
1
a hyperring structure and initiated the study of algebraic geometry based on hyperrings.
Connes and Consani also provided several evidences which show that hyperrings are algebraic structures which naturally appear in relation to number theory. For example, in their
archimedean analogue of the p-typical Witt construction, the role of hyperstructures is crucial
(see [8]). Furthermore, by viewing an underlying space of an affine algebraic group scheme
X = Spec A as a set of ‘K-rational points’ for the Krasner’s hyperfield K (Example 2.5),
Connes and Consani showed that X is a hypergroup when X = Gm and Ga (cf. [4]). This
result has been partially generalized by the author in [24]. The author also used a notion of
hyperrings to generalize the classical notion of valuations in [25].
Remark 1.1. Hyperrings were first considered by Connes and Consani as a possible replacement for monoids on the developments of F1 -geometry. One of main goals in F1 -geometry is
to enlarge the category of schemes to realize C := Spec Z as ‘a curve over F1 ’ in such a way
that a (suitably defined) Hasse-Weil zeta function of C over F1 becomes the complete Riemann
zeta function. A first natural candidate was the category of monoids (to enlarge the category
of schemes). However, a generalized scheme theory over monoids is essentially toric (cf. [14])
whereas one expects that Spec Z has an infinite genus (cf. [5], [31]). This motivated Connes
and Consani to search possible other candidates rather than monoids and the study of algebraic
geometry over more general algebraic objects by many others (cf. [3], [13], [28], [35], [37]).
In particular, B.Toën and M.Vaquié first introduced algebraic geometry over a closed monoidal
category by generalizing the functorial definition of schemes (see, [40]). For example, the classical scheme theory and monoid schemes as in A.Deitmar [13] agree with this picture. We
further note that, in [16], J.Giansiracusa and N.Giansiracusa showed that a scheme theory over semirings can be beautifully linked to tropical geometry by means of the theory of
monoid schemes. However, since the category of hyperrings is not a monoidal category, our
construction of hyperring schemes does not fit with the above generalized scheme theory.
Besides from the aforementioned Connes and Consani’s approaches, M.Marshall and P.Gladki
used multirings (which are more general objects than hyperrings) in relation to quadratic
forms, Artin-Schreier theory, and real algebraic geometry (cf. [17], [32]). J.Tolliver Studied
a relation between Krasner’s hyperfield approach and P.Deligne’s approach [15] to limits of
local fields. Also, O.Viro implemented a notion of hyperfield to search for a firm algebraic
foundation of tropical geometry (cf. [41], [42]). Note that supertropical algebras by Z.Izhakian
and L.Rowen (cf. [21], [22]) and blueprints by O.Lorscheid (cf. [28], [29]) provide natural algebraic framework for tropical geometry and have certain connections to hyperrings. Lastly,
we note that Connes and Consani proved that some class of hyperrings can be realized as an
s-algebra (an algebra over the sphere spectrum) in their recent paper [9].
The following is the organization of the paper:
In §2, we provide basic definitions and properties of hyperrings which will be used in sequel.
In §3, we investigate algebraic properties of hyperrings. We first start by resolving an issue
on the construction of quotients of hyperrings. A quotient of a hyperring by a (hyper) ideal
has been known only for a special class of (hyper) ideals called normal (hyper) ideals. We
first prove that in fact such construction is valid for any (hyper) ideal as follows:
Theorem A. (Proposition 3.5) Let R be a hyperring and I be a (hyper) ideal of R. Then
the set R/I of cosets has a canonical hyperring structure which generalizes the classical case.
The classical one-to-one correspondence between ideals and congruence relations is no longer
valid in the category of commutative semirings (Example 3.8). The difficulty stems from the
weakness of commutative semirings; lacking additive inverses. One may fix a such problem
at a price; by allowing an addition to be ‘multi-valued’, one obtains additive inverses. The
following theorem states that such a one-to-one correspondence is still valid for hyperrings.
2
Theorem B. (§3.2) Let R be a hyperring. Then there is a one-to-one correspondence between
the set of (hyper) ideals of R and the set of congruence relations on R.
This provides some evidence that hyperrings may be better suited to build an algebraic
foundation of tropical geometry.
In §3.3, we prove several propositions which are analogues to the classical propositions. These
results will be used in §4.2 to construct an integral hyperring scheme. In particular, we prove
the following version of Hilbert’s Nullstellensatz of hyperrings.
Theorem C. (Lemma 3.19) Let R be a hyperring and I be a hyperideal of R. Then the
following set
√
I := {r ∈ R | ∃n ∈ N such that r n ∈ I}
is the intersection of all prime hyperideals of R which contain I.
In §4, we study geometric aspects of hyperrings. In the first subsection, we reformulate a
tropical variety as a ‘positive part’ of an algebraic set over a hyperfield. This work has been
motivated by Viro’s paper [41]. Viro manifested that an algebraic foundation of tropical
geometry should be built on the framework of hyperfields and used some hyperfields to
realize his goal. For the Viro’s hyperfield approach over a semiring approach, see [1, §6]. The
novelty of our approach lies in the use of symmetrization process which is recently introduced
by S.Henry (cf. [20]). Briefly speaking, a symmetrization process ‘glues’ two semigroups to
obtain a hypergroup. In [23], we showed that such process can be generalized to a hyperring
and a semiring in some cases (also see Appendix A.2). In particular, one always derives a
hyperfield by patching two copies of a totally ordered idempotent semifield. By appealing
to these results, we may expect to benefit from advantages of both semirings and hyperrings
via a symmetrization. More precisely, after introducing a notion of an algebraic set over
hyperrings, we prove the following:
Theorem D. (Proposition 4.18) Let R := (Rmax )S be the hyperfield symmetrizing the tropical semifield Rmax = (R ∪ {−∞}, +, max} and sn be the coordinate-wise symmetrization map.
Let X be a tropical variety over Rmax in (Rmax )n . Then, there exist a (suitably defined) algebraic set XS over R and the following set bijection:
ϕ : X ≃ (Img(sn ) ∩ XS ).
In the second subsection, we take a scheme-theoretic point of view and construct an integral
hyperring scheme generalizing the classical notion of an integral scheme. One of the most
important results in classical scheme theory is that any commutative ring A is isomorphic to
the ring of global regular functions on the topological space X = Spec A. When one enlarges
the category of commutative rings to the category of commutative monoids or commutative
semirings, one still obtains the similar result. In other words, a commutative monoid (resp.
a commutative semiring) can be realized as the commutative monoid (resp. the commutative
semiring) of global regular functions on some topological space (see the references in Remark
1.1). Also, when A is a C∗ -algebra (not necessarily commutative), there is an analogue
known as Gelfand-Naimark-Segal construction. Therefore one is induced to ask the following
question.
Question. Can a hyperring R be realized as the hyperring of global regular functions on the
topological space X = Spec R?
We prove that the answer is affirmative when R is a hyperring without any (multiplicative)
zero-divisors as follows:
Theorem E. (Theorem 4.23, Proposition 4.31) Let R be a hyperring with no (multiplicative)
zero-divisor. Then, for the hyperring scheme (X = Spec R, OX ), we have Γ(X, OX ) ≃ R.
3
Furthermore, for each p ∈ X, the stalk OX,p of OX at p exists and is isomorphic to a
(suitably defined) localized structure Rp . In particular, the opposite category of hyperrings
without zero-divisors is equivalent to the category of integral affine hyperring schemes.
Remark. Note that a similar construction has been done in [36]. However, our goal is to
retrieve the classical important result: Γ(Spec A, OSpec A ) ≃ A for hyperrings whereas the
authors of [36] did not consider it.
We also construct a possible replacement FX of the structure sheaf OX and obtain the
following theorem.
Theorem F. (Remark 4.27, Propositions 4.28, 4.29) Let R be a hyperring and X = Spec R
be a topological space of prime spectra. We associate a presheaf FX of hyperrings on X which
satisfies the following:
(1) If R is a hyperring with no zero-divisors then FX is a sheaf of hyperrings isomorphic
to OX .
(2) If X is irreducible, then FX is a sheaf of hyperrings and the canonical map R −→
FX (X) is strict and injective homomorphism of hyperrings.
(3) If X is irreducible and R is local, then R ≃ FX (X).
The remaining part of the paper is devoted to explain how the theory of hyperring schemes
can be related with the classical scheme theory. It follows from the result (Proposition 2.14)
of Connes and Consani, one can naturally associate a hyperring to a commutative ring. By
using this, we prove the following:
Theorem G. (Proposition 4.35) Let A be an integral domain containing the field Q of
rational numbers. Let R := A/Q× , X = (Spec A, OX ), Y = (Spec R, OY ), and π : A −→
A/Q× be the canonical projection map. Then, as topological spaces, Y is homeomorphic to
X. Furthermore, for an open subset U ⊆ X, the hyperring of sections OY (U ) is obtained
from the ring of sections OX (U ) by the quotient construction of Connes and Consani as in
Proposition 2.14.
Acknowledgment
This is a part of the author’s Ph.D. thesis [23]. The author thanks his academic advisor
Caterina Consani. This paper has not been written without her support. The author also
thanks Jeffrey Tolliver for proof-reading the early version of the paper and providing helpful
comments.
2.
Review: Hyperrings
In this section, we quickly review the basic definitions and properties of hyperrings which will
be used in this paper. For more details regarding theory of hyperrings, we refer the readers
to [10].
In what follows, by a hyperoperation on a nonempty set H, we always mean a function
+ : H × H → P(H)∗ ,
where P(H)∗ is the set of nonempty subsets of H. Note that the reason why we use +
notation is that we will only consider the commutative case throughout this paper. We also
define, for any nonempty subsets ∀A, B ⊆ H, the following notation:
[
A + B :=
(a + b).
a∈A,b∈B
Therefore the notation (x + y) + z makes sense as the set A + {z} with A := x + y. Lastly, for
x, y ∈ H, when x + y contains a single element z we let x + y = z for the notational simplicity.
With these notations, for a subset A ⊆ H, and a ∈ H, by a + A we mean that {a} + A.
4
Definition 2.1. A canonical hypergroup (H, +) is a nonempty set equipped with a hyperoperation + with the following properties:
(1) + is commutative; x + y = y + x ∀x, y ∈ H.
(2) + is associative; (x + y) + z = x + (y + z) ∀x, y, z ∈ H.
(3) There is the identity element; ∃! 0H ∈ H such that 0H + x = x = x + 0H ∀x ∈ H.
(4) There is a unique inverse; ∀x ∈ H ∃!y ∈ H such that 0 ∈ x + y. We denote
y = −x.
(5) Reversibility; x ∈ y + z ⇐⇒ z ∈ x − y ∀x, y, z ∈ H.
We note that hypergroups are more general objects than canonical hypergroups. For example,
one does not assume the uniqueness of the identity or an inverse. However, in this paper we
only consider canonical hypergroups. From now on, we will simply say hypergroups rather
than canonical hypergroups for the sake of convenience.
Definition 2.2. A Krasner hyperring (R, +, ·, 0, 1) is a nonempty set R such that (R, +, 0)
is a hypergroup and (R, ·, 1) is a commutative monoid which satisfy the following conditions:
(1) + and · are compatible; ∀x, y, z ∈ R, x(y + z) = xy + xz, (x + y)z = xz + yz.
(2) 0 is an absorbing element; ∀x ∈ R, x · 0 = 0 = 0 · x.
(3) 0 6= 1.
When (R \ {0}, ·) is an abelian group, (R, +, ·, 0, 1) is said to be a hyperfield.
Throughout the paper, all hyperrings are assumed to be Krasner hyperrings unless otherwise
stated.
Definition 2.3. Suppose that (R1 , +1 , ·1 ), (R2 , +2 , ·2 ) are hyperrings. A map ϕ : R1 −→ R2
is said to be a homomorphism of hyperrings if
(1) ϕ is a homomorphism of hypergroups; ϕ(a +1 b) ⊆ ϕ(a) +2 ϕ(b) ∀a, b ∈ R1 .
(2) ϕ is a homomorphism of monoids; ϕ(a ·1 b) = ϕ(a) ·2 ϕ(b) ∀a, b ∈ R1 .
(3) ϕ is said to be strict if ϕ(a +1 b) = ϕ(a) +2 ϕ(b) ∀a, b ∈ R1 .
Definition 2.4. Let R be a hyperring. By a hyperring extension of R we mean a hyperring
L such that there is an injective homomorphism i : R −→ L of hyperrings. A sub-hyperring
H of R is a subset of R such that H itself is a hyperring with the induced addition and
multiplication.
Example 2.5. (cf. [6]) Let K := {0, 1} with the hyperoperation and the multiplication given
in the following tables:
+ 0
1
· 0 1
0 0
1
0 0 0
1 1 {0, 1}
1 0 1
One can easily check that K is a hyperfield. K is called the Krasner’s hyperfield.
Example 2.6. (cf. [6]) Let S = {−1, 0, 1}. A hyperoperation and a multiplication are commutative and given by
+
−1
0
1
−1
−1
−1 {−1, 0, 1}
0
−1
0
1
1 {−1, 0, 1} 1
1
· −1 0 1
−1 1 0 −1
0
0 0 0
1 −1 0 1
With the above operations, S becomes a hyperfield and is called the hyperfield of signs; one
can observe that the hyperoperation follows the rule of signs.
5
Example 2.7. (cf. [41]) Let T R := (R, +T , ·); the underlying set is the set of real numbers
and the multiplication is the usual multiplication of real numbers. The (hyper)addition is
given as follows:
{x}
if |x| > |y|
{y}
if |x| < |y|
x +T y =
{x}
if
x=y
[−|x|, |x|], a closed interval if x = −y
We call T R the Viro’s hyperfield; T R will play a role of the tropical semifield in §4.1.
Remark 2.8. The definition of a hyperring extension is subtle since any injective homomorphism of hyperrings does not have to be strict. However, in this paper, we stick with
Definition 2.4. With this definition, one can easily check that T R is a hyperfield extension
of S. In general, when H is a sub-hyperring of a hyperring R, for any a, b ∈ H, we have
a +H b ⊆ a +R b.
Next, we review the notion of (prime) ideals for hyperrings and the basic properties of them.
Definition 2.9. Let R be a hyperring.
(1) A nonempty subset I of R is a hyperideal if: ∀a, b ∈ I, ∀r ∈ R, we have a − r · b ⊆ I.
(2) A hyperideal I ( R is prime if I satisfies the following property: if xy ∈ I, then x ∈ I
or y ∈ I ∀x, y ∈ I.
(3) A hyperideal I ( R is maximal if I satisfies the following property: if J ( R is a
hyperideal of R which contains I, then I = J.
Proposition 2.10. (cf. [12, Proposition 2.12 and 2.13]) Let R be hyperring.
(1) Let I be a proper hyperideal of R, i.e. I 6= R. Then there exists a maximal hyperideal
m such that I ⊆ m.
(2) Any maximal hyperideal m is prime.
Definition 2.11. (cf. [36, §2]) Let R be a hyperring. We denote by Spec R the set of prime
hyperideals of R. One can impose the Zariski topology on Spec R as in the classical case. In
other words,
a subset A ⊆ Spec R is closed ⇐⇒ A = V (I) for some hyperideal I of R,
(1)
where V (I) := {p ∈ Spec R | I ⊆ p}.
The following proposition shows that Definition 2.11 indeed makes sense.
Proposition 2.12. (cf. [36, §2]) Let R be a hyperring and X = Spec R.
(1) Let {Ij }j∈J be a family of hyperideals of R. Then we have
[
\
Ij >),
V (Ij ) = V (<
(2)
j∈J
j∈J
S
S
where < j∈J Ij > is the smallest hyperideal containing j∈J Ij . Note that such
hyperideal exists since an arbitrary intersection of hyperideals is a hyperideal.
(2) Let I and I ′ be hyperideals of R, then we have
[
V (I) V (I ′ ) = V (I ∩ I ′ ).
(3)
Next, we review the notion of a localization of a hyperring. This construction also has been
promoted by R.Procesi-Ciampi and R.Rota in [36].
For a (multiplicative) submonoid S of a hyperring R, one defines the localization S −1 R as
follows: as a set, S −1 R is the set (R × S/ ∼) of equivalence classes, where
(r1 , s1 ) ∼ (r2 , s2 ) ⇐⇒ ∃x ∈ S
6
s.t. xr1 s2 = xr2 s1 .
(4)
Let [(r, s)] be the equivalence class of (r, s) ∈ R × S under the equivalence relation (4). A
hyperaddition of S −1 R is given by
[(r1 , s1 )] + [(r2 , s2 )] = [(r1 s2 + s1 r2 ), s1 s2 ] = {[(y, s1 s2 )] | y ∈ r1 s2 + s1 r2 }.
A multiplication is naturally given as follows:
[(r1 , s1 )] · [(r2 , s2 )] = [(r1 r2 , s1 s2 )].
We denote by rs an element [(r, s)]. Note that as in the classical case, the localization map,
S −1 : R −→ S −1 R sending r to 1r , is a homomorphism of hyperrings.
Proposition 2.13. (cf. [12, §3]) Let R be a hyperring and S be a (multiplicative) submonoid
of R.
(1) For a hyperideal I of R, the following set:
i
S −1 I := { | i ∈ I, s ∈ S}
s
is a hyperideal of S −1 R.
(2) If p is a prime hyperideal of R such that S ∩ p = ∅, then S −1 p is a prime hyperideal
of S −1 R.
(3) If S = R\p for some prime hyperideal p of R, then S −1 R has the unique maximal
hyperideal given by S −1 p.
The following theorem of Connes and Consani provides a useful way to construct hyperrings
from classical commutative algebras.
Theorem 2.14. (cf. [6, Proposition 2.6]) Let A be a commutative ring and
A× := {a ∈ A | ∃b ∈ A such that ab = 1}.
Let G be a (multiplicative) subgroup of A× and let A/G = {aG | a ∈ A} be the set of cosets.
(1) Multiplication:
m : A/G × A/G −→ A/G,
(xG, yG) 7→ xyG
∀x, y ∈ A.
(2) Addition:
+ : A/G × A/G −→ P(A/G)∗ ,
such that (xG, yG) 7→ {qG | q = xt + ys for some t, s ∈ G}
Then A/G is a hyperring which is called a quotient hyperring.
∀x, y ∈ A.
One can easily observe that for a field k with |k| ≥ 3, the Krasner’s hyperfield K is isomorphic
to the quotient hyperring k/k × .
3.
Algebraic aspects of Hyperrings
Hypergroups have been studied by many mathematicians in various fields; harmonic analysis
(cf. [27]), simple groups (cf. [2]), fuzzy logic (cf. [10]), and association schemes together with
Tits’ buildings (cf. [43]) to name a few. On the other hand, hyperrings have not brought much
attention after Krasner’s work until recently. In this section, we develop the basic algebraic
theory of hyperrings which will be used to develop the geometric theory in the next section.
7
3.1 Quotients of hyperrings. In algebra, the construction of a quotient object is usually essential to develop an algebraic theory. A particular case of quotient construction for
hyperrings has been studied by means of normal hyperideals (cf. [11], [12]). Recall that a
hyperideal I of a hyperring R is normal if
x+I −x⊆I
∀x ∈ R.
Remark 3.1. In [12], B.Davvaz and A.Salasi introduced the notion of a normal hyperideal
I of a hyperring R so that the following relation
x ≡ y ⇐⇒ (x − y) ∩ I 6= ∅
(5)
becomes an equivalence relation. One may observe that when R is a commutative ring, any
ideal of R is normal. In other words, in the classical case, the normal condition is redundant.
The definition of a normal hyperideal looks too restrictive for applications. For example, suppose that R is a hyperring extension of the Krasner hyperfield K. Then, for any x ∈ R, we
have x+x = {0, x}, therefore x = −x. It follows that the only (non-zero) normal hyperideal of
R is R itself. In this subsection, we prove that the relation (5) is, in fact, an equivalence relation without appealing to the normal condition on a hyperideal I. Furthermore, we show that
one can canonically construct a quotient hyperring R/I for any hyperideal I of a hyperring R.
Let R be a hyperring and I a hyperideal of R. We introduce the following relation on R
(cf. [12])
x ∼ y ⇐⇒ x + I = y + I,
(6)
S
where x + I := a∈I (x + a) and the equality on the right side of (6) is meant as an equality
of sets. Clearly, the relation (6) is an equivalence relation.
Remark 3.2. When R is a commutative ring, (6) is the classical equivalence relation obtained
from an ideal I: x ∼ y ⇐⇒ x − y ∈ I.
The following lemma provides an equivalent description of (6).
Lemma 3.3. Let R be a hyperring and I be a hyperideal of R. Let ∼ be the relation on R
as in (6). Then
x ∼ y ⇐⇒ (x − y) ∩ I 6= ∅, ∀x, y ∈ R.
(7)
Proof. Notice that (x − y) ∩ I 6= ∅ ⇐⇒ (y − x) ∩ I 6= ∅. Suppose that x ∼ y. Then by
definition we have x + I = y + I. By choosing 0 ∈ I, it follows that x + 0 = x ∈ y + I. Thus,
x ∈ y + a for some a ∈ I. By the reversibility property of R, we know that x ∈ y + a is
equivalent to a ∈ x − y. Thus we derive that a ∈ (x − y) ∩ I, hence (x − y) ∩ I 6= ∅.
Conversely, suppose that (x − y) ∩ I 6= ∅. We need to show that x + I = y + I. Since
the argument is symmetric, it is enough to show that x + I ⊆ y + I. For any t ∈ x + I,
there exists α ∈ I such that t ∈ x + α. Since (x − y) ∩ I 6= ∅, it follows that there exists
β ∈ (x − y) ∩ I. From the reversibility, this implies that x ∈ y + β. Therefore, we have
t ∈ x + α ⊆ (y + β) + α = y + (α + β). This implies that there exists γ ∈ (α + β) such that
t ∈ y + γ. But since α, β ∈ I we have γ ∈ I, thus t ∈ y + I.
Next, we use the equivalence relation (6) to define quotient hyperrings. We will use the
notations [x] and x + I interchangeably for the equivalence class of x under (6). We will also
use frequently the reversibility property of a hyperring without explicitly mentioning it.
Definition 3.4. Let R be a hyperring and I be a hyperideal of R. We define
R/I := {[x] | x ∈ R}
8
to be the set of equivalence classes of (6) on R. We impose on R/I two binary operations:
an addition:
[a] ⊕ [b] = (a + I) ⊕ (b + I) := {c + I | c ∈ a + b}
(8)
and a multiplication:
[a] ⊙ [b] := a · b + I.
(9)
Proposition 3.5. With the notation as in Definition 3.4, R/I is a hyperring with an addition
⊕ and a multiplication ⊙.
Proof. We first prove that operations ⊕ and ⊙ are well defined. For the addition, it is enough
to show that (a + I) ⊕ (b + I) = (a′ + I) ⊕ (b + I) for any [a] = [a′ ]. In fact, we only have to
show one inclusion since the argument is symmetric. Thus, we show that [a] ⊕ [b] ⊆ [a′ ] ⊕ [b].
If z + I ∈ (a + I) ⊕ (b + I), then we may assume z ∈ a + b. We need to show that there exists
w ∈ a′ + b such that [z] = [w]. But if z ∈ a + b = b + a then a ∈ z − b. In particular,
(a − a′ ) ⊆ (z − b) − a′ = z − (a′ + b).
(10)
Since [a] = [a′ ], it follows from Lemma 3.3 that there exists δ ∈ (a − a′ ) ∩ I. It also follows
from (10) that we have δ ∈ z − w for some w ∈ a′ + b and this implies (z − w) ∩ I 6= ∅.
Therefore, we have [z] = [w]. For the multiplication, we need to show that a · b + I = a′ · b + I.
Since (a − a′ ) ∩ I 6= ∅, we have δ ∈ (a − a′ ) ∩ I ⊆ (a − a′ ) which implies that (a − a′ )b ∩ I 6= ∅.
Therefore, [a · b] = [a′ · b] from Lemma 3.3. Hence, ⊕ and ⊙ are well defined.
Next, we prove that (R/I, ⊕) is a hypergroup. Clearly ⊕ is commutative. We claim that
X := ([a] ⊕ [b]) ⊕ [c] = {[d] = d + I | d ∈ a + b + c} := Y.
If [w] ∈ X, then [w] ∈ [r] ⊕ [c] for some [r] ∈ [a] ⊕ [b]. We may assume w ∈ r + c and r ∈ a + b.
Then we have w ∈ r + c ⊆ (a + b) + c = a + b + c. Thus [w] ∈ Y . Conversely, if [z] ∈ Y then
we may assume z ∈ a + b + c = (a + b) + c. This means z ∈ t + c for some t ∈ a + b. In turn,
this implies [z] ∈ [t] ⊕ [c], [t] ∈ [a] ⊕ [b]. Hence [z] ∈ X. It follows from the same argument
with [a] ⊕ ([b] ⊕ [c]) that the operation ⊕ is associative. The class [0] is the unique neutral
element. In fact, we have [0] ⊕ [x] = {[d] | d ∈ 0 + x = x} = [x]. Suppose that we have [w]
such that [w] ⊕ [x] = [x] for all [x] ∈ R/I. For x ∈ I, we have x ∈ w + x = x + w. Hence
w ∈ x − x ⊆ I. But one can see that [w] = [0] for all w ∈ I from Lemma 3.3. Therefore,
the neutral element is unique. Next, we claim that [0] ∈ [x] ⊕ [y] ⇐⇒ [y] = [−x]. Since
0 ∈ (x − x) we have [0] ∈ [x] ⊕ [−x]. Conversely, suppose that 0 + I ∈ (x + I) ⊕ (y + I) for
some y ∈ R. We need to show that y + I = −x + I. Since 0 + I ∈ (x + I) ⊕ (y + I), there exists
c ∈ x + y such that c + I = I. It follows that c ∈ I. Moreover, from c ∈ x + y = y − (−x),
we have that c ∈ (y − (−x)) ∩ I. Thus (y − (−x)) ∩ I 6= ∅ and [y] = [−x]. For the
reversibility property, if [x] ∈ [y] ⊕ [z], then we need to show that [z] ∈ [x] ⊕ [−y]. But
[x] ∈ [y] ⊕ [z] ⇐⇒ (x + I) ∈ (y + I) ⊕ (z + I) ⇐⇒ x + I = c + I for some c ∈ y + z. From
the reversibility property of R, z ∈ c − y. Thus [z] ∈ [c] ⊕ [−y]. But we have [x] = [c], hence
[z] ∈ [x] ⊕ [−y]. Finally, we only have to prove that ⊕, ⊙ are distributive. i.e.
([a] ⊕ [b]) ⊙ [c] = ([a] ⊙ [c]) ⊕ ([b] ⊙ [c]).
But this directly follows from that of R. This completes the proof.
In the sequel, we consider R/I as a hyperring with the addition ⊕ and the multiplication ⊙.
Proposition 3.6. Let R be a hyperring and I be a hyperideal of R. The projection map
π : R −→ R/I,
x 7→ [x]
is a strict, surjective homomorphism of hyperrings with Ker π = I.
9
Proof. Clearly, π is surjective and π(xy) = π(x)π(y). By the definition of a hyperaddition
(8), we have π(x+y) = [x+y] ⊆ [x]⊕[y]. This shows that π is a homomorphism of hyperrings.
For the strictness, take [c] ∈ [x] ⊕ [y]. Then there exists z ∈ x + y such that [z] = [c]. It
follows that π(z) = [z] = [c], thus π is strict. For the last assertion, suppose that π(x) = [0].
This implies that [x] = x + I = 0 + I = [0], hence x ∈ I. Therefore Ker π = I.
The next proposition shows that a quotient hyperring satisfies the universal property as in
the classical case.
Proposition 3.7. Let R and H be hyperrings and ϕ : R −→ H be a homomorphism of
hyperrings. Suppose that I is a hyperideal of R such that I ⊆ Ker ϕ. Then there exists a
unique hyperring homomorphism ϕ̃ : R/I −→ H such that ϕ = ϕ̃ ◦ π, where π : R −→ R/I
is the projection map as in Proposition 3.6.
Proof. Let us define
ϕ̃ : R/I −→ H,
ϕ̃([x]) = ϕ(x)
∀[x] ∈ R/I.
We first have to show that ϕ̃ is well defined. Let [x] = [y] for x, y ∈ R. Then we have
x + I = y + I, hence x ∈ y + c for some c ∈ I. Since c ∈ I ⊆ Ker ϕ, it follows that
ϕ(x) ∈ ϕ(y + c) ⊆ ϕ(y) + ϕ(c) = ϕ(y) + 0 = ϕ(y).
Therefore, ϕ(x) = ϕ(y) and ϕ̃ is well defined. Furthermore, since ϕ is a hyperring homomorphism, ϕ̃ is also a hyperring homomorphism. By the construction, we have ϕ = ϕ̃ ◦ π. The
uniqueness is clear.
3.2 Congruence relations. In this subsection, we define the notion of a congruence relation on a hyperring R and prove that there is a one-to-one correspondence between hyperideals
and congruence relations on R. Note that in the theory of semirings, this correspondence
fails in general as the following example shows. For the basic definitions and properties of
semirings we refer the readers to [18].
Example 3.8. Let M := Q≥0 be the semifield of nonnegative rational numbers with the usual
addition and the usual multiplication. Since M is a semifield, {0} and M are the only ideals
of M . One can easily see that {0} corresponds to the congruence relation:
x≡0
and M corresponds to the congruence relation:
x ≡ y ⇐⇒ x = y
∀x ∈ M
∀x, y ∈ M.
However there are more congruence relations. For example, one may consider the following
relation: ∀x, y ∈ M ,
x ≡2 y ⇐⇒ ∃k ∈ 2Z + 1 such that k(x − y) ∈ 2Z.
Clearly, ≡2 is reflexive and symmetric. Furthermore, suppose that x ≡2 y and y ≡2 z. Then
there exist odd integers k1 and k2 such that k1 (x − y), k2 (y − z) ∈ 2Z. One can easily check
that k1 k2 (x − z) ∈ 2Z. Therefore ≡2 is an equivalence relation. Next, when x ≡2 y and
α ≡2 β, ∃ odd integers k and t such that k(x − y), t(α − β) ∈ 2Z. It follows that
kt((x + α) − (y + β)) = tk(x − y) + kt(α − β) ∈ 2Z.
Also, one can easily see that kt(xα − yβ) ∈ 2Z. Hence, we conclude that
x ≡2 y and α ≡2 β =⇒ x + α ≡2 y + β and xα ≡2 yβ.
Therefore ≡2 is a congruence relation on M which does not have a corresponding ideal of
M . This example shows that an one-to-one correspondence between ideals and congruence
relations fails in this case. In fact, it is well known that if M is a semiring having no
10
nontrivial proper congruence relations then either M = B (the boolean semifield) or a field
(cf. [18, §7]).
We emphasize that in hyperring theory, a sum of two elements is no longer an element in
general but a set. Therefore, to define a congruence relation on a hyperring R, we need
a suitable notion stating when two subsets of R are equivalent. The following definition
provides such a notion.
Definition 3.9. Let R be a hyperring and ≡ be an equivalence relation on R. Let A, B be
two subsets of R. We write A ≡ B when the following condition holds:
∀a ∈ A, ∀b ∈ B
∃a′ ∈ A and ∃b′ ∈ B such that a ≡ b′ and a′ ≡ b.
(11)
Definition 3.10. Let R be a hyperring. A congruence relation ≡ on R is an equivalence
relation on R satisfying the following property:
∀x1 , x2 , y1 , y2 ∈ R,
x1 ≡ y 1 , x2 ≡ y 2
=⇒
x1 x2 ≡ y 1 y 2 ,
x1 + x2 ≡ y 1 + y 2 .
(12)
The following proposition shows that when a congruence relation ≡ is defined on R, there is
a canonical hyperring structure on the set R/ ≡ of equivalence classes. We let [r] denote an
equivalence class of r ∈ R under ≡.
Proposition 3.11. The set (R/ ≡) := {[r]| | r ∈ R} is a hyperring, where the addition is
defined by
[x] + [y] := {[t] | t ∈ x′ + y ′
and the multiplication law is given by
∀[x′ ] = [x], [y ′ ] = [y]}
[x] · [y] := [xy]
∀x, y, x′ , y ′ ∈ R,
∀x, y ∈ R.
(13)
(14)
Proof. Firstly, we prove that the addition and the multiplication are well defined. One easily
sees that (13) does not depend on representatives since it is already defined by all possible
representatives. Also it follows from (12) that the multiplication is well defined.
Secondly, we claim that (R/ ≡, +) is a hypergroup. We first show that + is associative by
proving the following equality
X := {[t] | t ∈ x′ + y ′ + z ′ , [x′ ] = [x], [y ′ ] = [y], [z ′ ] = [z]} = ([x] + [y]) + [z] := Y.
Indeed, if t ∈ x′ + y ′ + z ′ then t ∈ α + z ′ for some α ∈ x′ + y ′ . This implies that [t] ∈ [α] + [z]
and [α] ∈ [x]+[y], hence [t] ∈ Y . Conversely, if [t] ∈ ([x]+[y])+[z] then [t] ∈ [α]+[z] for some
[α] ∈ [x]+[y]. From (13), we have t ∈ α′ +z ′ for some α′ , z ′ ∈ R such that [α′ ] = [α], [z ′ ] = [z].
Also [α′ ] ∈ [x] + [y] since [α] = [α′ ]. This implies that α′ ∈ x′ + y ′ and t ∈ x′ + y ′ + z ′ for
some x′ , y ′ ∈ R such that [x′ ] = [x], [y ′ ] = [y]. The operations are trivially commutative. The
class [0] works as the zero element. Indeed, if [t] ∈ [x] + [0] then t ∈ x′ + y ′ with x′ ≡ x and
y ′ ≡ 0. It follows from (12) that x′ + y ′ ≡ x, hence t ≡ x. Thus [x] + [0] = [x]. An additive
inverse of [x] is [−x]. Indeed, since 0 ∈ x − x it is clear that [0] ∈ [x] + [−x]. Next, we show
that an inverse is unique. If [0] ∈ [x] + [y] then we have 0 ∈ x′ + y ′ with x′ ≡ x and y ′ ≡ y. It
follows that y ′ = −x′ and −x ≡ −x′ , therefore y ≡ y ′ = −x′ ≡ −x. Thus an additive inverse
uniquely exists. The reversibility property directly follows from that of R and the fact that
[x + y] ⊆ [x] + [y]. This proves that (R/ ≡, +) is a hypergroup.
Finally, one can observe that [1] works as the identity element. Therefore, all we have to
show is the distributive property:
[z]([x] + [y]) = [z][x] + [z][y] = [zx] + [zy],
∀[x], [y], [z] ∈ R/ ≡ .
If [α] ∈ [x] + [y], then α ∈ x′ + y ′ with [x′ ] = [x], [y ′ ] = [y]. This implies zα ∈ zx′ + zy ′ . But
since [zx′ ] = [zx], [zy ′ ] = [zy], it follows that [zα] ∈ [zx] + [zy]. Conversely if [t] ∈ [zx] + [zy]
then t ∈ α + β with [α] = [zx], [β] = [zy]. Thus α + β ≡ zx + zy = z(x + y), and t ≡ zγ for
some γ ∈ x + y. This completes the proof.
11
In what follows, for a hyperring R and a congruence relation ≡ on R, we always consider
R/ ≡ as a hyperring with the structure defined in Proposition 3.11.
Proposition 3.12. Let R be a hyperring and ≡ be a congruence relation on R. Then the
map
π : R −→ R/ ≡, r 7→ [r] ∀r ∈ R
is a strict surjective hyperring homomorphism.
Proof. The map π is clearly a surjective hyperring homomorphism. We prove that π is also
strict by showing that [x] + [y] ⊆ [x + y]. If [t] ∈ [x] + [y] then t ∈ x′ + y ′ for some x′ , y ′ ∈ R
such that x′ ≡ x and y ′ ≡ y. It follows from (12) that x + y ≡ x′ + y ′ . From (11), there exists
α ∈ x + y such that [α] = [t]. Therefore, [t] = [α] ∈ [x + y] and π is strict.
Proposition 3.13. Let π : R −→ R/ ≡ be the canonical projection as in Proposition 3.12.
Let I = Ker π. Then
ϕ : R/I −→ R/ ≡, < r >7→ [r] ∀r ∈ R
is an isomorphism of hyperrings, where < r > is an equivalence class of r in R/I under the
equivalence relation (6) and [r] is an equivalence class of r in R/ ≡ under ≡.
Proof. This follows from Proposition 3.12 and Proposition 2.11 of [12] which states that the
first isomorphism theorem for hyperrings holds when a given homomorphism is strict.
It follows from Proposition 3.13 that for a congruence relation ≡ on R, one can find a
hyperideal I of R such that R/I ≃ (R/ ≡). Conversely, in the next proposition, we prove
that for any hyperideal I, one can find a congruence relation ≡ such that R/I ≃ (R/ ≡).
Remark 3.14. Note that some of the algebraic properties of a hyperring differ greatly from
those of a commutative ring. For example, a hyperring does not satisfy the doubly distributive
property (cf. Remark 3.18). Thus one should be careful when generalizing classical results of
commutative rings to hyperrings.
Proposition 3.15. Let R be a hyperring and I be a hyperideal of R. Then the relation ≡
such that
x ≡ y ⇐⇒ x + I = y + I
is a congruence relation and R/I ≃ (R/ ≡).
Proof. Clearly ≡ is an equivalence relation. If x1 ≡ y1 and x2 ≡ y2 , we have
xi + I = yi + I,
i = 1, 2.
(15)
Thus we can find α, β ∈ I such that x1 ∈ y1 + α, x2 ∈ y2 + β. By multiplying these two, one
obtains
x1 x2 ∈ (y1 + α)(y2 + β) ⊆ y1 y2 + y1 β + y2 α + αβ.
Therefore, for any t ∈ I, we have x1 x2 + t ⊆ y1 y2 + (y1 β + y2 α + αβ + t). But since α, β, t ∈ I,
it follows that (y1 β + y2 α + αβ + t) ⊆ I. Hence, x1 x2 + t ⊆ y1 y2 + I and x1 x2 + I ⊆ y1 y2 + I.
Since the argument is symmetric, we have
x1 x2 + I = y1 y2 + I ⇐⇒ x1 x2 ≡ y1 y2 .
For the other condition of a congruence relation, we need to show (x1 + x2 ) ≡ (y1 + y2 ). It
is enough to show that ∀ t ∈ x1 + x2 , there exists y ∈ y1 + y2 such that t ≡ y. We can take
α, β ∈ I such that x1 ∈ y1 + α, x2 ∈ y2 + β from (15). It follows that
t ∈ (x1 + x2 ) ⊆ (y1 + y2 ) + (α + β).
Hence, t ∈ y + γ for some y ∈ y1 + y2 , γ ∈ α + β ⊆ I. This implies that t ≡ y from (7) and the
reversibility property of R. It is clear that in this case the kernel of a canonical projection
12
map π : R −→ R/ ≡ is I. It follows from the first isomorphism theorem of hyperrings
(cf. [12, Proposition 2.11]) that R/I ≃ R/ ≡ since π is strict.
Remark 3.16. Let R be a hyperring and I be a hyperideal of R. In a quotient hyperring
R/I, we defined the addition as
[a] ⊕ [b] = {[c]|c ∈ a + b}
and we proved that x ∼ y ⇐⇒ x + I = y + I is a congruence relation. In this case, we defined
the addition as
[a] + [b] = {[c] | c ∈ a′ + b′ ∀[a′ ] = [a], [b′ ] = [b]}.
At first glance, [a] ⊕ [b] and [a] + [b] seem different, but in fact they are the same sets. Clearly
[a] ⊕ [b] ⊂ [a] + [b]. Conversely, assume that t′ ∈ a′ + b′ for some [a′ ] = [a], [b′ ] = [b]. Since
a′ + I = a + I, b′ + I = b + I, we can find α, β ∈ I such that a′ ∈ a + α, b′ ∈ b + β. This
implies that t′ ∈ a′ + b′ ⊆ (a + b) + (α + β). But since (α + β) ⊆ I, it follows that t′ ∈ t + γ
for some t ∈ (a + b), γ ∈ I. By the reversibility property of R, γ ∈ t′ − t. In other words,
(t − t′ ) ∩ I 6= ∅, hence [t] = [t′ ]. This shows that [a] + [b] ⊆ [a] ⊕ [b].
3.3 Analogues of classical lemmas. In this subsection, we reformulate several basic
results in commutative algebra in terms of hyperrings. Throughout this subsection, we denote
by R a hyperring and by V (I) the set of of prime hyperideals of R containing a hyperideal
I. We also denote by N il(R) the intersection of all prime hyperideals of R.
Lemma 3.17. Let I ⊆ R be a hyperideal. Then the following set:
√
I := {r ∈ R | ∃n ∈ N such that r n ∈ I}
is a hyperideal.
√
√
Proof. Trivially we have 0 ∈ I. Suppose that a ∈ I, then an ∈ I for some
√ n ∈ N. Since I
is a hyperideal, for r ∈ R, we have r n an = (ra)n ∈ I. It follows that ra ∈√ I. Clearly, (−a)n
Finally,
suppose
is either an √
or −an . Since both an and −an are in I, it follows that −a ∈ I.P
l k l−k
a
b
⊆ I.
that a, b ∈ I and an , bm ∈ I. Then, for l ≥ (n + m), we have (a + b)l ⊆
k
√
√
This implies that (a + b) ⊆ I; therefore, I is a hyperideal.
Remark 3.18. In general, a hyperring does not satisfy the doubly distributive property (cf.
[41, pp 13 − 14]). In other words, the following identity:
(a + b)(c + d) = ac + ad + bc + bd
is in general not fulfilled. Instead, the following identity:
(a + b)(c + d) ⊆ ac + ad + bc + bd
holds.
Lemma 3.19. Let R be a hyperring and I a hyperideal of R. Then
\
√
I=
p.
p∈V (I)
√
Proof. Suppose that a ∈ I,√then an ∈ I ⊆ p for all p ∈ V (I). Since p is a prime hyperideal,
(I).
it follows that a ∈ p; hence, I ⊆ p for all p ∈ V √
T
Conversely, suppose that f ∈ p∈V (I) p and f 6∈ I. This implies that
S := {1, f, f 2 , ....} ∩ I = ∅.
Let Σ√be the set of hyperideals J of R such that S ∩ J = ∅ and I ⊆ J. Then Σ 6= ∅ since we
have I ∈ Σ. By Zorn’s lemma (ordered by inclusion), Σ has a maximal element q. Then
13
q is a prime hyperideal. Indeed, by definition, q is a hyperideal. Therefore, all we have to
prove is that q is prime. One can easily check, for x ∈ R, the following set:
[
q + xR := {a + b | a ∈ q, b ∈ xR}
is a hyperideal. If x, y 6∈ q then q is properly contained in q + xR and q + yR. Thus, q + xR,
q + yR 6∈ Σ from the maximality of q in Σ. It follows that f n ∈ q + xR and f m ∈ q + yR
for some n, m ∈ N. In other words, f n ∈ a1 + xr1 , f m ∈ a2 + yr2 for some a1 , a2 ∈ q and
r1 , r2 ∈ R. Therefore, we have
f n+m ∈ (a1 + xr1 )(a2 + yr2 ) ⊆ a1 a2 + a1 yr2 + a2 xr1 + xyr1 r2 ⊆ q + xyR.
This implies that xy 6∈ q because if xy ∈ q then f n+m ∈ q, and we assumed that f l 6∈ q for
all l ∈ N. It follows that q is a primeThyperideal containing I such that S ∩ q = ∅. However,
this is impossible since we took f ∈ p∈V (I) p. This completes the proof.
For a family {Xα }S
α∈J of subsets Xα ⊆ R, we denote by < Xα >α∈J the smallest hyperideal
of R containing ( α∈J Xα ). Note that < Xα >α∈J always exists since an intersection of
hyperideals is a hyperideal as in the classical case. We call < Xα >α∈J the hyperideal
generated by {Xα }α∈J .
Lemma 3.20. Let J be an index set.
(1) Let h ∈ R. Then the hyperideal generated by h is
hR := {hr | r ∈ R}.
(16)
(2) Suppose that Ii is the principal hyperideal generated by an element hi ∈ R for each
i ∈ J. Then
X
< Ii >i∈J = {r ∈ R | r ∈
bi hi , bi ∈ R, X ⊆ J, |X| < ∞}.
(17)
i∈X
(3) Let {Ii }i∈J be a family of hyperideals Ij ⊆ R. Then
X
< Ii >i∈J = {r ∈ R | r ∈
bi hi , bi ∈ R, hi ∈ Ii , X ⊆ J, |X| < ∞}.
(18)
i∈X
Proof. The proof is similar to the classical case.
Let R× := {r ∈ R | rr ′ = 1 for some r ′ ∈ R} and J(R) be the intersection of all maximal
hyperideals of R. The following lemma has been proven in [12].
Lemma 3.21. ( [12, Proposition 2.12, 2.13, 2.14])
(1) x ∈ J(R) ⇐⇒ (1 − xy) ⊆ R× ∀y ∈ R.
(2) For any hyperideal I ( R, we have V (I) 6= ∅.
One imposes the Zariski topology on the set Spec R of prime hyperideals of R as in the
classical case (cf. Definition 1, Proposition 2.12). In what follows, we consider X = Spec R
as a topological space equipped with the Zariski topology. Then, as in classical algebraic
geometry, we have the following.
Proposition 3.22. X = Spec R is a disconnected topological space if and only if R has a
(multiplicative) idempotent element different from 0, 1.
Proof. Suppose that e 6= 0, 1 is an idempotent element of R. Then we have e2 = e, and
it follows that 0 ∈ e(e − 1). Therefore there is an element f ∈ e − 1 such that ef = 0.
Moreover, f 6= 0 since e 6= 1. Similarly, f can not be 1 since ef = 0 and e 6= 0. Together
with Lemma 3.21, it follows that V (e) and V (f ) are nonempty subsets of X. Hence we have
X = Spec R = V (e) ∪ V (f ) since ef = 0.
14
Now we claim that V (e) ∩ V (f ) = ∅. Indeed, if p ∈ V (e) ∩ V (f ), then e, f ∈ p. This implies
that −e, −f ∈ p and (f − e) ⊆ p. However, we have
f ∈ e − 1 = −1 + e ⇐⇒ −1 ∈ f − e.
Thus we should have 1 ∈ p and it is impossible. It follows that {V (e)c , V (f )c } becomes the
disjoint open cover of X and hence X is disconnected.
Conversely, suppose that X = Spec R = U1 ∪ U2 , where U1 and U2 are disjoint open subsets
of X. In particular, U1 and U2 are also closed. We may assume the following (cf. Proposition
2.12): for some hyperideals I and J,
X = Spec R = V (I) ∪ V (J) = V (IJ), V (I) ∩ V (J) = V (< I, J >) = ∅.
T
Recall that N il(R) := p∈X p. It follows from Lemma 3.17 and 3.19 that N il(R) is the set
√
of all nilpotent elements of R. Since V (IJ) = X = V (N il(R)), we√have IJ = N il(R)
from Lemma
√ 3.19. Moreover, the fact V (< I, J >) = ∅ implies that < I, J > contains 1.
Otherwise, < I, J > does
√ not contain any unit element, and V (< I, J >) 6= ∅ from Lemma
3.21. It follows that 1 ∈ < I, J >, hence 1 ∈< I, J >. From Lemma 3.20, there exist a ∈ I
and b ∈ J such that 1 ∈ a + b.
We claim that a, b 6∈ N il(R). Indeed, suppose that a ∈ N il(R). Then a ∈ p for any p ∈ X.
This implies that for p ∈ V (J), p contains both a and b and hence 1 ∈ (a + b) ⊆ p.
Next, since
V (a) ⊇ V (I) ⇐⇒ D(a) ⊆ (V (I))c = V (J),
we have
D(a) ⊆ V (J), D(b) ⊆ V (I).
This implies
(V (a) ∪ V (b))c = D(a) ∩ D(b) ⊆ V (I) ∩ V (J) = ∅,
thus V (a) ∪ V (b) = X. Suppose that A =< a > and B =< b >. Then ab √
∈ A ∩ B, and it
follows that V (a) ∪ V (b) = V (A) ∪ V (B) = V (AB). Thus we have AB ⊆ AB = N il(R).
Therefore, ab ∈ N il(R), in turn, (ab)n = an bn = 0 for some n ∈ N. However, an , bn 6= 0 since
a, b 6∈ N il(R). We observe the following:
X
1 ∈ a + b =⇒ 1 ∈ (a + b)n ⊆
dk ak bn−k =⇒ 1 ∈ (an + bn ) + abf, for some f ∈ R. (19)
Since ab ∈ N il(R), clearly abf ∈ N il(R) ⊆ J(R). It follows from (19) that
1 ∈ α + abf,
for some α ∈ an + bn .
This implies that α ∈ 1 − abf . But since abf ∈ J(R), from Lemma 3.21, α is a unit. Let
β = α−1 . Then
α ∈ an + bn =⇒ αβ = 1 ∈ (an + bn )β = an β + bn β =⇒ bn β ∈ 1 − an β.
One observes that an β, bn β 6= 0 since an , bn 6= 0 and β is a unit. Furthermore, an β, bn β 6= 1.
Since an β = 1 ⇐⇒ an = β −1 = α, it would imply that an = α ∈ I. Therefore V (I) = ∅. But
we assumed that V (I) 6= ∅.
Finally let us define an element e = an β. Then we know, from the above, e 6= 0, 1. Furthermore, we have
e2 − e = e(e − 1) = an β(an β − 1).
Since we have bn β ∈ 1 − an β and bn βan β = an bn β 2 = 0, it follows that
0 ∈ e(e − 1) = e2 − e.
Hence, from the uniqueness of an inverse, we have e2 = e and e 6= 0, 1.
Proposition 3.23. X = Spec R is irreducible if and only if N il(R) is a prime hyperideal.
15
Proof. The proof is similar to the classical case.
4.
Geometric aspects of Hyperrings
In this section, we investigate geometric aspects of hyperrings. Our goal is to reformulate
tropical varieties in the framework of hyperfields and to realize a hyperring R (without zerodivisors) as the hyperring of global regular functions on the topological space X = Spec R.
Remark 4.1. It is noteworthy that finite hyperfield extensions of the Krasner hyperfield K
already have very interesting geometric aspect, namely a correspondence to incidence geometry. More precisely, in [6], Connes and Consani proved that for a finite hyperfield extension
R of the Krasner hyperfield K, one can associate an incidence geometry (with some extra
conditions) and vice versa. Furthermore, under their construction, the open problem of completely classifying finite hyperfield extensions of K is very closely related with a long-standing
conjecture on the existence of a non-Desarguesian finite projective plane with a simply transitive group of collineations. For more details, see [6], [38], [39]. Also, for more connections
between incidence geometry and hypergroups, we refer the reader to [10].
4.1 A tropical variety as an algebraic set over a hyperfield. In this subsection, we
recast a tropical variety as the ‘positive part’ of an algebraic set over a hyperfield. For more
details about tropical geometry we refer the reader to [30]. Note that our notion of a tropical
variety in this subsection (cf. Equation (28), Remarks 4.17) a bit differs from the one given
in [30]. However, such choice makes no difference in further study. Also, strictly speaking,
a tropical variety is the support of a polyhedral complex with a balancing condition, but we
only consider a tropical variety as a set without a balancing condition.
The main motivation for the study proposed in this section comes from the following observation: the definition of a tropical variety does not seem natural in the sense that it is
not defined as the set of solutions of polynomial equations, but as the set of points where a
maximum (or a minimum depending on the conventions) is attained at least twice. Recently,
there have been several attempts to build an algebraic foundation of tropical geometry: for
instance, [16], [22], [28], [34], [41]. We take up a hyperstructural approach and, in Proposition
4.18, we show that there exists a more natural description of a tropical variety by implementing a symmetrization procedure and an associated algebraic set over a hyperfield.
In this subsection, we use the multi-index notation: for I = (i1 , ..., in ) ∈ Nn , X I := xi11 · · · xinn .
Let us first define the notion of a polynomial equation with coefficients in a hyperring R.
Definition 4.2. Let R be a hyperring.
(1) By a monomial f with n variables over R we mean a single term:
f := aI X I ,
aI ∈ R,
I ∈ Nn .
(2) By a polynomial f with n variables over R we mean a finite formal sum:
X
f :=
aI X I ,
aI ∈ R, aI = 0 for all but finitely many I
(20)
(21)
I∈Nn
such that there is no repetition of monomials with the same multi-index I. We denote
by R[x1 , ..., xn ] the set of polynomials with n variables over R.
One can be easily misled in the above definition. For example, (x − x) is not a polynomial
over the hyperfield of signs S (Example 2.6) since the term x is repeated. The reason why we
do not want (x − x) to be a polynomial is that whenever a repetition of a monomial occurs,
an ambiguity follows. For instance, we may have (x − x) = (1 − 1)x = {−x, 0, x}. In other
words, (x − x) does not represent a single element. Furthermore, one can not perform the
16
basic arithmetic in general. For example, (x2 − 1) differs from (x + 1)(x − 1) as an element of
S[x] (cf. Example 4.3). Note that (x + 1)(x − 1) is not even a polynomial over S since it is not
of the form (21) Indeed it represents several different polynomials as you can see in Example
4.3. Therefore, for f, g ∈ R[x1 , ..., xn ], we say f = g only if they are entirely identical.
We directly generalize the classical P
addition and multiplication
of polynomial equations to
P
j ∈ R[x], the addition and the
R[x1 , ..., xn ]. For example, for f = ni=0 ai xi , g = m
b
x
j=0 j
multiplication of f and g are given by
m
n
X
X
i
bi x i ,
(ai + bi )x +
f + g :=
i=0
f g :=
n+m
X
(
X
i=0 r+l=i
i=n+1
ar bl )xi . (n ≤ m)
(22)
P
Notice that (ai + bi ) and r+l=i ar bl are not elements but subsets of R in general. Therefore,
the addition and the multiplication defined in this way are in general multi-valued as the
following example shows.
Example 4.3. Let R = S, the hyperfield of signs. For x − 1, x + 1 ∈ S[x], we have
(x + 1)(x − 1) = x2 + (1 − 1)x − 1 = {x2 − 1, x2 + x − 1, x2 − x − 1}.
(x + 1) + (x − 1) = (x + x) + (1 − 1) = {x, x + 1, x − 1}.
Remark 4.4. We emphasize that R[x1 , ..., xn ] is only a set with two multi-valued binary operations. However it appears, in some circumstance, that R[x1 , ..., xn ] behaves like a hyperring
as Example 4.13 shows.
Throughout this section, we will simply write a polynomial over R instead of a polynomial
with n variables when there is no possible confusion.
Definition 4.5. Let R be a hyperring and R[x1 , ..., xn ] be the set P
of polynomials over R. Let
L be a hyperring extension of R. By an evaluation f (α) of f = I aI X I ∈ R[x1 , ..., xn ] at
α = (α1 , ..., αn ) ∈ Ln we mean the following set:
X
aI αI ⊆ L.
(23)
f (α) :=
I
Example 4.6. Let R = L = S, the hyperfield of signs. Suppose that f = x2 − x ∈ R[x].
Then: f (1) = S, f (0) = {0}, f (−1) = {1}.
Let L = T R, Viro’s hyperfield (Example 2.7). Then: f (1) = [−1, 1], f (0) = {0}, f (−1) =
{1}.
An intuitive definition of a set of solutions of f ∈ R[x1 , ..., xn ] over a hyperring extension L
of R would be the following set:
{α = (α1 , ..., αn ) ∈ Ln | 0 ∈ f (α)}.
(24)
However, (24) may depend on the way one writes f (cf. [42, §5.2]). Moreover, for two different
elements f, g ∈ R[x1 , ..., xn ], we may have f (α) = g(α) ∀α ∈ Ln . For example, suppose that
f = x2 − 1, g = x4 − 1 ∈ S[x]. Then f (a) = g(a) ∀a ∈ T R, but f 6= g as elements of S[x]. To
resolve these issues, we introduce the following relation on R[x1 , ..., xn ].
Definition 4.7. Let R be a hyperring and R[x1 , ..., xn ] be the set of polynomials over R. Let
L be a hyperring extension of R. For f, g ∈ R[x1 , ..., xn ], we define
f ≡L g ⇐⇒ f (α) = g(α) (as sets)
∀α ∈ Ln .
(25)
Remark 4.8. The relation (25) depends on a hyperring extension L of R. However, we note
that if H is a hyperring extension of L then f ≡H g =⇒ f ≡L g.
17
The following statement is clear in view of the above definition.
Proposition 4.9. Let R be a hyperring and L be a hyperring extension of R. Then the
relation (25) on R[x1 , ..., xn ] is an equivalence relation.
Proof. This is straightforward since (25) is defined in terms of an equality of sets.
Under the equivalence relation (25), we can consider each polynomial f ∈ R[x1 , ..., xn ] as the
following function:
f : Ln −→ P ∗ (L),
α = (α1 , ..., αn ) 7→ f (α),
(26)
where P ∗ (L) is the set of nonempty subsets of L.
Definition 4.10. Let R be a hyperring and R[x1 , ..., xn ] be the set of polynomials over R.
Let L be a hyperring extension of R. By a solution of f over L we mean an element a =
(a1 , ..., an ) ∈ Ln such that 0 ∈ f (a) where we consider f as in (26) under ≡L . We denote by
VL (f ) the set of solutions of f over L.
Remark 4.11. Suppose that H is a hyperring extension of L. It clearly follows from the
definition that VL (f ) ⊆ VH (f ).
Let R[x1 , ..., xn ]/ ≡L be the set of equivalence classes of R[x1 , ..., xn ] under ≡L . When L is
doubly distributive (hence, so is R), the multiplication on R[x1 , ..., xn ]/ ≡L induced from the
multi-valued multiplication on R[x1 , ..., xn ] is well defined and single-valued (cf. Remark 3.18
for the definition of doubly distributive property). In fact, suppose that f, g ∈ R[x1 , ..., xn ]. If
h ∈ f · g then it follows from the doubly distributive property of L that h(α) = f (α) · g(α) for
all α ∈ Ln (cf. [41, the remark after Theorem 4.B.]). Therefore, under ≡L , the set f ·g becomes
a single equivalence class. This is one of the advantages of working with R[x1 , ..., xn ]/ ≡L
rather than working directly with R[x1 , ..., xn ].
Example 4.12. Let R = S, the hyperfield of signs and L = T R, Viro’s hyperfield. Then
L satisfies the doubly distributive property (cf. [41, Theorem 7.B.]). In Example 4.3, we
computed (x + 1)(x − 1) = {x2 − 1, x2 + x − 1, x2 − x − 1} in S[x]. One can easily see that
2
if |a| > 1
a
[−1, 1] if |a| = 1
∀a ∈ T R, (a2 − 1) = (a2 + a − 1) = (a2 − a − 1) =
−1 if |a| < 1
Therefore (x2 − 1) ≡T R (x2 + x − 1) ≡T R (x2 − x − 1), and x = 1, −1 are the only solutions
of the equivalence class of (x2 − 1) under ≡T R .
The following example shows that in general one can not expect R[x1 , ..., xn ]/ ≡L to be a
hyperring even when L satisfies the doubly distributive property.
Example 4.13. Let R = L = K, the Krasner’s hyperfield. Let [f ] be the equivalence class of
f ∈ K[x] under ≡K . Then any two non-constant polynomials over K with the same constant
term are equivalent under ≡K . It follows that
(K[x]/ ≡K ) = {[0], [1], [1 + x], [x]}.
For the notational convenience, let 0 := [0], 1 := [1], a := [1 + x], and b := [x]. Then we have
the following tables:
+
0
1
a
b
0
1
a
b
0
1
a
b
1 {0, 1}
{a,b}
a
a {a, b} {0, 1, a, b} {1, a}
b
a
{1, a}
{0, b}
18
·
0
1
a
b
0
0
0
0
0
1
0
1
a
b
a
0
a
a
b
b
0
b
b
b
One can check by using the above tables that (K[x]/ ≡K , +) is a hypergroup, but it fails to
satisfy the distributive law. For example, we have
a(1 + b) = {a} ⊆ a + ab = {1, a}.
However, (K[x]/ ≡K , +, ·) still satisfies the weak version of the distributive law: a(b + c) ⊆
ab + ac. It follows that (K[x]/ ≡K , +, ·) is not a hyperring but a multiring introduced in [32]
which assumes the same axioms with a hyperring but the distributive law is weaker as above.
We recall that if A and B are multirings, one defines a homomorphism of multirings as a
map ϕ : A −→ B such that
ϕ(a + b) ⊆ ϕ(a) + ϕ(b),
ϕ(ab) = ϕ(a)ϕ(b),
ϕ(1A ) = 1B ,
ϕ(0A ) = 0B
∀a, b ∈ A.
Consider the set A1K (K) := Hommulti (K[x]/ ≡K , K) of multiring homomorphisms from
K[x]/ ≡K to K. If ϕ ∈ A1K (K) then one has ϕ(a) = 1 since a + a = (K[x]/ ≡K ). One
can also easily check that ϕ(b) can be any point of K. It follows that A1K (K) = {ϕ0 , ϕ1 },
where ϕ0 (b) = 0 and ϕ1 (b) = 1. This suggests that one may consider K[x]/ ≡K as the
‘coordinate ring’ of an affine line over K.
In the sequel, we will always consider an element f of R[x1 , ..., xn ] under the equivalence relation (25) with a predesignated hyperring extension L of R. We will use the symmetrization
process introduced in [20] and the previous result of the author in [23]. Let us briefly explain
the symmetrization process. For more details, see Appendix A.
Let M be a totally ordered idempotent monoid (with a canonical order). We define
s(M ) := {(s, 1), (s, −1), 0 = (0, 1) = (0, −1) | s ∈ M \{0}}.
Also, we denote (s, 1) := s, (s, −1) := −s, and |(s, 1))| = |(s, −1)| = s. For any X = (x, p) ∈
s(M ), we define sign(X) = p. Then s(M ) is a hypergroup with the addition given by
if |x| > |y| or x = y
x
y
if |x| < |y| or x = y
x+y =
(27)
[−x, x]= {(t, ±1) | t ≤ |x|)} if y = −x
Let s : M −→ s(M ), s 7→ (s, 1) be the associated map.
In [23], we showed that the symmetrization process can be generalized, in some cases, to
semirings by imposing a coordinate-wise multiplication. The symmetrization of a totally
ordered idempotent semiring is not a hyperring in general, however, we showed that the symmetrization of a totally ordered idempotent semifield is a hyperfield. The symmetrization is
the main technique which we use in this section. Also, the essential idea is similar to that of
J.Giansiracusa and N.Giansiracusa’s construction of tropical schemes in [16].
Let Rmax := {R ∪ {−∞}, +, max} be the tropical semifield with a maximum convention. Recall that a tropical hypersurface trop(V (f )) defined by a polynomial equation f ∈ Rmax [x1 , ..., xn ]
is the following set:
{a ∈ (Rmax )n | the maximum of f is achieved at least twice}.
(28)
For the notational convenience, let M = Rmax and MS = R = s(Rmax ), the symmetrization
of M . Note that MS is a hyperfield since M is a totally ordered idempotent semifield as we
mentioned above.
P
Let f (x1 , ..., xn ) ∈ M [x1 , ..., xn ]. Write f (x1 , ..., xn ) = i mi (x1 , ..., xn ) as a sum of distinct
monomials and fix this presentation. Then we define
X
fî (x1 , ..., xn ) :=
mj (x1 , ..., xn ) ∈ M [x1 , ..., xn ] for each i.
j6=i
19
By identifying an element a ∈ M with the element (a, 1) ∈ MS = R, we define
X
f˜î (x1 , ..., xn ) := ( (mj (x1 , ..., xn ), 1)) + (mi (x1 , ..., xn ), −1) ∈ R[x1 , ..., xn ].
j6=i
With these notations we have the following description of a tropical hypersurface.
Proposition 4.14. With the same notation as above, we let
[
V (f˜ ) := {z ∈ Rn | 0R ∈ f˜ (z)}, HV (f ) :=
V (f˜ ).
î
î
î
i
With the coordinate-wise symmetrization map, ϕ = sn : M n −→ Rn , we have a set bijection:
trop(V (f )) ≃ (HV (f ) ∩ Img(ϕ)),
where trop(V (f )) is the tropical variety defined by f ∈ M [x1 , ..., xn ].
Remark 4.15. Even though we started by fixing one presentation of a polynomial equation
f ∈ M [x1 , ..., xn ], the set trop(V (f )) does not depend on the chosen presentation of fT. Therefore, even though HV (f ) may vary depending on a presentation of f , the set HV (f ) Img(ϕ)
is invariant of the presentation as long as there is no repetition of monomials.
Before we prove Proposition 4.14, we present an example to show how this procedure works.
Example 4.16. Let f (x, y) = x + y + 1 ∈ Rmax [x, y]. Then trop(V (f )) consists of three rays:
trop(V (f )) = {(x, y) ∈ Rmax × Rmax | 1 ≤ y = x, or y ≤ x = 1, or x ≤ y = 1}.
With the above notations, we have
fx (x, y) := y + 1,
f˜x (x, y) := (1, 1)y + (1, 1) + (1, −1)x.
(29)
Since we only consider the ‘positive’ solutions, x and y should be of the form (t, 1). Therefore,
in this case, we may assume that
f˜x (x, y) = (y, 1) + (1, 1) + (x, −1) = (y + 1, 1) + (x, −1).
(30)
By the definition of symmetrization, we have
0R ∈ f˜x (x, y) ⇐⇒ y + 1 = x in Rmax .
Thus, we obtain
{(x, y) ∈ Rmax × Rmax | 1 ≤ y = x, or y ≤ x = 1} = V (f˜x (x, y))
Similarly with fy (x, y) = x + 1 we have
0R ∈ f˜y (x, y) ⇐⇒ x + 1 = y in Rmax .
\
(s(Rmax ) × s(Rmax )).
This time, we obtain
{(x, y) ∈ Rmax × Rmax | 1 ≤ x = y, or x ≤ y = 1} = V (f˜y (x, y))
Finally, with f1 (x, y) = x + y, we have
0R ∈ f˜1 (x, y) ⇐⇒ x + y = 1 in Rmax .
\
(s(Rmax ) × s(Rmax )).
\
(s(Rmax ) × s(Rmax )).
This gives
{(x, y) ∈ Rmax × Rmax | y ≤ x = 1, or x ≤ y = 1} = V (f˜1 (x, y))
By taking the union of all three we recover
[
\
trop(V (f )) = (
V (f˜z (x, y))) (s(Rmax ) × s(Rmax )).
z∈{x,y,1}
Now we give the proof of Proposition 4.14.
20
Proof. When f is a single monomial, the result is clear since 0M and 0R will be the only
solution for each. Thus we may assume that f is not a monomial. If z = (z1 , ..., zn ) ∈
trop(V (f )) then there exist mi (x1 , ..., xn ), mj (x1 , ..., xn ) with i 6= j such that the value
mi (z) = mj (z) attains the maximum among all monomials mr (z). Then we have f (z) =
mi (z) = mj (z) ∈ M . It follows that
0R ∈ (f (z), 1) + (mi (z), −1) = (f (z), 1) + (mi (z), −1) = f˜ (ϕ(z)).
î
î
Thus we have ϕ(z) ∈ HV (f ).
Conversely, suppose that ϕ(z) ∈ HV (f ) ∩ Img(ϕ). Let ϕ(z) = (ϕ(zi )), where (zi , 1) ∈
S
M × {1} ⊆ R. Then, by the definition of HV (f ) = V (f˜î ), ϕ(z) is an element of V (f˜î ) for
some i. In other words,
X
0R ∈ ( (mj (z), 1)) + (mi (z), −1) = f˜î (ϕ(z)).
j6=i
Therefore, there exists some r 6= i such that
X
(mj (z), 1) = (mr (z), 1) = (mi (z), 1) and
j6=i
mj (z) ≤ mr (z) = mi (z)
∀j 6= i, r.
It follows that z ∈ trop(V (f )). So far we have showed that
ϕ(trop(V (f ))) = HV (f ) ∩ Img(ϕ).
Since ϕ is one-to-one, we conclude that trop(V (f )) ≃ HV (f ) ∩ Img(ϕ) as sets. In other
words, trop(V (f )) is the ‘positive’ part of HV (f ).
Remark 4.17. Our definition (28) of trop(V (f )) may contain a point a = (a1 , ..., an ) such
that ai = −∞(= 0M ) for some i. This is little different from the conventional definition
of a tropical hypersurface in which one excludes such points. However, from the proof of
Proposition 4.14, one can observe that the subset of trop(VT(f )) which does not have 0M at
any coordinate maps bijectively onto the subset of H(V (f )) Img(ϕ) which does not have 0R
at any coordinate.
When I is an ideal of Rmax [x1 , .., xn ] one defines a tropical variety defined by I as follows:
\
trop(V (I)) :=
trop(V (f )).
(31)
f ∈I
One has to be careful with (31) since the intersection is over all polynomials in I not just over
a set of generators of I (cf. [30]). To understand (31) as the ‘positive’ part of an algebraic
set over hyperstructures, we extend the previous proposition as follows.
Proposition 4.18. Let I be an ideal of Rmax [x1 , .., xn ]. Then, with the same notation as
Proposition 4.14, we have a set bijection via ϕ = sn :
\
\
\
trop(V (I)) :=
trop(V (f )) ≃ (
HV (f )) Img ϕ.
f ∈I
f ∈I
T
Proof. Take any z ∈ trop(V (I)) ⊆ (Rmax )n , then by definition, z ∈ f ∈I trop(V (f )). That is
z ∈ trop(V (f T
)) ∀f ∈ I. It follows
from the previous proposition that ϕ(z) ∈ HV (f ) ∀f ∈ I,
T
thus ϕ(z) ∈ ( f ∈I HV (f )) Img ϕ.
T
T
Conversely, if ϕ(z) ∈ ( f ∈I HV (f )) Img ϕ then ϕ(z) ∈ HV (f ) ∀f ∈ I. From the previous
proposition it follows that z ∈ trop(V (f )) ∀f ∈ I, hence z ∈ trop(V (I)). Thus we have
\
\
ϕ(trop(V (I))) = (
HV (f )) Img ϕ.
f ∈I
The conclusion follows from the injectivity of ϕ.
21
We close this section by providing a pictorial image which shows how our construction works.
Let R≥0 be a semiring with a underlying set as the set of nonnegative real numbers, an
addition is given by the maximum, and the multiplication is the usual multiplication of real
numbers. Then one can observe that there is an isomorphism ϕ of semirings as follows:
r 7→ tr and t−∞ := 0, where t is any real number greater than 1.
(32)
Under the symmetrization process, we have s(R≥0 ) = T R, a Viro’s hyperfield.
ϕ : Rmax −→ R≥0 ,
Example 4.19. Let f (x, y) := (0⊙x)⊕(0⊙y)⊕0 ∈ Rmax [x, y]. Then, under the isomorphism
ϕ, this becomes f (x, y) = x + y + 1 ∈ T R[x, y]. If we follow the notation as in Example 4.16,
we obtain:
f˜x (x, y) = y + 1 − x, f˜y (x, y) = x + 1 − y, f˜1 (x, y) = x + y − 1.
First, we obtain the following picture from f (x, y) in R2 :
R
X1 = {(a, b) ∈ R2 | 0 ≤ a = b}
X2 = {(a, b) ∈ R2 | a ≤ b = 0}
R
X3 = {(a, b) ∈ R2 | b ≤ a = 0}
Figure 1. Tropical Line in R2
On the other hand, we obtain the following picture in (T R)2 :
f˜x (x, y) = y + 1 − x
f˜y (x, y) = x + 1 − y
f˜1 (x, y) = x + y − 1
TR
TR
TR
TR
TR
Figure 2. Tropical Line in (T R)2
22
TR
4.2 Construction of hyperring schemes. In this subsection, we study several notions
of algebraic geometry over hyperrings from the scheme-theoretic point of view: we use results
in §3.3 to construct an integral hyperring scheme and prove that Γ(X, OX ) ≃ R for an affine
integral hyperring scheme (X, OX ).
In classical algebraic geometry, a scheme is a pair (X, OX ) of a topological space X and the
structure sheaf OX on X. The implementation of the notion of structure sheaf is essential to
link local and global algebraic data.
Let A be a commutative ring and (X = Spec A, OX ) be an affine scheme. One of important
results in classical algebraic geometry is the following:
OX (X) ≃ A.
(33)
In other words, a commutative ring A can be understood as the ring of global regular functions on the topological space X = Spec A. When we directly generalize the construction
of the structure sheaf of a commutative ring to a hyperring, (33) no longer holds (see, Example 4.24). Furthermore, in this case, OX does not even seem to be a sheaf of hyperrings
(see, Remark 4.20). To this end, we construct the structure sheaf on the topological space
X = Spec R only when R is a hyperring without (multiplicative) zero-divisors. We follow the
classical construction.
Let R be a hyperring and X = Spec R. For an open subset U ⊆ X, we define
G
OX (U ) := {s : U →
Rp },
(34)
p∈U
where s ∈ OX (U ) are sections such that s(p) ∈ Rp which also satisfy the following property:
for each p ∈ U , there exist a neighborhood Vp ⊆ U of p and a, f ∈ R such that
a
(35)
∀q ∈ Vp , f ∈
/ q and s(q) = in Rq .
f
A restriction map OX (U ) −→ OX (V ) is given by sending s to s ◦ i, where i : V ֒→ U is
an inclusion map. Then, clearly OX is a sheaf of sets on X. Moreover, one can define the
multiplication s · t of sections s, t ∈ OX (U ) as follows:
G
s·t: U →
Rp , p 7→ s(p)t(p).
(36)
Equipped with the above multiplication, one can easily see that OX becomes a sheaf of
(multiplicative) monoids on X. Furthermore, OX (U ) is equipped with the following hyperoperation:
s + t = {r ∈ OX (U ) | r(p) ∈ s(p) + t(p), ∀p ∈ U }.
(37)
Remark 4.20. This construction is essentially same as in [36]. However, in [36], the proof is
incomplete in the sense that the authors did not prove that (37) is associative and distributive
with respect to (36). Moreover, the main purpose of this subsection is to recover a hyperring
R as the hyperring of global sections on a topological space Spec R while the main goal of
the authors of [36] was to construct hyperring schemes. In Theorem 4.23, we prove that
when R does not have (multiplicative) zero-divisors, OX is indeed the sheaf of hyperrings,
and OX (X) ≃ R.
Definition 4.21. A hyperring R is called a hyperdomain if R does not have (multiplicative)
zero-divisors. In other words, for x, y ∈ R, if xy = 0 then either x = 0 or y = 0.
Let R be a hyperdomain and S := R\{0}. Clearly K := Frac(R) = S −1 R is a hyperfield
and the canonical homomorphism S −1 : R −→ K of hyperrings sending r to 1r is strict and
injective.
23
Let Sf = {1, f 6= 0, ..., f n , ...} be a multiplicative subset of R and Rf := Sf−1 R, then we have
the canonical homomorphisms of hyperrings R ֒→ Rf ֒→ K which are injective and strict.
Therefore, in the sequel, we consider Rf as a hyperring extension of R and K as a hyperring
extension of both R and Rf via the above canonical maps. For p ∈ Spec R, we denote by Rp
the hyperring S −1 R, where S = R\p.
Lemma 4.22. Let A be a set equipped with the two binary operations:
+A : A × A −→ P ∗ (A),
P ∗ (A)
·A : A × A −→ A,
where
is the set of nonempty subsets of A. Suppose that R is a hyperring and there
exists a set bijection ϕ : A −→ R such that ϕ(a +A b) = ϕ(a) + ϕ(b) and ϕ(a ·A b) = ϕ(a)ϕ(b).
Then, A is a hyperring isomorphic to R.
Proof. The proof is straightforward. For example, ϕ−1 (0R ) := 0A is the neutral element. In
fact, for a ∈ A, we have ϕ(0A +A a) = ϕ(0A ) + ϕ(a) = 0R + ϕ(a) = ϕ(a). Since ϕ is bijective,
it follows that 0A +A a = a ∀a ∈ A. Similarly, 1A := ϕ−1 (1R ) is the identity element. For
a ∈ A, we can write −ϕ(a) = ϕ(b) for some b ∈ A. Then, we have ϕ(a +A b) = ϕ(a) + ϕ(b) =
ϕ(a) − ϕ(a). It follows that 0R ∈ ϕ(a +A b), hence 0A ∈ a +A b. The other properties can be
easily checked and hence A is a hyperring. Clearly, A and R are isomorphic via ϕ.
Theorem 4.23. Let R be a hyperdomain, K = Frac(R), and X = Spec R. Let OX be the
sheaf of multiplicative monoids on X as in (34), equipped with the hyperaddition (37). Then,
the following hold
(1) OX (D(f )) is a hyperring isomorphic to Rf . In particular, if f = 1, we have
R ≃ OX (X)(= Γ(X, OX )).
(2) For each open subset U of X, OX (U ) is a hyperring. More precisely, OX (U ) is
isomorphic to the following hyperring:
a
/ p}.
OX (U ) ≃ Y (U ) := {u ∈ K | ∀p ∈ U, u = for some b ∈
b
Moreover, by considering the canonical map Rf ֒→ K, we have
\
OX (U ) ≃
OX (D(f )).
D(f )⊆U
(3) For each p ∈ X, the stalk OX,p exists and is isomorphic to Rp .
Proof. The proof is similar to the classical case (cf. [19]). We frequently use the results in
§3.3, in particular, Hilbert’s Nullstellensatz of hyperrings (Lemma 3.19).
(1) Consider the following map:
a
a
ψ : Rf → OX (D(f )),
7→ s, where s(p) = n in Rp .
(38)
n
f
f
Clearly, ψ is well defined since the map s defined as in (38) satisfies the condition
(35). It also follows from the definition that
ψ(
a
b
a
b
·
) = ψ( n ) · ψ( m ),
fn fm
f
f
ψ(
a
b
a
b
+
) ⊆ ψ( n ) + ψ( m ).
fn fm
f
f
First, we claim that ψ is one-to-one. Indeed, suppose that ψ( fan ) = ψ( fbm ). Then,
b
a
f n = f m as elements of Rp ∀p ∈ D(f ). Hence there is an element h 6∈ p such that
hf m a = hf n b in R. This implies that 0 ∈ hf m a − hf n b = h(f m a − f n b). However,
since h 6∈ p (hence, h 6= 0) and R is a hyperdomain, it follows that f m a = f n b. This
implies that fan = fbm in Rf , thus ψ is one-to-one.
24
Next, we claim that ψ is onto. Take s ∈ OX (D(f )). Then, we can cover D(f ) with
open sets Vi so that s is represented by a quotient agii on Vi with gi 6∈ p ∀p ∈ Vi
from (35). Since open subsets of the form D(h) form a basis, we may assume that
Vi = D(hi ) for some hi ∈ R. Let (hi ) and (gi ) be the hyperideals generated by
hi and gi . Since s is represented by agii on D(hi ), we have D(hi ) ⊆ D(gi ); hence,
p
p
V ((hi )) ⊇ V ((gi )). It follows from Lemma 3.19 that (hi ) ⊆
(gi ) and hence
n
n
hi ∈ (gi ) for some n ∈ N. Then, from Lemma 3.20, we have hi = cgi for some c ∈ R.
i
Hence agii = ca
. If we replace hi by hni (since D(hi ) = D(hni )) and ai by cai , we may
hn
i
assume that D(f ) is covered by the open subsets D(hi ) on which s is represented by
ai
hi . Moreover, as in the classical case, we observe that D(f ) can be covered by finitely
many D(hi ). In fact,
[
\
D(f ) ⊆
D(hi ) ⇐⇒ V ((f )) ⊇
V ((hi )).
(39)
Let Ii = (hi ), I =< Ii >, and J = (f ). Then, (39) can be written as follows:
\
[
V (Ii ) = V (I), D(f ) ⊆
D(hi ) ⇐⇒ V (J) ⊇ V (I).
(40)
√
√
It follows from Lemma 3.19 that J ⊆ I, thus f n ∈ I for some n ∈ N. Then, from
Lemma 3.20, we have
r
X
bi hi for some bi ∈ R.
(41)
fn ∈
i=1
We claim that D(f ) can be covered by D(h1 ) ∪ ... ∪ D(hr ). Indeed, this is equivalent
to
r
\
V ((hi )) = V (< (hi ) >i=1,...,r ).
V ((f )) ⊇
i=1
Let I :=< hi >i=1,...,r . Suppose that p ∈ V (I). Since I ⊆ p, it follows from (41) that
f n ∈ p, hence f ∈ p. This implies that (f ) ⊆ p, thus p ∈ V ((f )). From now on, we
fix the elements h1 , ..., hr such that D(f ) ⊆ D(h1 ) ∪ ... ∪ D(hr ). Then, on D(hi hj ),
a
we have two elements haii , hjj of Rhi hj which can be considered as (the restriction
of) the same element s. It follows from the injectivity of ψ (on D(hi hj )) that one
a
should have haii = hjj in Rhi hj . Therefore, (hi hj )n hj ai = (hi hj )n hi aj for some n ∈ N.
However, since R is a hyperdomain, we have (hi hj )n 6= 0. It follows that hj ai = hi aj
∀i, j = 1, ...,
P r from the uniqueness of an additive inverse.
Let f n ∈ ri=1 bi hi as in (41). Then, for each j ∈ {1, ..., r}, we have
X
X
X
X
f n aj ∈ (
bi hi )aj =
bi aj hi =
bi ai hj = (
bi ai )hj .
i
i
i
i
P
It follows that for each j = 1, ..., r, there exists βj ∈ i bi ai such that f n aj = βj hj .
Hence, we have
aj
βj
=
on D(hj ).
(42)
n
f
hj
a
We claim that βi = βj ∀i, j = 1, ..., r. Indeed, on D(hi hj ), we proved that hjj = haii .
Together with (42) and the injectivity of ψ, we have
βj
aj
ai
βi
=
=
= n on D(hi hj ).
n
f
hj
hi
f
Therefore, ∃m ∈ N such that (hi hj )m f n βj = (hi hj )m f n βi . Equivalently, we have
0 ∈ (hi hj )m f n (βj − βi ). However, we know that (hi hj )m , f n 6= 0 since hi hj , f 6= 0
25
and R is a hyperdomain. It follows that 0 ∈ βi − βj , thus βi = βj from the uniqueness
of an additive inverse. Let β be this common value βi . Then, we have f n aj = βhj
a
∀j = 1, ..., r. Therefore, fβn = hjj on D(hj ). In other words, ψ( fβn ) = s. This shows
that ψ is onto. This, however, does not complete the proof. We need to show that
ψ(ab) = ψ(a)ψ(b) and ψ(a + b) = ψ(a) + ψ(b), then the result follows from Lemma
4.22. Clearly, we have ψ(ab) = ψ(a)ψ(b) and ψ( fan + fbm ) ⊆ ψ( fan ) + ψ( f bm ). We show
the following:
a
b
a
b
ψ( n ) + ψ( m ) ⊆ ψ( n + m ).
f
f
f
f
b
a
Let s = ψ( f n ), t = ψ( f m ). Then, we have
s + t = {r ∈ OX (D(f )) | r(p) ∈ s(p) + t(p) ∀p ∈ D(f )}.
For r ∈ s + t, since ψ is onto, r = ψ( fcl ) for some
s(p) + t(p) that
c
fl
∈
a
fn
+
b
fm
c
fl
∈ Rf . It follows from r(p) ∈
in Rp , and this is equivalent to the following:
d
c
= n+m for some d ∈ (af m + bf n ) in Rp .
l
f
f
Therefore, ucf n+m = udf l for some u ∈ R\p. Since u 6= 0, we have cf n+m = df l .
d
d
Equivalently, fcl = f n+m
in Rf . However, f n+m
∈ fan + fbm , therefore s + t ⊆ ψ( fan +
b
f m ).
This shows the other inclusion. The conclusion follows from Lemma 4.22.
(2) One can easily see that Y (U ) is a hyperring (in fact, a sub-hyperring of K). We show
that there exists a bijection ϕ of sets from OX (U ) to Y (U ) such that ϕ(a + b) =
ϕ(a) + ϕ(b) and ϕ(ab) = ϕ(a)ϕ(b). Then, the first assertion will follow from Lemma
4.22. Indeed, if s ∈ OSX (U ), then from the same argument (in the proof of 1), we
can find a cover U = D(hi ) such that s = haii on D(hi ). However, since R has no
(multiplicative) zero-divisor, X = Spec R is irreducible from Proposition 3.23. Thus,
a
D(hi ) ∩ D(hj ) 6= ∅ ∀i, j. It follows that haii = hjj on D(hi ) ∩ D(hj ), equivalently
0 ∈ sij (ai hj − hi aj ) for some sij 6= 0 ∈ R. Since sij 6= 0 and R is a hyperdomain, it
follows that
ai
aj
0 ∈ (ai hj − hi aj ) ⇐⇒ ai hj = aj hi ⇐⇒
=
as elements of K = Frac(R).
hi
hj
Let u = ab be this common value in K. Then, for each p ∈ U , we have p ∈ D(hi ) for
some hi 6∈ p and ab = haii on D(hi ). It follows that u ∈ Y (U ). Since OX is a sheaf,
u ∈ Y (U ) is uniquely determined by s. We let ϕ(s) := u. Then, we have
a
ϕ : OX (U ) −→ Y (U ) := {u ∈ K | ∀p ∈ U, u = , b 6∈ p} ⊆ K.
b
ϕ is well defined and one-to-one since OX is a sheaf (of sets). We claim that ϕ is
F
′
onto. In fact, for u = ab ∈ Y (U ), we define s : U −→ p∈U Rp such that s(p) = ab = ab′
for b′ 6∈ p from the definition of Y (U ). Then s ∈ OX (U ). Next, it follows from the
definition that ϕ(s · t) = ϕ(s) · ϕ(t). Furthermore, we have ϕ(s + t) ⊆ ϕ(s) + ϕ(t).
Indeed, we have
α ∈ s + t ⇐⇒ α(p) ∈ s(p) + t(p)
∀p ∈ U.
(43)
However, since ϕ is bijective, each section is globally represented by an element of
K. Suppose that α, s, t are globally represented by fg , ha , mb respectively. Then, (43)
is equivalent to the following:
a
b
g
⇐⇒ ϕ(α) ∈ ϕ(s) + ϕ(t).
α(p) ∈ s(p) + t(p) ⇐⇒ ∈ +
f
h m
26
b
Conversely, for fg ∈ ϕ(s) + ϕ(t) = ha + m
, we have α ∈ OX (U ) such that α(p) = fg at
Rp . It follows that α ∈ s + t, and ϕ(s) + ϕ(t) ⊆ ϕ(s + t). This shows that OX (U ) is
isomorphic to Y (U ) via ϕ from Lemma 4.22.
Next, we prove the second assertion.
For D(f ) ⊆ U , we have Y (U ) ⊆ Y (D(f )) ⊆ K.
T
This implies that Y (U ) ⊆ D(f )⊆U Y (D(f )). Conversely, suppose that u = ab ∈
T
D(f )⊆U Y (D(f )). Then, for each p ∈ U , we have p ∈ D(f ) for some D(f ) ⊆ U .
Since u ∈ Y (D(f )), u can be written as xy so that y 6∈ p. It follows that u ∈ Y (U ),
T
and Y (U ) = D(f )⊆U Y (D(f )). One observes that Y (D(f )) = Rf ⊆ K. Thus, under
T
OX (D(f )) ≃ Rf , we have OX (U ) ≃ D(f )⊆U OX (D(f )).
(3) In general, direct limits do not exist in the category of hyperrings. Thus, one should
be careful. Since open sets of the form D(f ) form a basis of X, it is enough to show
that
lim OX (D(f )) = Rp .
−→
(44)
D(f )∋p
For each f ∈ R, let ψf : OX (D(f )) −→ Rp , s 7→ s(p). Then, we have the following
commutative diagram:
OX (D(f ))
▼▼▼
▼▼▼
▼▼
ψf ▼▼▼
▼&
ρ
Rp
/ OX (D(g)) ,
qq
qqq
q
q
qq ψg
x qq
q
where ρ is a restriction map of the structure sheaf OX . Let H be a hyperring and
suppose that we have another commutative diagram:
ρ
/ OX (D(g)) .
❁❁▼▼▼
ψ
ψg qqqq ✂✂
❁❁ ▼▼▼ f
qq ✂✂
❁❁ ▼▼▼
q
▼▼▼
q q ✂✂✂
❁❁
q
q
&
xq
❁❁
✂
Rp
✂✂ϕg
ϕf ❁❁
✂
❁❁
✂✂
❁❁
✂✂
❁❁
✂
✂✂
OX (D(f ))
H
By considering Rp as a sub-hyperring of K, let us define the map ψ as follows:
ψ : Rp −→ H,
b
b
7→ ϕt ( ),
t
t
where bt is considered as an element of OX (D(t)) such that bt (q) = bt in Rq for each
q ∈ D(t). If ψ is a well-defined homomorphism of hyperrings, the uniqueness of such
map easily follows. Indeed, suppose that ϕ : Rp −→ H is the homomorphism of
hyperrings such that ϕf = ϕ ◦ ψf ∀f ∈ R for p ∈ D(f ). A section s of OX (D(f )) is
represented by fbn (from the first part of the proposition). Therefore, ψf (s) = ψf ( fbn ),
and ϕf (s) = ϕ ◦ ψf (s) = ϕ ◦ ψf ( fbn ) = ϕ( fbn ). However, we have ϕf (s) = ψ ◦ ψf (s) =
ψ( fbn ). Thus, such ψ is unique if it exists.
′
Next, we show that ψ is well defined. Indeed, if bt = bt′ , then we have bt′ = b′ t since R
is a hyperdomain. It follows that D(bt′ ) = D(b′ t) ⊆ D(t), and we have the following
27
commutative diagram:
ρ
OX (D(t))
▼▼▼
▼▼▼
ϕt ▼▼▼▼
▼&
H
/ OX (D(b′ t)) = OX (D(bt′ )) .
❦❦❦
❦❦❦❦
❦
❦
❦
❦❦❦ ϕb′ t
u ❦❦❦
❦
′
From the similar argument with t′ , we have ϕt ( bt ) = ϕt′ ( bt′ ). This shows that ψ
does not depend on the choice of t, hence ψ is well defined. For fa , gb ∈ Rp , by
considering ϕf g , we have ψ( fa ) = ϕf g ( fa ), ψ( gb ) = ϕf g ( gb ), and ψ( fabg ) = ϕf g ( fabg ).
Thus, ψ( fa gb ) = ψ( fa )ψ( gb ). Similarly, we have ψ( fa + gb ) ⊆ ψ( fa ) + ψ( gb ). Hence, ψ is
a homomorphism of hyperrings. Finally, since {D(f )} is a basis of X, we have
lim OX (U ) = lim OX (D(f )).
−→
−→
U ∋p
D(f )∋p
Therefore, we conclude that OX,p ≃ Rp .
When R is a hyperdomain, we call the pair (X = Spec R, OX ) as in Theorem 4.23 an integral
affine hyperring scheme. The following example shows that if R has zero divisors, then in
general R 6= Γ(X, OX ).
Example 4.24. Consider the following quotient hyperring R:
R = Q ⊕ Q/G, where G = {(1, 1), (−1, −1)}.
Then, Spec R = {p1 , p2 } with p1 =< [(1, 0)] > and p2 =< [(0, 1)] >. Each pj becomes one
point open and closed subset of X and the intersection p1 ∩ p2 is empty. Furthermore, one
can easily check that Rpi ≃ Q/H, where H = {1, −1}. Therefore,
Γ(X, OX ) ≃ (Q/H) ⊕ (Q/H) 6= R.
Remark 4.25.
(1) One can construct other examples of affine hyperring schemes X =
Spec R for which R 6= Γ(X, OX ), but all such examples are disconnected. We do
not have yet any example of a connected topological space X = Spec R with R 6=
Γ(X, OX ). On the other hand, being connected is not a necessary condition. In fact,
let A = Z/12Z and G = {1, 5} ⊆ (Z/12Z)× . Then, with the quotient hyperring
R = A/G, the space X = Spec R is disconnected (consist of two points), however,
one can easily check that R ≃ Γ(X, OX ).
(2) What seems interesting is that by enhancing a multiplicative condition (hyperdomain)
we could fix an additive weakness (multi-valued addition). In fact, similar observation
has been made by O.Lorscheid in [28] stating that in the classical construction of
schemes, a multiplicative structure plays a major role while an additive structures is
less important.
Let R be a hyperring, X = Spec R, and OX be the structure sheaf (of sets) of X. Then, as we
previously mentioned in Remark 4.20, Γ(X, OX ) does not have to be a hyperring. Moreover,
even if Γ(X, OX ) is a hyperring, Example 4.24 shows that the natural map R −→ Γ(X, OX )
is not even an injective map in general. By appealing to the classical construction of Cartier
divisors, we define the presheaf FX of hyperrings on X = Spec R which slightly generalizes
OX (cf. Remark 4.27 and Proposition 4.29).
Let S := {α ∈ R | α is not a zero-divisor}. In other words, S is the set of regular elements of
R. Then, S 6= ∅ since 1 ∈ S. Furthermore, S is a multiplicative subset of R, therefore one can
define K := S −1 R. In what follows, we denote by R a hyperring, and S, K as above. Note
28
that by a sub-hyperring R of a hyperring L we mean a subset R of L which is a hyperring
with the induced operations.
Lemma 4.26. Let S be the set as above, and f ∈ S. Let ϕ : R −→ Rf , ϕ(a) = a1 be the
natural map of localization and ψ : Rf −→ K := S −1 R be a homomorphism of hyperrings
such that ψ( fan ) = fan . Then ϕ and ψ are strict, injective homomorphisms of hyperrings.
Proof. We only prove the case of ϕ since the proof for ψ is similar. If a1 = 1b then f n a = f n b
for some n ∈ N. This implies that 0 ∈ f n (a − b). Hence, we have c ∈ (a − b) such that
0 = f n c. Since f n ∈ S can not be a zero-divisor, we have c = 0, therefore a = b. This shows
that ϕ is injective. Furthermore, if γ ∈ ϕ(a) + ϕ(b) = a1 + 1b , then γ = 1t for some t ∈ a + b.
Therefore γ = ϕ(t), and ϕ is strict.
For each open subset U of X = Spec R, we define the following set:
a
(45)
FX (U ) := {u ∈ K | ∀p ∈ U, u = , b ∈ S ∩ pc }.
b
In other words, u ∈ K is an element of FX (U ) if u has a representative ab such that b 6∈ p for
each p ∈ U . The restriction map is given by the natural injection. i.e. if V ⊆ U , then we
have FX (U ) ֒→ FX (V ). Then, one can easily observe that FX (U ) is a hyperring. Thus, FX
becomes a presheaf of hyperrings on X = Spec R.
Remark 4.27. It follows from Theorem 4.23 that when R is a hyperdomain, we have
OX (U ) ≃ FX (U ) for each open subset U of X. Therefore, in this case, FX is indeed a
sheaf of hyperring and FX (X) = R.
Proposition 4.28. Let R be a hyperring. If X = Spec R is irreducible, then FX is a sheaf
of hyperrings.
Proof. Since
S FX (U ) is clearly a hyperring, we only have to prove that FX is a sheaf. Suppose
that U = Vi is an open covering of U . Firstly, if s ∈ FX (U ) is an element such that s|Vi = 0
for all i, then we have to show that s = 0. However, this is clear since the restriction map
is injective. Secondly, let si ∈ FX (Vi ) such that si |Vi ∩Vj = sj |Vi ∩Vj for all i, j. Since X is
irreducible, it follows that Vi ∩ Vj 6= ∅ ∀i, j. Moreover, the condition si |Vi ∩Vj = sj |Vi ∩Vj means
that si = sj as elements of K. Let s be this common element of K. Then, {si } can be glued
to s. Clearly, s is an element of FX (U ).
The following proposition shows that FX behaves more nicely than OX in some cases.
Proposition 4.29. Let R be a hyperring and assume that X = Spec R is irreducible. Then,
for f ∈ S, there exists a canonical injective and strict homomorphism ϕ : Rf −→ FX (D(f )).
In particular, R is a sub-hyperring of FX (X). Furthermore, if R has a unique maximal
hyperideal, then R ≃ FX (X).
Proof. From Lemma 4.26, there exists a canonical injective and strict homomorphism ψ :
Rf −→ K. From the definition of FX (D(f )), one sees that the image of ψ lies in FX (D(f )) ⊆
K. Therefore, ψ becomes our desired ϕ. When R has a unique maximal hyperideal, we have
to show that any element u of FX (X) is of the form a1 for some a ∈ R. Suppose that m is
the maximal ideal of R. Then, u ∈ FX (X) means that u = ab for some b ∈ S − m. Since m is
the only maximal hyperideal of R, it follows from Lemma 3.20 that 1 ∈ a + b for some a ∈ m.
−1
Therefore, b ∈ 1−a and b is a unit by Lemma 3.21. Thus, u = ab = ab1 and R ≃ FX (X).
Next, we prove that the opposite category of hyperdomains and the category of integral
affine hyperring schemes are equivalent via the contravariant functors, Spec and Γ. Note
that one can directly generalize the notion of a ringed space to define a hyperringed space.
However, the notion of a locally hyperringed space should be treated with greater care since
29
the category of hyperrings does not have (co)limits in general. Nevertheless, an integral
affine hyperring scheme (X, OX ) can be considered as a locally hyperringed space thanks to
Theorem 4.23. Thus, in what follows we consider (X, OX ) as a locally hyperringed space in
the sense of the direct generalization of the classical notion. We will simply write X instead
of (X, OX ) if there is no possible confusion. The following lemma has been proven in [12]
and [36], and will be mainly used.
Lemma 4.30. ( [12, Theorem 3.6], [36, Proposition 8]) Let ϕ : R −→ H be a homomorphism
of hyperrings. Then, for p ∈ Spec H, ϕ induces the following homomorphism ϕp of hyperrings:
ϕp : Rq −→ Hp ,
ϕ(a)
a
7→
,
b
ϕ(b)
where q = ϕ−1 (p)
such that if mp , mq are unique maximal hyperideals of Hp and Rq respectively, then ϕ−1
p (mp ) =
mq .
Proposition 4.31. Let R and H be hyperdomains, and X = Spec R, Y = Spec H. Then,
we have
Hom(R, H) = Hom(Y, X),
(46)
where Hom(R, H) is the set of homomorphisms of hyperrings and Hom(Y, X) is the set of
morphisms of locally hyperringed spaces.
Proof. Clearly, a homomorphism ϕ : R −→ H of hyperdomains induces the continuous map
f : Y = Spec H −→ X = Spec R,
p 7→ ϕ−1 (p).
Then, f induces the morphism of sheaves: f # : OX −→ f∗ OY . Indeed,
F for an open subset
−1
−1
V ⊆ X, we have f∗ OY (V ) := OY (f (V )) = {t | t : f (V ) −→ q∈f −1 (V ) Hq }, where t
satisfies the local condition (35). First, we define
G
G
G
ψV :=
ϕp :
Rϕ−1 (p) −→
Hp ,
(47)
p∈V
p∈f −1 (V )
p∈f −1 (V )
where ϕp is the map induced from ϕ at p as in Lemma 4.30. We also define
f # (V ) : OX (V ) −→ OY (f −1 (V )),
s 7→ t := ψV ◦ s ◦ f.
(48)
We need to check four things. Firstly, we have to show that t as in (48) is an element of
OY (f −1 (V )). Secondly, we have to show that f # is compatible with an inclusion V ֒→ U
of open sets of X; this is clear from the construction. Thirdly, we have to show that f # (V )
is a homomorphism of hyperrings. The proof of these three is similar to the classical case.
Finally, we have to show that f # (V ) is local. It easily follows from the third statement of
Theorem 4.23 and Lemma 4.30.
Conversely, suppose that a morphism
(g, g# ) : Y = (Spec H, OY ) −→ X = (Spec R, OX )
of integral affine hyperring schemes is given. Since R and H are hyperdomains, we can recover
a homomorphism of hyperrings ϕ : R −→ H by taking global sections thanks to Theorem
4.23. Therefore, all we have to prove is that the map (f, f # ) induced from ϕ as in (48) is
same as (g, g# ). The proof is similar to the classical case.
30
4.3 From hyperring schemes to the classical schemes. Next, we provide an example
showing that an integral hyperring scheme can be linked to the classical theory. Let A
be an integral domain containing the field Q of rational numbers, X = Spec A, and Y =
Spec(A/Q× ) = Spec(A ⊗Z K) (see Proposition 2.14 for the definition of A/Q× or [6] for the
scalar extension functor − ⊗Z K). We prove that there exists a canonical homeomorphism
ϕ : Y −→ X such that OY (ϕ−1 (U )) ≃ OX (U ) ⊗Z K for an open subset U ⊆ X. Indeed, such
homeomorphism is very predictable from the following observation: let B an integral domain
containing the field Q of rational numbers. Then, a polynomial f ∈ B[X1 , ..., Xn ] vanishes if
and only if qf vanishes ∀q ∈ Q× .
Lemma 4.32. Let A be an integral domain containing the field Q of rational numbers. Let
A ∋ f 6= 0 and f˜ be the image of f under the canonical projection map π : A −→ R := A/Q× .
Then, we have
Af /Q× ≃ Rf˜,
where R ˜ is the localization of R at f˜.
f
Proof. Since A is an integral domain, Af contains A (hence, contains Q). Thus Af /Q× is
well defined. Let us define the following map:
[a]
a
,
ψ : Af /Q× −→ Rf˜, [ n ] 7→
f
f˜n
where [ fan ] is the equivalence class of fan ∈ Af in Af /Q× and [a] is the equivalence class of
a ∈ A in R = A/Q× . One can easily show that ψ is a well-defined and strict homomorphism of
hyperrings which is also bijective. It follows from the first isomorphism theorem of hyperrings
(cf. [12, Proposition 2.11]) that ψ is an isomorphism of hyperrings.
Lemma 4.33. Let A be an integral domain containing the field Q of rational numbers and
R = A/Q× . Then, we have
Frac(A)/Q× ≃ Frac(R) = Frac(A/Q× ).
Proof. The proof is similar to Lemma 4.32. For the notational convenience, let us define the
following map:
b
[b]
ψ : Frac(A)/Q× −→ Frac(R), [ ] 7→
.
a
[a]
Again, we have to show that this is well defined, bijective, and a strict homomorphism of
hyperrings. The proof is almost identical to that of Lemma 4.32.
Lemma 4.34. Let A be a commutative ring, G ⊆ A× be a multiplicative subgroup, and A/G
be the quotient hyperring. Then, X = Spec A and Y = Spec(A/G) are homeomorphic (under
the Zariski topology).
Proof. If G = {1} then there is nothing to prove. Thus, we may assume that |G| ≥ 2. We
define the following map:
∼: X −→ Y,
q 7→ q̃ := {αG | α ∈ q}.
We claim that the map ∼ is well defined. Indeed, we have
αG = βG ⇐⇒ α = βu, u ∈ G ⊆ A× ⇐⇒ α, β ∈ q or α, β 6∈ q,
therefore q̃ is uniquely determined by q. Furthermore, q̃ is a hyperideal. In fact, we have 0G ∈
q̃. If aG ∈ q̃ then (−a)G = −(aG) ∈ q̃. For rG ∈ A/G and aG ∈ q̃, since (rG)(aG) = raG
and ra ∈ q, it follows that (rG)(aG) ∈ q̃. Suppose that aG, bG ∈ q̃. One can observe that
aG, bG ∈ q̃ ⇐⇒ a, b ∈ q since G ⊆ A× and q is a prime ideal. Therefore, for zG ∈ q̃, we may
31
assume that z = at + bh for some t, h ∈ G. It follows that z ∈ q, hence zG ∈ q̃. This shows
that q̃ is a hyperideal. Next, we show that q̃ is prime. Suppose that (aG)(bG) = (abG) ∈ q̃
and aG 6∈ q̃. This implies that ab ∈ q and au 6∈ q ∀u ∈ G, hence a 6∈ q. Since q is prime, this
implies that b ∈ q, and bG ∈ q̃.
Next, we claim that the map ∼ is continuous. Let ϕ :=∼ for the notational convenience. It
is enough to show that ϕ−1 (D(f G)) is open. We have the following:
ϕ−1 (D(f G)) = D(f ).
(49)
Indeed, if q ∈ D(f ), then ϕ(q) = q̃ can not contain f G by definition. Hence, D(f ) ⊆
ϕ−1 (D(f G)). Conversely, suppose that p ∈ ϕ−1 (D(f G)), then ϕ(p) ∈ D(f G). Since ϕ(p) =
p̃ = {αG | α ∈ p} and f 6∈ p, it follows that p ∈ D(f ), hence ϕ−1 (D(f G)) ⊆ D(f ). This
proves (49), hence ∼ is continuous.
Finally, we construct the inverse of the map ϕ =∼. The canonical projection map π : A −→
A/G induces the following canonical map:
p 7→ π −1 (p).
ψ : Y −→ X,
Clearly, ψ is continuous since ψ −1 (D(f )) = D(f G). We claim that ϕ and ψ are inverses to
each other. Since both ϕ and ψ are continuous, it is enough to show that ϕ is bijective and
ϕ ◦ ψ = idY . First, we show that ϕ is injective. Assume that ϕ(q) = ϕ(p) for p, q ∈ X. Then,
for x ∈ q, we have y ∈ p such that xG = yG. It follows that x = yg for some g ∈ G, hence
x ∈ p. Since the argument is symmetric, we have p = q.
For the surjectivity of ϕ, take an element ℘ ∈ Spec A/G. We consider αG as the subset
αG := {αg | g ∈ G} ⊆ A and define
[
αG.
p :=
αG∈℘
We have to show that p is a prime ideal of A. We have 0 ∈ p. Moreover, a ∈ p ⇐⇒ a ∈ αG
for some αG ∈ ℘. It follows that −αG ∈ ℘ and hence −a ∈ p. Furthermore, for a ∈ p
and r ∈ A, we have aG ∈ ℘ and rG ∈ A/G. It follows from (rG)(aG) = (raG) ∈ ℘ that
ra ∈ p. If a, b ∈ p then aG, bG ∈ ℘. This implies that aG + bG ⊆ ℘ and hence a + b ∈ p.
This proves that p is an ideal. We observe that p can not be A since that implies 1 ∈ p and
1G ∈ ℘, but ℘ 6= A/G. One further observes that p is prime since for ab ∈ p and a 6∈ p, we
have (aG)(bG) ∈ ℘ and aG 6∈ ℘. This implies that bG ∈ ℘, hence b ∈ p. Obviously, we have
ϕ(p) = ℘. This shows that ϕ is surjective. In fact, one can see that p = ψ(℘). Thus, we have
ϕ(p) = ϕ ◦ ψ(℘) = ℘ and therefore ϕ ◦ ψ = idY . This completes our proof.
Proposition 4.35. Let A be an integral domain containing the field Q of rational numbers.
Let R := A/Q× , X = (Spec A, OX ), Y = (Spec R, OY ), and π : A −→ A/Q× be the canonical
projection map. Then, the following hold.
(1)
ϕ : Spec R −→ Spec A, p 7→ π −1 (p)
is a homeomorphism.
(2) For an open subset U ⊆ X, we have
OY (ϕ−1 (U )) ≃ OX (U )/Q× .
Proof. The first assertion is proved in Lemma 4.34. Note that the map induced from the
canonical projection π : A −→ A/Q× is the desired homeomorphism ϕ.
For the second claim, we use the following classical identification:
\
OX (U ) =
Af ⊆ K := Frac(A).
(50)
D(f )⊆U
32
Each OX (U ) is an integral domain containing Q, hence OX (U )/Q× is well defined. From
(50), we may assume that
OX (U )/Q× and Af /Q× are subsets of K/Q× .
(51)
Also, from Lemma 4.32 and 4.33, we have
Af /Q× ≃ Rf˜,
K/Q× ≃ L := Frac(R).
(52)
It follows from (51) and (52) that
OX (U )/Q× = (
\
Af )/Q× =
D(f )⊆U
\
(Af /Q× ).
(53)
D(f )⊆U
In fact, the first equality simply follows from (50). It remains toTshow the second equality.
Indeed, we know that [a] ∈ OX (U )/Q× ⇐⇒ qa ∈ OX (U ) = D(f )⊆U Af ⇐⇒ qa ∈ Af
∀f ∈ A such that D(f ) ⊆ U for some q ∈ Q× . This implies that [a] ∈ Af /Q× ∀f ∈ A such
that D(f ) ⊆ U . It follows that
\
\
OX (U )/Q× = (
Af )/Q× ⊆
(Af /Q× ).
D(f )⊆U
D(f )⊆U
Conversely, if [a] ∈ D(f )⊆U (Af /Q× ) then [a] ∈ Af /Q× ∀f ∈ A such that D(f ) ⊆ U . In
other words, for each fT
, there exists qf ∈ Q× such that aqf ∈ Af . However, Q ⊆ Af ∀f ∈ A,
hence a ∈ Af and a ∈ D(f )⊆U Af . It follows that
\
[a] ∈ (
Af )/Q× = OX (U )/Q× ,
T
D(f )⊆U
and this shows (53).
Finally, let f˜ = π(f ) as in Lemma 4.32. Then, there is a one-to-one correspondence between
the following sets (cf. the proof of Lemma 4.34):
A := {f ∈ A | D(f ) ⊆ U },
B := {f˜ ∈ R | D(f˜) ⊆ ϕ−1 (U )},
(54)
where ϕ is the canonical homeomorphism in the first assertion. Therefore, together with
Theorem 4.23, we have
\
\
\
OX (U )/Q× =
(Af /Q× ) ≃
Rf˜ =
Rf˜ ≃ OY (ϕ−1 (U )).
D(f )⊆U
D(f )⊆U
D(f˜)⊆ϕ−1 (U )
This proves the second assertion.
Remark 4.36. Proposition 4.35 states that if A is an integral domain containing Q, then
X := Spec A ≃ XK = Spec AK = Spec(A/Q× ). In other words, the spaces are homeomorphic, but their functions(sections) are different. In fact, what the second assertion states is
that sections of XK can be derived from sections of X by tensoring them with K in the sense
of [6].
Appendix A.
A.1
From semistructures to hyperstructures
Basic definitions of semiring theory.
Definition A.1. A semigroup is a set M with a binary operation ∗ such that
a ∗ (b ∗ c) = (a ∗ b) ∗ c
∀a, b, c ∈ M.
If a ∗ b = b ∗ a ∀a, b ∈ M , we say that M is a commutative semigroup. If there exists a unique
element 1M ∈ M such that m ∗ 1M = 1M ∗ m = m ∀m ∈ M , then M is called a monoid.
33
Definition A.2. A semiring is a set S with two binary operations + and ∗ such that (S, +)
is a commutative monoid with the identity element 0S and (S, ∗) is a monoid with the identity
element 1S such that
(1) a ∗ (b + c) = a ∗ b + a ∗ c, (b + c) ∗ a = b ∗ a + c ∗ a, ∀a, b, c ∈ S.
(2) 0 ∗ a = 0 ∀a ∈ S.
If (S, ∗) is a commutative monoid then S is called a commutative semiring. If (S\{0}, ∗) is
an abelian group then S is called a semifield.
In what follows we always assume that all semirings and monoids are commutative.
Definition A.3.
(1) By an idempotent monoid we mean a monoid M such that a+a = a
∀a ∈ M . An (additively) idempotent semiring is a semiring (S, +, ∗) such that (S, +)
is an idempotent monoid.
(2) An idempotent monoid (or a semiring) M is called selective if x + y ∈ {x, y} ∀x, y ∈
M.
Example A.4. Let B := {0, 1} be a two point set. We impose the commutative multiplication
as 1 ∗ 0 = 0 and 1 ∗ 1 = 1. The commutative addition is given by 1 + 1 = 1 and 1 + 0 = 1. B
is the smallest idempotent semifield called the boolean semifield.
Let M be a monoid with the identity element 0. One can define the following canonical
partial order on M :
x ≤ y ⇐⇒ x + y = y.
(55)
By a partial order on M we mean a binary relation on M which is reflexive, transitive, and
antisymmetric. A partial order is said to be total if for any x, y ∈ M , we have x ≤ y or y ≤ x.
We claim that when M is selective, such order is total. In fact, we know that x + y = x or
x + y = y ∀x, y ∈ M , hence x ≤ y or y ≤ x. For an introduction to theory of semirings we
refer the readers to [18].
A.2 The symmetrization process. In his paper [20], S.Henry introduced the symmetrization process which generalizes in a suitable way the construction of the Grothendieck
group completion of a multiplicative monoid. This process allows one to encode the structure
of a B-semimodule as the ‘positive’ part of a hypergroup interpreted as a ‘S-hypermodule’.
We briefly recall this symmetrization process. Suppose that M is a totally ordered idempotent
monoid. We introduce the following notation
s(M ) := {(s, 1), (s, −1), 0 = (0, 1) = (0, −1) | s ∈ M \{0}}.
To minimize the notation we denote (s, 1) := s, (s, −1) := −s, and |(s, 1))| = |(s, −1)| = s.
For any X = (x, p) ∈ s(M ), we define sign(X) = p. Then s(M ) is a hypergroup with the
addition given by
if |x| > |y| or x = y
x
y
if |x| < |y| or x = y
x+y =
(56)
[−x, x]= {(t, ±1) | t ≤ |x|)} if y = −x
We denote with s : M −→ s(M ), s 7→ (s, 1) the associated map.
Let H be a hypergroup and M be a monoid. We say that a map f : M −→ H is additive if
f (0) = 0 and f (a + b) ∈ f (a) + f (b) ⊆ H
∀a, b, ∈ M.
Henry proved that the construction of s(M ) determines the minimal hypergroup associated
to a commutative monoid M as the following universal property states.
(Universal Property): Let M be a monoid such that the canonical order as in (55) is
total. Then, for any hypergroup K and an additive map g : M −→ K, there exists a unique
34
homomorphism h : s(M ) −→ K of hypergroups such that g = h ◦ s.
In fact, such h : s(M ) −→ K is given by h(X) = sign(X)g(x), ∀X = (x, p) ∈ s(M ) (cf. [20,
Theorem 5.1]).
Remark A.5.
(1) Let M be a monoid such that the canonical order as in (55) is total.
Assume also that M is equipped with a smallest element. Then M can be upgraded to
a semiring by defining the addition law as the maximum (with respect to the canonical
order) and the multiplication as the usual addition. For example, Rmax is the semifield
obtained from the (multiplicative) monoid (R ∪ {−∞}, +).
(2) The symmetrization process can be applied to a general class of monoids (cf. [20,
Theorem 5.1]). In fact, for a monoid M , s(M ) is a hypergroup if and only if M
satisfies the following condition: for all x, y, z, w ∈ M ,
x + b = z; b + w = y,
(57)
x + y = z + w =⇒ ∃b ∈ M s.t.
or x = z + b; w = b + y
Henry also proved (cf. [20, Proposition 6.2]) that when M is an idempotent monoid,
M fulfills the condition (57) if and only if the canonical order of M as in (55) is
total.
(3) As Connes and Consani pointed out in [7], the symmetrization process can be understood in terms of the functor ‘extension of scalars’.
If B = (B, +, ·) is a semiring such that the symmetrization process can be applied to the
additive monoid (B, +), the multiplicative structure of B induces the corresponding multiplicative structure on s(B) in the component-wise way. In other words, one can define the
multiplication law on s(B) such that:
(x, p) · (y, q) := (xy, pq),
p, q ∈ {−1, 1};
1 · 1 = (−1) · (−1) = 1,
1 · (−1) = −1.
Remark A.6. Let M be a semiring allowing for the symmetrization process. We will prove
that under the component-wise multiplication, s(M ) is not a hyperring but only a multiring
(cf. [32]). A multiring is a weaker version of a hyperring in the sense that a hyperring fulfills
the distributive law x(y + z) = xy + xz whereas the notion of a multiring only assumes the
weak distributive property
x(y + z) ⊆ xy + xz.
For example, let M be the semiring whose underlying set is Z≥0 with the addition given by
x + y := max{x, y}, and the multiplication given by the usual multiplication. Then s(M ) does
not satisfy the distributive law. For example, 2(3 − 3) 6= 6 − 6 = [−6, 6]. Indeed, we have
5 ∈ 6 − 6 = [−6, 6], but 5 can not be an element of 2(3 − 3) = 2 · [−3, 3] because 2 can not
divide 5 in usual sense. For s(M ) to satisfy the distributive law it seems necessary to add
a suitable divisibility condition on the multiplication of s(M ). A particular case is studied
in [23].
Lemma A.7. Let B be a selective semiring. Then the symmetrization BS := s(B) is a
multiring with the component-wise multiplication. In particular, if B is a totally ordered
(with the canonical order) semifield, then BS is a hyperfield.
Proof. We first note that x ≤ y implies xz ≤ yz for all z ∈ B. In fact, it follows from
x ≤ y ⇐⇒ x + y = y that xz + yz = yz ⇐⇒ xz ≤ yz. For the first assertion, all we have to
show is that X(Y + Z) ⊆ XY + XZ for all X, Y, Z ∈ BS . If X = 0, then there is nothing
to prove. Therefore we may assume that X 6= 0. Let Y = (y, p), Z = (z, q), X = (x, r).
When #(Y + Z) = 1, it follows from (56) that there are three possible cases. The first case
is when Y = Z. In this case, we have X(Y + Z) = XY = XY + XY = XY + XZ. The
second case is when p = q, but y 6= z. Since B is selective, we may further assume that
35
y > z. Therefore, we have xy ≥ xz and X(Y + Z) = XY ∈ XY + XZ. The final case is
when p 6= q and y 6= z. But, in this case, the similar argument as the second case shows that
X(Y + Z) ⊆ XY + XZ. When #(Y + Z) 6= 1, from (56), we may assume that Y = (y, 1),
Z = (y, −1), and X = (x, r). Take any T = (t, p) ∈ (Y + Z). It follows from (56) that t ≤ y,
hence xt ≤ xy. Therefore we have XT = (xt, pr) ∈ XY + XZ. When B is a semifield, each
non-zero element of BS has a multiplicative inverse. Therefore BS is a multifield and it is
well-known that any multifield is a hyperfield (and vice versa).
References
[1] Erwan Brugallé and Kristin Shaw. A bit of tropical geometry. The American Mathematical Monthly,
121(7):563–589, 2014.
[2] Howard Campaigne. Partition hypergroups. American Journal of Mathematics, pages 599–612, 1940.
[3] Alain Connes and Caterina Consani. On the notion of geometry over F1 . Journal of Algebraic Geometry
20 n.3 (2011), 525-557; S1056-3911(2010)00535-8, 2008.
[4] Alain Connes and Caterina Consani. From monoids to hyperstructures: in search of an absolute arithmetic. Casimir Force, Casimir Operators and the Riemann Hypothesis, de Gruyter, pages 147–198, 2010.
[5] Alain Connes and Caterina Consani. Schemes over F1 and zeta functions. Compos. Math, 146(6):1383–
1415, 2010.
[6] Alain Connes and Caterina Consani. The hyperring of adele classes. Journal of Number Theory,
131(2):159–194, 2011.
[7] Alain Connes and Caterina Consani. Projective geometry in characteristic one and the epicyclic category.
arXiv preprint arXiv:1309.0406, to appear on Nagoya Mathematical Journal, 2013.
[8] Alain Connes and Caterina Consani. The universal thickening of the field of real numbers. arXiv preprint
arXiv:1202.4377, to appear on Advances in the Theory of Numbers Thirteenth Conference of the Canadian
Number Theory Association, 2014.
[9] Alain Connes and Caterina Consani. Absolute algebra and Segals Γ-rings: au dessou de Spec(Z). arXiv
preprint arXiv:1502.05585, to appear in Journal of Number Theory, 2015.
[10] Piergiulio Corsini and Violeta Leoreanu-Fotea. Applications of hyperstructure theory, volume 5. Springer,
2003.
[11] Bijan Davvaz and Violeta Leoreanu-Fotea. Hyperring theory and applications, 2007.
[12] Bijan Davvaz and A Salasi. A realization of hyperrings. Communications in Algebra R , 34(12):4389–4400,
2006.
[13] Anton Deitmar. Schemes over F1 . In Number fields and function fieldstwo parallel worlds, pages 87–100.
Springer, 2005.
[14] Anton Deitmar. F1 -schemes and toric varieties. Contributions to Algebra and Geometry, 49(2):517–525,
2008.
[15] Pierre Deligne. Les corps locaux de caractéristique p, limites de corps locaux de caract éristique 0.
Représentations des groupes réductifs sur un corps loca, Travaux en cours, Hermann, Paris, pages 119–
157, 1984.
[16] Jeffrey Giansiracusa and Noah Giansiracusa. Equations of tropical varieties. arXiv preprint
arXiv:1308.0042, 2013.
[17] Pawel Gladki and Murray Marshall. Orderings and signatures of higher level on multirings and hyperfields.
Journal of K-theory: K-theory and its Applications to Algebra, Geometry, and Topology, 10(03):489–518,
2012.
[18] Jonathan Golan. Semirings and their applications, updated and expanded version of the theory of semirings, with applications to mathematics and theoretical computer science, 1999.
[19] Robin Hartshorne. Algebraic geometry. Number 52. Springer, 1977.
[20] Simon Henry. Symmetrization of monoids as hypergroups. arXiv preprint arXiv:1309.1963, 2013.
[21] Zur Izhakian, Manfred Knebusch, and Louis Rowen. Layered tropical mathematics. Journal of Algebra,
416:200–273, 2014.
[22] Zur Izhakian and Louis Rowen. Supertropical algebra. Advances in Mathematics, 225(4):2222–2286, 2010.
[23] Jaiung Jun. Algebraic geometry over semi-structures and hyper-structures of characteristic one. PhD
thesis, Johns Hopkins University, Baltimore, Maryland, USA, May 2015.
[24] Jaiung Jun. Hyperstructures of affine algebraic group schemes. arXiv preprint arXiv:1510.02979, 2015.
[25] Jaiung Jun. Valuations of semirings. arXiv preprint arXiv:1503.01392, 2015.
[26] Marc Krasner. Approximation des corps valués complets de caractéristique p 6= 0 par ceux de caractéristique 0. In Colloque dAlgébre Supérleure (Bruxelles), 1956.
36
[27] Grigory L Litvinov. Hypergroups and hypergroup algebras. arXiv preprint arXiv:1109.6596, 2011.
[28] Oliver Lorscheid. The geometry of blueprints: Part i: Algebraic background and scheme theory. Advances
in Mathematics, 229(3):1804–1846, 2012.
[29] Oliver Lorscheid. Scheme theoretic tropicalization. arXiv preprint arXiv:1508.07949, 2015.
[30] Diane Maclagan and Bernd Sturmfels. Introduction to tropical geometry, volume 161. American Mathematical Soc., 2015.
[31] Yuri Manin. Lectures on zeta functions and motives (according to deninger and kurokawa). Astérisque,
228(4):121–163, 1995.
[32] Murray Marshall. Real reduced multirings and multifields. Journal of Pure and Applied Algebra,
205(2):452–468, 2006.
[33] Frederic Marty. Sur une généralization de la notion de groupe. In 8th Congress Math. Scandinaves,
Stockholm, pages 45–49, 1934.
[34] Sam Payne. Analytification is the limit of all tropicalizations. Mathematical Research Letters, 16:543–556,
2009.
[35] Javier López Pena and Oliver Lorscheid. Mapping f 1-land: An overview of geometries over the field with
one element. arXiv preprint arXiv:0909.0069, 2009.
[36] R Procesi and R Rota Ciampi. The hyperring spectrum. Riv. Mat. Pura Appl, 1:71–80, 1987.
[37] Christophe Soulé. Les variétés sur le corps à un élément. Mosc. Math. J, 4(1):217–244, 2004.
[38] Koen Thas. Hyperfield extensions, characteristic one and the Connes-Consani plane connection. arXiv
preprint arXiv:1407.0607, 2014.
[39] Koen Thas and Don Zagier. Finite projective planes, Fermat curves, and Gaussian periods. J. Eur. Math.
Soc.(JEMS), 10(1):173–190, 2008.
[40] Bertrand Toën and Michel Vaquié. Au-dessous de spec. Journal of K-theory: K-theory and its Applications
to Algebra, Geometry, and Topology, 3(03):437–500, 2009.
[41] Oleg Viro. Hyperfields for tropical geometry i. hyperfields and dequantization. arXiv preprint
arXiv:1006.3034, 2010.
[42] Oleg Viro. On basic concepts of tropical geometry. Proceedings of the Steklov Institute of Mathematics,
273(1):252–282, 2011.
[43] Paul-Hermann Zieschang. Hypergroups. Max-Planck-Institut für Mathematik Preprint Series, 2010.
Department of Mathematical Sciences, Binghamton University, Binghamton, NY 13902, USA
E-mail address: [email protected]
37
| 0math.AC
|
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND
APPLICATIONS
arXiv:1705.06797v2 [math.FA] 9 Feb 2018
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
Abstract. The main result of this article is a rigidity result pertaining
to the spreading model structure for Banach spaces coarsely embeddable
into Tsirelson’s original space T ∗ . Every Banach space that is coarsely
embeddable into T ∗ must be reflexive and all its spreading models must
be isomorphic to c0 . Several important consequences follow from our
rigidity result. We obtain a coarse version of an influential theorem of
Tsirelson: T ∗ does not coarsely contain c0 nor ℓp for p ∈ [1, ∞). We
show that there is no infinite dimensional Banach space that coarsely
embeds into every infinite dimensional Banach space. In particular, we
disprove the conjecture that the separable infinite dimensional Hilbert
space coarsely embeds into every infinite dimensional Banach space. The
rigidity result follows from a new concentration inequality for Lipschitz
maps on the infinite Hamming graphs and taking values in T ∗ , and from
the embeddability of the infinite Hamming graphs into Banach spaces
that admit spreading models not isomorphic to c0 . Also, a purely metric
characterization of finite dimensionality is obtained.
Contents
1. Introduction
1.1. Motivation
1.2. Main results
2. Preliminaries
2.1. Nonlinear embeddings
2.2. Tree, symmetric difference and Hamming metrics
2.3. Spreading models
2.4. Tsirelson spaces T and T ∗
3. A metric characterization of finite dimensionality
4. The concentration inequalities
5. Final comments and open problems
References
2
2
4
5
5
6
7
9
10
13
18
19
2010 Mathematics Subject Classification. 46B20, 46B85, 46T99, 05C63, 20F65.
The first named author was partially supported by the National Science Foundation
under Grant Number DMS-1565826.
The second named author was supported by the French “Investissements d’Avenir”
program, project ISITE-BFC (contract ANR-15-IDEX-03) and as a participant of the
“NSF Workshop in Analysis and Probability” at Texas A&M University.
The third named author was supported by the National Science Foundation under
Grant Number DMS-1464713.
1
2
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
1. Introduction
1.1. Motivation. A natural and powerful approach to classify mathematical objects in a given category, is to discover properties that are invariant
with respect to the isomorphisms of the category. The notion of topological
dimension, which allows us to distinguish between R2 and R3 as topological
spaces, is a prime example. These invariants can take various forms, and
in metric geometry one usually seeks for properties that are stable under
embeddings of a certain type, e.g. Lipschitz, uniform, or coarse. In this
article we are concerned with a coarse invariant in the form of a concentration inequality for Lipschitz maps defined on infinite graphs with values into
certain infinite dimensional Banach spaces.
Discovering useful and tractable new concentration inequalities is usually
a challenge. In the Lipschitz category, several fundamental inequalities have
been discovered, partially motivated by the investigation of the Ribe program and its applications in geometric group theory or theoretical computer
science. We refer to the survey of A. Naor [25] for an enlightening account
of this theory, and to the recent work of Naor and Schechtman ([28], [27]
and [26]) where fundamental and powerful new inequalities were proved.
Strong interest in the coarse geometry of infinite dimensional Banach
spaces arose in connection with the Novikov conjecture in Topology and a
coarse version of the Baum-Connes conjecture in Noncommutative Geometry (see [30], [38], [13] and [36]). A very efficient and powerful technique for
proving the coarse geometric Novikov conjecture for a bounded geometry
metric space is to coarsely embed this space into a Banach space with some
strong enough convexity and smoothness properties. Indeed, G. Yu [39]
showed that a discrete metric space with bounded geometry that is coarsely
embeddable into a Hilbert space satisfies the coarse geometric Novikov conjecture. This result was later generalized by Kasparov and Yu [23] who
proved that the conclusion holds if the space is merely coarsely embeddable
into a super-reflexive Banach space. It is worth mentioning that the coarse
geometric Novikov conjecture implies Gromov’s conjecture that a uniformly
contractible Riemannian manifold with bounded geometry cannot have uniformly positive scalar curvature, and the zero-in-the spectrum conjecture
stating that the Laplace operator acting on the space of all L2 -forms of a
uniformly contractible Riemannian manifold has zero in its spectrum. Consequently, whether the separable Hilbert space is the Banach space into
which it is the hardest to embed, became a very natural and intriguing
question. More precisely, the following problem was raised (Problem 14 in
[15], Problem 11.17 in [34]).
Main Problem. Does ℓ2 coarsely embed into every infinite dimensional
Banach space?
There is rather strong evidence that ℓ2 is the space into which it is the
most difficult to embed. It was shown in [33] that every locally finite metric
subset of ℓ2 admits a bi-Lipschitz embedding into every infinite dimensional
Banach space, and in [6] that every proper subset of ℓ2 (i.e. whose closed
balls are compact) is almost Lipschitz embeddable into every infinite dimensional Banach space. Both proofs use Dvoretzky’s theorem [11] and the
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
3
barycentric gluing technique introduced in [3]. Moreover, using Gaussian
random variables one can embed ℓ2 linearly isometrically into Lp [0, 1] for
all p ∈ [1, ∞). Nowak [29] proved that ℓ2 coarsely embeds into ℓp for all
p ∈ [1, ∞) using Mazur maps. This technique was extended by Ostrovskii
[33] to embed ℓ2 coarsely into every Banach space with an unconditional
basis and non-trivial cotype using the Mazur maps constructed by Odell
and Schlumprecht in their solution to the distortion
[31]. There are
P problem
n ) , or c (which is
also Banach spaces with trivial cotype such as ( ∞
ℓ
0
n=1 ∞ ℓp
a universal space for separable spaces and bi-Lipschitz embeddings [1]) that
contain coarsely ℓ2 .
On the other hand, useful coarse invariants, such as asymptotic dimension or finite decomposition complexity, are not well suited to study the
coarse geometry of infinite dimensional Banach spaces, and the number of
properties at our disposal to prevent coarse embeddability between infinite
dimensional Banach spaces is rather limited. There are essentially three
obstructions. The first obstruction was discovered by Johnson and Randrianarivony [18] when they showed that ℓq does not coarsely embed into ℓ2
when q > 2. Their approach relies heavily upon the fact that the target
space is ℓ2 . The second obstruction is the metric cotype of Mendel and
Naor [24]. For instance, the metric cotype argument can be used to rule out
the coarse embeddability of ℓq into ℓp when q > max{p, 2}. The last one is
Kalton’s property Q [19] which serves as an obstruction to coarse embeddability into reflexive (or stable) Banach spaces1. Since ℓ2 has the smallest
possible cotype allowed for a Banach space, and is both reflexive and stable,
none of the above obstructions can provide a negative answer to the main
problem.
In this article the main problem is answered negatively, using the following approach. We show that there exists an infinite dimensional Banach
space in which the sequence (Hkω )k≥1 of k-dimensional Hamming graphs
(over a countable set) does not equi-coarsely embed. Since it is easy, and
well known, that (Hkω )k≥1 equi-coarsely embeds into ℓ2 , the claim follows.
Our work is inspired by a concentration inequality of Kalton and Randrianarivony [21]. They proved the following: Assume that f is a Lipschitz map
from Hkω into a reflexive Banach space Y , for which there exists p ∈ (1, ∞),
such that for any y in Y and any weakly null sequence (yn )∞
n=1 in Y we have
(1.1)
lim sup ky + yn kpY ≤ kykpY + lim sup kyn kpY .
n→∞
n→∞
Then for every two vertices m̄, n̄ in some subgraph isometric to Hkω , one has
(1.2)
kf (m̄) − f (n̄)kY ≤ 3Lip(f )k1/p .
Since diam(Hkω ) = k, inequality (1.2) says that a Lipschitz map on Hkω concentrates on a “full” subgraph. This concentration inequality was originally
used by Kalton and Randrianarivony as an obstruction to coarse Lipschitz
embeddability, but also provides information on compression exponents (cf.
[4] and [9]). Reflexivity is crucial to obtain inequality (1.2). Indeed, since c0
is Lipschitz universal, inequality (1.2) cannot hold for Lipschitz maps with
1A variant of Kalton’s property Q can be useful in the non-separable setting [20].
4
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
values in c0 , even though c0 satisfies inequality (1.1) for every p ∈ (1, ∞).
In order to prevent coarse embeddability one would like to have for some
C ∈ (0, ∞), which does not depend on k, the stronger inequality
kf (m̄) − f (n̄)k ≤ CLip(f ).
(1.3)
If we could find a reflexive Banach space Y satisfying inequality (1.2) for
p = ∞ that is,
(1.4)
lim sup ky + yn k ≤ max kyk, lim sup kyn k ,
n→∞
n→∞
then the concentration inequality (1.3) would hold for Y . Unfortunately, if a
Banach space Y satisfies inequality (1.4) then Y must contain an isomorphic
copy of c0 (see [22]). Thus it cannot be reflexive and moreover it contains
a bi-Lipschitz copy of every separable metric space, which precludes any
concentration inequality. As we shall see, this transition phase phenomenon
is not necessarily an obstruction. There are actually Banach spaces that
have a “c0 -like” asymptotic smoothness property which is strong enough to
obtain the concentration inequality (1.3) but weak enough to coexist with
reflexivity. A prime example of such space is Tsirelson’s original space.
Tsirelson’s original space T ∗ was the first example of a Banach space that
does not contain any isomorphic copies of ℓp or c0 [37]. The validity of
the concentration inequality (1.3) for T ∗ is pivotal to prove an unexpected
rigidity result which in turn implies that coarse and uniform versions of
Tsirelson’s theorem hold.
1.2. Main results. All unexplained notation can be found in Section 2.
The main result of this article is the following rigidity result pertaining to
the spreading model structure of Banach spaces coarsely embeddable into
Tsirelson’s original space T ∗ .
Theorem A. Let X be an infinite dimensional Banach space. If X coarsely
embeds into T ∗ , or if its unit ball BX uniformly embeds into T ∗ , then X is
reflexive and all its spreading models are isomorphic to c0 .
Our rigidity result shows that Tsirelson’s construction is actually extremely robust in the sense that nonlinear versions of Tsirelson’s theorem
hold. Indeed, Corollary B below is an immediate consequence of Theorem
A since for p ∈ (1, ∞) the canonical basis of ℓp generates a spreading model
that is linearly isometric to ℓp , while ℓ1 and c0 are not even reflexive.
Corollary B. c0 and ℓp for p ∈ [1, ∞) (resp. Bc0 and Bℓp ) do not coarsely
embed (resp. uniformly embed) into T ∗ .
We will say that a Banach space is coarsely minimal if it coarsely embeds into every infinite dimensional Banach space. The Main Problem asks
whether ℓ2 is coarsely minimal. In fact, Theorem A provides a much stronger
negative solution to the Main Problem. Indeed, a coarsely minimal Banach
space embeds into ℓ2 and it must have non-trivial cotype using Mendel and
Naor metric cotype notion [24], but it also embeds into T ∗ and by Theorem
A it must have trivial cotype, a contradiction.
Corollary C. There is no coarsely minimal infinite dimensional Banach
space.
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
5
The proof of Theorem A relies essentially on two ingredients. The main
ingredient, which is presented in Section 4, is a new concentration inequality
for Lipschitz maps on the k-dimensional infinite Hamming graph and taking
values into Tsirelson’s original space T ∗ .
Theorem D. Let k ∈ N and f : ([N]k , dH ) → T ∗ be a Lipschitz map. Then
there exists M ∈ [N]ω such that for all m̄, n̄ ∈ [M]k one has
(1.5)
kf (m̄) − f (n̄)k ≤ 5Lip(f ).
The concentration inequality above is strong enough to rule out coarse
and uniform embeddings, and can already be invoked to provide a negative
solution to the Main Problem without the full strength of Theorem A.
The other ingredient, presented in Section 3, uses the structure of spreading models of a Banach space to prove positive embeddability results. Such
results are obtained for three different metrics: the tree metric, the Hamming metric and the symmetric difference metric. Our embeddings are based
on the existence of spreading models that are isomorphic, or not isomorphic,
to c0 .
The proof of Theorem A, given in Section 4, is obtained by analyzing the
tension between the concentration inequality for the Hamming graphs and
the embeddability of the Hamming graphs into spaces which have spreading
models that are not isomorphic to c0 . For readers interested mostly in the
geometry of Hilbert space we included in Remark 4.6 a short proof (which
avoids the machinery of spreading models and only requires Theorem D) of
the ℓ2 -case of Corollary B.
The analysis of the spreading model structure of infinite dimensional Banach spaces in Section 3, also lead us to a metric characterization of finite dimensionality in terms of equi-coarse embeddability of the sequence of
countably branching trees of finite but arbitrarily large height, denoted by
(Tkω )k≥1 .
Theorem E. For a Banach space Y the following assertions are equivalent:
(1) Y is finite-dimensional.
(2) (Tkω )k≥1 does not equi-coarsely embed into Y .
We mention a few related open problems in the last section.
2. Preliminaries
2.1. Nonlinear embeddings. Let (X, dX ) and (Y, dY ) be two metric spaces
and f : X → Y . One defines
ρf (t) = inf dY (f (x), f (y)) : dX (x, y) ≥ t ,
and
ωf (t) = sup{dY (f (x), f (y)) : dX (x, y) ≤ t}.
Note that for every x, y ∈ X,
(2.1)
ρf (dX (x, y)) ≤ dY (f (x), f (y)) ≤ ωf (dX (x, y)).
6
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
The moduli ρf and ωf will be called the compression modulus and the
expansion modulus of the embedding, respectively. We adopt the convention sup(∅) = 0 and inf(∅) = +∞. The map f is a coarse embedding if
limt→∞ ρf (t) = ∞ and ωf (t) < ∞ for all t > 0. A map f : X → Y is said
to be a uniform embedding if limt→0 ωf (t) = 0 and ρf (t) > 0 for all t > 0,
i.e. f is an injective uniformly continuous map whose inverse is uniformly
continuous.
If one is given a family of metric spaces (Xi )i∈I , one says that (Xi )i∈I equicoarsely (resp. equi-uniformly) embeds into Y if there exist non-decreasing
functions ρ, ω : [0, ∞) → [0, ∞) and for all i ∈ I, maps fi : Xi → Y such that
ρ ≤ ρfi , ωfi ≤ ω, and limt→∞ ρ(t) = ∞ and ω(t) < ∞ for all t > 0 (resp.
limt→0 ω(t) = 0 and ρ(t) > 0 for all t > 0).
2.2. Tree, symmetric difference and Hamming metrics. In this section we define the metric spaces which will be needed and studied in the
sequel. For any infinite subset M of N, let [M]ω := {A ⊂ M : A infinite} and
[M]<ω := {A ⊂ M : A finite}. For k ∈ N, we put [M]≤k := {A ⊂ M : |A| ≤
k} and [M]k := {A ⊂ M : |A| = k} where |A| denotes the cardinality of the
set A. Elements of [M]ω and [M]<ω will always be listed in an increasing order, i.e., if we write m̄ = {m1 , m2 , . . .} ∈ [M]ω , or m̄ = {m1 , m2 , . . . , ml } ∈
[M]<ω we assume that m1 < m2 < . . . or m1 < m2 < . . . < ml , respectively.
We define three metrics, dT , d△ and dH on [N]<ω . The restrictions of these
metrics to the sets [M]≤k , [M]k for some M ∈ [N]ω and some k ∈ N, will still
be denoted dT , d△ and dH .
The tree metric. For m̄, n̄ ∈ [N]<ω , m̄ 6= n̄, let
(2.2)
dT (m̄, n̄) := |m̄| + |n̄| − 2|m̄ ∧ n̄|,
where m̄ ∧ n̄ := {m1 , m2 , . . . , ml } if m1 = n1 , m2 = n2 , . . . , ml = nl and
ml+1 6= nl+1 .
We denote by T the partial order of extension on [N]<ω , i.e. m̄ T n̄ if
|m̄| ≤ |n̄| and for all i ≤ |m̄| one has mi = ni . One can then define a graph
structure on [N]<ω , which is our set of vertices, by declaring two vertices m̄, n̄
to be adjacent if m̄ is the immediate predecessor of n̄ or n̄ is the immediate
predecessor of m̄. The graph obtained is the countably branching tree of
ω (N) or simply T ω . The countably branching tree
infinite height, denoted T∞
∞
≤k
of height k, [N] , will also be denoted Tkω . It is easy to see that the tree
metric coincides with the graph metric.
The symmetric difference metric. For m̄, n̄ ∈ [N]<ω , m̄ 6= n̄, let
(2.3)
d△ (m̄, n̄) := |m̄ △ n̄|,
where m̄ △ n̄ := (m̄ \ n̄) ∪ (n̄ \ m̄) is the symmetric difference. At some point,
it will be useful for us to notice that one can define a graph structure on [N]k
by defining an adjacency relation as follows: two vertices m̄, n̄ ∈ [N]k are
adjacent if and only if |m̄ \ n̄| = |n̄ \ m̄| = 1. This graph is the k-dimensional
Johnson graph over a countable set, and will be denoted Jkω (N) or simply
d
Jkω , and its graph metric dJ coincides with the metric 2△ .
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
7
The Hamming metric. For m̄, n̄ ∈ [N]<ω , let
(2.4)
dH (m̄, n̄) := {i ∈{1, 2, . . . , min(|m̄|, |n̄|)} : mi 6= ni }
+ max(|m̄|, |n̄|) − min(|m̄|, |n̄|).
When restricted to the set [N]k , the metric dH , can be seen as the graph
metric on the Hamming graph over a countable alphabet, denoted Hkω (N)
or simply Hkω , where two vertices are adjacent if they differ in exactly one
coordinate.
Remark 2.1. Both graphs Hkω and Jkω are infinite versions of finite graphs
that arose from the fundamental work from the mid-20th century in Coding
Theory of R. W. Hamming and S. M. Johnson, respectively.
2.3. Spreading models. For the convenience of the reader we briefly recall
a few useful results from the theory of spreading models initiated by Brunel
and Sucheston [10] in the 70’s. We shall follow the exposition in [8]. An application of Ramsey’s Theorem yields that every bounded sequence (xn )∞
n=1
in a separable Banach space admits a subsequence (yn )∞
n=1 such that for all
k ≥ 1, (ai )ki=1 ⊂ R,
N (a1 , a2 , . . . , ak ) :=
lim
n1 <n2 <···<nk →∞
ka1 yn1 + a2 yn2 + · · · + ak ynk k
exists, where the limit means that for all ε > 0, there is r ∈ N such that for
all r ≤ n1 < n2 < · · · < nk ,
(2.5)
ka1 yn1 + a2 yn2 + · · · + ak ynk k − N (a1 , a2 , · · · , ak ) < ε.
It is easy to see that that if (ei )∞
i=1 denotes the canonical basis of c00 ,
the vector space of sequences in R which eventually vanish, the formula
ka1 e1 +a2 e2 +· · ·+ak ek kE := N (a1 , a2 , · · · , ak ) defines a semi-norm, and that
k
(ei )∞
i=1 is a spreading sequence in the sense that for all k ≥ 1, (ai )i=1 ⊂ R,
and integers n1 < n2 < · · · < nk ,
(2.6)
ka1 en1 + a2 en2 + · · · + ak enk kE = ka1 e1 + a2 e2 + · · · + ak ek kE .
If (xn )∞
n=1 does not have a converging subsequence it is simple to show
that the semi-norm is actually a norm. The completion of c00 for the norm
k·kE is a Banach space E called a spreading model of X generated by the se∞
quence (xn )∞
n=1 , and we will refer to the sequence (ei )i=1 as the fundamental
sequence of the spreading model. The fundamental sequence (ei )∞
i=1 is not
∞
necessarily a basis of E but if the generating sequence (xn )n=1 is normalized
and weakly null then (ei )∞
i=1 is a normalized 1-suppression unconditional basis ([8, Proposition 1, p. 24]). We call a basic sequence (zj )∞
j=1 in a Banach
space Z c-suppression unconditional, for some c ≥ 1, if for all (ai )∞
i=1 ∈ c00
and all A ⊂ N
∞
X
X
ai zi ,
ai zi ≤ c
(2.7)
i∈A
i=1
and in this case we denote by cs the smallest number c satisfying (2.7). The
∞
basic sequence (zj )∞
j=1 is called c-unconditional if for all (ai )i=1 ∈ c00 and
8
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
all (εi )∞
i=1 ⊂ {−1, 1} we have
∞
X
(2.8)
i=1
εi ai zi ≤ c
∞
X
ai zi ,
i=1
and we denote in this case the smallest number c satisfying (2.8) by cu . It
is easy to see that basic sequences are suppression unconditional if and only
if they are unconditional, and that in this case
cs ≤ cu ≤ 2cs .
(2.9)
P
The function ϕE (k) := k ki=1 ei kE where (ei )∞
i=1 is the fundamental sequence of a spreading model E of a Banach space is usually called the
fundamental function of E. We deduce the following.
Proposition 2.2. Let (xn )∞
n=1 be a normalized weakly null sequence in a
Banach space X. Then for every ε > 0 there is a normalized weakly null
∞
basic subsequence (yn )∞
n=1 of (xn )n=1 with basis constant (1 + ε), generating
a spreading model E with fundamental function ϕE , such that for all k ≥ 1,
for all k ≤ n1 < n2 < · · · < nk , and for all (εi )ki=1 ⊂ {−1, 1} one has
k
X
1
1
(2.10)
ϕE (k) ≤
εi ei
2(1 + ε)
1+ε
i=1
≤
k
X
i=1
εi yni ≤ (1 + ε)
k
X
i=1
εi ei ≤ 2(1 + ε)ϕE (k).
We will be particularly interested in c0 -spreading models, where c0 is the
space of real valued sequences converging to 0 equipped with the sup-norm.
The following proposition follows from [8, Lemma 1, p. 73] and [8, Lemma
4, p. 75].
Proposition 2.3. X has a spreading model isomorphic to c0 if and only
if for all ε > 0, X has a spreading model E whose fundamental sequence
(ei )∞
i=1 is (1 + ε)-equivalent to the canonical basis of c0 , i.e. that for all
k ≥ 1, (ai )ki=1 ⊂ R,
(2.11)
1
sup |ai | ≤
(1 + ε) 1≤i≤k
k
X
ai ei
i=1
E
≤ (1 + ε) sup |ai |.
1≤i≤k
We will also need the following observation from [8].
Proposition 2.4. [8, Proposition 3, p. 79] Let (xn )∞
n=1 be a normalized
weakly null sequence in a Banach space X that generates a spreading model
E. Then E is not isomorphic to c0 if and only if there exists a subsequence
∞
(yn )∞
n=1 of (xn )n=1 such that,
(2.12)
lim
k→∞
inf
n1 <···<nk
inf
(εi )ki=1 ∈{−1,1}k
k
X
i=1
εi yni = +∞.
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
9
2.4. Tsirelson spaces T and T ∗ . We recall a construction of a space originally designed by Tsirelson [37], which is nowadays usually referred to as
T ∗ , while T denotes its dual space and was described by Figiel and Johnson in [14]. Doing so, convenient notation and terminology that will be
needed later on are introduced. Tsirelson’s original construction of T ∗ has
a geometric flavor while Figiel-Johnson construction of T is more analytic.
For E, F ∈ [N]<ω , and n ∈ N we write E < F if max(E) < min(F ),
and n ≤ E, or n < E, if n ≤ min E, or n < min E, respectively. Here
we set max(∅) = 0 and min(∅) = ∞. We call a sequence (Ej )nj=1 ⊂ [N]<ω
P
admissible if n ≤ E1 < E2 < · · · < En . For x = ∞
j=1 ξj ej ∈ c0 we call
supp(x) =
{j
∈
N
:
ξ
=
6
0}
the
support
of
x
and
for
E ∈ [N]<ω we write
j
P
n
E(x) =
j∈E ξj ej . We call a finite sequence (xj )j=1 of elements in c00
a block sequence if supp(xj−1 ) < supp(xj ), for all 2 ≤ j ≤
Pn. We define
inductively for each k ∈ N0 a norm k · kk on c00 . For x = ∞
j=1 ξj ej ∈ c00
we put kxk0 = maxj∈N |ξj |, and assuming that k · kk−1 has been defined for
some k ∈ N, we let for x ∈ c00
n
nX
o
1
kEj (x)kk−1 : (Ej )nj=1 is admissible .
kxkk = max kxkk−1 , max
2
j=1
Then we put for x ∈ c00
kxk = lim kxkk = max kxkk .
k→∞
k∈N
k · k is then a norm on c00 and T is defined to be the completion of c00
with respect to k · k. As observed in [14], k · k satisfies the following implicit
equation for all x ∈ T
n
o
nX
1
kEj (x)k : (Ej )nj=1 is admissible ,
kxk = max kxk0 , sup
2
j=1
and the unit vector basis is a 1-unconditional basis of T . Then it was proven
in [14] that T does not contain a subspace isomorphic to ℓ1 , which, together
with the easy observation that T certainly does not contain a subspace
isomorphic to c0 , yields by James’ Theorem [16, Theorem 2] that T must be
reflexive. Let us now denote the dual of T by T ∗ (the original Tsirelson’s
space). We denote the unit basis of T ∗ by (e∗j )∞
j=1 (which is a 1-unconditional
basis of T ∗ ) and its unit ball by BT ∗ . Next to the property that T ∗ is
reflexive, the following property of T ∗ will be essential for us:
(2.13)
n
X
j=1
x∗j ≤ 2,
whenever (x∗j )nj=1 ⊂ BT ∗ is a block sequence, with n ≤ supp(x∗1 ).
Indeed, assume that (x∗j )nj=1 is a block sequence in BT ∗ , with n ≤ supp(x∗1 ),
Pn
Pn
∗
∗
and let x ∈ T , kxk = 1, be such that
j=1 xj (x).SBy the
j=1 xj =
n
1-unconditionality of (ej )∞
j=1 in T we can assume that supp(x) ⊂
j=1 Ej ,
10
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
where Ej = supp(x∗j ), for j = 1, 2, . . . , n. Since (Ej )nj=1 is admissible we
obtain from (2.4) that
n
X
j=1
x∗j =
n
X
x∗j (x) =
j=1
n
X
j=1
n
X
Ej (x)k ≤ 2kxk = 2,
x∗j Ej (x) ≤
j=1
which proves our claim (2.13).
Tsirelson’s space T ∗ is not the only reflexive space satisfying inequality
(2.13). For every countable ordinal α, one can define Tα∗ (see [32]), a higher
order version of T ∗ , where the admissibility condition refers to Schreier
families of order α. These Tsirelson-type spaces are all reflexive and satisfy
inequality (2.13), and they are incomparable in the sense that Tα∗ does not
embed isomorphically into Tβ∗ whenever α 6= β.
Remark 2.5. In this article, all the results that are valid for T ∗ also hold for
every reflexive Banach space satisfying equation (2.13), so in particular for
Tα∗ for any countable ordinal α.
3. A metric characterization of finite dimensionality
We start by proving some coarse embeddability results for the metric
spaces introduced in Section 2.2. Here, we do not need that the metrics can
be seen as graph metrics. In Lemma 3.1 below we show that if Y is an infinite
dimensional Banach space admitting a spreading model E, generated by a
weakly null sequence, then there always exists a 1-Lipschitz map into Y ,
defined on any of the spaces ([N]<ω , dT ), ([N]<ω , dH ), or ([N]<ω , d△ ), whose
compression modulus is essentially bounded from below by the fundamental
function of the spreading model.
Lemma 3.1. Let (X, dX ) be any of the following metric spaces: ([N]<ω , d△ ),
([N]<ω , dT ), or ([N]k , dH ), for k ∈ N. Let Y be an infinite dimensional
Banach space admitting a spreading model E, generated by a normalized
weakly null sequence, then for every ν > 0 there exists a map fν : (X, dX ) →
Y such that for all x, y ∈ X,
1
ϕE (dX (x, y)) ≤ kfν (x) − fν (y)kY ≤ dX (x, y).
(3.1)
8(1 + ν)
Proof. Let ε > 0 such that (1 + ε)2 ≤ (1 + ν). By Proposition 2.2 there
is a weakly null normalized basic sequence (yn )∞
n=1 with basis constant not
larger than (1 + ε), and thus the bimonotonicity constant is at most 2(1 + ε),
generating a spreading model E, such that for all k ≥ 1, for all k ≤ n1 <
n2 < · · · < nk , and for all (εi )ki=1 ∈ {−1, 1}k one has
(3.2)
k
X
i=1
εi yni ≥
k
X
1
1
εi ei ≥
ϕE (k).
1+ε
2(1 + ε)
i=1
We now consider the three different cases.
The symmetric difference metric case. Let f : ([N]<ω , d△ ) → Y be
k
X
yni , for n̄ = {n1 , n2 , . . . , nk } ∈ [N]<ω . The map f is
defined by f (n̄) =
i=1
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
11
clearly 1-Lipschitz since after cancellations it follows for m̄, n̄ ∈ [N]<ω that
d△ (m̄,n̄)
X
kf (m̄) − f (n̄)k =
εi yqi
i=1
d△ (m̄,n̄)
(εi )i=1
for some q1 < q2 < · · · < qd△ (m̄,n̄) and
∈ {−1, 1}d△ (m̄,n̄) . If d =
d△ (m̄, n̄) = 2r is even, then qr+1 ≥ r, and hence (3.2) and the assumption
on the bimonotonicity constant of (yn )∞
n=1 yield
kf (m̄) − f (n̄)k ≥
1
2(1 + ε)
d
X
i=r+1
εi yqi ≥
1
2(1 + ε)2
d
X
εi ei
i=r+1
1
1
≥
ϕE (r) ≥
ϕE (d).
2
4(1 + ε)
8(1 + ε)2
If d is odd, the proof is similar, as will be seen while treating the next case.
<ω
<ω
The tree metric case. Let
XΦ : [N] → N be a bijection. Let f : ([N] , dT ) →
Y be defined by f (n̄) =
yΦ(ū) , for n̄ ∈ [N]<ω . The map f is clearly 1ūn̄
Lipschitz since after cancellations one has
dT (m̄,n̄)
kf (m̄) − f (n̄)k =
X
εi yqi ,
i=1
d (m̄,n̄)
T
for some q1 < q2 < · · · < qdT (m̄,n̄) and (εi )i=1
∈ {−1, 1}dT (m̄,n̄) . Here we
only detail the case when d = dT (m̄, n̄) is odd, say d = 2r − 1, with r ∈ N.
Note that qr ≥ r. So, again (3.2) and the assumption on the bimonotonicity
constant of (yn )∞
n=1 imply that
kf (m̄) − f (n̄)k ≥
d
d
X
X
1
1
εi ei
εi yni ≥
2(1 + ε)
2(1 + ε)2
i=r
i=r
1
1
1
≥
ϕE (r) ≥
ϕE (2r) ≥
ϕE (d).
2
2
4(1 + ε)
8(1 + ε)
8(1 + ε)2
The Hamming metric case. Let Φ : N × N → N be a bijection and let
k
1X
k
yΦ(i,ni ) . The map f is clearly
f : ([N] , dH ) → Y be defined by f (n̄) =
2
i=1
1-Lipschitz since after cancellations one has
1
kf (m̄) − f (n̄)k =
2
2dH (m̄,n̄)
X
εi yqi
i=1
2d (m̄,n̄)
for some q1 < q2 < · · · < q2dH (m̄,n̄) and (εi )i=1H
∈ {−1, 1}2dH (m̄,n̄) .
Necessarily qdH (m̄,n̄) ≥ dH (m̄, n̄), and hence, as in the other cases,
kf (m̄) − f (n̄)k ≥
1
4(1 + ε)
2dH (m̄,n̄)
X
i=dH (m̄,n̄)+1
εi yqi ≥
1
ϕE (dH (m̄, n̄)) .
8(1 + ε)2
12
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
Remark 3.2. We restricted in Lemma 3.1 the Hamming metric dH to [N]k ,
for k ∈ N, because dH is usually defined for strings of equal length. But it is
not hard to show the existence of a map f : [N]<ω → X, satisfying condition
(3.1), if we replace the value 8 by a larger number.
Lemma 3.1 yields a coarse embedding of ([N]<ω , dT ), ([N]<ω , d△ ), and
an equi-coarse embedding of the sequence ([N]k , dH )k∈N into every Banach
space which admits at least one spreading model, that is not isomorphic to
c0 and is generated by a weakly null sequence. Recall that Proposition 2.4 in
Section 2.3 insures that the associated fundamental function is unbounded.
Then, the following proposition follows simply from Lemma 3.1.
Corollary 3.3. The metric spaces ([N]<ω , dT ), ([N]<ω , d△ ), are all coarsely
embeddable, and the sequence ([N]k , dH )k∈N is equi-coarsely embeddable into
a Banach space that has a spreading model E, generated by a normalized
weakly null sequence, which is not isomorphic to c0 .
More precisely, in all three cases, for any ν > 0 there is a coarse embedding
fν into Y such that fν is 1-Lipschitz and the modulus of compression of
1
ϕE (t), for t > 0, where ϕE is the fundamental
fν satisfies ρfν (t) ≥ 8(1+ν)
function of the spreading model E.
For the tree metric we can deduce the following embeddability result, even
if all spreading models are equivalent to c0 .
Proposition 3.4. If an infinite dimensional Banach space Y has a spreading model isomorphic to c0 then (Tkω )k≥1 equi-bi-Lipschitzly embeds into Y .
More precisely, for every ν > 0 and every k ∈ N there exists a map
fk : Tkω → Y , such that for all x, y ∈ Tkω ,
(3.3)
1
dT (x, y) ≤ kfk (x) − fk (y)kY ≤ dT (x, y).
2(1 + ν)
Proof. By Proposition 2.3 one may assume that there exists a normalized
sequence (yn )∞
n=1 such that for all k ≥ 1, for all k ≤ n1 < n2 < · · · < nk ,
and for all (ai )ki=1 ⊂ R one has
(3.4)
1
sup |ai | ≤
(1 + ν) 1≤i≤k
k
X
i=1
ai yni ≤ (1 + ν) sup |ai |.
1≤i≤k
Let Φk : [N]≤k → {2k, 2kX
+X
1, . . . } be a bijection. For k ∈ N let fk : Tkω → Y
be defined by fk (n̄) =
yΦ(s̄) , for n̄ ∈ [N]≤k . After cancellations one
ūn̄ s̄ū
has
dT (m̄,n̄)
kfk (m̄) − fk (n̄)k =
X
αi yqi
i=1
d (m̄,n̄)
T
for some 2k ≤ q1 < · · · < qdT (m̄,n̄) and (αi )i=1
⊂ {1, . . . , dT (m̄, n̄)}
1
satisfying 2 dT (m̄, n̄) ≤ sup1≤i≤dT (m̄,n̄) |αi | ≤ dT (m̄, n̄). So
kfk (m̄) − fk (n̄)k ≥
1
1
sup
|αi | ≥
dT (m̄, n̄)
(1 + ν) 1≤i≤dT (m̄,m̄)
2(1 + ν)
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
13
and
kfk (m̄) − fk (n̄)k ≤ (1 + ν)
sup
1≤i≤dT (m̄,n̄)
|αi | ≤ (1 + ν)dT (m̄, n̄),
which proves our claim.
Remark 3.5. If an infinite dimensional Banach space Y has a spreading
model isomorphic to c0 then, using a duality argument [8, Proposition 1, p.
80], the relationship between the ℓ+
1 -weakly null index and the Szlenk index
[2, Theorem 4.2], and the embedding in [5, Theorem 2.6], one can show that
ω embeds bi-Lipschitzly into Y .
T∞
Moreover, since it is easy to see that every countable tree isometrically
ω we obtain that every countable tree coarsely embeds into
embeds into T∞
any infinite dimensional Banach space.
In the following theorem, which includes Theorem E, a purely metric
characterization of finite dimensionality in terms of graph preclusion in the
coarse category is given.
Theorem 3.6. Let Y be a Banach space. The following assertions are
equivalent:
(1) Y is finite dimensional.
(2) (Tkω )k≥1 does not equi-coarsely embed into Y .
ω does not coarsely embed into Y .
(3) T∞
Proof. We start with a self contained elementary proof of the main equivalence (1) ⇐⇒ (2), which avoids the use of Remark 3.5.
By compactness, a finite dimensional Banach space cannot contain an infinite sequence that is bounded and separated. On the other hand, if Y
is an infinite dimensional Banach space, then by Rosenthal’s ℓ1 -theorem, it
either contains ℓ1 isomorphically or it has a weakly Cauchy sequence which
ω , we may
is not norm converging. Since ℓ1 contains an isometric copy of T∞
assume that X contains a weakly Cauchy sequence (xn )∞
n=1 which is not
norm converging. After eventually passing to a subsequence of (xn )∞
n=1 , still
denoted (xn )∞
,
the
sequence
n=1
∞
x
2n−1 − x2n
(zn )∞
n=1 :=
kx2n−1 − x2n k n=1
is normalized and weakly null. Therefore, either (zn )∞
n=1 has a subsequence
that generates a spreading model isomorphic to c0 and we apply Proposition 3.4 or (zn )∞
n=1 has a subsequence that generates a spreading model not
isomorphic to c0 and we conclude with Corollary 3.3.
We now finish the circle of implications: (2) =⇒ (3) is obvious and
(3) =⇒ (1) follows from Corollary 3.3 and Remark 3.5.
4. The concentration inequalities
All the results of this section will be valid for the spaces [N]k , k ∈ N,
d
equipped either with the Johnson metric dJ = 2△ or the Hamming metric
dH . So throughout the section, d• will denote either dJ or dH . We start
with a general structural result on Lipschitz maps from ([N]k , d• ) into a
reflexive space with a basis. For two vectors x and y in a Banach space
14
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
with a basis, we shall use freely the convenient notation x ≺ y to mean that
supp(x) < supp(y), and k x when k ≤ min(supp(x)), where the supports
are with respect to the basis considered.
Proposition 4.1. Let Y be a reflexive Banach space with a bimonotone
ω
basis (ei )∞
i=1 . Then, for all k, r ∈ N, ε > 0, M ∈ [N] , and Lipschitz maps
k
′
ω
f : ([M] , d• ) → Y there exist M ∈ [N] and y ∈ Y satisfying the following:
(1)
(2)
(k)
For all m̄ ∈ [M′ ]k there exist r ym̄ ≺ ym̄ ≺ · · · ≺ ym̄ , all with finite
∞
supports with respect to (ei )i=1 , such that:
(4.1)
and
(4.2)
(i)
kym̄ k ≤ Lip(f ), for all i ∈ {1, 2, . . . , k},
(1)
(k)
(2)
kf (m̄) − (y + ym̄ + ym̄ + · · · + ym̄ )k < ε.
Proof. We will show this result by induction on k. The proof is based on
a standard gliding hump procedure and a few diagonal arguments that we
have chosen to detail. For n ∈ N, we shall denote Pn the basis projection
on the linear span of {e1 , . . . , en }.
For k = 1 assume that we are given r ∈ N, M ∈ [N]ω , a Lipschitz map
f : (M, d• ) → Y , and ε > 0. Since f is bounded and Y is reflexive, by weak
compactness, there exists y ∈ Y and M0 ∈ [M]ω such that (f (m))m∈M0
converges weakly to y. From the weak lower semi-continuity of the norm of
Y we deduce that kf (m)−yk ≤ Lip(f ) for all m ∈ M0 . Since (f (m)−y)m∈M0
is weakly null, one can find m1 ∈ M0 such that
ε
kPr (f (m) − y)k < , for all m ∈ M0 , with m ≥ m1 .
2
Then for all for all m ≥ m1 , m ∈ M0 , there exists sm > r such that
ε
k(I − Psm )(f (m) − y)k < .
2
(1)
We now set M′ := {m ∈ M0 : m ≥ m1 } and ym := (Psm − Pr )(f (m) − y),
(1)
(1)
for m ∈ M′ . For all m ∈ M′ , r ym , ym has finite support and
(1)
since the basis is bimonotone, kym k ≤ Lip(f ). Finally, we deduce that
(1)
kf (m) − (y + ym )k < ε, from the triangle inequality.
Assume now that our statement is proved for k ∈ N, and let r ∈ N, a Lipschitz map f : ([M]k+1 , d• ) → Y , and ε > 0 be given. Using that f ([M]k+1 )
is a bounded subset of the reflexive space Y and a diagonal argument, we
infer the existence of M0 ∈ [M]ω such that for all m̄ ∈ [M0 ]k , (f (m̄, n))n∈M0
converges weakly to some g(m̄) ∈ Y . Using again that the norm of Y
is weakly lower semi-continuous, we have that Lip(g) ≤ Lip(f ). This allows us to apply our induction hypothesis to the map g : [M0 ]k → Y and
to find M1 ∈ [M0 ]ω and y ∈ Y , such that for all m̄ ∈ [M1 ]k there exist
(1)
(2)
(k)
r ym̄ ≺ ym̄ ≺ · · · ≺ ym̄ , all with finite supports, such that for all
(i)
i ∈ {1, . . . , k} kym̄ k ≤ Lip(f ) and
ε
(1)
(2)
(k)
(4.3)
g(m̄) − (y + ym̄ + ym̄ + · · · + ym̄ ) < .
3
k
We now fix m̄ ∈ [M1 ] . Note that the weak lower semi-continuity of the
norm implies that for all n ∈ M1 , kf (m̄, n) − g(m̄)k ≤ Lip(f ). Denote
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
15
(k)
rm̄ = max(supp(ym̄ )). Since (f (m̄, n) − g(m̄))n∈M1 is weakly null, there
exists Nm̄ ∈ M1 such that for all n ∈ M1 , n ≥ Nm̄ :
ε
Prm̄ f (m̄, n) − g(m̄) < .
3
Then, for all n ∈ M1 , n ≥ Nm̄ , there exists sm̄,n > rm̄ so that
ε
(I − Psm̄,n ) f (m̄, n) − g(m̄) < .
3
(k+1)
We now set y(m̄,n) := (Psm̄,n − Prm̄ )(f (m̄, n) − g(m̄)). We have that for all
(k+1)
(k+1)
(k+1)
k ≺y
n ∈ M1 , n ≥ Nm̄ : y(m̄,n) has finite support, ym̄
(m̄,n) , ky(m̄,n) k ≤ Lip(f )
and
(1)
(k)
(k+1)
kf (m̄, n) − (y + ym̄ + · · · + ym̄ + y(m̄,n) )k < ε.
We conclude the proof with one last extraction argument.
If M1 = {m1 , m2 , . . . }, we define M′ := {m′1 , m′2 , . . . } recursively as follows.
For all i ∈ N, m′i := mΦ(i) where Φ(1) = 1, Φ(2) = 2, . . . , Φ(k) = k, and if
Φ(i) has been chosen for i ≥ k then Φ(i+1) is picked such that Φ(i+1) > Φ(i)
and mΦ(i+1) ≥ Nm̄ for all m̄ in the finite set [{m′1 , m′2 , . . . , m′i }]k . It is now
easy to check that M′ , y, and for all m̄ ∈ [M′ ]k and n ∈ M′ , n > max(m̄), the
(k+1)
(k)
(1)
(k)
(1)
vectors with finite support y(m̄,n) := ym̄ , . . . , y(m̄,n) := ym̄ , y(m̄,n) satisfy
the induction hypothesis for k + 1.
Remark 4.2. We have assumed that the basis of Y is bimonotone only for
convenience. In the general case, one gets a similar result, with the only
(i)
difference that kym̄ k ≤ CLip(f ), where C is the bimonotonicity constant.
Remark 4.3. Proposition 4.1 is actually true if we only assume that Y is a
Banach space with a boundedly complete Schauder finite dimensional decomposition. In that case, we just have to use weak∗ -compactness instead
of weak-compactness.
Our concentration result for Lipschitz maps with values in T ∗ follows
easily from Proposition 4.1. The following theorem is Theorem C when
d• = dH .
Theorem 4.4 (Concentration inequality). Let k ∈ N and f : ([N]k , d• ) → T ∗
be a Lipschitz map. Then there exists M′ ∈ [N]ω such that for all m̄, n̄ ∈
[M′ ]k one has
(4.4)
kf (m̄) − f (n̄)k ≤ 5Lip(f ).
Proof. We assume, as we may, that Lip(f ) > 0. Then it follows from
Proposition 4.1 with Y = T ∗ , r = k, M = N and ε = 12 Lip(f ) that
there exists M′ ∈ [N]ω , y ∈ T ∗ such that for all m̄ ∈ [M′ ]k there exist
(1)
(2)
(k)
k ym̄ ≺ ym̄ ≺ · · · ≺ ym̄ , all with finite supports, satisfying (4.1) and
(4.2). And hence, for all m̄, n̄ ∈ M′
(1)
(2)
(k)
(1)
(2)
(k)
kf (m̄) − f (n̄)k ≤ 2ε + kym̄ + ym̄ + · · · + ym̄ k + kyn̄ + yn̄ + · · · + yn̄ k
≤ 5Lip(f ),
where for the last inequality we use property (2.13) of T ∗ .
16
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
The concentration inequalities stated above immediately provide an obstruction to equi-coarse embeddability of the Johnson graphs (Jkω )k≥1 and
the Hamming graphs (Hkω )k≥1 . A rescaling argument can be used to provide
an obstruction to equi-uniform embeddability of the rescaled metrics.
Corollary 4.5.
(1) The sequence of Hamming graphs (Hkω )k≥1 (resp. ([N]k , dkH )k≥1 ) do
not equi-coarsely (resp. equi-uniformly) embed into T ∗ .
(2)The sequence of Johnson graphs (Jkω )k≥1 (resp. ([N]k , dkJ )k≥1 ) do not
equi-coarsely (resp. equi-uniformly) embed into T ∗ .
Proof. We only treat the case of the Hamming graphs, the case of the Johnson graphs being similar. Assume by contradiction that (Hkω )k≥1 equicoarsely embeds into T ∗ , i.e., that there exist non-decreasing functions
ρ, ω : [0, ∞) → [0, ∞) and for each k ∈ N a function fk : ([N]k , dH ) → T ∗ ,
such that ρ ≤ ρfk , ωfk ≤ ω, limt→∞ ρ(t) = ∞ and ω(t) < ∞, for all t > 0.
Since dH is the graph distance on Hkω , fk is actually ω(1)-Lipschitz. By
Theorem 4.4, for every k ≥ 1, there exists Mk ∈ [N]ω such that for all
m̄, n̄ ∈ [Mk ]k one has
(4.5)
kfk (m̄) − fk (n̄)k ≤ 5ω(1).
If m̄ and n̄ are chosen in [Mk ]k such that dH (m̄, n̄) = k then it implies that
for every k ≥ 1, ρ(k) ≤ 5ω(1), which contradicts the properties of ω and ρ
for k large enough.
Assume now by contradiction that ([N]k , dkH )k≥1 equi-uniformly embeds
into T ∗ , i.e., there exist non-decreasing functions ρ, ω : [0, ∞) → [0, ∞) and
for each k ∈ N a function fk : ([N]k , dkH ) → T ∗ , such that ρ ≤ ρfk , ωfk ≤ ω,
limt→0 ω(t) = 0 and ρ(t) > 0, for all t > 0. Denote the identity map from
([N]k , dH ) to ([N]k , dkH ) by ik . Then, Lip(fk ◦ ik ) = ωfk ◦ik (1) ≤ ω( k1 ). By
Theorem 4.4, for all k ≥ 1, there exist Mk ∈ [N]ω such that kfk (m̄) −
fk (n̄)k ≤ 5ω( k1 ), for all m̄, n̄ ∈ [Mk ]k . Therefore, considering again m̄ and
n̄ in [Mk ]k such that dH (m̄, n̄) = k, we get that for all k ≥ 1 one has
0 < ρ(1) ≤ 5ω( k1 ), which is a contradiction for large enough k.
Remark 4.6. As previously mentioned, Corollary B follows from the more
involved rigidity phenomenon depicted in Theorem A. Nevertheless, for
the reader only interested in the geometry of Hilbert space we include an
abridged proof of the ℓ2 -case. The proof follows simply from Corollary 4.5
and, in the coarse setting, the fact that the Johnson graphs equi-coarsely
embeds into ℓ2 . Indeed, for k ∈ N consider the map
k
fk : ([N] , dJ ) → ℓ2 ,
n̄ 7→
k
X
eni ,
i=1
where (en )∞
n=1 denotes the canonical basis of ℓ2 . Then
p
(4.6)
kfk (m̄) − fk (n̄)k2 = 2dJ (m̄, n̄).
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
17
For our claim in the uniform category define for k ∈ N
k
1 X
dJ
→ Bℓ2 , n̄ 7→ √
gk : [N]k ,
eni ,
k
k
i=1
then
r
2dJ (m̄, n̄)
,
k
and we again invoke Corollary 4.5 to conclude the proof of our claim.
(4.7)
kgk (m̄) − gk (n̄)k2 =
Reflexivity is usually not preserved under nonlinear embeddings, as witnessed by Ribe’sP
example in [35], where
P∞ he showed that the two separable
ℓ
Banach spaces ( ∞
)
and
(
n=1 pn ℓ2
n=1 ℓpn )ℓ2 ⊕ ℓ1 , where (pn )n≥1 is a sequence in (1, ∞) such that limn→∞ pn = 1, are uniformly homeomorphic.
Using the Kalton-Randrianarivony concentration inequality, it was shown
in [5] that if X coarse Lipschitz embeds into a reflexive Banach space that
is asymptotically uniformly smooth then X must be reflexive. Actually
X must have the Banach-Saks property as recently observed in [9]. Another consequence of the concentration inequality gives that reflexivity is
stable under coarse embeddability into Banach spaces possessing the “c0 like” asymptotic smoothness property of Tsirelson’s space T ∗ . This tradeoff, in order to preserve reflexivity, between faithfulness of the embedding
and smoothness properties of the target space is a rather interesting phenomenon.
Proposition 4.7. Let X be a Banach space. If X coarsely embeds, or BX
uniformly embeds, into T ∗ then X is reflexive.
Proof. Assume that X is not reflexive. Then, by James’ characterization of
reflexive spaces [17] there exists a sequence (xn )∞
n=1 ⊂ BX such that for all
k ≥ 1 and n̄ ∈ [N]2k ,
(4.8)
k
X
i=1
xn i −
2k
X
i=k+1
xn i ≥
k
.
2
P
For every k ∈ N, the map ϕk : Hkω (N) → X defined as ϕk (n̄) = ki=1 xni , for
n̄ ∈ [N]k , is clearly 2-Lipschitz. Assume by contradiction that there exists
a coarse embedding f : X → T ∗ . It follows again from the properties of
graph distances that f ◦ ϕk is ωf ◦ϕk (1)-Lipschitz. Since ωf ◦ϕk (1) ≤ ωf (2),
it follows from Theorem 4.4 that for all k ≥ 1, there exists Mk ∈ [N]ω , such
that kf ◦ ϕk (m̄) − f ◦ ϕk (n̄)k ≤ 5ωf (2), for all m̄, n̄ ∈ [Mk ]k . In particular
for all k ≥ 1, if m̄, n̄ ∈ [Mk ]k are such that m1 < m2 < · · · < mk < n1 <
n2 < · · · < nk , we obtain by (4.8) that ρf ( k2 ) ≤ 5ωf (2), which is impossible
for large enough k.
For the proof in the uniform setting, we use the map ϕk : Hkω → BX
P
defined as ϕk (n̄) = k1 ki=1 xni , which is clearly k2 -Lipschitz. Assume by
contradiction that there exists a uniform embedding f : BX → T ∗ . Then,
Lip(f ◦ ϕk ) ≤ ωf ( k2 ). By Theorem 4.4, for all k ≥ 1, there exist Mk ∈ [N]ω
such that kf ◦ ϕk (m̄) − f ◦ ϕk (n̄)k ≤ 5ωf ( k2 ), for all m̄, n̄ ∈ [Mk ]k . In
particular for all k ≥ 1, again if m̄, n̄ ∈ [Mk ]k are such that m1 < m2 <
18
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
· · · < mk < n1 < n2 < · · · < nk , we get by (4.8) that 0 < ρf ( 21 ) ≤ 5ωf ( k2 ),
which is again impossible if k is sufficiently large.
Remark 4.8. The conclusion of Proposition 4.7 can be slightly strengthened
as already observed in [9]. Indeed, using [7, Proposition 2, p. 273] one can
show that X has the Banach-Saks property.
Recall that Tsirelson’s space T ∗ has the following important properties:
• T ∗ is reflexive.
• T ∗ does not contain isomorphic copies of ℓp , for any p ∈ [1, ∞) nor
of c0 .
• All the spreading models of T ∗ are isomorphic to c0 .
Theorem A shows that Tsirelson’s construction is surprisingly rigid.
Proof of Theorem A. Assume first that g : X → T ∗ is a coarse embedding.
By Corollary 4.5 the sequence (Hkω )k≥1 does not equi-coarsely embed into X.
Since Proposition 4.7 insures that X is reflexive, it follows from Rosenthal’s
ℓ1 theorem that every spreading model of X can be generated by a weakly
null sequence. Finally it follows from Corollary 3.3 that every spreading
model of X is isomorphic to c0 .
Assume now that g : BX → T ∗ is a uniform embedding. Assume also
by contradiction that X admits a spreading model which is not isomorphic
to c0 . By Proposition 4.7, X is reflexive. Thus, it follows from Rosenthal’s
ℓ1 theorem and Proposition 2.4 that there exists a normalized weakly null
sequence (xn )∞
n=1 in X such that limk→∞ ψ(k) = ∞, where
(4.9)
ψ(k) =
inf
inf
n1 <n2 <···<nk (εi )k ∈{−1,1}
i=1
k
X
εi xni .
i=1
Define now ϕk : ([N]k , dkJ ) → T ∗ by
k
1 X
xni , for n̄ ∈ [N]k .
ψ(2k)
i=1
2
Note that ωϕk k1 ) ≤ ωg ( ψ(2k)
. For every k ∈ N there exists Mk ∈ [N]ω by
Theorem 4.4, such that
2
(4.11)
kϕk (m̄) − ϕk (n̄)k ≤ 5ωg
, whenever m̄, n̄ ∈ [Mk ]k .
ψ(k)
(4.10)
ϕk (n̄) = g
But, it follows from the definition of ψ that kϕk (m̄) − ϕk (n̄)k ≥ ρg (1),
2
whenever dJ (m̄, n̄) = k. So for all k ∈ N, we have 5ωg ( ψ(k)
) ≥ ρg (1) > 0.
This contradicts the fact that limk→∞ ψ(k) = ∞ and limt→0 ωg (t) = 0.
5. Final comments and open problems
Since T ∗ clearly has trivial cotype, it follows from Mendel-Naor metric
cotype obstruction that T ∗ is not coarsely embeddable into any Banach space
that coarsely embeds into a Banach space with non-trivial type. Therefore,
T ∗ and ℓp , p ∈ [1, ∞), are coarsely incomparable in the sense that T ∗ is
not coarsely embeddable into ℓp and ℓp is not coarsely embeddable into
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
19
T ∗ . To the best of our knowledge this provides the first pairs of coarsely
incomparable Banach spaces.
Corollary 5.1. Assume that X is a Banach space such that
(1) (Hkω )k≥1 (or (Jkω )k≥1 ) equi-coarsely embeds into X,
(2) X coarsely embeds into a Banach space with non-trivial type.
Then, X and T ∗ are coarsely incomparable.
Let us conclude with a few open questions.
Problem 5.1. Does ℓ2 coarsely embed into every super-reflexive Banach
space?
In view of Ostrovskii’s result [33], which we referenced in the introduction,
a counter-example to Problem 5.1 would have to be a Banach space that is
super-reflexive not containing any unconditional basic sequence. It seems
that the only known such space is Ferenczi’s space [12].
Problem 5.2. Does ℓ2 coarsely embed into any Banach space which has a
spreading model that is not equivalent to c0 ?
Acknowledgements. This work was completed while the second named
author was visiting Texas A&M University in College Station. He wishes to
thank the Mathematics Department of Texas A&M University for its warm
hospitality and the excellent working environment.
References
[1] I. Aharoni, Every separable metric space is Lipschitz equivalent to a subset of c+
0 ,
Israel J. Math. 19 (1974), 284–291.
[2] D. Alspach, R. Judd, and E. Odell, The Szlenk index and local l1 -indices, Positivity
9 (2005), no. 1, 1–44.
[3] F. Baudier, Metrical characterization of super-reflexivity and linear type of Banach
spaces, Arch. Math. 89 (2007), 419–429.
[4]
, Quantitative nonlinear embeddings into Lebesgue sequence spaces, J. Topol.
Anal. 8 (2016), no. 1, 117-150.
[5] F. Baudier, N. J. Kalton, and G. Lancien, A new metric invariant for Banach spaces,
Studia Math. 199 (2010), 73-94.
[6] F. Baudier and G. Lancien, Tight embeddability of proper and stable metric spaces,
Anal. Geom. Metr. Spaces 3 (2015), 140–156.
[7] B. Beauzamy, Banach-Saks properties and spreading models, Math. Scand. 44 (1979),
357–384.
[8] B. Beauzamy and J.-T. Lapresté, Modèles étalés des espaces de Banach, Travaux en
Cours. [Works in Progress], Hermann, Paris, 1984.
[9] B. M. Braga, Asymptotic structure and coarse Lipschitz geometry of Banach spaces,
Studia Math. 237 (2017), no. 1, 71–97.
[10] A. Brunel and L. Sucheston, On B-convex Banach spaces, Math. Systems Theory 7
(1974), 294–299.
[11] A. Dvoretzky, Some results on convex bodies and Banach spaces, Proc. Internat.
Sympos. Linear Spaces (Jerusalem, 1960), 1961, pp. 123–160.
[12] V. Ferenczi, A uniformly convex hereditarily indecomposable Banach space, Israel J.
Math. 102 (1997), 199–225.
[13] S. C. Ferry, A. Ranicki, and J. Rosenberg, A history and survey of the Novikov
conjecture, Novikov conjectures, index theorems and rigidity, Vol. 1 (Oberwolfach,
1993), 1995, pp. 7–66.
20
F. BAUDIER, G. LANCIEN, AND TH. SCHLUMPRECHT
[14] T. Figiel and W. B. Johnson, A uniformly convex Banach space which contains no
lp , Compositio Math. 29 (1974), 179–190.
[15] G. Godefroy, G. Lancien, and V. Zizler, The non-linear geometry of Banach spaces
after Nigel Kalton, Rocky Mountain J. Math. 44 (2014), no. 5, 1529–1583.
[16] R. C. James, Bases and reflexivity of Banach spaces, Ann. of Math. (2) 52 (1950),
518–527.
[17]
, Some self-dual properties of normed linear spaces, Symposium on InfiniteDimensional Topology (Louisiana State Univ., Baton Rouge, La., 1967), 1972,
pp. 159–175. Ann. of Math. Studies, No. 69.
[18] W. B. Johnson and N. L. Randrianarivony, lp (p > 2) does not coarsely embed into a
Hilbert space, Proc. Amer. Math. Soc. 134 (2006), 1045–1050.
[19] N. J. Kalton, Coarse and uniform embeddings into reflexive spaces, Quart. J. Math.
(Oxford) 58 (2007), 393–414.
[20]
, Lipschitz and uniform embeddings into ℓ∞ , Fund. Math. 212 (2011), no. 1,
53–69.
[21] N. J. Kalton and N. L. Randrianarivony, The coarse Lipschitz structure of ℓp ⊕ ℓq ,
Math. Ann. 341 (2008), 223–237.
[22] N. J. Kalton and D. Werner, Property (M ), M -ideals, and almost isometric structure
of Banach spaces, J. Reine Angew. Math. 461 (1995), 137–178.
[23] G. Kasparov and G. Yu, The coarse geometric Novikov conjecture and uniform convexity, Adv. Math. 206 (2006), 1–56.
[24] M. Mendel and A. Naor, Metric cotype, Ann. of Math.(2) 168 (2008), 247–298.
[25] A. Naor, An introduction to the Ribe program, Jpn. J. Math. 7 (2012), no. 2, 167–233.
[26]
, Discrete Riesz transforms and sharp metric Xp inequalities, Ann. of Math.
(2) 184 (2016), no. 3, 991–1016.
[27] A. Naor and G. Schechtman, Pythagorean powers of hypercubes, Ann. Inst. Fourier
(Grenoble) 66 (2016), no. 3, 1093–1116.
[28]
, Metric Xp inequalities, Forum Math. Pi 4 (2016), e3, 81pp.
[29] Nowak P. W., On coarse embeddability into lp -spaces and a conjecture of Dranishnikov, Fund. Math. 189 (2006), no. 2, 111–116.
[30] P. Nowak and G. Yu, Large scale geometry, EMS Textbooks in Mathematics, European Mathematical Society (EMS), Zürich, 2012.
[31] E. Odell and T. Schlumprecht, The distortion problem, Acta Math. 173 (1994), 259–
281.
[32] E. Odell, Th. Schlumprecht, and A. Zsák, Banach spaces of bounded Szlenk index,
Studia Math. 183 (2007), no. 1, 63–97.
[33] M. I. Ostrovskii, Coarse embeddability into Banach spaces, Topology Proc. 33 (2009),
163–183.
[34]
, Metric embeddings, De Gruyter Studies in Mathematics, vol. 49, De Gruyter,
Berlin, 2013. Bilipschitz and coarse embeddings into Banach spaces.
[35] M. Ribe, Existence of separable uniformly homeomorphic nonisomorphic Banach
spaces, Israel J. Math. 48 (1984), 139–147.
[36] J. Rosenberg, Novikov’s conjecture, Springer, [Cham], 2016.
[37] B. S. Tsirel′son, It is impossible to imbed lp of c0 into an arbitrary Banach space,
Funkcional. Anal. i Priložen. 8 (1974), 57–60 (Russian).
[38] A. Valette, Introduction to the Baum-Connes conjecture, Lectures in Mathematics
ETH Zürich, Birkhäuser Verlag, Basel, 2002. From notes taken by Indira Chatterji,
With an appendix by Guido Mislin.
[39] G. Yu, The coarse Baum-Connes conjecture for spaces which admit a uniform embedding into Hilbert space, Invent. Math. 139 (2000), 201–240.
THE COARSE GEOMETRY OF TSIRELSON’S SPACE AND APPLICATIONS
21
F. Baudier, Department of Mathematics, Texas A&M University, College
Station, TX 77843, USA
E-mail address: [email protected]
G. Lancien, Laboratoire de Mathématiques de Besançon, CNRS UMR-6623,
Université Bourgogne Franche-Comté, 16 route de Gray, 25030 Besançon
Cédex, France
E-mail address: [email protected]
Th. Schlumprecht, Department of Mathematics, Texas A&M University,
College Station, TX 77843-3368, USA, and Faculty of Electrical Engineering, Czech Technical University in Prague, Zikova 4, 16627, Prague, Czech
Republic
E-mail address: [email protected]
| 4math.GR
|
BOA for Nurse Scheduling
Li J and Aickelin U (2006): 'Bayesian Optimisation Algorithm for Nurse Scheduling', Scalable
Optimization via Probabilistic Modeling: From Algorithms to Applications (Studies in Computational
Intelligence), edited by M Pelikan, K Sastry and E Cantú-Paz, Chapter 17, pp 315-332, Springer,
Jingpeng Li and Uwe Aickelin
{jpl, uxa}cs.nott.ac.uk, School of Computer Science and IT, University of Nottingham, Nottingham, NG8
1BB, United Kingdom
Abstract: Our research has shown that schedules can be built mimicking a human scheduler by using a set
of rules that involve domain knowledge. This chapter presents a Bayesian Optimization Algorithm (BOA)
for the nurse scheduling problem that chooses such suitable scheduling rules from a set for each nurse’s assignment. Based on the idea of using probabilistic models, the BOA builds a Bayesian network for the set
of promising solutions and samples these networks to generate new candidate solutions. Computational results from 52 real data instances demonstrate the success of this approach. It is also suggested that the
learning mechanism in the proposed algorithm may be suitable for other scheduling problems.
Keywords: Nurse Scheduling, Bayesian Optimization Algorithm, Probabilistic Modeling.
1 Introduction
Personnel scheduling problems have been studied extensively over the past three decades (see survey papers by Baker 1976; Tien and Kamiyama 1982; Bradley and Martin 1990, Bechtold et al. 1991; Ernst et al.,
2004). Personnel scheduling is the problem of assigning employees to shifts or duties over a scheduling
period so that certain constraints (organizational and personal) are satisfied. It involves constructing a
schedule for each employee within an organization in order for a set of tasks to be fulfilled. In the domain
of healthcare, this is particularly challenging because of the presence of a range of different staff requirements on different days and shifts. In addition, unlike many other organizations, healthcare institutions
work twenty-four hours a day for every single day of the year. In this chapter, we focus on the development of new approaches for nurse scheduling.
Most nurse scheduling problems are extremely difficult and complex. Tien and Kamiyama (1982), for
example, say nurse scheduling is more complex than the traveling salesman problem. A general overview
of various approaches for nurse scheduling can be found in Sitompul and Randhawa (1990), Cheang et al.
(2003) and Burke et al. (2004). Early research (Warner and Prawda 1972; Miller et al. 1976; Trivedi and
Warner 1976) concentrated on the development of mathematical programming models. To reduce computational complexity, researchers had to restrict the problem dimensions and consider a smaller size of constraints in their models, resulting in solutions that are too simple to be applied in real hospital situations.
The above observations have led to other attempts, trying to solve the real nurse scheduling problems
within reasonable time. Besides heuristic approaches (e.g., Smith et al. 1979; Blau and Sear 1983), artifi-
cial intelligence approaches such as constraint programming (Okada and Okada 1988), expert systems
(Chen and Yeung 1993) and knowledge based systems (Scott and Simpson 1998) were investigated with
some success. Since the 1990’s, most papers have tackled the problem with metaheuristic approaches, including simulated annealing (Isken and Hancock 1991), tabu search (Dowsland 1998) and genetic algorithms (Aickelin and Dowsland 2003).
This chapter solves a nurse scheduling problem arising at a major UK hospital (Aickelin and Dowsland
2000; Dowsland and Thompson 2000). In this problem, the number of nurses is fixed (up to 30) and the
target is to create a weekly schedule by assigning each nurse one out of up to 411 predefined shift patterns
in the most efficient way. Due to the limitation of the traditional mathematical programming approach, a
number of meta-heuristic approaches have been explored for this problem. For example, in (Aickelin and
Dowsland 2000, 2003; Aickelin and White 2004) various approaches based on Genetic Algorithms (GAs)
are presented, in (Li and Aickelin 2004) an approach based on a learning classifier system is investigated,
in (Burke et al. 2003) a tabu search hyperheuristic is introduced, and in (Li and Aickelin 2003) a Bayesian
Optimization Algorithm (BOA) is described. In this chapter, we demonstrate a novel BOA approach.
In our proposed BOA, we try to solve the problem by applying a suitable rule, from a set containing a
number of available rules, for each nurse’s assignment. A schedule can be then created from a rule string
(or a rule sequence) corresponding to nurses from the first to the last. To evolve the rule strings Bayesian
networks (Pearl 1988) are used. The Bayesian network in our case is a directed acyclic graph with each
node corresponding to a nurse/rule pair, by which a schedule will be constructed step by step. The causal
relationship between two nodes is represented by a directed edge between the two nodes.
Based on the idea of using probabilistic models, the BOA builds a Bayesian network for the set of promising solutions and generates new candidate solutions by sampling these networks. (Pelikan et al. 1999; Pelikan 2005). In our proposed BOA for nurse scheduling, the conditional probabilities of all nurse/rule pairs
in the network are first computed according to an initial set of rule strings. Subsequently, new instances of
each node are generated by using conditional probabilities to obtain new rule strings. A new set of rule
strings is thus generated, some of which will replace previous strings based on roulette-wheel fitness selection. If the stopping criterion is not reached, the conditional probabilities for all nodes in the network are
updated again using the current set of rule strings.
2 The Nurse Scheduling Problem
The problem described in this chapter is that of creating weekly schedules for wards containing up to 30
nurses at a major UK hospital. These schedules must respect working contracts and meet the demand for a
given number of nurses of different grades on each shift, while being perceived to be fair by the nurse
themselves. The day is partitioned into three shifts: two day shifts called ‘earlies’ and ‘lates’, and a longer
night shift. Until the final scheduling stage, ‘earlies’ and ‘lates’ are merged into day shifts. Due to hospital
policy, a nurse would normally work either days or nights in a given week, and due to the difference in
shift length, a full week’s work normally includes more days than nights. For example, a full-time nurse
works 5 days or 4 nights, whereas a part-time nurse works 4 days or 3 nights, 3 days or 3 nights, or 3 days
or 2 nights. However, exceptions are possible and some nurses specifically must work both day- and nightshifts in one week.
As described in (Dowsland and Thompson 2000), the problem can be decomposed into three independent stages. The first stage uses a knapsack model to check if there are enough nurses to meet demand. If
not, additional nurses are allocated to the ward, so that the second stage will always admit a feasible solution. The second stage is the most difficult and is concerned with the actual allocations of days or nights to
be worked by each nurse. The third stage then uses a network flow model to assign those on days to ‘earlies’ and ‘lates’. The first and the third stages, and in this chapter we only concern with the highly constrained second stage.
The numbers of days or nights to be worked by each nurse defines the set of feasible weekly work patterns for that nurse. These will be referred to as shift patterns in the following. For example, (1111100
0000000) would be a pattern where the nurse works the first 5 days and no nights. Depending on the working hours of a nurse, there are a limited number of shift patterns available to her/him. Typically, there will
be between 20 and 30 nurses per ward, 3 grade-bands, 9 part time options and 411 different shift patterns.
Depending on the nurses’ preferences, the recent history of patterns worked, and the overall attractiveness
of the pattern, a penalty cost is allocated to each nurse-shift pattern pair. These values were set in close
consultation with the hospital and range from 0 (perfect) to 100 (unacceptable), with a bias to lower values.
Further details can be found in (Dowsland 1998).
This second stage problem is described as follows. Each possible weekly shift pattern for a given nurse
can be represented as a 0-1 vector with 14 elements, where the first 7 elements represent the 7 days of the
week and the last 7 elements the corresponding 7 nights of the week. A ‘1’/’0’ in the vector denotes a
scheduled day or night ‘on’/’off’. For each nurse there are a limited number of shift patterns available to
5
her/him. For instance, a full-time nurse working either 5 days or 4 nights has a total of 21 (i.e. C 7 ) feasi-
4
ble day shift patterns and 35 (i.e. C 7 ) feasible night shift patterns. Typically, there are between 20 and 30
nurses per ward, 3 grade-bands, 9 part time options and 411 different shift patterns. Depending on the
nurses’ preferences, the recent history of patterns worked, and the overall attractiveness of the pattern, a
penalty cost is allocated to each nurse-shift pattern pair. These values were set in close consultation with
the hospital and range from 0 (perfect) to 100 (unacceptable), with a bias to lower values. Further details
can be found in (Dowsland 1998).
This problem can be formulated as follows. The decision variable xij assumes 1 if nurse i works shift
pattern j and 0 otherwise. Let parameters m, n, p be the number of shift patterns, nurses and grades respectively. Parameter ajk is 1 if shift pattern j covers day/night k, 0 otherwise. qis is 1 if nurse i is of grade s or
higher, 0 otherwise. Furthermore, pij is the preference cost of nurse i working shift pattern j, and F(i) is the
set of feasible shift patterns for nurse i. Then we can use the following integer programming formulation
from (Dowsland 1998):
n
m
∑ ∑p
Min
s.t.
i =1 j∈F ( i )
ij
xij
(1)
∑ xij = 1, ∀i = {1,..., n}
j∈F ( i )
,
n
∑ ∑q
j∈F ( i ) i =1
is
(2)
a jk xij ≥ Rks , ∀k = {1,...,14}, s = {1,..., p}
.
(3)
xij ∈ {0,1}, ∀i, j.
The objective function (1) is to minimize total preference cost of all nurses. Constraint (2) ensures that
every nurse works exactly one shift pattern from his/her feasible set, and constrain (3) ensures that the demand for nurses is fulfilled for every grade on every day and night. This problem can be regarded as a
multiple-choice set-covering problem. The sets are given by the shift pattern vectors and the objective is to
minimize the cost of the sets needed to provide sufficient cover for each shift at each grade. The constraints described in (2) enforce the choice of exactly one pattern (set) from the alternatives available for
each nurse.
3 A BOA for Nurse Scheduling
In the nurse scheduling problem we are tackling, the number of nurses is fixed (approximately 30 depending on data instance), and the goal is to create weekly schedules by assigning each nurse one shift pattern in
the most efficient way. The problem can be solved by using a suitable rule, from a rule set that contains a
number of available rules, for each nurse’s assignment. Thus, a potential solution is represented as a rule
string, or a rule sequence connecting all nurses.
We chose this rule-base building approach, as the longer-term aim of our research is to model the explicit learning of a human scheduler. Human schedulers can provide high quality solutions, but the task is
tedious and often requires a large amount of time. Typically, they construct schedules based on rules learnt
during scheduling. Due to human limitations, these rules are typically simple. Hence, our rules will be
relatively simple, too. Nevertheless, human generated schedules are of high quality due to the ability of the
scheduler to switch between the rules, based on the state of the current solution. We envisage the proposed
BOA to perform this role.
3.1 The Construction of a Bayesian Network
Bayesian networks are also called directed graphical models, in which each node corresponds to one variable, and each variable corresponds to one position in the strings representing the solutions. The relationship between two variables is represented by a directed edge between the two corresponding nodes.
Bayesian networks are often used to model multinomial data with both discrete and continuous variables
by encoding the relationship between the variables contained in the modeled data, which represents the
N11
N12
...
N1,n
N21
N22
...
N2,n
N31
N32
...
N3,n
...
...
...
...
Nm-1,1
Nm-1,2
...
Nm-1,n
Nm,1
Nm,2
...
Nm,n
structure of a problem. Furthermore, they are used to generate new data instances or variables instances
with similar properties as those of given data.
Fig. 1. A Bayesian network for nurse scheduling
Figure 1 is the Bayesian network constructed for our nurse scheduling problem, which is a hierarchical
and acyclic directed graph representing the solution structure of the problem.
The node
N ij (i ∈ {1,2,..., m}; j ∈ {1,2,..., n})
in the network is a nurse/rule pair which denotes that nurse i is assigned by
rule j, where m is the number of nurses and n is the number of available rules. The directed edge from node
Nij to node Ni+1,j’ denotes a causal relationship of “Nij causing Ni+1,j’”, i.e. if nurse i is scheduled by rule j
then the next nurse will be scheduled by rule j’. In this network, a potential solution is shown as a directed
path from nurse 1 to nurse m connecting m nodes.
3.2 Learning based on the Bayesian Network
In BOAs, the structure of the Bayesian network can be either fixed (Pelikan et al 1999) or variable
(Mühlenbein and Mahnig 1999). In our proposed nurse scheduling model, we use a fixed network structure
because all variables are fully observed. In essence, our Bayesian network is a fixed nurse-size vector of
rules and the goal of learning is to find the variable values of all nodes Nij that maximize the likelihood of
the training data containing a number of independent cases.
Thus, the learning in our case amounts to ‘counting’ based on a multinomial distribution and we use the
symbol ‘#’ meaning ‘the number of’ in the following equations. It calculates the conditional probabilities
of each possible value for each node given all possible values of its parent nodes. For example, for node
Ni+1,j’ with a parent node Nij, its conditional probability is
P( N i +1, j ′ | N ij ) =
P( N i +1, j ′ , N ij )
P ( N ij )
=
# ( N i +1, j ′ = true, N ij = true)
# ( N i +1, j ′ = true, N ij = true)+ # ( N i +1, j ′ = false, N ij = true)
.
Note that nodes N1j have no parents. In this circumstance, their probabilities are computed as
P( N 1 j ) =
# ( N 1 j = true)
= true) + # ( N
# (N
= false )
=
(4)
# ( N 1 j = true)
# Training sets
1j
1j
.
(5)
To help the understanding of how these probabilities are computed, let us consider a simple example of
using three rules to schedule three nurses (shown in Figure 2). The scheduling process is repeated 50
times, and each time, rules are randomly used to get a solution, disregarding the feasibility of the solution.
The Arabic numeral adjacent to each edge is the total number of times that this edge has been used in the
50 runs. For instance, if a solution is obtained by using rule 2 to schedule nurse 1, rule 3 to nurse 2 and rule
1 to nurse 3, then there exists a path of “N12N23N31”, and the count of edge “N12 N23” and edge
“N23N31” are increased by one respectively.
N11
5
2
10
4
7
11
N21
5
9
5
11
1
10
4
N32
N31
3
N23
N22
3
7
N13
N12
3
0
N33
Fig. 2. A dummy problem with three nurses and three rules
Therefore, we can calculate the probabilities of each node at different states according to the above
count. For the nodes that have no parents, their probabilities are computed as:
P( N 11 ) =
10 + 2 + 3
15
20
15
=
P( N 12 ) =
P( N 13 ) =
(10 + 2 + 3) + (5 + 11 + 4) + (7 + 5 + 3) 50 ,
50 ,
50 .
For all other nodes (with parents), the conditional probabilities are:
10
10
2
3
=
P( N 22 | N11 ) =
P( N 23 | N11 ) =
10 + 2 + 3 15 ,
15 ,
15 ,
10
4
5
5
P( N 22 | N12 ) =
P( N 23 | N12 ) =
P( N 21 | N12 ) =
=
20 ,
20 ,
5 + 11 + 4 20 ,
5
3
7
7
P( N 22 | N 13 ) =
P( N 23 | N13 ) =
P( N 21 | N 13 ) =
=
15 ,
15 ,
7 + 5 + 3 15 ,
9
3
7
7
P( N 32 | N 21 ) =
P( N 33 | N 21 ) =
P( N 31 | N 21 ) =
=
19 ,
19 ,
7 + 9 + 3 19 ,
1
5
11
11
P( N 32 | N 22 ) =
P( N 33 | N 22 ) =
P( N 31 | N 22 ) =
=
17 ,
17 ,
11 + 1 + 5 17 ,
10
10
4
0
P( N 31 | N 23 ) =
=
P( N 32 | N 23 ) =
P( N 33 | N 23 ) =
10 + 4 + 0 14 ,
14 .
14 ,
P( N 21 | N 11 ) =
These probability values can be used to generate new rule strings, or new solutions. Since the first rule
in a solution has no parents, it will be chosen from nodes N1j according to their probabilities. The next rule
will be chosen from nodes Nij according to the probabilities conditioned on the previous nodes. This build-
ing process is repeated until the last node Nmj has been chosen, where m is number of the nurses. A path
from nurse 1 to nurse m is thus formed, representing a new potential solution. Since all probability values
for each nurse are normalized, we suggest the roulette-wheel method as a suitable strategy for rule selection
(Goldberg 1989).
For further clarity, consider the following example of scheduling five nurses with two rules (1: random
allocation, 2: allocate nurse to low-cost shifts). In the beginning of the search, the probabilities of choosing
rule 1 or rule 2 for each nurse is equal, i.e. 50%. After a few iterations, due to the selection pressure and reinforcement learning, we experience two solution pathways: Because pure low-cost or random allocation
produces low quality solutions, either rule 1 is used for the first 2-3 nurses and rule 2 on remainder or vice
versa. Therefore for this simple example, the Bayesian network learns ‘use rule 2 after 2-3 times of using
rule 1’ or vice versa.
3.3 Our BOA Approach
Based on the estimation of conditional probabilities, this section introduces a BOA for nurse scheduling. It
uses techniques from the field of modeling data by Bayesian networks to estimate the joint distribution of
promising solutions. The nodes, or variables, in the Bayesian network correspond to the individual
nurse/rule pairs by which a schedule will be built step by step.
The conditional probability of each variable in the Bayesian network is computed according to an initial
set of promising solutions. Subsequently, each new instance for each variable is generated by using the corresponding conditional probabilities, until all variables have been generated. Hence, in our case, a new
rule string has been obtained. Another set of rule strings will be generated in this way, some of which will
replace previous strings based on fitness selection. If the stopping criterion not reached, the conditional
probabilities for all nodes in the Bayesian network are updated again using the current set of promising rule
strings. In more detail, the steps of our BOA for nurse scheduling are described as follows:
1. Set t = 0, and generate an initial population P(0) at random;
2. Use roulette-wheel to select a set of promising rule strings S(t) from P(t);
3. Compute the conditional probabilities for the values of each node according to this set of promising
solutions;
4. For each nurse’s assignment, use the roulette-wheel method to select one rule according to the conditional probabilities of all available nodes for this nurse, thus obtaining a new rule string. A set of new
rule strings O(t) will be generated in this way;
5. Create a new population P(t+1) by replacing some rule strings from P(t) with O(t), and set t = t+1;
6. If the termination conditions not met, go to step 2.
3.4 Four Heuristic Rules for Solution Building
Using our domain knowledge of nurse scheduling, the BOA can choose from the following four rules. The
first two rules are quite simple, while the rest twos are a bit complex. Note that there are many more heuristic rules that could be used to build schedules.
1) ‘Random’ Rule
This ‘Random’ rule is used to assign shift patterns to individual nurses in a totally random manner. Its purpose is to introduce randomness into the search thus enlarging the search space, and most importantly to
ensure that the proposed algorithm has the ability to escape from local optima. This rule mirrors much of a
scheduler’s creativeness to come up with different solutions if required.
2) ‘k-Cheapest’ Rule
The second ‘k-Cheapest’ rule is purely towards the solution cost, and disregards the feasibility of the solution. For each nurse’s assignment, it randomly selects one candidate shift pattern from a set containing the
cheapest k shift patterns available to that nurse, in an effort to reduce the solution cost of a schedule as
much as possible.
3) ‘Overall Cover’ Rule
The ‘Overall Cover’ rule is designed to consider only the feasibility of the schedule. It schedules one nurse
at a time in such a way as to cover those days and nights with the most number of uncovered shifts.
This rule constructs solutions as follows. For each shift pattern in a nurse’s feasible set, it calculates the
total number of uncovered shifts that would be covered if the nurse worked that shift pattern. For instance,
assume that a shift pattern covers Monday to Friday nights. Further assume that the current requirements
for the nights from Monday to Sunday are as follows: (-4, 0, +1, -3, -1, -2, 0), where a negative number
means undercover and a positive overcover. The Monday to Friday shift pattern hence has a cover value of
8 as the sum of undercover is -8. In this example, a Tuesday to Saturday pattern would have a cover value
of 6 as the sum of undercover is -6.
For nurses of grade s, only the shifts requiring grade s nurses are counted as long as there is a single uncovered shift for this grade. If all these are covered, shifts of the next lower grade are considered and once
these are filled those of the next lower grade etc. This operation is necessary as otherwise higher graded
nurses might fill lower graded demand, whilst higher graded demand might not be met at all.
Due to the nature of this rule, nurses’ preference costs pij are not taken into account. Therefore, the
‘Overall Cover’ rule can be summarized as assigning the shift pattern with the largest amount of undercover for the nurse currently being scheduling.
4) ‘Contribution’ Rule
The fourth ‘Contribution’ rule is biased towards solution quality but includes some aspects of feasibility. It
cycles through all shift patterns of a nurse, assigns each one a score, and chooses the one with the highest
score. In case of more than one shift pattern with the best score, the first such shift pattern is chosen.
Compared with the third ‘Overall Cover’ rule, this ‘Contribution’ rule is more complex because it considers
the preferences of the nurses and also tries to look ahead a little.
The score of a shift pattern is calculated as the weighted sum of the nurse’s pij value for that particular
shift pattern and its contribution to the cover of all three grades. The latter is measured as a weighted sum
of grade one, two and three uncovered shifts that would be covered if the nurse worked this shift pattern,
i.e. the reduction in shortfall. Obviously, nurses can only contribute to uncovered demand of their own
grade or below.
More precisely, using the same notation as before, the score Sij of shift pattern j for nurse i is calculated
as
3
S ij = w p (100 − Pij ) +
14
∑ w q (∑ a
s
s =1
is
jk
d ks )
k =1
(6)
where parameter wp is the weight of the nurse’s pij value for the shift pattern and parameter ws is the
weight of covering an uncovered shift of grade s. Variable ajk is 1 if shift pattern j covers day k, 0 otherwise. Variable dks is 1 if there are still nurses needed on day k of grade s, 0 otherwise. Note that (100−pij)
must be used in the score, as higher pij values are worse and the maximum for pij is 100.
3.5 Fitness Function
Our nurse scheduling problem is complicated by the fact that higher qualified nurses can substitute less
qualified nurses but not vice versa. Furthermore, the problem has a special day-night structure as most of
the nurses are contracted to work either days or nights in one week but not both. These two characteristics
make the finding and maintaining of feasible solutions in any heuristic search is extremely difficult. Therefore, a penalty function approach is needed while calculating the fitness of completed solutions. Since the
chosen encoding automatically satisfies constraint set (2) of the integer programming formulation, we can
use the following formula to calculate the fitness of solutions (the fitter the solution, the lower its fitness
value):
n
m
∑∑ p
Min
i =1 j =1
ij
n m
14 p
xij + wdemand ∑∑ max Rks − ∑∑ qis a jk xij ,0
k =1 s =1
i =1 j =1
.
(7)
Note that only undercovering is penalized not overcovering, hence the use of the max function. The parameter wdemand is the penalty weight used to adjust the penalty that a solution has added to its fitness, and
this penalty is proportional to the number of uncovered shifts. For example, consider a solution with an objective function value of 22 that undercovers the Monday day shift by one shift and the Tuesday night shift
by two shifts. If the penalty weight was set to 20, the fitness of this solution would be 22 + (1+2)*20 = 82.
4 Computational Results
This section describes the computational experiments that are used to test the proposed BOA. For all experiments, 52 real data sets as given to us by the hospital are available. Each data set consists of one
week’s requirements for all shift and grade combinations and a list of available nurses together with their
preference costs pij and qualifications. The data was collected from three wards over a period of several
months and covers a range of scheduling situations, e.g. some data instances have very few feasible solutions whilst others have multiple optima.
For all data instances, we used the following set of fixed parameters to implement our experiments.
These parameters are based on our experience and intuition and thus not necessarily the best for each instance. We have kept them the same for consistency.
• Stopping criterion: number of generations = 200, or an optimal solution is found;
• Population size = 140;
• The number of solutions kept in each generation = 40;
• For the ‘k-Cheapest’ rule, k = 5;
• Weight set in formula (6): wp =1, w1 =8, w2 =2 and w3 =1;
• Penalty weight in fitness function (7): wdemand =200;
• Number of runs per data instance = 20;
The BOA was coded in Java 2, and all experiments were run on the same Pentium 4 2.0GHz PC with
512MB RAM under the Windows XP operating system. To test the robustness of the BOA, each data instance was run twenty times by fixing the above parameters and varying the pseudorandom number seed at
the beginning. The executing time per run and per data instance varies from half second to half a minute
depending on the difficulty of the individual data instance. For example, data instances 27, 29, 31 and 51
are harder to solve than others due to a shortage of nurses in these weeks.
The detailed computational results over these 52 instances are listed in Table 1, in which the last row
(headed with ‘Av.’) contains the mean values of all columns using following notations:
• IP: optimal solutions found with an integer programming software (Dowsland and Thompson, 2000);
• Rd-1: random search, i.e. only the first ‘Random’ rule is in use;
• Rd-2: random rule-selection, i.e. using four rules but every rule has an equal opportunity to be chosen all
the time for all nurses;
• Best: best result out of 20 runs of the BOA;
• Mean: average result of 20 runs of the BOA;
• Fea: number of runs terminating with the best solution being feasible;
• #: number of runs terminating with the best solution being optimal;
• ≤3: number of runs terminating with the best solution being within three cost units of the optimum.
The value of three units was chosen as it corresponds to the penalty cost of violating the least important
level of requests in the original formulation. Thus, these solutions are still acceptable to the hospital.
Table 1. Comparison of results over 52 instances
Data
01
02
03
04
05
IP
8
49
50
17
11
RD1
N/A
N/A
N/A
N/A
N/A
RD2
27
85
97
23
51
Best
8
56
50
17
11
Mean
8.1
74.5
72.0
17.0
12.4
Fea
20
20
20
20
20
#
19
0
2
20
8
≤3
20
0
5
20
16
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
Av.
2
11
14
3
2
2
2
2
3
3
37
9
18
1
7
0
25
0
1
0
48
2
63
15
35
62
40
10
38
35
32
5
13
5
7
54
38
22
19
3
3
3
4
27
107
74
58
21.1
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
51
80
62
44
12
12
47
17
102
9
55
146
73
135
53
19
56
119
4
3
222
158
88
31
210
253
102
30
95
118
130
28
130
44
51
87
188
86
70
34
196
11
35
69
162
197
135
82.9
2
14
15
14
2
2
3
3
4
4
38
9
19
10
7
1
26
1
1
0
52
28
65
109
38
159
43
11
41
46
45
7
25
8
8
55
41
23
24
6
7
3
5
30
109
171
67
29.7
2.4
16.8
17.5
17.3
4.9
2.8
7.8
3.6
6.5
5.1
38.8
11.3
20.8
12.0
8.3
1.6
27.5
1.6
1.0
0.2
66.8
35.6
65.5
169.2
83.5
175.0
80.4
15.6
63.8
60.2
47.6
8.2
33.0
10.8
8.8
56.2
73.3
24.2
77.4
8.1
9.4
3.4
5.7
47.8
175.5
173.6
98.6
39.8
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
20
17
0
0
0
2
2
0
0
0
0
0
4
0
0
5
0
0
0
20
18
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
13
0
0
0
0
0
2.5
17
3
11
0
10
20
2
20
7
20
20
11
20
0
19
20
15
20
20
20
1
0
3
0
3
0
4
8
2
0
0
7
0
3
10
15
1
13
0
2
0
20
10
2
1
0
0
8.5
According to the computational results in Table 1, one can see that using the ‘Random’ rule alone does
not yield a single feasible solution, as the ‘Rd1’ column shows. This underlines the difficulty of this problem. In addition, without learning, the results of randomly selecting one of the four rules at each move are
much weaker, as the ‘Rd2’ column shows. Thus, it is not simply enough to use the four rules to build
schedules that are acceptable to the hospital for practical use. For the proposed BOA, 38 out of 52 data in-
stances are solved to or near to optimality (i.e. within three cost units) and feasible solutions are always
found for all instances.
The behavior of an individual run of the BOA is as expected. Figure 3 depicts the BOA search process
for data instance 04,. In this figure, the x-axis represents the number of generations and the y-axis represents the solution cost of the best individual in each generation. As shown in Figure 3, the optimal solution, with a cost of 17, is produced at the generation of 57. The actual values may differ among various instances, but the characteristic shapes of the curves are similar for all seeds and data instances.
70
Cost
60
50
40
30
20
10
0
10
20
30
40
50
60
70
80
Generation
Fig. 3. a sample run of the BOA
To understanding the results further, we have developed a graphical interface to visualize the BOA’s
learning process, instead of simply outputting the final numerical values from a “black box”. The software
package is developed as a Java applet, which enables the BOA to run real problems online at
http://www.cs.nott.ac.uk/~jpl/BOA/BOA_Applet.html.
In our graphical interface, the conditional probability of choosing a rule for a specific nurse has been
mapped into a 256-gray value and the causal relationship between two nodes (i.e. nurse/rule pairs) is denoted by drawing an edge in this grey value to connect the nodes. For instance, a probability of 50% would
give a grey gradient of 128. The darker the edge, the higher the chance this edge (building block) will be
used.
Please note that in our online demonstration two further rules can be selected. These rules are still experimental and currently under evaluation. The two additional rules are the ‘Highest Cover’ rule and the
‘Enhanced Contribution’ rule. The ‘Highest Cover’ rule is very similar to the ‘Overall Cover’ rule described in 3.4, with the difference in their ways of calculating the undercover: a ‘Highest Cover’ rule is to
find shift patterns which cover those days and nights with the highest number of uncover, while a ‘Overall
Cover’ rule is to find the largest total number of uncovered shifts. The second ‘Enhanced Contribution’
rule is similar to the ‘Contribution’ rule (described in 3.4), and also uses formula (6) to assign each feasible
shift pattern a score. However, this rule uses a different definition of dks: dks is the actual number of required nurses if there are still nurses needed on day k of grade s, 0 otherwise.
Figure 4, 5 and 6 show the graphic results equivalent to the experiments carried out in this paper, i.e.
same parameter selection and 4 rules. The pictures show a single run for the ‘01’ data instance which involves the scheduling of 26 nurses. Figure 4 show the beginning, Figure 5 the intermediate and Figure 6
the final stages of the learning process respectively.
Fig. 4. Graphic display of the probability distributions at the initial generation
Fig. 5. Graphic display of the probability distributions after 50 generations
Fig. 6. Graphic display of the probability distributions after 100 generations
The graphic results in Figure 4, 5 and 6 are in accordance with our hypothesis. In the early process of
the BOA, any edge (i.e. any building block) can be used to explore the solution space as much as possible.
Thus, no particular edge stands out in Figure 4. With the BOA in progress, some edges are becoming more
remarkable, while some ill-fitting paths become diminishing (Shown in Figure 5). Eventually, the BOA
converges to several optimum solutions depicted as clear paths with some common parts, such as the segment of path from node 10 to node 21 (shown in Figure 6). This is particularly true for our problem instances because we know from the hospital that the optimum schedules are often not unique.
Based on the summary results of 20 runs on the same 52 benchmark instances, Table 2 give a comparison of our BOA with some other approaches, which are briefly described as follows:
• LCS: a learning classifier system.
• Basic GA: a GA with standard genetic operators;
• Adaptive GA: the same as the basic GA, but the algorithm also tries to self-learn good parameters during
the runtime;
• Multi-population GA: the same as the adaptive GA, but it also features competing sub-populations;
• Hill-climbing GA; the same as the multi-population GA, but it also includes a local search in the form of
a hill-climber around the current best solution;
• Indirect GA: a novel GA-based approach that first maps the constrained solution space into an unconstrained space, then optimizes within that new space and eventually translates solutions back into the
original space. Up to four different rules and hill-climbers are used in this algorithm.
Table 2. Summary results of various approaches
Algorithm
Reference
Best
Mean
Feasibility
Runtime
BOA
LCS
Optimal IP
Basic GA
Adaptive GA
Multi-population GA
Hill-climbing GA
Indirect GA
Li and Aickelin 2003
Li and Aickelin 2004
Dowsland and Thompson 2000
Aickelin and White 2004
Aickelin and White 2004
Aickelin and White 2004
Aickelin and White 2004
Aickelin and Dowsland 2003
29.7
35.5
21.4
79.8
65.0
37.1
23.2
22.0
39.8
47.7
21.4
88.6
71.4
59.8
44.7
25.5
100%
100%
100%
33%
45%
75%
89%
95%
23.0 secs
44.5 secs
>24 hrs
18.9 secs
23.4 secs
13.4 secs
14.9 secs
11.9 secs
Let us discuss the summary results in Table 2. Compared with the optimal results of the IP approach
which can take more than 24 hours runtime for each instances, the BOA’s results are still more expensive
on average but they are achieved within half minute. Compared with other meta-heuristic approaches
which are all executed quickly, our BOA performs best in terms of feasibility. In terms of solution quality,
in general our BOA performs better than the LCS, the basic GA, the adaptive GA and the multi-population
GA, and performs slightly worse than the hill-climbing GA and the indirect GA. However, it is worth mentioning that the hill-climbing GA includes the features of multiple populations and elaborate local search
which are not available in the proposed BOA, and the indirect GA uses the best nurses’ ordering together
with a local search to produce the final solution while our BOA only uses the ordering given in the original
data sets throughout the search.
5 Conclusions
A BOA is presented for the nurse scheduling problem. Unlike most existing rule-based approaches, the
proposed BOA has the ability to build schedules by using flexible, rather than fixed rules. Experimental results from real-world nurse scheduling problems have demonstrated the strength of the approach.
Although we have presented this work in terms of nurse scheduling, it is suggested that the main idea of
the BOA could be applied to many other scheduling problems where the schedules will be built systematically according to specific rules. It is also hoped that this research will give some preliminary answers
about how to include human-like learning into scheduling algorithms and thus may be of interest to practitioners and researchers in areas of scheduling and evolutionary computation.
Acknowledgements
The work was funded by the UK Government’s major funding agency, the Engineering and Physical Sciences Research Council (EPSRC), under grants GR/R92899/02.
References
Aickelin U, Dowsland K (2000) Exploiting problem structure in a genetic algorithm approach to a nurse
rostering problem. Journal of Scheduling 3: 139-153
Aickelin U, Dowsland K (2003) An indirect genetic algorithm for a nurse scheduling problem. Computers and Operations Research 31: 761-778
Aickelin U, White P (2004) Building better nurse scheduling algorithms. Annals of Operations Research
128: 159-177
Baker K (1976) Workforce allocation in cyclical scheduling problems: a survey. Operational Research
27: 155–167
Bechtold SE, Brusco MJ, Showalter MJ (1991) A comparative evaluation of labor tour scheduling
methods. Decision Sciences 22: 683–699
Blau R, Sear A (1983) Nurse scheduling with a microcomputer. Journal of Ambulance Care Management 6: 1–13
Bradley D, Martin J (1990) Continuous personnel scheduling algorithms: a literature review. Journal of
the Society for Health Systems 2: 8-23
Burke EK, De Causmaecker P, Vanden Berghe G, Van Landeghem H. (2004) The state of the art of
nurse rostering. Journal of Scheduling 7: 441-499
Burke EK, Kendall G, Soubeiga E (2003) A tabu-search hyperheuristic for timetabling and rostering.
Journal of Heuristics 9: 451-470
Cheang B, Li H, Lim A, Rodrigues B (2003) Nurse rostering problems – a bibliographic survey. European Journal of Operational Research 151: 447-460
Chen JG, Yeung T (1993) Hybrid expert system approach to nurse scheduling. Computers in Nursing
11: 183–192
Dowsland K (1998) Nurse scheduling with tabu search and strategic oscillation. European Journal of
Operational Research 106: 393-407
Dowsland K, Thompson JM (2000) Nurse scheduling with knapsacks, networks and tabu search. Journal
of Operational Research Society 51: 825-833
Ernst AT, Jiang H, Krishnamoorthy M, Sier D (2004) Staff scheduling and rostering: a review of applications, methods and models. European Journal of Operational Research 153: 3-27
Goldberg DE (1989) Genetic algorithms in search, optimization and machine leaning. Addison-Wesley
Li J, Aickelin U (2003) A Bayesian optimization algorithm for the nurse scheduling problem. In Proceedings of Congress on Evolutionary Computation. IEEE Press, pp 2149-2156
Li J, Aickelin U (2004) The application of Bayesian optimization and classifier systems in nurse scheduling. In Yao X et al. (eds) Proceedings of the 8th International Conference on Parallel Problem Solving
from Nature. Springer, LNCS 3242: 581-590
Miller HE, Pierskalla W, Rath G (1976) Nurse scheduling using mathematical programming. Operations
Research 24: 857–870
Mühlenbein H, Mahnig T (1999) FDA – a scalable evolutionary algorithm for the optimization of additively decomposed functions. Evolutionary Computation 7: 45-68
Okada M, Okada M (1988) Prolog-based system for nursing staff scheduling implemented on a personal
computer. Computers and Biomedical Research 21: 53–63
Pearl J (1988) Probabilistic reasoning in intelligent systems: networks of plausible inference. Morgan
Kaufmann
Pelikan M (2005) Hierarchical Bayesian optimization algorithm – toward a new generation of evolutionary algorithms. Springer, Berlin Heidelberg New York
Pelikan M, Goldberg D, Cantu-Paz E (1999) BOA: the Bayesian optimization algorithm. IlliGAL Report No 99003, University of Illinois
Scott S, Simpson RM (1998) Case-bases incorporating scheduling constraint dimensions: experiences in
nurse rostering. In Smyth S, Cunningham P (eds) Advances in case-based reasoning, Springer, LNAI 1488:
392–401
Sitompul D, Randhawa S (1990) Nurse scheduling models: a state-of-the-art review. Journal of the Society of Health Systems 2: 62-72
Smith LD, Bird D, Wiggins A (1979) A computerized system to schedule nurses that recognizes staff
preferences. Hospital Health Service Administration 24: 19–35
Tien JM, Kamiyama A (1982) On manpower scheduling algorithms. Society for Industrial and Applied
Mathematics 24: 275–287
Trivedi VM, Warner M (1976) A branch and bound algorithm for optimum allocation of float nurses.
Management Science 22: 972-981
Warner M, Prawda J (1972) A mathematical programming model for scheduling nursing personnel in a
hospital. Management Science 19: 411–422.
| 5cs.CE
|
arXiv:1505.06730v2 [math.AC] 8 May 2016
WEAKLY CLASSICAL PRIME SUBMODULES
HOJJAT MOSTAFANASAB, ÜNSAL TEKIR AND KÜRŞAT HAKAN ORAL
Abstract. In this paper, all rings are commutative with nonzero identity.
Let M be an R-module. A proper submodule N of M is called a classical
prime submodule, if for each m ∈ M and elements a, b ∈ R, abm ∈ N implies
that am ∈ N or bm ∈ N . We introduce the concept of “weakly classical
prime submodules”. A proper submodule N of M is a weakly classical prime
submodule if whenever a, b ∈ R and m ∈ M with 0 6= abm ∈ N , then am ∈ N
or bm ∈ N .
1. Introduction
Throughout this paper all rings are commutative with nonzero identity and all
modules are considered to be unitary. Several authors have extended the notion of
prime ideals to modules, see, for example [10, 13, 14]. Let M be a module over a
commutative ring R. A proper submodule N of M is called prime if for a ∈ R and
m ∈ M , am ∈ N implies that m ∈ N or a ∈ (N :R M ) = {r ∈ R | rM ⊆ N }.
Anderson and Smith [3] said that a proper ideal I of a ring R is weakly prime if
whenever a, b ∈ R with 0 6= ab ∈ I, then a ∈ I or b ∈ I. Weakly prime submodules
were introduced by Ebrahimi Atani and Farzalipour in [11]. A proper submodule
N of M is called weakly prime if for a ∈ R and m ∈ M with 0 6= am ∈ N ,
either m ∈ N or a ∈ (N :R M ). A proper submodule N of M is called a classical
prime submodule, if for each m ∈ M and a, b ∈ R, abm ∈ N implies that am ∈ N or
bm ∈ N . This notion of classical prime submodules has been extensively studied by
Behboodi in [6, 7] (see also, [8], in which, the notion of classical prime submodules
is named “weakly prime submodules”). For more information on classical prime
submodules, the reader is referred to [4, 5, 9].
The annihilator of M which is denoted by AnnR (M ) is (0 :R M ). Furthermore,
for every m ∈ M , (0 :R m) is denoted by AnnR (m). When AnnR (M ) = 0, M
is called a faithful R-module. An R-module M is called a multiplication module if
every submodule N of M has the form IM for some ideal I of R, see [12]. Note that,
since I ⊆ (N :R M ) then N = IM ⊆ (N :R M )M ⊆ N . So that N = (N :R M )M .
Finitely generated faithful multiplication modules are cancellation modules [17,
Corollary to Theorem 9], where an R-module M is defined to be a cancellation
module if IM = JM for ideals I and J of R implies I = J. Let N and K be
submodules of a multiplication R-module M with N = I1 M and K = I2 M for
some ideals I1 and I2 of R. The product of N and K denoted by N K is defined by
N K = I1 I2 M . Then by [2, Theorem 3.4], the product of N and K is independent
of presentations of N and K. Moreover, for m, m′ ∈ M , by mm′ , we mean the
product of Rm and Rm′ . Clearly, N K is a submodule of M and N K ⊆ N ∩ K
2010 Mathematics Subject Classification. Primary: 13A15; secondary: 13C99; 13F05.
Key words and phrases. Weakly prime submodule, Classical prime submodule, Weakly classical
prime submodule.
1
(see [2]). Let N be a proper submodule of a nonzero R-module M . Then the M radical of N , denoted by M -rad(N ), is defined to be the intersection of all prime
submodules of M containing N . If M has no prime submodule containing N , then
we say M -rad(N ) = M . It is shown in [12, Theorem 2.12] thatpif N is a proper
submodule of a multiplication R-module M , then M -rad(N ) = (N :R M )M . In
[15], Quartararo et al. said that a commutative ring R is a u-ring provided R has
the property that an ideal contained in a finite union of ideals must be contained in
one of those ideals; and a um-ring is a ring R with the property that an R-module
which is equal to a finite union of submodules must be equal to one of them. They
show that every Bézout ring is a u-ring. Moreover, they proved that every Prüfer
domain is a u-domain. Also, any ring which contains an infinite field as a subring
is a u-ring, [16, Exercise 3.63].
In this paper we introduce the concept of weakly classical prime submodules. A
proper submodule N of an R-module M is called a weakly classical prime submodule
if whenever a, b ∈ R and m ∈ M with 0 6= abm ∈ N , then am ∈ N or bm ∈ N .
Clearly, every classical prime submodule is a weakly classical prime submodule.
Among many results in this paper, it is shown (Theorem 2.17) that N is a weakly
classical prime submodule of an R-module M if and only if for every ideals I, J
of R and m ∈ M with 0 6= IJm ⊆ N , either Im ⊆ N or Jm ⊆ N . It is proved
(Theorem 2.19) that if N is a weakly classical prime submodule of an R-module M
that is not classical prime, then (N :R M )2 N = 0. It is shown (Theorem 2.25) that
over a um-ring R, N is a weakly classical prime submodule of an R-module M if
and only if for every ideals I, J of R and submodule L of M with 0 6= IJL ⊆ N ,
either IL ⊆ N or JL ⊆ N . Let R = R1 × R2 × R3 be a decomposable ring and
M = M1 × M2 × M3 be an R-module where Mi is an Ri -module, for i = 1, 2, 3. In
Theorem 2.38 it is proved that if N is a weakly classical prime submodule of M ,
then either N = {(0, 0, 0)} or N is a classical prime submodule of M . Let R be
a um-ring, M be an R-module and F be a faithfully flat R-module. It is shown
(Theorem 2.39) that N is a weakly classical prime submodule of M if and only if
F ⊗ N is a weakly classical prime submodule of F ⊗ M.
2. Properties of weakly classical prime submodules
First of all we give a module which has no nonzero weakly classical prime submodule.
Example 2.1. Let p be a fixed prime integer and N0 = N ∪ {0} . Then
r
E (p) := α ∈ Q/Z | α = n + Z for some r ∈ Z and n ∈ N0
p
is a nonzero submodule of the Z-module Q/Z. For each t ∈ N0 , set
r
Gt := α ∈ Q/Z | α = t + Z for some r ∈ Z .
p
Notice that for each t ∈ N0 , Gt is a submodule of E (p) generated by p1t + Z
for each t ∈ N0 . Each proper submodule of E (p) is equal to Gi for some i ∈
N0 (see, [16, Example 7.10]) . However, no Gt is a weakly classical prime submodule
2
1
of E (p) . Indeed, pt+2
+ Z ∈ E (p). Then 0 6= p2
1
1
+ Z = pt+1
+Z∈
/ Gt .
p pt+2
1
pt+2
+Z =
1
pt
+ Z ∈ Gt but
Theorem 2.2. Let M be an R-module and N a proper submodule of M .
(1) If N is a weakly classical prime submodule of M , then (N :R m) is a weakly
prime ideal of R for every m ∈ M \N with AnnR (m) = 0.
(2) If (N :R m) is a weakly prime ideal of R for every m ∈ M \N , then N is a
weakly classical prime submodule of M .
Proof. (1) Suppose that N is a weakly classical prime submodule. Let m ∈ M \N
with AnnR (m) = 0, and 0 6= ab ∈ (N :R m) for some a, b ∈ R. Then 0 6= abm ∈ N .
So am ∈ N or bm ∈ N . Hence a ∈ (N :R m) or b ∈ (N :R m).
(2) Assume that (N :R m) is a weakly prime ideal of R for every m ∈ M \N . Let
0 6= abm ∈ N for some m ∈ M and a, b ∈ R. If m ∈ N , then we are done. So we
assume that m ∈
/ N . Hence 0 6= ab ∈ (N :R m) implies that either a ∈ (N :R m) or
b ∈ (N :R m). Therefore either am ∈ N or bm ∈ N , and so N is weakly classical
prime.
We recall that M is a torsion-free R-module if and only if for every 0 6= m ∈ M ,
AnnR (m) = 0. As a direct consequence of Theorem 2.2 the following result follows.
Corollary 2.3. Let M be a torsion-free R-module and N a proper submodule of
M . Then N is a weakly classical prime submodule of M if and only if (N :R m) is
a weakly prime ideal of R for every m ∈ M \N .
Theorem 2.4. Let f : M → M ′ be a homomorphism of R-modules.
(1) Suppose that f is a monomorphism. If N ′ is a weakly classical prime submodule of M ′ with f −1 (N ′ ) 6= M , then f −1 (N ′ ) is a weakly classical prime
submodule of M .
(2) Suppose that f is an epimorphism. If N is a weakly classical prime submodule of M containing Ker(f ), then f (N ) is a weakly classical prime
submodule of M ′ .
Proof. (1) Suppose that N ′ is a weakly classical prime submodule of M ′ with
f −1 (N ′ ) 6= M . Let 0 6= abm ∈ f −1 (N ′ ) for some a, b ∈ R and m ∈ M . Since
f is a monomorphism, 0 6= f (abm) ∈ N ′ . So we get 0 6= abf (m) ∈ N ′ . Hence
f (am) = af (m) ∈ N ′ or f (bm) = bf (m) ∈ N ′ . Thus am ∈ f −1 (N ′ ) or bm ∈
f −1 (N ′ ). Therefore f −1 (N ′ ) is a weakly classical prime submodule of M .
(2) Assume that N is a weakly classical prime submodule of M . Let a, b ∈ R
and m′ ∈ M ′ be such that 0 6= abm′ ∈ f (N ). By assumption there exists m ∈ M
such that m′ = f (m) and so f (abm) ∈ f (N ). Since Ker(f ) ⊆ N , we have 0 6=
abm ∈ N . It implies that am ∈ N or bm ∈ N . Hence am′ ∈ f (N ) or bm′ ∈ f (N ).
Consequently f (N ) is a weakly classical prime submodule of M ′ .
As an immediate consequence of Theorem 2.4(2) we have the following corollary.
Corollary 2.5. Let M be an R-module and L ⊂ N be submodules of M . If N
is a weakly classical prime submodule of M , then N/L is a weakly classical prime
submodule of M/L.
Theorem 2.6. Let K and N be submodules of M with K ⊂ N ⊂ M . If K
is a weakly classical prime submodule of M and N/K is a weakly classical prime
submodule of M/K, then N is a weakly classical prime submodule of M .
3
Proof. Let a, b ∈ R, m ∈ M and 0 6= abm ∈ N . If abm ∈ K, then am ∈ K ⊂ N or
bm ∈ K ⊂ N as it is needed. Thus, assume that abm 6∈ K. Then 0 6= ab(m + K) ∈
N/K, and so a(m + K) ∈ N/K or b(m + K) ∈ N/K. It means that am ∈ N or
bm ∈ N , which completes the proof.
For an R-module M , the set of zero-divisors of M is denoted by ZR (M ).
Theorem 2.7. Let M be an R-module, N be a submodule and S be a multiplicative
subset of R.
(1) If N is a weakly classical prime submodule of M such that (N :R M )∩S = ∅,
then S −1 N is a weakly classical prime submodule of S −1 M .
(2) If S −1 N is a weakly classical prime submodule of S −1 M such that S ∩
ZR (N ) = ∅ and S ∩ ZR (M/N ) = ∅, then N is a weakly classical prime
submodule of M .
Proof. (1) Let N be a weakly classical prime submodule of M and (N :R M )∩S = ∅.
Suppose that 0 6= as11 as22 sm3 ∈ S −1 N for some a1 , a2 ∈ R, s1 , s2 , s3 ∈ S and m ∈ M .
Then there exists s ∈ S such that sa1 a2 m ∈ N . If sa1 a2 m = 0, then as11 as22 sm3 =
sa1 a2 m
0
ss1 s2 s3 = 1 , a contradiction. Since N is a weakly classical prime submodule,
−1
1m
N or
then we have a1 (sm) ∈ N or a2 (sm) ∈ N . Thus as11 sm3 = sa
ss1 s3 ∈ S
a2 m
sa2 m
−1
−1
=
∈
S
N
.
Consequently
S
N
is
a
weakly
classical
prime
submodule
s2 s3
ss2 s3
of S −1 M .
(2) Suppose that S −1 N is a weakly classical prime submodule of S −1 M and
S ∩ ZR (N ) = ∅ and S ∩ ZR (M/N ) = ∅. Let a, b ∈ R and m ∈ M such that
0
−1
N . If a1 1b m
0 6= abm ∈ N . Then a1 1b m
1 ∈S
1 = 1 , then there exists s ∈ S such that
0
sabm = 0 which contradicts S ∩ ZR (N ) = ∅. Therefore a1 1b m
1 6= 1 , and so either
am
bm
am
−1
−1
−1
N or 1 1 ∈ S N . We may assume that 1 1 ∈ S N . So there exists
1 1 ∈ S
u ∈ S such that uam ∈ N . But S ∩ ZR (M/N ) = ∅, whence am ∈ N . Consequently
N is a weakly classical prime submodule of M .
Darani [18] generalized the concept of prime submodules (resp. weakly prime
submodules) of a module over a commutative ring as following: Let N be a proper
submodule of an R-module M . Then N is said to be a 2-absorbing submodule
(resp. weakly 2-absorbing submodule) of M if whenever a, b ∈ R and m ∈ M with
abm ∈ N (resp. 0 6= abm ∈ N ), then am ∈ N or bm ∈ N or ab ∈ (N :R M ).
Proposition 2.8. Let N be a proper submodule of an R-module M .
(1) If N is a weakly prime submodule of M , then N is a weakly classical prime
submodule of M .
(2) If N is a weakly classical prime submodule of M , then N is a weakly 2absorbing submodule of M . The converse holds if in addition (N :R M ) is
a weakly prime ideal of R.
Proof. (1) Assume that N is a weakly prime submodule of M . Let a, b ∈ R and
m ∈ M such that 0 6= abm ∈ N . Therefore either bm ∈ N or a ∈ (N :R M ).
The first case leads us to the claim. In the second case we have that am ∈ N .
Consequently N is a weakly classical prime submodule.
(2) It is evident that if N is weakly classical prime, then it is weakly 2-absorbing.
Assume that N is a weakly 2-absorbing submodule of M and (N :R M ) is a weakly
prime ideal of R. Let 0 6= abm ∈ N for some a, b ∈ R and m ∈ M such that neither
4
am ∈ N nor bm ∈ N . Then 0 6= ab ∈ (N :R M ) and so either a ∈ (N :R M ) or
b ∈ (N :R M ). This contradiction shows that N is weakly classical prime.
The following example shows that the converse of Proposition 2.8(1) is not true.
L L
Z Z where p is a prime integer. ConExample 2.9. Let R = Z andL
M =L
Zp
sider the submodule N = {0} {0} Z of M . Notice that (0, 0, 0) 6= p(1, 0, 1) =
(0, 0, p) ∈ N , but (1, 0, 1) ∈
/ N . Also p(1, 1, 1) ∈
/ N which shows that p ∈
/ (N :Z M ).
Therefore N is not a weakly prime submodule of M . Now, we show that N is a
weakly classical prime submodule of M . Let m, n, z, w ∈ Z and x ∈ Zp be such that
(0, 0, 0) 6= mn(x, z, w) ∈ N . Hence mnx = 0 and mnz = 0. Therefore p|mnx and
z = 0. So p|m or p|nx. If p|m, then m(x, z, w) = (mx, 0, mw) = (0, 0, mw) ∈ N .
Similarly, if p|nx, then n(x, z, w) = (nx, 0, nw) = (0, 0, nw) ∈ N . Consequently N
is a weakly classical prime submodule of M .
Proposition 2.10. Let M be a cyclic R-module. Then a proper submodule N of M
is a weakly prime submodule if and only if it is a weakly classical prime submodule.
Proof. By Proposition 2.8(1), the “only if” part holds. Let M = Rm for some m ∈
M and N be a weakly classical prime submodule of M . Suppose that 0 6= rx ∈ N
for some r ∈ R and x ∈ M . Then there exists an element s ∈ R such that x = sm.
Therefore 0 6= rx = rsm ∈ N and since N is a weakly classical prime submodule,
rm ∈ N or sm ∈ N . Hence r ∈ (N :R M ) or x ∈ N . Consequently N is a weakly
prime submodule.
L L
Example 2.11. Let R = Z and M = Zp Zq
Q where p, q are two distinct
prime integers. One can easily see that the zero submodule of M is a weakly classical prime submodule. Notice that pq(1, 1, 0) = (0, 0, 0), but p(1, 1, 0) 6= (0, 0, 0)
and q(1, 1, 0) 6= (0, 0, 0). So the zero submodule of M is not classical prime. Hence
the two concepts of classical prime submodules and of weakly classical prime submodules are different in general.
Definition 2.12. Let N be a proper submodule of M and a, b ∈ R, m ∈ M. If N is
a weakly classical prime submodule and abm = 0, am ∈
/ N , bm ∈
/ N , then (a, b, m)
is called a classical triple-zero of N .
Theorem 2.13. Let N be a weakly classical prime submodule of an R-module M
and suppose that abK ⊆ N for some a, b ∈ R and some submodule K of M . If
(a, b, k) is not a classical triple-zero of N for every k ∈ K, then aK ⊆ N or
bK ⊆ N .
Proof. Suppose that (a, b, k) is not a classical triple-zero of N for every k ∈ K.
Assume on the contrary that aK 6⊆ N and bK 6⊆ N . Then there are k1 , k2 ∈ K
such that ak1 6∈ N and bk2 6∈ N . If abk1 6= 0, then we have bk1 ∈ N , because
ak1 6∈ N and N is a weakly classical prime submodule of M . If abk1 = 0, then since
ak1 ∈
/ N and (a, b, k1 ) is not a classical triple-zero of N , we conclude again that
bk1 ∈ N . By a similar argument, since (a, b, k2 ) is not a classical triple-zero and
bk2 ∈
/ N , then we deduce that ak2 ∈ N . From our hypothesis, ab(k1 + k2 ) ∈ N and
(a, b, k1 +k2 ) is not a classical triple-zero of N . Hence we have either a(k1 +k2 ) ∈ N
or b(k1 + k2 ) ∈ N . If a(k1 + k2 ) = ak1 + ak2 ∈ N , then since ak2 ∈ N , we have
ak1 ∈ N , a contradiction. If b(k1 + k2 ) = bk1 + bk2 ∈ N , then since bk1 ∈ N , we
have bk2 ∈ N , which again is a contradiction. Thus aK ⊆ N or bK ⊆ N .
5
Definition 2.14. Let N be a weakly classical prime submodule of an R-module
M and suppose that IJK ⊆ N for some ideals I, J of R and some submodule K
of M . We say that N is a free classical triple-zero with respect to IJK if (a, b, k)
is not a classical triple-zero of N for every a ∈ I, b ∈ J, and k ∈ K.
Remark 2.15. Let N be a weakly classical prime submodule of M and suppose
that IJK ⊆ N for some ideals I, J of R and some submodule K of M such that
N is a free classical triple-zero with respect to IJK. Hence, if a ∈ I, b ∈ J, and
k ∈ K, then ak ∈ N or bk ∈ N .
Corollary 2.16. Let N be a weakly classical prime submodule of an R-module M
and suppose that IJK ⊆ N for some ideals I, J of R and some submodule K of M .
If N is a free classical triple-zero with respect to IJK, then IK ⊆ N or JK ⊆ N .
Proof. Suppose that N is a free classical triple-zero with respect to IJK. Assume
that IK 6⊆ N and JK 6⊆ N . Then there are a ∈ I and b ∈ J with aK 6⊆ N and
bK 6⊆ N . Since abK ⊆ N and N is free classical triple-zero with respect to IJK,
then Theorem 2.13 implies that aK ⊆ N and bK ⊆ N which is a contradiction.
Consequently IK ⊆ N or JK ⊆ N .
Let M be an R-module and N a submodule of M . For every a ∈ R, {m ∈ M |
am ∈ N } is denoted by (N :M a). It is easy to see that (N :M a) is a submodule
of M containing N .
In the next theorem we characterize weakly classical prime submodules.
Theorem 2.17. Let M be an R-module and N be a proper submodule of M . The
following conditions are equivalent:
(1) N is weakly classical prime;
(2) For every a, b ∈ R, (N :M ab) = (0 :M ab) ∪ (N :M a) ∪ (N :M b);
(3) For every a ∈ R and m ∈ M with am ∈
/ N , (N :R am) = (0 :R am) ∪ (N :R
m);
(4) For every a ∈ R and m ∈ M with am ∈
/ N , (N :R am) = (0 :R am) or
(N :R am) = (N :R m);
(5) For every a ∈ R and every ideal I of R and m ∈ M with 0 6= aIm ⊆ N ,
either am ∈ N or Im ⊆ N ;
(6) For every ideal I of R and m ∈ M with Im * N , (N :R Im) = (0 :R Im)
or (N :R Im) = (N :R m);
(7) For every ideals I, J of R and m ∈ M with 0 6= IJm ⊆ N , either Im ⊆ N
or Jm ⊆ N .
Proof. (1)⇒(2) Suppose that N is a weakly classical prime submodule of M . Let
m ∈ (N :M ab). Then abm ∈ N . If abm = 0, then m ∈ (0 :M ab). Assume that
abm 6= 0. Hence am ∈ N or bm ∈ N . Therefore m ∈ (N :M a) or m ∈ (N :M b).
Consequently, (N :M ab) = (0 :M ab) ∪ (N :M a) ∪ (N :M b).
(2)⇒(3) Let am ∈
/ N for some a ∈ R and m ∈ M . Assume that x ∈ (N :R am).
Then axm ∈ N , and so m ∈ (N :M ax). Since am ∈
/ N , then m ∈
/ (N :M a).
Thus by part (2), m ∈ (0 :M ax) or m ∈ (N :M x), whence x ∈ (0 :R am) or
x ∈ (N :R m). Therefore (N :R am) = (0 :R am) ∪ (N :R m).
(3)⇒(4) By the fact that if an ideal (a subgroup) is the union of two ideals (two
subgroups), then it is equal to one of them.
(4)⇒(5) Let for some a ∈ R, an ideal I of R and m ∈ M , 0 6= aIm ⊆ N . Hence
I ⊆ (N :R am) and I * (0 :R am). If am ∈ N , then we are done. So, assume that
6
am ∈
/ N . Therefore by part (4) we have that I ⊆ (N :R m), i.e., Im ⊆ N .
(5)⇒(6)⇒(7) Have proofs similar to that of the previous implications.
(7)⇒(1) Is trivial.
Theorem 2.18. Let N be a weakly classical prime submodule of M and suppose
that (a, b, m) is a classical triple-zero of N for some a, b ∈ R, m ∈ M . Then
(1) abN = 0.
(2) a(N :R M )m = 0.
(3) b(N :R M )m = 0.
(4) (N :R M )2 m = 0.
(5) a(N :R M )N = 0.
(6) b(N :R M )N = 0.
Proof. (1) Suppose that abN 6= 0. Then there exists n ∈ N with abn 6= 0. Hence
0 6= ab(m+n) = abn ∈ N , so we conclude that a(m+n) ∈ N or b(m+n) ∈ N . Thus
am ∈ N or bm ∈ N , which contradicts the assumption that (a, b, m) is classical
triple-zero. Thus abN = 0.
(2) Let axm 6= 0 for some x ∈ (N :R M ). Then a(b + x)m 6= 0, because abm = 0.
Since xm ∈ N , a(b + x)m ∈ N . Then am ∈ N or (b + x)m ∈ N . Hence am ∈ N or
bm ∈ N , which contradicts our hypothesis.
(3) The proof is similar to part (2).
(4) Assume that x1 x2 m 6= 0 for some x1 , x2 ∈ (N :R M ). Then by parts (2)
and (3), (a + x1 )(b + x2 )m = x1 x2 m 6= 0. Clearly (a + x1 )(b + x2 )m ∈ N . Then
(a + x1 )m ∈ N or (b + x2 )m ∈ N . Therefore am ∈ N or bm ∈ N which is a
contradiction. Consequently (N :R M )2 m = 0.
(5) Let axn 6= 0 for some x ∈ (N :R M ) and n ∈ N . Therefore by parts (1)
and (2) we conclude that 0 6= a(b + x)(m + n) = axn ∈ N . So a(m + n) ∈ N or
(b + x)(m + n) ∈ N . Hence am ∈ N or bm ∈ N . This contradiction shows that
a(N :R M )N = 0.
(6) Similart to part (5).
A submodule N of an R-module M is called a nilpotent submodule if (N :R
M )k N = 0 for some positive integer k (see [1]), and we say that m ∈ M is nilpotent
if Rm is a nilpotent submodule of M .
Theorem 2.19. If N is a weakly classical prime submodule of an R-module M
that is not classical prime, then (N :R M )2 N = 0 and so N is nilpotent.
Proof. Suppose that N is a weakly classical prime submodule of M that is not
classical prime. Then there exists a classical triple-zero (a, b, m) of N for some
a, b ∈ R, m ∈ M . Assume that (N :R M )2 N 6= 0. Hence there are x1 , x2 ∈ (N :R
M ) and n ∈ N such that x1 x2 n 6= 0. By Theorem 2.18, 0 6= (a+x1 )(b+x2 )(m+n) =
x1 x2 n ∈ N . So (a + x1 )(m + n) ∈ N or (b + x1 )(m + n) ∈ N . Therefore am ∈ N
or bm ∈ N , a contradiction.
Remark 2.20. Let M be a multiplication R-module and K, L be submodules of
M . Then there are ideals I, J of R such that K = IM and L = JM . Thus
KL = IJM = IL. In particular KM = IM = K. Also, for any m ∈ M we define
Km := KRm. Hence Km = IRm = Im.
Corollary 2.21. If N is a weakly classical prime submodule of a multiplication
R-module M that is not classical prime, then N 3 = 0.
7
Proof. Since M is multiplication, then N = (N :R M )M . Therefore by Theorem
2.19 and Remark 2.20, N 3 = (N :R M )2 N = 0.
Assume that Nil(M ) is the set of nilpotent elements of M . If M is faithful, then Nil(M ) is a submodule
of M and if M is faithful multiplication, then
T
Nil(M ) = Nil(R)M = Q (= M -rad({0})), where the intersection runs over all
prime submodules of M , [1, Theorem 6].
Theorem 2.22. Let N be a weakly classical prime submodule of M . If N is not
classical prime, then
p
p
(1) (N :R M ) = AnnR (M ).
(2) If M is multiplication, then M -rad(N )=M -rad({0}). If in addition M is
faithful, then M -rad(N ) = Nil(M ).
Proof. (1) Assume that N is not classical prime. By Theorem 2.19, (N :R M )2 N =
0. Then
(N :R M )3 = (N :R M )2 (N :R M )
⊆ ((N :R M )2 N :R M )
= (0 :R M ),
p
p
p
M ) ⊆ (0 :R M ). Hence, we have (N :R M ) = (0 :R M ) =
and so (N :R
p
AnnR (M ).
p
p
(2) By part (1), M -rad(N ) = (N :R M )M = (0 :R M )M = M -rad({0}) =
Nil(M ).
Corollary 2.23. Let R be a ring and I be a proper ideal of R.
(1)
R I is a weakly classical prime submodule of R R if and only if I is a weakly
prime ideal of R.
(2) Every proper ideal of R is weakly prime if and only if for every R-module
M and every proper submodule N of M , N is a weakly classical prime
submodule of M .
Proof. (1) Let R I be a weakly classical prime submodule of R R. Then by Theorem
2.2(1), (I :R 1) = I is a weakly prime ideal of R. For the converse, notice that R I
is a weakly prime submodule of R R if and only if I is a weakly prime ideal of R.
Now, apply part (1) of Proposition 2.8.
(2) Assume that every proper ideal of R is weakly prime. Let N be a proper
submodule of an R-module M . Since for every m ∈ M \N , (N :R m) is a proper
ideal of R, then it is a weakly prime ideal of R. Hence by Theorem 2.2(2), N is
a weakly classical prime submodule of M . We have the converse immediately by
part (1).
Regarding Remark 2.20 we have the next proposition.
Proposition 2.24. Let M be a multiplication R-module and N be a proper submodule of M . The following conditions are equivalent:
(1) N is a weakly classical prime submodule of M ;
(2) If 0 6= N1 N2 m ⊆ N for some submodules N1 , N2 of M and m ∈ M , then
either N1 m ⊆ N or N2 m ⊆ N .
8
Proof. (1)⇒(2) Let 0 6= N1 N2 m ⊆ N for some submodules N1 , N2 of M and
m ∈ M . Since M is multiplication, there are ideals I1 , I2 of R such that N1 = I1 M
and N2 = I2 M . Therefore 0 6= N1 N2 m = I1 I2 m ⊆ N , and so either I1 m ⊆ N or
I2 m ⊆ N . Hence N1 m ⊆ N or N2 m ⊆ N .
(2)⇒(1) Suppose that 0 6= I1 I2 m ⊆ N for some ideals I1 , I2 of R and some m ∈ M .
It is sufficient to set N1 := I1 M and N2 := I2 M in part (2).
Theorem 2.25. Let R be a um-ring, M be an R-module and N be a proper submodule of M . The following conditions are equivalent:
(1) N is weakly classical prime;
(2) For every a, b ∈ R, (N :M ab) = (0 :M ab) or (N :M ab) = (N :M a) or
(N :M ab) = (N :M b);
(3) For every a, b ∈ R and every submodule L of M , 0 6= abL ⊆ N implies that
aL ⊆ N or bL ⊆ N ;
(4) For every a ∈ R and every submodule L of M with aL * N , (N :R aL) =
(0 :R aL) or (N :R aL) = (N :R L);
(5) For every a ∈ R, every ideal I of R and every submodule L of M , 0 6=
aIL ⊆ N implies that aL ⊆ N or IL ⊆ N ;
(6) For every ideal I of R and every submodule L of M with IL * N , (N :R
IL) = (0 :R IL) or (N :R IL) = (N :R L);
(7) For every ideals I, J of R and every submodule L of M , 0 6= IJL ⊆ N
implies that IL ⊆ N or JL ⊆ N .
Proof. Similar to that of Theorem 2.17.
Remark 2.26. The zero submodule of the Z-module Z4 , is a weakly classical prime
submodule (weakly prime ideal) of Z4 , but (0 :Z Z4 ) = 4Z is not a weakly prime
ideal of Z.
Proposition 2.27. Let R be a um-ring, M be an R-module and N be a proper
submodule of M . If N is a weakly classical prime submodule of M , then (N :R L) is
a weakly prime ideal of R for every faithful submodule L of M that is not contained
in N .
Proof. Assume that N is a weakly classical prime submodule of M and L is a
faithful submodule of M such that L * N . Let 0 6= ab ∈ (N :R L) for some
a, b ∈ R. Then 0 6= abL ⊆ N , because L is faithful. Hence Theorem 2.25 implies
that aL ⊆ N or bL ⊆ N , i.e., a ∈ (N :R L) or b ∈ (N :R L). Consequently (N :R L)
is a weakly prime ideal of R.
Proposition 2.28. Let M be an R-module and N be a weakly classical prime
submodule of M .Then
(1) For every a, b ∈ R and m ∈ M , (N :R abm) = (0 :R abm) ∪ (N :R
am) ∪ (N :R bm);
(2) If R is a u-ring, then for every a, b ∈ R and m ∈ M , (N :R abm) = (0 :R
abm) or (N :R abm) = (N :R am) or (N :R abm) = (N :R bm).
Proof. (1) Let a, b ∈ R and m ∈ M . Suppose that r ∈ (N :R abm). Then
ab(rm) ∈ N . If ab(rm) = 0, then r ∈ (0 :R abm). Therefore we assume that
ab(rm) 6= 0. So, either a(rm) ∈ N or b(rm) ∈ N . Thus, either r ∈ (N :R am) or
r ∈ (N :R bm). Consequently (N :R abm) = (0 :R abm) ∪ (N :R am) ∪ (N :R bm).
(2) Apply part (1).
9
Theorem 2.29. Let R be a um-ring, M be a faithful multiplication R-module and
N be a proper submodule of M . The following conditions are equivalent:
(1) N is a weakly classical prime submodule of M ;
(2) If 0 6= N1 N2 N3 ⊆ N for some submodules N1 , N2 , N3 of M , then either
N1 N3 ⊆ N or N2 N3 ⊆ N ;
(3) If 0 6= N1 N2 ⊆ N for some submodules N1 , N2 of M , then either N1 ⊆ N
or N2 ⊆ N ;
(4) N is a weakly prime submodule of M ;
(5) (N :R M ) is a weakly prime ideal of R.
Proof. (1)⇒(2) Let 0 6= N1 N2 N3 ⊆ N for some submodules N1 , N2 , N3 of M .
Since M is multiplication, there are ideals I1 , I2 of R such that N1 = I1 M and
N2 = I2 M . Therefore 0 6= I1 I2 N3 ⊆ N , and so by Theorem 2.25, I1 N3 ⊆ N or
I2 N3 ⊆ N . Thus, either N1 N3 ⊆ N or N2 N3 ⊆ N .
(2)⇒(3) Is easy.
(3)⇒(4) Suppose that 0 6= IK ⊆ N for some ideal I of R and some submodule K
of M . It is sufficient to set N1 := IM and N2 = K in part (3).
(4)⇒(1) By part (1) of Proposition 2.8.
(1)⇒(5) By Proposition 2.27.
(5)⇒(4) Let 0 6= IK ⊆ N for some ideal I of R and some submodule K of M .
Since M is multiplication, then there is an ideal J of R such that K = JM . Hence
0 6= IJ ⊆ (N :R M ) which implies that either I ⊆ (N :R M ) or J ⊆ (N :R M ).
If I ⊆ (N :R M ), then we are done. So, suppose that J ⊆ (N :R M ). Thus
K = JM ⊆ N .
Proposition 2.30. Let R be a um-ring. Let M be a finitely generated faithful
multiplication R-module and N a submodule of M. Then the following conditions
are equivalent:
(1) N is a weakly classical prime submodule;
(2) (N :R M ) is a weakly prime ideal of R;
(3) N = IM for some weakly prime ideal I of R.
Proof. (1) ⇔ (2) . By Theorem 2.29.
(2) ⇒ (3) Since (N :R M ) is a weakly prime ideal and N = (N :R M ) M, then
condition (3) holds.
(3) ⇒ (2) Suppose that N = IM for some weakly prime ideal I of R. Since M is a
multiplication module, we have N = (N : M ) M. Therefore N = IM = (N : M ) M
and so I = (N : M ), because by [17, Corollary to Theorem 9] M is cancellation.
Theorem 2.31. Let M1 , M2 be R-modules and N1 be a proper submodule of M1 .
Then the following conditions are equivalent:
(1) N = N1 × M2 is a weakly classical prime submodule of M = M1 × M2 ;
(2) N1 is a weakly classical prime submodule of M1 and for each r, s ∈ R and
m1 ∈ M1 we have
rsm1 = 0, rm1 ∈
/ N1 , sm1 ∈
/ N1 ⇒ rs ∈ AnnR (M2 ).
Proof. (1)⇒(2) Suppose that N = N1 × M2 is a weakly classical prime submodule
of M = M1 × M2 . Let r, s ∈ R and m1 ∈ M1 be such that 0 6= rsm1 ∈ N1 . Then
(0, 0) 6= rs(m1 , 0) ∈ N . Thus r(m1 , 0) ∈ N or s(m1 , 0) ∈ N , and so rm1 ∈ N1 or
sm1 ∈ N1 . Consequently N1 is a weakly classical prime submodule of M1 . Now,
10
assume that rsm1 = 0 for some r, s ∈ R and m1 ∈ M1 such that rm1 ∈
/ N1
and sm1 ∈
/ N1 . Suppose that rs ∈
/ AnnR (M2 ). Therefore there exists m2 ∈ M2
such that rsm2 6= 0. Hence (0, 0) 6= rs(m1 , m2 ) ∈ N , and so r(m1 , m2 ) ∈ N or
s(m1 , m2 ) ∈ N . Thus rm1 ∈ N1 or sm1 ∈ N1 which is a contradiction. Consequently rs ∈ AnnR (M2 ).
(2)⇒(1) Let r, s ∈ R and (m1 , m2 ) ∈ M = M1 × M2 be such that (0, 0) 6=
rs(m1 , m2 ) ∈ N = N1 × M2 . First assume that rsm1 6= 0. Then by part (2),
rm1 ∈ N1 or sm1 ∈ N1 . So r(m1 , m2 ) ∈ N or s(m1 , m2 ) ∈ N , and thus we are
done. If rsm1 = 0, then rsm2 6= 0. Therefore rs ∈
/ AnnR (M2 ), and so part (2)
implies that either rm1 ∈ N1 or sm1 ∈ N1 . Again we have that r(m1 , m2 ) ∈ N or
s(m1 , m2 ) ∈ N which shows N is a weakly classical prime submodule of M .
The following two propositions have easy verifications.
Proposition 2.32. Let M1 , M2 be R-modules and N1 be a proper submodule of
M1 . Then N = N1 × M2 is a classical prime submodule of M = M1 × M2 if and
only if N1 is a classical prime submodule of M1 .
Proposition 2.33. Let M1 , M2 be R-modules and N1 , N2 be proper submodules of
M1 , M2 , respectively. If N = N1 × N2 is a weakly classical prime (resp. classical
prime) submodule of M = M1 × M2 , then N1 is a weakly classical prime (resp.
classical prime) submodule of M1 and N2 is a weakly classical prime (resp. classical
prime) submodule of M2 .
Example 2.34. Let R = Z, M = Z × Z and N = pZ × qZ where p, q are two
distinct prime integers. Since pZ, qZ are prime ideals of Z, then pZ, qZ are weakly
classical prime Z-submodules of Z. Notice that (0, 0) 6= pq(1, 1) = (pq, pq) ∈ N ,
but neither p(1, 1) ∈ N nor q(1, 1) ∈ N . So N is not a weakly classical prime
submodule of M . This example shows that the converse of Proposition 2.33 is not
true.
Let Ri be a commutative ring with identity and Mi be an Ri -module, for i = 1, 2.
Let R = R1 × R2 . Then M = M1 × M2 is an R-module and each submodule of M
is in the form of N = N1 × N2 for some submodules N1 of M1 and N2 of M2 .
Theorem 2.35. Let R = R1 × R2 be a decomposable ring and M = M1 × M2 be
an R-module where M1 is an R1 -module and M2 is an R2 -module. Suppose that
N = N1 × M2 is a proper submodule of M . Then the following conditions are
equivalent:
(1) N1 is a classical prime submodule of M1 ;
(2) N is a classical prime submodule of M ;
(3) N is a weakly classical prime submodule of M .
Proof. (1)⇒(2) Let (a1 , a2 )(b1 , b2 )(m1 , m2 ) ∈ N for some (a1 , a2 ), (b1 , b2 ) ∈ R and
(m1 , m2 ) ∈ M . Then a1 b1 m1 ∈ N1 so either a1 m1 ∈ N1 or b1 m1 ∈ N1 which shows
that either (a1 , a2 )(m1 , m2 ) ∈ N or (b1 , b2 )(m1 , m2 ) ∈ N . Consequently N is a
classical prime submodule of M .
(2)⇒(3) It is clear that every classical prime submodule is a weakly classical prime
submodule.
(3)⇒(1) Let abm ∈ N1 for some a, b ∈ R1 and m ∈ M1 . We may assume that
0 6= m′ ∈ M2 . Therefore 0 6= (a, 1)(b, 1)(m, m′ ) ∈ N . So either (a, 1)(m, m′ ) ∈ N
or (b, 1)(m, m′ ) ∈ N . Therefore am ∈ N1 or bm ∈ N1 . Hence N1 is a classical
prime submodule of M1 .
11
Proposition 2.36. Let R = R1 × R2 be a decomposable ring and M = M1 × M2
be an R-module where M1 is an R1 -module and M2 is an R2 -module. Suppose that
N1 , N2 are proper submodules of M1 , M2 , respectively. If N = N1 × N2 is a weakly
classical prime submodule of M , then N1 is a weakly prime submodule of M1 and
N2 is a weakly prime submodule of M2 .
Proof. Suppose that N = N1 × N2 is a weakly classical prime submodule of M . By
hypothesis, there exist x ∈ M1 \N1 and y ∈ M2 \N2 . First we show that N1 is a
weakly prime submodule of M1 . Let 0 6= am1 ∈ N1 for some a ∈ R1 and m1 ∈ M1 .
Then 0 6= (1, 0) (a, 1) (m1 , y) ∈ N1 × N2 = N . Notice that if (a, 1) (m1 , y) ∈
N1 × N2 = N , then y ∈ N2 which is a contradiction. So we get (1, 0) (m1 , y) ∈
N1 × N2 = N . Thus m1 ∈ N1 . Hence N1 is a weakly prime submodule of M1 . A
similar argument shows that N2 is a weakly prime submodule of M2 .
The following example shows that the converse of Proposition 2.36 is not true in
general.
Example 2.37. Let R = M = Z × Z and N = pZ × qZ where p, q are two distinct
prime integers. Since pZ, qZ are prime ideals of Z, then pZ, qZ are weakly prime
(weakly classical prime) Z-submodules of Z. Notice that (0, 0) 6= (p, 1)(1, q)(1, 1) =
(p, q) ∈ N , but neither (p, 1)(1, 1) ∈ N nor (1, q)(1, 1) ∈ N . So N is not a weakly
classical prime submodule of M .
Theorem 2.38. Let R = R1 × R2 × R3 be a decomposable ring and M = M1 ×
M2 × M3 be an R-module where M1 is an R1 -module, M2 is an R2 -module and
M3 is an R3 -module. If N is a weakly classical prime submodule of M , then either
N = {(0, 0, 0)} or N is a classical prime submodule of M .
Proof. Since {(0, 0, 0)} is a weakly classical prime submodule in any module, we may
assume that N = N1 × N2 × N3 6= {(0, 0, 0)}. We assume that N is not a classical
prime submodule of M and reach a contradiction. Without loss of generality we may
assume that N1 6= 0 and so there is 0 6= n ∈ N1 . We claim that N2 = M2 or N3 =
M3 . Suppose that there are m2 ∈ M2 \N2 and m3 ∈ M3 \N3 . Get r ∈ (N2 :R2 M2 )
and s ∈ (N3 :R3 M3 ). Since (0, 0, 0) 6= (1, r, 1)(1, 1, s)(n, m2 , m3 ) = (n, rm2 , sm3 ) ∈
N , then (1, r, 1)(n, m2 , m3 ) = (n, rm2 , m3 ) ∈ N or (1, 1, s)(n, m2 , m3 ) = (n, m2 , sm3 ) ∈
N . Therefore either m3 ∈ N3 or m2 ∈ N2 , a contradiction. Hence N = N1 ×M2 ×N3
or N = N1 × N2 × M3 . Let N = N1 × M2 × N3 . Then (0, 1, 0) ∈ (N :R M ).
Clearly (0, 1, 0)2 N 6= {(0, 0, 0)}. So (N :R M )2 N 6= {(0, 0, 0)} which is a contradiction, by Theorem 2.19. In the case when N = N1 × N2 × M3 we have that
(0, 0, 1) ∈ (N :R M ) and similar to the previous case we reach a contradiction.
Theorem 2.39. Let R be a um-ring and M be an R-module.
(1) If F is a flat R-module and N is a weakly classical prime submodule of
M such that F ⊗ N 6= F ⊗ M, then F ⊗ N is a weakly classical prime
submodule of F ⊗ M.
(2) Suppose that F is a faithfully flat R-module. Then N is a weakly classical
prime submodule of M if and only if F ⊗ N is a weakly classical prime
submodule of F ⊗ M.
Proof. (1) Let a, b ∈ R. Then by Theorem 2.25, either (N :M ab) = (0 :M ab)
or (N :M ab) = (N :M a) or (N :M ab) = (N :M b). Assume that (N :M ab) =
12
(0 :M ab). Then by [5, Lemma 3.2],
(F ⊗ N :F ⊗M ab) = F ⊗ (N :M ab) = F ⊗ (0 :M ab)
= (F ⊗ 0 :F ⊗M ab) = (0 :F ⊗M ab) .
Now, suppose that (N :M ab) = (N :M a). Again by [5, Lemma 3.2],
(F ⊗ N :F ⊗M ab) = F ⊗ (N :M ab) = F ⊗ (N :M a)
= (F ⊗ N :F ⊗M a) .
Similarly, we can show that if (N :M ab) = (N :M b), then (F ⊗ N :F ⊗M ab) =
(F ⊗ N :F ⊗M b) . Consequently by Theorem 2.25 we deduce that F ⊗ N is a weakly
classical prime submodule of F ⊗ M.
(2) Let N be a weakly classical prime submodule of M and assume that F ⊗ N =
⊆
F ⊗ M . Then 0 → F ⊗ N → F ⊗ M → 0 is an exact sequence. Since F is
⊆
a faithfully flat module, 0 → N → M → 0 is an exact sequence. So N = M ,
which is a contradiction. So F ⊗ N 6= F ⊗ M . Then F ⊗ N is a weakly classical
prime submodule by (1). Now for the converse, let F ⊗ N be a weakly classical prime submodule of F ⊗ M . We have F ⊗ N 6= F ⊗ M and so N 6= M .
Let a, b ∈ R. Then (F ⊗ N :F ⊗M ab) = (0 :F ⊗M ab) or (F ⊗ N :F ⊗M ab) =
(F ⊗ N :F ⊗M a) or (F ⊗ N :F ⊗M ab) = (F ⊗ N :F ⊗M b) by Theorem 2.25. Suppose that (F ⊗ N :F ⊗M ab) = (0 :F ⊗M ab). Hence
F ⊗ (N :M ab) = (F ⊗ N :F ⊗M ab) = (0 :F ⊗M ab)
= (F ⊗ 0 :F ⊗M ab) = F ⊗ (0 :M ab) .
⊆
Thus 0 → F ⊗ (0 :M ab) → F ⊗ (N :M ab) → 0 is an exact sequence. Since F is a
⊆
faithfully flat module, 0 → (0 :M ab) → (N :M ab) → 0 is an exact sequence which
implies that (N :M ab) = (0 :M ab). With a similar argument we can deduce that
if (F ⊗ N :F ⊗M ab) = (F ⊗ N :F ⊗M a) or (F ⊗ N :F ⊗M ab) = (F ⊗ N :F ⊗M b),
then (N :M ab) = (N :M a) or (N :M ab) = (N :M b). Consequently N is a weakly
classical prime submodule of M by Theorem 2.25.
Corollary 2.40. Let R be a um-ring, M be an R-module and X be an indeterminate. If N is a weakly classical prime submodule of M , then N [X] is a weakly
classical prime submodule of M [X].
Proof. Assume that N is a weakly classical prime submodule of M . Notice that
R[X] is a flat R-module. Then by Theorem 2.39, R[X] ⊗ N ≃ N [X] is a weakly
classical prime submodule of R[X] ⊗ M ≃ M [X].
References
[1] M. M. Ali, Idempotent and nilpotent submodules of multiplication modules, Comm. Algebra,
36 (2008), 4620–4642.
[2] R. Ameri, On the prime submodules of multiplication modules, Inter. J. Math. Math. Sci.,
27 (2003), 1715–1724.
[3] D. D. Anderson and E. Smith, Weakly prime ideals, Houston J. Math., 29 (2003), 831–840.
[4] A. Azizi, On prime and weakly prime submodules, Vietnam J. Math., 36(3) (2008) 315–325.
[5] A. Azizi, Weakly prime submodules and prime submodules, Glasgow Math. J., 48 (2006)
343–346.
[6] M. Behboodi, A generalization of Bears lower nilradical for modules, J. Algebra Appl., 6 (2)
(2007) 337-353.
13
[7] M. Behboodi, On weakly prime radical of modules and semi-compatible modules, Acta Math.
Hungar., 113 (3) (2006) 239-250.
[8] M. Behboodi and H. Koohy, Weakly prime modules, Vietnam J. Math., 32 (2) (2004) 185-195.
[9] M. Behboodi and S. H. Shojaee, On chains of classical prime submodules and dimension
theory of modules, Bull. Iranian Math. Soc., 36(1) (2010) 149–166.
[10] J. Dauns, Prime modules, J. Reine Angew. Math., 298 (1978) 156–181.
[11] S. Ebrahimi Atani and F. Farzalipour, On weakly prime submodules, Tamk. J. Math., 38 (3)
(2007), 247–252.
[12] Z. A. El-Bast and P. F. Smith, Multiplication modules, Comm. Algebra, 16 (1988), 755–779.
[13] C.-P. Lu, Prime submodules of modules, Comm. Math. Univ. Sancti Pauli, 33 (1984), 61-69.
[14] R. L. McCasland and M. E. Moore, Prime submodules, Comm. Algebra, 20 (1992), 18031817.
[15] P. Quartararo and H. S. Butts, Finite unions of ideals and modules, Proc. Amer. Math. Soc.,
52 (1975), 91-96.
[16] R.Y. Sharp, Steps in commutative algebra, Second edition, Cambridge University Press, Cambridge, 2000.
[17] P. F. Smith, Some remarks on multiplication modules, Arch. Math., 50 (1988), 223–235.
[18] A. Yousefian Darani and F. Soheilnia, On 2-absorbing and weakly 2-absorbing submodules,
Thai J. Math., 9 (2011), 577-584
Hojjat Mostafanasab
Department of Mathematics and Applications,
University of Mohaghegh Ardabili,
P. O. Box 179, Ardabil, Iran.
Email: [email protected], [email protected]
Ünsal Tekir
Department of Mathematics,
Marmara University,
Ziverbey, Goztepe, Istanbul 34722, Turkey.
Email: [email protected]
Kürşat Hakan Oral
Department of Mathematics,
Yildiz Technical University,
Davutpasa Campus, Esenler, Istanbul, Turkey.
Email: [email protected]
14
| 0math.AC
|
arXiv:1801.01145v1 [cs.IT] 30 Dec 2017
Vectorial Boolean functions and linear codes in the context of
algebraic attacks
M. Boumezbeur∗, S. Mesnager†, K. Guenda‡
Abstract
In this paper we study the relationship between vectorial (Boolean) functions and
cyclic codes in the context of algebraic attacks. We first derive a direct link between
the annihilators of a vectorial function (in univariate form) and certain 2n -ary cyclic
codes (which we prove that they are LCD codes) extending results due to Rønjom and
Helleseth. The knowledge of the minimum distance of those codes gives rise to a lower
bound on the algebraic immunity of the associated vectorial function. Furthermore, we
solve an open question raised by Mesnager and Cohen. We also present some properties
of those cyclic codes (whose generator polynomials determined by vectorial functions)
as well as their weight enumerator. In addition we generalize the so-called algebraic
complement and study its properties.
Keywords: Algebraic attacks, vectorial functions, algebraic immunity, annihilators, cyclic
codes, LCD codes, weight distribution.
1
Introduction
Error correcting codes are widely studied by researchers and employed by engineers. They
have long been known to have applications in computer and communication systems, data
storage devices (starting from the use of Reed Solomon codes in CDs) and consumer electronics. On the other hand, substitution boxes (S-boxes) are fundamental parts of block
ciphers. They play an important role in their robustness, by providing confusion. Mathematically, S-boxes are vectorial (multi-output) Boolean functions, that is, functions from
the vector space Fn2 (of all binary vectors of length n) to the vector space Fm
2 , for given
positive integers n and m. These functions are called (n, m)-functions and include the
∗
USTHB, Laboratory of Algebra and number theory, BP 32 El Alia, Bab Ezzouar, Algeria. E-mail:
[email protected]
†
University of Paris VIII (department of mathematics), University of Paris XIII (LAGA) and Telecom
ParisTech, France. Email: [email protected]
‡
USTHB, Laboratory of Algebra and number theory, BP 32 El Alia, Bab Ezzouar, Algeria. Email:
[email protected]
1
(single-output) Boolean functions (which correspond to the case m = 1). When they are
used as S-boxes in block ciphers, their number m of output bits equals or approximately
equals the number n of input bits. They can also be used in stream ciphers, with m significantly smaller than n, in the place of Boolean functions to speed up the ciphers. We recall
that to resist against known attacks, the cryptographic Boolean and vectorial function
involving in the cryptosystems should be designed properly. In particular Boolean functions should be balanced to resist against statistical attacks, high algebraic degree to resist
against Berlekamp-Massey’s attack, high order correlation immunity against correlation attacks (in the combiner model of stream ciphers) and should have high algebraic immunity
to resist against algebraic attacks. For multi-output Boolean functions F , various measures
of nonlinearity have been widely discussed in the literature from coding point of view (see
for instance [16]. In particular, it has been shown that their corresponding linear code and
their corresponding punctured code are related to some special linear codes. The aim of
this paper is to exhibit further links between vectorial Boolean functions and linear codes.
We are dealing with vectorial functions in the context of algebraic attacks [2, 9, 11, 21]
(which principe goes back to Shanon’s work [25]) which are among the most efficient attacks for secret key cryptosystems introduced in 2003 by Courtois and Meier in [9] and
Courtois and Pieprzyk in [11]. The concept of algebraic immunity of Boolean functions
was proposed by Meier et al. in [21]. Next, it has been generalized to vectorial functions
n
by Armknecht and Krause in [1]. The authors of [10] proved that ⌈ ⌉ is the maximal value
2
of the algebraic immunity of n−variable Boolean functions. In [1] the authors gave an
upper bound on the algebraic immunity of vectorial function with n inputs and m outputs.
Some constructions of Boolean and vectorial functions with optimal algebraic immunity
were proposed in [1], [5], [6] and [29]. In [6] Carlet and Feng gave a construction of an
infinite class of Boolean functions with optimal algebraic immunity, their construction was
obtained from the BCH bound of cyclic codes, this gives a link between the algebraic immunity of Boolean functions and cyclic codes. Furthermore, by considering the univariate
representation of Boolean functions, Rønjom and Helleseth [15] gave a direct link between
the annihilators of Boolean functions and some q−ary cyclic codes. They proved that the
algebraic immunity of Boolean functions is equal to the minimum distance of those codes.
Another direct link between Boolean functions and cyclic codes was presented in [22], where
Mesnager and Cohen gave a lower bound on algebraic immunity of Boolean functions by
studying the properties of some cyclic codes. In this paper we firstly derive a direct link
between the annihilators of vectorial functions and some cyclic codes. The knowledge of
the minimum distance of those codes gives rise to a lower bound on the algebraic immunity
of the associated vectorial function. We present some properties of those cyclic codes for
which their generator polynomials are determined by vectorial functions.
This paper is organized as follows. In section 2 we give some preliminaries. In section 3
we present a direct link between the annihilators of a vectorial functions and certain cyclic
codes generalizing the link between Boolean functions and cyclic codes given in [15]. More
2
specifically, we show that any annihilator of an (n, m)-function F belong to an 2n -ary cyclic
code with generator polynomial GF (Theorem 8). Next, we describe the set of zeros of such
cyclic code C(F −1 (b))b∈F2m from which we deduce that the algebraic immunity is equal to
the minimal weight-heigh of a codeword of C(F −1 (b))b∈F2m (Corollary 10). Furthermore,
we extend in Subsection 3.2 the results given in [22] to (n, m)-functions F and derive a
lower bound on the lowest algebraic degree of nonzero annihilators of F (Theorem 12).
In subsection 3.3 we give a relation between the spectral immunity, algebraic immunity
of an (n, m)-vectorial function F and the minimal weight of C(F −1 (b))b∈F2m (Theorem
13). In section 4 we extend the notion of algebraic complement for vectorial functions
and present several results. In particular the relationship between the annihilators of the
(n, m)-function F and the annihilators of its algebraic complement is given. In Subsection
4.3, we study the lowest algebraic degree of nonzero annihilators of Boolean and vectorial
functions and answer to an open problem raised in [22]. In section 5 we show that the cyclic
code C(F −1 (b))b∈F2m associate to the (n, m)-function F is a linear code with complementary
duals (abbreviated LCD) which is a linear code whose intersection with its dual is trivial
[19]. Further, we present the parameters of C(F −1 (b))b∈F2m as well as some of its properties.
2
Preliminaries and notation
Throughout this paper, |E| will denote the cardinality of a finite set E.
2.1
Some background on vectorial Boolean functions
Given two positive integers n and m, a mapping from Fn2 to Fm
2 , is called an (n, m)-vectorial
function. Boolean functions correspond to the case where m = 1. An (n, m)-function F
being given, the coordinate functions of F are the Boolean functions f1 , · · · , fm defined by
F (x) = (f1 (x), · · · , fm (x)) at every x ∈ Fn2 . The component functions of F are the linear
combinations of its coordinate functions, with non all-zero coefficients.
In cryptography, the most used representation of an (n, m)-function is the so-called Algebraic Normal Form (ANF):
F (x1 , · · · , xn ) =
X
c(u)
u∈Fn
2
n
Y
i=1
xui i
!
,
c(u) ∈ Fm
2 .
The algebraic degree deg(F ) of any (n, m)-function F is by definition the global degree of
its ANF. It also equals the maximum algebraic degree of the coordinate functions of F or of
its component functions. Let F2n be the Galois field of order 2n . By identifying Fn2 to F2n ,
(n, m)-functions can be viewed as functions from F2n to F2m . In this case, the component
functions are the functions T r1m (vF (x)) where T r1m is the absolute trace from F2m to F2 :
P
m
2i
T r1m (x) = m−1
i=0 x , the function (x, y) 7→ T r1 (xy) being an inner product in F2m .
3
Every mapping from F2n into F2n admits a unique univariate polynomial representation:
F (x) =
n −1
2X
δi xi , δi ∈ F2n .
i=0
The algebraic degree of F is equal to maxj | cj 6=0 w2 (j) where we recall that w2 (j) is
the 2-weight of j, that is, the number of nonzero coefficients js in the binary expansion
Pn−1
s
s=o js 2 of j.
For every integer m dividing n, an (n, m)-function F can be viewed as a function from
F2n to itself and, therefore admits a unique univariate polynomial representation, which
n (P2n −1 c xj ), where T r n (x) = x + x2m + x22m + x23m +
can be represented in the form T rm
j
m
j=0
n−m
2
n
m
is the trace function from F2 to F2 (but without uniqueness if we do not add
···+x
restrictions on the polynomial inside the brackets).
Given a Boolean function f : F2n → F2 , its support denoted by supp(f ) is defined as
supp(f ) := {x ∈ F2n , f (x) = 1}.
Definition 1 (Walsh Hadamard transform). The Walsh Hadamard transform of f : F2n →
F2 at u ∈ F2n is defined as
Wf (u) =
X
n
(−1)f (x)+T r1 (ux) .
(1)
x∈F2n
Definition 2 (Annihilator [13]). Let H be a subset of Fn2 , any function g : Fn2 → F2 of
the set I(H) = {g 6= 0, g(x) = 0, ∀x ∈ H}, is called an annihilator of H.
Given an (n, m)−vectorial function F in univariate form, with m divides n, we shall
use the following definition of annihilators of F :
Definition 3. [8]. The Boolean function g 6= 0 is an annihilator of F if g(x)F (x) = 0, for
all x ∈ F2n , where for b ∈ F2m , g(x) = 0 for all x ∈ F −1 (b) := {a ∈ F2n , F (a) = b}.
In [1] Armknecht and Krause have introduced the algebraic immunity of (n, m)-functions
and gave the following definition.
Definition 4 (Algebraic immunity [1]). Let F be an (n, m)-function. The algebraic immunity of F is defined as follows
−1
AI(F ) = min{deg(g) : g 6= 0; ∃b ∈ Fm
(b)}
2 | g(x) = 0, ∀x ∈ F
We denote by LDA(F ) the lowest algebraic degree of nonzero annihilators of F . For
more about Boolean and vectorial functions, we send the reader to the excellent chapters
[3, 4].
4
2.2
Linear binary sequences and spectral immunity
A binary sequence s = (st )t≥0 is called linear recurring sequence of order n over F2 , if the
sequence satisfies
st+n + hn−1 · st+n−1 + hn−2 · st+n−2 + · · · + h0 · st = 0,
where n is a positive integer, t ≥ 0 and h0 , · · · , hn−1 ∈ F2 . The initial state vector is
the state vector s0 = (s0 , s1 , · · · , sn ).
The sequence s is generated with the characteristic polynomial
h(x) = xn + hn−1 xn−1 + · · · + h1 x + h0 .
If h(x) is with the least degree, then it is called the minimal polynomial of s and its
degree is the linear complexity of the sequence and is denoted by lc(s).
If there exists a positive integer T such that st+T = st , then the sequence is called periodic
of period T .
Definition 5. (Annihilator of a binary sequence [27]) For a binary sequence s of period
T , the binary sequence a 6= 0 is also of period T satisfying a · s = 0 is called an annihilator
of s.
Definition 6. (Spectral immunity [14, 15]) The spectral immunity of the binary sequence
s, denoted by SI(s), is the lowest linear complexity of all nonzero binary annihilators of s,
that is
SI(s) =
min lc(a).
{a6=0; a·s=0}
A filter generator is a key-stream generator composed of one LFSR. If zt is the sequence
generated the filter generator, then the output sequence is given by
st = f (zt+γ1 , zt+γ2 , · · · , zt+γn ).
2.3
Cyclic codes
In following we recall some basic notations and definitions on cyclic codes.
Definition 7. (Cyclic codes) A linear [n, k]−code C over F2 is called cyclic if for any codeword c = (c0 , · · · , cn−1 ) ∈ C, the vector c = (cn−1 , c0 , · · · , cn−2 ) is also a codeword of C.
Let Ai denote the number of codewords in C with Hamming weight i. The weight distribution of C is the sequence the (A1 , A2 , · · · , An ) and the weight enumerator is defined by
Q(z) = 1 + A1 z + A2 z 2 + · · · + An z n .
The dual code C ⊥ of a cyclic code C is the cyclic code defined by
C ⊥ = {v ∈ Fn2 ; v · c = 0, ∀c ∈ C};
5
where v · c is the usual inner product of v = (v1 , · · · , vn ) and c = (c1 , · · · , cn ) in Fn2 ,
v·c=
n
P
vi · ci .
i=1
n
Let S be a subset of F2n , let C(S) be the set of all tuples (a1 , · · · , a2n −1 ) of F22n −1 such
that
2n
−1
P
ai xi = 0 for all x ∈ S. Then C(S) is a cyclic code of length 2n − 1.
i=1
The generator polynomial g(x) of a cyclic code C is a factor of xn − 1 in F2n [x] and
every codeword in the code can be expressed as a multiple of g.
Let p(x) = (xn − 1)/g(x), then the generator polynomial of C ⊥ is
g⊥ = xk p(x−1 )/p(0).
(2)
The defining set of the cyclic code C is the set {i ∈ Z, g(αi ) = 0}. For more about cyclic
codes, we send the reader to the excellent book [17].
3
Vectorial functions and cyclic codes in the context of algebraic attacks
In this section we shall extend the results of [15] to (n, m)-functions and we give bounds
on the algebraic immunity and the spectral immunity of those functions.
3.1
Algebraic immunity of vectorial functions and cyclic codes
In [15] Rønjom and Helleseth gave a connection between Boolean functions and certain
2n −ary cyclic codes and then related the problem of the estimation of the algebraic immunity to the low-weight height codewords in cyclic codes. We recall that the weight height
of a polynomial g is defined to be
wh(g) = max{wt(i); ci 6= 0},
where wt(i) is the Hamming weight of the vector i, defined as the number of its nonzero
entries (see [15]).
The next result shows that any annihilator of a Boolean function f belongs to certain
2n −ary cyclic code of length 2n − 1 and gives an estimation of Al(f ).
Theorem 8. [15] Let f be a Boolean function in univariate form, any annihilator of f
belongs to the 2n -ary cyclic code with generator polynomial Gf such that
n −1
Gf = gcd(f (x) + 1, x2
6
+ 1).
The algebraic immunity of f is equal to the minimal weight-height of a codeword
g=
2n
−2
P
i=0
ci xi in the cyclic codes generated by Gf and G1+f .
In following we extend Theorem 8 to (n, m)-functions and derive a direct link between
(n, m)-functions and 2n −ary cyclic codes.
Theorem 9. Let F be an (n, m)-function given in its univariate representation. Then any
annihilator of F belongs to the 2n -ary cyclic code with generator polynomial GF given by
Y
GF =
n −1
gcd(F (x) − a, x2
+ 1).
(3)
a∈F∗2n
Proof. Let g be an annihilator of F such that g(x)F (x) = 0 for all x ∈ F2n . This holds for
any x ∈ F∗2n and we obtain
n −1
g(x)F (x) ≡ 0 mod (x2
+ 1),
which implies
n
g(x) ≡ 0 mod (
x2 −1 + 1
).
gcd(F (x), x2n −1 + 1
n −2
Since the (n, m)-function F (x) takes values in F2n , that is, F (x) ∈ {0, 1, α, · · · , α2
any x ∈ F2n , where α is a primitive element of F2n , then
n −1
x2
+1=
Q
n −1
gcd(F (x) − a, x2
} for
+ 1).
a∈F2n
n −1
n
Since that F 2 (x) = F (x) mod (x2
+ 1)
2n −1
Then
Q
x
+1
n
=
gcd(F (x) − a, x2 −1 + 1).
n −1
2
∗
gcd(F (x), x
+ 1) a∈F2n
g(x) ≡ 0 mod (
Q
a∈F∗2n
n −1
gcd(F (x) − a, x2
+ 1)),
hence g(x) is a codeword of the 2n −ary cyclic code with generator polynomial GF .
The following lemma describes the set of zeros of the cyclic with generator polynomial
GF given by (3) in Theorem 9.
Lemma 10. Let F be an (n, m)-function in univariate form, the nonzero elements of
F −1 (b), for b ∈ F2m , are the zeros of the cyclic code generated by GF denoted by C(F −1 (b))b∈F2m .
7
Proof. Let g(x) =
2n
−1
P
ai xi be an annihilator of F , then there exist b ∈ F2m such that
i=0
g(x) = 0 for every x ∈ F −1 (b). Since g is associated to a codeword of the code generated
2n
−1
P
2n
−1
P
ai xi = x
by GF , then for x 6= 0,
i=1
ai xi−1 = 0 that is
2n
−2
P
aj+1 xj = 0. Thus it holds
j=0
i=1
for every codeword of the cyclic code.
An immediate result from Theorem 9 is the following.
Corollary 11. The algebraic immunity of an (n, m)-function F is equal to the minimal
weight-heigh of a codeword of the 2n −ary cyclic code C(F −1 (b))b∈F2m with generator polynomial GF given by (3).
Proof. Suppose g ∈ AN (F ) with algebraic degree equal to the algebraic immunity of F .
Then by Theorem 9, g is associated to a codeword (a1 , · · · , a2n −1 ) of C(F −1 (b))b∈F2m .
Since that the degree of g is equal to
max
wt2 (j), then for all aj 6= 0, such that
n
j=0,··· ,2 −1; aj 6=0
0 ≤ j ≤ 2n − 1), wt2 (j) is equal to the weight-height of g.
3.2
Bound on the algebraic immunity of vectorial functions
In [22], the authors gave the following lower bound on the LDA(f ) by studying the properties of the 2n −ary cyclic codes C(supp(f )).
Theorem 12. ([22]) Let f : F2n → F2 and δ be the !
minimum distance of C(supp(f )),
P
n
where e is the lowest positive integer such that ei=0
≥ δ. Then LDA(f ) ≥ e.
i
The next result is an extension of Theorem 12 providing a bound on the algebraic
immunity of (n, m)-functions.
Theorem 13. Let F be an (n, m)-function in univariate form, b ∈ F2m and δ be the
minimum distance of C(F −1 (b))b∈F2m , hence the following hold.
!
e
P
n
(i) If e is the lowest positive integer such that verifying
< δ, then F has no nonzero
i=0 i
annihilator of algebraic degree less than or equal to e.
!
e
P
n
(ii) If e is the lowest positive integer such that verifying
≥ δ, then LDA(F ) ≥ e.
i=1 i
Proof. Let g(x) =
2n
−1
P
ai xi be an annihilator of F of algebraic degree at most e, that is,
i=0
−1
ai = 0 for every i of wt2 (i) > e. Let c be a codeword
! of C(F (b))b∈F2m , thus the number
e
P
n
of the nonzero components of c is at most
, that is the weight of c is less than δ.
i=0 i
8
Hence c is the null codeword which prove that any annihilator of F is of algebraic degree
! at
e
P
n
least e + 1. For the (ii) part, suppose that e is the smallest integer such that
≥ δ.
i=0 i
Then
e−1
P
i=0
3.3
n
i
!
< δ and then LDA(F ) ≥ e.
Spectral immunity of vectorial functions
In the following we give a relation between the spectral immunity and algebraic immunity
of an (n, m)-vectorial function F and the weight of C(F −1 (b))b∈F2m .
Let x ∈ F2n be the initial state, then the filter generator produce the key-stream zt , which
is a sequence over Fn2 , such that zt = F (xαt ) for t = 0, 1, · · · .
Let g be a Boolean function in univariate form, so that g(xαt ) = ut is an annihilator of
zt . According to Definition 6, the spectral immunity of zt is the lowest linear complexity
of ut which is the number of nonzero coefficients in
g(x) =
2n
−2
P
i=0
δi xi ∈ F2n [x].
Theorem 14. Let F be an (n, m)-functions function given in its univariate representation.
Then under the hypothesis of Theorem 9, the following hold
(i) The spectral immunity of F is equal to the minimal weight of the 2n −ary cyclic code
with generator polynomial GF .
(ii) The relation between the spectral immunity and the algebraic immunity of F is given
by the following inequality
SI(F ) ≤
AI(F
P)
i=0
!
n
.
i
Proof. By definition, the algebraic immunity of an (n, m)-functions function is the lowest
algebraic degree of its nonzero annihilator g. Let g(x) =
2nP
−1
ai xi be a Boolean func-
i=0
tion of algebraic degree at least AI(F ) associated to a codeword c = (a1 , · · · , a2n −1 )
of C(F −1 (b))b∈F2m! such that ai = 0 for all i with wt2 (i) ≤ AI(F ). Therefore g has
AI(F
P) n
at least
nonzero coefficients in its univariate representation, then SI(F ) ≤
i
i=0
Al(F
P)
i=0
4
!
n
.
i
The algebraic complement of vectorial functions
In this section, we present a generalization of the notion of algebraic complement to (n, m)functions. Next we give a bound on the lowest algebraic degree annihilator of F c .
9
We start by giving some results about the algebraic complement of Boolean functions.
4.1
The algebraic complement of Boolean functions
In [28], the authors extended the concept of algebraic immunity of Boolean function f and
then introduced the notion of algebraic complement f c as given in the following definition
Definition 15. [28] Given a Boolean function f on Fn2 , we define an algebraic complement
of f , denoted by f c , as the function that contains all monomials xa11 , · · · , xann , where each
aj ∈ F2 , that are not in ANF of the function f . That is any pair of functions (f, f c ) does
not have any monomials in common.
In [26] and [28] the authors have exhibited some properties of the algebraic complement
of Boolean function. Set ∆(x) =
n
Q
(1 + xi ) where x = (x1 , · · · , xn ) ∈ Fn2 . Then, they
i=1
have shown that f (x)∆(x) = 0 for f (0) = 0 and f (x)∆(x) = ∆(x) for f (0) = 1. Moreover,
they have proved that function f c satisfies f c = f + ∆ (noted that f c (x) = f (x) for all
x ∈ Fn2 \ {0}).
The next proposition gives an explicit relationship between AN (f ) and AN (f c ).
Proposition 16. [26] Let f be a Boolean function such that f (0) = 1 then AN (f c ) =
AN (f ) ∪ AN (f )c .
4.2
The algebraic complement of vectorial functions
The notion of algebraic complement can be extended to (n, m)-functions as follows.
Definition 17. Let F = (f1 , · · · , fm ) be an (n, m)-vectorial function. The algebraic
c ), where f c is the algebraic
complement of F , denoted by F c , is defined as F c = (f1c , · · · , fm
i
complement of fi , for 1 ≤ i ≤ m.
In the following, we give some properties of F c as well as a relationship between AN (F )
and AN (F c ).
Lemma 18. Let F be an (n, m)-function, then
1. F c (x) = F (x) for all nonzero x ∈ Fn2 .
2. F c (x) = F (x) + ∆(x)(1, · · · , 1) for all x ∈ Fn2 .
Proof. From [28, Lemma 2] we have fi (x) = fic (x) for all nonzero x ∈ Fn2 , 1 ≤ i ≤ m, then
for all nonzero x ∈ Fn2 we have that
c
F c (x) = (f1c (x), · · · , fm
(x)) = (f1 (x), · · · , fm (x)) = F (x).
10
From [28, Lemma 2] we have for all x ∈ Fn2 , f c (x) = f (x) + ∆(x), then
c
F (x)c = (f1c (x), · · · , fm
(x))
= (f1 (x) + ∆(x), · · · , fm (x) + ∆(x)) = F (x) + ∆(x)(1, · · · , 1).
The proof is completed.
We recall that a Boolean function g is said to be an annihilator of F if there exists
−1 (b) = {a ∈ Fn ; f (a) = b , 1 ≤ i ≤ m}.
b ∈ Fm
i
i
2
2 such that g(x) = 0 for all x ∈ F
In the following we give a relation between F −1 (b) and (F c )−1 (b).
Lemma 19. Let F be an (n, m)-function and F c its algebraic complement, the following
results hold
(i) If 0 ∈ F −1 (b) then (F c )−1 (b) = F −1 (b) ∩ Fn2 \ {0}.
(ii) If 0 ∈
/ F −1 (b) then (F c )−1 (b) = F −1 (b) ∪ {0}.
(iii) If 0 ∈
/ F −1 (b) and 0 ∈
/ (F c )−1 (b) then F −1 (b) = (F c )−1 (b).
Proof. Let g ∈ AN (F ), g(x) = 0 for all x ∈ F −1 (b). From [28, Lemma 2, item 1] we have
fi (x) = fic (x) for a nonzero x ∈ Fn2 , thus (F c )−1 (b) = F −1 (b) for all nonzero a ∈ Fn2 .
Let 0 ∈ F −1 (b), then 0 ∈
/ (F c )−1 (b), therefore (F c )−1 (b) = F −1 (b) ∩ Fn2 \ {0}.
−1
Let 0 ∈
/ F (b), then 0 ∈ (F c )−1 (b), therefore (F c )−1 (b) = F −1 (b) ∪ {0}.
Lemma 20. Let g ∈ AN (F ), with the previous notations set AN (F )c = {gc , g ∈ AN (F )},
then
(i) g ∈ AN (F c ) for F −1 (b) = (F c )−1 (b).
(ii) AN (F ) ⊂ AN (F c ) with 0 ∈ F −1 (b).
(iii) AN (F )c ⊂ AN (F c ) with 0 ∈
/ F −1 (b).
Let g ∈ AN (F c ), then
(iv) AN (F c ) ⊂ AN (F ) for 0 ∈
/ F −1 (b).
(v) AN (F c )c ⊂ AN (F ) for 0 ∈ F −1 (b).
−1 (b).
Proof. Let g be an annihilator of F , for a given b ∈ Fm
2 , g(x) = 0 for all x ∈ F
−1
c
−1
First suppose 0 ∈
/ F (b) and 0 ∈
/ (F ) (b), then by item (iv) of Lemma 19, F −1 (b) =
(F c )−1 (b), which gives g(x) = 0 for all x ∈ F −1 (b) = (F c )−1 (b), then g is an annihilator
of F c . Now if 0 ∈ F −1 (b) then 0 ∈
/ (F c )−1 (b), from item (iii) of Lemma 19, F −1 (b) =
c
−1
(F ) (b) ∪ {0}. Then g(x) = 0 for all x ∈ (F c )−1 (b) ∪ {0}. Thus g(x) = 0 for all
x ∈ (F c )−1 (b). So g ∈ AN (F c ) which gives AN (F ) ⊂ AN (F c ).
Since that 0 ∈
/ F −1 (b), then from item (iii) of Lemma 19, F −1 (b) = (F c )−1 (b) ∩ Fn2 − {0},
which gives g(x) = 0 for all x ∈ (F c )−1 (b)∩Fn2 \{0}. So gc ∈ AN (F c ) since that gc takes the
same value of g for all nonzero x ∈ Fn2 and the value 0 for x = 0. So AN (F )c ⊂ AN (F c ).
The relation between the annihilators of the (n, m)-function F and the annihilators of
its Algebraic complement is given by the following theorem.
11
Theorem 21. Let F be an (n, m)-function and F c its algebraic complement, then
(i) AN (F c ) = AN (F ) ∪ AN (F )c for 0 ∈ F −1 (b).
(ii) AN (F ) = AN (F c ) ∪ AN (F c )c for 0 ∈
/ F −1 (b).
c
−1
c
(iii) AN (F ) = AN (F ) for F (b) = (F )−1 (b).
Proof. Let AF = {g ∈ AN (F ), 0 ∈ F −1 (b)} ⊂ AN (F ), AF c = {g ∈ AN (F c ), 0 ∈
(F c )−1 (b)} ⊂ AN (F c ) and let 0 ∈ F −1 (b), g ∈ AN (F ). This means that g(x) = 0 for all
x ∈ F −1 (b), then g(0) = 0, therefore AF = AN (F ).
By item (i) of Lemma 20, g ∈ AF c for all g ∈ AF which gives AF c ⊂ AF . By item (iii) of
Lemma 20, g ∈ AF for all g ∈ AF c , then AF ⊂ AF c , this implies that AF = AN (F ) = AF c .
Let BF c = {g ∈ AN (F c ), 0 ∈
/ (F c )−1 (b)}, then #AN (F c ) = #AF c + #BF c . By item (ii)
of Lemma 20, g ∈ AF then gc ∈ BF c . But AF = AF c , then g ∈ BF c for all g ∈ AF c . Thus
#AF c = #BF c .
Let ϕ be a mapping from AF c to BF c , such that for all p ∈ AF c , ϕ(p) = pc ∈ BF c . To
prove that ϕ is injective we take pc1 , pc2 such that pc1 = pc2 , then p1 + ∆(x) = p2 + ∆(x) this
implies that p1 = p2 .
Since #AF c = #BF c , then ϕ is surjective. Then we have that AN (F c ) = AF c ∪ BF c =
AN (F )∪AN (F )c . Which proves the theorem’s first item. When 0 ∈ F −1 (b), 0 ∈
/ (F c )−1 (b).
We prove item (ii) of the Theorem in same way as item (i).
Let F −1 (b) = (F c )−1 (b). By item (v) of Lemma 20, if g ∈ AN (F ) then g ∈ AN (F c ), thus
AN (F c ) ⊂ AN (F ), if g ∈ AN (F c ) then g ∈ AN (F ), thus AN (F ) ⊂ AN (F c ). That gives
AN (F ) = AN (F c ), which completes the proof.
4.3
On the lowest algebraic degree of nonzero annihilators of Boolean
and vectorial functions
In this subsection, we are firstly interested on the lowest algebraic degree of nonzero annihilators of Boolean functions.
Let α be a primitive element of F2n , we define the set of t consecutive zeros of C(supp(f ))
by V (α, l, t) = {αl , αl+1 , . . . , αl+t−1 }.
Theorem 22. [17] Let α be a primitive element of F2n and let r, k and t be non-negative
integers, m a positive integer relatively prime to n. Let C ⊂ F2n a cyclic code having
αr , αr+1 , . . . , αr+t−1 , αr+m , αr+m+1 , . . . , αr+m+t−1 , . . . ,
αr+km , αr+km+1 , . . . , αr+km+t−1 . Then the minimum distance of C δ is greater then t + k.
In [23], the following result was proved.
Theorem 23. Let f : F2n → F2 be a Boolean function and l, k and δ be positive integers,
let m be a positive integer which is relatively prime to n. Suppose Sf contains V (α, l, δ −
1) ∪ V (α, l + m, δ − 1) ∪ . . . ∪ V (α,!l + km, δ − 1). Then LDA(f ) ≥ e, where e is the smallest
e
P
n
positive integer such that
≥ δ + k.
i=1 i
12
In following we give a lower bound for LDA(1 + f ), where f is defined as in Theorem
23.
Theorem 24. Let f : F2n → F2 be a Boolean function, let l, k and δ be positive integers,
m be a positive integer relatively prime to n. Then
(i) Supp(1 + f ) contains V (α, l + δ − 1, 2n − δ) ∪ V (α, l + m + δ − 1, 2n − δ) ∪ . . . ∪ V (α, l +
km + δ − 1, 2n − δ).
!
e
P
n
(ii) LDA(1+f ) ≥ e−1, where e is the smallest possible positive integer such that
≥
i=1 i
2n − δ + k.
Proof. Let f be a Boolean function and C(supp(f )) the cyclic code with generator polynomial Gf .
supp(1 + f ) = zeros(f ) ∩ F∗2n , then
n −2
S1+f = {1, α, . . . , αl−1 , αl+δ−1 , . . . , α2
n −2
} ∪ {1, α, . . . , αl+m−1 , αl+m+δ−1 , . . . , α2
2n −2
∪ . . . ∪ {1, α, . . . , αl+km−1 , αl+km+δ−1 , . . . , α
l+δ−1
= {α
l+δ−1+(2n −δ)−1
,...,α
l+m+δ−1
} ∪ {α
l+km+δ−1+(2n −δ)−1
∪ {αl+km+δ−1 , . . . , α
}
}.
n −δ)−1
, . . . , αl+m+δ−1+(2
} ∪ ...
}.
n
= V (α, l + δ − 1, 2 − δ) ∪ V (α, l + m + δ − 1, 2n − δ) ∪ . . . ∪ V (α, l + km+
δ − 1, 2n − δ).
Thanks to [22, Corollary 02] and!Theorem 22, LDA(1 + f ) ≥ e − 1, where e is the smallest
e
P
n
positive integer such that
≥ 2n − δ + k.
i=1 i
As a consequence, one can immediately derive a bound on the algebraic immunity of
Boolean function.
Corollary 25. Let f and 1 + f be defined as in Theorem 23 and Theorem
24, respectively.
!
e
P
n
Then AI(f ) ≥ e − 1, where e is the smallest integer such that
≥ min(δ + k, 2n −
i
i=1
δ + k).
Now, we are interested in vectorial functions. The following result gives a link between
the lowest algebraic degree of nonzero annihilators of vectorial functions and of its algebraic
complement.
Proposition 26. Let F be an (n, m)-function and F c be its algebraic complement. Then
we have
LDA(F ) − 1 ≤ LDA(F c ) ≤ LDA(F ) + 1.
13
Proof. Let g be an annihilator of F of degree LDA(F ). Since that F (x) = F c (x) for all
x ∈ F∗2n , xg(x)F c (x) = 0 for all x ∈ F2n . Then LDA(F c ) ≤ LDA(F ) + 1.
Let g be an annihilator of F c , clearly xg(x)F (x) = 0 for all x ∈ F2n . Then LDA(F ) − 1 ≤
LDA(F c ).
5
Further properties of the cyclic C(F −1(b))b∈F2m associate to
the vectorial function F
In this section we firstly prove that the constructed cyclic code C(F −1 (b))b∈F2m is a linear
code with complementary dual (abbreviated LCD). Recall that LCD codes are linear codes
whose intersection with their dual are trivial [19]. When they are binary, they play an
important role in armoring implementations against side-channel attacks and fault injection
attacks [7]. In following, we recall some basic results on LCD cyclic codes [18].
Definition 27. (Self reciprocal polynomial) The polynomial G(x) will be called selfdeg(G) G(x−1 ).
reciprocal if and only if G(x) = G−1
0 x
Proposition 28. [18] Let C be a cyclic code generated by a monic polynomial G(x), then
the following statements are equivalent.
(i) C is an LCD code.
(ii) C is generated by the self-reciprocal polynomial G(x).
As shown in Theorem 9, for certain b ∈ Fm
2 , the annihilators of an (n, m)-function in
n
univariate form belong to the 2 −ary cyclic code C(F −1 (b))b∈F2m with generator polynomial
GF . Let C ⊥ be the dual of C(F −1 (b))b∈F2m . Then we can deduce easily the generator
polynomial of C ⊥ .
Corollary 29. The dual code C ⊥ of the cyclic code C(F −1 (b))b∈F2m is generated by the
polynomial
deg(G) G(x−1 ),
G⊥
F =x
n −1
where G(x) = gcd(F (x), x2
+ 1).
In other words, the generator polynomial of C ⊥ is the reciprocal polynomial of G(x).
By definition, the zeros of C ⊥ are the inverses of the non zeros of C(F −1 (b))b∈F2m , so from
Lemma 10, are the inverses of the elements of zero of F .
In the following, we present the parameters of the LCD C(F −1 (b))b∈F2m code.
Theorem 30. For a given b ∈ F2m , C(F −1 (b))b∈F2m is an LCD cyclic code with parameters
1
[2n − 1, 2n−1 − Wφb (0), d].
2
14
Proof. Let βi (βi be a primitive element of F2n ) be a zero of G(x), then βi−1 is a zero of
⊥ −1
⊥
G⊥
F (x). Which implies that G(βi ) = GF (βi ) = 0. Thus G(x) = GF (x) for all x ∈ {x ∈
F2n , F (x) = 0}, this implies that {β1 , . . . , βr } = {β1−1 , . . . , βr−1 }. Then from Proposition
28, C(F −1 (b))b∈F2m is an LCD code.
By definition, the degree of the generator polynomial of C(F −1 (b))b∈F2m is equal to |D|,
where D the defining set of C(F −1 (b))b∈F2m . From Lemma 10, D is the set F −1 (b) for b in
F2m .
Let φb be the indicator function of F −1 (b) defined over F2n by φb (x) = 1 if F (x) = b and
φb (x) = 0 otherwise. Then |F −1 (b)| = |{x ∈ F2n , φb (x) = 1}|. Hence,
|D| = |{x ∈ F2n , φb (x) = 1}|
= 2n − |{x ∈ F2n , φb (x) = 0}|
1X X
= 2n −
(−1)yφb (x)
2 y∈F x∈F n
2
2
1
= 2n−1 − Wφb (0).
2
1
Then the dimension of C(F −1 (b))b∈F2m is equal to 2n−1 − Wφb (0).
2
Since the set of annihilators of an (n, m)-function containing the annihilators of its
algebraic complement (the inverse is true), then we give the link between the cyclic codes
associated to F and F c as follows.
5.0.1
′
Link between C(F −1 (b))b∈F2m and C((F c )−1 (b ))b′ ∈F2m
Let F be an (n, m)-function in univariate form and F c be its algebraic complement. In the
′
′
following we give a relation between C(F −1 (b))b∈F2m and C((F c )−1 (b )), for some b, b ∈
F2m .
′
Lemma 31. Let F : F2n → F2m and F c its algebraic complement, b, b ∈ F2m . Let
′
C(F −1 (b))b∈F2m and C((F c )−1 (b )) be the 2n −ary cyclic codes generated by GF and GF c ,
then
′
(i) C(F −1 (b)) ⊂ C((F c )−1 (b )).
(ii) ∃ h ∈ F2n [x] such that GF = GF c .h.
′
Proof. Let 0 ∈ F −1 (b), which means that 0 ∈
/ (F c )−1 (b ). Since F c (x) = F (x) for all
′
−1
c
−1
∗
∗
x ∈ F2n then F (b) = (F ) (b ) ∩ F2n .
Suppose that g ∈ AN (f ) associated to a codeword (a1 , . . . , a2n −1 ) of C(F −1 (b))b∈F2m , then
2n
−1
P
ai xi = 0 for all x ∈ F −1 (b), this implies that a0 = 0 for all x ∈ F −1 (b). Thus
i=0
2n
−1
P
i=1
′
ai xi = 0 for all x ∈ F −1 (b) ∩ F∗2n . Then (a1 , . . . , a2n −1 ) ∈ C((F c )−1 (b )). Thanks to
15
[12, Theorem 2.29, item 1], GF c divide GF , so it exists a polynomial h(x) ∈ F2n [x] such
that GF = GF c .h(x).
5.0.2
The weight distribution of C(F −1 (b))b∈F2m
The non zeros weights of the LCD code associated to an (n, m)-function F are given as
below. First, we recall that an upper bound on the algebraic immunity of (n, m)-functions
was given in [1] in the next lemma.
Lemma 32.
! For an (n, m)-function F , Al(F ) ≤ d where d is the smallest integer such
d
P n
that
> 2n−m .
i=0 i
By considering the case where n = m, the following proposition holds.
Proposition 33. Let F be an (n, m)-function, then Al(F ) ≤ 1.
Proof. From Lemma 32, Al(F ) ≤ d where d is the smallest integer such that
d
P
i=0
2n−m . Let n = m, then d is the smallest integer such that
d
P
i=0
d = 1 and Al(F ) ≤ 1.
n
i
!
n
i
!
>
> 2n−n = 1, that is
In the following, we give the non-zeros weights of the LCD code .
−1
Theorem 34. Let F be an (n, m)-function
! and C(F (b))b∈F2m of minimal distance δ. Let
e
P
n
e be the smallest integer such that
≥ δ, then
i=0 i
e ≤ Al(F ) ≤ 1
and we have
(i) If e = 1, then C(F −1 (b))b∈F2m has codewords of non zero weights ≥ 1 + n.
(ii) If e = 0, then C(F −1 (b))b∈F2m is the trivial LCD code.
Proof. Suppose that g =
2n
−1
P
is an annihilator of F of algebraic degree LDA(F ), g is
i=0
associated to a codeword c of
C(F −1 (b))b∈F2m ,
then wt(c) is equal to
LDA(F
P )
i=0
From Theorem 13, LDA(F ) ≥ e where e is the smallest integer such that
e
P
i=0
from Proposition 33 we have !that e ≤ Al(F ) ≤ 1. Thus e ∈ {0, 1}.
1
P
n
If e = 1, then wt(c) ≥
= 1 + n.
i=0 i
If e = 0, the result is obvious.
16
!
n
.
i
!
n
≥ δ and
i
6
Conclusions
In this paper a direct link between the annihilators of (n, m)-functions (in univariate form)
and certain 2n −ary cyclic codes was given. The knowledge of the minimum distance of
those codes give rise to a lower bound on the algebraic immunity of the (n, m)-functions.
Some properties of cyclic codes (whose generator polynomials were determined by vectorial
functions) as well as their weight enumerator were given. Further, we gave a generalization
and some properties of the so-called algebraic complement as well as a lower bound on the
lowest algebraic degree of nonzero annihilators.
References
[1] F. Armknecht and M. Krause.: "Constructing single-and multi-output Boolean functions with maximal immunity", ICALP (2), pp. 180-191, 2006.
[2] L.M. Batten.: Algebraic attacks over GF(q), Lecture Notes in Computer Science 3348,
Springer, ISBN 3-540-24130-2, pp. 84-91, 2004.
[3] C. Carlet, "Boolean Functions for Cryptography and error-correcting codes", Boolean
Models and Methods in Mathematics, Computer Science, and Engineering. Y. Crama,
P.L. Hammer. pp. 257-397, 2010.
[4] C. Carlet.: Vectorial Boolean Functions for Cryptography, Boolean Models and Methods in Mathematics, Computer Science, and Engineering. Y. Crama, P.L. Hammer.
pp. 398-470, 2010.
[5] C. Carlet, D.K. Dalai, K.C. Gupta and S. Maitra.: Algebraic immunity for cryptographically significant Boolean functions : Analysis and construction, IEEE Trans.
Inform. Theory 52(7), pp. pp. 3105-3121, 2006.
[6] C. Carlet and K. Feng.: An infinite class of balanced functions with optimal algebraic immunity, good immunity to fast algebraic attacks and good nonlinearity, In
ASIACRYPT, LNCS 5350. Springer-Verlag, pp. 425-440, 2008.
[7] C. Carlet and S. Guilley.: Complementary dual codes for counter-measures to sidechannel attacks, In: E. R. Pinto et al. (eds.), Coding Theory and Applications, CIM
Series in Mathematical Sciences, vol. 3, pp. 97-105, Springer Verlag, 2014 and Journal
Adv. in Math. of Comm. 10(1), pp. 131-150, 2016.
[8] N. Courtois.: Algebraic attacks on stream ciphers, Extended slides by Courtois.
www.nicolascourtois/papers/toyolili-slides.pdf.
17
[9] N. Courtois and W. Meier.: Algebraic attacks on stream ciphers with linear feedback,
In Advances cryptology-Eurocrypt 2003. Lecture Notes in computer science. SpringerVerlag, vol. 2656, pp. 345-359, 2003.
[10] N. Courtois and W. Meier.: Algebraic attacks on stream ciphers with linear feedback,
extended version of the Eurocrypt 2003.
[11] N. Courtois and J. Pieprzyk.: Cryptanalysis of block ciphers with over-defined systems
of equations, Proceedings of ASIACRYPT 2002, Lecture Notes in Computer Science
2501, pp. 267-287, 2003.
[12] C. Ding.: Codes from Difference sets, World Scientific Publishing Company, 2014.
[13] K. Feng, Q. Liao and J. Yang.: Maximal values of generalized algebraic immunity,
Designs, Codes and Cryptography 50(2), pp. 243-252, 2009.
[14] G. Gong, S. Rønjom, T. Helleseth and H. Hu.: Fast Discrete Fourier Spectra Attacks
on stream Ciphers, IEEE Trans. Inform. Theory 57(8), pp. 5555-5565, 2011.
[15] T. Helleseth and S. Rønjom.: Simplifying Algebraic Attacks with Univariate Analysis,
In Proceedings of Information Theory and Applications Workshop (ITA), pp. 1-7,
2011.
[16] J. Liu, S. Mesnager, L. Chen.: On the nonlinearity of S-boxes and linear codes. Journal
Cryptography and Communications 9(3): 345-361, 2017.
[17] F.J. MacWilliams and N.J. Sloane.: The theory of error correcting codes, Amsterdam,
North Holland 1977.
[18] J.L. Massey, "Reversible codes", Information&Control, Vol. 7. pp. 369-380, Sept 1964.
[19] J. L. Massey.: Linear codes with complementary duals, Discrete Math., vol. 106-107,
pp. 337-342, 1992.
[20] M. Matsui.: Linear cryptanalysis method for DES cipher, advences in cryptologyEUROCRYPT 1993. LNCS, vol. 765, Springer-verlag, 1993, pp. 386-397.
[21] W. Meier, E. Pasalic and C. Carlet.: Algebraic attacks and decomposition of Boolean
functions, in Eurocrypt 2004. Lecture notes in computer science, volume 3027, 2004.
[22] S. Mesnager, G.D. Cohen.: Cyclic codes and algebraic immunity of Boolean functions,
ITW: pp. 1-5, 2015.
[23] S. Mesnager.: A note on linear codes and algebraic immunity of Boolean functions,
21-st International Symposium on Mathematical Theory of Networks and Systems
July 7-11, 2014.
18
[24] K. Nyberg.: Perfect nonlinear S-Boxes, in advences in cryptology-EUROCRYPT, New
York, NY, USA, Springer- Verlag, pp. 378-385, 1991.
[25] C.E. Shannon.: Communication theory of secrecy systems, Bell system technical journal, 28, pp. 656-715, 1949. .
[26] C.P. Wang and X.S. Chen.: On extended algebraic immunity", Des. Codes Cryptography 57(3), pp. 271-281, 2010.
[27] J. Wang, K. Chen and S. Zhu.: Annihilators of Fast Discrete Fourier Spectra Attacks," Advances in Information and Computer security, 7-th international workshop
on security, IWSEC 2012 Fokuoka, Japan, pp. 182-196, 2012.
[28] Xian-Mo Zhang, J. Pieprzyk and Y. Zheng.: On algebraic Immunity and Annihilators,
Lecture note in computer Science, Information Security and Cryptology-ICISC, pp.
65-77, 2006.
[29] J. Zhang, S. Song, J. Du and Q. Wen.: On the construction of multi-output Boolean
functions with optimal algebraic immunity, science China Information Sciences 55(7),
pp. 1617-1623, 2012.
19
| 7cs.IT
|
IEEE TRANSACTIONS ON POWER SYSTEMS
1
Can Attackers with Limited Information Exploit
Historical Data to Mount Successful False Data
Injection Attacks on Power Systems?
arXiv:1703.07500v1 [cs.SY] 22 Mar 2017
Jiazi Zhang, Student Member, IEEE, Zhigang Chu, Student Member, IEEE, Lalitha Sankar, Senior Member, IEEE,
and Oliver Kosut, Member, IEEE
Abstract—This paper studies physical consequences of unobservable false data injection (FDI) attacks designed only with
information inside a sub-network of the power system. The goal
of this attack is to overload a chosen target line without being
detected via measurements. To overcome the limited information,
a multiple linear regression model is developed to learn the
relationship between the external network and the attack subnetwork from historical data. The worst possible consequences of
such FDI attacks are evaluated by solving a bi-level optimization
problem wherein the first level models the limited attack resources, while the second level formulates the system response to
such attacks via DC optimal power flow (OPF). The attack model
with limited information is reflected in the DC OPF formulation
that only takes into account the system information for the attack
sub-network. The vulnerability of this attack model is illustrated
on the IEEE 24-bus RTS and IEEE 118-bus systems.
Index Terms—Cyber-physical system, Cyber-security, false
data injection attacks, state estimation, multiple linear regression,
bi-level optimization.
I. I NTRODUCTION
HE electric power system is monitored via an extensive
network of sensors in tandem with data processing algorithms, i.e., an intelligent cyber layer, that enables continual
observation and control of the physical system. In recent
years, several incidents [1]–[4] demonstrate that the cyber
layer of power system is vulnerable to cyber-attacks that
impact the system operation status and lead to serious physical
consequences. Therefore, it is crucial to fully understand the
potential consequences of such attacks.
In this paper, we focus on false data injection (FDI) attacks,
in which the attacker replaces a subset of measurements
with counterfeits. The most effective FDI attacks are those
which are unobservable to state estimation (SE). It has been
established in [5]–[7] that FDI attacks can bypass both DC SE
and AC SE when designed appropriately. Such attacks can be
designed to specifically have an impact either on electricity
markets (e.g., [8], [9]) or the physical system (e.g., [10],
[11]) via a bi-level optimization problem wherein the first
level problem models the attacker’s goal and the second level
problem models the system response.
The attack optimization problem designed to find the worstcase attack in [8]–[11] requires the attacker to know systemwide information including topology, generation cost and
T
The authors are with the School of Electrical, Computer, and Energy
Engineering, Arizona State University, Tempe, AZ 85281 USA (e-mail:
{jzhan188,zchu2,lalithasankar,okosut}@asu.edu).
capacity, and load data. In practice, obtaining all the required
information can be difficult for the attacker. Recently, [12]–
[14] have demonstrated that it is possible to design FDI attacks
against SE with inaccurate or limited topology information.
However, physical consequences of the worst-case limited
information FDI attacks have not been analyzed.
In this paper, we assume the attacker only has access to
information inside an attack sub-network and absolutely no
knowledge of the outside network. In order to overcome the
limited information, we suppose that the attacker infiltrates
the sub-network long before it executes its attack, so that it
can observe the natural behavior of the system in order to
predict the effect of an attack. In particular, we assume that
the attacker has access to historical data inside the sub-network
that includes loads, costs, capacities, status, and dispatches of
generators, and locational marginal prices. We suppose that
the attacker uses multiple linear regression method to learn
the relationship between the external network and the attack
sub-network from historical data. Furthermore, we predict the
response of the control center under such attacks in a local
sub-network via a bi-level optimization problem.
The limited information attack problem is similar to the
"seamless" market problem [15] which aims to achieve maximum social welfare across several adjacent markets, while
allowing each market to model its own system, exchanging
boundary information with its neighbors. In order to predict
the behavior of the adjacent market, in [15], [16], linear regression model is used. However, in contrast to the market problem
that requires perfect prediction of the external network, the
attacker only needs partial prediction to overload a target line.
In this paper, we demonstrate that even if the prediction of the
external network re-dispatch is inaccurate, the attacker can still
cause overflow on the target line.
This paper builds upon our prior work [17], where we
consider limited information such that the attacker can learn
system parameters perfectly inside a sub-network and imperfectly outside of the sub-network. In [17], we demonstrate that
such attacks can result in line overflows with both accurate and
inaccurate information outside of the attack sub-network.
The contributions of this paper are as follows:
1) Introduce a method to compute power flows inside the
attack sub-network with only localized information by
approximating the external effects via pseudo-boundary
injections.
IEEE TRANSACTIONS ON POWER SYSTEMS
2
2) Develop a multiple linear regression model, allowing
the attacker to learn the relationship between pseudoboundary injections and power injections in the subnetwork from historical data.
3) Introduce a bi-level optimization problem to maximize
the power flow on a target line wherein the first level
models attacker’s limited resources and the second level
models system response via a modified DC OPF. Such
DC OPF formulation only takes into account localized
information and the pseudo-boundary injections.
4) Prove the existence of a linear relationship between
pseudo-boundary injections and power injections inside
the attack sub-network for certain topologies. Furthermore, we show that even if the attacker cannot exactly
predict the physical consequences, it can compute an
upper bound on the power flow of the target line.
5) Demonstrate that an attacker can cause line overflows in
the IEEE 24-bus RTS system and IEEE 118-bus system
using this bi-level attack optimization problem.
The paper is organized as follows. The system model is
introduced in Sec. II. Prior work on perfect information FDI
attack is reviewed in Sec. III. The limited information attack
model is presented in Sec. IV. Justifications for the proposed
attack strategy is given in Sec. V, followed by numerical results
in Sec. VI and conclusions in Sec. VII.
II. S YSTEM AND ATTACK M ODEL
In this section, we introduce the mathematical formulation
for SE and OPF. Throughout, we assume there are nb buses,
nbr lines, ng generators, and nz measurements in the system.
We assume that the system uses DC SE, and DC OPF. The
topology of the entire power system is denoted by G.
A. Measurement Model and State Estimation
The DC measurement model can be written as
z = Hx + e
(1)
where z is the nz × 1 measurement vector; x is the nb × 1
voltage angle state vector; H denotes the nz × nb dependency
matrix between measurements and states; e is the nz × 1 measurement error vector assumed to be composed of independent
Gaussian random variables.
We use weighted least-squares (WLS) to solve this problem
[18]. Subsequent to SE, a bad data detector uses χ2 -test to
detect and eliminate noisy measurements.
B. Optimal Power Flow
minimize
PG
X
g=1
In this section, we briefly review a closely related work [10]
on unobservable FDI attacks assuming an attacker with perfect
knowledge. As in [10], we distinguish between two types
of buses in the network: load buses that have load directly
connected to that bus, and non-load buses with no load. The
knowledge (denoted K1) and capabilities (denoted C1) of the
attacker in [10] is described below:
K1. The attacker has knowledge of (i) the complete network
topology; (ii) the cost, capacity, and operational status of
all generators in the system; and (iii) historical load data
of the entire network.
C1. The attacker may choose a small area S, which is a subgraph of the entire network G, i.e., a sub-network chosen
in certain manner (see below for description) and bounded
by load buses. The attacker may replace measurements
inside S.
A. Attack Design with Perfect Information
In [10], a bi-level optimization problem is introduced to
find the FDI attack that maximizes the power flow on a
chosen target line. In [10], the authors use B-θ method to
formulate the second level DC OPF, in which the power flow
vector is computed as a linear function of voltage angle, θ.
In contrast, in this paper we equivalently formulate the DC
OPF using PTDF, where the line power flow is calculated as
the product of PTDF matrix and power injection. The bi-level
attack optimization problem is as follows:
maximize
c,P
Pl − ζ kck0
(6)
subject to
P = K(GPG∗ − PD )
(7)
kck0 ≤ N0
(8)
− τ PD ≤ Hc ≤ τ PD
(
(9)
PG
Cg (PGg )
g=1
ng
subject to
III. P RIOR W ORK : P ERFECT K NOWLEDGE FDI ATTACKS
)
ng
X
{PG∗ } = arg minimize
Cg (PGg )
The DC OPF problem can be written as:
ng
X
where PG is the ng × 1 active power generation vector; G
is the nb × ng generator-to-bus connectivity matrix; Cg (·) is
the quadratic cost function for generator g; K is the nbr × nb
power transfer distribution factor (PTDF) matrix; the nb ×
1 real power load vector PD ; the maximum and minimum
generation limit vectors PG,max and PG,min , respectively; the
nbr × 1 thermal limit vector Pmax .
The objective (2) is to minimize the total costs of all generators. Constraints (3)−(5) represent power balance, thermal
limits, and generation limits.
PGg =
nb
X
(2)
PDi
(3)
i=1
− Pmax ≤ K(GPG − PD ) ≤ Pmax
PG,min ≤ PG ≤ PG,max
subject to
ng
nb
X
X
PDi
PGg =
g=1
(4)
(5)
(10)
g=1
(11)
i=1
− Pmax ≤ K(GPG − PD + Hc) ≤ Pmax
PG,min ≤ PG ≤ PG,max
(12)
(13)
IEEE TRANSACTIONS ON POWER SYSTEMS
where P is a nbr × 1 vector of power flow, c is a nb × 1
attack vector; τ is the load shift factor; N0 is the l0 -norm
constraint limit; H is the nb × nb dependency matrix between
power injection measurements and state variables; and ζ is the
weight of the norm of attack vector c.
The objective of the optimal attack problem is to maximize
the power flow on the target line l while changing as few
states as possible. In the first level, the attack vector is chosen
subject to the l0 -norm constraint of the attack vector in (8)
and the load shift limitation in (9). In the second level, the
system response to the attack determined in the first level is
modeled via DC OPF in (10)−(13).
The bi-level optimization problem introduced above is nonlinear and non-convex. For tractability, several constraints are
modified to convert the original formulation into an equivalent
mixed-integer linear program (MILP). The modifications include: (a) relaxing the l0 -norm constraint in (8) to an l1 -norm
constraint with limit N1 and linearizing it by introducing a
slack vector u; (b) replacing the second level DC OPF problem
by its Karush-Kuhn-Tucker (KKT) optimality conditions; and
(c) linearizing the complementary slackness conditions in
KKT by introducing a new vector δ of binary variables for
dual variables and a large constant M .
B. Attack Implementation
Once the attack vector c is determined, the attacker can
identify the buses with state changes to enable the attack, i.e.,
the buses corresponding to non-zero entries of c. We refer to
these buses as center buses. The attack subgraph S includes all
center buses as well as the lines and buses connecting to every
center bus. The non-center buses connected to a center bus are
all load buses. This method ensures that nothing is changed
outside the attack subgraph S while the changes needed at the
non-center buses are presented as load changes.
Given attacker’s knowledge K1 and capabilities C1, the
authors in [10] introduce the FDI attacks as follows:
(
zi ,
i∈
/J
a
zi =
(14)
zi + Hi c ,
i∈J
where J denotes the set of measurements inside S.
IV. O PTIMAL L INE OVERFLOW ATTACKS WITH
L OCALIZED I NFORMATION
In this section, we build upon the attack in [10] by replacing
assumptions K1 and C1 with the following limited assumptions on the attacker’s knowledge (K2) and capability (C2):
K2 Within a sub-network L, the attacker has perfect knowledge of the topology, historical load data, generator data
including operational status, capacity, cost, and historical dispatch information, and locational marginal price
(LMP). This sub-network L is bounded by load buses.
C2 The attacker may modify measurements within an attack
sub-graph S within L, i.e., S ⊆ L.
An example attack sub-network in the IEEE 24-bus RTS
system is shown in Fig. 1.
Notation: The area outside L, where the attacker has no
knowledge, is denoted as the external network, i.e., E = G \ L.
3
We define the set of boundary buses in L as B, such that
each bus in B is connected to at least one bus in E. The
set of remaining buses in L is defined as the internal bus
set I = L \ B. For any vector or matrix associated with the
entire network G such as c, G, H, K, P, PG , and PD , we write
the equivalent parameters corresponding to the sub-network
¯ For example, H̄ refers to the dependency matrix
L with (·).
between power injection measurements and state variables
only inside L. Sub-vectors are denoted by subscripts with the
corresponding set of elements (buses, lines, or generators).
Sub-matrices are denoted with a subscript giving the set of
rows, and a superscript giving the set of columns.
A. System Power Flow with Localized Information
According to assumption K2, the attacker only has knowledge inside L. Therefore, the attacker cannot calculate the
line power flow inside L with (7) since both the PTDF matrix
K of the network G and the subset of power injections in
external network E are unavailable to attacker. To form the
line power flow with K2, we introduce a vector of pseudoboundary injection P̄I,B . The ith entry of P̄I,B , namely P̄I,i ,
corresponding to boundary bus i, represents the sum of power
flows delivered from L to E at boundary bus i, i ∈ B, as
X
P̄I,i =
Pk
(15)
k∈WE
i
where WEi represents the lines located in E that are connected
to boundary bus i.
Using (15), the vector of line power flows in L can be
written as
P̄ = K̄ I (ḠI P̄G − P̄D,I ) + K̄ B (ḠB P̄G − P̄D,B − P̄I,B ) (16)
where K̄ is split into column-wise sub-matrices K̄ I and K̄ B ,
and Ḡ is split into row-wise sub-matrices ḠI and ḠB , both
corresponding to buses in I and B, respectively. This equation
can be further simplified as
P̄ = K̄(ḠP̄G − P̄D ) − K̄ B P̄I,B .
(17)
B. Multiple Linear Regression
The optimal line overflow attack introduced in Sec. III-A
involves determining the attack vector in the first level and
estimating the system response to the attack via the whole
system DC OPF in the second level. However, due to limited
knowledge, the attacker must predict the response of the OPF
using only local knowledge. The OPF may be reformulated
to include power balance, thermal limit, and generation limit
constraints only in L, and apply (17) to capture all effects in
the external network through the pseudo-boundary injections
P̄I,B . However, with this formulation, the attacker still cannot
predict how the attack affects P̄I,B since it depends on both
power injections in L and E. Therefore, before the attack is
executed, the attacker cannot estimate the system re-dispatch
after the attack accurately.
If the attacker can obtain a large amount of historical
power injections and pseudo-boundary injections data in L
(for example, by observing the system over a long time), it
IEEE TRANSACTIONS ON POWER SYSTEMS
4
21
~
~
16
19
fi,k
where fi,j , j = 0, ..., k, are regression coefficients for boundary bus i, and εi,t is random error. In the following, we let
F̂i = [fi,1 fi,2 ... fi,k ] be the coefficient vector for boundary
bus i.
Consider a problem with an m × 1 observed output vector
yi , and an m × k + 1 input matrix X = [1 x1 ... xk ]. The
relationship in (20) can be written in matrix notation as
h
iT
yi = X fˆi,0 F̂i + εi .
(21)
~
~
where Gi is the set of generators connected to bus i, and WL
i
is the set of lines in L that connected to bus i. We suppose
the attacker uses multiple linear regression to learn [fˆ0 F̂ ].
Multiple linear regression is a statistical method to find a
linear relationship between multiple inputs and single output
[19]. Take boundary bus i for an example. Let the output yi =
P̄I,i and inputs xT = ḠP̄G − P̄D . At one instance of time t,
yi,t satisfies
fi,1
fi,2
yi,t = x1,t x2,t . . . xk,t . + fi,0 + εi,t (20)
..
20
14
15
13
11
12
~
where [fˆ0 F̂ ] represent an affine relationship, and P̄ˆI,B is
the attacker’s prediction of pseudo-boundary injection by
capturing the functional relationship via a linear model. Note
that the historical pseudo-boundary injections can be computed
with data in L as
X
X
P̄k
(19)
P̄I,i =
P̄G,g − P̄D,i −
k∈WL
i
22
23
17
24
g∈Gi
~
~
~
18
can learn a functional relationship between pseudo-boundary
injection, P̄I,B , and power injections inside L. The attacker can
then predict the pseudo-boundary injections with the power
injection in L as
P̄ˆI,B = F̂ ḠP̄G − P̄D + fˆ0
(18)
3
9
10
4
8
5
1
We repeatedly use this process to obtain [fˆi,0 F̂i ] for each
i, i ∈ B. Thus, the attacker can use historical data to obtain
the estimate F̂ , such that [fˆ0 F̂ ] = [fˆi,0 F̂i ], ∀i ∈ B. The
dimension of [fˆ0 F̂ ] is nB × (k + 1), where nB is the number
of boundary buses. The attacker now can approximate (17) as
P̄ = K̄ ḠP̄G − P̄D − K̄ B F̂ ḠP̄G − P̄D + fˆ0 . (23)
In the following subsection, (23) is used to evaluate the
vulnerability to attacker with limited information.
C. Attack Optimization Problem under Localized Information
In this section, we introduce a bi-level attack optimization
problem to formulate the limited information attack. The first
level determines the attack vector in L that maximize target
line flow and the second level represents system re-dispatch
~
~
~
2
7
Boundary Bus
Fig. 1. IEEE 24-bus RTS system decomposed into attack sub-network and
attack external network.
after attack via DC OPF formulated with only information
in L. However, since the attacker does not have knowledge
of either the topology or the generator information in E,
we assume that the attacker only minimizes the total cost
of generation in L and approximates the effect of the total
generation cost in E as the total cost of the pseudo-boundary
injections in the second level modified OPF. For boundary bus
i, this cost is estimated as the product of the LMP, λi , and the
pseudo-boundary injection at bus i. The limited information
bi-level attack optimization problem is as follows:
maximize P̄l − ζ kc̄k0
(24)
c̄,P̄
subject to
∗
P̄ = K̄ ḠP̄G∗ − P̄D − K̄ B P̄I,B
kc̄k0 ≤ N0 ,
(25)
c̄B = 0
(26)
− τ P̄D ≤ H̄ c̄ ≤ τ P̄D
(27)
X
∗ ∗
X
P̄G , P̄I,B = arg minimize Cg P̄Gg +
λi P̄I,i
P̄G ,P̄I,B
g∈L
Least squares estimation (LSE) can be used to estimate the
regression coefficients F̂i in (21) as
−1 T
[fˆi,0 F̂i ]T = XT X
X yi .
(22)
6
i∈B
(28)
subject to
P̄I,B = F̂ ḠP̄G − P̄D + H̄ c̄ + fˆ0
X
X
X
P̄G,g −
P̄I,i =
P̄D,i
g∈L
i∈B
(29)
(30)
i∈L
− P̄max ≤ K̄ ḠP̄G − P̄D + H̄ c̄ − K̄ B P̄I,B ≤ P̄max (31)
P̄G,min ≤ P̄G ≤ P̄G,max
(32)
where (28) captures the modified OPF objective as the first
term represents the total cost of generation in L and the second
term is the total cost of pseudo-boundary injections. Constraint (29) represents the attacker’s prediction of the pseudoboundary injection after attack resulting from the counterfeit
loads. Note that, in the cyber system (OPF with attack vector),
the power injections in L is ḠP̄G − P̄D + H̄ c̄, thus, the
corresponding pseudo-boundary injection should respond to
these injections with attack. In addition, we have changed the
constraint on the attack vector in (26) to limit the attack to be
within the sub-network L.
IEEE TRANSACTIONS ON POWER SYSTEMS
As with the bi-level optimization problem for perfect information, (24)−(32) is non-linear and non-convex. We employ
the same modifications as detailed in Sec. III-A to convert it
into a MILP.
V. J USTIFICATION OF THE L OCALIZED I NFORMATION FDI
ATTACKS
In this section, we make a distinction between the physical
system, as it actually exists, and the cyber system, as seen by
the control center, which may differ from the physical system
due to the FDI attack. We use the superscripts p and c to
denote the physical and cyber power flows, respectively. Due
to limited information, the attacker can only use data in L to
compute the physical and cyber power flows which may be different from the actual values. Therefore, we refer the physical
and cyber power flows computed by the attacker as attackercomputed physical and cyber power flows, respectively.
We prove that: (i) there exists a linear relationship F
between pseudo-boundary injection and power injections in
L under certain circumstances; and (ii) even if F̂ does not
accurately predict the system response after attack, the attacker
can still compute an upper bound on the physical power flow
with limited information.
The following assumptions are made about the historical
data available to the attacker: (i) the topology for all the
historical data remains the same, (ii) each instance of historical
data satisfies OPF, and (iii) there exists a subset of buses
Z in E, for which power injections remain constant in the
historical data. The subset of remaining buses in E is denoted
as Y = E \ Z. In our prior work [10], [11], [17], [20], we
have shown that congested lines are more vulnerable to line
overflow FDI attacks. Analogously, in this work, we assume
the target line is congested.
A. Validation of Multiple Linear Regression Method
In this subsection, we prove the existence of linear relationship between pseudo-boundary injections and power injections
in L under certain circumstances.
For simplicity, we define the set of lines in the network
G that are the congested for each instance of historical data
as C, where C+ and C− are the subsets in C for which the
power flow directions are positive and negative, respectively.
We assume there are nc congested lines in C, nL , nE , and nY
buses in L, E, and Y, respectively.
In order to evaluate the performance of the coefficient
matrix F̂ , we define a matrix B = [KCY ; 1T ], where KCY is
the sub-matrix of K whose rows correspond to the congested
lines in C and columns correspond to the buses in Y.
Theorem 1. The coefficient matrix F̂ perfectly predicts the
pseudo-boundary injections with power injections in L linearly
if and only if B is full column rank.
Proof. We denote the vector of power injections in G as v; that
is v = GPG − PD ; the vectors vL , vE , vY , and vZ represent
the subsets of v corresponding to buses in L, E, Y, and Z,
respectively. We define Wi as the set of lines connecting to
E
boundary bus i, i ∈ B, where WL
i and Wi are the subsets
5
of lines located in L and E, respectively. We define a vector
Ji as the sum of row
P vectors in K corresponding to lines
in WEi ; that is Ji = k∈WE Kk . The matrices JiL , JiY , and
i
JiZ are the sub-matrices of J in which the columns of the
matrices corresponding to buses in L, Y, and Z, respectively.
As introduced in Sec. IV, the pseudo-boundary injection at
bus i is a linear combination of power injections at each bus
in L, Y, and Z is given by
P̄I,i = JiL vL + JiY vY + JiZ vZ .
(33)
Note that vZ is a constant across all instances of historical
data. Since each instance of historical data resulted from an
converged OPF, vL and vY satisfy the following:
KkL vL + KkY vY = Pk,max − KkZ vZ
KrL vL + KrY vY
1T vL + 1T vY
= −Pr,max −
KrZ vZ
T
= −1 vZ
∀k ∈ C+
(34)
−
(35)
∀r ∈ C
(36)
where (34) and (35) are the thermal limit constraints for
congested lines in C+ and C− , respectively, and (36) is the
power balance constraint.
Equations (34)−(36) can be collected as
AvL + BvY = d
(37)
where A = [KCL ; 1T ] and d = [SPC,max − KCZ vZ ; −1T vZ ].
The matrix S is a nc × nc diagonal matrix with Skk = 1,
∀k ∈ C+ , and Skk = −1, ∀k ∈ C− .
The dimensions of A and B are (nc + 1) × nL and
(nc + 1) × nY , respectively. Note that the number of columns
in B represents the total number of buses in Y.
Suppose that B is full column rank. Thus, B T B is
+
non-singular;
−1 T that is, there+ exists a pseudoinverse B =
T
B B
B , such that B B = I. Therefore, applying B +
to (37), the vector vY can be rewritten as
vY = −B + AvL + B + d.
(38)
The pseudo-boundary injection P̄I,i in (33) can be written as
P̄I,i = JiL − JiY B + A vL + JiY B + d + JiZ vZ .
(39)
Therefore, the linear coefficient Fi between P̄I,i and vL is
Fi = JiL − JiY B + A
fi,0 = JiY B + d + JiZ vZ .
(40)
From (40), we see that Fi is unique and is the perfect linear
predictor. The linear coefficient matrix between P̄I,B and vL
is F = [Fi ], ∀i ∈ B.
Suppose that B is not full column rank. Thus there exist infinitely many of vY satisfying (37), i.e., vY cannot be uniquely
determined by vL . Therefore, the multiple linear regression
will not perfectly predict the pseudo-boundary injections.
In Sec. VI, we provide a test case in IEEE 24-bus system
for which B is full column rank. We demonstrate that the F̂
obtained with multiple linear regression method does indeed
lead to perfect prediction of P̄I,B . We also provide three
counter examples (one in IEEE 24-bus system and the others in
IEEE 118-bus system). For these illustrated counter-examples,
B satisfies (nc + 1) < nY , which indicates that B is not
IEEE TRANSACTIONS ON POWER SYSTEMS
6
full column rank. However, even for a case with B satisfying
(nc + 1) ≥ nY , B cannot be assumed to be full column rank.
An example is a system with 3 buses in Y and 2 parallel
congested lines. For this system, rank KCY = 1 since the row
vectors in K for the parallel lines are the same. The matrix
B, hence, is not a full rank matrix since rank (B) ≤ 2 and by
Theorem 1, F̂ cannot result in an accurate prediction.
computed by the attacker, since it requires knowledge only of
the local network L and the attack vector c̄∗ .
Note that from (25) and (31), the attacker can compute the
difference between the physical and cyber power flows on
target line l (P̄lp and P̄lc , respectively) solved with limited
information attack optimization as
P̄lp − P̄lc = −K̄l H̄ c̄∗ .
B. Upper Bound on Physical Consequences of Attack
Although F̂ in general cannot accurately predict P̄I,B when
B is not full column rank, the attacker can still utilize F̂ in
the bi-level attack optimization problem (24)−(32) to predict
the physical power flow on target line. However, the attackercomputed physical power flow may not match the physical
power flow. The following theorem shows that even so, the
attacker can compute an upper bound Plub on the physical
power flow on the target line subsequent to an attack.
Theorem 2. The physical power flow on the target line l
resulting from attack vector c̄∗ is upper bounded by
Plub = Pl,max − K̄l H̄ c̄∗ .
(41)
Proof. Solving the attack optimization problem (24)−(32), the
attacker can obtain the optimal attack vector c̄∗ . The resulting
attack vector for the whole system is c∗ , where c∗i = c̄∗i for
i ∈ L and c∗i = 0 for i ∈ E. Injecting c∗ in the system
will result in a system re-dispatch determined by (10)−(13).
The difference between the physical and cyber power flows
(Plp and Plc , respectively) on target line l after the post-attack
system re-dispatch is
Plp − Plc = −Kl Hc∗ .
(42)
Thus, the physical power flow on target line l satisfies
Plp = Plc − Kl Hc∗ ≤ Pl,max − Kl Hc∗ .
(43)
where the upper bound follows from the thermal limit constraint on Plc in (12). Note that Kl and H are unknown to
the attacker with limited information. However, the attacker
has the knowledge of K̄l and H̄. We now show that the upper
bound in (43) is equivalent to Plub defined in (41).
The PTDF matrices K̄ and K satisfy the following
K̄ = Γ̄H̄ +
(44)
K = ΓH +
(45)
where Γ and Γ̄ are the dependency matrices between power
flow measurements and voltage angle states in G and L,
respectively, H + and H̄ + are the pseudoinverse of H and
H̄, respectively. Note that for target line l, both Γl and Γ̄l
have only two non-zero elements corresponding to the two
end buses of l (denoted as buses lf and lt , respectively). In
l
l
particular, Γlf = Γ̄lf = −Γllt = −Γ̄llt = x1l , where xl is the
line impedance of line l. Thus,
K̄l H̄ c̄∗ = Γ̄l c̄∗ = Γl c∗ = Kl Hc.
(46)
Therefore, the right-hand side of (43) is exactly equal to Plub .
This proves the upper bound in (41). Moreover, Plub can be
(47)
Thus, the difference between physical and cyber power flows
seen by the attacker and the system are the same
P̄lp − P̄lc = Plp − Plc .
(48)
VI. N UMERICAL R ESULTS
In this section, we illustrate the efficacy of the attacks
designed with the method proposed in Sec. IV. To this end,
we first compute the coefficient matrix with historical data
using the multiple linear regression method. Subsequently,
we solve the optimization problem to find the optimal attack
vector c̄∗ inside L. Finally, we test the physical consequences
of the attack vector c̄∗ on the entire network G. The test
systems include the IEEE 24-bus reliability test system (RTS)
and the IEEE 118-bus system. The whole network DC OPF
and limited information attack algorithm is implemented with
Matlab. The optimization problem is solved with CPLEX.
To model realistic power systems, we assume that there are
congested lines prior to the attack and the attacker chooses one
of them in L as the target to maximize power flow. This is
achieved in simulation by uniformly reducing all line ratings
by 50% for IEEE 24-bus RTS and 45% for IEEE 118-bus
system.
We illustrate our results for the following choice of parameters: the weight of the l1 -norm of attack vector in (24), ζ, is
set to 1% of the original power flow value of the target line;
and the load shift factor in (27), τ , is set to 10%. We assume
that the attacker can obtain 200 instances of historical data
inside L.
We focus on two scenarios for the historical data:
• Scenario 1 - Constant Loads in E: In each instance of
data, loads in E remain unchanged while loads in L varies
as a percent p of the base load, where p is independent
N (0, 10%). That is, power injections vary only at buses
with marginal generators (denoted EM ). Therefore, in this
scenario, the set Y introduced in Sec. V is given by Y =
EM . The number of buses in EM is denoted by nEM
• Scenario 2 - Varying Loads in the entire network G:
In each instance of data, loads in both L and E vary as a
percent p of the base load, with p chosen independently
for each load as N (0, 10%). In this scenario, power
injections at all buses in E vary in the historical data,
i.e., Y = E.
Note that the data in both scenarios also satisfy the following
assumptions: (i) the topology for all the historical data remains
the same, (ii) the historical generation dispatches data in both
scenarios satisfies OPF, and (iii) there exists a subset of buses
Z in E, for which power injections remain constant in the
historical data. An example of attack is that an attacker hacks
IEEE TRANSACTIONS ON POWER SYSTEMS
7
M a x im a l P F
S c e n a r io 2 (% )
M a x im a l P F
S c e n a r io 1 (% )
In this subsection, we present attack consequences on the
IEEE 24-bus RTS system for Scenarios 1 and 2. The subnetwork L is illustrated in Fig. 1. In each scenario, we
compare the attack consequences on target line 28 determined
by the optimization problems for two cases: (i) complete
system knowledge as in [10] (identified as global case), and
(ii) limited system knowledge (henceforth identified as local
case). For local case, we compare the physical power flow
Plp and the attacker-computed physical power flow P̄lp . The
results of attacks are illustrated in Fig. 2. We illustrate the
difference between the physical and the attacker-computed
pseudo-boundary injections in Fig. 3.
1 1 0
1 0 5
G lo b a l C a s e
L o c a l, P h y s ic a l P F
L o c a l, A tta c k e r-c o m p u te d P h y s ic a l P F
1 0 0
0 .0 0
0 .0 2
0 .0 4
(a )
0 .0 6
0 .0 8
0 .1 0
1 1 0
1 0 5
G lo b a l C a s e
L o c a l, P h y s ic a l P F
L o c a l, A tta c k e r-c o m p u te d P h y s ic a l P F
L o c a l, U p p e r B o u n d
1 0 0
0 .0 0
0 .0 2
0 .0 4
l1-n o r m
(b )
0 .0 6
0 .0 8
0 .1 0
c o n s tr a in t (r a d )
P se u d o -b o u n d a r y P se u d o -b o u n d a r y
I n je c tio n E r r o r
I n je c tio n E r r o r
S c e n a r io 2 ( p .u .) S c e n a r io 1 ( p .u .)
Fig. 2. The maximum power flow (PF) v.s. the l1 -norm constraint (N1) when
target line is 28 of IEEE 24-bus system for (a) Scenario 1, and (b) Scenario
2 historical data.
0 .0
0 .0
0 .0
-0 .0
-0 .0
0 .0
0 .0
0 .0
0 .0
0 .0
-0 .0
4
2
0
2
B u s 1 9
4
0 .0 2
0 .0 4
(a )
B u s 1 5
0 .0 6
B u s 1 4
0 .0 8
4
0 .1 0
1
B u s 1 4
0
1
0 .0 2
0 .0 4
l1-n o r m
(b )
0 .0 6
0 .0 8
1 0 6
1 0 4
L o c a l, P h y s ic a l P F
L o c a l, A tta c k e r-c o m p u te d P h y s ic a l P F
L o c a l, U p p e r B o u n d
1 0 2
1 0 0
0 .0
0 .2
0 .4
(a )
0 .6
0 .8
1 .0
1 0 8
1 0 6
1 0 4
L o c a l, P h y s ic a l P F
L o c a l, A tta c k e r-c o m p u te d P h y s ic a l P F
L o c a l, U p p e r B o u n d
1 0 2
1 0 0
0 .0
0 .2
0 .4
(b )
0 .6
0 .8
1 .0
c o n s tr a in t (r a d )
Fig. 4. The maximum power flow (PF) v.s. the l1 -norm constraint (N1) when
target line is 5 of IEEE 118-bus system for (a) Scenario 1, and (b) Scenario
2 historical data.
B u s 1 5
2
1 0 8
l1-n o r m
B u s 1 9
3
B. Results for IEEE 118-bus System
In this subsection, we test the consequences of attacks on the
IEEE 118-bus system. The details of sub-network L are listed
in Table II. The results of attacks designed with historical data
in Scenarios 1 and 2 are illustrated in Fig. 4 with sub-plots (a)
and (b), respectively. The difference between the physical and
the attacker-computed pseudo-boundary injections at 3 of 18
boundary buses, buses 23, 70, and 80, for both scenarios are
illustrated in Fig. 5 with sub-plots (a) and (b), respectively.
The parameters of the test system are also summarized in
M a x im a l P F
S c e n a r io 1 (% )
A. Results for IEEE 24-bus RTS system
perfectly predicted by the multiple linear regression method,
which explains why the attacker-computed system response
post-attack is the same as the actual response, as illustrated in
Figs. 2(a) and 3(a).
Furthermore, Table I shows that for the historical data
in Scenario 2, rank(B) < nE . Thus, by Theorem 1, the
predictions of pseudo-boundary injections by the multiple
linear regression are not accurate and there will be mismatches
between the actual and the attacker-computed system response
post-attack. This is verified by the non-zero pseudo-boundary
power injection differences shown in Fig. 3(b). In Fig. 2(b),
in addition to plotting the attacker-computed physical power
flow, we also plot the upper bound on physical power flow.
From Fig. 2(b), we observe that although there are mismatches
between the actual and attacker-computed system response
under attack, the upper bound found in Sec. V-B exactly
matches the physical power flow.
M a x im a l P F
S c e n a r io 2 (% )
into the system and collects the data from 12:00 p.m. to 2:00
p.m. for the entire month of July and then launches a FDI
attack at the end of the month.
0 .1 0
c o n s tr a in t (r a d )
Fig. 3. The pseudo-boundary power injection error v.s. the l1 -norm constraint
(N1) when target line is 28 of IEEE 24-bus system for (a) Scenario 1, and
(b) Scenario 2 historical data.
In Figs. 2(a) and (b), we note that the solutions for the
local case is sub-optimal relative to that for the global case.
The reason is that as N1 is relaxed, getting a larger overflow
on the target line requires measurements in both L and E to be
modified. Therefore, the constraint on limited attack resources
prevents any further increase in the maximal target line flow
for the local case.
The parameters of the test system are summarized in Table
I. The historical data in Scenario 1 satisfies rank(B) = nEM .
Thus, by Theorem 1, the pseudo-boundary power injections are
Table I. Note that for historical data in both scenarios, B
does not have full column rank. Therefore, Theorem 1 predicts
a mismatch between physical and attacker-computed pseudoboundary injections. This is verified by Fig. 5, which shows
the pseudo-boundary injection error. In Figs. 4(a) and (b),
we find that in both scenarios, both the attacker-computed
physical power flow and the upper bound match the physical
power flow. This case demonstrates that even though there are
mismatches between physical and attacker-computed pseudoboundary injections, the attacker-computed physical power
flow can still be correct. Note that, in this case, both the cyber
power flow and the attacker-computed cyber power flow reach
the limit post-attack since the target line is congested before
attack. Therefore, from (48), the attacker-computed physical
power flow is the same as the physical power flow.
IEEE TRANSACTIONS ON POWER SYSTEMS
8
Table I
S UMMARY OF THE T EST S YSTEMS
Scenario 1
P se u d o -b o u n d a r y P se u d o -b o u n d a r y
I n je c tio n E r r o r
I n je c tio n E r r o r
S c e n a r io 2 ( p .u .) S c e n a r io 1 ( p .u .)
0 .1
0 .1
0 .0
0 .0
-0 .0
-0 .1
-0 .1
-
Scenario 2
Test System
# of Congested
Lines (nc )
# of Buses in
EM (nEM )
rank(B)
# of Congested
Lines (nc )
# of Buses
in E (nE )
rank(B)
24-bus
2
2
2
2
16
3
118-bus
3
5
4
3
71
4
5
B u s 2 3
0
R EFERENCES
B u s 8 0
5
0
5
B u s 7 0
0
5
0 .1
0 .2
0 .3
0 .4
0 .5
0 .6
0 .7
0 .8
0 .9
1 .0
(a )
0 .2
0 .1
0 .1
0 .0
0 .0
0 .0
0 .1
0 .1
0 .2
0
5
0
B u s 2 3
5
0
B u s 8 0
0
B u s 7 0
5
5
0
0 .1
0 .2
0 .3
0 .4
0 .5
0 .6
0 .7
0 .8
0 .9
1 .0
(b )
l1-n o r m
c o n s tr a in t (r a d )
Fig. 5. The pseudo-boundary power injection error v.s. the l1 -norm constraint
(N1) when target line is 5 of IEEE 118-bus system for (a) Scenario 1, and
(b) Scenario 2 historical data.
Table II
S UMMARY OF THE ATTACK S UB - NETWORK IN IEEE 118- BUS S YSTEM
Buses
1-14, 16, 17, 23, 25-27, 30, 33-35, 37-40, 47, 49, 59-66,
68-70, 75, 77, 80, 81, 116, 117
Lines
1-17, 20, 22, 31-33, 36-38, 47, 48, 50-55, 65, 88-100, 102,
104-108, 115, 116, 119, 120, 123, 124, 126, 127, 183, 184
Boundary
Buses
13, 14, 17, 23, 27, 33-35, 40, 47, 49, 59, 62, 66, 70, 75,
77, 80
VII. C ONCLUSION
In this paper, we have studied the physical system consequences of unobservable FDI attacks that are designed only
using information inside an attack sub-network. We have
introduced pseudo-boundary injections to represent the power
flows delivered from the external network and developed a
multiple linear regression model for the attacker to learn the relationship between pseudo-boundary injections and the power
injections inside the attack sub-network. We have formulated
a bi-level optimization problem to maximize the power flow
on a chosen target line with attacker’s perfect information
in the attack sub-network as well as the predicted pseudoboundary injections. Our results illustrate that the attacker can
overload transmission lines with the proposed bi-level attack
optimization problems. In conclusion, one must be concerned
that even with limited information, an attacker with access to
historical data can take advantage of it to the detriment of
reliable system operations.
ACKNOWLEDGMENT
This material is based upon work supported by the National
Science Foundation under Grant No. CNS-1449080.
[1] M. Zeller, “Myth or reality – does the aurora vulnerability pose a
risk to my generator?” in 64th Annual Conference for Protective Relay
Engineers, April 2011, pp. 130–136.
[2] “The Stuxnet worm: A cyber-missile aimed at Iran,” The Economist,
Tech. Rep., 24 September 2010. [Online]. Available: http://www.
economist.com/blogs/babbage/2010/09/stuxnet_worm
[3] T. Espiner, “Siemens: Stuxnet infected 14 industrial plants,”
http://www.zdnet.com/article/siemens-stuxnet-infected-14-industrialplants/, September 2010.
[4] K. Zetter, “Inside the cunning, unprecedented hack of Ukraine’s power
grid,”
http://www.wired.com/2016/03/inside-cunning-unprecedentedhack-ukraines-power-grid/, March 2016.
[5] Y. Liu, P. Ning, and M. K. Reiter, “False data injection attacks against
state estimation in electric power grids,” in Proceedings of the 16th ACM
Conference on Computer and Communications Security, ser. CCS ’09,
Chicago, Illinois, USA, 2009, pp. 21–32.
[6] G. Hug and J. A. Giampapa, “Vulnerability assessment of AC state
estimation with respect to false data injection cyber-attacks,” IEEE
Transactions on Smart Grid, vol. 3, no. 3, pp. 1362–1370, 2012.
[7] J. Liang, O. Kosut, and L. Sankar, “Cyber-attacks on AC state estimation: Unobservability and physical consequences,” in IEEE PES General
Meeting, Washington, DC, July 2014.
[8] Y. Yuan, Z. Li, and K. Ren, “Modeling load redistribution attacks in
power systems,” Smart Grid, IEEE Transactions on, vol. 2, no. 2, pp.
382–390, June 2011.
[9] ——, “Quantitative analysis of load redistribution attacks in power
systems,” Parallel and Distributed Systems, IEEE Transactions on,
vol. 23, no. 9, pp. 1731–1738, Sept 2012.
[10] J. Liang, L. Sankar, and O. Kosut, “Vulnerability analysis and consequences of false data injection attack on power system state estimation,”
IEEE Transactions on Power Systems, vol. 31, no. 5, pp. 3864–3872,
Sept 2016.
[11] J. Zhang and L. Sankar, “Physical system consequences of unobservable
state-and-topology cyber-physical attacks,” IEEE Transactions on Smart
Grid, vol. 7, no. 4, pp. 2016–2025, July 2016.
[12] M. A. Rahman and H. Mohsenian-Rad, “False data injection attacks with
incomplete information against smart power grids,” in 2012 IEEE Global
Communications Conference (GLOBECOM), Dec 2012, pp. 3153–3158.
[13] X. Liu and Z. Li, “Local load redistribution attacks in power systems
with incomplete network information,” IEEE Transactions on Smart
Grid, vol. 5, no. 4, pp. 1665–1676, July 2014.
[14] ——, “False data attacks against AC state estimation with incomplete
network information,” IEEE Transactions on Smart Grid, vol. PP, no. 99,
pp. 1–10, 2016.
[15] “PSERC
final
project
report:
Electricity
market
structures to reduce seams and enhance investment,” February
2010. [Online]. Available: https://pdfs.semanticscholar.org/2828/
e5f13d9cfca4662b2c662862b849ac60a366.pdf
[16] M. Farrokhseresht, N. Farrokhseresht, and M. R. Hesamzadeh, “Optimizing the net transfer capacity of a cross-border interconnector by
reactive power planning,” in 2016 IEEE Power and Energy Society
General Meeting (PESGM), July 2016, pp. 1–5.
[17] J. Zhang, Z. Chu, L. Sankar, and O. Kosut, “False data injection attacks
on power system state estimation with limited information,” in 2016
IEEE Power and Energy Society General Meeting (PESGM), July 2016,
pp. 1–5.
[18] A. Abur and A. G. Exposito, Power System State Estimation: Theory
and Implementation. New York: CRC Press, 2004.
[19] D. C. Montgomery, E. A. Peck, and G. G. Vining, Introduction to Linear
Regression Analysis, 5th ed. Wiley, 2012.
[20] Z. Chu, J. Zhang, O. Kosut, and L. Sankar, “Evaluating power system
vulnerability to false data injection attacks via scalable optimization,”
IEEE TRANSACTIONS ON POWER SYSTEMS
in 2016 IEEE International Conference on Smart Grid Communications
(SmartGridComm), Nov 2016, pp. 260–265.
9
| 3cs.SY
|
Fiber-Flux Diffusion Density for White Matter
Tracts Analysis: Application to Mild Anomalies
Localization in Contact Sports Players
arXiv:1709.06122v1 [cs.CV] 18 Sep 2017
Itay Benou1,3 , Alon Friedman2,3,4 , and Tammy Riklin Raviv1,3
1
4
Department of Electrical Engineering,
Ben-Gurion University of the Negev, Beer-Sheva, Israel
2
Department of Physiology and Cell Biology,
Ben-Gurion University of the Negev, Beer-Sheva, Israel
3
The Zlotowski Center for Neuroscience,
Ben-Gurion University of the Negev, Beer-Sheva, Israel
Departments of Medical Neuroscience and Brain Repair Centre,
Dalhousie University, Faculty of Medicine, Halifax, Canada
Abstract. We present the concept of fiber-flux density for locally quantifying white matter (WM) fiber bundles. By combining scalar diffusivity
measures (e.g., fractional anisotropy) with fiber-flux measurements, we
define new local descriptors called Fiber-Flux Diffusion Density (FFDD)
vectors. Applying each descriptor throughout fiber bundles allows alongtract coupling of a specific diffusion measure with geometrical properties,
such as fiber orientation and coherence. A key step in the proposed framework is the construction of an FFDD dissimilarity measure for sub-voxel
alignment of fiber bundles, based on the fast marching method (FMM).
The obtained aligned WM tract-profiles enable meaningful inter-subject
comparisons and group-wise statistical analysis. We demonstrate our
method using two different datasets of contact sports players. Alongtract pairwise comparison as well as group-wise analysis, with respect
to non-player healthy controls, reveal significant and spatially-consistent
FFDD anomalies. Comparing our method with along-tract FA analysis shows improved sensitivity to subtle structural anomalies in football
players over standard FA measurements.
1
Introduction
WM tractography from diffusion tensor imaging (DTI) is an efficient tool for
longitudinal analysis and group studies, in particular when standard magnetic
resonance imaging (MRI) is not sufficiently sensitive to detect subtle structural
anomalies, such as in mild traumatic brain injury (mTBI) [24]. The fiber bundles
rendered by tractography, in the form of streamline 3D coordinates, can be represented by geometrical properties as well as diffusivity measures (e.g., fractional
anisotropy - FA, mean diffusivity - MD, axial diffusivity - AD, radial diffusivity
- RD). Nevertheless, coherent mathematical modeling of the bundles, for alongtract pair-wise comparison and group-wise analysis, is a challenging task. The
main difficulty is finding a common parameterization to faithfully represent the
many fibers within a single bundle, and to match different bundles.
A straight-forward parameterization considers the natural grid of the images. Often, voxel-based registration of the MRI volumes is performed prior to
the modeling. However, whole-brain registration does not guarantee an optimal
alignment between corresponding fiber tracts due to large topological differences [12,21,28]. Therefore, most along-tract analysis approaches use arc-length
(equidistant) re-parameterization prior to quantitative analysis [7,16,19,26], and
sometimes use anatomical landmarks [17] or crop the tract edges [28] to refine
the alignment. Alternatively, tractography-based registration methods directly
align sets of fibers based on their geometry and shape, using their streamline 3D
coordinates, e.g., [12,20].
A different paradigm considers parameterization that is intrinsic to specific
bundles. Yushkevich et al. [31] used a parametric medial-surface representation
of thin sheet-like fiber structures, by projecting the volumetric data into a 2manifold. In a similar manner, tube-like shaped fiber bundles were modeled by
their average (midline) trajectory in [5,8,11]. A more recent method suggests
using manifold learning to achieve joint parameterization of fiber bundles, by
mapping corresponding tracts across subjects into a latent bundle core [15].
Other approaches circumvent the parameterization problem altogether. In [9], a
metric on WM fiber bundles is defined by the path integral of the fibers modeled as currents with an optimally constructed vector field. This approach has
been extended in [4], using varifolds. However, [4,9] do not provide along-tract
analysis.
The contribution of the proposed framework is two-fold, referring to both
fiber bundle modeling and alignment. Aiming to perform quantitative along-tract
analysis, we introduce the concept of Fiber-Flux Diffusion Density (FFDD) descriptors that couple the bundle’s geometry with local diffusivity measures. This
allows diffusion-related features to be accounted for, as well as structural variations along tracts, which may not be reflected by diffusion scalars alone. Fiber
bundle modeling, in the form of tract-profiles, is obtained by application of these
descriptors along the mean trajectory of the bundle. Fiber tracts alignment is addressed as a curve matching problem between the mean trajectories of the tracts.
The proposed dissimilarity measure is based on FFDD tract profiles, thus utilizing both diffusional and geometrical information for the alignment task, rather
than relying exclusively on geometrical properties (e.g., arc-length and curvature)
as in classical curve matching algorithms [22,30], or on scalar measurements as in
FA-based registration methods [1,25]. Moreover, unlike traditional curve matching approaches [6,30], we do not map one curve into another. Instead, we adapt
the FMM5 for curve alignment [10], to symmetrically match pairs of tracts with
sub-voxel accuracy, based on FFDD dissimilarities as an inverse speed map. The
proposed alignment framework plays a key role in the construction of standardized FFDD profiles that can be considered as a bundle-specific atlas. This atlas
5
The FMM was proposed by Sethian [23] for solving boundary value problems of the
Eikonal equation.
facilitates group-wise statistical analysis for the assessment and localization of
abnormalities in WM fiber tracts.
We demonstrate the validity of our method by performing a tract-specific
longitudinal analysis of a basketball player diagnosed with occipital mTBI and a
frontal hemorrhage, having scans one week and 6 months post-injury. We further
conduct a cross-sectional study, comparing 13 professional American-football
players with possible mTBIs, with 17 normal control (NC) subjects. The analysis includes five major white matter tracts: the left and right fronto-occipital
fasciculus (IFOF), left and right corticospinal tract (CST), and the forceps minor tract (FMT). Substantial FFDD abnormalities were found in several football
players compared to controls, mostly located at the occipital part of the IFOF
and at the central part of the forceps minor. The same regions also demonstrate
statistically significant FFDD differences between the groups, indicated by low
p-values and high standard deviation (STD). For some players, repeated scans
revealed consistent and increased FFDD anomalies with time, even after a few
weeks off-season. Results are in line with mTBI findings from DTI [14]. We also
demonstrate that the proposed FFDD method provides improved sensitivity to
subtle structural anomalies compared to along-tract FA analysis, due to the use
of additional geometric information.
The rest of the paper is organized as follows: Section 2 presents the FFDD
descriptors, followed by an introduction of the proposed framework for fiber bundles alignment and statistical analysis. Section 3 presents experimental results
for two different datasets of contact-sport players. We conclude in section 4.
2
2.1
Methods
Fiber-Flux Diffusion Density Descriptor
A fiber bundle B can be thought of as a set of similar trajectories with a common
origin and destination, along which water molecules are diffused [13]. In the
spirit of this notion, we define a local measure for quantifying the fiber-flux of
B through a given plane π, with normal n̂π (p) at point p ∈ π, i.e.,
Np
1 X
τ̂i (xi ) · n̂π (p) ,
FB (π; p) =
Np i=1
(1)
where Np is the number of intersected fibers, ξ = {xi } is the set of intersection
points between the plane and the fiber bundle, and {τ̂i (xi )} are the tangents
of the fibers at those points. We call FB (π; p) the fiber-flux density (FFD) of
bundle B at point p. The plane π is oriented such that the fiber-flux is maximized, i.e., n̂π (p) = arg max FB (π; p). We use an iterative approach to solve
n̂π
this maximization problem in the spirit of [27]. We further introduce diffusivity
properties into our model by extending the FFD measure. Let S (xi ) define a
diffusivity scalar of choice (FA, MD, AD, or RD), associated with the point xi .
We define the fiber-flux diffusion density (FFDD) as follows:
JB (π; p) =
Np
1 X
S (xi ) τ̂i (xi ) · n̂π (p)
Np i=1
(2)
In practice, we refer to the FFDD as a vector J B (p) = JB (p)n̂ (p) to account
for the local orientation of the fiber bundle. Note that the set of four FFDD
descriptors (each assigned with a different diffusivity measure) couples diffusion
measures with local geometrical features of the bundle. For example, local differences in orientation are taken into account, and regions with “incoherent” fiber
orientations are “punished” by having lower FFDD values.
2.2
Along Tract Profiles
We calculate the mean fiber of the bundle c(s) = (x (s) , y (s) , z (s)), where s
is its arc-length parameter, based on Fourier descriptor [5]. According to this
method, individual streamline fibers are represented by the coefficients of cosine series expansions, which are computed from tractography data using least
squares estimation. The mean fiber is then optimally obtained by averaging the
representation coefficients and applying the inverse transformation. The locations of the planar cross-sections along the bundle are determined by equidisM
tant sampling points along the mean fiber {pm } = {c(sm )}m=1 . Tract-profiles
{J B (pm )}m are obtained by applying the FFDD descriptors along the tract, over
these points.
2.3
Fiber Bundles Alignment
We address the alignment of two bundles B1 and B2 as a curve-matching problem
between their mean fibers c1 (s1 ) and c2 (s2 ), where s1 ∈ [0 , L1 ] and s2 ∈ [0 , L2 ]
are the respective arc-length parameterizations. We adapt the FMM-based symmetrical curve matching framework of [10] to allow sub-sampling resolution of
the alignment. Nevertheless, rather than using geometrical properties alone for
the construction of the inverse speed map F (s1 , s2 ), we propose a new dissimilarity measure which relies on the FFDD profiles:
F (s1 , s2 ) = kJ B1 (s1 ) − J B2 (s2 )k + λ
(3)
where λ is a scalar used for regularization, set as in [10]. Given F (s1 , s2 ), the
FMM solves the Eikonal equation |∇T (s1 , s2 )| = F (s1 , s2 ) ∀ s1 , s1 , providing as
output the weighted distance matrix T (s1 , s2 ) . Fig.1a-b present F (s1 , s2 ) and
T (s1 , s2 ), respectively. The optimal alignment is then defined by the shortest
path in F (s1 , s2 ) from the starting point (0, 0) to the endpoint (L1 , L2 ). The
alignment path α (τ ) = (s1 (τ ) , s2 (τ )) defines pairs of matching points between
the bundles, and is computed with sub-voxel resolution as follows:
α (τ − ε) = α (τ ) − ε∇T (s1 , s2 ) ; α (L) = (L1 , L2 )
(4)
(a) F (s1 , s2 )
(b) T (s1 , s2 )
(c) Compute α
(d) Aligned Curves
Fig. 1. Alignment using FMM. (a) Local dissimilarities F (s1 , s2 ) based on FFDD
profiles. (b) T (s1 , s2 ) is obtained by solving the Eikonal equation. (c) The alignment
path α (τ ) is computed by backtracking along the gradients of T (s1 , s2 ). (d) Resulting
alignment (color-coded).
as illustrated in Fig. 1c. The step size ε is usually set to some small value
M
(ε 1). For uniformity, we re-sample α into M samples, i.e., {α (τ m )}m=1 , such
M
m
that the aligned mean fibers are obtained by C̃1 = {c1 (s1 (τ ))}m=1 and C̃2 =
M
{c2 (s2 (τ m ))}m=1 (see Fig. 1d), and their tract-profiles are aligned accordingly:
M
M
J˜B1 = {J B1 (c1 (s1 (τ m )))}m=1 and J˜B2 = {J B2 (c2 (s2 (τ m )))}m=1 .
2.4
Along-Tract Variability Analysis
Pairwise Comparison: Let J˜B1 and J˜B2 be a pair of aligned tract-profiles to
be compared, e.g., of a subject-specific tract in two longitudinal scans. We define
a pointwise dissimilarity measure between them as follows:
dJ (B1 , B2 ; α (τ m )) = kJ B1 (c1 (s1 (τ m ))) − J B2 (c2 (s2 (τ m )))k
(5)
Although we focus here on computing local dissimilaritiesZalong the two bundles,
global dissimilarity can also be calculated: DJ (B1 , B2 ) =
dJ (B1 , B2 ; α (τ )) dα.
α
Group-Wise Statistical Analysis: Alignment of multiple fiber tracts for
Ng
group-wise analysis is performed as follows. Let Jg = {J Bn (s)}n=1
denote the
Ng
set of Ng tract-profiles of a group of subjects, and let Cg = {Cn (s)}n=1 denote
their respective mean fibers with a joint arc-length parameterization s. We define
a reference tract profile, with its corresponding mean fiber as follows:
J ref (s) =
Ng
Ng
1 X
1 X
J Bn (s) , Cref (s) =
Cn (s)
Ng n=1
Ng n=1
(6)
Alignment of the tract-profiles is obtained by first mapping each of them to the
reference tract as discussed in Section 2.3. We then interpolate the resulting
Ng
alignment paths {α (τn )}n=1
such that they all contain the same set of M samM
ples of the reference tract {Cref (sm )}m=1 . We construct a bundle-specific atlas
by pointwise averaging the aligned tract-profiles. This atlas represents the standardized tract-profile of the group, which is used as a benchmark for group-wise
statistical analysis.
3
Experimental Results
We demonstrate our FFDD method on two different datasets of contact-sports
players. Normal control (NC) group includes scans of healthy age-matched males.
Diffusion weighted images (DWI) of all subjects were acquired on a 3T Philips
Ingenia scanner using a single-shot, spin-echo, echo-planar imaging (EPI) sequence (TE=106 ms, TR=9000 ms, FOV = 224x224x120 mm). A total of 60
2[mm]-thick slices were acquired with 33 different gradient directions (b=1000
s/mm2 ) with a voxel resolution of 1.75×1.75×2 mm. Pre-Processing included
rigid alignment to the SPM MNI T1-template; motion and eddy currents correction; DTI tensor model fitting [2]; and Tractography of five major tracts: left
and right IFOF, left and right CST, and the FMT [29]. All performed by DSI
Studio software (http://dsi-studio.labsolver.org/). The tracts were delineated by placing multiple regions of interest (ROIs) from the JHU WM atlas [18].
3.1
Longitudinal Case Study
We performed pairwise comparison between two scans of a 32-year-old basketball player, diagnosed with mild occipital traumatic brain injury and frontal
hemorrhage due to contrecoup impact, acquired one week and 6 months postinjury. The hemorrhagic lesion at the frontal right hemisphere of the player is
no longer visible in the FLAIR image acquired 6 months after injury (Fig. 2a).
Local differences between corresponding, longitudinal FA- and MD-FFDD profiles of the FMT (chosen due to its proximity to the lesion area) are shown in
Fig. 2d. Figs. 2b-c present color-coded FMT to visually demonstrate these differences. Results show significant longitudinal variability at the right hemisphere
part of the tract, corresponding to the lesion area, and relatively minor differences along the rest of the tract. These results should be considered as a proof of
concept, validating the FFDD analysis results for the detection and localization
of mTBI-related variabilities between fiber bundles.
3.2
Football Players Study
We analyzed 13 active professional American-football players (mean age = 28.3,
STD = 6.4), with respect to 17 NCs (mean age = 26.1, STD = 2.3). For each
subject, four FFDD tract-profiles were computed (based on FA, MD, RD, and
AD), for each of the five examined tracts. The standardized FFDD profiles of
NCs are shown in Fig. 3. Note that although FFDD values vary along the tracts,
their profiles are consistent across subjects.
Fig. 4 presents pointwise group-average and STD of MD-FFDD profiles of
football players, demonstrating increased values at the occipital part of the left
IFOF, and at the central part of the FMT, compared to NCs. Note that the football group also exhibits higher STD values compared to NCs, at the same areas
ONE WEEK
MD-FFDD
6 MONTHS
FA-FFDD
(a) FLAIR
(b) FA-FFDD (c) MD-FFDD
(d) Local Differences
Fig. 2. Longitudinal FFDD analysis of the FMT. Right Panel: FA- and MD-based
tract-profiles of both scans. Local differences are color-coded on the graph. Left Panel:
FLAIR scans (axial slices), one week post-injury (top) and 6 months post-injury
(bottom). (a) Highlighted boxes around the hemorrhaging area - lesion is no longer
visible 6 months after injury. In (b) and (c) the tracts are colored-coded by the magnitude of their FFDD profiles. Regions with high longitudinal variability (marked in red
and green ellipses) correspond to the lesion area.
along the tracts with increased group-average values. This statistical spread indicates that only a subset of the football players group has abnormal FFDD values,
as expected. Our method also demonstrates statistically significant differences
between the football and control groups (p-values < 0.05) at these regions of
the IFOF and FMT, as shown in Fig. 5. The left panel of the figure presents
an along-tract p-values analysis of the two tracts, calculated pointwise based on
four different FFDD tract profiles and corrected for multiple comparisons using
false detection rate (FDR) [3]. As reference, the right panel of the figure presents
a p-values analysis based on whole tract average of conventional diffusivity measures extracted via DSI Studio, calculated using an unpaired T-test, which also
shows statistically significant differences between the groups in the IFOF and
FMT for some diffusion measures. These findings are further supported by a
group-wise statistical analysis (mean and STD) of whole-tract average diffusivity measures (MD, AD, and RD), presented in Fig. 6. Results are in line with
the FFDD analysis, demonstrating increased group-average diffusivity in the left
IFOF and FMT of the football group compared to NCs. The figure also indicates the maximal value measured within the football group for each diffusivity
measure. Note that for the left IFOF, player #11 demonstrates maximal values
across all diffusivity measures, while the same applies to the FMT of player #12.
We note that the CST did not present significant differences between the groups,
in both FFDD and conventional analysis.
Right IFOF
Left CST
Right CST
Forceps Minor
RD-FFDD
AD-FFDD
MD-FFDD
FA-FFDD
Tract
Left IFOF
Fig. 3. Top row: The 5 examined tracts. Bottom rows: Four types of standardized
tract-profiles (colored): FA-FFDD, MD-FFDD, AD-FFDD, and RD-FFDD constructed
from the aligned tract-profiles of NCs (gray).
∆Mean
Pointwise STD
∆STD
Left IFOF
FMT
Pointwise Mean
Fig. 4. Group-wise statistical analysis based on MD-FFDD. Pointwise comparison of
within-group average profile (left) and STDs (right) are presented for the FMT (top)
and the left IFOF (bottom).
Left IFOF
Forceps Minor
Tracts-Average Analysis
Fig. 5. Group-wise p-values analysis. Left and middle: pointwise (corrected) p-values
along the left IFOF and FMT, based on FFDD profiles. Most statistically significant
differences between the groups (p-value<0.05) are located in the occipital part of the
IFOF (samples #60 to #90) and central part of the FMT (samples #40 to #60).
Right: scalar p-values based on tract-average diffusion measures (FA, MD, AD, and
RD) of five different tracts. Statistically significant differences between the groups are
shown for the left IFOF and FMT, in MD, AD, and RD measures.
AD Analysis
RD Analysis
Forceps Minor
Left IFOF
MD Analysis
Fig. 6. Box-plot group-wise statistical analysis of average diffusivity measures in the
left IFOF (top) and FMT (bottom). Each box represents the distribution of values
within the group: the red line represents the group-average value, the edges of the
box represent the 25th and 75th precentiles of the group, and the edges of the dashed
line represent the maximal and minimal values. The x symbol represents the maximal
value measured in the football group. Note that for both tracts, the football group
consistently demonstrates higher values across all diffusivity measures.
Experiments also showed significant FFDD longitudinal changes between
mid-season and post-season scans in some football players. Fig. 7 presents midand post-season MD-FFDD profiles comparison of the left IFOF of one of the
players, showing increased irregularities over time at the occipital part of the
tract. Fig. 8 presents a similar MD-FFDD longitudinal analysis of the FMT of a
different player, showing increased irregularities at the central part of the tract.
Mid vs. Post
Mid-season
Post-season
STDs
Fig. 7. MD-FFDD longitudinal changes in the left IFOF of player #11. Left: Midseason and post-season tract-profiles in comparison to normative values (±1 STD from
standardized profile of NC). Middle and Right: Mid- and post-season abnormalities
are color-coded along the tract (in units of #STDs from standardized profile). In the
occipital area, mid-season profile demonstrates moderate abnormality (up to 2.5 STDs)
while in post-season substantial abnormality (up to 6.5 STDs) is shown.
Mid vs. Post
Mid-season
Post-season
STDs
Fig. 8. MD-FFDD longitudinal changes along the FMT of player #12. Left: Midseason and post-season tract-profiles are compared to normative values (±1 STD from
standardized profile of NCs). Middle and Right: Mid- and post-season abnormalities
are color-coded along the tract (in units of #STDs from standardized profile). While
mid-season profile demonstrates abnormality (up to 3.5 STDs) in a small region around
the center of the tract, the post-season profile shows expansion in both magnitude (up
to 4.5 STDs) and location of abnormality along the tract.
3.3
Sensitivity Analysis
In order to demonstrate the improved sensitivity of the proposed method in
anomalies detection, we compared our FFDD groupwise analysis to an existing
approach of along-tract analysis based on standard FA measurements. Similar
to the FFDD analysis, normative values of standard FA were obtained by computing the pointwise mean and STD along the aligned FA profiles of NCs. The
average FA profile of the football players group is then compared to these normative values. Fig. 9 presents a comparison between standard FA analysis and
FA-FFDD analysis for the left and right IFOF. The comparison shows that while
both methods yield similar results at the frontal and central parts of the tracts,
the FA-FFDD analysis shows higher variation from NCs (1 STD) compared to
Left IFOF
(b) Variability from NCs
Right IFOF
(a) Football vs. NCs
(b) Variability from NCs
FA-FFDD
(a) Football vs. NCs
Standard FA
STDs
Fig. 9. Comparison between groupwise statistical analysis based on FA-FFDD (top
row) and standard FA (bottom row) for the left and right IFOF. (a) Pointwise average
profile of post-season football group compared to normative values (±1 STD from
standardized profile of NCs). The red arrow points to the area in which FA-FFDD
presents improved sensitivity in comparison to standard FA. (b) Deviation from NCs
is color-coded along the tract (in units of #STDs from standardized profile).
Left IFOF
(a) Football vs. NCs
(b) Pointwise differences
Right IFOF
(a) Football vs. NCs
(b) Pointwise differences
Fig. 10. Groupwise statistical analysis based on geometric fiber-flux density (FFD)
alone, with no diffusion measurements, for the left and right IFOF. (a) Pointwise comparison of within-group average FFD profiles between post-season football players and
NCs. (b) Pointwise differences between the two profiles are color-coded along the tracts.
The comparison shows increased flux-density variability between the groups at the occipital part of the tracts, allowing for increased sensitivity of the FFDD analysis in this
area.
standard FA (0.5 STD) at the occipital part of the tracts. Note that this finding is
demonstrated symmetrically for the left and right IFOF, at a spatially-consistent
location with the cross-sectional and longitudinal results presented earlier in this
section. The improved sensitivity of FA-FFDD in this region over standard FA
is due to the additional geometric information provided by the proposed descriptor, as demonstrated in Fig. 10, which shows the group-average fiber-flux
density (FFD, no coupling with diffusion measurements) of football players and
NCs along the left and right IFOF. Note that higher FFD variability between
the groups is indeed located at the occipital part of the tracts.
4
Summary and Conclusion
We presented a novel concept of FFDD descriptors that combine geometrical and
diffusivity properties of WM fiber bundles, for local quantification of pairwise and
group-wise differences. A sub-voxel alignment of tract profiles is accomplished
by considering local FFDD dissimilarities as an FMM inverse speed map. This
allows the construction of bundle-specific atlases for statistical analysis. Our
method is demonstrated on two datasets of contact-sports players, revealing local
WM tract anomalies. In a group-wise comparison between active football players
and normal (non-players) controls, our method revealed statistically significant
differences between the groups, at spatially-consistent areas within the IFOF and
FMT tracts. Furthermore, our method presented improved sensitivity to subtle
structural anomalies in football players compared to along-tract FA analysis.
The obtained results suggest the proposed method as a promising tool for mTBI
assessment and localization.
Acknowledgment This research is partially supported by the Israel Science Foundation (T.R.R. 1638/16) and the IDF Medical Corps (T.R.R.).
References
1. Andersson, Jenkinson, Smith, et al. Non-linear registration, aka spatial normalisation fmrib technical report tr07ja2. FMRIB Analysis Group of the University of
Oxford, 2, 2007.
2. P. Basser, J. Mattiello, and D. LeBihan. Estimation of the effective self-diffusion
tensor from the NMR spin echo. JMR, pages 247–254, 1994.
3. Y. Benjamini and Y. Hochberg. Controlling the false discovery rate: a practical
and powerful approach to multiple testing. Journal of the royal statistical society.
Series B (Methodological), pages 289–300, 1995.
4. N. Charon and A. Trouvé. The varifold representation of nonoriented shapes for
diffeomorphic registration. SIIMS, pages 2547–2580, 2013.
5. Chung, Adluru, Lee, et al. Cosine series representation of 3D curves and its application to white matter fiber bundles in diffusion tensor imaging. Statistics and its
interface, page 69, 2010.
6. I. Cohen, N. Ayache, and P. Sulger. Tracking points on deformable objects using
curvature information. In ECCV, pages 458–466, 1992.
7. Colby, Soderberg, Lebel, et al. Along-tract statistics allow for enhanced tractography analysis. Neuroimage, pages 3227–3242, 2012.
8. I. Corouge, S. Gouttard, and G. Gerig. Towards a shape model of white matter
fiber bundles using diffusion tensor MRI. In ISBI, pages 344–347, 2004.
9. Durrleman et al. Registration, atlas estimation and variability analysis of white
matter fiber bundles modeled as currents. NeuroImage, pages 1073–1090, 2011.
10. M. Frenkel and R. Basri. Curve matching using the fast marching method. In
EMMCVPR, pages 35–51, 2003.
11. Garyfallidis, Brett, Correia, et al. Quickbundles, a method for tractography simplification. Frontiers in neuroscience, page 175, 2012.
12. E. Garyfallidis, O. Ocegueda, D. Wassermann, and M. Descoteaux. Robust and
efficient linear registration of white-matter fascicles in the space of streamlines.
NeuroImage, pages 124–140, 2015.
13. L. Heimer. The human brain and spinal cord: functional neuroanatomy and dissection guide. Springer Science & Business Media, 2012.
14. Hulkower, Poliak, Rosenbaum, et al. A decade of DTI in traumatic brain injury:
10 years and 100 articles later. AJNR, pages 2064–2074, 2013.
15. Khatami, Schmidt-Wilcke, Sundgren, et al. Bundlemap: Anatomically localized
classification, regression, and hypothesis testing in diffusion mri. Pattern Recognition, pages 593–600, 2017.
16. Klein, Hermann, Konrad, et al. Automatic quantification of dti parameters along
fiber bundles. In Bildverarbeitung für die Medizin 2007, pages 272–276. 2007.
17. Mårtensson, Nilsson, Ståhlberg, et al. Spatial analysis of diffusion tensor tractography statistics along the inferior fronto-occipital fasciculus with application in
progressive supranuclear palsy. MAGMA, pages 527–537, 2013.
18. Mori, Oishi, Jiang, et al. Stereotaxic white matter atlas based on diffusion tensor
imaging in an ICBM template. Neuroimage, pages 570–582, 2008.
19. O’Donnell et al. Tract-based morphometry for white matter group analysis. Neuroimage, pages 832–844, 2009.
20. O’Donnell, Wells, Golby, et al. Unbiased groupwise registration of white matter
tractography. In MICCAI, pages 123–130, 2012.
21. L. O’Donnell and O. Pasternak. Does diffusion MRI tell us anything about the
white matter? an overview of methods and pitfalls. Schiz res, pages 133–141, 2015.
22. Sebastian, Klein, et al. On aligning curves. TPAMI, pages 116–125, 2003.
23. J. Sethian. A fast marching level set method for monotonically advancing fronts.
PNAS, pages 1591–1595, 1996.
24. Shenton, Hamoda, Schneiderman, et al. A review of magnetic resonance imaging
and diffusion tensor imaging findings in mild traumatic brain injury. Brain imaging
and behavior, pages 137–192, 2012.
25. Smith, Jenkinson, Johansen-Berg, et al. Tract-based spatial statistics: Voxelwise
analysis of multi-subject diffusion data. NeuroImage, 31:1487–1505, 2006.
26. Stamile et al. A sensitive and automatic white matter fiber tracts model for longitudinal analysis of diffusion tensor images in multiple sclerosis. PloS one, 2016.
27. Tagliasacchi, Zhang, and Cohen-Or. Curve skeleton extraction from incomplete
point cloud. In ACM Transactions on Graphics, page 71, 2009.
28. Yeatman, Dougherty, Myall, et al. Tract profiles of white matter properties: automating fiber-tract quantification. PloS one, 2012.
29. Yeh, Verstynen, Wang, et al. Deterministic diffusion fiber tracking improved by
quantitative anisotropy. PloS one, 2013.
30. L. Younes. Computable elastic distances between shapes. SIIMS, pages 565–586,
1998.
31. Yushkevich, Zhang, Simon, et al. Structure-specific statistical mapping of white
matter tracts. NeuroImage, pages 448–461, 2008.
| 1cs.CV
|
arXiv:1604.04405v1 [math.ST] 15 Apr 2016
Multiscale inference for a multivariate density with
applications to X-ray astronomy
Konstantin Eckle, Nicolai Bissantz, Holger Dette
Katharina Proksch
Ruhr-Universität Bochum
Institut für Mathematische Stochastik
Fakultät für Mathematik
Georg-August-Universität Göttingen
44780 Bochum, Germany
37077 Göttingen, Germany
Sabrina Einecke
Technische Universität Dortmund
Fakultät Physik
44221 Dortmund, Germany
April 18, 2016
Abstract
In this paper we propose methods for inference of the geometric features of a multivariate density. Our approach uses multiscale tests for the monotonicity of the density
at arbitrary points in arbitrary directions. In particular, a significance test for a mode at
a specific point is constructed. Moreover, we develop multiscale methods for identifying
regions of monotonicity and a general procedure for detecting the modes of a multivariate density. It is is shown that the latter method localizes the modes with an effectively
optimal rate. The theoretical results are illustrated by means of a simulation study and
a data example. The new method is applied to and motivated by the determination
and verification of the position of high-energy sources from X-ray observations by the
Swift satellite which is important for a multiwavelength analysis of objects such as Active
Galactic Nuclei.
Keywords and Phrases: multiple tests, modes, multivariate density, X-ray astronomy
AMS Subject Classification: 62G07, 62G10, 62G20
1
1
Introduction
This work is concerned with the development of a statistical toolbox which is useful for data
analysis in many problems of applied sciences. As a specific example we consider a problem
from X-ray astronomy, namely the determination of the positions of objects of BL Lacertae
type with statistical significance. Those objects form a specific subclass of blazars and hence
active Galactic Nuclei (AGN), where high-energy relativistic jets arise perpendicular to the
accretion disc and (in this case) point in the general direction of the Earth. They are among
the brightest extragalactic sources in X-rays and gamma-rays in the sky. Determination and
verification of the position of such objects from an observed distribution of origin positions of
X-ray photons from the object is of paramount importance for a multiwavelength analysis of
the object to understand and compare the appearance of photons of different energies in the
object.
From a statistical point-of-view, problems of this type are fundamental and refer to a proper
understanding of the shape of a density f based on a sample of multivariate observations.
Numerous authors have worked on the detection of qualitative features, such as modes and
regions of monotonicity of a density, in particular on tests for the existence and the localization
of modes. For example, it was pointed out by Chan and Tong (2004) that the presence of modes
can yield to a less precise forecasting. Similarly, a precise localization of modes can be used for
non-parametric clustering [see for example Pollard (1981) for an early reference and Li et al.
(2007), Chacón and Duong (2013) and Chaudhuri et al. (2014) for more recent work].
As pointed out by Romano (1988) and Grund and Hall (1995), estimation of modes of a density is a very complex problem, even more difficult than the estimation of the density itself.
In fact, the problem is closely related to the estimation of the first derivative of the density.
There exists a large amount of literature about statistical inference on modes of a density in
the univariate setting, which can be roughly divided into four different categories: tests on the
number of modes, the localization of modes, significance testing of candidate modes and tests
that allow for inference about monotonicity. Donoho and Liu (1991) provide the minimax rate
for estimating a single mode. The problem of estimating the number of modes is considered in
Silverman (1981) and Hall and York (2001). These authors investigate a test that uses bootstrap methods based on the so-called critical bandwidth of a kernel density estimator [see also
Mammen et al. (1991) and Chan and Tong (2004) for an asymptotic analysis and an extension
to the dependent case]. In Hartigan and Hartigan (1985), the distance of the empirical distribution function to the best-fitting unimodal density is used as test statistic. Hartigan (1987) and
Müller and Sawitzki (1991) propose the excess mass approach for statistical inference of (multi-)
modality, which is also used by Polonik (1995); Minnotte (1997) and Fisher and Marron (2001)
to construct nonparametric tests for the existence of modes. Chaudhuri and Marron (1999)
introduce the SiZer-map as a graphical tool for the analysis of the local monotonicity properties of a density. In this paper, the derivative of a kernel density estimator is tested locally
for a significant deviation from zero. A particular characteristic of the SiZer map is that these
2
tests are performed simultaneously over a fixed range of bandwidths. A multicsale test for the
monotonicity of a univariate density, which allows simultaneous confidence statements about
regions of increase and decrease, can be found in Dümbgen and Walther (2008). In the univariate deconvolution model, Schmidt-Hieber et al. (2013) propose a multiscale test for qualitative
features of a density such as regions of monotonicity.
On the other hand, for multivariate densities there are just a few results on modality and even
less on monotonicity. Tsybakov (1990) proves that the optimal minimax rate for mode detection over a β-Hölder class is n−(β−1)/(2β+d) (β > 0). The excess mass approach can also be used
in the multivariate case, but most authors concentrate on one dimension because - as pointed
out by Burman and Polonik (2009) - there is usually a trade-off between practical feasibility
and theoretical justification. Abraham et al. (2004) use kernel smoothing to construct consistent estimators of the single mode of a multivariate density, while Klemelä (2005) suggests an
adaptive estimate which achieves the optimal rate. Burman and Polonik (2009) do not prespecify the total number of modes and propose a method for locating modal regions by means
of formal testing for the presence of anti-modes. A rate-optimal algorithm for the localization
of the modes of a multivariate density based on a k-nearest neighbour estimator of the density
can be found in a recent paper of Dasgupta and Kpotufe (2014).
A test about local monotonicity properties of a bivariate density can be found in Godtliebsen et al.
(2002) generalizing the SiZer-map. In a multivariate setting Duong et al. (2008) test locally
whether the norm of the gradient of the density vanishes using a kernel density estimate with
a fixed bandwidth. In a recent paper Genovese et al. (2015) suggest an algorithm for mode
estimation of a d-dimensional density. These authors construct non-parametric confidence intervals for the eigenvalues of the Hessian at modes of a density estimate, which can be used for
the construction of a significance test. The method is based on a sample splitting, where the
first half of the data is used to localize the modes by means of the mean-shift algorithm and
the second half of the data is used for the significance test. Genovese et al. (2015) also point
out that the multiscale approach of Dümbgen and Walther (2008) for constructing confidence
intervals for modes is only applicable to one-dimensional densities.
The goal of the present paper is to fill this gap by providing a multiscale method to identify
regions of monotonicity of a multivariate density. In Section 2 we briefly review the approach
of Dümbgen and Walther (2008). We also define a concept of monotonicity in the multivariate
case and introduce a multiscale test for this property at a pre-specified point x0 ∈ Rd . The main
idea is to investigate monotonicity properties of the density in “various” directions e ∈ Rd by
projecting observations from a wedge centered at x0 onto the line {x0 +te | t ≥ 0}. A multiscale
test is provided that allows for a simultaneous inference of the monotonicity properties at a given
confidence level α. Section 3 extends the approach to the situation where no prior information
regarding the location and the number of the modes is available. The theoretical results of this
paper establish the consistency of this approach and show that modes can be detected with the
optimal rate (up to a logarithmic factor). The finite sample properties of the multiscale test are
3
investigated in Section 4 and in Section 5 we apply our proposed method to the determination
and verification of the positions of the blazars Markarian 501 and S3 0218+35. Finally, all
technical details and proofs are deferred to Section 6.
2
Local testing for a mode
In this Section, we present a test for the presence of a mode of the density f at a pre-specified
candidate point x0 ∈ Rd based on a sample of independent random variables X1 , . . . , Xn with
density f . We begin with a brief review of the work of Dümbgen and Walther (2008), who
investigate regions of monotonicity of a univariate density (that is d = 1).
2.1
Multiscale inference about a univariate density revisited
For one-dimensional independent identically distributed random variables X1 , . . . , Xn with density f let X(1) ≤ . . . ≤ X(n) denote the corresponding order statistics and consider the associated
local spacings
X(i) − X(j)
X(i;j,k) =
, j ≤ i ≤ k.
X(k) − X(j)
Dümbgen and Walther (2008) propose to use the local spacings
Tjk (X) =
k−1
X
β(X(i;j,k))
i=j+1
to construct a test statistic for (local) monotonicity of the density f on the interval (X(j) , X(k) ),
where the function β is defined by β(x) := (2x − 1)1(0,1) (x). Note that Tjk (X) has mean zero
if f is constant on (X(j) , X(k) ). Let X denote a random variable with density f independent of
X1 , . . . , Xn , and denote by
F̃ (x) = P X ≤ x | X ∈ [X(1) , X(n) ]
the conditional distribution function of X given X ∈ [X(1) , X(n) ]. Define U(i) = F̃ (X(i) ),
then U(2) , . . . , U(n−1) correspond in distribution to the order statistics of a sample of (n − 2)
independent uniformly distributed random variables on the interval [0, 1] (note that U(1) = 0
and U(n) = 1). It can be shown that the statistic
Tjk (U ) =
k−1
X
i=j+1
satisfies
Tjk (X)
β(U(i;j,k)) for 1 ≤ j < k ≤ n, k − j > 1,
(
≥ Tjk (U ), if f is increasing on (X(j) , X(k) ),
≤ Tjk (U ), if f is decreasing on (X(j) , X(k) ).
4
(2.1)
q
),
Define Γ(δ) := 2 log( exp(1)
δ
Tn (U ) =
max
1≤j<k≤n,k−j>1
r
3
k−j
,
|Tjk (U )| − Γ
k−j−1
n−1
(2.2)
and denote by κn (α) the (1 − α)-quantile of the statistic Tn (U ). The multiscale test for monotonicity proposed by Dümbgen and Walther (2008) now concludes that the density f is not
increasing on every interval (X(j) , X(k) ) with
r
k − j − 1 k − j
Tjk (X) < −cjk (α) :=
Γ
+ κn (α) , 1 ≤ j < k ≤ n, k − j > 1
3
n−1
and that f is not decreasing on every interval (X(j) , X(k) ) with Tjk (X) > cjk (α). The overall
risk of at least one false-positive decision within the simultaneous tests on all scales (i.e. for
1 ≤ j < k ≤ n, k − j > 1) is at most α.
2.2
Assumptions and geometrical preparations
Throughout this paper kxk denotes the Euclidean norm of a vector x ∈ Rd . The function
f : Rd → R has a mode at the point x0 , if for every vector e ∈ Rd with kek = 1 the function
fe : t 7→ f (x0 + te), t ≥ 0, is strictly decreasing in a neighbourhood of t = 0. The aim of the
test for the presence of a mode defined below is to investigate the monotonicity of functions of
this type in different directions e. The number of directions is determined by the sample size n.
As the set {x0 + te | t ≥ 0} has Lebesgue measure 0, we also have to consider observations in a
neighbourhood of this line for inference about monotonicity of the function fe . For this purpose,
we introduce a signed distance of the projection of a point x ∈ Rd onto the line {x0 + te | t ∈ R}
and introduce so-called wedges. For the following discussion we denote by {e1 , . . . , ed−1 } an
arbitrary but fixed orthonormal basis of (span{e})⊥ , hx, yi is the standard inner product of the
d
vectors x, y ∈ Rd and “=” denotes equality in distribution.
Definition 2.1. Let x0 ∈ Rd and e ∈ Rd with kek = 1.
(1) The projected signed distance of a point x ∈ Rd from x0 in direction e on {x0 + te | t ∈ R}
is defined as
Pe x := hx − x0 , ei.
(2) The wedge with vertex x0 , direction e, length l > 0 and angle ϕ ∈ (0, π2 ) is defined as
n
K ≡ K(x0 , e, ϕ) := x ∈ Rd 0 < Pe x ≤ l and hx − x0 , ei i ∈ [− tan(ϕ)Pe x, tan(ϕ)Pe x]
o
for i = 1, . . . , d − 1 .
5
(3) For a wedge K ⊂ Rd let X(1) , . . . , X(N ) be those random variables among X1 , . . . , Xn
which are located in K, arranged in ascending order with respect to their signed projected
distances from x0 , i.e. X(j) ∈ K for j = 1, . . . , N and Pe X(1) ≤ . . . ≤ Pe X(N ) . The wedge
KN is defined as KN := x ∈ K : 0 < Pe x ≤ Pe X(N ) .
x0
×X(N )
ϕ
KN
l
×
Pe X(N )
e
K
Figure 1: The wedges K and KN for d = 2.
A typical wedge is displayed in Figure 1 in the case d = 2. We are now able to define monotonicity properties of the function f on the wedge K that will play a crucial role in following
discussion.
Definition 2.2.
(i ) The function f is called increasing on the wedge K, if f (x0 + t̃2 e0 ) ≥ f (x0 + t̃1 e0 ) for all
e0 ∈ Rd with ke0 k = 1 and t̃2 > t̃1 ≥ 0 such that x0 + t̃2 e0 ∈ K.
(ii ) The function f is called decreasing on the wedge K, if f (x0 + t̃2 e0 ) ≤ f (x0 + t̃1 e0 ) for all
e0 ∈ Rd with ke0 k = 1 and t̃2 > t̃1 ≥ 0 such that x0 + t̃2 e0 ∈ K.
2.3
A local test for modality
Let x0 ∈ Rd denote a candidate position for a mode. The construction of a local test for the
hypothesis that f has a mode at x0 is based on an investigation of the monotonicity properties
of f on pairwise disjoint wedges K i (i = 1, . . . , Mn ) with common vertex x0 . We begin with the
case Mn = 1 and use the notation K := K 1 for the sake of simplicity. Throughout this paper
1A denotes the indicator function of a set A.
Theorem 2.3. Let X be a d-dimensional random variable with density f independent of
X1 , . . . , Xn and denote by
F̃ (z) = P(Pe X ≤ z|N, X ∈ KN , X(N ) ) (0 < z ≤ Pe X(N ) )
6
P
the distribution function of Pe X conditional on N = ni=1 1K (Xi ), {X ∈ KN } and X(N ) . Then,
conditionally on N, F̃ (Pe X(1) ), . . . , F̃ (Pe X(N −1) ) are distributed as the order statistics of N − 1
independent uniformly distributed random variables on the interval [0, 1].
The first step in the construction of a test for a mode at the point x0 is to investigate monotonicity in the sense of Definition 2.2. For this purpose, we use a comparison of the projected
distances Pe X(1) , . . . , Pe X(N −1) with the distribution of projected distances of random variables
U1 , . . . , UN −1 which are uniformly distributed on the wedge KN . For a random variable U which
is uniformly distributed on K and independent of X1 , . . . , Xn , U1 , . . . , UN −1 , we have
F̃U (z) = P(Pe U ≤ z|N, U ∈ KN , X(N ) ) =
zd
(0 < z ≤ Pe X(N ) ),
(Pe X(N ) )d
Pe Uj d
, j = 1, . . . , N −1, has a uniform
and by Theorem 2.3, the random variable F̃U (Pe Uj ) = Pe X
(N)
distribution on the interval [0, 1], conditionally on N and the event {Uj ∈ KN }. Consequently,
we propose the test statistic
N
−1
X
(Pe X(j) )d
(2.3)
TK =
β
(Pe X(N ) )d
j=1
for testing monotonicity properties of the density f on the wedge K, where β(z) = (2z −
(Pe X(N−1) )d
(P X(1) )d
,
.
.
.
,
have the same
1)1(0,1) (z). If f is constant on K, we have E[TK ] = 0 as (PeeX(N)
d
)
(Pe X(N) )d
distribution as an order statistic of uniformly distributed random variables on the interval [0, 1].
On the other hand, if f is increasing on the wedge K, the observations in KN tend to have large
projected distances from x0 , which results in positive values of the test statistic TK . Similarly,
if f is decreasing on K, it is more likely that the test statistic is negative.
Theorem 2.4. Let F̃ denote the conditional distribution function defined in Theorem 2.3, TK
PN −1
β(F̃ (Pe X(j) )).
be defined in (2.3) and TKU := j=1
(i) If f is increasing on K, then TKU ≤ TK (a.s.) conditionally on N.
(ii) If f is decreasing on K, then TKU ≥ TK (a.s.) conditionally on N.
By Theorem 2.3, conditionally on N, the statistic TKU has the same distribution as the random
PN −1
variable j=1
β(U(j) ), where U1 , . . . , UN −1 are independent uniformly distributed random variables on the interval [0, 1]. Therefore, Theorem 2.4 is the key result to obtain critical values for
a multiscale test.
In the second step, we combine test statistics of the form TK for different wedges to construct
a test for a mode at the point x0 . For this purpose, define
ln :=
1
log(n) d+4
n
7
(2.4)
and construct a family Kn of Mn pairwise disjoint wedges K 1 , . . . , K Mn with common vertex
d−1
x0 , length C1 log(n) d+4 ln and angle ϕn := C22 log(n)−1 (for some constants C1 , C2 > 0) and by
n
specifying the central directions {e1n , . . . , eM
n } as follows
(1) Choose a direction e1n with ke1n k = 1
(2) If e1n , . . . , ei−1
have been specified, then - whenever possible - choose a vector ein with
n
kein k = 1 such that for some ε > 0
√
|angle(ein , ejn )| ≥ (2 + ε) arctan d − 1 tan(ϕn ) for all j < i.
Note that this procedure does not define Kn in a unique way. However, if Kn has been fixed
n
for the central directions e1n , . . . , eM
n , then the following property holds. For any normalized
vector e ∈ Rd there exists a direction ein such that
√
|angle(e, ein )| < (2 + ε) arctan d − 1 tan(ϕn ) = O log(n)−1 .
This can be seen easily by deriving a contradiction from the opposite assertion using the expansion tan(z) = z(1 + o(1)) = arctan(z) for z → 0.
Now define for each K i ∈ Kn
n
X
i
N :=
1K i (Xj )
j=1
as the number of observations in the wedge K i and consider the corresponding statistics TK i
and TKU i defined in (2.3) and Theorem 2.4, respectively. An application of Theorem 2.3 on each
wedge K i shows that, conditionally on N 1 , . . . , N Mn ,
d
TKU i =
i −1
N
X
i
β(U(j)
) (i = 1, . . . , Mn ),
(2.5)
j=1
where {Uji | j = 1, . . . , N i − 1, i = 1, . . . , Mn } are independent uniformly distributed random
variables on the interval [0, 1]. In particular, the statistics TKU 1 , . . . , TKU Mn are conditionally
independent.
By means of the representation (2.5), the quantile κ̃n (α) defined by the condition
N i
r 3
1
Mn
U
=1−α
(2.6)
≤
κ̃
(α)
N
,
.
.
.
,
N
|T
|
−
Γ
P max
i
n
i=1,...,Mn
Ni − 1 K
n−1
can be obtained by numerical simulation, as soon as the numbers of observations N 1 , . . . , N Mn
Ni
)
in the wedges K 1 , . . . , K Mn have been specified. We note that a calibration by the term Γ( n−1
i
for various scales (i.e. different values of N ) is necessary to show that the quantile κ̃n (α) is
asymptotically bounded [see Section 6 for details].
8
In a third step, we consider on each of the wedges K 1 , . . . , K Mn two hypotheses, that is
incr
H0,i
: f is increasing on K i
versus
incr
H1,i
: f is not increasing on K i
(2.7)
versus
decr
H1,i
: f is not decreasing on K i
(2.8)
(i = 1, . . . , Mn ), and
decr
H0,i
: f is decreasing on K i
(i = 1, . . . , Mn ), where the the notation of an increasing (decreasing) function on the wedge K i
is introduced in Definition 2.2. The ith hypothesis in (2.7) is rejected, whenever
TK i < −c̃K i (α)
q
i
N −1
(i = 1, . . . , Mn ), where the quantile c̃K i (α) is defined by c̃K i (α) =
3
(i = 1, . . . , Mn ). Similarly, the ith hypothesis in (2.8) is rejected, whenever
TK i > c̃K i (α)
(2.9)
Ni
κ̃n (α) + Γ( n−1
)
(2.10)
(i = 1, . . . , Mn ). The final result of this section specifies the error of at least one false decision
among these 2Mn local level α-tests on monotonicity.
Theorem 2.5. Assume that all tests (2.9) and (2.10) for the hypotheses (2.7) and (2.8) are
performed (i = 1, . . . , Mn ). The probability of at least one false rejection of any of the tests is
at most α.
We conclude this section by showing that the results presented so far can be used to obtain
a consistent multiscale test for the hypothesis that the density f has a mode at a given point
x0 ∈ Rd . The test decides for the presence of a mode at x0 if every test (2.9) for the null
hypothesis that f is increasing on the wedge K i ,q
i = 1, . . . , Mn , rejects the null. Note that in
i
Ni
) in (2.9), where κ̃′n (α)
this case we use the one-sided quantiles c̃′K i (α) = N 3−1 κ̃′n (α) + Γ( n−1
is defined by the condition
N i
r 3
1
Mn
′
U
= 1 − α.
(2.11)
≤ κ̃n (α) N , . . . , N
T i −Γ
P max −
i=1,...,Mn
Ni − 1 K
n−1
Theorem 2.6. Assume that the density f is twice continuously differentiable in a neighbourhood
of x0 with f (x0 ) 6= 0, gradient ∇f (x0 ) = 0 and a Hessian Hf (x0 ) satisfying e⊤
0 Hf (x0 )e0 ≤ −c <
d
0 for all e0 ∈ R with ke0 k = 1. Consider the family of wedges Kn defined in Section 2.3 with
constants C1 , C2 satisfying
4D 2 f (x0 )
C1d+4 C2d−1 > 2
,
(2.12)
c d+4
where
√
2(2d + 2)(d + 2)
D=
(2.13)
h
(d+2)/d
i1/2 .
d2
d
2d+2 2 1/2
1 − 2(2d+2)2 − 1 + 1 + 4( d )
1 − d+2
Then, all Mn tests defined in (2.9) (using the quantiles c̃′K i (α) instead of c̃K i (α) (i = 1, . . . , Mn ))
reject the null hypothesis with asymptotic probability one as n → ∞.
9
Note that the constant D in (2.13) depends only on the dimension d. Hence, the lower bound
on the constants C1 and C2 is determined by the shape of the modal region (more precisely
the largest eigenvalue of the Hessian Hf (x0 ) at x0 ) as well as by the value of the density at the
point x0 .
3
Global inference on monotonicity
In this section we extend the local inference on modality at a fixed point to the situation where
no specific candidate position for the mode can be defined in advance. This is particularly
important since there exist several applications where at most approximate information about
the position of the modes is available. As in the previous section, let X1 , . . . , Xn denote independent d-dimensional random variables with density f . The proposed test for the detection
of modes proceeds in several steps.
The first step consists in a selection of the candidate modes. Here, we choose these as the
vertices of an equidistant grid in Rd . Secondly, we introduce a generalization of the multiscale
test on monotonicity presented in Section 2, where we divide the wedges in subsections that
are determined by the data. The latter approach can be very useful in settings without a priori
knowledge about the modes, as a true mode obviously has not to be located at the vertex
of a wedge. Figure 2 provides a graphical representation of the results of the global test on
modality in the bivariate case where the multiscale generalization has been omitted. Here,
on every dotted wedge K, the test has rejected that f is decreasing on K. Accordingly, the
cross-hatches refer to a rejection that f is increasing on K. Non-marked wedges indicate that
no significant result has been found. For a detailed description of the settings used to provide
Figure 2 and an analysis of the results, we refer to the end of this section.
3.1
Geometrical preparations
Throughout this section let bj denote the jth unit vector in Rd (j = 1, . . . , d) and define
⌈x⌉ := inf{z ∈ Z| z ≥ x}. Recall the definition of ln in (2.4) and denote (for given constants
C1 , ε > 0) by Gn the grid consisting of the vertices
d
X
j=1
(ij (2 + ε)C1 log(n)ln − n)bj
i1 . . . id ∈ 0, . . . , ⌈ (2+ε)C2n
⌉
. Note that the grid Gn covers the cube [−n, n]d and that
1 log(n)ln
the sequence log(n)ln (which determines the order of the mesh size) is chosen such that the test
of modality defined below is consistent.
We now define by Kn a family of wedges (cf. Definition 2.1) with length l = C1 log(n)ln , an angle
ϕ = ϕn = C22 log(n)−1 for a given constant C2 > 0, vertex in Gn , and a direction contained in
the set of given directions {e1n , . . . , enMn } (cf. Section 2.3). For an arbitrary but fixed element K
10
4
3
2
1
0
-1
-2
-4
-3
-2
-1
0
1
2
3
4
Figure 2: Example of a global map for monotonicity of a density.
of Kn let X(1) , . . . , X(N ) denote those random variables among {X1 , . . . , Xn } which are located
in K and ordered with respect to their signed projected distances from the vertex x0 of K. For
0 ≤ j < k ≤ N with k − j > 1 we define
K(j, k) := {x ∈ Rd x ∈ K and Pe X(j) < Pe x ≤ Pe X(k) }
as a subsection of the wedge K, where X0 := x0 and x0 denotes the vertex of K. A typical set
is depicted in Figure 3. We conclude this section with a definition of a concept of monotonicity
on subsections of a wedge.
Definition 3.1. Let K be a wedge with vertex x0 and K(j, k) ⊆ K be a subsection. The
function f : Rd → R is
(i) increasing on K(j, k), if f (x0 + t̃2 e0 ) ≥ f (x0 + t̃1 e0 ) for all e0 ∈ Rd with ke0 k = 1 and all
t̃2 > t̃1 ≥ 0 such that x0 + t̃ℓ e0 ∈ K(j, k) (ℓ = 1, 2).
(i) decreasing on K(j, k), if f (x0 + t̃2 e0 ) ≤ f (x0 + t̃1 e0 ) for all e0 ∈ Rd with ke0 k = 1 and all
t̃2 > t̃1 ≥ 0 such that x0 + t̃ℓ e0 ∈ K(j, k) (ℓ = 1, 2).
3.2
Regions of monotonicity and mode detection
The approach proposed here consists of simultaneous tests for monotonicity of the density f on
every subsection of every wedge in Kn . For the definition of these tests we will proceed similarly
11
×X(j)
×
P X(j)
x0
l
×
P X(k)
e
×X(k)
K(j, k)
K
Figure 3: The subsection K(j, k) for d = 2.
as in Section 2.3. We begin by introducing a multiscale test statistic on the subsection K(j, k)
of a wedge K ∈ Kn which is defined by
TK(j,k) :=
k−1
X
β
l=j+1
(Pe X(l) )d − (Pe X(j) )d
(Pe X(k) )d − (Pe X(j) )d
,
(3.1)
where 0 ≤ j < k ≤ N, k − j > 1. Note that TK(0,N ) = TK , where TK is the test statistic defined
in (2.3).
Now, let Kn = {K i | i = 1, . . . , Ln } denote the family of wedges defined in Section 3.1. For
the multiscale approach, we use for each subsection K i (ji , ki ) of the wedge K i the test statistic
TK i (ji ,ki ) defined by (3.1) (0 ≤ ji < ki ≤ N i , ki − ji > 1, i = 1, . . . , Ln ) and consider
kX
i −1
TKU i (ji ,ki) =
l=ji +1
β
F̃ i (P i X ) − F̃ i (P i X )
e (l)
e (ji )
F̃ i (Pei X(ki ) ) − F̃ i (Pei X(ji ) )
,
where Pei denotes the signed projected distance and F̃ i denotes the conditional distribution
function with respect to K i (0 ≤ ji < ki ≤ N i , ki − ji > 1, i = 1, . . . , Ln ). Using similar
arguments as in Section 2.3, it follows that (conditionally on N i )
TKU i (ji ,ki ) ≤ TK i (ji ,ki ) (a.s.) if f is increasing on K(ji , ki ),
TKU i (ji ,ki ) ≥ TK i (ji ,ki ) (a.s.) if f is decreasing on K(ji , ki )
(0 ≤ ji < ki ≤ N i , ki − ji > 1, i = 1, . . . , Ln ). Moreover,
TKU i (ji ,ki)
d
=
kX
i −1
l=ji +1
β
Ui − Ui
(l)
(ji )
i
U(k
− U(ji i )
i)
conditional on N 1 , . . . , N Ln ,
where {Ujii | ji = 1, . . . , N i − 1, i = 1, . . . , Ln } are independent random variables which are
i
i
uniformly distributed on the interval [0, 1], and U(1)
≤ . . . ≤ U(N
i −1) is the order statistics of
12
U1i , . . . , UNi i −1 (i = 1, . . . , Ln ). Finally, let κn (α) denote the (1 − α)-quantile of the conditional
distribution of the random variable
r
k − j
3
i
i
U
max
max
(3.2)
|TK i (ji ,ki ) | − Γ
i
i=1,...,Ln 0≤ji <ki ≤N , ki −ji >1
ki − ji − 1
n−1
given N 1 , . . . , N Ln . We consider on each subsection K i (ji , ki ) of the wedge K i the hypotheses
incr
incr
H0,i,j
: f is increasing on K i (ji , ki ) versus H1,i,j
: f is not increasing on K i (ji , ki ), (3.3)
i ,ki
i ,ki
decr
decr
H0,i,j
: f is decreasing on K i (ji , ki ) versus H1,i,j
: f is not decreasing on K i (ji , ki )
i ,ki
i ,ki
incr
(0 ≤ ji < ki ≤ N i , ki − ji > 1, i = 1, . . . , Ln ). The hypothesis H0,i,j
is rejected if
i ,ki
TK i (ji ,ki ) < −cK i(ji ,ki) (α)
(0 ≤ ji < ki ≤ N i , ki − ji > 1, i = 1, . . . , Ln ), where cK i(ji ,ki) (α) :=
decr
Similarly, H0,i,j
is rejected if
i ,ki
(3.4)
q
ki −ji −1
3
i −ji
κn (α) + Γ( kn−1
) .
TK i (ji ,ki ) > cK i(ji ,ki) (α)
(3.5)
(0 ≤ ji < ki ≤ N i , ki − ji > 1 i = 1, . . . , Ln ). Following the line of arguments in the proof of
Theorem 2.5, we obtain the following result.
Theorem 3.2. If all tests (3.4) and (3.5) are performed simultaneously (0 ≤ ji < ki ≤ N i , ki −
ji > 1, i = 1, . . . , Ln ), then the probability of at least one false rejection is at most α.
3.3
Mode detection
We consider the following asymptotic regime. For n ∈ N let Kn denote the family of wedges
defined in Section 3.1 and define In as the set of indices i corresponding to the wedges Kni ∈ Kn
whose vertices x0,n fulfill C1 log(n)ln ≤ kxn0 − x0 k ≤ mn C1 log(n)ln for a mode x0 of f and
5
mn = (log(n)) 2 and whose direction en fulfills angle(xn0 − x0 , en0 ) = O(log(n)−1 ). Then, every
i
incr
test (3.4) for the hypothesis H0,i,0,N
i defined (3.3) (i.e. f is increasing on Kn ) with i ∈ In ⊆
{1, . . . , Ln }, rejects the null with asymptotic probability one.
Theorem 3.3. Let
√
2 2(2d + 1)(d + 1)
D=
(d+1)/d h
d
d2
1 − d+1
1 − 2(2d+1)
)2
− 1 + 1 + 4( 2d+1
2
d
1
2
i 12
.
Assume that for any mode x0 ∈ Rd the density f satisfies c1 ≥ f (x0 ) > 0 and that there exist
functions gx0 : Rd → R, f˜x0 : R → R such that the density f has a representation of the form
f (x) ≡ (1 + gx0 (x))f˜x0 (kx − x0 k)
13
(3.6)
(in a neighbourhood of x0 ). Furthermore, let gx0 be differentiable in a neighbourhood of x0
with gx0 (x) = o(1) and h∇gx0 (x), e0 i = o(kx − x0 k1+γ ) (for some γ > 0) if x → x0 and all
e0 ∈ Rd with ke0 k = 1. In addition, let f˜x0 be differentiable in a neighbourhood of 0 with
f˜x′ 0 (h) ≤ −ch(1 + o(1)) for h → 0. If Kn is the family of wedges defined in Section 3.1 with
C1d+4 C2d−1 >
D 2 c1
,
c2 d + 4
(3.7)
then every mode x0 of f will be detected with asymptotic probability one as n → ∞.
Theorem 3.3 shows that the proposed procedure can find all modes with (up to the logarithmic
factor) optimal rate. Note that we proceed in two steps: the verification of the presence of a
mode and its localization. With probability one the presence of every mode will be detected
(by means of the asymptotic regime introduced at the beginning of this section). The rate for
the localization of a mode is given by the mesh size of the grid Gn , which is determined by the
length of the wedges.
4
Finite sample properties
In this section we illustrate the finite sample properties of the proposed multiscale inference. In
particular, we study the power of the local test for a mode at a given point x0 ∈ Rd . We also
present an example illustrating how the results of Section 3 can be used to obtain a graphical
representation of the local monotonicity properties of the density.
4.1
Local test for modality
Here, we investigate the finite sample properties of the local test for a two-dimensional density,
where the level is given by α = 0.05. The corresponding quantiles κ̃′n (0.05) defined in (2.11) are
determined from 1000 simulation runs based on independent and uniformly distributed random
variables on the interval [0, 1] and are listed in Table 1 for the sample sizes n = 100, 500, 5000 in
the situation considered in Table 2 (note that κ̃′n (0.05) depends on the number of observations in
every wedge and hence both on the number and on the size of the wedges). By its construction,
observations
100
500
5000
κ̃′n (0.05)
0.126
-0.319
-0.854
Table 1: Simulated quantiles κ̃′n (0.05) in the situation considered in Table 2.
the local test is conservative, and therefore we also investigate a calibrated version of the new
14
test. The quantiles of the calibrated test are chosen such that the level of the test coincides
with α = 0.05 for the data obtained from a uniform distribution on the set [−2.5, 2.5]2 . Note
that this calibration does not require any knowledge about the unknown density f . However,
the procedure requires the choice of the length and the angle of the wedges and according to
Theorem 2.6 we used
1
log(n) d+4
d−1
C2
ln = C 1
log(n) d+4 , ϕn =
log(n)−1 ,
n
2
where C1 , C2 > 0 are constants. In the following, the power and level of the test with respect
to different choices of C1 and C2 is investigated. We also consider different numbers Mn of
wedges in our study. Recall from the discussion in Section 2.3 that the constants C1 and C2
have to satisfy (2.12) in order to guarantee consistency of the test. All results presented below
are based on 1000 simulation runs.
We begin with a comparison of the test introduced in Section 2 (based on the critical values
κ̃′n (0.05)) and a calibrated version of this test. In Table 2 we present the simulated level and
power of the local test for a mode at the point x0 = (0, 0)⊤ for different sample sizes. The
constants in the definition of the length and the angle are chosen as C1 = 2 and C2 = 9.65.
For the investigation of the level we consider a uniform distribution on the square [−2.5, 2.5]2 ,
since it represents a “worst” case scenario. For the calculation of the power, we sample from the
standard normal distribution. We observe that the test proposed in Section 2 is conservative
but it has reasonable power with increasing sample size. On the other hand, the calibrated
version of the multiscale test keeps its nominal level and rejects the null hypothesis of no mode
at x0 in nearly all cases.
observations
100
500
5000
ln
1.54
1.31
0.99
Mn
3
4
5
level
0.0
0.0
0.0
power
36.8
50.0
72.7
level (cal.)
4.8
4.5
5.0
power (cal.)
97.6
98.4
100
Table 2: Simulated level and power of the local test for a mode at x0 = (0, 0)⊤ of a 2-dimensional
density.
Next we investigate the influence of the shape of the modal region on the power of the local
test. To this end, we sample from normal distributions with expectation (0, 0)⊤ and covariance
matrix Σ 6= I2 . The results for
0 0.5
−0.5 1
Σ1 = −1 1.5 and Σ2 = −2 2.5
(4.1)
are presented in Table 3. We conclude that the shape of the modal region (determined by the
absolute values of the eigenvalues of the covariance matrix) has a strong influence on the power
of the test. In the case Σ1 (eigenvalues 0.5 and 1), the absolute values of both eigenvalues
15
observations
100
500
5000
ln
1.54
1.31
0.99
Mn
3
4
5
power
65.4
95.6
97.8
Σ1
power (cal.)
98.7
100
100
power
38.8
80.1
92.1
Σ2
power (cal.)
94.3
99.6
99.7
Table 3: Influence of the shape of the modal region on the power of the local test. The matrices
Σ1 and Σ2 are given in (4.1).
are smaller than one. For Σ2 the eigenvalues are given by 0.5 and 1.5. Hence, we observe a
(slight) decrease in power in comparison to the first case. However, due to the existence of an
eigenvalue with absolute value smaller than one, the test still performs better as in the case of
a standard normal distribution.
As the local test requires the specification of the point x0 , we next investigate the influence of
its incorrect specification on the power of the test. For this purpose, we consider the same data
(two-dimensional standard normal distribution) and perform the tests under the assumption
that the modes are given by x0 = (0.2, 0.2)⊤ and x0 = (0.7, 0.7)⊤ , respectively (which has to be
compared to the true position of the mode at (0, 0)⊤ ). The corresponding results are shown in
Table 4, and we conclude that a “small” deviation of the candidate mode from the true mode has
a very small effect on the power of the tests. In the case x0 = (0.7, 0.7)⊤ , the distance between
the candidate and the true mode is very large in comparison to the length of the wedges. For
n = 100 observations the length of the wedges is still substantially larger than the distance
between the candidate and the true mode. Hence, the test detects the presence of a mode, but
we observe a decrease in its power. However, for n = 5000 observations the distance between
the candidate position and the true mode is approximately equal to the length of the wedges.
As a consequence, the multiscale test is performed with a finer triangulation and (correctly)
does not indicate the existence of a mode at the point x0 = (0.7, 0.7)⊤ .
observations
100
500
5000
ln
1.54
1.31
0.99
Mn
3
4
5
x0 = (0.2, 0.2)⊤
power power (cal.)
32.4
96.6
43.1
97.8
47.2
98.3
x0 = (0.7, 0.7)⊤
power power (cal.)
2.8
75.6
1.2
57.1
0.1
10.8
Table 4: Influence of a misspecification of the mode on the power of the local test.
In the remaining part of this section we investigate the influence of the choice of the parameters
C1 and C2 on the power of the test. Note that the volume of every wedge is proportional to
lnd , where ln is the length of the wedge. This means that dividing the length in half yields a
wedge with a volume which is 2−d times smaller than the volume of the original wedge. Thus,
the number of observations in the smaller wedge is substantially smaller than the number of
observations in the larger wedge. Therefore, we expect that the constant C1 has an impact on
the power of the test. These theoretical considerations are reflected by the numerical results
16
in Table 5, which show the power for a fixed sample size n = 500, different choices of C1
(represented by the different lengths) and fixed parameter C2 . We observe a loss of power of
both tests with decreasing length of the wedge.
On the other hand, decreasing the constant C2 such that the number of wedges doubles has
the effect that the number of observations in every wedge decreases approximately by 50%.
In Table 6 we show the power for a fixed sample size n = 500, a fixed constant C1 = 2 and
varying values of C2 (represented by the different number of wedges). Here, the picture is not
so clear. While we observe a loss in power of the non-calibrated tests with an increasing number
of wedges, the power of the calibrated test changes only slightly. In both cases, the calibrated
version still performs rather well opposite to its uncalibrated version.
observations
500
500
500
ln
1.31
0.98
0.65
Mn
4
4
4
level
0.0
0.0
0.0
power
50.0
1.5
0.0
level (cal.)
4.5
4.9
5.4
power (cal.)
98.4
74.4
37.2
Table 5: Influence of the length of the wedges on the power.
observations
500
500
500
ln
1.31
1.31
1.31
Mn
4
6
8
level
0.0
0.0
0.0
power
50.0
0.7
0.0
level (cal.)
4.5
5.3
4.9
power (cal.)
98.4
92.7
89.8
Table 6: Influence of the number of directions tested on the power.
4.2
Identifying local monotonicity of a multivariate density
In this section we demonstrate how the results of Section 3 can be used to obtain a graphical
representation of the local monotonicity behaviour of the density (in the case d = 2). We
conduct the procedure to detect regions of monotonicity as proposed in Section 3.2. For the
sake of convenience, we use only the largest scales in the test statistic (3.1) (i.e. we test on
the entire wedges and not on the subsections introduced in Section 3.1). The significance level
is α = 0.05. We chose an equidistant grid covering [−3, 3] × [−1, 3] with points (i, j)⊤ , i =
−3, . . . , 3, j = −1, . . . , 3, the length of any wedge is l = 21 and all angles are given by ϕ = π4 .
Figure 2 presents the map of the local monotonicity properties on the basis of n = 100000
observations from a mixture of three normal distributions (i.e. f has three modes of different
shape) [see Figure 4]. Here, a cross-hatched wedge indicates that the local test rejected the
hypothesis that the density is increasing on the respective wedge. Similarly, a dotted wedge
implies that the test rejected the hypothesis that the density is decreasing. Non-significant
wedges are not marked.
The map indicates the existence of modes close to the grid points (−2, 0)⊤ and (2, 0)⊤ and in
a weaker sense indicates a mode close to the grid point at (0, 2)⊤ . The marked geometrical
17
Figure 4:
The density of a (uniform) mixture of
N ((−1.9, −0.07)⊤ , 0.2I) and N ((2, −0.1)⊤ , 0.25I) distribution.
a
N ((−0.05, 2.1)⊤, 0.5I),
objects around these grid points are shown in Figure 5. In the grid point at (0, 2)⊤ we obtain
not so many significant rejections as in the wedges with vertex (2, 0)⊤ . Still, the dotted wedges
show that there is a significant increase towards the mode which gives an indication for the
presence of a mode as well. An improved procedure with a direct focus on the modes will be
discussed in the following section.
Figure 5: Indications for the presence of a mode. Left panel: grid points (−2, 0)⊤ and (2, 0)⊤ .
Right panel: grid point (0, 2)⊤ .
4.3
Mode detection
In this section we demonstrate how the multiscale test can be successfully used for the localization of modes if the inference on the local monotonicity behaviour of the density is not included
in the test statistic. More precisely, we consider the grid introduced in Section 4.2. Similar
to the local test on modality, we conclude that the density has a mode close to a grid point
if every test on every wedge whose vertex is given by the grid point rejects that the density
is increasing on the respective wedge. We again recommend a calibrated version of the global
18
test, where the quantiles are chosen such that the probability of the discovery of a non-existing
mode is approximately 5% if the data comes from a two-dimensional uniform distribution. The
following results are based on n = 2500 observations and 1000 simulation runs.
We have investigated two densities, a constant density on the square [−3.5, 3.5] × [−1.5, 3.5]
and the density with three modes presented in Figure 4. For the uniform distribution the test
found a mode in 4.6% of the simulation runs. For the the tri-modal density the test detected in
78.9% cases a mode in the point (−2, 0)⊤ , in 53.8% cases a mode in (2, 0)⊤ and in 7.4% cases
a mode in the point (0, 2)⊤ .
5
Real data example
Active Galactic Nuclei (AGN) consist of a supermassive black hole at their centre surrounded
by an accretion disc. For some AGN classes, high-energy relativistic jets perpendicular to the
disc are produced. In case this jet is pointing in the general direction of the Earth, the object
is referred to as a blazar. The BL Lacertae type object Markarian (Mrk) 501 is such a blazar
and one of the closest (in a distance of 4.7 × 108 light years) and brightest extragalactic sources
in X-rays and gamma-rays. It is known as one of the most extreme blazars and features very
strong and fast variability, making Mrk 501 a perfect candidate for probing AGN. Due to the
strong emission over the entire electromagnetic spectrum, correlation studies between different
energy bands (parts of the electromagnetic spectrum) are particularly interesting and will give
insights into the processes inside an AGN, e.g. the emission models or the particle populations,
since different spatial regions of the object may emit radiation of diverse energies. Therefore,
the determination of the position of the radiation in a certain energy regime is of paramount
importance.
Here, 19 individual observations of Mrk 501 in the year 2015, performed in photon counting
(PC) mode by the Swift-XRT on board the Swift satellite, are analysed. The Swift satellite
was launched in 2004 and is a multiwavelength space observatory with three instruments on
board, the X-ray Telescope (XRT) being one of them, which is capable to observe X-rays in
the 0.3 − 10 keV energy regime. For each considered observation, the High Energy Astrophysics
Science Archive Research Center (HEASARC1 ) provides an image, based on Level 2 event files
that have been calibrated and screened by a standard pipeline. These images with a size of
1000 × 1000 pixels contain the information how many photons (i.e. which X-ray flux) have been
recorded in each pixel during the exposure time. The exposure times of the analysed images
range from about 100 s to 1000 s. Due to different positions of the satellite in space and different
alignments of its main axis, each image shows a slightly different region of the sky. Figure 6
provides an illustration of the data obtained from one observation.
By a combination of the 19 individual observations, we have at our disposal 49248 observations
of X-ray photons with known positions of origin on the sky. Our aim is the precise localiza1
NASA/Goddard Space Flight Center, https://heasarc.gsfc.nasa.gov/.
19
Figure 6: Observation of Mrk 501 from Sep. 4, 2015.2
tion of the mode of the distribution. To this end, we conduct the test presented in Section
4.3 for a significance level of 0.01. We chose an equidistant grid covering [253.446◦ , 253.586◦] ×
[39.64◦ , 39.88◦] (corresponding to [16 h 53 min 47.04 s, 16 h 54 min 20.64 s]×[39◦ 38′ 24′′ , 39◦ 52′ 48′′ ])
consisting of 961 grid points with mesh size 0.008◦ . The length of any wedge is l = 0.004◦ and
all angles are given by ϕ = π4 . Again, we used a calibrated version of the test where the quantiles are chosen such that a non-existing mode for a uniform distribution on [253◦ , 253.8◦ ] ×
[39.5◦ , 40.1◦] has been found in less than 1% of the simulation runs (based on 1000 simulation
runs). Our test detected the mode of the distribution in (253.466◦ , 39.760◦) (corresponding to
(16 h 53 min 51.84 s, 39◦ 45′ 36′′ )). The precision regarding the location of this mode is given
by the mesh size 0.008◦ .
In contrast to the high confidence detection possible within short exposure times for the
brightest extragalactic sources, such as Mrk 501, the detection of fainter point sources is more
challenging. Often only few photons reach the detector due to, e.g., the large distance to the
source and the absorption of photons. Within the sky region of one Swift-XRT image, there
might be multiple point sources in X-rays, but even background fluctuations can look like faint
point sources. The study of the population of these point sources, the correlation to other
energy bands, and variability studies contribute enormously to the understanding of the X-ray
sky. This requires reliable methods for the detection and the determination of the position,
including the confidence of a given calculation. In the following, the capability to determine
multiple modes of faint point sources in Swift-XRT images is demonstrated. 18 images provided
2
This picture has been created using HEAsoft, http://heasarc.nasa.gov/docs/software/lheasoft/.
20
Figure 7: Observation of the sky region around the blazar S3 0218+35 from Sep. 8, 2012.3
by HEASARC of individual observations of the sky region around the blazar S3 0218+35 in
the years 2005, 2012, 2014 and 2016 performed in PC mode by the Swift-XRT are analysed.
The exposure times of the images provided by HEASARC range from 3000 s to 5000 s. Figure 7
shows one of these images. The two point sources are marked with a square. Figure 8 provides
detailed images of the two point sources.
For the detection and the localization of the two point sources we conduct the test presented
in Section 4.3 at a significance level of 0.01. Here, we have at our disposal 18061 observations.
For this application, we chose an equidistant grid covering [35.2◦ , 35.32◦] × [35.825◦, 35.945◦ ]
(corresponding to [2 h 20 min 48 s, 2 h 21 min 16.8 s] × [35◦ 49′ 30′′ , 35◦ 56′ 42′′ ]) consisting of
961 grid points with mesh size 0.004◦ . The length of any wedge is l = 0.002◦ and all angles
are given by ϕ = π4 . The quantiles are chosen such that a non-existing mode for a uniform
distribution on [35◦ , 35.5◦ ] × [35.7◦ , 36.1◦ ] has been found in less than 1% of the simulation runs
(based on 1000 simulation runs). Our test detected the two point sources at (35.212◦ , 35.829◦)
(corresponding to (2 h 20 min 50.88 s, 35◦ 49′ 44.4′′ )) and at (35.272◦, 35.937◦) (corresponding
to (2 h 21 min 5.28 s, 35◦ 56′ 13.2′′ )) at a precision of 0.004◦.
Acknowledgements This research has made use of data obtained through the High Energy
Astrophysics Science Archive Research Center Online Service, provided by the NASA/Goddard
Space Flight Center. The authors would like to thank Martina Stein, who typed parts of this
3
This picture has been created using HEAsoft, http://heasarc.nasa.gov/docs/software/lheasoft/.
21
Figure 8: Observations of the two points sources around S3 0218+35 from Sep. 8, 2012.4
manuscript with considerable technical expertise. This work has been supported in part by the
Collaborative Research Center “Statistical modeling of nonlinear dynamic processes” (SFB 823,
Teilprojekt C1, C4) of the German Research Foundation (DFG).
References
Abraham, C., Biau, G., and Cadre, B. (2004). On the asymptotic properties of a simple estimate of
the mode. ESAIM Probab. Stat., 8:1–11 (electronic).
Burman, P. and Polonik, W. (2009). Multivariate mode hunting: data analytic tools with measures
of significance. J. Multivariate Anal., 100(6):1198–1218.
Chacón, J. and Duong, T. (2013). Data-driven density derivative estimation, with applications to
nonparametric clustering and bump hunting. Electronic Journal of Statistics, 7:499–532.
Chan, K. S. and Tong, H. (2004). Testing for multimodality with dependent data. Biometrika,
91(1):113–123.
Chaudhuri, K., Dasgupta, S., Kpotufe, S., and von Luxburg, U. (2014). Consistent procedures for
cluster tree estimation and pruning. IEEE Trans. Inform. Theory, 60(12):7900–7912.
Chaudhuri, P. and Marron, J. S. (1999). SiZer for exploration of structures in curves. J. Amer. Statist.
Assoc., 94(447):807–823.
Dasgupta, S. and Kpotufe, S. (2014). Optimal rates for k-NN density and mode estimation. Advances
in Neural Information Processing Systems, 27:2555–2563.
Donoho, D. L. and Liu, R. C. (1991). Geometrizing rates of convergence. III. Ann. Statist., 19(2):633–
667, 668–701.
4
These pictures have been created using HEAsoft, http://heasarc.nasa.gov/docs/software/lheasoft/.
22
Dümbgen, L. and Walther, G. (2008). Multiscale inference about a density. Ann. Statist., 36(4):1758–
1785.
Duong, T., Cowling, A., Koch, I., and Wand, M. P. (2008). Feature significance for multivariate kernel
density estimation. Comput. Statist. Data Anal., 52(9):4225–4242.
Fisher, N. I. and Marron, J. S. (2001). Mode testing via the excess mass estimate. Biometrika,
88(2):499–517.
Genovese, C., Perone-Pacifico, M., Verdinelli, I., and Wasserman, L. A. (2015). Nonparametric inference for density modes. J. R. Stat. Soc. Ser. B. Stat. Methodol.
Godtliebsen, F., Marron, J. S., and Chaudhuri, P. (2002). Significance in scale space for bivariate
density estimation. J. Comput. Graph. Statist., 11(1):1–21.
Grund, B. and Hall, P. (1995). On the minimisation of Lp error in mode estimation. Annals of
Statistics, 23:2264–2284.
Hall, P. and York, M. (2001). On the calibration of Silverman’s test for multimodality. Statistica
Sinica, 11:515–536.
Hartigan, J. A. (1987). Estimation of a convex density contour in two dimensions. Journal of the
American Statistical Association, 82(397):267–270.
Hartigan, J. A. and Hartigan, P. M. (1985). The dip test of unimodality. Ann. Statist., 13(1):70–84.
Klemelä, J. (2005). Adaptive estimation of the mode of a multivariate density. J. Nonparametr. Stat.,
17(1):83–105.
Li, J., Ray, S., and Lindsay, B. G. (2007). A nonparametric statistical approach to clustering via mode
identification. J. Mach. Learn. Res., 8:1687–1723.
Mammen, E., Marron, J. S., and Fisher, N. (1991). Some asymptotics for multimodality tests based
on kernel density estimates. Probab. Theory Relat. Fields, 91:115–132.
Minnotte, M. C. (1997). Nonparametric testing of the existence of modes. Ann. Statist., 25(4):1646–
1660.
Müller, D. W. and Sawitzki, G. (1991). Excess mass estimates and tests for multimodality. J. Amer.
Statist. Assoc., 86(415):738–746.
Pollard, D. (1981). Strong consistency of k-means clustering. Ann. Statist., 9(1):135–140.
Polonik, W. (1995). Measuring mass concentrations and estimating density contour clusters—an excess
mass approach. Ann. Statist., 23(3):855–881.
Romano, J. (1988). On weak convergence and optimality of kernel density estimates of the mode.
Annals of Statistics, 16:629–647.
Schmidt-Hieber, J., Munk, A., and Dümbgen, L. (2013). Multiscale methods for shape constraints in
deconvolution: confidence statements for qualitative features. Ann. Statist., 41(3):1299–1328.
Silverman, B. W. (1981). Using kernel density estimates to investigate multimodality. J. Roy. Statist.
Soc. Ser. B, 43(1):97–99.
Tsybakov, A. B. (1990). Recurrent estimation of the mode of a multidimensional distribution. Problemy
Peredachi Informatsii, 26(1):38–45.
23
6
Proofs
The assertions of most of our results are split up in two parts, one for densities that are
increasing and one for decreasing densities. Often, the proof for one case can be transferred in
a straightforward way to the other one. In this situation, we only prove one case as the other
case follows immediately by similar arguments. Only if this transfer is not obvious, we give
details for both situations.
6.1
Proofs of the results of Section 2.3
Proof of Theorem 2.3: It is well-known that, given X(N ) and N, the random variables
F̃ (Pe X(1) ), . . . , F̃ (Pe X(N −1) ) have the same distribution as the order statistics of N − 1 uniformly distributed and independent random variables on the interval [0, 1]. By an application
of the law of iterated expectations, the assertion follows.
Proof of Theorem 2.4: We only consider the case where the density f is increasing on K.
As F̃ (Pe X(j) ) 6= 0 almost surely for j = 1, . . . , N − 1 (cf. Theorem 2.3) and Pe X(j) = Pe X(N )
implies F̃ (Pe X(j) ) = 1, it is sufficient to prove
F̃ (z) ≤ F̃U (z) =
zd
(Pe X(N ) )d
(5.1)
for all z ∈ (0, Pe X(N ) ]. For this purpose, notice that the distribution function F̃ in Theorem
2.3 is given by
F̃ (z) =
P(Pe X ≤ z, X ∈ KN |N, X(N ) )
G(z)
=
,
P(X ∈ KN |N, X(N ) )
G(Pe X(N ) )
where the function G is defined by
G(z) =
Z
z
Z
t tan(ϕ)
...
−t tan(ϕ)
0
Z
t tan(ϕ)
f (x0 + te + s1 e1 + . . . + sd−1 ed−1 ) ds1 . . . dsd−1 dt .
−t tan(ϕ)
We now prove (5.1) by contradiction and assume that there exits z ∗ ∈ (0, Pe X(N ) ] such that
Z
z∗
0
(z ∗ )d
d
G(z ∗ )
>
=
f˜(t) dt =
G(Pe X(N ) )
(Pe X(N ) )d
(Pe X(N ) )d
Z
z∗
td−1 dt,
(5.2)
0
where f˜(t) := G(PGe X(t)(N) ) is the density of F̃ . From (5.2) and the monotonicity of the integral it
follows that there exists a point t0 ∈ (0, z ∗ ] with
′
f˜(t0 ) >
dtd−1
0
.
(Pe X(N ) )d
24
(5.3)
For the following discussion, we introduce an alternative parametrization of the wedge K. As
{e, e1 , . . . , ed−1 } is an orthonormal basis of Rd , every x ∈ K can be represented as x = x0 +
t1 e + s1 e1 + . . . + sd−1 ed−1 for some t1 ∈ (0, l] and si ∈ [−t1 tan(ϕ), t1 tan(ϕ)] for i = 1, . . . , d − 1.
With the notation
t1 e + s1 e1 + . . . + sd−1 ed−1
and t̃1 := kt1 e + s1 e1 + . . . + sd−1 ed−1 k,
(5.4)
e0 :=
kt1 e + s1 e1 + . . . + sd−1 ed−1 k
we have x = x0 + t̃1 e0 , and the mapping of (t1 , s1 , . . . , sd−1 ) to (t̃1 , e0 ) defines a bijection. Hence,
any x ∈ K can also be uniquely represented by the vector e0 and the scalar t̃1 (see Figure 9).
t̃1
e0
x
×
x0
s1
e
t1
K
Figure 9: Representation of x ∈ K for d = 2.
Let t1 ∈ [t0 , Pe X(N ) ] and consider a point
x = x0 + t1 e + s1 e1 + . . . + sd−1 ed−1 = x0 + t̃1 e0 ∈ KN ,
where e0 and t̃1 are defined in (5.4). Let t̃0 =
y := x0 + t̃0 e0 = x0 + t̃0
t0
kt e
t1 1
+ s1 e1 + . . . + sd−1 ed−1 k ≤ t̃1 and define
t1 e + s1 e1 + . . . + sd−1 ed−1
.
kt1 e + s1 e1 + . . . + sd−1 ed−1 k
A straightforward calculation shows that y = x0 + t0 e + s̃1 e1 + . . . + s̃d−1 ed−1 , where s̃i = tt10 si ,
(i = 1, . . . , d − 1). Note that Pe y = t0 and that f (y) ≤ f (x), as f is increasing on the wedge
K. We therefore obtain
Z t1 tan(ϕ)
Z t1 tan(ϕ)
f (x0 + t1 e + s1 e1 + . . . + sd−1 ed−1 )
ds1 . . . dsd−1
f˜(t1 ) =
...
G(Pe X(N ) )
−t1 tan(ϕ)
−t1 tan(ϕ)
Z t1 tan(ϕ)
Z t1 tan(ϕ)
f (x0 + t0 e + tt10 s1 e1 + . . . + tt10 sd−1 ed−1 )
ds1 . . . dsd−1 .
≥
...
G(Pe X(N ) )
−t1 tan(ϕ)
−t1 tan(ϕ)
d−1 Z t0 tan(ϕ)
Z t0 tan(ϕ)
t1
f (x0 + t0 e + s̃1 e1 + . . . + s̃d−1 ed−1 )
=
ds̃1 . . . ds̃d−1
...
t0
G(Pe X(N ) )
−t0 tan(ϕ)
−t0 tan(ϕ)
d−1
t1
f˜(t0 ).
=
t0
25
d−1 dtd−1
dtd−1
0
1
=
for any point t1 ∈ [t0 , Pe X(N ) ], and
Using (5.3) we have f˜(t1 ) > tt10
d
(Pe X(N) )
(Pe X(N) )d
from (5.2) we conclude
Z Pe X(N)
Z Pe X(N)
Z Pe X(N)
Z z∗
d−1
d−1
dtd−1
dt
dt
1=
dt
+
dt
=
dt = 1,
f˜(t) dt >
(Pe X(N ) )d
(Pe X(N ) )d
(Pe X(N ) )d
0
z∗
0
0
which is a contradiction and proves the first assertion of Theorem 2.4.
Proof of the representation (2.5): It follows from Theorem 2.3 that for fixed i, the random
variables F̃ i (Pein X1i ), . . . , F̃ i (Pein XNi i −1 ) are independent and uniformly distributed on the interval [0, 1], given {Xji ∈ KNi i } (j = 1, . . . , N i − 1) and N i . Here, F̃ i denotes the conditional
distribution function of Pein X on K i . Recall that the wedges KNi i , i = 1, . . . , Mn , are disjoint. Standard arguments and the law of iterated expectations yield stochastic independence
of observations lying in different wedges.
Proof of Theorem 2.5: Let a ∈ {−1, 1, 0}Mn be the vector determined by
if f is increasing on K i ,
1,
ai =
−1,
0,
if f is decreasing on K i ,
else.
The probability of at least one false rejection among all tests in (2.9) and (2.10) can be estimated
by Theorem 2.4, that is
p := Pf ∃i ∈ {i | ai 6= 0} : ai TK i < −c̃K i (α) N 1 , . . . , N Mn
≤ Pf ∃i ∈ {i | ai 6= 0} : ai TKU i < −c̃K i (α) N 1 , . . . , N Mn
≤ 1 − P |TKU i | ≤ c̃K i (α) for all i = 1, . . . , Mn N 1 , . . . , N Mn .
Using (2.6) we further deduce
r 3
N i
U
1
Mn
T
N
,
.
.
.
,
N
p ≤ 1−P
≤
κ̃
(α)
for
all
i
=
1,
.
.
.
,
M
−
Γ
i
n
n
Ni − 1 K
n−1
r 3
N i
U
1
Mn
= 1 − P max
= α.
T
≤
κ̃
(α)
N
,
.
.
.
,
N
−
Γ
i
n
i=1,...,Mn
Ni − 1 K
n−1
6.2
Further Results
This section provides a general consistency result which is the main ingredient for the proof of
Theorems 2.6 and 3.3. The consistency result stated in Theorem 6.3 below is a more general
result. The following notation is used throughout this section for the precise statement of the
assumptions on the density f .
26
Definition 6.1. Let K be a wedge with vertex x0 and j ∈ {1, 2}. We define
f (x0 + t̃e0 ) − f (x0 + s̃e0 )
,
t̃j − s̃j
x0 +t̃e0 ∈K,0≤s̃<t̃
f (x0 + t̃e0 ) − f (x0 + s̃e0 )
:=
sup
.
t̃j − s̃j
x0 +t̃e0 ∈K,0≤s̃<t̃
inf f (j) :=
K
sup f (j)
K
(5.5)
inf
For a better interpretation of Definition 6.1, let f be twice continuously differentiable in a
neighbourhood of x0 . A straightforward application of the mean value theorem shows
inf f (1) =
K
inf
x0 +t̃e0 ∈K
h∇f (x0 + t̃e0 ), e0 i,
sup f (1) =
K
sup h∇f (x0 + t̃e0 ), e0 i.
x0 +t̃e0 ∈K
If we have a mode in x0 , the gradient ∇f (x0 ) vanishes and supK f (1) vanishes as well. Similarly,
a Taylor expansion of order 2 yields
inf f (2) =
K
sup f (2) =
K
1
inf he0 , Hf (x0 )e0 i {x0 + t̃e0 , t̃ ≥ 0} ∩ K 6= ∅}
2
1
sup he0 , Hf (x0 )e0 i {x0 + t̃e0 , t̃ ≥ 0} ∩ K 6= ∅}
2
+ o(1)
+ o(1)
for l → 0, where Hf (x0 ) denotes the Hessian of f in x0 . In the situation of Theorem 2.6, the
d
condition e⊤
0 Hf (x0 )e0 ≤ −c < 0 holds for all e0 ∈ R with ke0 k = 1. Thus, only the case j = 2
is relevant for its proof. However, if the assumption is violated and the Hessian is indefinite,
the results can be generalized in a straightforward manner by considering orders j ≥ 3 as well.
For the proof of Theorem 3.3, the case j = 1 will be used.
To simplify notation, let
Z
π
ϕ̃ = tan(ϕ) for ϕ ∈ 0,
and F (K) =
f (x) dx.
2
K
The quantities
H+j (f, K) :=
(2ϕ̃)d−1 ld+j inf K f (j)
(2ϕ̃)d−1 ld+j supK f (j)
p
p
, H−j (f, K) :=
F (K)
F (K)
(j = 1, 2)
depend on the size of the wedge K through its length l and angle ϕ and on the monotonicity
of f on K and are the key objects in the following discussion. We begin by showing that the
quantiles κ̃n (α) and κn (α) defined in (2.6) and (3.2) are bounded from above by a constant
independent of Mn , N i , i = 1, . . . , Mn , and n. As a consequence, the same holds for the quantiles
κ̃′n (α) defined in (2.11).
Theorem 6.2. There exists a constant A > 0, such that max{κ̃n (α), κn (α)} ≤ A.
27
Proof: We only consider κ̃n as the result for κn can be shown similarly. From the discussion
in Section 2.3, it follows that
d
TKU i =
i −1
N
X
β(Uji ) for i = 1, . . . , Mn
j=1
(conditionally on N 1 , . . . , N Mn ), where Uji , j = 1, . . . , N i − 1, i = 1, . . . , Mn , are independent
uniformly distributed random variables on the interval [0, 1]. Recall the definition of Tjk (U ) in
(2.1), then we will show a the end of the proof that, conditionally on N 1 , . . . , N Mn ,
Pi
TPi−1 N k ,Pi
k=1
k=1
N k (U )
=
ℓ=
k −1
N
X
k=1
Pi−1
k=1
β U(ℓ;Pi−1 N k ,Pi
k=1
k=1
N k)
N k +1
d
=
i −1
N
X
β(Uji )
(5.6)
j=1
(i = 1, . . . , Mn ). As the statistic Tn (U ) defined in (2.2) calculates the maximum over more
scales than the statistic defined in (2.6), we obtain κ̃n ≤ κn , where κn is the (1 − α)-quantile of
the statistic Tn (U ) defined in (2.2). By Theorem 3.1 in Dümbgen and Walther (2008), there
exists an upper bound A > 0 for κn , which is independent of n. This completes the proof.
For a proof of (5.6), we finally note that for a sample of independent uniformly distributed
random variables U1 , . . . , Un on the interval [0, 1] and fixed 1 ≤ j < k ≤ n with k − j > 1,
the random variables U(l;j,k) = (U(l) − U(j) )/(U(k) − U(j) ) (l = j + 1, . . . , k − 1) have the
same distribution as the order statistic of k − j − 1 independent uniformly distributed random
variables on the interval [0, 1].
q
Following the notation from Section 2.3 (recall that Γ(δ) =
2 log( exp(1)
)), we define the
δ
quantiles
r
N i
Ni − 1
A+Γ
cKni (α) =
3
n−1
(i = 1, . . . , Mn ) and provide a general consistency result for locally increasing and decreasing
densities.
Theorem 6.3. Let j ∈ {1, 2} and Kn = {Kn1 , . . . , KnMn } be a family of Mn pairwise disjoint
wedges with length ln > 0 and angle ϕn ∈ (0, π2 ).
i
(i) If (for all i = 1, . . . , Mn ) the condition H+j (f, Kni ) ≥ D Γ(F√(K2 n )) + bn √1n holds for some
constant
√
j(d + j)2 2
D>
(5.7)
12
1
1
(2d+j)2 2
2j 2
j2
2
(2d + j) − 1 + 1 + (2d+j)2
1 − 2(2d+j)2 −1 + 1 + 4 j 2
and a positive sequence bn satisfying bn → ∞ and bn = o(
P TKni
p
log(n)) as n → ∞, then
> cKni (α) for all i = 1, . . . , Mn → 1.
28
i
(ii) If (for all i = 1 . . . , Mn ) the condition H−j (f, Kni ) ≤ −D Γ(F√(K2 n )) + bn √1n holds for some
constant
√
2 2(2d + j)(d + j)
D>
(d+j)/d h
i1/2
2d+j 2 1/2
d2
d
1 − 2(2d+j)2 − 1 + 1 + 4( d )
j 1 − d+j
P TKn
p
log(n)) as n → ∞, then
< −cKn (α) for all i = 1, . . . , Mn → 1.
and a positive sequence bn satisfying bn → ∞ and bn = o(
Remark 6.4. It follows from the proof of Theorem 6.3 below, that the bounds on H+j (f, Kn )
resp. H−j (f, Kn ) imply in particular that Mn = o(n) as n → ∞.
The proof of Theorem 6.3 is divided into eight parts: seven technical lemmas (Lemma 6.5 Lemma 6.11) and the main part of the proof. We first state and prove the technical lemmas
and finally combine the results in order to complete the proof of Theorem 6.3. For each of the
lemmas, we assume that the conditions of Theorem 6.3 hold. At first, we consider only one
fixed wedge K with length ln = l and angle ϕn = ϕ. To simplify notation, let
d−1
d+j
2d−1
2
j
z d −
z
F ± (z) := z + c±
d+j
d+j
for j ∈ {1, 2} and z ∈ [0, 1] with
cj+
(tan(ϕ))d−1 (Pe X(N ) )d+j inf K f (j)
:=
F (KN )
and
cj−
(tan(ϕ))d−1 (Pe X(N ) )d+j supK f (j)
:=
.
F (KN )
Lemma 6.5.
j
(i) If inf K f (j) > 0, then F̃ (z) ≤ F +
j
(ii) If supK f (j) < 0, then F̃ (z) ≥ F −
zd
(Pe X(N) )d
zd
(Pe X(N) )d
for z ∈ (0, Pe X(N ) ].
for z ∈ (0, Pe X(N ) ].
Proof: We only prove the first part and define the auxiliary function
j
f + (s1 , . . . , sd−1 , t) :=1KN (x0 + te + s1 e1 + . . . + sd−1 ed−1 )
inf f (j)
1
d
K
tj −
+
(Pe X(N ) )j ,
·
|KN |
F (KN )
d+j
R
where |KN | := KN 1 dx denotes the volume of KN . Note that
Z z Z t tan(ϕ)
Z t tan(ϕ)
j
f + (s1 , . . . , sd−1 , t) ds1 . . . dsd−1 dt
...
0
−t tan(ϕ)
d
z
−t tan(ϕ)
(j)
z
d
j
j
d−1 inf K f
t
−
+
(P
X
)
dt
(2
tan(ϕ)t)
e
(N
)
(Pe X(N ) )d
F (KN )
d+j
0
2d−1
2d−1
z d+j
zd
zd
zd
j
= F+
.
+ c+
−
=
(Pe X(N ) )d
d + j (Pe X(N ) )d+j
d + j (Pe X(N ) )d
(Pe X(N ) )d
=
Z
29
(5.8)
j
j
In particular, as F + (1) = 1, the function f + defines a density on KN . We now prove assertion
(i) by contradiction and assume that there exists z ∗ ∈ (0, Pe X(N ) ], such that
∗
F̃ (z ) =
Z
z∗
f˜(t)dt =
0
(z ∗ )d
G(z ∗ )
j
> F+
=
G(Pe X(N ) )
(Pe X(N ) )d
Z
z∗
hj+ (t)dt,
(5.9)
0
where f˜ and
hj+ (t)
:=
Z
t tan(ϕ)
...
−t tan(ϕ)
j
F +,
denote the density of F̃ and
exists a point t0 ∈ (0, z ∗ ] with
Z
t tan(ϕ)
j
−t tan(ϕ)
f + (s1 , . . . , sd−1 , t) ds1 . . . dsd−1
respectively. Due to the monotonicity of the integral, there
f˜(t0 ) > hj+ (t0 ),
(5.10)
f˜(t1 ) > hj+ (t1 ) for all t1 ∈ [t0 , Pe X(N ) ].
(5.11)
which implies
For a proof of (5.11), let t1 > t0 and x = x0 + t1 e + s1 e1 + . . . + sd−1 ed−1 ∈ KN . As in the proof
of Theorem 2.4, we use the representation x = x0 + t̃1 e0 , where t̃1 and e0 are defined in (5.4).
For t̃0 := tt01 t̃1 < t̃1 , let
y := x0 + t̃0 e0 = x0 + t0 e0 +
t0
t0
s1 e1 + . . . + sd−1 ed−1 ∈ KN .
t1
t1
Using inf K f (j) > 0, we find
f (x) − f (y) = f (x0 + t̃1 e0 ) − f (x0 + t̃0 e0 ) = (t̃j1 − t̃j0 )
f (x0 + t̃1 e0 ) − f (x0 + t̃0 e0 )
(5.12)
t̃j1 − t̃j0
tj
≥ (t̃j1 − t̃j0 ) inf f (j) = t̃j1 1 − j0 inf f (j) ≥ (tj1 − tj0 ) inf f (j) ,
K
K
t1 K
where the last estimate follows since t̃1 ≥ t1 . Recall that G(Pe X(N ) ) = F (KN ), then we obtain
t1 tan(ϕ)
t1 tan(ϕ)
f (x0 + t1 e + s1 e1 + . . . + sd−1 ed−1 )
ds1 . . . dsd−1
G(Pe X(N ) )
−t1 tan(ϕ)
−t1 tan(ϕ)
Z t1 tan(ϕ)
Z t1 tan(ϕ)
1
≥
...
f (x0 + t0 e + tt10 s1 e1 + . . . + tt10 sd−1 ed−1 )
F
(K
)
N
−t1 tan(ϕ)
−t1 tan(ϕ)
+ (tj1 − tj0 ) inf f (j) ds1 . . . dsd−1 .
f˜(t1 ) =
Z
...
Z
K
30
A change of variables yields
d−1 Z t0 tan(ϕ)
Z t0 tan(ϕ)
t1
f (x0 + t0 e + s̃1 e1 + . . . + s̃d−1 ed−1 )
˜
f (t1 ) ≥
...
ds̃1 . . . ds̃d−1
t0
F (KN )
−t0 tan(ϕ)
−t0 tan(ϕ)
+ (2 tan(ϕ)t1 )d−1 (tj1 − tj0 )
inf K f (j)
F (KN )
d−1
inf K f (j)
t1
f˜(t0 ) + (2 tan(ϕ)t1 )d−1 (tj1 − tj0 )
=
,
t0
F (KN )
and straightforward calculations show that
hj+ (t1 )
d−1
inf K f (j)
t1
hj+ (t0 ) + (2 tan(ϕ)t1 )d−1 (tj1 − tj0 )
=
t0
F (KN )
d−1
t1
inf K f (j)
<
f˜(t0 ) + (2 tan(ϕ)t1 )d−1 (tj1 − tj0 )
≤ f˜(t1 ),
t0
F (KN )
where we used (5.10) to obtain the strict inequality. From (5.9) and (5.11) we also get
Z
Pe X(N)
f˜(t) dt >
0
Z
z∗
hj+ (t) dt
+
0
Z
Pe X(N)
z∗
which contradicts the condition 1 = F̃ (Pe X(N ) ) =
Lemma 6.5.
R Pe X(N)
0
j
hj+ (t) dt = F + (1) = 1,
f˜(t) dt. This completes the proof of
Lemma 6.6.
(i) If inf K f (j) > 0, we have
(ii) If supK f (j) < 0, we have
PN −1
i=1
j
β (F + )−1 (F̃ (Pe X(i) )) ≤ TK .
PN −1
i=1
j
β (F − )−1 (F̃ (Pe X(i) )) ≥ TK .
j
Proof: We only prove the first part and begin showing that the function z 7→ F + (z) is strictly
increasing for z ∈ [0, 1]. Recalling the representation (5.8), it is sufficient to prove that the
inequality
j
f + (s1 , . . . , sd−1 , t) > 0 for t ∈ (0, Pe X(N ) ]
holds for all (s1 , . . . , sd−1 ) ∈ [−t tanϕ, t tanϕ]. For the sake of simplicity, we suppress the
j
j
j
dependence of f + on (s1 , . . . , sd−1 ) and note that the function t 7→ f + (t) = f + (t, s1 , . . . , sd−1 )
j
is strictly increasing. Therefore, it remains to show f + (0) ≥ 0. We prove this inequality by
j
contradiction and assume that f + (0) < 0. For x = x0 + te + s1 e1 + . . . + sd−1 ed−1 ∈ KN , it
follows (using t̃0 = 0 in (5.12))
f (x) − f (x0 )
inf K f (j)
f (x)
j
j
j
≥
≥ tj
= f + (t) − f + (0) > f + (t).
F (KN )
F (KN )
F (KN )
31
(5.13)
Integrating both sides of (5.13) leads to a contradiction. Consequently, the map z 7→ F̄+j (z) is
strictly increasing on [0, 1], which implies (using Lemma 6.5 and the monotonicity of the function
j
(Pe X(j) )d
β) that β((F + )−1 (F̃ (Pe X(j) ))) ≤ β Pe X(N)
, whenever F̃ (Pe X(j) ) 6= 0 and Pe X(j) 6= Pe X(N ) .
)d
However, it is easy to see that these cases correspond to Pe X(j) = 0 and F̃ (Pe X(j) ) = 1, where
there is in fact equality. Thus the proof of the first part is completed.
The conditional expectation considered in the following lemma is used to derive a bound on
the corresponding conditional probability via the Hoeffding inequality in Lemma 6.11 below.
Lemma 6.7.
(i)
(ii)
P
j −1
N −1
E
i=1 β (F + ) (F̃ (Pe X(i) )) |N, X(N ) =
P
j −1
N −1
E
i=1 β (F − ) (F̃ (Pe X(i) )) |N, X(N ) =
2d−1 j(N −1) j
c .
(2d+j)(d+j) +
2d−1 j(N −1) j
c̃ .
(2d+j)(d+j) −
Proof: We only prove the first part. Let U1 , . . . , UN −1 be independent uniformly distributed
random variables on the interval [0, 1]. Theorem 2.3 yields
N
−1
X
i=1
β
j
(F + )−1 (F̃ (Pe X(i) ))
d
=
N
−1
X
i=1
j
β (F + )−1 (Ui ) ,
given N and X(N ) . The assertion now follows from
Z 1
j −1
j
β(x)g+
(x) dx =
E β((F + ) (U1 ))|N, X(N ) =
0
j
where g+
(x) :=
j
d
F (x)
dx +
= 1 + cj+
2d−1 d+j
x d −1
d
−
2d−1
d+j
2d−1 j
cj ,
(2d + j)(d + j) +
.
In the following, we consider a sequence of wedges (Kn )n∈N given by vertices xn0 , directions en ,
lengths ln > 0 and angles ϕn ∈ (0, π2 ). Furthermore, we denote by en1 , . . . , end−1 the orthonormal
basis of (span{en })⊥ and let δn = F (Kn ), δNn = F (KNn ). Lemma 6.8 and Lemma 6.9 below
ensure the feasibility of our procedure in an asymptotic sense. They show that the random
wedge KNn is of similar size than the deterministic, predefined wedge Kn , i.e. that its complement Kn \KNn is small. Note that the test procedure can only be consistent if the wedge Kn
contains a sufficiently large number of observations. Therefore, we introduce for γ ∈ (0, 21 ] and
0 < ε < 1 the conditional probability PNn given the event
n
o
|KNn |
Nn = Nn ≥ (1 − γ)nδn ,
≥1−ε .
(5.14)
|Kn |
The results of Lemma 6.11 below are only shown for PNn . However, the following Lemmas 6.8
and 6.9 demonstrate that these conditions are asymptotically negligible. For example, Lemma
6.8 shows that, with increasing n, the wedge KNn ⊆ Kn approximates Kn in probability at an
exponential rate.
32
Kn
xn0
ln
en
1
ln (1 − ε) d
Ir n
Figure 10: The section Irn .
Lemma 6.8. Let j ∈ {1, 2} and 0 < ε < 1.
(i) If the assumptions of Theorem 6.3 (i) are satisfied, then
r
d+j
D
|KNn |
nδn
< 1 − ε ≤ exp −
Γ(δn ) 1 − (1 − ε) d
.
P Nn = 0 or
|Kn |
d+j
2
(ii) If the assumptions of Theorem 6.3 (ii) are satisfied, then
r
d+j
|KNn |
d+j
nδn
D
d
P Nn = 0 or
+
< 1 − ε ≤ exp −
Γ(δn ) (1 − ε)
ε−1 .
|Kn |
d+j
2
d
Proof: We only prove the first part. Since
1
d
ln (1 − ε) , we obtain
|KNn |
|Kn |
=
(Pen X(Nn ) )d
d
ln
< 1 − ε if and only if Pen X(Nn ) <
o
n
1
|KNn |
< 1 − ε = Nn = 0 or Pen X(Nn ) < ln (1 − ε) d .
Nn = 0 or
|Kn |
Define
n
o
1
Irn := Kn ∩ x ∈ Rd : Pen x ≥ ln (1 − ε) d
|K
|
(see Figure 10), then {Nn = 0 or |KNnn| < 1 − ε ⊆ { no observation in Irn }. Now, recall for a
proof of part (i) from (5.12) that (with t0 = 0 and t1 = t) f (xn0 + ten + s1 en1 + . . . + sd−1 end−1 ) ≥
tj inf Kn f (j) . Thus,
Z
pf,Irn : =
f (x) dx
I rn
ln
=
≥
Z
1
ln (1−ε) d
Z
ln
1
ln (1−ε) d
Z
tϕ̃n
...
−tϕ̃n
tϕ̃n
Z
−tϕ̃n
...
Z
tϕ̃n
−tϕ̃n
tϕ̃n
Z
−tϕ̃n
f (xn0 + ten + s1 en1 + . . . + sd−1 end−1 ) ds1 . . . dsd−1 dt
j
t inf f
Kn
(j)
d+j
(2ϕ̃n )d−1 d+j
ds1 . . . dsd−1 dt =
ln 1 − (1 − ε) d inf f (j) .
Kn
d+j
33
)
√ n and therefore pf,I
The assumptions of Theorem 6.3 (i) imply that H+j (f, Kn ) ≥ D Γ(δ
rn ≥
2n
q
d+j
D
δn
Γ(δn )(1 − (1 − ε) d ). As the variables Zi = 1Irn (Xi ) are Bernoulli distributed with
d+j
2n
parameters pf,Irn , we have
q
d+j
D
.
P( no observation in Irn ) ≤ exp − d+j nδ2n Γ(δn ) 1 − (1 − ε) d
As the number Nn of observations in Kn is Bin(n, δn )-distributed, we obtain the following result
from Chernoff’s Inequality.
2
Lemma 6.9. P (Nn ≤ (1 − γ)nδn ) ≤ exp − nδn γ2 for any γ ∈ (0, 12 ].
Lemma 6.10. Let j ∈ {1, 2}.
(i) If the assumptions of Theorem 6.3 (i) are satisfied, then
2
D 2 Γ(δn )2
j Γ(δn )
=:
c̃
,
+
(d + j)2 2
2
(a)
nδn ≥
(b)
nδn ≥ Ln c̃j+ log(exp(1)n), where Ln ≥ 1 − o(1) for n → ∞.
(ii) If the assumptions of Theorem 6.3 (ii) are satisfied, then
2
j 2 D 2 Γ(δn )2
j Γ(δn )
=:
c̃
.
−
(d(d + j))2 2
2
(a’)
nδn ≥
(b’)
nδn ≥ Ln c̃j− log(exp(1)n), where Ln ≥ 1 − o(1) for n → ∞.
Proof: We only prove the first part. As in the proof of Lemma 6.8, we obtain
Z
(2ϕ̃n )d−1 d+j
f (x) dx ≥
ln inf f (j) .
δn =
Kn
d+j
Kn
Hence,
H+j (f, Kn ) =
p
(2ϕ̃n )d−1 lnd+j inf Kn f (j) p
(2ϕ̃n )d−1 lnd+j inf Kn f (j)
√
≤ (2ϕ̃ )d−1 d+j
δn = (d + j) δn .
n
δn
ln inf Kn f (j)
d+j
Therefore, it follows from the assumption H+j (f, Kn ) that nδn ≥
consequence of Lemma 7.5 in Dümbgen and Walther (2008).
D 2 Γ(δn )2
.
(d+j)2
2
Part (b) is a
The following Lemma provides the key to prove consistency. Note that for the construction of
the test statistic TKn , it is necessary that at least two observations are contained in the wedge
Kn . Given the event Nn , we have that Nn ≥ 2 if nδn ≥ 4. If the assumptions of Theorem 6.3 (i)
hold, it follows from (a) that nδn ≥ 4 is fulfilled for D ≥ 2(d + j). Similarly, if the assumptions
of Theorem 6.3 (ii) hold, then (a’) yields the condition D ≥ 2d(d+j)
.
j
34
Lemma 6.11. Let j ∈ {1, 2} and PNn denote the probability conditional on the event Nn defined
in (5.14), and define for γ ∈ (0, 21 ], 0 < ε < 1 and η > 0 the constant
√
!
(2d+j)(d+j)
1
√
√
√
2
2
log
(1
−
γ)
−
κ
(α)
+
η
n
nδn
qj 3
Dj (η, δn ) :=
−
.
2+ 2
d+j
2
1
Γ(δ
)
Γ(δ
)
n
n
d
(1 − ε)
(1 − γ) − nδn
(5.15)
)
√ n for some constant D ≥ Dj (n, δn ) ∨ 2(d + j), then
(i) If H+j (f, Kn ) ≥ D Γ(δ
2n
2
η
.
PNn TKn ≤ cKn (α) Nn ≤ exp −
6
2d(d+j)
)
√ n for some constant D ≥ Dj (n, δn ) ∨
(ii) If H−j (f, Kn ) ≤ −D Γ(δ
, then
j
2n
2
η
PNn TKn ≥ −cKn (α) Nn ≤ exp −
.
6
Proof: We only prove the first part and define
cjn,+ :=
d+j
inf Kn f (j)
ϕ̃d−1
n (Pen X(Nn ) )
.
δNn
Then a tedious but straightforward calculation shows that the inequality
r
2d−1 jcjn,+ (Nn − 1)
Nn − 1
−η
≥ cKn (α)
(2d + j)(d + j)
3
|KNn |
|Kn |
≥ 1 − ε and Nn ≥ (1 − γ)nδn . This implies
PNn TKn ≤ cKn (α) Nn = E PNn TKn ≤ cKn (α)|Nn , X(Nn )
r
h
i
2d−1 jcjn,+ (Nn − 1)
Nn − 1
≤ E PNn TKn ≤
Nn , X(Nn )
−η
(2d + j)(d + j)
3
r
j
i
h
2d−1 jcn,+ (Nn − 1)
Nn − 1
Nn , X(Nn ) ,
−η
≤ E PNn RKn ≤
(2d + j)(d + j)
3
PNn −1
j
β((F + )−1 (F̃n (Pen X(i) ))). Therefore,
where we used Lemma 6.6 and the notation RKn = i=1
the assertion follows from Lemma 6.7 and Hoeffding’s inequality.
holds for
Proof of Theorem 6.3: For a proof of the first part we proceed in two steps: firstly, we
will find an upper bound for the probability that the test will not reject for one single wedge.
Secondly, we will consider the probability for simultaneous rejection on every wedge in Kn . For
a fixed wedge Kn ∈ Kn , 0 < ε < 1 and γ ∈ (0, 21 ], we have
P TKn ≤ cKn (α) for any single Kn ∈ Kn ≤ PNn TKn ≤ cKn (α) for any single Kn ∈ Kn +P(Nnc ),
35
where the event Nn is defined in (5.14). Notice that the assumptions of Theorem 6.3 imply
those of Lemma 6.11 and recall that F (Kn ) = δn . We have from the assumption in Theorem
6.3 (i) with Kn = Kni
√
)
√n ,
H+j (f, Kn ) ≥ D 1 + bn Γ(δ2n ) Γ(δ
2n
which relaxes the assumption on the constant D in Lemma 6.11 as follows. Let ηn,Kn > 0 and
o
n
√
D 1 + bn Γ(δ2n ) ≥ max Dj (ηn,Kn , δn ), 2(d + j) ,
(5.16)
where Dj is defined in (5.15). Therefore, it follows from Lemma 6.11, Lemma 6.8 and Lemma
6.9 that the probability under consideration can be bounded by
q
d+j
2
η2 n
nδn
D
exp − n,K
+
exp
−
Γ(δn )(1 − (1 − ε) d ) + exp − nδn γ2 ,
(5.17)
6
d+j
2
which concludes the proof for any single wedge.
We now consider the union of all wedges of Kn and define δ̃n := inf Kn ∈Kn F (Kn ). As the wedges
in Kn are pairwise disjoint, it follows from Lemma 6.10 that
1
= o(n).
Mn = #{Kn : Kn ∈ Kn } ≤
δ̃n
(5.18)
Therefore, Kn consists of a finite number of wedges. From (5.17) and the monotonicity of the
function δ 7→ δ log exp(1)
, we obtain the estimate
δ
P TKn ≤ cKn (α) for at least one Kn ∈ Kn
q
X
2
d+j
ηn,K
D
n
+ Mn exp − d+j n2δ̃n Γ(δ̃n )(1 − (1 − ε) d ) + exp −
exp − 6
≤
Kn ∈Kn
(5.19)
nδ̃n γ 2
,
2
if condition (5.16) is fulfilled for every Kn ∈ Kn . We now show that the right-hand side of
(5.19) vanishes as n → ∞ by investigating the asymptotic behaviour of every summand. For
1
the first summand, let ηn,Kn := 6 log δ1n + bn 2 , then
X
Kn ∈Kn
exp −
2
ηn,K
n
6
=
X
Kn ∈Kn
δn exp −
bn
6
= o(1),
P
because Kn ∈Kn δn ≤ 1 and bn → ∞ as n → ∞. Next, we consider the second summand in
(5.19). An application of Lemma 6.10 (b) gives
q
q
nδ̃n
Γ(δ̃n ) ≥ c̃j+ log(exp(1)n)(1 − o(1)).
2
Hence, by (5.18), if
D
d+j
q
D2
(d+j)2
1 − (1 − ε)
36
d+j
d
> 1,
(5.20)
Mn exp −
D
d+j
q
c̃j+ log(exp(1)n) 1 − (1 − ε)
Finally, by (5.18) and Lemma 6.10 (b), we have
o(1) exp − nδ̃n
if
d+j
d
(1 − o(1)) = o(1).
γ2
γ2
+ log(n) ≤ o(1) exp − (1 − o(1))c̃j+ log(exp(1)n) + log(n)
2
2
2
γ
≤ o(1) exp − log(n) c̃j+ − 1 + o(1)
= o(1),
2
D2 γ 2
> 1.
(d + j)2 2
(5.21)
In this case, the third term vanishes as well as n → ∞.
It remains to show that condition (5.16) is fulfilled for every Kn ∈ Kn . With κn (α) ≤ A, we
have to prove that
√
(5.22)
D 1 + bn Γ(δ2n )
√
√
√ A + ηn,Kn 2 2 log 1 − γ − nδ1n
(2d + j)(d + j)
q
−
.
≥√
2+ 2
d+j
Γ(δn )
Γ(δn )2
3j(1 − ε) d
(1 − γ) − 1
nδn
From Lemma 6.10 (b) it follows that nδn → ∞ for n → ∞ for all Kn ∈ Kn . Thus, using
1
ηn,Kn := 6 log δ1n + bn 2 , we find that for sufficiently large n an upper bound for the right
hand side of (5.22) is given by
√
√
A + {6 log δ1n }1/2 + bn
(2d + j)(d + j)
2+
√
1/2
d+j p
3j(1 − ε) d
(1 − γ) − o(1)
log δ1n
√
√
√
( 2 + 6)(2d + j)(d + j)
(2d + j)(d + j)(A + bn )
≤√
+√
(5.23)
1/2 p
d+j p
d+j
3j(1 − ε) d
1 − γ − o(1)
3j(1 − ε) d log δ1n
1 − γ − o(1)
2√2(2d + j)(d + j)
o(bn )
≤ 1+
.
d+j p
{log δ1n }1/2 j(1 − ε) d
1 − γ − o(1)
Combining (5.20), (5.21) and (5.23), we obtain the following condition
n
o
√
√
d+j −1/2
D
2 2(2d+j)
2
d
> max 1 − (1 − ε)
.
, γ ,
d+j √
d+j
j(1−ε) d
1−γ
(5.24)
In order to minimize the restrictions imposed by condition (5.24), we now determine 0 < ε < 1
and γ ∈ (0, 21 ], such that the lower bound on D is as small as possible. Balancing the second
and third terms in (5.24) we obtain
p
−j 2 + j 4 + 4j 2 (2d + j)2
j
1
γ=
<
≤ ,
2
2(2d + j)
2d + j
2
37
where we used (2d + j)2 ≥ 9 (note that d ≥ 1) for the first inequality. For the choice of ε we
d+j
introduce the notation a := (1 − ε) d and balance the first and third expression in (5.24) and
obtain
q
2
2j 2
a = (2d+j)
−
1
+
.
1
+
j2
(2d+j)2
Finally, inserting our choice of ε and γ in (5.24), we find the condition (5.7) since our calculations
also show that D is larger than all three terms of (5.24) simultaneously in this case.
6.3
Proof of Theorem 2.6 and 3.3
For the sake of simplicity, we prove both results for the case C2 = 1 and C := C1 . The general
case follows by exactly the same arguments with an additional amount of notation.
Proof of Theorem 2.6: We note that it follows from Theorem 6.2 that cK i (α) ≥ c̃K i (α)
for i = 1, . . . , Mn . Hence, it remains to show that the assumptions for Theorem 6.3 (ii) are
satisfied. By assumption on f , we have supK i f (2) ≤ − 2c + o(1) for n → ∞ (i = 1, . . . , Mn ).
Moreover, from the approximation tan(x) = x(1 − o(1)) for x → 0, we have
d
log(n) d+4
d−1
1
(1 + o(1)) (i = 1, . . . , Mn ).
(5.25)
F (K i ) = f (x0 ) C d log(n)−d+1+d d+4
d
n
Hence,
d+2
d+4
d−1
1
(c/2 − o(1))
log(n)−d+1+(d+2) d+4 C d+2 log(n)
d
2
i
+2 log(n) 2 c/2 − o(1)
n
2
q
−H− (f, K ) ≥ q
=
C
d
2(d+4)
d−1
−d+1
n
d
(1
+
o(1))
f (x0 ) 1d C 2 log(n) 2 +d 2(d+4) log(n)
f (x0 ) d1
n
p
(i = 1, . . . , Mn ). Furthermore, we obtain from (5.25) and the assumption bn = o( log(n)) that
q
i ))
bn
d log n
√
+
D Γ(F√(K
=
D
(1 + o(1)).
n
d+4 n
2n
Therefore, the assumptions of Theorem 6.3 (ii) (for j = 2) are fulfilled as the constant C
2
q
f (x0 ) d+4
by (2.12) and hence the assertion follows by an application of
satisfies C > 2D
c
d+4
Theorem 6.3.
2
Proof of Theorem 3.3: Note that it is sufficient to prove consistency for the largest scale.
By Theorem 6.2, we have to prove that the assumptions for Theorem 6.3 (ii) are satisfied for
the family of wedges {Kni | i ∈ In } introduced in Section 3.3. Let Kn ∈ {Kni | i ∈ In }. We
begin with the determination of an upper bound for the quantity supKn f (1) defined in (5.5).
For this purpose, consider a point xn0 + t̃e0 ∈ Kn with e0 ∈ Rd (ke0 k = 1) and t̃ > 0. Now, the
representation (3.6) and an application of the mean value theorem yields for 0 ≤ s̃ < t̃
f (xn0 + t̃e0 ) − f (xn0 + s̃e0 )
t̃ − s̃
n
= f˜x0 (kx0 + s̃e0 − x0 k)h∇gx0 (ξ2 ), e0 i + (1 + gx0 (xn0 + t̃e0 ))f˜x′ 0 (ξ1 )Rn
∆n =
38
(5.26)
for some kxn0 + s̃e0 − x0 k ≤ ξ1 ≤ kxn0 + t̃e0 − x0 k and ξ2 ∈ [xn0 + s̃e0 , xn0 + t̃e0 ], where
kxn0 + t̃e0 − x0 k − kxn0 + s̃e0 − x0 k
.
t̃ − s̃
A further application of the mean value theorem gives
Pd
Pd
n
2
n
2
j=1 (x0,j + t̃e0,j − x0,j ) −
j=1 (x0,j + s̃e0,j − x0,j )
√
Rn =
2 ξ(t̃ − s̃)
Pd
Pd
n
2
2
n
j=1 2(x0,j − x0,j )e0,j (t̃ − s̃) + t̃ − s̃
j=1 (x0,j − x0,j )e0,j
√
√
=
≥
ξ
2 ξ(t̃ − s̃)
n
n
1
cos(angle(x0 − x0 , e0 ))kx0 − x0 k
√
≥ C log(n)ln (1 − o(1)) √
=
ξ
ξ
Rn =
for some kxn0 + s̃e0 − x0 k2 ≤ ξ ≤ kxn0 + t̃e0 − x0 k2 . Moreover, we have
2
ξ ≤ kxn0 + t̃e0 − x0 k2 ≤ kxn0 − x0 k + t̃ ≤ ln2 (mn C log(n))2 (1 + o(1)).
Hence, Rn ≥ 1−o(1)
. With the same arguments as used before, one shows that ξ1 ≥ C log(n)ln
mn
and kξ2 − x0 k ≤ Cmn log(n)ln (1 + o(1)). Finally, by assumption on gx0 and f˜x0 and (5.26), this
yields
∆n ≤ (1 + o(1))(−c)C log(n)ln Rn + f (x0 )o((mn log(n)ln )1+γ )
cC
ln
+ o((mn log(n)ln )1+γ ) = −
log(n)ln (1 − o(1)),
≤ (1 − o(1))(−c)C log(n)
mn
mn
γ
as mlnn o((mn ln )1+γ ) = o(m2+γ
n ln ) = o(1) as n → ∞ by the choice of mn and ln . Consequently,
cC
log(n)ln (1 − o(1)). As tan(x) = x(1 − o(1)) (x → 0), we have
supKn f ′ ≤ − m
n
d
n 1 d
−d+1+d log(n) d+4
F (Kn ) = f (x0 ) C (log(n))
(1 + o(1))
d
n
d
log(n) d+4
1
= f (x0 ) C d log(n)
(1 + o(1)).
d
n
It follows from the conditions c1 ≥ f (x0 ) ≥ 0
d+2
d+1
3 log(n) d+4 cC
(1 − o(1))
C
(log(n))
n
mn
H−1 (f, Kn ) ≤ − q
d
2(d+4)
d
1
f (x0 ) d1 C 2 (log(n)) 2 log(n)
(1 + o(1))
n
log(n) 12 1 − o(1) (log(n)) 52
log(n) 12 1 − o(1)
d
d
+2
+2
2
2
= −cC
.
d
d
≤ −cC
√
√
n
c1
mn
n
c1
√
If bn = o( log n), we have
q
bn
d log n
n ))
√
=
D
+
(1 + o(1)),
D Γ(F√(K
n
d+4 n
2n
and the assumptions of Theorem 6.3 (ii) are fulfilled if the constant C satisfies C >
which is a direct consequence of (3.7).
39
p
D
c
c1
d+4
2
d+4
2
,
This figure "TestonMonotonicity.png" is available in "png" format from:
http://arxiv.org/ps/1604.04405v1
| 10math.ST
|
A study on topological descriptors for the analysis of 3D surface texture
Matthias Zeppelzauera , Bartosz Zielińskib , Mateusz Judab , Markus Seidla
a Media
Computing Group, Institute of Creative Media Technologies,
St. Pölten University of Applied Sciences,
Matthias-Corvinus Strasse 15, 3100 St. Pölten, Austria
b The Institute of Computer Science and Computer Mathematics,
Faculty of Mathematics and Computer Science, Jagiellonian University,
ul. Łojasiewicza 6, 30-348 Kraków, Poland
arXiv:1710.10662v1 [cs.CV] 29 Oct 2017
Abstract
Methods from computational topology are becoming more and more popular in computer vision and have shown to improve the
state-of-the-art in several tasks. In this paper, we investigate the applicability of topological descriptors in the context of 3D surface
analysis for the classification of different surface textures. We present a comprehensive study on topological descriptors, investigate
their robustness and expressiveness and compare them with state-of-the-art methods including Convolutional Neural Networks
(CNNs). Results show that class-specific information is reflected well in topological descriptors. The investigated descriptors can
directly compete with non-topological descriptors and capture complementary information. As a consequence they improve the
state-of-the-art when combined with non-topological descriptors.
Keywords: Surface Texture Analysis, 3D Surface Classification, Surface Topology Analysis, Surface Representation, Persistent
Homology, Persistence Diagram, Persistence Image
1. Introduction
In recent years, methods for sparse and dense reconstruction of 3D scenes have progressed strongly due to the availability of inexpensive off-the-shelf hardware like Microsoft Kinect
and the development of robust 3D reconstruction algorithms
(e.g. structure from motion techniques, SfM) [1, 2]. Thus,
the amount of available 3D data is rising constantly. Furthermore, the reconstruction accuracy is increasing strongly, which
enables 3D reconstructions with sub-millimeter resolution [3].
The high resolution enables the accurate description of the geometric micro-structure of a surface, which opens up new opportunities for search and retrieval in 3D scenes, such as the recognition of objects by their specific surface properties as well as
the distinction of different types of materials for improved scene
understanding.
The geometric micro-structure determines the haptic appearance of a surface (e.g. in terms of roughness, waviness, and lay)
and is also referred to as surface texture1 [5, 6, 7]. The geometric structure of a surface is closely related to its topology and
thus, topological descriptors are promising candidates for its
representation. In this article, we investigate the capabilities of
topological analysis of describing and classifying 3D surfaces
according to their geometric micro-structure and present a first
extensive study on topological descriptors in this context.
Email addresses: [email protected] (Matthias
Zeppelzauer), [email protected] (Bartosz Zieliński)
1 Note that surface texture is different from the apparent texture or visual
texture of a surface which refers to the visual (radiometric) appearance of the
surface in terms of color variations, brightness, and reflectivity [4].
Preprint submitted to Elsevier
For our study we employ high-resolution 3D reconstructions of natural rock surfaces from the archaeological domain,
see Fig. 1a which exhibit different surface textures. We map the
surfaces to depth maps (Fig. 1b) [8] and analyze the maps in
a patch-wise manner. For each surface patch we extract topological descriptors and traditional non-topological descriptors.
Next, we train a classifier and try to predict the class of texture
from each patch. The result is a class label for each surface
patch. We employ precise maps with ground-truth labels (see
Fig. 1c) to evaluate the classification accuracy.
This article is an extended version of the work in [9] where
we reported first preliminary results for topological surface texture analysis. This article significantly extends this work to a
comprehensive and in-depth study on topological descriptors
for 3D surface analysis. We evaluate different extensions and
variants to the proposed topological descriptors and study in
detail their robustness and discriminativity. In particular the
contribution of this paper over the previous one is an investigation of: (i) the sensitivity of the topological descriptors to
their computational parameters; (ii) the influence of different
normalization variants; (iii) the role of outliers in persistence
diagrams [10]; (iv) the extension of topological descriptors with
additional pre-filtering operations (v) the robustness and the explanatory power of the descriptors and (vi) strategies to reduce
the dimension of the descriptors’ feature vectors. Moreover,
we verify the validity of our results on an extended large-scale
dataset [11]. To provide the complete picture of our analysis we
include the results from [9] in this paper.
Our study reveals that topological descriptors are robust 3D
surface descriptors which can compete and partly outperform
October 31, 2017
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
Z
X
Y
(a)
y
y
x
x
(b)
(c)
Figure 1: 3D surface data: (a) the 3D point cloud of the surface; (b) the depth
projection of the surface; (c) ground truth labeling that specifies areas with
different surface texture (blue and green).
traditional (non-topological) descriptors. Moreover our investigations clearly show that topological descriptors capture complementary information to non-topological descriptors and that
this information is necessary to obtain peak performance in
classification. Remarkably, our results on the extended dataset
show that our combinations of topological and non-topological
descriptors yield a similar performance level (and partly even
higher performance) as deep learning in [11].
The remaining article is organized as follows. In Section 2
we review related work on 3D surface analysis and relevant
topological approaches in this context. Section 3 introduces the
investigated topological descriptors together with their properties and outlines our topological approach for surface texture
analysis. Section 4 summarizes the experimental setup and
Section 5 presents the study on topological descriptors together
with the obtained results. We summarize the major conclusions
of our study in Section 6.
2. Related work
Approaches for 3D surface analysis can be partitioned into
two classes: methods that operate in 3D space directly and
methods that operate on derived image-space representations.
The following sections review related work on both approaches.
2.1. Surface analysis in 3D space
The natural domain for the analysis of surfaces is the 3D domain. A popular approach is to analyze the mesh or point cloud
of a given surface reconstruction with local 3D surface descriptors that describe the local geometry around a 3D point. Numerous descriptors have been proposed for this purpose in recent
years. Johnson and Hebert propose spin images as local descriptors for the dense description of meshes for object recognition in 3D scenes [12]. Darom and Keller extract SIFT features
from local depth images of a points’ neighborhood to model the
surface geometry around it [13]. Zaharescu et al. introduce a local surface descriptor for meshes (MeshHOG) that can be computed from an arbitrary scalar function (e.g. curvature) defined
over the surface [14]. Steder et al. propose the normal-aligned
radial feature (NARF) which captures local depth variations in
range data [15]. Frome et al. extend the well-known 2D shape
context descriptor [16] to 3D point clouds (3DSC) and show
that it outperforms spin images [17]. Rusu et al. propose two
local 3D point cloud descriptors, namely persistent point feature histogram (PFH) [18] and an accelerated version fast PFH
(FPFH) [19]. Both build upon the relations between surfels, i.e.
the combination of a point and its surface normal to describe the
surface geometry [20]. Tombari et al. propose a 3D descriptor
(SHOT) based on the point normals that is defined in a robust
local reference frame [21].
Local descriptors are particularly well-suited for the sparse
description of a surface, i.e. descriptors are computed at salient
feature points only. The analysis of 3D surfaces by their surface texture - as investigated in this paper – requires the dense
analysis and representation of a surface. Previous experiments
have shown that local 3D descriptors are impractical for this
purpose, especially due to computational reasons, because descriptors have to be computed at each surface point to obtain a
dense representation, which becomes highly demanding when
a point cloud has several millions of points [8].
2.2. Surface analysis in image-space
To circumvent the problems raised by 3D descriptors in the
context of dense surface analysis, Othmani et al. propose to
map 3D surfaces to image-space and then apply efficient image
analysis techniques there to process the projected surfaces [22].
Usually a 2D depth map is derived from the 3D surface (piecewise or globally, depending on the amount of curvature of the
surface) to obtain an image-space representation. If the surface
does not contain self-occlusions, the depth map completely represents the 3D surface, except for inaccuracies or losses in resolution due to rasterization [8].
The analysis of 3D surface texture in image-space can be
considered as texture analysis task on the depth map. Thus,
approaches of image texture analysis become applicable. Popular methods for image texture analysis include histograms of
vector quantized filter responses [23] and later generalizations
such as the bag-of-visual-words model for textures [24] and the
Fisher vector [25]. Recently, deep learning-based approaches
for texture analysis have been introduced for the problem of
image texture analysis [26], which outperform many existing
methods.
An alternative approach to texture analysis (both for image
texture and surface texture) is to consider the underlying data
(depth map in our case) as a geometric object and to analyze its
topological properties. For this purpose topological data analysis (TDA) is a promising approach which currently gains increasing importance in data science. TDA provides methods
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
to study qualitative properties of objects based on the theory of
size functions introduced by Frosini et al. [27, 28] and persistent
homology by Edelsbrunner et al. [29]. TDA methods have recently been successfully applied to computer vision problems,
such as shape and texture analysis. Li et al. combine the bagof-features approach with persistence diagrams (PDs) for shape
retrieval [30]. Reininghaus et al. proposes an SVM kernel for
persistence diagrams to make them compatible with statistical
machine learning methods [31]. The authors apply topological descriptors together with the novel kernel to shape retrieval
and texture classification in images. Another approach to leverage topological information for machine learning tasks is the
persistent image (PI) proposed by Adams et al. [32]. The PI
is a vector-based representation of PD which is built upon earlier work on size functions [33, 34]. The PI can be employed
directly as a feature vector in conventional machine learning
techniques and can easily be combined with other descriptors.
Our work mainly builds upon the approach of Adams et al.
In our study we investigate different topological descriptors, including PI, for the domain of surface texture analysis. The difference to the investigations in [31] and [32] is that we try
to enrich the PI representation by pre-filtering, normalization,
and feature selection and that we analyze in-depth the information captured by the PI descriptor, its redundancy and discriminativity. Furthermore, we investigate the sensitivity of the
representation to its computational parameters. Beyond this we
combine them with non-topological state-of-the-art descriptors
to investigate synergy effects.
3. Topological approach
3.1. Overview
Fig. 2 provides an overview of the approach that we propose
for the analysis of 3D surfaces by persistence homology features. The input 3D surfaces are first projected to image-space
to obtain depth maps as proposed in [8]. Next, depth maps are
z-standardized (see Section 3.5) and split into square patches.
Patches are normalized (optional step, see Section 3.5) and prefiltered by Schmid or MR filters or CLBP (optional step, see
Section 3.4). Next, topological descriptors are extracted as described in Sections 3.2 and 3.3. Finally feature selection is performed (optionally) before patch-wise classification to decrease
the dimensionality of the feature vector and to reduce redundancies. The result is a class label for each surface patch, which
were mapped back to the surface in Fig. 2 for illustration.
In the following, we describe the individual components of
the approach and start with its core component, namely the persistent homology.
3.2. Persistent homology
By mathematical standards, topology with its 120 years of
history, is a relatively young discipline. It grew out of Poincares
seminal work on the stability of the solar system as a qualitative tool to study the dynamics of differential equations without
explicit formulas for solutions [35, 36, 37]. Due to the lack of
useful analytic methods, topology soon became a purely theoretical discipline. However, in the last few years we observe
a rapid development of topological data analysis tools, which
open new applications for topology.
Topological spaces appearing in data analysis are typically
constructed from small pieces called cells. A natural tool in the
study of multidimensional images with topological methods are
hypercubes (points, edges, squares, cubes etc.), e.g. a pixel in
a 2-dimensional image is equivalent to a square, a voxel in a
3 dimensional volume is equivalent to a cube. Hypercubes are
building blocks for structures called cubical complexes. Such
representations give topology a combinatorial flavour and make
it a natural tool in the study of multi-dimensional data sets.
Intuitively, the rank of the nth homology group, the so called
nth Betti number denoted by βn , counts the number of n-dimensional
holes in the topological space. In particular, β0 counts the number of connected components. As an example consider the image of the digit 8. In this image there is one connected component and two holes, hence β0 = 1 and β1 = 2. For a hollow
sphere we have β0 = 1, β1 = 0, β2 = 1. For a tube in a tire
(torus) we have β0 = 1, β1 = 2, β2 = 1.
Betti numbers do not differentiate between small and large
holes. In consequence, the holes resulting from the noise in
the data cannot be distinguished from the holes indicative for
the nature of the data. For instance, in a noisy image of the
digit 8 one can get easily β0 > 1. A remedy for this drawback
is persistent homology, a tool invented at the beginning of the
21st century [29]. Persistent homology tracks the holes from
birth to death when the topological space is gradually built by
adding cubes in some prescribed order.
If X is a cubical complex, one can add cubes step by step.
Typically, the construction goes through different scales, starting from the smallest pieces. However, in general an arbitrary
function f : X → R, called the Morse function or the measurement function, may be used to control the order in which the
complex is built, starting from low values of f and increasing
subsequently. This way we obtain a sequence of topological
spaces, called a filtration,
∅ = Xr0 ⊂ Xr1 ⊂ Xr2 ⊂ · · · ⊂ Xrn = X,
where Xr := f −1 ((−∞, r]) and ri is a growing sequence of values
of f at which the complex changes. As the space is gradually
constructed, holes are born, persist for some time and eventually may die. The length of the associated birth-death intervals (persistence intervals) indicates if the holes are relevant or
merely noise. The lifetime of holes is usually visualized by
the persistence diagram (PD). Persistence diagrams constitute
the main tool of topological data analysis. They visualize geometric properties of a multidimensional object X in a simple
two-dimensional diagram.
Fig. 3a shows the depth map of a 3D surface where colors
correspond to different depths (blue refers to low depth, yellow to high depth). In this case pixels are represented as 2dimensional cells of a cubical complex. For the complex we
can obtain a filtration Xr using a measuring function which has
a value for a 2-dimensional cube equal to depth (pixel color).
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
Image Space
Projection
Z
Y
Z-Standardization
X
y
3D surface
x
Schmid
Filters
MR
Filters
Depth Map
CLBP
Patch-based
Normalization
Pre-filtering
Patching
y
x
Surface Patches
1 2 3
Topological
Analysis
4
5
6
7 8
9 10 11 12
PD_AGG
Feature
Selection
Classification
x
Class Labels
PI
Figure 2: Overview of the proposed approach.
elements, minimum, maximum, mean, standard deviation,
variP√ P
ance, 1-quartile, median, 3-quartile, and norms
di , di ,
P
and (di )2 . The result is a 12-dimensional feature vector.
10
Death time
20
40
60
80
5
5
0
10
100
120
20
40
60
(a)
80 100 120
-5
-5
15
0
5
10
Birth time
(b)
5
10
15
(c)
Figure 3: Example surface patch with derived topological descriptors: (a) the
depth map of the original 3D surface; (b) the corresponding persistent diagram;
(c) and the persistence image (resolution 16 × 16 pixels, σ x = σy = 0.001 and
exponential weighting function g).
For a lower dimensional cell (a vertex or an edge) we can set
the function value as a minimum from the higher-dimensional
neighborhoods of the cell. Fig. 3b shows the persistence diagram for Xr .
The PD is an expressive representation of the surface topography. Unfortunately, the data structure of the PD is complex and cannot easily be combined with traditional machine
learning methods which require fixed-length input vectors. This
hinders the application of computational topology to computer
vision and machine learning tasks. There is still no specific answer on how and when the tools of computational topology and
machine learning should be used together.
A first attempt to obtain a fixed-length descriptor of a topological space filtration is to compute elementary statistics of
persistence intervals (or equivalently persistence diagrams). Let
I := {[b1 , e1 ], [b2 , e2 ], . . . , [bn , en ]}
be a set of persistence intervals. Let D := {di := (ei −bi )}ni=1 be a
set of the interval lengths. We build an aggregated descriptor of
D, denoted by PD AGG, using following measures: number of
3.3. Persistence image
Aside from the PD AGG descriptor described above, which
can be used with standard classification methods, there are also
attempts to use PD directly with appropriately modified classifiers. Reininghaus et al. [31] proposed a multi-scale kernel for
PDs, which can be used with a support vector machine (SVM).
While this kernel is well-defined in theory, in practice it becomes highly inefficient when the number of training vectors
becomes large, as the entire kernel matrix must be computed
explicitly (note that this applies to all kernel-based methods).
There are alternative methods that aim at representing the PD
as a vectorial representation. An example is a persistence landscape (PL) [38] which is a functional representation of a PD.
The representation lies in a Banach space and is stable. Another approach to obtain a vectorial representation from PD is
persistent image introduced by Adams et al. [32], which according to Makarenko et al. [39] is especially well-suited for texture
data.. Persistence image (PI) is based on concepts known from
the theory of size functions [33, 34]. The method scales well
to large training data. In contrast to PL, PI lives in Euclidean
space, which makes is directly compatible to a broad set of machine learning techniques.
The PI is derived by mapping a PD to an integrable function
G p : R2 → R, which is a sum of Gaussian functions centered at
each point of the PD. Taking a discretization of a subdomain of
G p defines a grid. An image can be created by computing the
integral of G p on each grid box, thus defining a matrix of pixel
values. Formally, the value of each pixel p = (x, y) within a PI
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
5
20
20
10
20
40
40
40
20
60
60
60
80
80
80
100
100
100
30
10
40
120
15
40
10
15
20
(a)
40
10
10
15
15
10
15
(c)
60
80 100 120
20
40
60
80 100 120
(c)
5
5
5
10
10
10
15
5
10
15
(d)
5
10
15
15
5
10
(e)
15
5
10
15
(f)
Figure 5: The patch from Fig. 3a preprocessed with: (a) Schmid filter of size =
49, σ = 6 and τ = 1; (b) MR filter of size = 49, σ x = 2 and σy = 6; (c) CLBP
of type riu2, n = 8 and r = 3; together with their PIs (d-f).
(d)
is defined by the following equation:
− 12
1
e
g(bi , ei )
2πσ x σy
[b ,e ]∈I
X
i
40
(b)
15
Figure 4: PIs for patch from Fig. 3a computed with different parameter values:
(a) original PI from Fig. 3b; (b) resolution 64 × 64 pixels instead of 16 × 16; (c)
σ x = σy = 0.01 instead of 0.001; (d) no weighting function.
p
120
20
60
5
ZZ
80 100 120
(b)
5
5
60
(a)
60
5
PI(p) =
120
20
50
(x−bi ) (y−ei )
+ 2
σ2x
σy
!
dy dx, (1)
may be applied before topological analysis as frequently observed in related work [31].
The patch from Fig. 3a processed with Schmid filters, MR
filters, and with CLBP is presented in Fig. 5, together with the
respective PIs. It can be observed that different pre-processing
leads to strongly different representations. Again, it is not clear
which representation is most robust and beneficial with respect
to classification accuracy.
i
3.5. Normalization
where g(bi , ei ) is a weighting function, which depends on the
distance from the diagonal. Points close to the diagonal are
usually considered as noise and therefore get lower weights by
the function g. Variables σ x and σy are the standard deviations of the Gaussians in x and y direction. The resulting image
(see Fig. 3c) is vectorized to achieve a standardized vectorial
representation which is compatible to a broad range of machine
learning techniques. The pixels under the diagonal are not taken
into account in case of PI (they all are zeros).
The advantage of PIs compared to aggregated PD descriptors is that the spatial information from the PD are preserved,
which may lead to higher classification accuracies [32]. However, the computation of PI requires numerous parameters like
the weighting function g, σ x and σy of the employed Gaussians
and the final image resolution. Changes in values of the parameters may result in strongly different PIs, which is illustrated in
Fig. 4. It is a priori not clear which parameters yield the best
representation for a specific task and how robust PI is to these
parameters in practice.
3.4. Pre-filtering
Different types of pre-processing can be applied to the input data before computing topological descriptors, in order to
extract more complex topological information. We take into
consideration the Schmid filter bank [40] and the Maximum
Response (MR) filter bank [41], as they are very popular and
rotation invariant. Moreover, robust texture descriptors such
as Local Binary Patterns (LBP) or Complete LBP (CLBP)[42]
Especially when working with depth maps normalization
plays an important role for persistent homology analysis. The
input depth maps have different depth value ranges depending
on the characteristics of the 3D surfaces (e.g. deep valleys and
high hills versus flat surfaces). Normalization can be performed
at a global or local level.
Global normalization:. To compute comparable topological descriptors the depth maps need to be mapped to a similar value
range first, i.e. the distribution of the depth values needs to be
standardized. We apply z-standardization to the depth values of
each depth map separately. As a result all depth maps have zero
mean and unit variance and thus cover a similar value range
(except for outliers).
Local normalization:. Aside from standardizing the value distribution of the depth maps additional normalization can be applied locally across the individual patches of the surface. Different types of normalization add different invariants to the resulting feature representations which may result in more robust
classification. By performing a patch-based normalization we
remove the global reference between the patches, i.e. each
patch’s value distribution is normalized individually, independently from the neighboring patches. Thus, after normalization
the absolute depth reference is removed. What remains is pure
topological surface information. This means that two patches
with similar topological structure at different depth levels of
the surface become similar and are represented by similar PDs.
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
5
20
5
20
40
40
60
0
80
60
0
80
100
100
120
20 40 60 80 100 120
-5
120
20 40 60 80 100 120
(a)
-5
(b)
5
5
10
10
15
15
5
10
15
5
(c)
10
15
(d)
5
20
20
40
40
60
60
80
80
100
100
120
0
120
20
40
60
80 100 120
20 40 60 80 100 120
(e)
-5
(f)
5
5
10
10
15
15
5
10
(g)
a measurement function (in our case represented by a patch)
changes sligthly, then its persistence diagram do not change too
much as well. There are two major metrics in PDs space: bottleneck distance (dB ) and Wasserstein distance (dW ). CohenSteiner et al. prove in [43, 44] results about stability of PD
under dB and dW metrics for classes of the measurement functions. Especially, by [44, Wasserstein Stability Theorem] we
know that the computation of persistence diagrams for Lipschitz functions is stable under dW .
In the case of depth map it is easy to observe that the measurement function is Lipschitz (the size and pixel values are
bounded). In [32] it is proven that PIs are stable under dW , as
well. Hence, we consider the process of obtaining PIs from
surface patches being stable in theory. In our experiments we
analyze this the stability of the resulting descriptors in practice.
15
5
10
15
(h)
Figure 6: The effect of local normalization: two patches (a) and (b) which
are located at different depth levels of the surface together with their PIs (c)
and (d). The two patches become more similar after applying local min-max
normalization (e) and (f). A similar effect is observed for the corresponding PIs
(g) and (h).
Without patch-based normalization the absolute depth information is retained and the two patches yield different (shifted) PDs,
see Fig. 6 for an example.
Given a patch P, different types of local normalization can
be applied, such as:
• z-standardization: P = (P − mean(P))/std(P),
• min-max normalization: P = (P − min(P))/(max(P) −
min(P)) and
• positional standardization: P = (P−median(P))/mad(P),
where std refers to the standard deviation and mad to the median absolute deviation.
3.6. Stability
A crucial property of persistent homology is the stability of
persistence diagrams. Briefly speaking, we expect that when
4. Datasets and evaluation protocol
For our experiments, we employ a recently released dataset
of high-resolution 3D reconstructions from the archaeological
domain with a resolution of approximately 0.1 mm [11]. The
dimensions of the scanned surfaces ranges from approx. 20×30
cm to 30 × 50 cm. The reconstructions represent natural rock
surfaces that exhibit human-made engravings (so-called rock
art). The engravings represent animals, symbols and figures
that have been engraved by humans in ancient times. See Fig. 1
for an example surface. The engraved regions in the surface
exhibit a different surface texture than the surrounding natural
rock surface. In our experiments we aim at automatically separating the engraved areas from the natural rock surface.
For each surface a precise ground truth has been generated
by domain experts that labels all engravings on the surface.
The dataset contains two classes of surface topographies: class
1 represents engraved areas and class 2 represents the natural
rock surface. Class priors are imbalanced. Class 1 represents
approximately 19% of the data and is thus underrepresented. A
corresponding ground truth labeling is depicted in Fig. 1c.
In total 26 high-resolution surface reconstructions have been
acquired as described above. To accelerate experiments, we
form two datasets of different sizes:
Small-scale dataset: . The small scale dataset contains depth
maps from 4 surface reconstructions with a total number of 12.3
millions of points. This dataset has also been employed for the
experiments in [8] and thus represents a baseline dataset. We
employ this dataset for our initial experiments to reduce computation time and to enable a broader set of experiments in our
evaluation. Furthermore, by using this dataset our results become comparable to previous work. In our evaluation the depth
maps of the first two surface reconstructions are employed for
training and the remaining two surfaces represent the independent evaluation set.
Large-scale dataset: . This dataset contains depth maps of all
the 26 surface reconstructions that have been acquired. This
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
0
3
0
3
2
1
64
128
0
64
-2
-2
(a) ideal natural
0
128
64
128
0
3
2
1
1
64
0
-1
-1
-2
-2
-3
64
-3
0
2
0
0
128
(b) ideal engraved I
3
64
0
-1
128
128
(d) natural+noise
128
-3
0
64
128
(e) engraved I+noise
10000
10
8000
20
2
1
64
0
-1
-2
128
-3
0
64
128
(c) ideal engraved II
0
3
2
1
64
0
10
8000
-2
128
-3
0
64
128
(f) engraved II+noise
10000
10
8000
20
6000
30
6000
30
40
4000
40
4000
40
4000
50
2000
50
2000
50
2000
60
60
60
0
20
40
60
(g) PI natural
0
20
40
60
(h) PI engraved I
Evaluation Protocol. Our evaluation protocol is defined as follows: from the input depth maps, we extract overlapping square
image patches which are input to feature extraction (either for
topological features or non-topological features). The resulting
feature vectors for each patch are then input to classification.
The classification performance is measured by the Dice Similarity Coefficient (DSC). DSC measures the matching between
the ground truth and the predicted class labels, i.e. the mutual
overlap between an automatic labeling X of a depth map and a
manual (ground truth) labeling Y:
-1
10000
20
6000
30
3
1
64
-1
-3
0
0
2
0
20
40
60
(i) PI engraved II
Figure 7: Synthetic dataset: “natural” and “engraved” surfaces without noise
(a-c); the same surfaces with noise (d-f); and their PIs (g-i). PI generated for
the “natural” surface is unimodal, while PIs generated for “engraved” surfaces
are bimodal and less coherent.
dataset is employed in our evaluation to demonstrate the generalization ability of the proposed approach and to verify if conclusions drawn from the small-scale dataset can also be transferred for the larger and more complex one. For classification
experiments we randomly select 13 surfaces as a training set
and employ the remaining ones for evaluation.
We further perform experiments on synthetic data to provide an experimental proof-of-concept for our topological approach.
Synthetic dataset: . The synthetic dataset serves as a proofof-concept to simulate the process of engraving a natural rock
surface and the effect of the resulting surface topographies on
the topological descriptors. We start with an ideally flat surface
that represents a “natural” rock surface without any irregularities, see Fig. 7a. All values of this surface are constant and
equal to zero. In Figs. 7b and c we simulate differently dense
and coarse engravings by adding negative Gaussian distributions to the flat surface. The spacing between Gaussians equals
20 ± 5 and 30 ± 5 pixels. In a next step, we add noise to all
surfaces to simulate natural irregularities and effects of longtime exposition of the surfaces to weather. Noise is generated
using the surface generator of [45]2 . Fig. 7d shows the resulting noise, i.e. the simulated flat natural surface. We further add
the same noise to the simulated engraved surfaces from Fig. 7b
and c, see Figs. 7e and f. The depth values of all surfaces are
z-standardized.
DSC(X, Y) =
2|X ∩ Y|
.
|X| + |Y|
DSC is between 0 and 1, where 1 means a perfect classification.
For each classification experiment we first select a subset S i
of patches from the training set T . Next, we estimate suitable
classifier parameters by 5-fold cross-validation on the subset S i .
With the optimized parameters we train a classifier on the entire
subset S i . This classifier is finally tested on the evaluation set.
There is no parameter tuning on the evaluation set.
Each classification experiment is repeated 10 times for different randomly selected subsets S i from the training set T . In
each repetition other patches from the training set T are selected
and thus the actual training data varies over the 10 repetitions.
The result are 10 classification results (DSC values) for each experiment. We report the mean and the standard deviation over
all 10 repetitions. By conducting each classification experiment
with different subsets of the training set we reduce the dependency of our results on the training data.
As mentioned above, the class priors in our dataset are imbalanced. Skewed datasets pose problems to most classification techniques and often yield suboptimal models as one class
dominates the other classes. A classifier especially designed
for imbalanced datasets is Random Undersampling Boosting
(RUSBoost) [46]. RusBoost builds upon AdaBoost [47] which
is an ensemble method that combines the weighted decisions
of weak classifiers to obtain a final decision for a given input
patch. RUSBoost extends this concept by a data sampling strategy that enforces similar class priors. During each training iteration the majority class in the training set is undersampled
in a random fashion to balance the resulting class priors. In
this manner, the weak classifiers can be learned from balanced
datasets without being biased from the skewed class distribution. In previous experiments RUSBoost has already achieved
robust results and high classification performance and has outperformed among others cost-sensitive SVMs [48] with linear
and RBF kernels which are also suitable for imbalanced data.
The code used in our study was implemented in Matlab.
Most of the descriptors were extracted with VLFeat library [49],
except for PD AGG and PI which were computed with the CAPD::
RedHom library [50, 51] with the PHAT [52, 53] algorithm for
persistence homology.
5. Experiments and results
2 implementation
available from:
http://www.mysimlabs.com/
surface_generation.html, last visited June 2017
We first investigate the expressive power of topological descriptors for our synthetic data set and then perform a compre-
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
hensive study on the robustness and expressiveness of the topological descriptors for the analysis of real-world 3D surfaces.
We investigate different variants of their computation and compare them with traditional non-topological descriptors. Finally,
we investigate the effect of combining topological with nontopological descriptors. For each set of experiments we first
describe the experimental setup and the investigated aspects in
the context of the experiment and then report on the obtained
results. Unless otherwise stated the experiments on the 3D reconstructions were performed on the small-scale dataset.
5.1. Experiments on synthetic data
We generated PI descriptors for the synthetic surfaces using the baseline configuration (see Tab. 1). The PI generated
for the natural surface in Fig. 7g shows a bright cluster in the
center of PI. This is plausible, because most of the holes are
born and die at the same time. The PIs of the two engraved
surfaces in Figs. 7h and i exhibit a completely different spatial
distribution. The reason is that there are two types of holes on
the surfaces. One group of holes corresponds to the natural irregularities (noise) and the other one to the engravings. Holes
from both groups are born and die at different times. Thus the
PIs become bimodal. This is strongly different from the natural
surface for which PI is unimodal and more coherent.
These experiments demonstrate that the topological descriptors capture different surface topographies well. In the following sections we analyze real-world 3D surfaces and perform
a detailed analysis of topological and non-topological descriptors.
5.2. Topological baseline descriptors
In an initial experiment we extract topological descriptors
which serve as a baselines for later comparison. From the depth
maps in the training and evaluation sets we extract topological descriptors in a patch-based manner. The patch size is
128 × 128 pixels (i.e. 10.8 × 10.8 mm) and the step size between
two patches is 16 pixels (1.35 mm). The baseline descriptors include: (i) aggregated descriptors from the persistence diagram
(PD AGG), see Section 3.2 and (ii) descriptors obtained from
the persistence image (PI), see Section 3.3.
For both descriptors we select those parameters for which
they have yielded best results in our previous investigation [9].
See Tab. 1 for the parameter selections for the baseline configurations and an overview of all possible parameter values.
PD AGG results in an 12-dimensional feature vector (see
Section 3.2), while the feature vector derived from PI has a dimension D of 136 which originates from the fact that all values under the diagonal are zero and are skipped. Thus, D =
(R2 + R)/2 where R refers to the resolution of the PI (R = 16 for
the baseline configuration).
As already mentioned in Section 3.5 the input depth maps
have different depth value ranges and are z-standardized to make
different surfaces comparable. After z-standardization the depth
maps have zero mean and unit variance and thus cover a similar value range (except for outliers). Nevertheless we unify
the PDs by defining a parameter “diagram limits” as an interval
Descriptor
PD AGG
PI
Parameter
Local normalization
Diagram limits
Outlier removal
Mapping
Pre-filtering
Local normalization
Diagram limits
Sigma
Resolution
Weighting
Outlier removal
Mapping
Pre-filtering
Possible values
{none, z-std, p-std, minmax}
[min, max]
{yes, no}
{none, CLBP}
{none, Schmid, MR}
{none, z-std, p-std, minmax}
[min, max]
R+
N+
{none, linear, exponential}
{yes, no}
{none, CLBP}
{none, Schmid, MR}
Baseline configuration
none
[-5, 5]
no
none
none
none
[-5, 5]
0.001
16
none
no
none
none
Table 1: Topological descriptors, their computation parameters and the specification of the baseline configuration. The abbreviations “z-std” stands for zstandardization, “p-std” for positional standardization, and “minmax” for minmax normalization.
Descriptor
PD AGG
PI
DSC
σ = 0.00025
σ = 0.0005
σ = 0.001
σ = 0.002
R=8
0.701 ± 0.003
0.712 ± 0.005
0.722 ± 0.006
0.723 ± 0.002
R = 16
0.729 ± 0.003
0.730 ± 0.002
0.733 ± 0.003
0.728 ± 0.003
R = 32
0.732 ± 0.002
0.731 ± 0.004
0.734 ± 0.004
0.730 ± 0.003
0.655 ± 0.012
R = 64
0.725 ± 0.007
0.730 ± 0.003
0.735 ± 0.004
0.730 ± 0.003
Table 2: DSC obtained for PD AGG and PI with various resolutions and sigmas. Numbers in bold refer to the baseline descriptors.
[min, max]. Both values [min, max] define the axis limits of the
underlying PD. We set these values to [−5, 5] which can safely
be done as after z-standardization the depth values most likely
lie in the interval [−5, 5] except for individual outliers. Note
that if patch-based normalization is used (see Section 3.5), the
diagram limits need to be adapted according to the type of normalization which may change the depth value range, e.g. in the
case of min-max normalization depth values are moved to the
interval [0, 1].
The baseline configuration of PD AGG yields a DSC of
0.655 ± 0.012. The baseline for PI outperforms PD AGG with a
DSC of 0.733±0.003. A reason for the better performance of PI
is that PI preserves spatial information from the PD as already
mentioned in Section 3.3 while PD AGG captures only overall
statistics. We repeat each experiment 10 times with different
training data (see Section 4 for details). The standard deviation
across all the 10 experiments is low (0.012 and 0.003) which
shows that the dependency on the training data is low as well
for both descriptors.
5.3. PI with different resolutions and sigmas
To investigate the sensitivity of resolution and sigma for the
computation of PI, we extract PIs for different resolutions (8,
16, 32, 64 pixels) and standard deviations (0.00025, 0.0005,
0.001, 0.002). The remaining parameters of the baseline configuration remain constant.
The results are presented in Tab. 2. The performance of the
baseline descriptors is highlighted bold. Results show that the
sensitivity of PIs to both parameters is low. With increasing resolution the performance slightly improves (especially by going
from 8x8 to 16x16 pixels). Higher resolutions do not provide
significant improvements. Since the computation time grows
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
quadratically with resolution it must be stressed that it is desirable to keep the resolution as small as possible.
Experiments for different sigmas show that the best results
are obtained with σ = 0.001 independently from the resolution. Thus both parameters can be optimized independently
from each other which saves computation time in parameter
tuning and model selection in practice.
5.4. Different normalization variants
We perform experiments with different normalization variants with PI and PD AGG and compare them to the baseline
feature configurations. Variants include:
0.75
0.70
0.65
0.60
1. No normalization at all: we do not provide diagram limits for the computation of the PD. Thus each PD adapts
to the value range of the underlying image patch and thus
PD AGG and PI adapt completely to the value range.
2. Globally defined diagram limits: for each input patch the
same diagram limits for PD are applied. This corresponds
to the baseline features.
3. Patch-based normalization with z-standardization: each
input patch is z-standardized individually before feature
extraction.
4. Patch-based normalization with positional standardization: same as above with positional standardization.
5. Patch-based normalization with min-max normalization:
same as above with minmax normalization.
The first variant demonstrates what happens if we apply the
topological features directly on the data without any adaptions
and is expected to represent a lower bound of performance. In
variant 2 the depth relations (i.e. the absolute depth differences)
between patches are retained. In variants 3-5 depth relations
between patches are removed and the only the pure topological
information inside each patch is retained.
Results for the different normalization variants are presented
in Fig. 8. The best accuracy is obtained with globally defined
diagram limits, which preserve the depth relations between the
patches. Applying no normalization at all is suboptimal for PI.
The three patch-based normalization variants all yield lower results. This shows that the depth relations between patches are
important for the classification task.
Interestingly, positional standardization which is robust to
outliers in the depth distribution leads to the weakest results.
This indicates that the outliers in the distribution are of high
importance for the task.
Due to the different normalizations the above descriptors
(especially variant 2 and the patch-based normalization variants) may represent complementary information that may be
beneficial in combination. To investigate this further, we evaluate different combinations of PD AGG and PI with different
normalization types. We select patch-based z-standardization
in the following experiments. Finally, we investigate the effect
of combining PI with PD AGG (with similar and different normalizations). For the combination of features we concatenate
the respective feature vectors before feeding them into the classifier (early fusion). Results are summarized in Tab. 3.
0.55
0.50
no
normaliza!on
global
limits
patch-based
(z-std)
PD_AGG
patch-based
(p-std)
patch-based
(min-max)
PI
Figure 8: The effect of different normalization methods for PI (yellow line,
square markers) and PD AGG (orange line with circular markers). Patch-based
normalizations (z-std, p-std, and min-max) are outperformed by global normalization. Applying no normalization at all (first experiment) is not beneficial for
3D surface analysis.
Descriptor(s)
PD AGG
PD AGG
PI
PI
PI+PD AGG
PI+PD AGG
PI+PD AGG
Normalization
Global limits (baseline)
Global limits + z-std
Global limits (baseline)
Global limits + z-std
Global limits (baselines)
z-std
Global limits + z-std
DSC
0.655 ± 0.012
0.671 ± 0.008*
0.733 ± 0.002
0.735 ± 0.004
0.730 ± 0.003
0.723 ± 0.005
0.735 ± 0.002*
Table 3: DSC for combining PI with PD AGG (with the same and different
normalization types). Asterisks (*) correspond to results significantly better
than corresponding baseline topological descriptors with p < 0.05.
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
Death time
5
5
5
10
10
0
-5
15
-5
0
Birth time
(a)
5
15
5
10
(b)
15
5
10
15
(c)
Figure 9: The effect of outliers on the PI. A PD containing 4 sample points
[−2, −1.5], [−1, −0.5], [0, 0.5] and [3, 6] (a) and the corresponding PIs generated without (b) and with (c) outlier removal, with diagram limits [−5, 5].
Although the point [3, 6] lies outside of the domain of the PI (and is thus considered an outlier), the Gaussian centered around it still extends into the domain
of the PI when outliers are not explicitly removed.
The only result in Tab. 3 where the combination of different
normalized descriptors improves results is for PD AGG (DSC
increases to 0.671 ± 0.008 with p < 0.053 ). In case of PI only
a small improvement by 0.002 is observed which is not significant. Interestingly, the combination of PI with PD AGG does
not improve results over the baseline of PI alone. A reason for
this may be that PI implicitly captures the information that is
represented in PD AGG.
5.5. Outlier removal
Outliers are usually considered as unwanted artifacts. In
persistence homology, however the role of outliers is different.
Outliers in the PD, i.e. points with a large distance to the diagonal, are usually considered as the most meaningful and strong
topological components of the underlying data.
In our approach outliers are those points in the PD which
extend the pre-defined diagram limits. There are two alternatives: to keep them or to remove these outliers completely
from the further computation. For the two investigated topological descriptors this has different consequences. In the case of
PD AGG the outliers are simply included in the derived statistics, which is the same as extending the diagram limits to the
farthest outliers. For PI this means that although the outliers lie
outside of the domain of the PI the Gaussian centered around an
outlier during PI construction may still extend into the domain
of the PI (see Fig. 9c).
We compare the effect of outlier removal by computing respective variants of our baseline features. The results show that
removing outliers decreases results for PI from 0.733 ± 0.003 to
0.731 ± 0.003 and for PD AGG from 0.655 ± 0.012 to 0.639 ±
0.011. In case of PD AGG the decrease is significant with
p < 0.05). This indicates that outliers play an important role
for the topological descriptors.
5.6. Weighting in PI computation
The original formulation of PI includes a weighting function for the individual Gaussian components which grows exponentially with distance to the diagonal of the PD (function g
in Equation 1 in Section 3.3). The intuition behind this is to
3 Statistical significance is computed with the Wilcox signed rank test, as
most of the samples do not pass the Shapiro-Wilk normality test.
emphasize stronger topological components in the PD (further
away from the diagonal) and to attenuate the influence of noise
which is often located close to the diagonal. Previous experiments have shown that the weighting is in practice not always
beneficial [9]. To further investigate the role of weighting in
the PI computation, we compare the performance of PI with
different weighting functions namely: constant weighting, linear weighting and exponential weighting.
Experiments reveal that weighting significantly decreases
performance with p < 0.005: from 0.733 ± 0.003 to 0.714 ±
0.008 in case of linear weighting and to 0.715 ± 0.007 in case
of exponential weighting. Best results are obtained when all
Gaussians are equally weighted. This indicates that short intervals close to the diagonal play an important role in the context
of surface texture description. This contradicts with the original assumptions made in the design of PI. We investigate this
aspect further in Section 5.13.
5.7. Pre-filtering
We investigate different ways to pre-filter the surface patches
before extracting topological features. This has recently been
proposed to obtain more robust and expressive representations [30,
31]. We employ three different methods for pre-filtering (as described in Section 3.4): (i) Schmid filter bank [40], (ii) MR filter
bank [41] and (iii) CLBP encoding [42].
For MR and Schmid filter banks we employ the default parameters proposed by the respective authors. The only exception is that in case of Schmid filter banks we use two times
smaller τ, as in preliminary experiments there is hardly any difference in the filter responses for τ > 2. Each patch is filtered
with all filters from the respective filter bank. For each filtered
patch topological features are extracted separately and are then
concatenated into one high-dimensional feature vector.
Additionally, we extract CLBP features [42] from the raw
patches as proposed in [30] and [31] and then extract topological descriptors from the CLBP S and CLBP M maps. For
CLBP computation we employ the implementation of [54] with
different encodings: rotation invariant LBP (ri) and rotation
invariant uniform LBP (riu2) as well as different numbers of
neighborhood samples n = {8, 16} and different radii r = {3, 5}.
The results of experiments are presented in Tab. 4. It can be
observed that prefiltering with Schmid and MR filter banks does
not improve performance, both for PD AGG and PI. CLBP,
however, improves DSC significantly in case of PD AGG (from
0.655 ± 0.012 to 0.708 ± 0.005, p < 0.01 and CLBP with “riu2”
encoding, r = 5 and n = 8). The effect of pre-filtering is further
analyzed in Section 5.12. From the experiments so far it seems
that for PI the best strategy is to apply it on the raw data directly, while for PD AGG which are much simpler and coarser
descriptors the additional pre-filtering can be beneficial.
Finally, we combine our baseline features with the respective Schmid, MR and CLBP filtered variants by early fusion
(concatenation) to investigate if the descriptors can benefit from
each other. Experiments show that most of the combinations decrease performance (see Fig. 10), which is consistent with the
pre-filtering results presented above. The only positive effect in
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
Descriptor
PD AGG
PD AGG
PD AGG
Pre-filtering
(baseline)
Schmid
MR
PD AGG
CLBP
CLBP
CLBP
CLBP
(baseline)
Schmid
MR
PI
PI
PI
PI
CLBP
CLBP
CLBP
CLBP
DSC
riu2
ri
riu2
ri
r
r
r
r
=3
=5
=3
=5
n=8
0.671 ± 0.004**
0.708 ± 0.005**
0.697 ± 0.006**
0.700 ± 0.006**
r
r
r
r
=3
=5
=3
=5
n=8
0.712 ± 0.014
0.707 ± 0.014
0.709 ± 0.009
0.706 ± 0.006
0.655 ± 0.012
0.643 ± 0.023
0.607 ± 0.044
n = 16
0.674 ± 0.009**
0.678 ± 0.008**
0.670 ± 0.019*
0.687 ± 0.006**
0.733 ± 0.002
0.702 ± 0.007
0.698 ± 0.004
n = 16
0.725 ± 0.004
0.719 ± 0.006
0.722 ± 0.007
0.706 ± 0.005
Table 4: Results for pre-filtering depth maps before extraction of topological features. Asterisks (*) and (**) correspond to results significantly better than corresponding baseline topological descriptor with p < 0.05 and p < 0.01, respectively. Bold represent the best results for PD AGG and PI.
strategies:
0.76
0.73
0.70
0.67
0.64
0.61
0.58
0.55
Baseline
Baseline + Baseline
with Schmid
PD_AGG
Baseline + Baseline
with MR
Baseline + Baseline
with CLBP
PI
Figure 10: DSC for combining baseline topological descriptors with the respective Schmid, MR and CLBP filtered variants. For PI the combination is
not able to further improve results. PD AGG benefits from adding the CLBPfiltered variant but not significantly.
performance is achieved by the combination of PD AGG with
the respective CLBP filter variant. This is also consistent with
the previous results where CLBP improved PD AGG. The result of the combined descriptors is, however, weaker than that
of PD AGG based on CLBP alone (0.674 vs 0.708). This shows
that there is no synergy attained by the combination.
5.8. Feature selection
By applying pre-filtering (as in the previous Section 5.7)
or by using PIs with high resolutions the dimension of the resulting feature vectors grows linearly and quadratically, respectively. Thus, strategies to reduce the dimension may be useful
in practice. We apply different feature selection strategies to
the PI descriptors to investigate to which degree dimensionality
reduction affects classification performance and how much redundancy the descriptors exhibit. We apply the following three
1. Gini importance: we select Gini importance [55] as a
strategy because it is a by-product of the employed classifier (RUSBOOST, see Section 4) and it can be obtained
without additional cost. It directly reflects the importance
of each descriptor entry for distinguishing between the
classes.
Computation: we compute the Gini importance for each
entry of the PI descriptor during training of the classifier and keep only the N most important (highest values)
entries.
2. Fisher criterion: While Gini takes dependencies between
descriptor entries into account (through the classifier tree
hierarchy), the Fisher criterion [56] is completely independent from the classifier and evaluates the importance
of every entry independently. For this reason we decided
to use is as a complementary strategy.
Computation: we compute the Fisher criterion for each
descriptor entry during training of the classifier and keep
only the N most discriminative entries (highest Fisher criterion)
3. Combined Gini and Fisher criterion: we combine both
strategies to see if there are synergy effects between them.
Computation: for each descriptor entry we compute Gini
importance and Fisher criterion and average the rankings
obtained by both measures. The N entries with the highest combined ranking are retained.
Additionally, to provide a baseline, we select features randomly to investigate the actual benefit of feature selection strategies.
For all three feature selection approaches we vary N from 2,
4, 8, 16, 32, 64 to 100% of the PI descriptor (136 dimensions).
Results are illustrated in Fig. 11. Note, that the x-axis in Fig. 11
is not linear. By selecting only 2% of the components, Gini
yields a DSC of 0.6015 ± 0.0197. With the Fisher criterion even
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
of EGHS+ESDD (DSC of 0, 728 ± 0.005). The performance
of our two baseline descriptors are shown with a dotted line
(PD AGG) and a dashed line (PI) in black for comparison. The
PD AGG descriptor achieves similar results to LBP, EH, GLCM,
and EGHS. Remarkably, PI alone already outperforms all nontopological features except ESDD. This attests a strong explanatory power of the PI for the description of surface texture while
PD AGG rather lacks in expressiveness.
0,74
0,72
0,70
0,68
0,66
0,64
0,62
0,60
0,58
0,56
0,54
0,52
0,50
5.10. Combination with non-topological features
2%
4%
Gini
8%
Fisher
16%
32%
Combina!on
64%
100%
Random
Figure 11: Classification performance (DSC) with different feature selection
strategies for differently large subsets of the baseline PI descriptor.
a DSC of 0.6560 ± 0.0213 is achieved. Both values strongly
outperform the random baseline (0.535 ± 0.06). With increasing number of features the performance increases continuously
for Gini. The Fisher criterion is less stable and is outperformed
in most cases by Gini as can be seen in Fig. 11. A reason for
the better performance of Gini is that Gini is computed during classifier training and expresses the importance of a given
entry for the classifier training. Fisher, in contrast, is computed independently from the classifier from the training data
directly. The combined scheme is located approximately between the two approaches which may stem from averaging the
respective rankings of the individual measures. Interestingly,
feature selection is only reasonable up to a certain number of
features. By selection of more than 16% of features the random
baseline outperforms the selection strategies. The reason for
this behavior is that the feature selection strategies do not take
the inter-correlations between the entries into account and thus
redundant components are easily selected when the number of
selected features increases. We further investigate the importance and discriminativity of individual entries in Section 5.13.
5.9. Comparison with non-topological features
From the patches of the depth map we extract a number of
non-topological image descriptors in the same patch-wise manner as the topological descriptors. Non-topological descriptors include: MPEG-7 Edge Histogram (EH) [57], Dense SIFT
(DSIFT) [58], Local Binary Patterns (LBP) [59], Histogram
of Oriented Gradients (HOG) [60], Gray-Level Co-occurrence
Matrix (GLCM) [61], Global Histogram Shape (GHS) [8], Spatial Depth Distribution (SDD) [8], as well as enhanced versions
of GHS and SDD (short EGHS and ESDD) that apply additional enhancements to the depth map described in [8].
For the extraction of non-topological features, we employ
publicly available implementations if available, such as VLFEAT
for SIFT and HOG[62], the LBP implementation of [63], and
the GLCM implementation of [64]. GHS, SDD, EGHS, and
ESDD are implemented as in [8].
Results for the individual descriptors are shown in Fig. 12
(blue line). The best individual non-topological descriptor is
ESDD (DSC of 0.743 ± 0.002) followed by the combination
In previous investigations the addition of topological descriptors to other (non-topological) descriptors has improved
classification performance [9]. We investigate the effect of combining both types of descriptors by adding the best configurations of PI and PD AGG to the evaluated non-topological descriptors from Section 5.9. The results are shown as additional
curves in Fig. 12. The orange curve (square markers) shows the
performance when non-topological descriptors are combined
with PD AGG and the yellow curve (diamond-shaped markers)
shows the performance in combination with PI. The combination with PD AGG improves all descriptors except for EH. PI
improves all descriptors and some in fact by a large margin.
Even the performance of the best non-topological descriptor
(ESDD) can be further improved when combined with PI (from
0.743 ± 0.002 to 0.788 ± 0.005 with p < 0.01). A corresponding example result is shown in Figure 13. The surface shows
a horse with a rider that holds a long spear in its hand. The
first row shows the original rock-surface, the depth map, and
the ground-truth labeling (engraved areas in red, remaining areas uncolored). The second row shows the results of ESDD, PI
and ESDD+PI. Classification results are color coded: red refers
to true positive detections of class 1 (engraved area), green to
false positive detections of class 1 and blue to false negative
detections of class 1. Uncolored areas are true positive detections of class 2. The results for ESDD and PI are comparable
in overall quality but show different strengths and weaknesses
of the descriptors. ESDD better captures fine details than PI
(e.g. the rear legs) but produces more false detections (green)
at natural surface irregularities like the vertical and diagonal
cracks in the rock surface. PI is more robust to such irregularities and captures the shape of the engraved ares well but generates more false detections along the boundary of the engraved
areas (e.g. along the spear and the horse’s tail). In combination
the mutual weaknesses compensate each other (e.g. the shaft
of the spear and the arm of the rider are better detected) and
yield a more accurate overall segmentation. These results illustrate well that topological descriptors represent complementary
information about the surface texture compared to traditional
texture descriptors and that their combination is beneficial.
5.11. Large-scale experiments
Experiments so far were performed on the small-scale dataset
which was chosen to save computation time and in turn gave us
the opportunity to broaden our evaluation to many different aspects, parameter values, configurations etc. A central question
in our evaluation is how topological descriptors compete with
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
0.75
0.65
0.55
0.45
0.35
0.25
non-topological feature
with PD_AGG
with PI
PI Baseline
PD_AGG Baseline
Figure 12: Classification results (DSC) for different non-topological descriptors (blue line with circular markers). The orange (square marker) and yellow
(diamond marker) lines represent the performance when the features are combined with PD AGG and PI, respectively. The baselines for PD AGG and PI
are plotted as dotted and dashed lines, respectively.
on the same dataset. In contrast to the investigation in [11]
which uses 4-fold cross-validation on the entire dataset, we employ 5-fold cross-validation on the training set and additionally
use half of the dataset as a completely independent test set. The
use of an independent test set makes our evaluation protocol
is more demanding that that of [11]. Reference [11] reports
results for two approaches: Random Forests (RF) and Convolutional Neural Networks (CNNs). For the CNN the fully convolutional net by [65] has been fine-tuned on randomly selected
patches from the training data, see [11] for details. RF yields
a DSC of 0.568. CNN outperforms RF with a DSC of 0.667.
With a DSC of 0.654 our best topological feature (PI) yields a
similar performance level as CNNs in [11]. PI combined with
non-topological descriptors even outperforms the results of the
CNN with a peak performance of 0.681. Thus, the performance
of the investigated descriptors can be considered state-of-theart.
5.12. Robustness and sensitivity
non-topological descriptors which currently represent the stateof-the-art for 3D surface and texture analysis [11]. To answer
this question more profoundly, we run the experiments from
Sections 5.9 and 5.10 on the large-scale dataset from Section 4.
We evaluate all descriptors without further parameter tuning to
obtain objective and unbiased results, i.e. no fine tuning is performed on the large-scale dataset at all. The topological descriptors are then compared to a selection of the most promising non-topological descriptors. Furthermore, we evaluate the
performance of the best descriptor combinations evaluated so
far.
Results for the large-scale dataset (together with the corresponding results for the small-scale dataset) are presented in
Fig. 14. The overall performance of all descriptors drops for
the large-scale dataset. This is expected and results from the
higher complexity of the dataset (more heterogeneous data).
The absolute performance difference between the two datasets
is, however, quite consistent between all descriptors (∆ = 0.082
with a quite low standard deviation of ±0.014), which shows
that the classifier can generalize well from all descriptors. The
best topological descriptor is again PI with 0.654 ± 0.002 which
is similar to the best non-topological descriptor (ESDD) with
0.661 ± 0.001 (dashed line). The smaller difference in peak performance between topological and non-topological descriptors
compared to the small-scale dataset shows that the topological
descriptors better generalize to the more complex data in the
large dataset.
The combination of topological and non-topological descriptors yields a performance gain in most cases. Peak performance is obtained when all evaluated descriptors are combined:
0.681 ± 0.002. This improvement is significant with p < 0.01
compared to the best combination of non-topological descriptors. The two horizontal lines in Fig. 14 represent the peak performance without and with topological information. The vertical spacing between the lines represents the gain obtained by
adding topological information in classification (+0.02). Note
that the gain for the small-scale dataset is even larger (+0.045).
We further compare our results with those obtained in [11]
A crucial property of persistent homology is stability. In
the following experiments we evaluate if the theoretical conclusions about stability from Section 3.6 hold in practice. For
this purpose, we randomly select number of patches and add
different levels of random Gaussian noise (see Fig. 15a-c for an
example). Next, we compute the differences between the raw
patches and their noisy variants as well as the differences between the PIs computed from the raw and noisy patches. The
normalized difference d between two patches is computed as:
P
p∈P |P(p) − N(p)|
,
d(P, N) = P
p∈P |P(p)| + |N(p)|
where N and P is the original patch with and without random
Gaussian noise. The normalized difference between two PIs is
computed analogously to that of patches by replacing N and P
with the corresponding PIs. In Fig. 16 we plot the differences
between the patches (x-axis) vs. the differences between the PIs
(y-axis). Color encodes the signal-to-noise ratio (SNR).
With decreasing SNR (increasing noise level) the differences between patches and PIs increases, see Fig. 16a. Interestingly, the differences between the PIs exceeds the differences
between the raw patches (i.e. points are above the diagonal)
which shows that there is a strong sensitivity to the introduced
noise.
Different types of pre-filtering (described in Section 3.4)
can also influence the stability of the resulting representations.
To evaluate this further, we perform a similar analysis for patches
filtered with Schmid filter, MR filter and CLBP (see Fig. 16bd). Experiments reveal that the difference between PIs grows
much slower in case of Schmid filter and MR filter (nearly all
points are below the diagonal). This is due to the fact, that
both filters are defined based on Gaussian functions and have a
low-pass characteristic. Thus, noise is removed to a wide extent in both cases. For CLBP the behavior is similar to the raw
patches. However, the absolute value of difference d is much
lower (d < 0.6 instead of d < 0.9) than for the raw patches
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
(a) Surface
(b) Depth map
(c) Expert ground truth
(d) Result ESDD
(e) Result PI
(f) Result ESDD+PI
Figure 13: Classification results for a surface from the test set: (a) the original rock surface; (b) the depth map of the surface (in false color and contrast enhanced);
(c) the ground truth labeling obtained by domain experts (engraved areas are highlighted red); (d,e,f) show the classification results of ESDD, PI and the combination
of both features. Red areas correspond to correctly classified engraved areas (class 1), green areas are false positives detections of class 1 and blue refers to false
negatives (non-detected engraved areas). All remaining (uncolored) pixels are true positives for class 2.
combined
descriptors
0.7
0.65
0.6
0.55
0.5
PD_AGG
PI
Baseline Baseline
EGHS
ESDD
EGHS + ESDD + ESDD + PI EGHS + EGHS + EGHS +
ESDD
PDAGG
ESDD + ESDD + PI ESDD +
PDAGG
PDAGG +
PI
Small-Scale Dataset
Large-Scale Dataset
Raw
1
25
0.8
20
0.6
15
0.4
10
0.2
0
5
0
0.1
0.2
0.3
0.4
0.5
Normalized difference between patches
Normalized difference between PIs
Figure 14: Performance comparison between small-scale and large-scale
dataset in terms of DSC. Also for the large-scale dataset the topological descriptors add additional information and are necessary to obtain peak performance.
20
20
20
40
40
40
60
60
60
80
80
80
100
100
100
120
120
120
60
(a)
80 100 120
20
40
60
(b)
80 100 120
25
0.8
20
0.6
15
0.4
10
0.2
0
5
0
0.1
0.2
MR
1
25
0.8
20
0.6
15
0.4
10
0.2
0
5
0
0.1
0.2
0.3
(c)
20
40
60
80 100 120
(c)
Figure 15: The patch from Fig. 3a with random Gaussian noise. SNR equals:
(a) 5; (b) 10; (c) and 15.
0.3
0.4
0.5
Normalized difference between patches
(b)
0.4
0.5
Normalized difference between patches
40
Schmid
1
(a)
best non-topological descriptor
best combina!on
20
Normalized difference between PIs
non-topological
descriptors
Normalized difference between PIs
topological
descriptors
Normalized difference between PIs
0.8
0.75
CLBP
1
S
25
0.8
20
0.6
15
0.4
10
0.2
0
5
0
0.1
0.2
0.3
0.4
0.5
Normalized difference between patches
(d)
Figure 16: The noise analysis: abscissa represents difference between the original patch and the same patch with random Gaussian noise; ordinate represents
normalized difference between their PIs; color corresponds to value of SNR.
The dashed line is the diagonal ( f (x) = x) of the plot. Plots correspond to: (a)
original patches; as well as patches filtered with (b) Schmid filter; (c) MR filter;
(d) and CLBP.
Raw
0.5
0.4
0.3
0.2
0.1
0
4 8
16
32
64
Normalized difference between PIs
Normalized difference between PIs
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
Schmid
0.5
0.4
0.3
0.2
0.1
0
16
0.3
0.2
0.1
32
Displacement (in pixels)
(c)
64
Normalized difference between PIs
Normalized difference between PIs
CLBP
0.4
16
2
-0.5
4
6
-1
6
8
-1.5
8
10
-2
10
12
-2.5
12
14
-3
14
16
-3.5
16
10
15
-3.5
-4
-4.5
-5
-5.5
5
10
15
64
(a)
(b)
(b)
MR
4 8
32
Displacement (in pixels)
(a)
0
0
4
5
4 8
Displacement (in pixels)
0.5
2
0.5
Figure 18: Importance of PI pixels in the classification process, measured with:
(a) Fisher discriminant; (b) and Gini importance. The scale in both cases is
logarithmic (base 10). Brighter pixels are more important (discriminative) in
classification. Red arrows correspond to three groups of important pixels.
S
0.4
0.3
0.2
0.1
0
4 8
16
32
64
Displacement (in pixels)
(d)
Figure 17: Displacement analysis: abscissa represents displacement between
the original patch and its neighbor; ordinate represents normalized difference
between their PIs. Each curve corresponds to a single patch. Middle black
curve represents mean value, while the upper and lower black lines presents
the standard deviation, i.e. the grayish area represents the interval of mean ±
standard deviation. Plots correspond to: (a) original patches; as well as patches
filtered with (b) Schmid filter; (c) MR filter; (d) and CLBP.
which shows that CLBP contributes to the robustness of the resulting PI.
Furthermore, we investigate how sensitive PIs are towards
displacement of the underlying patches. For this purpose we
compute the differences in the PIs stemming from a pair of
neighboring and overlapping patches. Again, we randomly select number of patches. Next we shift the patches by 4, 8, 16,
32, and 64 pixels. From the resulting 5 pairs of patches we compute PIs and compute their normalized difference as above. In
Fig. 17a-d we plot the differences in PIs for all displacements
as separate lines for each patch. The trend of the curves shows
that the difference between the PIs grows with the amount of
displacement in all four cases. Interestingly, the differences are
noticeably smaller in the case of CLBP, which indicates that it
introduces an additional robustness to displacements to the PI
which is not the case for Schmid and MR filters.
We observe, that our experiments confirm theoretical results
from Section 3.6. Furthermore, the computational results show
that the constants in the stability theorems have only little influence on the results.
5.13. Discriminativity
An important question in the context of PI is if all entries
in the PI (pixels) are of equal importance or if some pixels are
more important than others. Experiments of feature selection
in Section 5.8 have indicated that a small set of pixels exist
that are more important than the other pixels. To determine the
importance of the individual pixels we compute for each pixel
(a) the Fisher discriminant and (b) the Gini importance for the
entire training set. Fig 18 shows the resulting images where
each pixel in the PI is colored according to its importance.
It turns out that the most important pixels are those located
in the center of the PI quite near to the diagonal. According to
Fisher criterion there are also two smaller groups of important
pixels, above and below the center. This is less visible but also
present in case of Gini, see red arrows in Fig. 18b.
In order to further understand those results, we generated
average PIs for the investigated classes of surface texture separately, see Fig. 19. The averaged PIs clearly show that the
two classes exhibit highly different spatial patterns in the PIs
which strongly correlate with the three groups of pixels identified with Fisher discriminant and Gini importance. This shows
that PI captures the different classes of surface textures well.
The results are further consistent with those obtained for
the synthetic dataset (see Section 5.1). Class 1 representing engraved areas has bimodal distribution, while class 2 representing the natural rock surface is unimodal.
The analysis on discriminativity further provides insights
into the question of weighting in PI computation (see Section
5.6). In general, PI requires a proper weighting function for
stability reasons (see [32]). Our analysis on discriminativity
shows, however, that the components near the diagonal are most
important and thus most probably do not represent noise. The
reason for this is that the 3D surfaces have been smoothed and
filtered for outliers prior to our processing. Thus noise (which
usually appears along the diagonal of PD) has already been removed and all the remaining points actually represent important
topological information. In our case, points near to the diagonal represent small surface deviations that correspond to the
fine-structure of the surface and that are of high importance for
surface texture classification (see Fig. 18). Thus, using weighting would be counterproductive in our case. In other situations,
it is possible to avoid instability issues with PI by using another
approach for constructing PI, such as the one described by Eqn.
(9) in [31], which reduces the influence of the low-persistence
points close to the diagonal when the scale σ increases.
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
0.4
2
2
0.45
4
0.4
0.35
4
0.3
6
0.25
8
0.2
10
0.15
12
0.1
14
0.05
16
5
10
(a)
15
0
0.35
6
0.3
8
0.25
10
0.2
12
0.15
0.1
14
0.05
16
5
10
15
0
(b)
Figure 19: Averaged PIs of the training patches from (a) class 1 and (b) class 2.
6. Conclusions
In this article we have presented an extensive study on topological descriptors for 3D surface texture analysis. From patches
of the surfaces’ depth maps we extract different topological descriptors which build upon the persistence diagram. The descriptors have fixed dimensions and can thus easily be combined with conventional classifiers as well as with traditional
non-topological descriptors. We perform experiments on surface texture classification and investigate different aspects of
the topological descriptors, such as sensitivity to parameters,
robustness to noise, their discriminative power and the information density. Moreover, we propose and evaluate different
normalization strategies, pre-filtering, and alternative weighting
functions and investigate the complementary nature of topological and non-topological descriptors. The major conclusions
drawn from our study are:
• PD AGG and PI capture relevant and discriminative information on surface textures, whereby our study clearly
shows that the spatial information from the PD preserved
in PI leads to a more powerful descriptor.
• PIs exhibit a low sensitivity to changes in resolution and
sigma whereas a certain minimum resolution must be ensured to obtain an expressive descriptor. Sigma shows to
be independent from resolution to a large extent.
• For 3D surfaces with varying depth ranges, depth normalization is essential to obtain comparable topological descriptors. In our experiments global normalization leads
to the best results.
• While related works propose LBP-based representations
our study shows that such a pre-filtering does not necessarily improve results in practice.
• Contrary to our expectations, it turned out that the most
important information in PI is located near to the diagonal of the PD. A deeper analysis reveals that the most
expressive information is spatially clustered in the PI and
that these clusters are characteristic for the investigated
classes.
A number of open questions and future research directions
were identified in our study. Major questions include: (i) how
can weighting in PI construction be improved in a way that
characteristics of the underlying data are taken into account?
(ii) how can the resolution of PI be improved in areas where
the most important information is concentrated to increase the
expressiveness of PI? (iii) do more powerful pre-filter functions
that build upon topological attributes such as [66] improve the
descriptors and can we learn them from the data?
Overall our investigation shows that topological descriptors
can compete with non-topological descriptors for surface texture classification. A comparison with [11] shows that our approach can compete with (and even slightly outperform) the
performance of a CNN on the same dataset. Topological and
non-topological descriptors achieve a significant improvement
of performance when they are combined with each other. This
confirms that topological descriptors capture relevant information that is not captured by traditional descriptors. Thus, they
are a promising approach for further research.
7. Acknowledgments
Parts of the work for this article have been carried out in
the project 3D-Pitoti which is funded from the European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement no 600545; 2013-2016. Other parts of the
work from this article were supported by the National Science
Centre, Poland under grant agreement no 2015/19/D/ST6/01215;
2016-2019.
References
References
[1] D. Crandall, A. Owens, N. Snavely, D. Huttenlocher, Discrete-continuous
optimization for large-scale structure from motion, in: Computer Vision
and Pattern Recognition (CVPR), 2011 IEEE Conference on, IEEE, pp.
3001–3008.
[2] C. Wu, Towards linear-time incremental structure from motion, in:
3DTV-Conference, 2013 International Conference on, IEEE, pp. 127–
134.
[3] J. Wohlfeil, B. Strackenbrock, I. Kossyk, Automated high resolution 3d
reconstruction of cultural heritage using multi-scale sensor systems and
semi-global matching, International Archives of the Photogrammetry,
Remote Sensing and Spatial Information Sciences, XL-4 W 4 (2013) 37–
43.
[4] M. Tuceryan, A. Jain, Texture analysis, The handbook of pattern recognition and computer vision 2 (1998) 207–248.
[5] ASME, ASME B46.1-1995 Surface texture (surface roughness, waviness
and lay): An American National Standard, The Anmerican Society of
Mechanical Engineers (ASME), New York, USA, 1996.
[6] L. Blunt, X. Jiang, Advanced techniques for assessment surface topography: development of a basis for 3D surface texture standards “surfstand”,
Elsevier, 2003.
[7] J. Barcelo, V. Moitinho de Almeida, Functional analysis from visual and
non-visual data. an artificial intelligence approach, Mediterranean Archaeology and Archaeometry 12 (2012) 273–321.
[8] M. Zeppelzauer, M. Seidl, Efficient image-space extraction and representation of 3d surface topography, in: Proceedings of the IEEE International
Conference on Image Processing (ICIP), IEEE, Quebec, Canada, 2015.
[9] M. Zeppelzauer, B. Zieliński, M. Juda, M. Seidl, Topological descriptors
for 3d surface analysis, in: International Workshop on Computational
Topology in Image Context, Springer, 2016, pp. 77–87.
[10] B. T. Fasy, F. Lecci, A. Rinaldo, L. Wasserman, S. Balakrishnan,
A. Singh, et al., Confidence sets for persistence diagrams, The Annals of
Statistics 42 (2014) 2301–2339.
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
[11] G. Poier, M. Seidl, M. Zeppelzauer, C. Reinbacher, M. Schaich, G. Bellando, A. Marretta, H. Bischof, The 3d-pitoti dataset: A dataset for highresolution 3d surface segmentation, in: Proceedings of the 15th International Workshop on Content-Based Multimedia Indexing, ACM, Firenze,
Italy, 2017.
[12] A. E. Johnson, M. Hebert, Using spin images for efficient object recognition in cluttered 3d scenes, Pattern Analysis and Machine Intelligence,
IEEE Transactions on 21 (1999) 433–449.
[13] T. Darom, Y. Keller, Scale-invariant features for 3-d mesh models, Image
Processing, IEEE Transactions on 21 (2012) 2758–2769.
[14] A. Zaharescu, E. Boyer, K. Varanasi, R. Horaud, Surface feature detection
and description with applications to mesh matching, in: Computer Vision
and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, IEEE,
pp. 373–380.
[15] B. Steder, R. B. Rusu, K. Konolige, W. Burgard, Point feature extraction
on 3d range scans taking into account object boundaries, in: IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2011,
pp. 2601–2608.
[16] S. Belongie, J. Malik, J. Puzicha, Shape matching and object recognition
using shape contexts, Pattern Analysis and Machine Intelligence, IEEE
Transactions on 24 (2002) 509–522.
[17] A. Frome, D. Huber, R. Kolluri, T. Bülow, J. Malik, Recognizing objects in range data using regional point descriptors, in: Computer VisionECCV 2004, Springer, 2004, pp. 224–237. 00401.
[18] R. B. Rusu, Z. C. Marton, N. Blodow, M. Beetz, Persistent point feature
histograms for 3d point clouds, in: Proc 10th Int Conf Intel Autonomous
Syst (IAS-10), Baden-Baden, Germany, IOS Press, 2008, pp. 119–128.
[19] R. B. Rusu, N. Blodow, M. Beetz, Fast point feature histograms (FPFH)
for 3d registration, in: Robotics and Automation, 2009. ICRA’09. IEEE
International Conference on, IEEE, 2009, pp. 3212–3217. 00320.
[20] E. Wahl, U. Hillenbrand, G. Hirzinger, Surflet-pair-relation histograms: a
statistical 3d-shape representation for rapid classification, in: 3-D Digital
Imaging and Modeling, 2003. 3DIM 2003. Proceedings. Fourth International Conference on, IEEE, 2003, pp. 474–481.
[21] F. Tombari, S. Salti, L. Di Stefano, Unique signatures of histograms for
local surface description, in: Computer Vision–ECCV 2010, Springer,
2010, pp. 356–369. 00160.
[22] A. Othmani, L. Lew Yan Voon, C. Stolz, A. Piboule, Single tree species
classification from terrestrial laser scanning data for forest inventory, Pattern Recognition Letters 34 (2013) 2144–2150.
[23] T. Leung, J. Malik, Detecting, localizing and grouping repeated scene
elements from an image, in: European Conference on Computer Vision,
Springer, 1996, pp. 546–555.
[24] G. Csurka, C. Dance, L. Fan, J. Willamowski, C. Bray, Visual categorization with bags of keypoints, in: Workshop on statistical learning in
computer vision, ECCV, 1-22, Prague, 2004, pp. 1–2.
[25] F. Perronnin, C. Dance, Fisher kernels on visual vocabularies for image categorization, in: 2007 IEEE Conference on Computer Vision and
Pattern Recognition, IEEE, pp. 1–8.
[26] M. Cimpoi, S. Maji, I. Kokkinos, A. Vedaldi, Deep filter banks for texture recognition, description, and segmentation, International Journal of
Computer Vision 118 (2016) 65–94.
[27] P. Frosini, Measuring shapes by size functions, in: Intelligent Robots and
Computer Vision X: Algorithms and Techniques, International Society
for Optics and Photonics, 1992, pp. 122–133.
[28] A. Verri, C. Uras, P. Frosini, M. Ferri, On the use of size functions for
shape analysis, Biological Cybernetics 70 (1993) 99–107.
[29] H. Edelsbrunner, D. Letscher, A. Zomorodian, Topological persistence
and simplification, Discrete and Computational Geometry 28 (2002) 511–
533.
[30] C. Li, M. Ovsjanikov, F. Chazal, Persistence-based structural recognition, in: IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), IEEE, 2014, pp. 2003–2010.
[31] J. Reininghaus, S. Huber, U. Bauer, R. Kwitt, A stable multi-scale kernel
for topological machine learning, in: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4741–4748.
[32] H. Adams, T. Emerson, M. Kirby, R. Neville, C. Peterson, P. Shipman,
S. Chepushtanova, E. Hanson, F. Motta, L. Ziegelmeier, Persistence images: A stable vector representation of persistent homology, Journal of
Machine Learning Research 18 (2017) 1–35.
[33] M. Ferri, P. Frosini, A. Lovato, C. Zambelli, Point selection: A new
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
comparison scheme for size functions (with an application to monogram
recognition), in: Computer Vision — ACCV’98: Third Asian Conference
on Computer Vision Hong Kong, China, January 8–10, 1998 Proceedings,
Volume I, Springer, 1998, pp. 329–337.
P. Donatini, P. Frosini, A. Lovato, Size functions for signature recognition, in: Proc. SPIE, volume 3454, pp. 178–183.
H. J. Poincaré, Sur le probleme des trois corps et les équations de la
dynamique, Acta Mathematica 13 (1890) 1–270.
H. J. Poincaré, Les méthodes nouvelles de la mécanique céleste,
Gauthiers-Villars, Paris (1892, 1893, 1899).
H. J. Poincaré, Analysis situs, J. Éc. Polytech., ser. 2 1 (1895) 1–123.
P. Bubenik, Statistical topological data analysis using persistence landscapes., Journal of Machine Learning Research 16 (2015) 77–102.
N. Makarenko, M. Kalimoldayev, I. Pak, A. Yessenaliyeva, Texture
recognition by the methods of topological data analysis, Open Engineering 6 (2016).
C. Schmid, Constructing models for content-based image retrieval, in:
Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on, volume 2, IEEE,
pp. II–39.
J.-M. Geusebroek, A. W. Smeulders, J. Van de Weijer, Fast anisotropic
gauss filtering, IEEE Transactions on Image Processing 12 (2003) 938–
943.
Z. Guo, L. Zhang, D. Zhang, A completed modeling of local binary
pattern operator for texture classification, IEEE Transactions on Image
Processing 19 (2010) 1657–1663.
D. Cohen-Steiner, H. Edelsbrunner, J. Harer, Stability of persistence diagrams, Discrete & Computational Geometry 37 (2007) 103–120.
D. Cohen-Steiner, H. Edelsbrunner, J. Harer, Y. Mileyko, Lipschitz functions have lp-stable persistence, Foundations of Computational Mathematics 10 (2010) 127–139.
N. Garcia, E. Stoll, Monte carlo calculation for electromagnetic-wave
scattering from random rough surfaces, Physical review letters 52 (1984)
1798.
C. Seiffert, T. M. Khoshgoftaar, J. Van Hulse, A. Napolitano, Rusboost:
A hybrid approach to alleviating class imbalance, Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on 40 (2010)
185–197.
Y. Freund, R. E. Schapire, A decision-theoretic generalization of on-line
learning and an application to boosting, Journal of computer and system
sciences 55 (1997) 119–139.
G. Fumera, F. Roli, Cost-sensitive learning in support vector machines,
VIII Convegno Associazione Italiana per LIntelligenza Artificiale (2002).
A. Vedaldi, B. Fulkerson, Vlfeat: An open and portable library of computer vision algorithms, in: Proceedings of the International Conference
on Multimedia, ACM, 2010, pp. 1469–1472.
M. Juda, M. Mrozek, P. Brendel, H. Wagner, et al., Capd::redhom,
http://redhom.ii.uj.edu.pl, 2010-2015.
M. Juda, M. Mrozek, Capd:: Redhom v2-homology software based on
reduction algorithms, in: Mathematical Software–ICMS 2014, Springer,
2014, pp. 160–166.
U. Bauer, M. Kerber, J. Reininghaus, Phat - persistent homology algorithms toolbox, https://bitbucket.org/phat-code/phat, 2013.
U. Bauer, M. Kerber, J. Reininghaus, H. Wagner, Phat - persistent homology algorithms toolbox, in: H. Hong, C. Yap (Eds.), Mathematical
Software - ICMS 2014, volume 8592 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2014, pp. 137–143.
G. Zhenhua, L. Zhang, D. Zhang, Clbp implementation for matlab, http:
//www.comp.polyu.edu.hk/~cslzhang/code/CLBP.rar, 2010.
L. Breiman, Random forests, Machine learning 45 (2001) 5–32.
R. O. Duda, P. E. Hart, et al., Pattern classification and scene analysis,
volume 3, Wiley New York, 1973.
ISO-IEC, Information Technology - Multimedia Content Description Interface, 15938, ISO/IEC, Moving Pictures Expert Group, 1st edition,
2002.
D. G. Lowe, Distinctive image features from scale-invariant keypoints,
International journal of computer vision 60 (2004) 91–110.
T. Ojala, M. Pietikäinen, D. Harwood, A comparative study of texture measures with classification based on featured distributions, Pattern
recognition 29 (1996) 51–59.
N. Dalal, B. Triggs, Histograms of oriented gradients for human detec-
Preprint of Article in Elsevier Journal on Computer Vision and Image Understanding (CVIU):
https://doi.org/10.1016/j.cviu.2017.10.012
[61]
[62]
[63]
[64]
[65]
[66]
tion, in: Computer Vision and Pattern Recognition, 2005. CVPR 2005.
IEEE Computer Society Conference on, volume 1, IEEE, pp. 886–893.
R. M. Haralick, K. Shanmugam, I. H. Dinstein, Textural features for
image classification, Systems, Man and Cybernetics, IEEE Transactions
on (1973) 610–621.
A. Vedaldi, B. Fulkerson, VLFeat: An open and portable library of computer vision algorithms, http://www.vlfeat.org/, 2008.
M. Heikkilae, T. Ahonen, A general local binary pattern (lbp) implementation for matlab, http://www.cse.oulu.fi/CMV/Downloads/
LBPMatlab, 2014.
A. Uppuluri, Glcm texture features, http://de.mathworks.com/
matlabcentral/fileexchange/22187-glcm-texture-features,
2008.
J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for semantic segmentation, in: Proceedings of the International Conference on
Computer Vision and Pattern Recognition (CVPR), pp. 3431–3440.
T. P. Nguyen, A. Manzanera, W. G. Kropatsch, Impact of topology-related
attributes from local binary patterns on texture classification, in: Computer Vision - ECCV 2014 Workshops: Zurich, Switzerland, September
6-7 and 12, 2014, Proceedings, Part II, Springer International Publishing,
2015, pp. 80–93.
| 1cs.CV
|
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
1
Latent Parameter Estimation in Fusion Networks
Using Separable Likelihoods
arXiv:1708.00842v2 [cs.SY] 2 Jan 2018
Murat Üney, Member, IEEE, Bernard Mulgrew, Fellow, IEEE, Daniel E. Clark, Member, IEEE
Abstract—Multi-sensor state space models underpin fusion
applications in networks of sensors. Estimation of latent parameters in these models has the potential to provide highly
desirable capabilities such as network self-calibration. Conventional solutions to the problem pose difficulties in scaling with the
number of sensors due to the joint multi-sensor filtering involved
when evaluating the parameter likelihood. In this article, we
propose a separable pseudo-likelihood which is a more accurate
approximation compared to a previously proposed alternative
under typical operating conditions. In addition, we consider
using separable likelihoods in the presence of many objects
and ambiguity in associating measurements with objects that
originated them. To this end, we use a state space model with
a hypothesis based parameterisation, and, develop an empirical
Bayesian perspective in order to evaluate separable likelihoods
on this model using local filtering. Bayesian inference with this
likelihood is carried out using belief propagation on the associated
pairwise Markov random field. We specify a particle algorithm
for latent parameter estimation in a linear Gaussian state space
model and demonstrate its efficacy for network self-calibration
using measurements from non-cooperative targets in comparison
with alternatives.
Index Terms—sensor networks, hidden Markov models,
Markov random fields, pseudo-likelihood, simultaneous localisation and tracking, Monte Carlo algorithms, dynamical Markov
random fields
I. I NTRODUCTION
wide range of sensing applications including wide area
surveillance is underpinned by state space models which
are capable of representing a variety of dynamic phenomena
such as spatio-temporal (see, e.g., [1]) processes. In fusion
(or, object tracking [2]) networks, multi-sensor versions of
stochastic state space models, also known as hidden Markov
models [3], are used to estimate object trajectories in a
surveillance region.
These models, however, are often specified by some latent
parameters [4] some of which are unknown in practice and
need to be estimated based on measurements from the state
processes (or, objects). Examples of this problem setting in
fusion networks include estimation of noise parameters [5],
sensor biases [6], [7] and localisation/calibration of sensors in
a GPS denying environment (e.g., in underwater sensing [8])
A
This work was supported by the Engineering and Physical Sciences
Research Council (EPSRC) Grant number EP/K014277/1 and the MOD
University Defence Research Collaboration (UDRC) in Signal Processing.
Murat Üney and Bernard Mulgrew are with the Institute for Digital
Communications, School of Engineering, University of Edinburgh, EH9 3FB,
Edinburgh, UK (e-mail: [email protected], [email protected]).
Daniel E. Clark is with Départment CITI, Telecom-SudParis, 9, rue
Charles Fourier 91011, EVRY Cedex, France (e-mail: [email protected]).
using point detections from non-cooperative targets [9], [10].
Another example is the estimation of the orientations and
positions of nodes in a camera network based on feature
detections [11].
Such problems fall in the domain of parameter estimation in
state space models (see, e.g., [12] for a review). The parameter
likelihood of the multi-sensor problem, however, does not
scale well with the number of sensors which specifies the
dimensionality of the unknown, or, the length of the measurement window that will be used for estimation. In the presence
of multiple objects, the scalability issue is exacerbated by
the measurement origin (or, data association) uncertainties
that arise. Exact evaluation of the likelihood in this case has
combinatorial complexity with the number of sensors [13],
and, in general multiple object models, it is intractable even for
a single sensor [14]. Estimation using a maximum likelihood
(ML) or a Bayesian approach requires repeated evaluation of
this likelihood (see, e.g., [12], [15], [16]) necessitating the use
of efficient approximation strategies.
Intractable or computationally prohibitive likelihoods have
motivated a number of lines of work in the statistics literature including likelihood free methods, or, approximate
Bayesian computation [17], and, composite likelihood/pseudolikelihood approaches [18]. Likelihood free methods can be
used for sampling from the parameter posterior in state space
models [19] including those capable of modelling multiple
objects [20]. The latter approach is based on developing
surrogates to replace the original likelihood, e.g., block based
approximations in maximum likelihood [21]. The pseudolikelihood perspective has been useful in networked settings
in which constraints on i) the availability of parts of data,
and/or, ii) scalability in processing with the number of sources
arise. Examples include surrogates built upon local functions
for estimation of parametric probability measures (e.g., exponential family distributions) from distributedly stored high
dimensional samples [22]–[24].
It is not straightforward to find such pseudo-likelihoods for
parameter estimation in state space models, however, that can
resolve these two issues that arise when there are multiple data
sources (or, sensors). It is worthwhile to develop and analyse
surrogates that provide scalability with the number of sources,
and, are suitable to local computations (e.g., local filtering).
In [25], we proposed a pseudo-likelihood which is a product
of “dual-term” approximations replacing their intractable exact
counterparts. These approximations are separable in that they
can be evaluated using single sensor filtering. This feature
underpins scalability with the number of sensors. In [26], we
have investigated the quality of the dual-term approximation,
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
and, related it to the level of uncertainty in the prediction and
estimation of the underlying state process.
In this work, we propose an alternative pseudo-likelihood
which is provably a more accurate approximation for parameter estimation in multi-sensor state space models, under typical
operating conditions. This approximation is also separable
in that it is a scaled product of quadruple terms each of
which can be found using single sensor filtering. In order
to exploit this quad-term likelihood when there are multiple
objects, extra attention should be paid to the handling of the
data association uncertainties. We propose to use a hypothesis
based parameterisation for the multi-object state space model
as detailed in [14]& [27] in order to facilitate the use of
the quad-term surrogate in this setting. In the parameterised
model, we explicitly point out the combinatorial complexity
of exact likelihood evaluation with the number of sensors.
Then, we introduce an empirical Bayesian [28] interpretation
of local filtering that facilitates the use of separable likelihoods
within this model. These modelling aspects detailing the use of
separable likelihoods in hypothesis based multi-object models
constitute the second contribution of this work.
Separable likelihoods fit well in distributed fusion archictectures in which locally filtered distributions are transmitted
in the network, as opposed to sensor measurements [29].
Moreover, they facilitate parameter estimation using a message passing computational structure which is desirable in
networked problems. Specifically, the proposed likelihood
surrogate together with independent parameter priors lead to
a pairwise Markov random field (MRF) posterior model. The
marginal distributions of this model approximates posterior
marginals of the latent parameters to be estimated. We estimate
these marginals iteratively using Belief Propagation (BP) [30]
which consists of successive message passings among neighbouring nodes and updating of local marginals based on
these messages. This computational structure lends itself to
decentralised estimation, as well as scalable computation at
fusion centre.
As an indication of the approximation quality, we consider
the Kullback-Leibler divergence (KLD) [31] of the quad-term
likelihood with respect to the actual pairwise likelihood and
relate it to the uncertainties in predicting and estimating the
underlying state using individual and joint sensor histories.
We show that with more accurate local filters the approximation quality improves and the proposed quad-term separable
likelihood has an improved error bound compared to the
aforementioned dual-term approximation.
We provide a Monte Carlo algorithm for sensor selfcalibration in this framework for linear Gaussian state space
(LGSS) models. The algorithm is based on the nonparametric
BP approach [32] and involves sampling from the updated
marginals followed by quad-term likelihood evaluations in
the message passing stage. As BP iterations converge to a
fixed point, the empirical average of the samples from the
marginals constitute (an approximate) minimum mean squared
error (MMSE) estimate of the latent parameters. The edge
potential are evaluated using the entire measurement history
within a selected time period in an offline fashion which
is a strategy similar to particle Markov chain Monte Carlo
2
(MCMC) algorithms [15]. As such, we differ from [26] in
which windowing of measurements are used for enabling
online processing.
Preliminary results of the proposed pseudo-likelihood can
be found in [33]. This article provides a complete account of
our solution strategy in multiple object models and is structured as follows: Section II provides the probabilistic model
and the problem statement. Then, we detail pairwise pseudolikelihoods in parameterised multi-object models, and, relate
this perspective to latent parameter estimation via inference
over pairwise MRFs, in Section III. The proposed quad-term
node-wise separable likelihood approximation is detailed in
Section IV. Section V details the structural and computational
properties of the quad-term approximation when the unknowns
are respective quantities. Based on these results, we propose
a distributed sensor localisation algorithm in linear Gaussian
multi-object state space models in Section VI. The efficacy of
this algorithm is demonstrated in comparison to the approach
in [26], in Section VII. Finally, we conclude in Section VIII.
II. P ROBLEM D EFINITION
A. Probabilistic model
Let us consider a set of sensors V “ t1, . . . , N u networked
over communication links listed by E Ă V ˆ V. The graph
G “ pV, Eq is undirected (i.e., the links are bi-directional),
connected, and, might contain cycles.
Next, let us consider a single object with state evolution
modelled as a Markov process Xk for time index k ě 1.
This process is specified by an initial state distribution and a
transition density. The state space model with parameters θ is
then specified as follows [12]: The state value xk is a point
in the state space X and is generated by the chain
Xk |pX1:k´1 “ x1:k´1 q „ πpxk |xk´1 ; θq,
X1 „ πb px1 ; θq,
(1)
where .|. denotes conditioning. A measured value zki P Zi at
sensor i P V is generated independently in accordance with
the likelihood model
i
i
Zki |pX1:k “ x1:k , Z1:k
“ z1:k
q „ gi pzki |xk ; θq
(2)
where subscript 1 : k indicates a vector concatenation over
time.
In fusion scenarios, there are multiple such objects denoted
by a multi-object state
Xk fi rXk,1 , . . . , Xk,Mk s ,
(3)
that induce measurements according to the above state space
model resulting with sensors collecting a multitude of measurements
”
ı
i
i
Zki fi Zk,1
, . . . , Zk,O
,
(4)
i
k
where Mk is the number of objects and Oki is the number of
measurements collected at sensor i at time k. Here, the origin
i
s are unknown, i.e., the data associations which encode
of Zk,j
a mapping from these measurement (random) variables to the
elements of Xk (and, equivalently to the previously collected
measurements from the same objects) are not known [2].
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
In the general multi-object tracking model, Mk and Oki are
random variables with laws determined by probability models
regarding how these objects appear in the surveillance region
and disappear (which is often referred to as their birth and
death, respectively), the law for the false alarms, etc. For
the sake of simplicity and ease of presentation in the limited
space especially when relating computational complexity to
the number of sensors and objects in the following discussion,
we assume that all of the objects that exist at time step
k “ 1 remain in the scene for the time window considered
and there are no missed detections and false alarms in sensor
measurements which imply that Mk “ M and Oki “ Mk ,
respectively, for some positive integer M with probability one.
In this simplified “closed world” model the multi-object
state transition is given by
πpXk |Xk´1 ; θq “
M
ź
πpxk,m |xk´1,m ; θq.
(5)
m“1
The likelihood of the measurements collected by sensor i is
conditioned not only on the multi-object state Xk , but, also on
a (data association) hypothesis τki that encodes the association
of measurements to the objects within Xk :
li pZik |Xk , τki ; θq “
M
ź
i
gi pzk,o
|xk,τki poq ; θq,
(6)
o“1
and, the prior on τki assigns equal probability to all M !
permutations of r1, . . . , M s that τki can take, i.e.,
ppτki q “
1
.
M!
(7)
B. Statement of the problem
We are interested in estimating θ P B using the measurements collected across the network by sensors i P V for a
time window of length t. The parameter likelihood of the
problem quantifies how well these measurements fit into the
state space model with the selected value of the parameter,
and, is evaluated via multi-sensor filtering [4, Sec.IV]:
`
˘
l Z11:t , . . . , ZN
1:t |θ “
t
ź
`
˘
1
N
p Z1k , . . . , ZN
k |Z1:k´1 , . . . , Z1:k´1 , θ , (8)
k“1
where the time updates on the right hand side are given by
˘
`
1
N
p Z1k , . . . , ZN
k |Z1:k´1 , . . . , Z1:k´1 , θ “
ÿ ÿ
¨ ¨ ¨ ppτk1 q ˆ . . . ˆ ppτkN q
τk1
ˆ
ż
XM
τkN
1
N
lpZ1k , . . . , ZN
k |Xk , τk , . . . , τk ; θq
ˆ ppXk |Z11:k´1 , . . . , ZN
1:k´1 ; θqdXk , (9)
and, the multi-sensor likelihood inside the integration factorises as
ź
1
N
lpZ1k , . . . , ZN
li pZik |Xk , τki ; θq.
k |Xk , τk , . . . , τk ; θq “
iPV
(10)
3
where the terms in the product are given by (6).
Here, (8) follows from the chain rule of probabilities. The
term in (9) is the contribution at time step k which updates
the likelihood of the previous time step and is found using the
Markov property that the sensor measurements are mutually
independent of the measurement histories, conditioned on
the current state for any value of θ. Let us denote this
j
relation by Zkj K
KZ1:k´1
|Xk , θ for i P V (see, e.g., [34], for
this notation). (10) follows from that the measurements of
different sensors are mutually independent, i.e., Zki K
KZkj |Xk , θ
for pi, jq P V ˆ V.
This likelihood can be used in a MMSE estimator of θ P B,
in principle, for a random variable Θ associated with a prior
density ppθq. This estimate is given by the expected value of
the posterior distribution
1
N
ppθ|Z11:t , . . . , ZN
(11)
1:t q 9 lpZ1:t , . . . , Z1:t |θq ppθq,
ż
θ̂ “
θ ppθ|Z11:t , . . . , ZN
1:t q dθ. (12)
B
The MMSE estimate can be computed by generating L
samples from the posterior distribution in (11) using, for
example, MCMC methods [15] and using these samples to
find a Monte Carlo estimate of the integral in (12). In both this
approach and maximum likelihood (ML) solutions aiming to
maximise (8) with iterative optimisation, repeated evaluations
of the likelihood are required.
The evaluation of this likelihood is intractable, however,
not only because of the pM !qN summations in (9), but, also
because of the complexity in finding the integrations involved.
The integrands here are i) the multi-sensor likelihood in (10),
and, ii) the prediction density for Xk based on the network’s
entire measurement history up to time k. In other words,
(9) is the scale factor for the posterior density of Bayesian
recursions, or, the “centralised” filter given by
ppXk , τk1:N |Z11:k , . . . , ZN
1:k ; θq “
1:N
lpZ1k , . . . , ZN
; θq
k |Xk , τk
` 1
˘ ˆ ppτk1:N q
N
1
p Zk , . . . , Zk |Z1:k´1 , . . . , ZN
,
θ
1:k´1
ˆppXk |Z11:k´1 , . . . , Zj1:k´1 , θq,
ppXk |Z11:k´1 , . . . , ZN
1:k´1 , θq
ÿ
1
τk´1
¨¨¨
ÿ ż
N
τk´1
(13)
“
πpXk |Xk´1 , θq
XM
1:N
ˆppXk´1 , τk´1
|Z11:k´1 , . . . , ZN
1:k´1 , θqdXk´1 ,
(14)
where we denote by τk1:N the concatenation of τki s and it has
pM !qN different configurations. Here, both the prediction (14)
and update (13) are OppM !qN q.
In order to address these challenges, multi-object filtering
(or, tracking) algorithms often employ two approximations:
First, they aim to find the most probable data association
hypothesis in (13) denoted by τ̄ 1:N
k , instead of both evaluating
this expression for all possible associations and storing them.
The benefits of doing so are that i) one can generate tracks
(or, object trajectories) as simply marginals of ppXk , τk1:N “
τ̄ 1:N
k | . q for k “ 1, ..., t, and, ii) evaluations of the integral
in (14) in the next time step can be restricted to this value of
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
the hypothesis variable. Equivalently, the posterior distribution
in (13) is factorised as
ppXk , τk1:N |Z11:k , . . . , ZN
1:k ; θq
1:N
“ ppXk |Z11:k , . . . , ZN
qppτk1:N |Z11:k , . . . , ZN
1:k , θ, τk
1:k , θq
1:N
“ ppXk |Z11:k , . . . , ZN
qppτk1:N |Z1k , . . . , ZN
1:k , θ, τk
k , θq(15)
where the association variables appear as model parameters
in the first term and the second term is similar to a prior
distribution on these models with the difference that it is
conditioned on the current measurements. At time k ´ 1, let
us select this “empirical prior” as
1:N
1:N
ppτk´1
|Z1k´1 , . . . , ZN
pτk´1
q
k´1 , θq Ð δτ̄ 1:N
k´1
(16)
where δ is Kronecker’s delta function and Ð denotes assignment. The second approximation follows from the first one:
Evaluation of the prediction stage in (14) reduces to evaluation
of the Chapman-Kolmogorov equation for only the most likely
value of the association parameters. This approach is often
referred to as empirical Bayes [28], and is used to facilitate
approximate solutions to otherwise intractable problems.
A similar approximation can be used when evaluating
the parameter posterior in (11). This leads to the following
likelihood
˘
`
1:N
“
l Z11:t , . . . , ZN
1:t |θ, τ1:t
t
ź
`
˘
1
N
pτk1:N Z1k , . . . , ZN
k |Z1:k´1 , . . . , Z1:k´1 , θ , (17)
k“1
1:N
conditioned on τ1:t
where the factors are defined by
`
˘
1
N
pτk1:N Z1k , . . . , ZN
k |Z1:k´1 , . . . , Z1:k´1 , θ fi
ż
1:N
lpZ1k , . . . , ZN
; θq
k |Xk , τk
XM
ˆ ppXk |Z11:k´1 , . . . , ZN
1:k´1 ; θqdXk . (18)
1:N
“ τ̄ 1:N
This likelihood evaluated at τ1:k
1:k replaces the one
in (8) when the empirical (model) prior is selected as in (16)
(see Appendix A for details). We will refer to (17) as the
empirical likelihood. Note that (18) is the integral term in (9).
The empirical likelihood update term is computationally
more convenient, however, alone it is not sufficient for scalais equivbility with the number of sensors N : Finding τ̄ 1:N
k
alently an N ` 1-dimensional assignment problem which is
NP hard even for N “ 2 sensors [35] which partly underlies
the local filtering paradigm for multi-sensor processing and
our interest in compatible solutions. For a moment, let us
consider the problem for a single object, i.e., for M “ 1.
In this case τki for i “ 1, . . . , N have only one possible
configuration (i.e., there is no data association uncertainty).
Because the dimensionality of θ is specified by N and (8)
will be evaluated for roughly N L samples (when estimating
(12) (see, e.g., [36])) each of which costing – in the simplest
linear Gaussian measurements case (9) 1 – at the least OpN 2 tq,
1 Specifically, for linear Gaussian measurements with no data association
uncertainty, the marginal parameter likelihood involves computation of the
innovation covariance for the so called group-sensor measurements in joint
multi-sensor filtering.
4
the computational cost will be cubic in the number of sensors
which can easily become prohibitive for large N .
The networked setting has additional constraints to take
into account: The sensors perform local filtering of their
measurements and exchange filtered (track) distributions over
G as opposed to transmitting their measurements [29]. As a
result, the network-wide measurements are not available to
evaluate the likelihood of the problem. Instead, local distributions we denote by ppXk , τkj “ τ̂kj |Zj1:k q are made available to
neighbouring nodes where τ̂kj is an approximation to the most
probable association configuration τ̄ jk found locally, based on
only the local sensor measurements at sensor j. There are
computationally efficient algorithms for finding such solutions
for the single sensor problem (see, e.g., [35] and the references
therein). Therefore, a viable solution needs to build upon these
densities and local data associations τ̂ki as opposed to joint
multi-sensor filtering in the network.
The problem we address in this work is the design of
scalable approximations to (8) for estimating θ in a networked setting based on local filtering results at the nodes.
The proposed approach also addresses the aforementioned
computational bottleneck at fusion centres in centralised multisensor architectures with a designated node receiving unfiltered sensor measurements.
It is also worth noting that the parameter vector θ P B
can be used to represent a wide variety of parameters of
the global model some of which can be intrinsic to sensors
i P V individually such as parameters pertaining to local
noise models. We are particularly interested in a second class
of parameters which have dependencies among sensors such
as respective parameters such as sensor locations and similar
“calibration” parameters. In the former setting, the estimation
of local parameters decouple into independent estimation
problems which can be solved using a suitable approach (see,
e.g., [27], [37], [38]).
In our setting, θ fi rθ1 , . . . , θN s where θi is associated with
i P V and its estimation does not decouple and depends on
all measurements across the network due to the dependencies
of parameters, which, in turn, brings forward the multi-sensor
aspects of the problem this work aims to address. Because
local filtering is performed, on the other hand, local estimation
of data association τ̂ki is available independent of θ, which we
discuss in detail later in Section V.
III. A P SEUDO - LIKELIHOOD AND
A PAIRWISE MRF
POSTERIOR FOR DECENTRALISED ESTIMATION
Pseudo-likelihoods are constructed from likelihood like
functions which are computationally convenient and defined
typically over smaller subsets of the data to overcome difficulties posed by intractable likelihoods over the entire set of
data (see, e.g., [18] and the references therein). Let us denote
the network-wide data set by
Z fi rZ11:t , . . . , ZN
1:t s.
A fairly general form for a pseudo-likelihood is given
by [18]
ź
˜lpZd |Zc , θqωs
l̃pZ|θq “
(19)
s
s
sPS
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
5
summarise the information neighbours have gained on these
variables. This is described for all i P V by
ż
ź
mji pθi q “ ψij pθi , θj q ψj pθj q
mi1 j pθj q dθj , (24)
i1 Pnepjqzi
p̃i pθi q 9 ψi pθi q
ź
mji pθi q.
(25)
jPnepiq
Fig. 1. A multi-sensor state space - or, hidden Markov- model (black dashed
box on the right representing a chain over k) and a Markov Random field
model of the parameter posterior (the blue edges on the left).
where S is an index set, ωs is a positive real number,
and, Zds and Zcs are mutually exlusive subsets of Z (for
example, Zd1 “ Zi2 and Zc1 “ Zj1 , etc.). These sets can
be selected in various ways ensuring that the factors are
computationally convenient functions, for example, marginals
and/or conditional densities, and, estimates based on l̃pZ|θq
are sensible. Note that (8) is also in this form, however, with
difficult to evaluate factors.
Let us consider θ “ rθ1 , ..., θN s and a pseudo-likelihood
surrogate for the empirical likelihood in (17):
ź
˜
lτ1:t
1:N pZ|θq “
(20)
lτ i,j pZi , Zj |θi,j q,
pi,jqPE
“
In BP iterations, nodes simultaneously send messages to
their neighbours using (24) (often using constants as the
previously received messages during the first step) and update
their local “belief” using (25). If G contains no cycles (i.e., G
is a tree), p̃i s are guaranteed to converge to the marginals of
(22), in a finite number of steps [30]. For the case in which G
contains cycles, iterations of (24) and (25) are known as loopy
BP (LBP). For the case, convergence does not have general
guarantees, nevertheless LBP has been been very successful
in computing approximate marginals in a distributed fashion,
in fusion, self-localisation and tracking problems in sensor
networks [39]–[41]. In our problem setting, we assume that
the models over spanning trees of a loopy G are consistent
in that they lead to “similar” marginal parameter distributions,
which suggests the existence of LBP fixed points [42] that will
be converged when initial beliefs are selected reasonably [43].
1:t
t
ź ź
pτ i,j pZik , Zjk |Zi1:k´1 , Zjk´1 , θi,j q.(21)
k
pi,jqPE k“1
Here, E is essentially the set of sensor pairs whose likelihoods
we would like to incorporate into the pseudo-likelihood,
and, it is convenient to choose them as those that share a
communication link, in a networked setting (Section II-A).
The pairwise structure above is beneficial to use with the
MMSE estimator in (12). Note that the MMSE estimate is the
concatenation of the expected values of posterior marginals,
i.e., ppθi |Zq for i “ 1, . . . , N . These distributions can be found
using message passing algorithms over G when the surrogate
(20) is used in (11) together with independent but arbitrary a
priori distributions selected for Θi s. Specifically, the parameter
posterior corresponding to such a selection of prioirs and the
pseudo-likelihood (20) is a pairwise Markov random field over
G “ pV, Eq [34]:
ź
ź
ppθ|Zq 9
ψi pθi q
ψij pθi , θj q,
(22)
iPV
pi,jqPE
ψi pθi q “ p0,i pθi q,
ψij pθi , θj q “ lτ i,j pZi , Zj |θi , θj q,
1:t
(23)
where the node potential functions (i.e., ψi s) are the selected
priors (e.g., uniform distributions over bounded sets θi s take
values from) and the edge potentials (i.e., ψij s) are the pairwise likelihoods for the pairs pi, jqs. This model is illustrated
in Fig. 1.
The pairwise MRF model in (22) allows the computation
of posterior marginal ppθi |Zq through iterative local message
passings such as Belief Propagation (BP) [30]. In BP, the
nodes maintain distributions over their local variables and
update them based on messages from their neighbours which
IV. Q UAD - TERM
NODE - WISE SEPARABLE LIKELIHOODS
The pseudo-likelihood introduced in Section III leads to a
parameter posterior that admits a pairwise MRF model. This
is advantageous in providing a means for decentralised estimation through message passing algorithms in a network. The
edge potentials (23) of this model, however, are i) conditioned
jointly on two sensors’ measurements simultaneous access to
which is infeasible in a networked setting, and, ii) conditioned
on association variables for two sensors and ideally should be
i,j
evaluated at its most probable configuration τ̄ i,j
1 , . . . , τ̄ t each
of which is NP-hard to find, as explained in Section II.
In order to overcome these difficulties, we introduce an
approximation which factorises into terms local to nodes, i.e.,
a node-wise separable approximation. Let us consider the
“centralised” pairwise likelihood update term in (21) given
some configuration τki,j for k “ 1, . . . , t, and, drop them from
the subscript for the sake of simplicity in notation, as well as
the i, j subscript in θ, in the following discussion. This term
factorises in alternative ways as follows:
ppZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq
“ ppZik |Zi1:k´1 , Zj1:k , θqppZjk |Zi1:k´1 , Zj1:k´1 , θq
“
(26)
ppZjk |Zi1:k , Zj1:k´1 , θqppZik |Zi1:k´1 , Zj1:k´1 , θq
(27)
¯1{2
“ ppZik |Zi1:k´1 , Zj1:k , θqppZjk |Zi1:k´1 , Zj1:k´1 , θq
´
¯1{2
ˆ ppZjk |Zi1:k , Zj1:k´1 , θqppZik |Zi1:k´1 , Zj1:k´1 , θq (28)
´
In the first and second lines above, the chain rule is used.
The third equality can be found by taking the geometric mean
of the first two expressions. All four factors in Eq.(28) are
conditioned on the measurement histories of both sensors to
which one cannot have simulatenous access in a networked
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
setting. We would like to aviod this by leaving out the history
of sensor i (sensor j) in the first two (last two) terms of
(28), i.e.,
qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq
¯1{2
1 ´
fi
ppZik |Zj1:k , θqppZjk |Zj1:k´1 , θq
κk pθq
´
¯1{2
ˆ ppZjk |Zi1:k , θqppZik |Zi1:k´1 , θq
κk pθq “
ż ż ´
1
(29)
1
ppZik , Zjk |Zj1:k´1 , θq
¯1{2
1
1
1
1
ˆppZik , Zjk |Zi1:k´1 , θq
dZik dZjk
(30)
where κk pθq is the normalisation constant that guarantees
q to integrate to unity. Note that κk is a function of the
parameters θ.
The appeal of this quadruple term is that its factors depend
on single sensor histories. As such, they require filtering of
sensor histories of i and j individually enabling the evaluation
of their product in a network. This point is discussed later in
this section.
A. Approximation quality
We consider the difference between the original centralised
update term in (28) and the quad-term approximation introduced in (29). Because these terms are probability densities
over sensor measurements, their “divergence” can be quantified using the KLD [31]:
Proposition 4.1: The KLD between the centralised update
and the node-wise separable approximation in (29) is bounded
by the average of the mutual information (MI) [31] between
the current measurement pair and a single sensor’s history
conditioned on the history of the other sensor, i.e.,
DpppZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq||qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θqq
1
j
i
|Z1:k´1
, Θq
ď IpZki , Zkj ; Z1:k´1
2
1
j
i
` IpZki , Zkj ; Z1:k´1
|Z1:k´1
, Θq. (31)
2
The proof can be found in Appendix B 2 . The upper bound
in (31) measures the departure of the current pair of measurements, and, one of the sensor histories from a state
of conditional independence when they are conditioned on
the history of the other sensor. Note that these variables,
when conditioned on Xk , are conditionally independent, i.e.,
j
pZki , Zkj qK
KZ1:k´1
|Xk , Θ holds and consequently
j
i
IpZki , Zkj ; Z1:k´1
|Xk , Θq “ IpZki , Zkj ; Z1:k´1
|Xk , Θq “ 0.
Similarly, the average MI term on the right hand
j
i
side of (31) is zero if pZki , Zkj qK
KZ1:k´1
|Z1:k´1
, Θ and
j
j
i
i
pZk , Zk qK
KZ1:k´1 |Z1:k´1 , Θ hold simultaneously. This condition is satisfied, for example, in the case that either of
2 Note that the results presented in this section are valid for any selection
i,j
of τ1:t
as they relate random variables which are conditioned on the data
association. The divergences and bounds, nevertheless, are more relevant
i,j
for τ1:t
“ τ̄ i,j
1:t .
6
j
i
the measurement histories Z1:k´1
and Z1:k´1
are sufficient
statistics for Xk (i.e., it can be predicted by both sensors with
probability one). This level of accuracy should not be expected
as the transition density of state space models introduce some
uncertainty. Therefore, it is instructive to relate the KLD in
(31) further to the uncertainty on Xk given the sensor histories:
Corollary 4.2: The KLD considered in Proposition 4.1 is upper bounded by the weighted sum of uncertainty reductions in
the local target prediction and posterior distributions achieved
when the other sensor’s history is included jointly:
DpppZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq||qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θqq
˜
¯
1 ´
j
j
i
HpXk |Z1:k´1
, Θq ´ HpXk |Z1:k´1
, Z1:k´1
, Θq
ď
2
¸
´
¯
j
i
i
` HpXk |Zk´1 , Θq ´ HpXk |Z1:k´1 , Z1:k´1 , Θq
˜
¯
1 ´
j
j
i
`
HpXk |Z1:k
, Θq ´ HpXk |Z1:k
, Z1:k´1
, Θq
2
¸
´
¯
j
i
i
` HpXk |Z1:k , Θq ´ HpXk |Z1:k , Z1:k´1 , Θq
, (32)
where H denotes the Shannon differential entropy [31].
The proof is provided in Appendix C. Corollary 4.2 relates the
approximation quality of the quad-term node-wise separable
updates to the uncertainties in the target state prediction and
posterior distributions when individual node histories and their
combinations are considered. The difference terms on the RHS
of (32) quantify the difference in uncertainty between estimating the target state Xk using only the local measurements,
and, also taking into account the other sensor’s measurements.
Overall, a better quality of approximation should be expected
when the local filtering densities involved concentrate around
a single point in the state space.
B. The quad-term pairwise likelihood
The quad-term update in (29) leads to a separable approximate likelihood given by
t
´
¯ ź
˜l Zi , Zj |θ “
qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq
1:t
1:t
(33)
k“1
We refer to this term as the quad-term separable likelihood
as it can also be expressed as a (scaled) product of four factors
each of which are the products of the four factors of (29)
over k. Let us define
k
rij
pZik , θq fi ppZik |Zj1:k , θq,
skj pZjk , θq fi ppZjk |Zj1:k´1 , θq.
Then, the quad-term update in (33) is given by
qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq “
¯1{2´
¯1{2
1 ´ k i
k
,
rij pZk , θqskj pZjk , θq
rji
pZjk , θqski pZik , θq
κk pθq
(34)
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
where the normalisation factor is given in (30), and, equivalently in terms of the four factors above as
ż ż ´
¯1{2
1
1
k
κk pθq “
rij
pZik , θqskj pZjk , θq
¯1{2
´
1
1
1
1
k
dZik dZjk .
ˆ rji
pZjk , θqski pZik , θq
Corollary 4.3: The KLD between the parameter likelihood
in (8) and the node-wise separable approximation in (33) is
bounded by the terms on the right hand sides of (31) and (32)
summed over k “ 1, . . . , t as
7
simplifications of the expressions involved which provide computational benefits. In particular, parameters such as respective
location and bearing angles relate the local coordinate frames
of the sensors which collect measurements in their local frame.
The local filtering distributions are hence over the space of
state vectors in the local frame. A point xk P X (Section II-A)
is implicitly in the Earth coordinate frame (ECF), and, associated with its representation in the jth local frame rxk sj through
a coordinate transform T with the following properties
rxk sj “ T pxk ; θj q,
(37)
´1
rxk si “ T pT prxk sj ; θj q; θi q.
t
´ ´
¯ ´
¯¯ ÿ
D l Zi1:t , Zj1:t |θ ||l̃ Zi1:t , Zj1:t |θ “
D pp||qq . (35) As an example, when xk is a location on the Cartesian plane,
and, θj is the position of sensor j, T is given by
k“1
Proof. Eq. (35) can easily be found after expanding the
KLD term explicitly and expressing the logarithm of products
involved as sums over logarithms of the factors. Boundedness
follows from non-negativity of KLDs and summing both sides
of (31) and (32) over k “ 1, ..., t.
As a conclusion, when t is not large – e.g., on the order
of tens which is typical in fusion applications– the proposed
approximation can be used for parameter estimation via local
filtering. Sensors that are more accurate in inferring the underlying state process result with a smaller KLD in (35), which
in turn leads to a more favourable estimation performance.
One other approximation based on local filtering distributions was studied in [26] which has the following dual-term
product form
upZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq
fi ppZik |Zj1:k´1 , θqppZjk |Zi1:k´1 , θq.
(36)
In Appendix D, we shown that Dpp||qq ă Dpp||uq, when
sensors are equivalent. The entropy bound given in (32) is also
smaller than that for the dual-term approximation. In other
words, the quad-term approximation is more accurate compared to the dual-term approximation, under typical operating
conditions.
The scaling factor of the dual term approximation is unity
regardless of θ, on the other hand, admitting a significant
amount of flexibility in the range of the distributions and
likelihoods that can be accommodated in the state space model.
For example, the dual-term pseudo-likelihood is used with
random finite set variables (RFS) in [26], [44], which, in a
sense, have the association variables marginalised out making
it possible to avoid multi-dimensional assignment problems in
the general multi-object tracking model. For RFS distributions,
however, it is not straightforward to compute the scaling
factor in (30) for the quad-term. In this article, we consider
a parametric model instead, which is effectively configured
through association variables.
V. Q UAD - TERM
LIKELIHOOD FOR SENSOR CALIBRATION
PARAMETERS
The results presented so far are fairly general and do not
depend on the nature of Θ. When Θ represents respective
parameters such as calibration parameters, there are certain
T pxk ; θj q fi xk ´ θj ,
T pT
´1
prxk sj ; θj q; θi q “ rxk sj ` θj ´ θi .
For simplicity in notation, we will denote T pT ´1 p.; θj q; θi q by
Tθ p.q when semantics is clear from the context3 .
In this section, it is revealed how local filtering distributions
are used in the quad-term update. When θ are respective quantities, these distributions become independent of θ because
both the state and the measurement variables are in the same
local coordinate frame. In other words, at sensor j
ppXk , τkj |Zj1:k , θj q ” pprXk sj , τkj |Zj1:k q,
(38)
holds for the filtering posterior, for any configuration of τkj .
In the prediction stage of filtering, the multi-object transition
kernel in (5) also becomes independent of Θ, so, the ChapmanKolmogorov equation for finding the prediction density at
sensor j (together with the empirical Bayes selection of
association priors during iterations as explained in Section II)
becomes
ż
j
pprXk sj |Z1:k´1 q “ πprXk sj |rXk´1 sj q
j
ˆ pprXk´1 sj , τk´1
“ τ̄ jk´1 |Zj1:k´1 qdrXk sj . (39)
Note also that the entropy terms in (32) that are conditioned
on a single sensor’s measurements measure the uncertainty of
the above densities. Consequently, they also become indepenj
j
dent of Θ, i.e., HpXk |Z1:k´1
, Θq equals to HpXk |Z1:k´1
q for
example, which highlights its relevance to the local prediction
accuracy.
A. The quad-term update for calibration
Now, let us expand the quad-term time updates in (34), and,
explicitly show the aforementioned simplifications. We start
with skj which is the scale factor of the local Bayesian filter
at sensor j:
skj pZjk , θq “ pτ j pZjk |Zj1:k´1 , θq,
ż k
“ lj pZjk |Xk , τkj , θqppXk |Zj1:k´1 , θqdXk ,
ż
“ lj pZjk |rXk sj , τkj qpprXk sj |Zj1:k´1 qdrXk sj , (40)
3 Note that, when the calibration parameters also include orientation angles,
Tθ p.q involves rotation matrices accordingly.
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
where in the last line independence from θ is asserted. Because
skj does not depend on θ, we denote skj pZjk , θq by skj pZjk q in
the rest of the article.
k
Next, let us consider rij
which has terms in different
coordinate frames:
k
rij
pZik , θq “ pτ i,j pZik |Zj1:k , θq
k
ż
i
“ li pZk |Xk , τki , θqppXk , τkj |Zj1:k , θqdXk
ż
¯
˘ ´
`
“ li Zik |T pXk ; θi q, τki p T ´1 pXk ; θj q, τkj |Zj1:k dXk
ż
¯
˘ ´
`
“ li Zik |Tθ prXk sj q, τki p rXk sj , τkj |Zj1:k drXk sj . (41)
In the third line above, the coordinate transformations are
substituted explicitly. The last line follows from that the
filtering distribution (38) is in the jth local frame.
8
given by
pm px1 |Zj1:k´1 q fi
skj pZjk q “
M
ź
j
skj,o pzk,o
q
(42)
o“1
j
skj,o pzk,o
q fi
ż
j
gj pzk,o
|x1k qpm1 px1k |Zj1:k´1 qdx1k
where m1 “ τ̂kj poq and the density in the integral is the m1 th
marginal of the local prediction density, the mth of which is
“
ppXk “ xk,1 , . . . , xk,m´1 , x1 , xk,m`1 ,
xk,m`2 , . . . , xk,M s ,τkj “ τ̂kj |Zj1:k´1 q
dxk,1 . . . dxk,m´1 dxk,m`1 . . . dxk,M .
k
k
The term rij
in (41) (or, rji
) is also computed based on these
local filtering distributions. The integration in the RHS of (41)
implicitly assumes that the ordering of individual objects in the
local multi-object vectors are the same. In a networked setting,
however, this is not necessarily the case and the identities of
the fields in the state vector may differ [45]. In order to tackle
with this unknown correspondance, we introduce an additional
permutation random variable γk (see, e.g., [46]) for relating
the fields of a multi-object vector Xk as ordered locally at
sensor i and j, such that the mth field of the state vector at
sensor i refers to the same object in the γk pmqth field of the
state vector at sensor j. For example,
B. Evaluation of the quad-term update based on single sensor
filtering distributions
Here, we discuss the evaluation of the quad-term likelihood
given local filtering distributions and single sensor association
configurations which we denote for sensor j by
pprXk sj , τkj “ τ̂kj |Zj1:k q and τ̂kj , respectively, as explained in
Section II-B.
Instead of considering evaluation for the most probable
association hypothesis τ̄ i,j
k which is infeasible to find, we
propose to use the local results τ̂ki,j fi pτ̂ki , τ̂kj q as a reasonable approximation to this configuration and substitute them
in (40)–(41). These approximations can be found regardless
of θ as discussed earlier in this section by using one of the
well studied algorithms in the literature [2] such as solving
a 2 ´ D association problem at each time step [35] to find
τ̂kj . We detail this approach for a linear Gaussian state space
model in Section VI.
Given these local results and their exchange over the network, one can consider an in-network computation scheme
for evaluating (40)–(41). Specifically, these terms (and, the
other factors of the quad-term update which are obtained by
replacing i and j in these expressions) can be found at the
sensor platform where the measurements to be substituted for
evaluation are stored, i.e., sensors j and i, respectively, for
k
. Substitution of the measurement histories on the
skj and rij
conditioning side will have been carried out by local filtering.
More explicitly, skj in (40) (or, ski ) becomes a product of
similar terms when τ̂jk is substituted in (40), and, its computation is carried out during the local filtering of sensor j’s (or,
sensor i’s) measurements using
ż
rxk,m si “ Tθ prxk,γk pmq sj q.
Suppose that an estimate γ̂k of this quantity is provided.
After substituting in (41) together with τ̂ki,j one obtains
k
rij
pZik , θq “
M
ź
k
i
rij,o
pzk,o
, θq,
(43)
o“1
k
i
rij,o
pzk,o
, θq fi
ż
` i
˘
gi zk,o
|Tθ px1k q pm1 px1k |Zj1:k qdx1k
where m1 “ γ̂k pτ̂ki poqq and the density inside the integral is the
m1 th marginal of the filtering distribution local to sensor j. In
Appendix E, we show that (43) replaces the likelihood for γk
when m1 “ γk pτ̂ki poqq and an ML estimate γ̂k can be found in
a way similar to solving the data association problem in local
filtering, which is detailed later in Section VI-A.
Finally, the scale factor (30) is computed. This involves finding the measurement distributions in (30) using the prediction
distribution in (39) for both sensors i and j together with their
likelihoods. This leads to the following two decomposition:
The first term in (30) is found as
1
1
ppZik , Zjk |Zi1:k´1 , θq “
M
ź
1
1
po pz ik,o , z jk,ξpoq |Zi1:k´1 , θq
żo“1
po pz i , z j |Zik´1 , θq fi gi pz i |x1k qpm1 px1k |Zi1:k´1 qdx1k
ż
(44)
ˆ gj pz j |Tθ ´1 px1k qqpm1 px1k |Zi1:k qdx1k
where ξpoq “ τ̂kj´1 ˝ τ̂ki poq maps the oth measurement at
sensor i to the corresponding one in sensor j, and, m1 “ τ̂ki poq
in the second line.
The second term in (30) is found as
1
1
ppZik , Zjk |Zj1:k´1 , θq “
i
po pz , z
j
|Zjk´1 , θq
M
ź
1
1
po pz ik,o , z jk,ξpoq |Zjk´1 , θq
żo“1
fi gj pz j |x1k qpm1 px1k |Zj1:k´1 qdx1k
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
ˆ
ż
gi pz i |Tθ px1k qqpm1 px1k |Zj1:k q dx1k
(45)
where m1 “ γ̂k ˝ τ̂ki poq in the last line is the object that
corresponds to the oth measurement at sensor i.
Consequently, the scale factor is found as
κk pθq “
M
ź
κk,o pθq
(46)
żo“1
`
po pz i , z j |Zi1:k´1 , θq
κk,o pθq fi
¯1{2
dz i dz j ,
ˆpo pz i , z j |Zj1:k´1 , θq
using the densities found in (44) and (45).
The expressions above describe the evaluation of the quadterm likelihood in terms of single sensor filtering distributions
that can be obtained using any filtering algorithm with individual measurement histories. The use of the local filtering
distributions provides scalability with the number of sensors
for parameter estimation in the state space model in Section II.
These computations can be distributed in the pair pi, jq
as follows: Both sensors i and j perform local filtering and
exchange the resulting posterior densities at every step, as
well as si and sj , respectively, found using (42). Based
on the received densities, sensors i and j evaluate rij and
rji , respectively, using (43). As part of the filtering process,
they realise the Chapman-Kolmogorov equation in (39) with
their local posterior, as well as the remote posterior recently
received. These densities are then used in (44), (45), and, (46)
to compute the scale factor for the next step. The scale factor
hence found in the previous step, therefore, is substituted
in (34) together with the four terms computed. This is repeated
for k “ 1, . . . , t, and, the quad-term likelihood (33) is
computed, as a result.
VI. A M ONTE C ARLO LBP ALGORITHM FOR SENSOR
CALIBRATION IN LINEAR G AUSSIAN STATE SPACE MODELS
In this section, we consider a linear Gaussian state
space (LGSS) model within the probabilistic graphical model
in Fig. 1, and, specify an algorithm for estimation of θs
that combines the quad-term calibration likelihood evaluation
detailed in Section V with BP message passing on the resulting
pairwise MRF model (Section III). This algorithm uses Monte
Carlo methods for realising BP [32] and facilitates scalability
by building upon single sensor filtering as required by the
quad-term approximation. As a result, an efficient inference
scheme over the model in Fig. 1 is achieved.
The state space model we consider is specified by a linear
state transition with process noise that is additive and Gaussian, and, linear measurements with independent Gaussian
measurement noise, i.e.,
πpxk |xk´1 q “ N pxk ; Fxk´1 , Qq
(47)
gj pzkj |xk ; θj q “ N pzkj ; Hj rxk sj , Rj q
(48)
for j “ 1, . . . , N , where N p.; µ, Pq is a multi-dimensional
Gaussian density with mean vector µ and covariance matrix P.
Here, xk is the concatenation of position and velocity (on
a 2 ´ D Euclidean plane, without loss of generality). The
9
matrices F and Q model motion with unknown acceleration
(equivalently, manouevres), and, are selected as
«
ff
«
ff
I, ∆T ˆ I
q1 I, q2 I
2
F“
, Q“σ
0,
I
q2 I, q3 I
where I and 0 are the 2 ˆ 2 identity and zero matrices,
respectively. ∆T is the time difference between consecutive
steps. Q is positive definite and parameterised with σ 2 , and,
0 ă q1 ă q2 ă q3 ă 1 specifying the magnitude of
the uncertainty, and, contributions of higher order terms,
respectively 4 .
In the measurement model, Rj is the measurement noise
covariance, and, Hj is the observation matrix which we
assume forms an observable pair with F (e.g., Hj “ rI, 0s).
A. Local single sensor filtering
We now focus on filtering and provide explicit formulae that
adopts the recursions in (13)–(18) for a single sensor in the
LGSS model. We use the empirical Bayes approach explained
in Section II-B for scaling with time under data association
uncertainties. This approach corresponds to the single frame
data association solution in multi-object tracking [35].
First, let us consider the prediction stage at time k in which
we are given a filtering density evaluated at the most likely
j
data association hypothesis τ̂k´1
of the previous step 5 . The
latter is a product of its marginals
j
j
ppXk´1 , τk´1
“ τ̂k´1
|Zj1:k´1 q “
M
ź
j
j
pm pxk´1,m , τk´1
“ τ̂k´1
|Zj1:k´1 q
m“1
where
j
j
j,m
pm pxk´1,m , τk´1
“ τ̂k´1
|Zj1:k´1 q “ ppxk´1,m |z1:k´1
q.
i,m
Here, z1:k´1
denotes the measurements induced by object m
from step 1 to k ´ 1, i.e.,
˙
ˆ
j,m
j
j
j
,
z1:k´1
fi zk´1,ρ
,
z
,
.
.
.
,
z
j
pmq k´2,ρj pmq
1,ρj pmq
k´1
k´2
1
where ρ is the inverse of τ , i.e., ρ˝τ is the identity permutation.
Consequently, the mth marginal of the posterior at k ´1 is a
Gaussian density that can be obtained equivalently by Kalman
j,m
filtering [36] over z1:k´1
, i.e.,
j,m
pm pxk´1,m |z1:k´1
q “ N pxk´1,m ; x̂jk´1,m , Pjk´1,m q,
(49)
with the mean and covariance matrices over time specifying
the mth “track.”
4 It can easily be shown that this state transition model is invariant under
the selected coordinate frame for xk s.
5 Note that the empirical prior on τ j
k´1 is selected as in (16) leading to this
j
j
.
other than τ̂k´1
posterior be identically zero for all values of τk´1
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
Hence, the prediction density in (39) (Section V-A) evaluj
j
ated at τk´1
“ τ̂k´1
for the state transition in (47) is given by
ppXk |Zj1:k´1 q “
x̂jk|k´1,m
Pjk|k´1,m
“
“
M
ź
N pxk,m ; x̂jk|k´1,m , Pjk|k´1,m q (50)
m“1
Fx̂jk´1,m ,
FPjk´1,m FT
` Q,
where the last two lines are Kalman prediction equations with
p.qT denoting matrix transpose.
Next, let us consider the update stage in which we use the
prediction density (50) with M measurements concatenated in
Zjk and the measurement likelihood. This likelihood is found
by substituting (48) in (6). First, we use this term within
the likelihood for the association variable τkj which –as it is
j
mutually independent from τ1j , . . . , τk´1
– is given by
lj pZj1:k |τkj q 9 pτ j pZjk |Zj1:k´1 q
żk
“ lpZjk |Xk , τkj qppXk |Zj1:k´1 qdXk
“
“
M ż
ź
o“1
M ż
ź
(51)
j
gj pzk,o
|x1k qpτ j poq px1k |Zj1:k´1 qdx1k
k
j,τ j poq
o“1
The first line above follows from that the joint distribution of
j
Z1:k´1
and τkj is independent of the latter.
The prior distribution for τkj is non-informative as given
in (7), so, the ML estimate using (51) coincides with the MAP
estimate and it is given by
τkj PSM
(52)
where SM is the set of M -permutations.
An equivalent problem is found by taking the logarithm
of the objective function in the combinatorial optimisation
problem above as follows:
M
ÿ
`
˘
c o, m “ τkj poq
τ̂kj “ arg max
j
(53)
τk PSM o“1
cpo, mq fi log
ż
j
j,m
gj pzk,o
|x1k qppx1k |z1:k´1
qdx1k ,
for o, m “ 1, . . . , M .
This cost for the LGSS model is explicitly found using the
prediction distribution (50) and the measurements within the
KF innovations [36] as
j
j
cpo, mq “ log N pzk,o
; ẑk,m
, Sjk,m q
j
ẑk,m
“
Hj x̂jk|k´1,m ,
Sjk,m
“ Rj `
putation of the M 2 cost matrix usually has the predominant
computational time.
Next, we consider the state distribution update (see (15) in
Section II-B) and assert the empirical (model) prior in (16).
As a result, the filtering density at k becomes a product of its
marginals each of which is a Gaussian as in (49) found by the
KF update [36], i.e.,
ppXk , τkj “ τ̂kj |Zj1:k q “
M
ź
N pxk,m ; x̂k,m , Pk,m q
(55)
m“1
where,
j
x̂k,m “ x̂k|k´1,m ` Kk,m pzk,ρ
´ Hj x̂k|k´1,m q
j
pmq
k
Pk,m “
Kk,m fi
pI ´ Kk,m Hj q Pjk|k´1,m
´1
Pjk|k´1,m HTj Sjk ,m .
Note that, because the posterior density is now non-zero only
for τkj “ τ̂kj , the marginalisation over τkj (see, e.g., (14)) in
the following prediction stage reduces to (39) and (50).
B. Evaluation of the calibration quad-term in the LGSS Model
j
k
gj pzk,o
|x1k qppx1k |z1:k´1
qdx1k .
τ̂kj “ arg max lj pZj1:k |τkj q
10
(54)
Hj Pjk|k´1,m HTj .
The optimisation in (53) is a 2 ´ D assignment problem
which can be solved in polynomial time with M (despite that
the search space SM has a factorial size) using one of the
well known solvers [47] including the auction algorithm [48].
This algorithm operates over a matrix of costs obtained by
C “ rcpo, mqs to iteratively find the M pairs corresponding
to the best permutation τ̂kj in the ML problem (52). Here, com-
Let us consider the evaluation of skj given by (42) using the
formulae for the LGSS model introduced in Section VI-A. By
comparison with the cost term in (53) and (54), it can easily
be seen that skj is the exponential of the association cost for
τ̂kj , i.e.,
M
ÿ
`
˘
skj pZjk q “ exp
c o, τ̂kj poq .
(56)
o“1
k
Next, let us consider (43) for evaluating rij
. Evaluation of
this term involves finding the object identity correspondance
γ̂k by solving a 2-D assignment as explained in Appendix E.
In the LGSS model, the assignment cost matrix D “ rdpo, mqs
is found as
i
i
dpo, mq “ log N pzk,o
; ẑk,m
, Sik,m q
i
ẑk,m
Sik,m
“
(57)
Hi Tθ px̂jk,m q,
“ Ri ` Hi Tθ pPjk,m qHTi .
for o, m “ 1, . . . , M . Here, the second order statistics Pjk,m
is also transformed by applying any rotations involved in Tθ
to its eigenvectors. The best assignment which here encodes
γ̂k is found using the auction algorithm [48] as well, similar
to the assignment in the Bayesian filtering update (Sec. VI-A).
Using this estimate, the quad-term factor is computed using
k
rij
pZik , θq “ exp
M ´
ÿ
`
˘¯
d o, γ̂k τ̂ki poq .
(58)
o“1
In order to evaluate the other factors of the quad-term upk
, similar computations are used. It suffices
date, i.e., sik and rji
to replace i in the subscripts/superscripts of the expressions
above with j, and, vice versa.
Finally, let us consider the scale factor in (46). Let us use the
notation introduced in the previous section for expressing the
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
densities inside the integration. Starting with (44), one obtains
po pz i , z j |Zi1:k´1 , θq “
“
µ1 “
Σ1 “
i
i,τ̂k´1
poq
ppz i , z j |z1:k´1
, θq
i j T
N prz , z s ; µ1 , Σ1 q
ff
«
i
ẑk,m
Hj Tθ´1 px̂ik,m q
«
(59)
0
Rj ` Hj Tθ´1 pPik,m qHTj
ff
i
i
where m “ τ̂k´1
poq, and, ẑk,m
and Sik,m are computed as
in (54) with i substituted in place of j.
The second density in (46) conditioned on sensor j’s history,
i.e., (45), is similarly found as
po pz i , z j |Zj1:k´1 , θq “ N prz i , z j sT ; µ2 , Σ2 q
«
ff
Hi Tθ px̂jk,m q
µ2 “
j
ẑk,m
«
Ri ` Hi Tθ pPjk,m qHTi
Σ2 “
0
Consider the BP message from node j to i in (24). Suppose
that independent identically distributed (i.i.d.) samples from
the (scaled) product of the jth local belief and the incoming
messages from all neighbours except i are given, i.e.,
ź
plq
θ̄j „ p̃j pθj q
mi1 j pθj q for l “ 1, ..., L.
(63)
i1 Pnepjq{i
Sik,m
0
11
These samples are used with kernel approximations in order
to represent the message from node j to i (scaled to one), in
the NBP approach [32]. We use Gaussian kernels leading to
the approximation given by
m̂ji pθi q “
L
ÿ
plq
plq
ωji N pθi ; θji , Λji q,
(64)
l“1
plq
plq
plq
plq
θji “ T pT ´1 pθ̄j ; θj q; θi q,
plq
plq
plq
ωji “ řL
ψi,j pθi , θj q
l1 “1
0
Sjk,m
ff
j
i
poq and, ẑk,m
and Sjk,m are given
where m “ γ̂k´1 ˝ τ̂k´1
in (54).
Using the densities above and integration rules for Gaussians, the oth term of the scale factor in (46) is found as
κk,o pθq “
`ˇ ´1 ˇ ˇ ´1 ˇ˘1{4
"
ˇΣ ˇ ˇΣ ˇ
˘
1 ` T ´1
2
1
T ´1
ˇ ´1 ´1 ˇ1{2 exp ´ 4 µ1 Σ1 µ1 ` µ2 Σ2 µ2
ˇ Σ1 `Σ2 ˇ
ˇ
ˇ
2
˘T ` ´1
˘´1
1`
´1
Σ1 ` Σ´1
` Σ´1
1 µ1 ` Σ 2 µ2
2
4
`
˘(
´1
ˆ Σ´1
. (60)
1 µ1 ` Σ 2 µ2
In a distributed setting, the scale factor expressions above
are computed both at sensors i and j, for which the prediction
stage given in (50) is carried out for both the local posterior
and the posterior recevied from the other sensor at time k ´ 1.
pl1 q
pl1 q
ψi,j pθi , θj q
,
where the kernel weights are the normalised edge potentials.
Λji is related to a bandwidth parameter that can be found using
Kernel Density Estimation (KDE) techniques. In particular, we
use the rule-of-thumb method in [50] and find
ˆ
˙2{pd`4q
4
Λji “
Ĉji ,
p2d ` 1qL
ÿ ÿ pl1 q plq pl1 q
plq
ωji ωji pθji ´ m̂ji qpθji ´ m̂ji qT ,
Ĉji “
m̂ji “
l1 l
L
ÿ
plq plq
ωji θji
l“1
where m̂ji and Ĉji are the empirical mean and covariance of
the samples, respectively, and d is the dimensionality of θji s.
Given these messages, let us consider sampling from the
updated marginal in (25). We use the weighted bootstrap (also
known as sampling/importance resampling) [51] with samples
generated from the (scaled) product of Gaussian densities
with mean and covariance found as the empirical mean and
covariance of the particle sets, respectively. In other words,
given m̂ji and Ĉji as above, we generate
plq
C. Sampling from the calibration marginals using nonparametric BP
θi „ f pθi q,
l “ 1, . . . , L,
ź
f pθi q 9 N pθi ; m̂i , Ĉi q
N pθi ; m̂ji , Ĉji q.
jPnepiq
In this section, we introduce particle based representations and Monte Carlo computations [49] for the realisation of (loopy) BP message passings. Note that Sections VI-A and VI-B specify the evaluation of edge potentials
j
j
i
i
given in (33) for pτ1:t
, τ1:t
q “ pτ̂1:t
, τ̂1:t
q.
For sampling from the marginal parameter posteriors, we
adopt the approach detailed in [26, Sec.VI] for carrying out
LBP belief update and messaging in (25) and (24), respectively. Given L equally weighted samples from p̃i pθi q, i.e.,
plq
θi
„ p̃i pθi q,
(61)
for l “ 1, . . . , L, the edge potentials are evaluated to obtain
¯
´
plq plq
plq plq
(62)
ψij pθi , θj q “ l̃ Zi1:t , Zj1:t |θ “ pθi , θj q .
The particle weights for these samples to represent the
updated marginal is given by
plq
plq
ωi “ ω̂i {
L
ÿ
pl1 q
ω̂i
1
plq
ω̂i
l “1
´
¯
ź
plq
plq
plq
“ p0,i pθi q
m̂ji pθi q {f pθi q
jPnepiq
where p0,i is the prior density selected for θi (and, the node
potential in (22)). Thus, the local calibration marginal is
estimated by
P̂i pdθi q “
L
ÿ
l“1
plq
ωi δθplq pdθi q.
i
(65)
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
Algorithm 1 Pseudo-code for estimation of θ using the quadterm separable likelihood within Belief Propagation.
2:
3:
Find ppXk , τkj “ τ̂kj |Zj1:k q in (55) as described in Section VI-A
4:
Find
5:
end for
in (56)
1000
S14
S13
T4
S5
500
0
-1000
-1500
-1500
S4
S3
S1
S2
S12
T1
S6
S11
T2
S7
-1000
S8
-500
0
S9
500
1000
S10
1500
2000
2500
East (m)
7: for all j P V do
Sample
S15
T3
1500
-500
6: end for
8:
S16
2000
Ź Local filtering
for k “ 1, . . . , t do
skj pZjk q
2500
North (m)
1: for all j P V do
12
plq
θi
Ź Sample from priors
„ p0,i pθi q for l “ 1, . . . , L as in (61)
Fig. 2. Example scenario: 16 Sensors collect measurements from 4 objects
(T1-T4) with association uncertainties. Initial positions of the objects are
denoted by black squares. Trajectories for 60 time steps are depicted. The
blue lines depict the edges of the MRF model used for estimation.
9: end for
10: for s “ 1, ..., S do
11:
12:
Ź S-steps of LBP
for all pi, jq P E do
Ź Evaluate edge potentials
for l “ 1, . . . , L do
plq
plq
plq
plq
13:
k pZi , θ “ pθ , θ qq using (57), (58) for k “ 1, . . . t
Find rij
i
j
k
14:
k pZj , θ “ pθ , θ qq for k “ 1, . . . t
Find rji
i
j
k
15:
Find
16:
Find
plq plq
κk pθ “ pθi , θj qq using (46), (59)–(60) for
plq plq
qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θ “ pθi , θj qq
k “ 1, . . . , t
using (34) for
k “ 1, . . . , t
17:
18:
plq
end for
19:
end for
20:
for all pi, jq P E do
21:
Ź Find LBP message
Find the kernel representation m̂ji pθi q in (64)
22:
end for
23:
for all i P V do
24:
plq
Find ψi,j pθi , θj q in (62) as the quad-term likelihood in (33)
Find the updated P̂i in (65) and sample
plq
1 řL
l“1 θi
L
25:
θ̂i Ð
26:
end for
Ź Update local marginals
plq
θi
„ p̃i pθi q
27: end for
plq
plq
As the final step of the bootstrap, tθi , ωi uM
l“1 is resampled (with replacement) leading to equally weighted particles
plq
from p̃i pθi q, i.e., tθi uL
l“1 . We follow similar bootstrap steps
in order to generate the samples in (63).
After nodes iterate the BP computations described above
for S times, each node estimates its location by finding the
plq
empirical mean of tθi uL
l“1 . These steps are summarised in
Algorithm 1.
θ “ rθ1 , . . . , θ16 s with the prior distribution for θ1 selected as
Dirac’s delta, i.e., p0,1 pθ1 q “ δpθ1 q. For the other nodes, the
localisation prior, i.e., p0,i pθi q for i “ 2, . . . , 16, is a uniform
distribution over the sensing region.
We use Algorithm 1 specified in Section VI for estimating θ.
The MRF model we consider is specified by the pairwise graph
G in Fig. 2 (blue edges). We use L “ 100 points in (61) to
represent the local belief densities. We select the sensor data
time window length as t “ 10 (starting at time 21 until 30 in
the scenario in Fig. 2). We follow the steps in Algorithm 1
for S “ 16 iterations.
A typical run is illustrated in Fig. 3. Here, the scatter plot
of particles from marginal posteriors are given over iterations.
Note that the network coordinate system is established by
sensor 1 through its informative prior, and, LBP emanates this
information towards the outer nodes while learning the edge
potentials using node-wise separable likelihood evaluations
plq plq
ψi,j pθi , θj q given in (23) and (33).
For performance assesment, first, we consider the mean
squared error for θ̂ output by our algorithm, as we have
built our discussion on MMSE estimators in (12). We find
this value empirically by taking the average of the squared
norm of estimation errors over 100 Monte Carlo simulations.
In Fig. 4, we present a semi-log plot of this quantity over
iterations (blue line). Note that, convergence occurs in less
VII. E XAMPLE : S ELF - LOCALISATION IN LGSS MODELS
In this example, we demonstrate the quad-term node-wise
separable likelihood in sensor self-localisation. The LGSS
model given by (47) and (48) is used with process noise
parameters selected as σ “ 0.5, q1 “ 1{4, q2 “ q3 “ 1{2,
q4 “ 1. The measurement model for sensor i is given by
Hi “ rI, 0s and Ri “ σn2 I with σn “ 10 modelling noisy
position measurements in the local coordinate frame.
Let us consider the multi-object multi-sensor scenario depicted in Fig. 2. 16 sensors observe 4 objects moving with
data association uncertainties. The locations of the sensors are
to be estimated with respect to sensor 1 which is selected
as the origin of the network coordinate system. Therefore
Fig. 3. Node beliefs in LBP iterations: Marginal posterior estimates after
iteration 4 (upper-left), 6 (upper-right), 8 (lower-left), and, 10 (lower-right).
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
105
q u a d -t e r m
d u a l-t e r m
log MSE
104
103
102
4
6
8
10
12
14
16
iteration
Fig. 4. Log-normalised error margin versus the iteration number n.
than ten iterations which is a favourable feature. We compare
this algorithm with the RFS based dual-term pseudo-likelihood
proposed in [26]. When evaluating this term, we use a Poisson
multi-object model output by using the Gaussian mixture
probability hypothesis density (GM-PHD) filter [52] with
the LGSS model. The averaged MSE performance for the
case that dual-term likelihoods are used as edge potentials is
depicted with the green dashed line in Fig. 4. The quad-term
approximation is seen to provide faster convergence with both
pseudo-likelihoods leading to an on par accuracy in the steady
regime, in this example. The edge update time for the quadterm update averages to 0.601 per edge per particle compared
to 1.312 for the dual-term update demonstrating its relative
efficiency.
Note that, the MSE is a network-wide term and the local
error norms are smaller. The localisation miss-distance averaged over sensors is given in Fig. 5. The average error (˘
one standard deviation) in the final step is 2.60 ˘ 0.70m with
a maximum value of 4.96 which is less than 0.5% of the
edge distances of 1000m. These results demonstrate that the
proposed scheme is capable of providing self-localisation with
favourable accuracy and small error margins.
13
it using local filtering operations, hence, scale with the number
of sensors.
In order to use the proposed approximation in the case
of multiple objects, we employed a parameterised multiobject state space model in which different configurations
of the parameter specify different hypothesis of object-tomeasurement and object-to-object associations. Specifically,
we introduced an empirical Bayesian perspective for evaluating separable likelihoods in this model, using only local
Bayesian filtering. This approach substitutes the estimates of
the hypothesis variables when evaluating the likelihood of the
latent parameter, and, decouples these two inference tasks if
the hypothesis variables can be estimated locally. Therefore, it
can be extended to general hypothesis variables that capture,
for example, variable number of objects Mk , less than one
probability of detection, i.e., PD ă 1, and, measurements with
false alarm, by adapting the assignment problems involved
accordingly (see, for example [35]).
The associated posterior distribution is a MRF over which
distributed inference is possible using message passing algorithms such as LBP. We specified a particle message passing
algorithm for sampling from latent parameter marginals for
a linear Gaussian state space model with multiple objects.
This algorithm is demonstrated in simulations for sensor selflocalisation using point measurements from non-cooperative
objects in a potentially GPS denying environment. It is
possible to estimate unkown orientation angles as well, by
appropriately defining the transform in (37) and taking into
account the rotations implied by θi s in the LBP steps of
Algorithm 1 (additional details can be found in [53]).
A PPENDIX
A. Empirical Bayes parameter likelihood
VIII. C ONCLUSIONS
In this work, we have addressed the prohibitive complexity
of latent parameter estimation in state space models when there
are many sensors collecting measurements. We proposed a
pseudo-likelihood, namely the quad-term node-wise separable
likelihood, as an accurate surrogate to the actual likelihood
which is extremely costly to evaluate. The separable structure
of this quad-term approximation makes it possible to evaluate
35
average error norm (m)
30
25
20
15
The empirical Bayes parameter posterior follows from the
decomposition of the posterior in (11) using the chain rule of
probabilities as
ppθ|Z11:t , . . . , ZN
(66)
1:t q “
ÿ
ÿ
1
N
1:N
1:N
1
N
ppθ|Z1:t , . . . , Z1:t , τ1:t qppτ1:t |Z1:t , . . . , Z1:t q.
¨¨¨
1
τ1:t
N
τ1:t
The first term inside the summations is the parameter posterior conditioned on the association variables and the second
term is similar to a prior with respect to the first term. The
fact that this term is conditioned on the measurements makes
an empirical selection possible as discussed in Section II-B.
Let us use a similar empirical prior selection approach as used
in (16), i.e.,
10
1:N
ppτ1:t
|Z11:t , . . . , ZN
1:t q “
5
0
4
6
8
10
12
14
16
iteration
Fig. 5. Localisation miss-distance averaged over nodes versus the iteration
number. 100 Monte Carlo runs displayed with the boxes centered at the
median (red). Edges (blue) indicate the 25th and 75th percentiles.
t
ź
ppτk1:N |Z1k , . . . , ZN
k q
k“1
1:N
ppτk1:N |Z1k , . . . , ZN
pτk´1
q.
k q Ð δτ̄ 1:N
k´1
(67)
After substituting from (67) in (66), the parameter posterior is
found as
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
ppθ|Z11:t , . . . , ZN
1:t q
1:N
1:N
“ ppθ|Z11:t , . . . , ZN
1:t , τ1:t “ τ̄ 1:t q
1:N
1:N
9 lpZ11:t , . . . , ZN
1:t |θ, τ1:t “ τ̄ 1:t qppθq
where the likelihood in the last line is given by (17) and (18)
in Section II-B.
B. Proof of Proposition 4.1
Proof. Let us expand the KLD term in (31) by substituting its
arguments given in (9), (29) and (30):
¯
´
D ppZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq||qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq
ż
“ dZi1:k dZj1:k dθ ppZi1:k , Zj1:k , θq
ˆ log
“
ż
ppZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq
qpZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq
dZi1:k dZj1:k dθ ppZi1:k , Zj1:k , θq
˜
ppZik , Zjk , Zi1:k´1 |Zj1:k´1 , θq
1
ˆ
log
2
ppZik , Zjk |Zj1:k´1 , θqppZi1:k´1 |Zj1:k´1 , θq
` log
ppZik , Zjk , Zj1:k´1 |Zi1:k´1 , θq
14
and,
j
i
IpZki , Zkj ; Z1:k´1
|Z1:k´1
, Θq “
j
j
i
i
IpZki ; Z1:k´1
|Z1:k´1
, Θq ` IpZkj ; Z1:k´1
|Z1:k
, Θq. (72)
The MI terms on the RHSs of the equations above are for
random variables which form Markov chains with the current
the state variable Xk . Consider the (conditional) chains Zkj Ø
j
j
i
i
Xk Ø Z1:k´1
|Z1:k´1
, Θ and Zki Ø Xk Ø Z1:k´1
|Z1:k´1
,Θ
for the RHS of Eq.(71). The Data Processing Inequality [31]
applied to these terms lead to
j
i
IpZki , Zkj ; Z1:k´1
|Z1:k´1
, Θq
j
j
i
i
|Zkj , Z1:k´1
, Θq
ď IpXk ; Z1:k´1
|Z1:k´1
, Θq ` IpXki ; Z1:k´1
j
j
i
“ HpXk |Z1:k´1
, Θq ´ HpXk |Z1:k´1
, Z1:k´1
, Θq
j
`HpXk |Zkj , Z1:k´1
, Θq
´
(73)
j
i
HpXk |Zkj , Z1:k´1
, Z1:k´1
, Θq
A similar break down of Eq.(72) results with
j
i
IpZki , Zkj ; Z1:k´1
|Z1:k´1
, Θq
j
i
i
ď HpXk |Z1:k´1
, Θq ´ HpXk |Z1:k´1
, Z1:k´1
, Θq
i
`HpXk |Zki , Z1:k´1
, Θq
´
(74)
j
i
HpXk |Zki , Z1:k´1
, Z1:k´1
, Θq.
Substituting from (73) and (74) into (31) results with (32).
ppZik , Zjk |Zi1:k´1 , θqppZj1:k´1 |Zi1:k´1 , θq
¯
`2 log κk pθq
(68) D. Comparison of the quad-term and dual-term updates
´
1
j
Let us compare the KLDs of the quad-term and dual-term
i
“
IpZkj , Zki ; Z1:k´1
|Z1:k´1
, Θq
2
updates. The KLD of the dual-term in (36) is given by [26]
¯
j
i
`IpZkj , Zki ; Z1:k´1
|Z1:k´1
, Θq ` Etlog κk pθqu. (69)
DpppZik , Zjk |Zi1:k´1 , Zj1:k´1 , θq||upZik , Zjk |Zi1:k´1 , Zj1:k´1 , θqq
j
j
i
i
In the equations above, κk pθq is a normalisation constant
“ IpZkj ; Z1:k´1
|Z1:k´1
, Θq ` IpZki ; Z1:k´1
|Z1:k´1
, Θq
given by (30). Eq.(68) is obtained after multiplying both
j
j
i
i
` IpZk ; Zk |Z1:k´1 , Z1:k´1 , Θq. (75)
the numerator and the denominator of the quotient inside
j
j
i
i
the logarithm by ppZ1:k´1 |Z1:k´1 , θqppZ1:k´1 |Z1:k´1 , θq and The MI terms on the RHS of the above equation can be
a rearrangement of the terms. The definition of MI [31] expanded using that IpA; B|Cq “ HpA|Cq ´ HpA|C, Bq [31].
results with the first two terms in Eq.(69). The last term is The third term can be expanded in alternative ways as follows:
the expectation of the normalisation constant over the joint
j
i
j
i
IpZki ; Zkj |Z1:k´1
, Z1:k´1
, Θq
distribution of the sensor histories Z1:k´1
and Z1:k´1
, and, Θ.
(a)
j
j
i
i
i
Let us now consider the normalisation constant:
, Z1:k
, Θq (76)
“ HpZk |Z1:k´1 , Z1:k´1 , Θq ´ HpZki |Z1:k´1
ż
¯1{2
´
j
j
j
j
(b)
i
i
i
i
j
i
i
i
κk pθq “ dZk dZk ppZk , Zk |Z1:k´1 , θqppZk , Zk |Z1:k´1 , θq
“ HpZkj |Z1:k´1
, Z1:k´1
, Θq ´ HpZkj |Z1:k´1
, Z1:k
, Θq (77)
˙1{2
ˆż
After decomposing the first two terms on the RHS of (75)
ď
dZik dZjk ppZik , Zjk |Zj1:k´1 , θq
similarly and adding to the average of (76) and (77) (which
equals to the third term), we obtain
ˆż
˙1{2
ˆ
dZik dZjk ppZik , Zjk |Zi1:k´1 , θq
“ 1.
The inequality (70) follows from Hölder’s Inequality. Consequently, the last term in (69) is non-positive, and, (31) is
obtained.
C. Proof of Corollary 4.2
Proof. We apply the chain rule of information to the MI terms
on the RHS of (31) leading to
j
i
IpZki , Zkj ; Z1:k´1
|Z1:k´1
, Θq “
j
j
i
i
IpZkj ; Z1:k´1
|Z1:k´1
, Θq ` IpZki ; Z1:k´1
|Z1:k
, Θq, (71)
(70)Dpp||uq “ ∆u` ´ ∆u´
`
i
HpZkj |Z1:k´1
, Θq
(78)
j
HpZki |Z1:k´1
, Θq
`
”
1
j
i
, Z1:k´1
, Θq
∆u´ fi ´ HpZki |Z1:k´1
2
j
i
`HpZkj |Z1:k´1
, Z1:k´1
, Θq
∆u fi
ı
j
j
i
i
`HpZki |Z1:k´1
, Z1:k
, Θq ` HpZkj |Z1:k
, Z1:k´1
, Θq
Now, let us consider (31). After substituting (71) and (72)
in the RHS and expanding the MI terms as above, one obtains
Dpp||qq ď ∆q ` ´ ∆u´
(79)
”
1
j
i
∆q ` fi
HpZkj |Z1:k´1
, Θq ` HpZki |Z1:k´1
, Θq
2
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
ı
j
i
`HpZkj |Z1:k
, Θq ` HpZki |Z1:k
, Θq
Now, let us compare (79) and (78): The negative weighted
terms are equal, so, the difference of the positive weighted
terms are considered:
j
i
∆u` ´ ∆q ` “ HpZkj |Z1:k´1
, Θq ´ HpZkj |Z1:k´1
, Θq (80)
j
i
`HpZki |Z1:k´1
, Θq ´ HpZki |Z1:k´1
, Θq ` ǫ
i
i
ǫ fi HpZkj |Z1:k´1
, Θq ´ HpZkj |Z1:k
, Θq
j
j
, Θq ´ HpZki |Z1:k
, Θq
`HpZki |Z1:k´1
Properties of differential entropy [31] suggest that ǫ ě 0
regardless of the problem setting (as conditioning reduces
entropy). Under normal sensing conditions it is reasonable to
expect that ǫ ą 0 holds as sensor measurements are highly
informative on all the variables at time k. For sensors with
similar sensing capabilities, it is also reasonable to expect that
their measurement history are interchangeable. In other words,
j
i
HpZki |Z1:k´1
, Θq ` HpZkj |Z1:k´1
, Θq “
j
i
HpZkj |Z1:k´1
, Θq ` HpZki |Z1:k´1
, Θq (81)
holds in which case (80) is nonzero and Dpp||uq ą Dpp||qq.
This condition can be relaxed for a difference of ǫ between
the RHS and LHS of (81).
Comparison in terms of entropy upper bounds is more
straightforward. Let us consider (32) and Corollary 4.2 in [26]
which we repeat here for convenience:
Hu`
Dpp||uq ď
´ Hu´ ,
(82)
j
`
i
Hu “ HpXk |Z1:k´1 , Θq ` HpXk |Z1:k´1 , Θq,
j
i
Hu´ “ HpXk |Z1:k´1
, Z1:k´1
, Θq `
j
j
i
i
maxtHpXk |Z1:k´1 , Z1:k , Θq, HpXk |Z1:k´1
, Z1:k
, Θqu.
For sensors of identical capabilities, both terms in the maximisation should be identical as the accuracy of the state
estimate should not differ for using either of the sensors’
current measurement in addition to the histories of both. As a
result, we can replace Hu´ with
j
i
Hu´ “ HpXk |Z1:k´1
, Z1:k´1
, Θq`
´
¯
1
j
j
i
i
HpXk |Z1:k´1
, Z1:k
, Θq ` HpXk |Z1:k´1
, Z1:k
, Θq .
2
Now, note that the sum of the positive weighted terms in the
RHS of (32) (let us denote by Hq` ) is smaller than Hu` owing
to that conditioning reduces entropy [31]. The sum of the
negative weighted terms (let us denote by Hq´ ), for the case
,equals to Hu´ , leading to
Hu` ´ Hu´ ´ pHq` ´ Hq´ q ą 0,
which indicates that the entropy bound of the quad-term update
is smaller than that for the dual-term update.
E. Local ML estimate of object correspondances
The semantic of the object correspondance γk is slightly
different from that of the random permutation variables used
in identity management [45] in a way closer to data association
15
because local (track) identities are synonymous with the measurements they are associated with instead of signal features
etc. Therefore, in our model, γk is uniquely defined when
given τki and τkj , and, γ1 , . . . , γk are mutually independent.
The ML estimate of γk using the likelihood for data set
pZik , Zj1:k q is also conditioned on the association configuration
pτ̂ki , τ̂ki q and θ is given by
γ̂k “ arg max log lpZik , Zj1:k |γk , τki “ τ̂ki , τkj “ τ̂kj , θq (83)
γk PSM
where SM is the set of M -permutations. This likelihood can
be decomposed using the chain rule of probability as
lpZik , Zj1:k |γk , τki , τkj , θq
“ ppZik , Zjk |Zj1:k´1 , γk , τki , τkj , θqppZj1:k´1 |γk , τki , τkj , θq,
“ ppZik , Zjk |Zj1:k´1 , γk , τki , τkj , θqppZj1:k´1 |θq,
(84)
where the second equality follows from the independence of
sensor j’s measurements up to time k ´ 1 from the association
variables at time k. The first term on the RHS is easily
identified as (43) evaluated for γk when m1 “ γk pτki poqq.
This term, when subsituted in (83), leads to a 2-D assignment
problem given by
γ̂k “ arg max
γk PSM
M
ÿ
dpo, m1 “ γk pτ̂ki poqqq
o“1
k
i
dpo, m1 q fi log rij,o
pzk,o
, θq,
k
for o, m1 “ 1, . . . , M , where rij,o
is given in (43).
2
After finding the M costs above, this problem can be
solved using the auction algorithm [48] in polynomial time
with M . This algorithm finds the M pairs corresponding to
the best permutation γ̂k . We use a similar approach for the data
association problem in Bayesian filtering in Section VI-A.
ACKNOWLEDGMENT
The authors would like to thank Prof. Arnaud Doucet
for discussions on pseudo-likelihood methods in state space
models, and, Dr. Sinan Yildirim for discussions on the general
multi-object tracking model in [14] and [27]8.
R EFERENCES
[1] S. Sarkka, A. Solin, and J. Hartikainen, “Spatiotemporal learning
via infinite-dimensional Bayesian filtering and smoothing: A look at
Gaussian process regression through Kalman filtering,” IEEE Signal
Processing Magazine, vol. 30, no. 4, pp. 51–61, July 2013.
[2] B.-N. Vo, M. Mallick, Y. Bar-Shalom, S. Coraluppi, R. Osborne, R.
Mahler, B.-T. Vo, and J.G. Webster, Wiley Encyclopedia of Electrical
and Electronics Eng., Multitarget Tracking, J. Wiley & Sons, Inc., 2015.
[3] O. Cappé, E. Moulines, and T. Ryden, Inference in Hidden Markov
Models, Springer Series in Statistics. Springer New York, 2006.
[4] O. Cappé, S. J. Godsill, and E. Moulines, “An overview of existing
methods and recent advances in sequential Monte Carlo,” Proceedings
of the IEEE, vol. 95, pp. 899–924, 2007.
[5] S. S. Singh, N. Whiteley, and S. J. Godsill, “Approximate likelihood
estimation of static parameters in multi-target models,” in Bayesian
Time Series Models, D. Barber, A. T. Cemgil, and S. Chiappa, Eds., pp.
225–244. Cambridge University Press, 2011.
[6] X. Lin, Y. Bar-Shalom, and T. Kirubarajan, “Multisensor multitarget bias
estimation for general asynchronous sensors,” IEEE Trans. on Aerospace
and Electronic Systems, vol. 41, no. 3, pp. 899–921, July 2005.
[7] B. Ristic, D.E. Clark, and N. Gordon, “Calibration of multi-target
tracking algorithms using non-cooperative targets,” IEEE Journal of
Selected Topics in Signal Processing, vol. 7, no. 3, pp. 390–398, 2013.
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
[8] M. Erol-Kantarci, H.T. Mouftah, and S. Oktug, “A survey of architectures and localization techniques for underwater acoustic sensor
networks,” Communications Surveys Tutorials, IEEE, vol. 13, no. 3, pp.
487–502, Third Quarter 2011.
[9] N.K. Kantas, S.S. Singh, and A. Doucet, “Distributed maximum
likelihood for simultaneous self-localization and tracking in sensor
networks,” IEEE Trans. on Signal Proc., vol. 60, no. 10, pp. 5038–
5047, 2012.
[10] M. Uney, B. Mulgrew, and D. Clark, “Target aided online sensor
localisation in bearing only clusters,” in SSPD’14, Sept 2014, pp. 1–5.
[11] D. Devarajan, Z. Cheng, and R.J. Radke, “Calibrating distributed camera
networks,” Proc. of the IEEE, vol. 96, no. 10, pp. 1625–1639, 2008.
[12] N. Kantas, A. Doucet, S. S. Singh, J. Maciejowski, and N. Chopin,
“On particle methods for parameter estimation in state-space models,”
Statistical Science, vol. 30, no. 3, pp. 328–351, 08 2015.
[13] L. Chen, M.J. Wainwright, M. Çetin, and A.S. Willsky, “Data association
based on optimization in graphical models with application to sensor
networks,” Math. Comput. Model., vol. 43, no. 9-10, pp. 1114–1135,
May 2006.
[14] L. Jiang, S. S. Singh, and S.Yildirim, “Bayesian tracking and parameter
learning for non-linear multiple target tracking models,” IEEE Transactions on Signal Processing, vol. 63, no. 21, pp. 5733–5745, Nov 2015.
[15] C. Andrieu, A. Doucet, and R. Holenstein, “Particle markov chain
monte carlo methods,” Journal of the Royal Statistical Society: Series
B (Statistical Methodology), vol. 72, no. 3, pp. 269–342, 2010.
[16] J. Ala-Luhtala, N. Whiteley, K. Heine, and R. Piche, “An introduction
to twisted particle filters and parameter estimation in non-linear statespace models,” IEEE Trans. on Signal Proc., vol. PP, no. 99, pp. 1–1,
2016.
[17] J.-M. Marin, P. Pudlo, C. P. Robert, and R. J. Ryder, “Approximate
bayesian computational methods,” Statistics and Computing, vol. 22,
no. 6, pp. 1167–1180, 2012.
[18] C. Varin, N. Reid, and D. Firth, “An overview of composite likelihood
methods,” Statistica Sinica, vol. 21, pp. 5–42, 2011.
[19] E. Ehrlich, A. Jasra, and N. Kantas, “Gradient free parameter estimation
for hidden Markov models with intractable likelihoods,” Methodology
and Computing in Applied Prob., vol. 17, no. 2, pp. 315–349, 2015.
[20] S. Yildirim, S. S. Singh, T. Dean, and A. Jasra, “Parameter estimation
in hidden Markov models with intractable likelihoods using sequential
Monte Carlo,” Journal of Computational and Graphical Statistics, vol.
24, no. 3, pp. 846–865, 2015.
[21] C. Andrieu, A. Doucet, and V. B. Tadic, “On-line parameter estimation
in general state-space models,” in Proceedings of the 44th IEEE
Conference on Decision and Control, Dec 2005, pp. 332–337.
[22] A. Wiesel and A.O. Hero, “Distributed covariance estimation in gaussian
graphical models,” Signal Processing, IEEE Transactions on, vol. 60,
no. 1, pp. 211–220, Jan 2012.
[23] Q. Liu and A. Ihler, “Distributed parameter estimation via pseudolikelihood,” in Proceedings of the 29th International Conference on
Machine Learning (ICML-12), John Langford and Joelle Pineau, Eds.,
New York, NY, USA, July 2012, ICML ’12, pp. 1487–1494, Omnipress.
[24] Y. D. Mizrahi, M. Denil, and N. D. Freitas, “Distributed parameter
estimation in probabilistic graphical models,” in Advances in NIPS
27, Z. Ghahramani, M. Welling, C. Cortes, N.d. Lawrence, and K.q.
Weinberger, Eds., pp. 1700–1708. Curran Assoc., Inc., 2014.
[25] M. Uney, B. Mulgrew, and D. Clark, “Cooperative sensor localisation
in distributed fusion networks by exploiting non-cooperative targets,” in
IEEE Workshop on Stat. Signal Proc. (SSP)’14, June 2014, pp. 516–519.
[26] M. Uney, B. Mulgrew, and D. Clark, “A cooperative approach to sensor
localisation in distributed fusion networks,” IEEE Transactions on Signal
Processing, vol. 64, no. 5, pp. 1187–1199, March 2016.
[27] S. Yildirim, L. Jiang, S. S. Singh, and T. A. Dean, “Calibrating the
Gaussian multi-target tracking model,” Statistics and Computing, vol.
25, no. 3, pp. 595–608, May 2015.
[28] B. P. Carlin and T. A. Louis, Bayes and empirical Bayes methods for
data analysis, Chapman and Hall/CRC, 2nd edition, 2000.
[29] D. Hall, C.-Y. Chong, J. Llinas, and M. Liggins II, Eds., Distributed
Data Fusion for Network-Centric Operations, CRC Press, 2013.
[30] J. S. Yedidia, W. T. Freeman, and Y. Weiss, “Understanding belief
propagation and its generalizations,” Tech. Rep. TR-2001-22, Mitsubishi
Electric Research Laboratories, San Francisco, CA, USA, Jan. 2002.
[31] T. M. Cover and J. A. Thomas, Elements of Information Theory, John
Wiley & Sons, 1991.
[32] E. B. Sudderth, A. T. Ihler, M. Isard, W. T. Freeman, and A. S. Willsky,
“Nonparametric belief propagation,” Commun. ACM, vol. 53, no. 10,
pp. 95–103, Oct. 2010.
16
[33] M. Uney, B. Mulgrew, and D. Clark, “Distributed estimation of latent
parameters in state space models using separable likelihoods,” in
Proceedings of the ICASSP 2016, 2016.
[34] M. J. Wainwright and M. I. Jordan, “Graphical models, exponential
families, and variational inference,” Found. Trends Mach. Learn., vol.
1, no. 1-2, pp. 1–305, Jan. 2008.
[35] A. B. Poore and S. Gadaleta, “Some assignment problems arising from
multiple target tracking,” Mathematical and Computer Modelling, vol.
43, no. 9 - 10, pp. 1074 – 1091, 2006, Optimization and Control for
Military Applications.
[36] B. Ristic, Arulampalam S., and N. Gordon, Beyond the Kalman Filter;
particle filters for tracking applications, Artech House, 2004.
[37] J. Kokkala and S. Sarkka, “Combining particle MCMC with RaoBlackwellized Monte Carlo data association for parameter estimation
in multiple target tracking,” Digital Signal Processing, vol. 47, pp. 84
– 95, 2015, Special Issue in Honour of W. J. (Bill) Fitzgerald.
[38] I. Schlangen, D. E. Clark, and E. D. Delande, “Single-cluster PHD filter
methods for joint multi-object filtering and parameter estimation,” arXiv
preprint arXiv:1705.05312, 2017.
[39] A.T. Ihler, J.W. Fisher, R.L. Moses, and A.S. Willsky, “Nonparametric
belief propagation for self-localization of sensor networks,” Selected
Areas in Communications, IEEE Journal on, vol. 23, no. 4, pp. 809–
819, 2005.
[40] M. Cetin, Lei Chen, J.W. Fisher, A.T. Ihler, R.L. Moses, M.J. Wainwright, and A.S. Willsky, “Distributed fusion in sensor networks: A
graphical models perspective,” IEEE Signal Processing Magazine, vol.
23, no. 4, pp. 42–55, 2006.
[41] M. Uney and M. Cetin, “Graphical model-based approaches to target
tracking in sensor networks: An overview of some recent work and
challenges,” in Image and Signal Processing and Analysis, 2007. ISPA
2007. 5th International Symposium on, 2007, pp. 492–497.
[42] M.J. Wainwright, T.S. Jaakkola, and A.S. Willsky, “Tree-based reparameterization framework for analysis of sum-product and related algorithms,” IEEE Trans. on Info. Theory, vol. 49, no. 5, pp. 1120–1146,
May 2003.
[43] J. S. Yedidia, W.T. Freeman, and Y. Weiss, “Constructing free-energy
approximations and generalized belief propagation algorithms,” IEEE
Trans. on Info. Theory, vol. 51, no. 7, pp. 2282–2312, July 2005.
[44] M. Üney, B. Mulgrew, and D. Clark, “Distributed localisation of sensors
with partially overlapping field-of-views in fusion networks,” in 2016
19th International Conference on Information Fusion (FUSION), July
2016, pp. 1340–1347.
[45] L. J. Guibas, “The identity management problem- a short survey,” in
11th Int. Conf. on Info. Fusion (Fusion’08), June 2008, pp. 1–7.
[46] J. Huang, C. Guestrin, and L. Guibas, “Fourier theoretic probabilistic
inference over permutations,” Journal of Machine Learning Research,
vol. 10, pp. 997–1070, May 2009.
[47] R. Burkard, M. Dell’Amico, and S. Martello, Assignment Problems,,
Society for Industrial and Applied Mathematics, 2012.
[48] D. P. Bertsekas, “The auction algorithm: A distributed relaxation method
for the assignment problem,” Annals of Operations Research, vol. 14,
no. 1, pp. 105–123, 1988.
[49] G. Casella and C. P. Robert, Monte Carlo Statistical Methods, Springer,
second edition, 2005.
[50] B.W. Silverman, Density Estimation for Statistics and Data Analysis,
Chapman and Hall, 1986.
[51] A. F. M. Smith and A. E. Gelfand, “Bayes statistics without tears: A
sampling-resampling perspective,” The American Statistician, vol. 46,
no. 2, pp. 84–88, May 1992.
[52] B. N. Vo and W. K. Ma, “The gaussian mixture probability hypothesis
density filter,” IEEE Transactions on Signal Processing, vol. 54, no. 11,
pp. 4091–4104, Nov 2006.
[53] M. Üney, K. Copsey, S. Page, B. Mulgrew, and P. Thomas, “Enabling
self-configuration of fusion networks via scalable opportunistic sensor
calibration,” in Proceedings of the SPIE Defence+Security Symposium,
Signal Processing, Sensor/Information Fusion, and Target Recognition
XXVII, 2018, to appear.
| 7cs.IT
|
Unsupervised Image-to-Image Translation Networks
Thomas Breuel, Jan Kautz
NVIDIA
{mingyul,tbreuel,jkautz}@nvidia.com
arXiv:1703.00848v5 [cs.CV] 15 Feb 2018
Ming-Yu Liu,
Abstract
Unsupervised image-to-image translation aims at learning a joint distribution of
images in different domains by using images from the marginal distributions in
individual domains. Since there exists an infinite set of joint distributions that
can arrive the given marginal distributions, one could infer nothing about the joint
distribution from the marginal distributions without additional assumptions. To
address the problem, we make a shared-latent space assumption and propose an
unsupervised image-to-image translation framework based on Coupled GANs.
We compare the proposed framework with competing approaches and present
high quality image translation results on various challenging unsupervised image
translation tasks, including street scene image translation, animal image translation,
and face image translation. We also apply the proposed framework to domain
adaptation and achieve state-of-the-art performance on benchmark datasets. Code
and additional results are available in https://github.com/mingyuliutw/unit.
1
Introduction
Many computer visions problems can be posed as an image-to-image translation problem, mapping
an image in one domain to a corresponding image in another domain. For example, super-resolution
can be considered as a problem of mapping a low-resolution image to a corresponding high-resolution
image; colorization can be considered as a problem of mapping a gray-scale image to a corresponding
color image. The problem can be studied in supervised and unsupervised learning settings. In the
supervised setting, paired of corresponding images in different domains are available [8, 15]. In the
unsupervised setting, we only have two independent sets of images where one consists of images
in one domain and the other consists of images in another domain—there exist no paired examples
showing how an image could be translated to a corresponding image in another domain. Due to
lack of corresponding images, the UNsupervised Image-to-image Translation (UNIT) problem is
considered harder, but it is more applicable since training data collection is easier.
When analyzing the image translation problem from a probabilistic modeling perspective, the key
challenge is to learn a joint distribution of images in different domains. In the unsupervised setting,
the two sets consist of images from two marginal distributions in two different domains, and the task is
to infer the joint distribution using these images. The coupling theory [16] states there exist an infinite
set of joint distributions that can arrive the given marginal distributions in general. Hence, inferring
the joint distribution from the marginal distributions is a highly ill-posed problem. To address the
ill-posed problem, we need additional assumptions on the structure of the joint distribution.
To this end we make a shared-latent space assumption, which assumes a pair of corresponding images
in different domains can be mapped to a same latent representation in a shared-latent space. Based on
the assumption, we propose a UNIT framework that are based on generative adversarial networks
(GANs) and variational autoencoders (VAEs). We model each image domain using a VAE-GAN. The
adversarial training objective interacts with a weight-sharing constraint, which enforces a sharedlatent space, to generate corresponding images in two domains, while the variational autoencoders
relate translated images with input images in the respective domains. We applied the proposed
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
E1
Z : shared latent space
x̃1!1
1
x1
z
E1
D1
G1
E2
G1
T/F
x̃2!1
2
G2
x2
x1
X2
X1
x̃1!2
1
x2
(a)
z
E2
T/F
x̃2!2
2
G2
(b)
D2
Figure 1: (a) The shared latent space assumption. We assume a pair of corresponding images (x1 , x2 ) in
two different domains X1 and X2 can be mapped to a same latent code z in a shared-latent space Z. E1 and
E2 are two encoding functions, mapping images to latent codes. G1 and G2 are two generation functions,
mapping latent codes to images. (b) The proposed UNIT framework. We represent E1 E2 G1 and G2 using
CNNs and implement the shared-latent space assumption using a weight sharing constraint where the connection
weights of the last few layers (high-level layers) in E1 and E2 are tied (illustrated using dashed lines) and the
connection weights of the first few layers (high-level layers) in G1 and G2 are tied. Here, x̃1→1
and x̃2→2
1
2
1→2
2→1
are self-reconstructed images, and x̃1
and x̃2
are domain-translated images. D1 and D2 are adversarial
discriminators for the respective domains, in charge of evaluating whether the translated images are realistic.
Table 1: Interpretation of the roles of the subnetworks in the proposed framework.
Networks
Roles
{E1 , G1 }
{E1 , G2 }
{G1 , D1 }
{E1 , G1 , D1 }
VAE for X1 Image Translator X1 → X2 GAN for X1 VAE-GAN [14]
{G1 , G2 , D1 , D2 }
CoGAN [17]
framework to various unsupervised image-to-image translation problems and achieved high quality
image translation results. We also applied it to the domain adaptation problem and achieved state-ofthe-art accuracies on benchmark datasets. The shared-latent space assumption was used in Coupled
GAN [17] for joint distribution learning. Here, we extend the Coupled GAN work for the UNIT
problem. We also note that several contemporary works propose the cycle-consistency constraint
assumption [29, 10], which hypothesizes the existence of a cycle-consistency mapping so that an
image in the source domain can be mapped to an image in the target domain and this translated image
in the target domain can be mapped back to the original image in the source domain. In the paper, we
show that the shared-latent space constraint implies the cycle-consistency constraint.
2
Assumptions
Let X1 and X2 be two image domains. In supervised image-to-image translation, we are given
samples (x1 , x2 ) drawn from a joint distribution PX1 ,X2 (x1 , x2 ). In unsupervised image-to-image
translation, we are given samples drawn from the marginal distributions PX1 (x1 ) and PX2 (x2 ). Since
an infinite set of possible joint distributions can yield the given marginal distributions, we could infer
nothing about the joint distribution from the marginal samples without additional assumptions.
We make the shared-latent space assumption. As shown Figure 1, we assume for any given pair
of images x1 and x2 , there exists a shared latent code z in a shared-latent space, such that we
can recover both images from this code, and we can compute this code from each of the two
images. That is, we postulate there exist functions E1∗ , E2∗ , G∗1 , and G∗2 such that, given a pair of
corresponding images (x1 , x2 ) from the joint distribution, we have z = E1∗ (x1 ) = E2∗ (x2 ) and
∗
conversely x1 = G∗1 (z) and x2 = G∗2 (z). Within this model, the function x2 = F1→2
(x1 ) that
∗
∗
∗
maps from X1 to X2 can be represented by the composition F1→2 (x1 ) = G2 (E1 (x1 )). Similarly,
∗
∗
x1 = F2→1
(x2 ) = G∗1 (E2∗ (x2 )). The UNIT problem then becomes a problem of learning F1→2
∗
∗
∗
and F2→1
. We note that a necessary condition for F1→2
and F2→1
to exist is the cycle-consistency
∗
∗
∗
∗
constraint [29, 10]: x1 = F2→1
(F1→2
(x1 )) and x2 = F1→2
(F2→1
(x2 )). We can reconstruct
the input image from translating back the translated input image. In other words, the proposed
shared-latent space assumption implies the cycle-consistency assumption (but not vice versa).
To implement the shared-latent space assumption, we further assume a shared intermediate representation h such that the process of generating a pair of corresponding images admits a form of
2
x1
.
(1)
x2
Consequently, we have G∗1 ≡ G∗L,1 ◦ G∗H and G∗2 ≡ G∗L,2 ◦ G∗H where G∗H is a common high-level
generation function that maps z to h and G∗L,1 and G∗L,2 are low-level generation functions that map
h to x1 and x2 , respectively. In the case of multi-domain image translation (e.g., sunny and rainy
image translation), z can be regarded as the compact, high-level representation of a scene ("car in
front, trees in back"), and h can be considered a particular realization of z through G∗H ("car/tree
occupy the following pixels"), and G∗L,1 and G∗L,2 would be the actual image formation functions
in each modality ("tree is lush green in the sunny domain, but dark green in the rainy domain").
∗
∗
∗
∗
Assuming h also allow us to represent E1∗ and E2∗ by E1∗ ≡ EH
◦ EL,1
and E2∗ ≡ EH
◦ EL,2
.
z→h
%
&
In the next section, we discuss how we realize the above ideas in the proposed UNIT framework.
3
Framework
Our framework, as illustrated in Figure 1, is based on variational autoencoders (VAEs) [13, 22, 14]
and generative adversarial networks (GANs) [6, 17]. It consists of 6 subnetworks: including two
domain image encoders E1 and E2 , two domain image generators G1 and G2 , and two domain
adversarial discriminators D1 and D2 . Several ways exist to interpret the roles of the subnetworks,
which we summarize in Table 1. Our framework learns translation in both directions in one shot.
VAE. The encoder–generator pair {E1 , G1 } constitutes a VAE for the X1 domain, termed VAE1 . For
an input image x1 ∈ X1 , the VAE1 first maps x1 to a code in a latent space Z via the encoder E1 and
then decodes a random-perturbed version of the code to reconstruct the input image via the generator
G1 . We assume the components in the latent space Z are conditionally independent and Gaussian with
unit variance. In our formulation, the encoder outputs a mean vector Eµ,1 (x1 ) and the distribution
of the latent code z1 is given by q1 (z1 |x1 ) ≡ N (z1 |Eµ,1 (x1 ), I) where I is an identity matrix. The
reconstructed image is x̃1→1
= G1 (z1 ∼ q1 (z1 |x1 )). Note that here we abused the notation since
1
we treated the distribution of q1 (z1 |x1 ) as a random vector of N (Eµ,1 (x1 ), I) and sampled from it.
Similarly, {E2 , G2 } constitutes a VAE for X2 : VAE2 where the encoder E2 outputs a mean vector
Eµ,2 (x2 ) and the distribution of the latent code z2 is given by q2 (z2 |x2 ) ≡ N (z2 |Eµ,2 (x2 ), I). The
reconstructed image is x̃2→2
= G2 (z2 ∼ q2 (z2 |x2 )).
2
Utilizing the reparameterization trick [13], the non-differentiable sampling operation can be reparameterized as a differentiable operation using auxiliary random variables. This reparameterization trick
allows us to train VAEs using back-prop. Let η be a random vector with a multi-variate Gaussian
distribution: η ∼ N (η|0, I). The sampling operations of z1 ∼ q1 (z1 |x1 ) and z2 ∼ q2 (z2 |x2 ) can be
implemented via z1 = Eµ,1 (x1 ) + η and z2 = Eµ,2 (x2 ) + η, respectively.
Weight-sharing. Based on the shared-latent space assumption discussed in Section 2, we enforce a
weight-sharing constraint to relate the two VAEs. Specifically, we share the weights of the last few
layers of E1 and E2 that are responsible for extracting high-level representations of the input images
in the two domains. Similarly, we share the weights of the first few layers of G1 and G2 responsible
for decoding high-level representations for reconstructing the input images.
Note that the weight-sharing constraint alone does not guarantee that corresponding images in two
domains will have the same latent code. In the unsupervised setting, no pair of corresponding images
in the two domains exists to train the network to output a same latent code. The extracted latent
codes for a pair of corresponding images are different in general. Even if they are the same, the same
latent component may have different semantic meanings in different domains. Hence, the same latent
code could still be decoded to output two unrelated images. However, we will show that through
adversarial training, a pair of corresponding images in the two domains can be mapped to a common
latent code by E1 and E2 , respectively, and a latent code will be mapped to a pair of corresponding
images in the two domains by G1 and G2 , respectively.
The shared-latent space assumption allows us to perform image-to-image translation. We can
translate an image x1 in X1 to an image in X2 through applying G2 (z1 ∼ q1 (z1 |x1 )). We term such
an information processing stream as the image translation stream. Two image translation streams exist
in the proposed framework: X1 → X2 and X2 → X1 . The two streams are trained jointly with the
two image reconstruction streams from the VAEs. Once we could ensure that a pair of corresponding
3
images are mapped to a same latent code and a same latent code is decoded to a pair of corresponding
images, (x1 , G2 (z1 ∼ q1 (z1 |x1 ))) would form a pair of corresponding images. In other words, the
∗
composition of E1 and G2 functions approximates F1→2
for unsupervised image-to-image translation
∗
discussed in Section 2, and the composition of E2 and G1 function approximates F2→1
.
GANs. Our framework has two generative adversarial networks: GAN1 = {D1 , G1 } and GAN2 =
{D2 , G2 }. In GAN1 , for real images sampled from the first domain, D1 should output true, while
for images generated by G1 , it should output false. G1 can generate two types of images: 1) images
from the reconstruction stream x̃1→1
= G1 (z1 ∼ q1 (z1 |x1 )) and 2) images from the translation
1
stream x̃2→1
=
G
(z
∼
q
(z
|x
)).
Since
the reconstruction stream can be supervisedly trained, it
1 2
2 2 2
2
is suffice that we only apply adversarial training to images from the translation stream, x̃2→1
. We
2
apply a similar processing to GAN2 where D2 is trained to output true for real images sampled from
the second domain dataset and false for images generated from G2 .
Cycle-consistency (CC). Since the shared-latent space assumption implies the cycle-consistency
constraint (See Section 2), we could also enforce the cycle-consistency constraint in the proposed
framework to further regularize the ill-posed unsupervised image-to-image translation problem. The
resulting information processing stream is called the cycle-reconstruction stream.
Learning. We jointly solve the learning problems of the VAE1 , VAE2 , GAN1 and GAN2 for the
image reconstruction streams, the image translation streams, and the cycle-reconstruction streams:
min
max LVAE1 (E1 , G1 ) + LGAN1 (E1 , G1 , D1 ) + LCC1 (E1 , G1 , E2 , G2 )
E1 ,E2 ,G1 ,G2 D1 ,D2
LVAE2 (E2 , G2 ) + LGAN2 (E2 , G2 , D2 ) + LCC2 (E2 , G2 , E1 , G1 ).
(2)
VAE training aims for minimizing a variational upper bound In (2), the VAE objects are
LVAE1 (E1 , G1 ) =λ1 KL(q1 (z1 |x1 )||pη (z)) − λ2 Ez1 ∼q1 (z1 |x1 ) [log pG1 (x1 |z1 )]
LVAE2 (E2 , G2 ) =λ1 KL(q2 (z2 |x2 )||pη (z)) − λ2 Ez2 ∼q2 (z2 |x2 ) [log pG2 (x2 |z2 )].
(3)
(4)
where the hyper-parameters λ1 and λ2 control the weights of the objective terms and the KL
divergence terms penalize deviation of the distribution of the latent code from the prior distribution.
The regularization allows an easy way to sample from the latent space [13]. We model pG1 and pG2
using Laplacian distributions, respectively. Hence, minimizing the negative log-likelihood term is
equivalent to minimizing the absolute distance between the image and the reconstructed image. The
prior distribution is a zero mean Gaussian pη (z) = N (z|0, I).
In (2), the GAN objective functions are given by
LGAN1 (E1 , G1 , D1 ) = λ0 Ex1 ∼PX1 [log D1 (x1 )] + λ0 Ez2 ∼q2 (z2 |x2 ) [log(1 − D1 (G1 (z2 )))]
LGAN2 (E2 , G2 , D2 ) = λ0 Ex2 ∼PX2 [log D2 (x2 )] + λ0 Ez1 ∼q1 (z1 |x1 ) [log(1 − D2 (G2 (z1 )))].
(5)
(6)
The objective functions in (5) and (6) are conditional GAN objective functions. They are used to
ensure the translated images resembling images in the target domains, respectively. The hyperparameter λ0 controls the impact of the GAN objective functions.
We use a VAE-like objective function to model the cycle-consistency constraint, which is given by
LCC1 (E1 , G1 , E2 , G2 ) =λ3 KL(q1 (z1 |x1 )||pη (z)) + λ3 KL(q2 (z2 |x1→2
))||pη (z))−
1
λ4 Ez2 ∼q2 (z2 |x1→2
[log
p
(x
|z
)]
G
1
2
)
1
1
λ3 KL(q1 (z1 |x2→1
))||pη (z))−
2
LCC2 (E2 , G2 , E1 , G1 ) =λ3 KL(q2 (z2 |x2 )||pη (z)) +
λ4 Ez1 ∼q1 (z1 |x2→1
) [log pG2 (x2 |z1 )].
2
(7)
(8)
where the negative log-likelihood objective term ensures a twice translated image resembles the
input one and the KL terms penalize the latent codes deviating from the prior distribution in the
cycle-reconstruction stream (Therefore, there are two KL terms). The hyper-parameters λ3 and λ4
control the weights of the two different objective terms.
Inheriting from GAN, training of the proposed framework results in solving a mini-max problem
where the optimization aims to find a saddle point. It can be seen as a two player zero-sum game.
The first player is a team consisting of the encoders and generators. The second player is a team
consisting of the adversarial discriminators. In addition to defeating the second player, the first player
has to minimize the VAE losses and the cycle-consistency losses. We apply an alternating gradient
4
Accuracy
0.7
0.7
0.56
0.56
0.42
0.42
0.28
0.28
0.14
0
1
6𝑳 Dis
5𝑳 Dis
4𝑳 Dis
3𝑳 Dis
2
3
4
# of shared layers in gen.
(a)
0.14
(b)
0
1000
Method
𝞴1=10
1 = 10
𝞴1=0.1
1 = 0.1
100
𝞴1=1
1 =1
𝞴1=0.01
1 = 0.01
10
1
Accuracy
Weight 0.569±0.029
Sharing
Cycle 0.568±0.010
Consistenc
Proposed 0.600±0.015
𝞴22
(c)
(d)
Figure 2: (a) Illustration of the Map dataset. Left: satellite image. Right: map. We translate holdout satellite
images to maps and measure the accuracy achieved by various configurations of the proposed framework.
(b) Translation accuracy versus different network architectures. (c) Translation accuracy versus different
hyper-parameter values. (d) Impact of weight-sharing and cycle-consistency constraints on translation accuracy.
update scheme similar to the one described in [6] to solve (2). Specifically, we first apply a gradient
ascent step to update D1 and D2 with E1 , E2 , G1 , and G2 fixed. We then apply a gradient descent
step to update E1 , E2 , G1 , and G2 with D1 and D2 fixed.
Translation: After learning, we obtain two image translation functions by assembling a subset of the
subnetworks. We have F1→2 (x1 ) = G2 (z1 ∼ q1 (z1 |x1 )) for translating images from X1 to X2 and
F2→1 (x2 ) = G1 (z2 ∼ q2 (z2 |x2 )) for translating images from X2 to X1 .
4
Experiments
We first analyze various components of the proposed framework. We then present visual results on
challenging translation tasks. Finally, we apply our framework to the domain adaptation tasks.
Performance Analysis. We used ADAM [11] for training where the learning rate was set to 0.0001
and momentums were set to 0.5 and 0.999. Each mini-batch consisted of one image from the first
domain and one image from the second domain. Our framework had several hyper-parameters. The
default values were λ0 = 10, λ3 = λ1 = 0.1 and λ4 = λ2 = 100. For the network architecture,
our encoders consisted of 3 convolutional layers as the front-end and 4 basic residual blocks [7] as
the back-end. The generators consisted of 4 basic residual blocks as the front-end and 3 transposed
convolutional layers as the back-end. The discriminators consisted of stacks of convolutional layers.
We used LeakyReLU for nonlinearity. The details of the networks are given in Appendix A.
We used the map dataset [8] (visualized in Figure 2), which contained corresponding pairs of images
in two domains (satellite image and map) useful for quantitative evaluation. Here, the goal was to
learn to translate between satellite images and maps. We operated in an unsupervised setting where
we used the 1096 satellite images from the training set as the first domain and 1098 maps from the
validation set as the second domain. We trained for 100K iterations and used the final model to
translate 1098 satellite images in the test set. We then compared the difference between a translated
satellite image (supposed to be maps) and the corresponding ground truth maps pixel-wisely. A pixel
translation was counted correct if the color difference was within 16 of the ground truth color value.
We used the average pixel accuracy over the images in the test set as the performance metric. We
could use color difference for measuring translation accuracy since the target translation function
was unimodal. We did not evaluate the translation from maps to images since the translation was
multi-modal, which was difficult to construct a proper evaluation metric.
In one experiment, we varied the number of weight-sharing layers in the VAEs and paired each
configuration with discriminator architectures of different depths during training. We changed the
number of weight-sharing layers from 1 to 4. (Sharing 1 layer in VAEs means sharing 1 layer for
E1 and E2 and, at the same time, sharing 1 layer for G1 and G2 .) The results were reported in
Figure 2(b). Each curve corresponded to a discriminator architecture of a different depth. The x-axis
denoted the number of weigh-sharing layers in the VAEs. We found that the shallowest discriminator
architecture led to the worst performance. We also found that the number of weight-sharing layer
had little impact. This was due to the use of the residual blocks. As tying the weight of one layer, it
effectively constrained the other layers since the residual blocks only updated the residual information.
In the rest of the experiments, we used VAEs with 1 sharing layer and discriminators of 5 layers.
5
We analyzed impact of the hyper-parameter values to the translation accuracy. For different weight
values on the negative log likelihood terms (i.e., λ2 , λ4 ), we computed the achieved translation
accuracy over different weight values on the KL terms (i.e., λ1 , λ3 ). The results were reported in
Figure 2(c). We found that, in general, a larger weight value on the negative log likelihood terms
yielded a better translation accuracy. We also found setting the weights of the KL terms to 0.1 resulted
in consistently good performance. We hence set λ1 = λ3 = 0.1 and λ2 = λ4 = 100.
We performed an ablation study measuring impact of the weight-sharing and cycle-consistency
constraints to the translation performance and showed the results in Figure 2(d). We reported average
accuracy over 5 trials (trained with different initialized weights.). We note that when we removed
the weight-sharing constraint (as a consequence, we also removed the reconstruction streams in the
framework), the framework was reduced to the CycleGAN architecture [29, 10]. We found the model
achieved an average pixel accuracy of 0.569. When we removed the cycle-consistency constraint
and only used the weight-sharing constraint1 , it achieved 0.568 average pixel accuracy. But when we
used the full model, it achieved the best performance of 0.600 average pixel accuracy. This echoed
our point that for the ill-posed joint distribution recovery problem, more constraints are beneficial.
Qualitative results. Figure 3 to 6 showed results of the proposed framework on various UNIT tasks.
Street images. We applied the proposed framework to several unsupervised street scene image
translation tasks including sunny to rainy, day to night, summery to snowy, and vice versa. For each
task, we used a set of images extracted from driving videos recorded at different days and cities. The
numbers of the images in the sunny/day, rainy, night, summery, and snowy sets are 86165, 28915,
36280, 6838, and 6044. We trained the network to translate street scene image of size 640×480. In
Figure 3, we showed several example translation results . We found that our method could generate
realistic translated images. We also found that one translation was usually harder than the other.
Specifically, the translation that required adding more details to the image was usually harder (e.g.
night to day). Additional results are available in https://github.com/mingyuliutw/unit.
Synthetic to real. In Figure 3, we showed several example results achieved by applying the proposed
framework to translate images between the synthetic images in the SYNTHIA dataset [23] and the
real images in the Cityscape dataset [2]. For the real to synthetic translation, we found our method
made the cityscape images cartoon like. For the synthetic to real translation, our method achieved
better results in the building, sky, road, and car regions than in the human regions.
Dog breed conversion. We used the images of Husky, German Shepherd, Corgi, Samoyed, and Old
English Sheep dogs in the ImageNet dataset to learn to translate dog images between different breeds.
We only used the head regions, which were extracted by a template matching algorithm. Several
example results were shown in Figure 4. We found our method translated a dog to a different breed.
Cat species conversion. We also used the images of house cat, tiger, lion, cougar, leopard, jaguar,
and cheetah in the ImageNet dataset to learn to translate cat images between different species. We
only used the head regions, which again were extracted by a template matching algorithm. Several
example results were shown in Figure 5. We found our method translated a cat to a different specie.
Face attribute. We used the CelebA dataset [18] for attribute-based face images translation. Each face
image in the dataset had several attributes, including blond hair, smiling, goatee, and eyeglasses. The
face images with an attribute constituted the 1st domain, while those without the attribute constituted
the 2nd domain. In Figure 6, we visualized the results where we translated several images that do not
have blond hair, eye glasses, goatee, and smiling to corresponding images with each of the individual
attributes. We found that the translated face images were realistic.
Domain Adaptation. We applied the proposed framework to the problem for adapting a classifier
trained using labeled samples in one domain (source domain) to classify samples in a new domain
(target domain) where labeled samples in the new domain are unavailable during training. Early
works have explored ideas from subspace learning [4] to deep feature learning [5, 17, 26].
We performed multi-task learning where we trained the framework to 1) translate images between
the source and target domains and 2) classify samples in the source domain using the features
extracted by the discriminator in the source domain. Here, we tied the weights of the high-level
layers of D1 and D2 . This allows us to adapt a classifier trained in the source domain to the target
domain. Also, for a pair of generated images in different domains, we minimized the L1 distance
1
We used this architecture in an earlier version of the paper.
6
Figure 3: Street scene image translation results. For each pair, left is input and right is the translated image.
Input
Old Eng.
Sheep Dog
Input
Cougar
Husky
German
Shepherd
Corgi
Input
Husky
Corgi
Input
Leopard
Figure 4: Dog breed translation results.
Cheetah
Leopard
Lion
Tiger
Figure 5: Cat species translation results.
Input
+Blond Hair +Eyeglasses
+Goatee
+Smiling
Input
+Blond Hair +Eyeglasses
Figure 6: Attribute-based face translation results.
7
+Goatee
+Smiling
Table 2: Unsupervised domain adaptation performance. The reported numbers are classification accuracies.
Method
SA [4]
DANN [5]
DTN [26]
CoGAN
UNIT (proposed)
SVHN→ MNIST
MNIST→ USPS
USPS→ MNIST
0.5932
-
0.7385
-
0.8488
-
0.9565
0.9315
0.9053
0.9597
0.9358
between the features extracted by the highest layer of the discriminators, which further encouraged
D1 and D2 to interpret a pair of corresponding images in the same way. We applied the approach to
several tasks including adapting from the Street View House Number (SVHN) dataset [20] to the
MNIST dataset and adapting between the MNIST and USPS datasets. Table 2 reported the achieved
performance with comparison to the competing approaches. We found that our method achieved a
0.9053 accuracy for the SVHN→MNIST task, which was much better than 0.8488 achieved by the
previous state-of-the-art method [26]. We also achieved better performance for the MNIST↔SVHN
task than the Coupled GAN approach, which was the state-of-the-art. The digit images had a small
resolution. Hence, we used a small network. We also found that the cycle-consistency constraint was
not necessary for this task. More details about the experiments are available in Appendix B.
5
Related Work
Several deep generative models were recently proposed for image generation including GANs [6],
VAEs [13, 22], and PixelCNN [27]. The proposed framework was based on GANs and VAEs but it
was designed for the unsupervised image-to-image translation task, which could be considered as a
conditional image generation model. In the following, we first review several recent GAN and VAE
works and then discuss related image translation works.
GAN learning is via staging a zero-sum game between the generator and discriminator. The quality
of GAN-generated images had improved dramatically since the introduction. LapGAN [3] proposed
a Laplacian pyramid implementation of GANs. DCGAN [21] used a deeper convolutional network.
Several GAN training tricks were proposed in [24]. WGAN [1] used the Wasserstein distance.
VAEs optimize a variational bound. By improving the variational approximation, better image
generation results were achieved [19, 12]. In [14], a VAE-GAN architecture was proposed to improve
image generation quality of VAEs. VAEs were applied to translate face image attribute in [28].
Conditional generative model is a popular approach for mapping an image from one domain to
another. Most of the existing works were based on supervised learning [15, 8, 9]. Our work differed
to the previous works in that we do not need corresponding images. Recently, [26] proposed the
domain transformation network (DTN) and achieved promising results on translating small resolution
face and digit images. In addition to faces and digits, we demonstrated that the proposed framework
can translate large resolution natural images. It also achieved a better performance in the unsupervised
domain adaptation task. In [25], a conditional generative adversarial network-based approach was
proposed to translate a rendering images to a real image for gaze estimation. In order to ensure
the generated real image was similar to the original rendering image, the L1 distance between
the generated and original image was minimized. We note that two contemporary papers [29, 10]
independently introduced the cycle-consistency constraint for the unsupervised image translation.
We showed that that the cycle-consistency constraint is a natural consequence of the proposed
shared-latent space assumption. From our experiment, we found that cycle-consistency and the
weight-sharing (a realization of the shared-latent space assumption) constraints rendered comparable
performance. When the two constraints were jointed used, the best performance was achieved.
6
Conclusion and Future Work
We presented a general framework for unsupervised image-to-image translation. We showed it
learned to translate an image from one domain to another without any corresponding images in two
domains in the training dataset. The current framework has two limitations. First, the translation
model is unimodal due to the Gaussian latent space assumption. Second, training could be unstable
due to the saddle point searching problem. We plan to address these issues in the future work.
8
References
[1] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. arXiv preprint arXiv:1701.07875, 2017.
[2] M. Cordts, M. Omran, S. Ramos, T. Scharwächter, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and
B. Schiele. The cityscapes dataset. Conference on Computer Vision and Pattern Recognition Workshop,
2015.
[3] E. L. Denton, S. Chintala, R. Fergus, et al. Deep generative image models using a laplacian pyramid of
adversarial networks. Advances in Neural Information Processing Systems, 2015.
[4] B. Fernando, A. Habrard, M. Sebban, and T. Tuytelaars. Unsupervised visual domain adaptation using
subspace alignment. International Conference on Computer Vision, 2013.
[5] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky.
Domain-adversarial training of neural networks. Journal of Machine Learning Research, 2016.
[6] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio.
Generative adversarial nets. Advances in Neural Information Processing Systems, 2014.
[7] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. Computer Vision and
Pattern Recognition, 2016.
[8] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial
networks. Conference on Computer Vision and Pattern Recognition, 2017.
[9] J. Johnson, A. Alahi, and L. Fei-Fei. Perceptual losses for real-time style transfer and super-resolution.
European Conference in Computer Vision, 2016.
[10] T. Kim, M. Cha, H. Kim, J. Lee, and J. Kim. Learning to discover cross-domain relations with generative
adversarial networks. International Conference on Machine Learning, 2017.
[11] D. Kingma and J. Ba. Adam: A method for stochastic optimization. International Conference on Learning
Representations, 2015.
[12] D. P. Kingma, T. Salimans, and M. Welling. Improving variational inference with inverse autoregressive
flow. Advances in Neural Information Processing Systems, 2016.
[13] D. P. Kingma and M. Welling. Auto-encoding variational bayes. International Conference on Learning
Representations, 2014.
[14] A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther. Autoencoding beyond pixels using a
learned similarity metric. International Conference on Machine Learning, 2016.
[15] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz,
Z. Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network.
Conference on Computer Vision and Pattern Recognition, 2017.
[16] T. Lindvall. Lectures on the coupling method. Courier Corporation, 2002.
[17] M.-Y. Liu and O. Tuzel. Coupled generative adversarial networks. Advances in Neural Information
Processing Systems, 2016.
[18] Z. Liu, P. Luo, X. Wang, and X. Tang. Deep learning face attributes in the wild. International Conference
on Computer Vision, 2015.
[19] L. Maaløe, C. K. Sønderby, S. K. Sønderby, and O. Winther. Auxiliary deep generative models. International Conference on Machine Learning, 2016.
[20] Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y. Ng. Reading digits in natural images with
unsupervised feature learning. Advances in Neural Information Processing Systems workshop, 2011.
[21] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional
generative adversarial networks. International Conference on Learning Representations, 2016.
[22] D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and variational inference in
deep latent gaussian models. International Conference on Machine Learning, 2014.
[23] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. Lopez. The SYNTHIA Dataset: A large collection
of synthetic images for semantic segmentation of urban scenes. Conference on Computer Vision and
Pattern Recognition, 2016.
[24] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques for
training gans. Advances in Neural Information Processing Systems, 2016.
[25] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang, and R. Webb. Learning from simulated and
unsupervised images through adversarial training. Conference on Computer Vision and Pattern Recognition,
2017.
[26] Y. Taigman, A. Polyak, and L. Wolf. Unsupervised cross-domain image generation. International
Conference on Learning Representations, 2017.
[27] A. van den Oord, N. Kalchbrenner, L. Espeholt, O. Vinyals, A. Graves, et al. Conditional image generation
with pixelcnn decoders. Advances in Neural Information Processing Systems, 2016.
[28] X. Yan, J. Yang, K. Sohn, and H. Lee. Attribute2image: Conditional image generation from visual
attributes. European Conference in Computer Vision, 2016.
[29] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using cycle-consistent
adversarial networks. International Conference on Computer Vision, 2017.
9
A
Network Architecture
The network architecture used for the unsupervised image-to-image translation experiments is given
in Table 3. We use the following abbreviation for ease of presentation: N=Neurons, K=Kernel size,
S=Stride size. The transposed convolutional layer is denoted by DCONV. The residual basic block is
denoted as RESBLK.
Table 3: Network architecture for the unsupervised image translation experiments.
Layer
1
2
3
4
5
6
µ
Layer
1
2
3
4
5
6
7
Layer
1
2
3
4
5
6
B
Encoders
Shared?
CONV-(N64,K7,S1), LeakyReLU
CONV-(N128,K3,S2), LeakyReLU
CONV-(N256,K3,S2), LeakyReLU
RESBLK-(N512,K3,S1)
RESBLK-(N512,K3,S1)
RESBLK-(N512,K3,S1)
RESBLK-(N512,K3,S1)
Generators
No
No
No
No
No
No
Yes
Shared?
RESBLK-(N512,K3,S1)
RESBLK-(N512,K3,S1)
RESBLK-(N512,K3,S1)
RESBLK-(N512,K3,S1)
DCONV-(N256,K3,S2), LeakyReLU
DCONV-(N128,K3,S2), LeakyReLU
DCONV-(N3,K1,S1), TanH
Discriminators
Yes
No
No
No
No
No
No
Shared?
CONV-(N64,K3,S2), LeakyReLU
CONV-(N128,K3,S2), LeakyReLU
CONV-(N256,K3,S2), LeakyReLU
CONV-(N512,K3,S2), LeakyReLU
CONV-(N1024,K3,S2), LeakyReLU
CONV-(N1,K2,S1), Sigmoid
No
No
No
No
No
No
Domain Adaptation
MNIST↔USPS. For the MNIST and USPS adaptation experiments, we used the entire training sets
in the two domains (60000 training images for MNIST and 7291 training images for USPS) for
learning and reported the classification performance on the test sets (10000 test images for MNIST
and 2007 test images for USPS). The MNIST and USPS images had different sizes, we resized them
to 28×28 for facilitating the experiments. We trained the Coupled GAN algorithm [17] using the
same setting for a fair comparison.
The encoder and generator architecture was given in Table 4. For the discriminators and classifiers,
we used an architecture akin to the one used in the Coupled GAN paper, which is given in Table 5.
SVHN→MNIST. For the SVHN and MNIST domain adaptation experiment, we used the extra
training set (consisting of 531131 images) in the SVHN dataset for the source domain and the
training set (consisting of 60000 images) in the MNIST dataset for the target domain as in the DTN
work [26]. The test set consists of 10000 images in the MNIST test dataset. The MNIST images were
in gray-scale. We converted them to RGB images and performed a data augmentation where we also
used the inversions of the original MNIST images for training. All the images were resized to 32×32
for facilitating the experiment. We also found spatial context information was useful. For each input
image, we created a 5-channel variant where the first three channels were the original RGB images
and the last two channels were the normalized x and y coordinates.
The encoder and generator architecture was the same as the one used in the MNIST↔USPS domain
adaptation. For the discriminators and the classifier, we used a network architecture akin to the one
used in the DTN paper. The details are given in Table 6. We used dropout at every layer in the
classifier for avoiding over-fitting.
10
Table 4: Encoder and generator architecture for MNIST↔USPS domain adaptation.
Layer
1
2
3
4
µ
Layer
1
2
3
4
5
Encoders
Shared?
CONV-(N64,K5,S2), BatchNorm, LeakyReLU
CONV-(N128,K5,S2), BatchNorm, LeakyReLU
CONV-(N256,K8,S1), BatchNorm, LeakyReLU
CONV-(N512,K1,S1), BatchNorm, LeakyReLU
CONV-(N1024,K1,S1)
No
Yes
Yes
Yes
Yes
Generators
Shared?
DCONV-(N512,K4,S2), BatchNorm, LeakyReLU
DCONV-(N256,K4,S2), BatchNorm, LeakyReLU
DCONV-(N128,K4,S2), BatchNorm, LeakyReLU
DCONV-(N64,K4,S2), BatchNorm, LeakyReLU
DCONV-(N3,K1,S1), TanH
Yes
Yes
Yes
No
No
Table 5: Discriminator architecture for MNIST↔USPS domain adaptation.
Layer
1
2
3
4a
4b
Discriminators
Shared?
CONV-(N20,K5,S1), MaxPooling-(K2,S2)
CONV-(N50,K5,S1), MaxPooling-(K2,S2)
FC-(N500), ReLU, Dropout
FC-(N1), Sigmoid
FC-(N10), Softmax
No
Yes
Yes
Yes
Yes
Table 6: Discriminator architecture for SVHN→MNIST domain adaptation.
Layer
1
2
3
4
5a
5b
Discriminators
Shared?
CONV-(N64,K5,S1), MaxPooling-(K2,S2)
CONV-(N128,K5,S1), MaxPooling-(K2,S2)
CONV-(N256,K5,S1), MaxPooling-(K2,S2)
CONV-(N512,K5,S1), MaxPooling-(K2,S2)
FC-(N1), Sigmoid
FC-(N10), Softmax
11
No
Yes
Yes
Yes
Yes
Yes
| 1cs.CV
|
Noname manuscript No.
(will be inserted by the editor)
Multi-task, multi-label and multi-domain learning with
residual convolutional networks for emotion recognition
arXiv:1802.06664v1 [cs.CV] 19 Feb 2018
Gerard Pons · David Masip
Received: date / Accepted: date
Abstract Automated emotion recognition in the wild
from facial images remains a challenging problem. Although recent advances in Deep Learning have supposed a significant breakthrough in this topic, strong
changes in pose, orientation and point of view severely
harm current approaches. In addition, the acquisition of
labeled datasets is costly, and current state-of-the-art
deep learning algorithms cannot model all the aforementioned difficulties. In this paper, we propose to apply a multi-task learning loss function to share a common feature representation with other related tasks.
Particularly we show that emotion recognition benefits from jointly learning a model with a detector of
facial Action Units (collective muscle movements). The
proposed loss function addresses the problem of learning multiple tasks with heterogeneously labeled data,
improving previous multi-task approaches. We validate
the proposal using two datasets acquired in non controlled environments, and an application to predict compound facial emotion expressions.
Keywords Facial emotion recognition · Facial Action
Units · Multi-task learning · Convolutional neural
networks
1 Introduction
Images of faces provide relevant information for emotion perception. As humans we can infer an accurate
first impression of somebody’s emotions by observing
their face. Multiple applications benefit from automated
G. Pons and D.Masip
Department of Computer Science, Universitat Oberta de
Catalunya, Spain
E-mail: {gponsro, dmasipr}@uoc.edu
facial emotion recognition, such as human computer interaction (HCI) [4], student engagement estimation [43],
emotionally aware devices [41], or the improvement of
expression production in autism disorder patients [8].
The state-of-the-art in automated facial expression
analysis shows excellent performance in the controlled
scenario, where images are acquired in studio environments. Nevertheless, the categorization of emotions in
the wild, is still an unsolved problem. Besides the strong
intra-class variability, facial expression algorithms in
the wild must also deal with strong local changes in
the illumination conditions, out of plane rotations, large
variations in pose and point of view, and low resolution
imaging.
Recent advances in computer vision and particularly in object recognition suggest that new methods
based on Deep Learning can improve the facial expression recognition task. Convolutional Neural Networks
(CNNs) have represented a relevant breakthrough, especially since the last improvements on the ImageNet
Challenge [28]. However, the amount of available data
for this task is small, especially in all the possible configurations of pose, illumination and resolution. This
supposes an inconvenience to exploit the training capacity of these networks, which need large amounts of
training data. In this context the introduction of multitask learning is particularly relevant, as it proved to
successfully boost the performance of an individual task
with the inclusion of other correlated tasks in the training process [20, 24, 47]. Thus, tasks with small amounts
of data available can benefit from being trained simultaneously with other tasks, sharing a common feature
representation and transferring knowledge between different domains. One of the main difficulties with multitask approaches using different databases is the fact
that not all the samples are labeled for all the tasks.
2
In order to deal with this, classical approaches define
different loss functions for each task and train alternatively for the different domains. This penalizes the
tasks where sample labels are not available. One approach to solve this problem is a selective joint loss,
but it only predicts probabilities for the label set to
which the image belongs. Hence, in order to generate
predictions for all the tasks, in this paper we propose
a multi-label database-wise joint loss to overcome this
problem.
This paper makes the following contributions: (i) we
formalize a novel dataset-wise selective sigmoid crossentropy loss function to simultaneously train a multitask, multi-label and multi-domain model. (ii) We validate that this new proposal outperforms single task
CNNs and the classical multi-task approach using different databases. (iii) We show that the results of the
joint learning of an unlabeled task are coherently correlated to the labeled task in the case of the emotion
recognition problem.
2 Related work
Automated emotion recognition methods from facial
expression analysis in images use a subset of discrete
basic emotions, as defined in [15]. Although the study
of emotional states from faces is grounded in the late
19th century [12], Ekman and Friesen [15] defined a set
of six basic emotions that are shared among all cultures, namely happiness, surprise, anger, sadness, fear,
disgust. The production of these emotions in faces depends on specific facial muscle movements. The Facial
Action Coding System (FACS) [16] defines a set of Action Units (AUs) that atomically group muscle movements. Each expression of emotion can be encoded as
a combination of AU activations.
Facial expression recognition is a multidisciplinary
research field, studied in machine learning, computer
vision, cognitive science, psychology, neuroscience and
applied health sciences. The number of computer vision researchers working in the field of facial expression
analysis has increased since the early 90s, and a large
amount of published works in the topic exist. Depending on the features used for the recognition task, we
can distinguish two prevailing methodologies: geometric
based approaches and appearance based approaches. In
the first case, algorithms focus on localizing and tracking specific fiducial facial landmarks, in order to train
a classifier based on distances and relative positions of
these landmarks. Kotsia and Pitas [27] track a set of
119 key points to classify emotions from sequences, and
Jeni et al. [25] apply a Procrustes transformation on
Gerard Pons, David Masip
a reduced subset of 117 landmarks for emotion recognition. In appearance based emotion recognition, a set
of features is extracted from pixel images to train a
classifier. Classic examples are Gabor Wavelets [4], Histogram of Oriented Gradients (HOG) [29], and Local
Binary Patterns (LBP) [46]. More complex models use
Gaussian processes [17], hybrid methods using mid-level
features [38] or patches from specific key locations [21].
More recently, deep learning methods have outperformed
most state-of-the-art algorithms in the literature. Liu
et al. [30] apply deep learning to a geometric model
of facial regions for facial expression analysis, and Lu
et al. [31] use Convolutional Neural Networks (CNN)
on facial appearance. Recent models obtain improved
results focusing on facial expression recognition in the
wild by using committees of several CNN classifiers [26,
36]. Although most of the previous models focus on
static images, facial expression analysis can benefit from
temporal information. Cohen et al. [9] use Hidden Markov
models on video sequences. Deep learning methods for
emotion perception modeling in video include the use
of CNNs for feature extraction and LSTM for learning
the temporal dynamics [42]. Recent surveys on facial
expression analysis can be found at [11, 33, 39].
Despite the current improvements, emotion recognition in the wild remains an open problem for the
computer vision community. Results from the Emotion
Recognition in the Wild (EmotiW 2015) reveal modest
accuracies [13]. EmotionNet is the biggest Challenge in
terms of data available, with more than 1 million images
(2,000 labeled emotions and 950K unlabeled samples).
The first edition of the challenge concluded that nonfrontal faces still pose major difficulties to automated
algorithms, and recognition rates decrease as a function
of pitch and yaw rotations [5].
Deep learning methods benefit from large training
datasets, but the labeling process of facial expression
data requires expert FACs coders and it is a costly task.
Therefore, most of the current available data sets are
reduced and difficult to acquire. In this paper we conjecture that emotion recognition can be improved if we
transfer knowledge from other related tasks. Particularly we focus on Multi-task learning (MTL). MTL was
first studied by Caruana in [7], where he proposes to
jointly learn parallel tasks sharing a common representation, and transferring part of the knowledge learned
to solve one task to improve the learning of the other
related tasks. Detailed analysis on the topic is carried
out in some recent surveys [45, 34].
Ranjan et al. [37] propose a multi-task approach for
face detection, landmark localization, pose estimation
and gender recognition. This method exploits the benefits of multi-task to improve the performance of each
Multi-task, multi-label and multi-domain learning with residual convolutional networks for emotion recognition
individual tasks. However, they use a single source of
data, hence all the images must be labeled with all the
tasks involved.
The closest work to this paper was published by
Fourure et al. [19]. They propose a multi-task CNN for
semantic segmentation of outdoor images in order to
learn different tasks using images from different databases. They use a selective soft-max cross entropy with
the objective of not penalizing the training of a task
when feeding the model with images from another task.
This approach though does not learn from similar labels. For instance, when training a database A with a
class labeled Grass, and database B with a class labeled Vegetation, the method would predict the label
corresponding to the database where the input image
belonged to (if the image is from database A will predict
it as Grass). In this paper we hypothesize that general
accuracies will improve if we learn from both labels at
the same time, and take full benefit of training related
tasks using different non-homogeneous datasets.
3 Proposed approach
We propose a novel multi-label loss function that can
be integrated into convolutional neural networks to improve the training of the emotion recognition task incorporating complementary tasks and data from different
sources. This proposal can be extended to other multilabel problems, but in this article we focus on the multitask training of emotion recognition and facial Action
Units (AUs).
Emotion recognition is a problem where the model
classifies the input image as one of the 7 emotion classes
[15] (including the neutral class). Therefore, when addressing this task using a CNN, the computation of
the output probabilities is performed by using soft-max
function in the values of the last layer. This setting is
not valid in our multi-task approach, since AU recognition is a multi-label problem. The classical setting
for multi-label approaches is the use of the sigmoid
function for the calculation of the output probabilities.
However, using images from different sources and for
different tasks implies having partially unlabeled data
when performing the multi-task training, i.e. not all
data is labeled with respect to all the tasks. Thus, this
approach is not valid when dealing with images from
different databases, since the sigmoid function penalizes
the absence of labels in the related task. For example,
optimizing the Happy class using samples with no AU
annotations would penalize the AU6 and AU12 classes
provoking a non-desired effect since these AUs together
form the emotion Happy. We thus propose the Selective Joint Multi-task (SJMT) approach which defines a
3
novel dataset-wise selective sigmoid cross-entropy loss
function to address multi-task, multi-label and multidomain problems.
Given a set of images from k different datasets,
where the label spaces are different, we define yjk the
actual label of the j-th class in the space Lk . The aim
is to learn a nonlinear mapping presented as a CNN
which minimizes a cross-entropy loss function for each
individual sample. This function is defined as follows:
E(ŷ, y, k) =
−1 X
[yj log yˆj + (1 − yˆj ) log (1 − yˆj )](1)
N
k
j∈L
where N is the number of classes, k is the number of
different datasets, each defined on its own label set Lk .
The term yj is the actual label of the j-th class, and yˆj
is a sigmoid function defined as:
yˆj =
1
1 + e−p̂j
(2)
where p̂j is the output of the last layer of the CNN
defined as f (W h + b), f is the activation function of
the layer, W and b the weights and bias of the layer
and h the hidden representation of the last layer.
Figure 1 shows the different approaches compared in
this article. First Figure 1(a) shows the approach where
individual CNN are tailored to each task and database.
Second Figure 1(b) shows the classical multi-task approach where the filters of the network are shared during training among the different tasks, taking into account that each task has an specific cross-entropy loss
function and output layer, and the training is performed
alternatively feeding the model with batches of images
from each dataset. Finally, Figure 1(c) shows the approach proposed in this paper, where the whole network
is shared among the different tasks and images from
different databases are feed indistinctly thanks to the
selective cross-entropy function defined in equation 1.
Even though this cross-entropy function could be incorporated in any CNN architecture, in this paper we
used a ResNet [22] to implement the multi-task, multilabel and multi-database approach. Residual networks
have shown great robustness in different fields [18, 35,
44] given their framework which eases the training of
these networks, and enables them to be substantially
deeper and, thus, obtain better performances than other
networks.
4 Experimental results
In this paper we used the following databases for the
experimental validation:
4
Gerard Pons, David Masip
(a)
(b)
(c)
Fig. 1 Different CNN strategies used in this paper. (a) shows the traditional approach with different CNNs for different tasks.
(b) shows the classical multi-task approach with different output layers and loss functions for each individual task. (c) shows
the proposed multi-task approach with a single output layer and the selective sigmoid cross entropy function.
– SFEW2.0 database [13], released for a competition in the 3rd Emotion Recognition In the Wild
2015 (EmotiW2015) challenge. The database was
created by extracting frames from film clips with
emotional content in order to obtain images in closeto-real world conditions. The images are labeled with
7 expressions (angry, disgust, fear, happy, sad, surprise, and neutral). The database consists of 958
images for training, 436 for validation, and 372 for
testing. Since the labels of the test dataset are not
provided, the validation set was used as test set for
the experiments in this work. The database also provides the same data after being processed with a
face alignment algorithm. We used this set in the
experiments of this paper.
– EmotioNet database [6], released for the EmotioNet Challenge 2017 [5]. The database consists
of 25,000 images of facial expressions with manual
AU annotations. These annotations were given by
expert human coders and cross referenced for verification of accuracy. From these images, a subset
of 2,000 images were annotated with the 7 emotion
expressions and 9 compound emotions made by the
combination of two basic emotions. Since the test
dataset is private and was only available for participants of the challenge, the test set for the experiments in this work was generated randomly selecting
1,000 images (the same in all the experiments).
– Extended Cohn-Kanade database (CK+) [32],
consists of 593 posed video sequences recorded from
123 university students ranging from 18 to 30 years
old. The subjects were asked to express a series of
facial displays including single or combined action
units. The database also provides labels for the presence of emotions in these sequences. In this paper,
we randomly generated a subset of 100 images for
testing the methods. This database was acquired in
controlled environments, and it has been used only
to determine the parameters used for the neural networks. Theses paremeters were consistent in all the
experiments.
For ResNet [22] we used a training batch size of
128 images, and learning rate of 0.05 and exponential
decay every 10,000 steps, for a total of 80,000 steps.
Batch normalization and data augmentation by means
of padding and random cropping images during training were used to prevent overfitting. Input data of size
32 × 32 was used to pre-load CIFAR10 weights [23].
Multi-task, multi-label and multi-domain learning with residual convolutional networks for emotion recognition
In the comparison among individual networks we also
used VGG [40]. For this network we used a training
batch size of 32 images, and an exponentially decayed
learning rate of 0.00001 during 2,000 steps. To prevent
the network from overfitting a dropout rate of 0.5 was
used. In order to pre-load the trained weights used by
the VGG team in the ILSVRC-2014 competition, input
images of size 224 × 224 were used.
All the networks trained in this paper were pretrained using the Facial Expression Recognition (FER)
2013 database [1], which consists of 28,709 examples
for training, 3,589 for validation and 3,589 for testing.
All the experiments in this work were developed using
TensorFlow [2] on NVIDIA GeForce GTX Titan GPU.
With regard to the detection of AUs and recognition of emotion categories, the evaluation criteria used
was accuracy. Accuracy measures the number correctly
classified examples, indicating whether our algorithm is
able to discriminate between sample images with a certain AU/emotion present. In statistics, this difference
between the measured and true value is generally called
observational error, and it is defined as follows:
accuracyi =
T Pi + T Ni
N
5
Table 1 Results of different methods for the emotion and
AU recognition tasks in percentage of accuracy.
Emotion
AUs
Method
SFEW
CK+
EmotioNet
CK+
94.5%
Baseline
35.9%
-
-
OpenFace
-
-
87.7%
-
Azure
33.8%
-
-
-
VGG-16
37.5%
93.0%
93.2%
95.0%
ResNet-32
37.8%
98.0%
93.2%
99.0%
ResNet-110
41.3%
99.0%
93.9%
99.1%
used in the networks we needed to train (VGG and
ResNets).
Table 1 shows the accuracy obtained for each CNN
for the different tasks and databases. Note that the images were not pre-processed specifically to enhance the
performance in one database or another, and the same
parameters were used in all the databases for a fare
comparison. Hence, the most noticeable fact are the results obtained by ResNet-110, which are superior to the
rest of the methods in the comparison.
(3)
where i specifies the class, i.e., AUi or the i-th emotion
category, T Pi (true positives) are correctly identified
test instances of class i, T Ni (true negatives) are test
images correctly labeled as not belonging to class i, and
N is the total number of test images.
4.1 Emotion and AU recognition with individual
networks
Before validating the proposed approach in this paper
we needed to validate the performance of each task
trained with dedicated individual networks for such task
and database. In addition, the baseline accuracy for
each database is reported if known.
Thus, we trained specific networks for emotion recognition (see Figure 1(a)) with the SFEW and CK+ databases, and for AU detection with the EmotioNet and
CK+ databases. Given their outstanding results in different fields, we decided to compare two of the most
used architectures in image processing, VGG-16 and
ResNet (with different sizes: 32 and 110). In this comparison we also added the results of other state-of-the
art commercial methods, such as the Microsoft Azure
emotion recognition API [10] and the OpenFace framework for AU detection [3]. Note that these methods
were not validated with the CK+ database, since we
used this database only for defining the hyper-parameters
4.2 Multi-task learning for emotion and AU
recognition
In this section we want to validate the hypothesis that
the proposed SJMT method (see Figure 1(c)) outperforms the individual networks compared in the previous section as well as the classical multi-task approach
(see Figure 1(b)). Given the results from the individual network comparison we decided to use the two best
architectures in this comparison: ResNet 32 and 110.
For this experiment we jointly trained the multitask approaches for emotion and AU recognition with
the SFEW and EmotioNet databases respectively and
the obtained results are summarized in Table 2. As occurred in the previous experiment, ResNet 110 generally obtained better results than ResNet 32. Comparing
both the SJMT and classical multi-task approaches, the
proposed method outperforms the classical approach
in all the experiments except for AU recognition with
ResNet 32. Finally, SJMT with ResNet 110 obtained
the best results for both emotion and AU recognition,
compared to the classical multi-task, and the single task
approaches shown in Table 1
In order to analyze the benefits of using the proposed SJMT approach with related tasks, we focused
on the predictions for AUs when feeding the model with
images from the SFEW database, which only provides
labels for emotion recognition. Thus, we grouped the
images by the predicted emotion category and com-
6
Gerard Pons, David Masip
Table 2 Results of multi-task learning for the emotion and
AU recognition tasks in percentage of accuracy. Multi-task
refers to the classical approach (see Figure 1(b)) and SJMT
refers to the Selective Joint Multi-task (SJMT) approach (see
Figure 1(c))
Method
Multi-task
SJMT
Network
SFEW
EmotioNet
(Emotion)
(AUs)
ResNet-32
39.4%
93.7%
ResNet-110
40.3%
93.6%
ResNet-32
40.6%
93.5%
ResNet-110
45.9%
93.9%
Table 3 Description of the recognized AUs for each emotion.
Emotion
Angry
Disgust
Fear
Happy
Fig. 2 Mean scores for AUs in SFEW images where only
emotion labels are provided using the SJMT approach.
puted the mean scores for each AU as shown in Figure 2.
This figure shows that when the model recognizes an
emotion category, is simultaneously capable to identify
the corresponding AUs in most of the cases. In order to
simplify the interpretation of the figure, Table 3 shows
the description of the AUs predicted by SJMT along
with each emotion and a representative image from the
test dataset of the SFEW database.
The table shows the coherence of the approach when
recognizing the AUs in SFEW images. Given the fact
that only 11 AU labels are used in the EmotioNet database (which is the database used to train the SJMT
approach), some of the AUs theoretically forming the
emotions cannot be recognized. A clear example is the
Sad emotion, where the most discriminant AU in theory
is 15 (Lip Corner Depressor) and the model recognizes
it as 12 (Lip Corner Puller).
4.3 Compound emotion recognition
Most of the research on emotion recognition has focused on the study of the seven basic categories. We use
the term “basic” to refer to the fact that such emotion
categories cannot be decomposed into smaller semantic labels. However, it has been shown [14] that many
Sad
Surpise
AU
Description
4
Brow Lowerer
9
Nose Wrinkler
25
Lips Part
1
Inner Brow Raiser
4
Brow Lowerer
6
Cheek Raiser
5
Upper Lid Raiser
25
Lips Part
6
Cheek Raiser
12
Lip Corner Puller
25
Lips Part
12
Lip Corner Puller
17
Chin Raiser
2
Outer Brow Raiser
25
Lips Part
26
Jaw Drop
Example
more facial expressions of emotion exist and are used
regularly by humans. Compound emotions are those
that can be constructed by combining basic component
categories to create new ones. Hence, in this section
we will validate our method on the extended problem
of recognizing the following compound emotions from
the EmotioNet database: angrily disgusted, angrily surprised, fearfully angry, fearfully surprised, happily disgusted, happily surprised, sadly angry, and sadly disgusted. The number of available samples is scarce for
most of the classes, and we hypothesize that this emotion prediction task can take benefit from a joint learning process along with an auxiliary task such as AU
recognition.
Following the experiments of the previous sections,
here we compare the performance of the SJMT approach for compound emotion images with a single ded-
Multi-task, multi-label and multi-domain learning with residual convolutional networks for emotion recognition
icated task network, in both cases training a ResNet
110. Given the unbalanced number of images per class
(see Table 4), we decided to train the networks with
a training set of 160 images consisting of 15 images in
each class.
Table 4 Results of single and multi-task learning for compound emotion recognition tasks in percentage of accuracy.
Compound
emotion
Number
Single
SJMT
images
ResNet-110
ResNet-110
Angrily disgusted
19
15.8%
42.1%
Angrily surprised
25
36.0%
64.0%
Fearfully angry
19
15.8%
63.1%
Fearfully surprised
17
58.8%
70.6%
Happily disgusted
486
47.6%
78.6%
Happily surprised
36
61.1%
72.2%
Sadly angry
15
13.3%
20.0%
Sadly disgusted
105
31.4%
64.7%
Mean compounds
-
34.9%
59.4%
All images
722
43.4%
73.1%
Table 4 shows that also when dealing with more
complex emotions the SJMT network outperforms the
single one, globally and for each of the classes. The
improvement on the performance of the classes with
the most examples (happily disgusted and sadly disgusted) also explains the large difference between methods when comparing globally, since much more images
from the total are correctly classified. However, when
comparing the mean of all the class accuracies we can
see that SJMT with 59.4% considerably improves the
single network performance (34.9%).
5 Conclusions
In this paper we presented the Selective Joint Multi-task
approach which defines a novel dataset-wise selective
sigmoid cross-entropy loss function to address multitask, multi-label and multi-database problems. Specifically, we addressed this proposal to overcome one of the
challenges with emotion recognition in the wild, which
is the lack of large public labeled databases by including
AU recognition in a multi-task approach.
Our proposed approach was assessed using the SFEW
database for emotion recognition and the EmotioNet
database for AU detection, and the results were compared to single task networks dedicated to each task
individually and to the classical multi-task approach.
SJMT obtained the best results in terms of accuracy
for all the experiments. Results also showed the benefits of learning multiple correlated tasks simultaneously
7
by demonstrating visually that even for images without AU labels, the model is capable of inferring them
in a coherent way. Finally, we evaluated the proposed
method in a database of compound emotions, again obtaining improved results.
Future work will consist of studying and developing the inclusion of the landmark detection task in this
multi-task approach. We believe that the addition of
structural and geometrical cues to the appearance based
CNNs might boost the performance of the emotion recognition task, but a new reformulation of the selective
cross-entropy function needs to be done in order to handle such a problem in a joint training.
Acknowledgements This research was supported by TIN201566951-C2-2-R grant from the Spanish Ministry of Economy
and Competitiveness and NVIDIA Hardware grant program.
References
1. Facial
Expression
Recognition
2013
database.
https://www.kaggle.com/c/
challenges-in-representation-learning-facial-expression-recogn
data (2013)
2. Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen,
Z., Citro, C., Corrado, G.S., Davis, A., Dean, J., Devin,
M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G.,
Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M.,
Levenberg, J., Mané, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B.,
Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V., Viégas, F., Vinyals, O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., Zheng, X.: TensorFlow:
Large-scale machine learning on heterogeneous systems
(2015). URL http://tensorflow.org/. Software available from tensorflow.org
3. Baltruaitis, T., Mahmoud, M., Robinson, P.: Crossdataset learning and person-specific normalisation for automatic action unit detection. In: 2015 11th IEEE International Conference and Workshops on Automatic Face
and Gesture Recognition (FG), vol. 06, pp. 1–6 (2015)
4. Bartlett, M.S., Littlewort, G., Fasel, I., Movellan, J.R.:
Real time face detection and facial expression recognition: Development and applications to human computer
interaction. In: Computer Vision and Pattern Recognition Workshop, 2003. CVPRW’03. Conference on, vol. 5,
pp. 53–53. IEEE (2003)
5. Benitez-Quiroz, C.F., Srinivasan, R., Feng, Q., Wang, Y.,
Martinez, A.M.: Emotionet challenge: Recognition of facial expressions of emotion in the wild. arXiv preprint
arXiv:1703.01210 (2017)
6. Benitez-Quiroz, C.F., Srinivasan, R., Martinez, A.M.:
Emotionet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the
wild. In: 2016 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pp. 5562–5570 (2016)
7. Caruana, R.: Multitask learning. In: Learning to learn,
pp. 95–133. Springer (1998)
8. Cockburn, J., Bartlett, M., Tanaka, J., Movellan, J.,
Pierce, M., Schultz, R.: Smilemaze: A tutoring system in
real-time facial expression perception and production in
8
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
Gerard Pons, David Masip
children with autism spectrum disorder. In: ECAG 2008
Workshop Facial and Bodily Expressions for Control and
Adaptation of Games, p. 3. Citeseer (2008)
Cohen, I., Sebe, N., Garg, A., Chen, L.S., Huang, T.S.:
Facial expression recognition from video sequences: temporal and static modeling. Computer Vision and image
understanding 91(1-2), 160–187 (2003)
Copeland, M., Soh, J., Puca, A., Manning, M., Gollob,
D.: Microsoft Azure: Planning, Deploying, and Managing
Your Data Center in the Cloud, 1st edn. Apress, Berkely,
CA, USA (2015)
Corneanu, C.A., Oliu, M., Cohn, J.F., Escalera, S.: Survey on rgb, 3d, thermal, and multimodal approaches for
facial expression recognition: History, trends, and affectrelated applications. IEEE Transactions on Pattern Analysis and Machine Intelligence (99), 1 (2016). DOI
10.1109/TPAMI.2016.2515606
Darwin, C.: The expression of emotion in animals and
man. London, England: Murray (1872)
Dhall, A., Ramana Murthy, O., Goecke, R., Joshi, J.,
Gedeon, T.: Video and image based emotion recognition
challenges in the wild: Emotiw 2015. In: Proceedings
of the 2015 ACM on International Conference on Multimodal Interaction, pp. 423–426. ACM (2015)
Du, S., Tao, Y., Martinez, A.M.: Compound facial expressions of emotion. Proceedings of the National Academy
of Sciences 111(15), E1454–E1462 (2014)
Ekman, P., Friesen, W.V.: Constants across cultures in
the face and emotion. Journal of personality and social
psychology 17(2), 124 (1971)
Ekman, P., Friesen, W.V.: Facial action coding system: A
technique for the measurement of facial movement. CA:
Consulting Psychologists Press. Ellsworth, PC, & Smith,
CA (1988). From appraisal to emotion: Differences among
unpleasant feelings. Motivation and Emotion 12, 271–302
(1978)
Eleftheriadis, S., Rudovic, O., Pantic, M.: Discriminative shared gaussian processes for multiview and viewinvariant facial expression recognition. IEEE transactions
on image processing 24(1), 189–204 (2015)
Feichtenhofer, C., Pinz, A., Wildes, R.: Spatiotemporal
residual networks for video action recognition. In: D.D.
Lee, M. Sugiyama, U.V. Luxburg, I. Guyon, R. Garnett
(eds.) Advances in Neural Information Processing Systems 29, pp. 3468–3476. Curran Associates, Inc. (2016)
Fourure, D., Emonet, R., Fromont, E., Muselet, D.,
Neverova, N., Trémeau, A., Wolf, C.: Multi-task, multidomain learning: application to semantic segmentation
and pose regression. Neurocomputing 251, 68–80 (2017)
Ganin, Y., Lempitsky, V.: Unsupervised domain adaptation by backpropagation. In: Proceedings of the 32Nd
International Conference on International Conference on
Machine Learning - Volume 37, ICML’15, pp. 1180–1189
(2015)
Happy, S., Routray, A.: Automatic facial expression
recognition using features of salient facial patches. IEEE
transactions on Affective Computing 6(1), 1–12 (2015)
He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning
for image recognition. arXiv preprint arXiv:1512.03385
(2015)
He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in
deep residual networks. arXiv preprint arXiv:1603.05027
(2016)
Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. arXiv preprint 1503.02531
(2015)
25. Jeni, L.A., Takacs, D., Lorincz, A.: High quality facial
expression recognition in video streams using shape related information only. In: Computer Vision Workshops
(ICCV Workshops), 2011 IEEE International Conference
on, pp. 2168–2174. IEEE (2011)
26. Kim, B.K., Lee, H., Roh, J., Lee, S.Y.: Hierarchical committee of deep cnns with exponentially-weighted decision
fusion for static facial expression recognition. In: Proceedings of the 2015 ACM on International Conference
on Multimodal Interaction, pp. 427–434. ACM (2015)
27. Kotsia, I., Pitas, I.: Facial expression recognition in image
sequences using geometric deformation features and support vector machines. Image Processing, IEEE Transactions on 16(1), 172–187 (2007)
28. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet
classification with deep convolutional neural networks.
In: Advances in Neural Information Processing Systems,
p. 2012
29. Li, Z., Imai, J.i., Kaneko, M.: Facial-component-based
bag of words and phog descriptor for facial expression
recognition. In: Systems, Man and Cybernetics, 2009.
SMC 2009. IEEE International Conference on, pp. 1353–
1358. IEEE (2009)
30. Liu, M., Li, S., Shan, S., Wang, R., Chen, X.: Deeply
learning deformable facial action parts model for dynamic
expression analysis. In: Asian conference on computer
vision, pp. 143–157. Springer (2014)
31. Lu, G.m., He, J.l., Yan, J.j., Li, H.: Convolutional neural
network for facial expression recognition. Journal of Nanjing University of Posts and Telecommunications 36(1),
16–22 (2016)
32. Lucey, P., Cohn, J.F., Kanade, T., Saragih, J., Ambadar,
Z., Matthews, I.: The extended cohn-kanade dataset
(ck+): A complete dataset for action unit and emotionspecified expression. In: 2010 IEEE Computer Society
Conference on Computer Vision and Pattern Recognition
- Workshops, pp. 94–101 (2010)
33. Martinez, B., Valstar, M.F., Jiang, B., Pantic, M.: Automatic analysis of facial actions: A survey. IEEE Transactions on Affective Computing (2017)
34. Pan, S.J., Yang, Q.: A survey on transfer learning. IEEE
Transactions on knowledge and data engineering 22(10),
1345–1359 (2010)
35. Pohlen, T., Hermans, A., Mathias, M., Leibe, B.: Fullresolution residual networks for semantic segmentation
in street scenes. In: 2017 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pp. 3309–3318
(2017)
36. Pons, G., Masip, D.: Supervised committee of convolutional neural networks in automated facial expression
analysis. IEEE Transactions on Affective Computing
(2017)
37. Ranjan, R., Patel, V.M., Chellappa, R.: Hyperface: A
deep multi-task learning framework for face detection,
landmark localization, pose estimation, and gender recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)
38. Sanchez-Mendoza, D., Masip, D., Lapedriza, A.: Emotion
recognition from mid-level features. Pattern Recognition
Letters 67, 66–74 (2015)
39. Sariyanidi, E., Gunes, H., Cavallaro, A.: Automatic analysis of facial affect: A survey of registration, representation, and recognition. IEEE transactions on pattern analysis and machine intelligence 37(6), 1113–1133 (2015)
40. Simonyan, K., Zisserman, A.: Very deep convolutional
networks for large-scale image recognition.
CoRR
abs/1409.1556 (2014)
Multi-task, multi-label and multi-domain learning with residual convolutional networks for emotion recognition
41. Soleymani, M., Pantic, M.: Emotionally aware tv. In:
Proceedings of TVUX-2013: Workshop on Exploring and
Enhancing the User Experience for TV at ACM CHI
(2013)
42. Sun, B., Wei, Q., Li, L., Xu, Q., He, J., Yu, L.: Lstm for
dynamic emotion and group emotion recognition in the
wild. In: Proceedings of the 18th ACM International Conference on Multimodal Interaction, pp. 451–457. ACM
(2016)
43. Whitehill, J., Serpell, Z., Lin, Y.C., Foster, A., Movellan,
J.R.: The faces of engagement: Automatic recognition of
student engagementfrom facial expressions. IEEE Transactions on Affective Computing 5(1), 86–98 (2014)
44. Yu, L., Chen, H., Dou, Q., Qin, J., Heng, P.A.: Automated melanoma recognition in dermoscopy images via
very deep residual networks. IEEE Transactions on Medical Imaging 36(4), 994–1004 (2017)
45. Zhang, Y., Yang, Q.: A survey on multi-task learning.
arXiv preprint arXiv:1707.08114 (2017)
46. Zhao, G., Pietikainen, M.: Dynamic texture recognition
using local binary patterns with an application to facial
expressions. IEEE transactions on pattern analysis and
machine intelligence 29(6), 915–928 (2007)
47. Zhu, X., Ramanan, D.: Face detection, pose estimation,
and landmark localization in the wild. In: 2012 IEEE
Conference on Computer Vision and Pattern Recognition, pp. 2879–2886 (2012)
9
| 1cs.CV
|
NON-INNER AMENABILITY OF THE THOMPSON
GROUPS T AND V
arXiv:1609.05086v1 [math.OA] 16 Sep 2016
UFFE HAAGERUP AND KRISTIAN KNUDSEN OLESEN
Abstract. In this paper we prove that the Thompson groups T and V
are not inner amenable. In particular, their group von Neumann algebras do not have property Γ. Moreover, we prove that if the reduced
group C ∗ -algebra of T is simple, then the Thompson group F is nonamenable. Furthermore, we give a few new equivalent characterizations
of amenability of F .
1. Introduction
The Thompson groups F , T and V where introduced by Richard Thompson in 1965. They are countable discrete groups of piecewise linear bijections
of the half open interval [0, 1) onto itself and satisfy F ⊆ T ⊆ V . We refer
to the paper [5] by Cannon, Floyd and Parry for a detailed introduction
to the subject. It is well-known that the Thompson groups T and V are
non-amenable, but it is a major open problem to decide whether or not
F is amenable. All three groups are ICC, that is, all conjugacy classes of
non-trivial elements are infinite, and hence their von Neumann algebras LF ,
LT and LV are all factors of type II1 . Jolissaint [14] proved that F is inner amenable in the sense of Effros [9]. He later strengthened this result
by proving that LF is a McDuff factor; see [15]. In particular, this means
that LF has property Γ, and hence, by a result of Effros [9], F is inner
amenable. The first and main result of this paper is that T and V are not
inner amenable, and hence, by the same result of Effros, LT and LV do not
have property Γ. Therefore, they are also not McDuff factors. The fact that
F is inner amenable was proved in a different way by Ceccherini-Silberstein
and Scarabotti [6], a few years later.
Another result of this paper connects amenability of F with simplicity
of the reduced group C ∗ -algebra of T . More precisely, we prove that if the
reduced group C ∗ -algebra Cr∗ (T ) is simple, then F is non-amenable. Shortly
after the results of this paper where announced by the first named author
at the Fields Institute in Toronto in March 2014, this result was re-obtained
by Breuillard, Kalantar, Kennedy and Ozawa [3], using completely different methods. Very recently, this result was also re-proved by Le Boudec and
Matte Bon [19]. They also showed the converse of this statement. This result
Date: September 19, 2016.
Both authors are supported by the ERC Advanced Grant no. OAFPG 247321 and the
Danish National Research Foundation through the Centre for Symmetry and Deformation
(DNRF92). The first author is also supported by the Danish Natural Science Research
Council.
1
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
2
is not the only connection between amenability of F and simplicity of its reduced group C ∗ -algebra. Indeed, it has been well-known for some time that
F is amenable if and only if its reduced group C ∗ -algebra has a unique tracial
state, and it has been a long-standing open problem set forth by de la Harpe
(see, for example, [7]) to decide whether the reduced group C ∗ -algebra of a
group is simple if and only if it has a unique tracial state. During the last few
years tremendous progress has been made on the topic of C ∗ -simple groups
(that is, groups whose reduced group C ∗ -algebra is simple) and groups with
the unique trace property (that is, groups whose reduced group C ∗ -algebra
has a unique tracial state). This new development began when Kalantar
and Kennedy [16] gave new characterizations of C ∗ -simple groups in terms
of boundary actions. Immediately after, Breuillard, Kalantar, Kennedy and
Ozawa [3] proved several striking results, including that C ∗ -simplicity implies
the unique trace property. Very recently, Le Boudec, [18], gave counterexamples to the remaining implication of the long-standing open problem of de
la Harpe by providing examples of groups which are not C ∗ -simple, but have
the unique trace property. This settled completely the relationship between
C ∗ -simplicity and the unique trace property. Further examples of non-C ∗ simple groups with the unique trace property were provided by Ivanov and
Omland [13].
In addition to the results mentioned thus far, we investigate the C ∗ -algebras generated by the images of F , T and V via a representation discovered
by Nekrashevych [20]. More precisely, we prove that these C ∗ -algebras are
distinct, and that that one generated by the image of V is the Cuntz algebra O2 . Furthermore, we give new characterizations of amenability of F in
terms of the size of certain explicit ideals in the reduced group C ∗ -algebras
of F and T , as well as in terms of whether certain convex hulls contain zero.
We end the introduction by giving the definition of the Thompson groups.
The Thompson group V is the set of all piecewise linear bijections of [0, 1)
which are right continuous, have finitely many points of non-differentiability,
all being dyadic rationals, and have a derivative which is a power of 2 at each
point of differentiability. It is easy to see that these maps fix the set of dyadic
rationals in [0, 1). The Thompson group T consists of those elements of V
which define a homeomorphism of [0, 1), when identified with R/Z in the
standard way. Equivalently, T is the set of elements in V which have at
most one point of discontinuity with respect to the usual topology on [0, 1).
Finally, the Thompson group F consists of those elements of V which are,
in fact, homeomorphisms of [0, 1), or, equivalently, the set of element g ∈ T
satisfying g(0) = 0. All three groups are finitely generated and also finitely
presented in the following generators: F is generated by A and B; T is
generated by A, B and C; and V is generated by A, B, C and π0 , where
A, B, C and π0 are defined in [5]. Besides these standard generators, we
will also consider the following element D of T , given by D(x) = x + 34 ,
for x ∈ [0, 14 ), and D(x) = x − 41 for x ∈ [ 14 , 1). Note that C 3 = D 4 = 1.
Moreover, the elements C and D also generate T as a group, as it is easily
seen that A = D 2 C 2 and B = C 2 DA.
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
3
2. The isomorphism between T and PPSL(2, Z)
In order to prove that T and V are not inner amenable, we shall use that T
is isomorphic to a group of piecewise fractional linear transformations. This
fact was originally proved by Thurston, as explained in [5, §7]. However,
Thurston realized T as Möbius transformations of the interval [0, 1], whereas
we will use a slightly modified version of Thurston’s result, given by Imbert
in [12, Theorem 1.1], where T is realized as Möbius transformations of R ∪
{∞} instead. The construction is essentially the same, and we shall explain
the difference in Remark 2.3 below. A description of the isomorphism is also
given by Fossas in [10]. The second named author would like to thank Vlad
Sergiescu for several fruitful discussions and his help in sorting out the origin
of the particular version of Thurston’s result we are presenting here.
Recall that the group PSL(2, Z) = SL(2, Z)/{±1} acts in a natural way
on R ∪ {∞} via Möbius transformations, that is, for x ∈ R ∪ {∞},
ax + b
a b
(1)
g(x) =
{±1} ∈ PSL(2, Z).
,
when g =
c d
cx + d
This action is faithful, in the sense that only the neutral element acts trivially.
Recall also that the elements of PSL(2, Z) act by homeomorphisms in an
orientation-preserving way.
Definition 2.1. We denote by PPSL(2, Z) the group of homeomorphisms
of R∪{∞} which are piecewise in PSL(2, Z), that is, piecewise of the form (1),
and which have only finitely many breakpoints, all them being in Q ∪ {∞}.
In the above definition, a breakpoint of an element g ∈ PPSL(2, Z) should
be understood as a point x ∈ R ∪ {∞} for which there is no open neighbourhood U so that g acts on U as an element of PSL(2, Z).
Theorem 2.2 (Thurston). There exists a homeomorphism φ of R ∪ {∞}
onto R/Z = S 1 such that the map Φ : PPSL(2, Z) → T given by
g 7→ φ ◦ g ◦ φ−1
is an isomorphism.
As mentioned, Thurston’s construction of the homeomorphism, see [5,
§7], is different than the one in the theorem above, which is briefly explained
in [12]. More details are given by Fossas in the proof of Theorem 2.2 in [10].
More precisely, one first defines an order preserving homeomorphism ψ of
[−∞, ∞] onto [− 12 , 12 ], and next obtains φ : R ∪ {∞} → R/Z from ψ by
identifying the endpoints of each of the two intervals, that is, φ(x) = ψ(x)+Z,
for all x ∈ R, and φ(∞) = 12 + Z. The map ψ is denoted by “?” in [10]. It is
closely related to (but different from) the Minkowski question mark function,
as explained in Remark 2.3 below. To construct ψ, one writes each element
in Q ∪ {±∞} as a reduced fraction pq with p, q ∈ Z and q ≥ 0, using the
convention that
−∞ = −1
and
∞ = 10 .
0
1
Two reduced fractions pq and rs (possibly −1
0 and 0 ) are called consecutive
Farey numbers if |ps − rq| = 1. Given two consecutive Farey numbers one
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
4
defines
p r
p+r
⊕ =
.
q s
q+s
Let now ψ be defined on Q ∪ {±∞} by letting ψ(−∞) = − 12 , ψ(0) = 0,
ψ(∞) = 12 , and then recursively let ψ( pq ⊕ rs ) = 12 ψ( pq ) + ψ( rs ) , whenever
p
r
q and s are consecutive Farey numbers. As explained in [10], this map
is well-defined, and it is a strictly increasing bijection of Q ∪ {±∞} onto
Z[ 12 ] ∩ [− 21 , 12 ]. Therefore, it has a unique extension to a strictly increasing
homeomorphism, also called ψ, of [−∞, ∞] onto [− 21 , 12 ].
Remark 2.3. Let ? : [0, 1] → [0, 1] denote the Minkowski question mark function (see, for example, [22]). Then ? is constructed recursively in the same
way as ψ, but on the interval [0, 1]. However, one starts with the assignments
?(0) = 0 and ?(1) = 1,
while in the case of ψ (on the interval [0, 1]) we have
ψ(0) = 0 and ψ(1) = 14 .
Therefore ψ(x) = 41 ?(x), for 0 ≤ x ≤ 1. Moreover, it is not hard to check that
ψ satisfies the following symmetries ψ(−x) = −ψ(x) and ψ( x1 ) = 12 − ψ(x),
for x ∈ [0, ∞]. Hence, ψ can be expressed in terms of the Minkowski question
mark function by
1 1
− + ?(− x1 ) x ∈ [−∞, −1]
21 4
x ∈ [−1, 0]
− ?(−x)
.
ψ(x) = 1 4
x ∈ [0, 1]
4 ?(x)
1 1 1
x ∈ [1, ∞]
2 − 4 ?( x )
Thurston’s version of Theorem 2.2 states that T is isomorphic to the
group of homeomorphisms of [0, 1] with the endpoints identified, which are
piecewise in PSL(2, Z), have only finitely many breakpoints, all of which
being in Q ∩ [0, 1]. This isomorphism is realized by conjugating by the
Minkowski question mark function ?, instead of ψ.
It is well-known (see [23, Example 1.5.3]) that PSL(2, Z) is isomorphic to
Z2 ∗ Z3 = ha, b | a2 = b3 = 1i, with generators
0 −1
0 −1
{±1}.
{±1}
and
b=
a=
1 1
1 0
Note that PSL(2, Z) is a subgroup of PPSL(2, Z). Hence by [10, Remark 2.3],
we have the following result.
Proposition 2.4. Let Φ : PPSL(2, Z) → T be the isomorphism from Theorem 2.2, and let Λ denote the subgroup Φ(PSL(2, Z)) of T . Then
D 2 = CA = Φ(a)
and
C = Φ(b)
∼ Z2 ∗ Z3 , of order 2 and 3, respectively.
are free generators of Λ =
We will also need the following well-known (and easy to prove) fact about
the action of the group PSL(2, Z) on R ∪ {∞} by Möbius transformations.
Proposition 2.5. An element f ∈ PSL(2, Z) is uniquely determined by its
value on two distinct points in Q ∪ {∞}.
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
5
Remark 2.6. Note that we can translate the above result to one about Λ, by
using the homeomorphism φ. Indeed, φ restricts to a bijection from Q ∪ {∞}
to [0, 1) ∩ Z[ 21 ], so an element in Λ is uniquely determined by its value on
two points in [0, 1) ∩ Z[ 21 ].
3. Non-inner amenability of T and V
In this section we prove that the Thompson groups T and V are not inner
amenable. This answers a question that Ionut Chifan raised at a conference in Alba-Iulia in 2013. We start by recalling the fundamentals of inner
amenability.
The reduced group C ∗ -algebra associated to a discrete group G, denoted
by Cr∗ (G), is the C ∗ -algebra generated by the image of the left regular representation of G, λ : G → B(ℓ2 (G)), while the group von Neumann algebra
associated to G, denoted by LG, is the weak operator closure of Cr∗ (G) inside
B(ℓ2 (G)). The group von Neumann algebra is a type II1 -factor if and only
if the group is ICC.
The von Neumann algebra LG of a discrete ICC group G is said to have
property Γ, if there exists a net (ui )i∈I of unitaries in LG such that hui δe |
δe i = 0, for all i ∈ I, and limi∈I k(ui x − xui )δe k = 0, for all x ∈ LG.
Property Γ can be defined for general finite factors, but we are only concerned
here with the case where the von Neumann algebra is a group von Neumann
algebra.
Recall that a discrete group G is said to be inner amenable, if there exists
a state m on ℓ∞ (G) such that m(δe ) = 0 and m(f ) = m(α(g)f ), for all
f ∈ ℓ∞ (G) and g ∈ G, where (α(g)f )(h) = f (g −1 hg), for all h ∈ G. Inner
amenability was introduced by Effros in 1975 in an attempt to characterize
property Γ for group von Neumann algebras in terms of a purely group
theoretic property; see [9]. Therein he proved the following theorem.
Theorem 3.1 (Effros). For a discrete ICC group G, the following statements
are equivalent:
(i) G is inner amenable.
(ii) There exists a net (ηi )i∈I of unit vectors in ℓ1 (G) satisfying ηi (e) = 0,
for all i ∈ I, and limi∈I kα(g)ηi − ηi k1 = 0, for all g ∈ G.
(iii) There exists a net (ξi )i∈I of unit vectors in ℓ2 (G) satisfying ξi (e) = 0,
for all i ∈ I, and limi∈I kα(g)ξi − ξi k2 = 0, for all g ∈ G.
Moreover, if the group von Neumann algebra of G has property Γ, then G is
inner amenable.
Effros conjectured that inner amenability of a discrete ICC group G was,
in fact, equivalent to property Γ of LG. In 2012, Vaes [24] provided a counterexample.
Recall that an action of a discrete group G on a set X is a homomorphism
from G to the set of permutations of X, and that such an action is said to be
amenable if there exists a finitely additive probability measure on X which
is invariant under the action. It straightforward to check that a discrete
group G is inner amenable if and only if the conjugation action of the group
on G \ {e} is amenable.
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
6
A key result for our proof of non-inner amenability of T and V is the
following result due to Rosenblatt [21].
Proposition 3.2 (Rosenblatt). Let G be a non-amenable discrete group
acting on a set X. If the stabilizer of each point is amenable, then the action
of G is itself non-amenable.
From this we directly get the following sufficient condition for non-inner
amenability.
Corollary 3.3. Let G be discrete a group. If G has a non-amenable subgroup H such that {g ∈ H : ghg −1 = h} is amenable, for each h ∈ G \ {e},
then G is non-inner amenable.
Notice that, in the corollary above, any subgroup of G containing H will
automatically fail to be inner amenable, as well.
Theorem 3.4. The Thompson groups T and V are not inner amenable.
Proof. Recall from Section 2 that Λ denotes the subgroup of T generated
by C and D 2 , and that conjugation by the map φ from Theorem 2.2 restricts
to an isomorphisms between Λ and PSL(2, Z). Clearly Λ is not amenable, so,
by Corollary 3.3, it suffices to prove that {g ∈ Λ : gf = f g} is amenable, for
all f ∈ V \ {e}. This will imply that both T and V are non-inner amenable.
We will consider separately the cases where f ∈ Λ \ {e}, f ∈ T \ Λ and
f ∈ V \ T , respectively. Fix f ∈ V \ {e} and let us denote the subgroup
{g ∈ Λ : gf = f g} by H.
First, suppose that f ∈ Λ \ {e}. Then H is the centralizer of f in Λ, which
is cyclic by Theorems 2.3.3 and 2.3.5 in [17]. In particular, it is amenable.
Now, suppose that f ∈ T \ Λ. It is easy to see that, with f˜ denoting
−1
φ f φ, we have φ−1 Hφ = {h ∈ PSL(2, Z) : hf˜) = f˜h}. Let us show
that this group is amenable, since then H will be amenable as well. Since
f ∈
/ Λ, we know that f˜ ∈
/ PSL(2, Z), and, in particular, f˜ has at least
two breakpoints. Let y1 , . . . , yn denote these breakpoints, and note that
they all are in Q ∪ {∞}. Suppose that h ∈ PSL(2, Z) with hf˜h−1 = f˜.
Clearly, the breakpoints of hf˜h−1 are h(y1 ), . . . , h(yn ), so since hf˜h−1 = f˜,
these must also be breakpoints of φ−1 f φ. In other words, h permutes the
breakpoints of f˜. Now, because f has at least two breakpoints, it follows
from Proposition 2.5 that h is uniquely determined by the corresponding
permutation of y1 , . . . , yn . Since there are only finitely many permutations
of n elements, we deduce that φ−1 Hψ is finite. In particular, H is also finite,
and therefore amenable.
Last, suppose that f ∈ V \ T . Then f is discontinuous, and since f is
a bijection of [0, 1), it is easy to see that f must have at least two points
of discontinuity. Since these points of discontinuity are all dyadic rationals
and the elements of Λ are uniquely determined by their values of two dyadic
rational numbers, by Remark 2.6, we can use the same argument as the one
for T \ Λ to conclude that H is finite, since every element in Λ must permute
points of discontinuity of f . This proves that {g ∈ Λ : gf = f g} is amenable,
for all f ∈ V \ {e}. By Corollary 3.3, the conclusion follows.
The proof of the above theorem shows that, in fact, every subgroup of V
containing Λ is non-inner amenable.
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
7
By the results of Effros (see Theorem 3.1), it follows that neither LT
nor LV has property Γ, and hence they are not McDuff factors.
4. Simplicity of Cr∗ (T ) implies non-amenability of F
In this section we prove that the Thompson group F is non-amenable if
the reduced group C ∗ -algebra of T is simple. First, we recall the notion of
weak containment of group representations.
By a representation of a group, we always mean a unitary representation,
that is, a homomorphism from the given group to the unitary group of some
Hilbert space. We denote the full group C ∗ -algebra of a discrete group G by
C ∗ (G). If π : G → B(H) is a representation of G on a Hilbert space H, then
it extends uniquely to a representation of C ∗ (G). We use the same symbol
to denote this representation. If π : G → B(H) and ρ : G → B(K) are two
representations of the group G, then we say that π is weakly contained in ρ
if, for every ξ ∈ H and ε > 0, there exist η1 , . . . , ηn ∈ K such that
hπ(g)ξ | ξi −
n
X
hρ(g)ηi | ηi i < ε.
i=1
This is equivalent to the fact that there exists a ∗-homomorphism σ from the
C ∗ -algebra generated by ρ(G) to the C ∗ -algebra generated by π(G) such that
π = σ ◦ ρ. We need the following well-known result about weak containment.
Proposition 4.1. Let G be a discrete group acting on a set X. If all stabilizers of the action are amenable, then the representation of G on ℓ2 (X) is
weakly contained in the left regular representation of G.
For more about weak containment and connections to amenability see [1].
The following result is a well-known characterization of amenable actions.
For a proof in the case where the action is left translation of the group on
itself, the reader may consult [4, Theorem 2.6.8].
Proposition 4.2. Suppose that G is a discrete group acting on a set X, and
let π denote the induced representation on ℓ2 (X). Then the action of G on X
is non-amenable if and only if there exist g1 , . . . , gn in G so that
n
1X
π(gk ) < 1.
n
k=1
In P
fact, if the action is non-amenable and G is finitely generated, then
k n1 nk=1 π(gk )k < 1, for any set of elements g1 , g2 , . . . , gn generating G.
In the following we will make use of the Cuntz algebra O2 . Recall that
O2 is the universal C ∗ -algebra generated by two isometries s1 and s2 with
orthogonal range projections summing up to the identity. In other words,
O2 is the universal C ∗ -algebra generated by elements s1 and s2 satisfying the
relations s∗1 s1 = 1, s∗2 s2 = 1 and s1 s∗1 + s2 s∗2 = 1. There is a canonical way
to realize the Thompson groups as subgroups of the unitary group of O2 .
This was discovered by Nekrashevych in [20], as kindly pointed out to us by
Wojciech Szymanski.
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
8
Let us describe the concrete model we will use for the Cuntz algebra O2 .
We think of it as a specific set of bounded operators on ℓ2 (X) where X denotes
the set Z[ 21 ] ∩ [0, 1). Define the operators s1 and s2 on ℓ2 (X) by
s1 δx = δx/2
and
s2 δx = δ(1+x)/2 ,
for all x ∈ X. It is straightforward to check that s1 and s2 are isometries
satisfying s1 s∗1 + s2 s∗2 = 1, so they generate a copy of the Cuntz algebra O2 .
The groups F , T and V act by definition on the set X, and we denote the
induced representations on B(ℓ2 (X)) by π. That is, π(g)δx = δg(x) , for all
x ∈ X and g ∈ V . We use π to denote this representation when restricted to
F and T , as well. As it so happens, the image of π is contained in O2 . In
fact, one can check the following explicit identities:
π(D) = s2 s2 s∗1 s∗1 + s1 s1 s∗2 s∗1 + s1 s2 s∗1 s∗2 + s2 s1 s∗2 s∗2 ,
π(C) = s2 s2 s∗1 + s1 s∗1 s∗2 + s2 s1 s∗2 s∗2 ,
π(D 2 ) = s2 s∗1 + s1 s∗2 ,
π(π0 ) = s2 s1 s∗1 + s1 s∗1 s∗2 + s2 s2 s∗2 s∗2 .
We denote the C ∗ -algebras generated by π(F ), π(T ) and π(V ) inside O2
by Cπ∗ (F ), Cπ∗ (T ) and Cπ∗ (V ), respectively.
Proposition 4.3. With the notation above, we have
Cπ∗ (F ) ( Cπ∗ (T ) ( Cπ∗ (V ) = O2 .
Proof. By construction, Cπ∗ (F ) ⊆ Cπ∗ (T ) ⊆ Cπ∗ (V ) ⊆ O2 . We will prove that
the first two inclusions are proper and that the last one is an equality.
It is easy to see that Cπ∗ (F ) 6= Cπ∗ (T ), since Cδ0 is a Cπ∗ (F )-invariant
subspace which is not Cπ∗ (T )-invariant. Hence Cπ∗ (F ) 6= Cπ∗ (T ).
For the rest, let us first prove that Cπ∗ (V ) = O2 , and afterwards that
∗
Cπ (T ) 6= O2 . Our strategy for showing that Cπ∗ (V ) = O2 is to prove that
Cπ∗ (V ) contains s1 and s2 . In fact, since s2 = π(D 2 )s1 , it suffices to show
that it contains s1 . To ease notation, let us denote ℓ2 (Z[ 21 ] ∩ [0, 12 )) and
ℓ2 (Z[ 12 ] ∩ [ 12 , 1)) by H1 and H2 , respectively. Consider the subgroup V0 of V
consisting of elements g ∈ V satisfying g(x) = x, for all x ∈ [0, 21 ). Note
that, for g ∈ V0 , both H1 and H2 are invariant subspaces for π(g), so that
we can write π(g) = π(g)|H1 ⊕ π(g)|H2 . It is easy to see using Theorem 3.2
that the action of V0 on [ 21 , 1) is non-amenable. Thus, by Proposition 4.2,
P
there exist elements g1 , . . . , gn in V0 , such that k n1 nk=1 π(gk )|H2 k < 1.
P
P
Since n1 nk=1 π(gk )|H1 = 1H1 , we deduce that ( n1 nk=1 π(gk ))m converges
in norm, as m → ∞, to the projection of ℓ2 (X) onto H1 , that is, to s1 s∗1 .
Hence s1 s∗1 , and therefore also s2 s∗2 = 1 − s1 s∗1 , belong to Cπ∗ (V ). It is
straightforward to check that
s1 = π(A)s1 s∗1 + π(D 2 A−1 )s2 s∗2 .
This implies that s1 ∈ Cπ∗ (V ), and we conclude that Cπ∗ (V ) = O2 .
Last, let us prove that Cπ∗ (T ) 6= O2 . We do this by exhibiting two
different states φ0 and φ1 on O2 which agree on Cπ∗ (T ). Let φ0 denote
the vector state given by δ0 . It is well-known that elements of the form
si1 si2 · · · sin s∗i1 s∗i2 · · · s∗im span a dense subspace of O2 . Moreover, it is easy
to check that φ0 (sn1 (s∗1 )m ) = 1, for all n, m ≥ 0, while φ0 is zero on the rest
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
9
of these elements. Let φ1 denote the composition of φ0 with the automorphism of O2 that interchanges s1 and s2 . Then, for all n, m ≥ 0, φ1 satisfies
φ1 (si1 · · · sin s∗i1 · · · s∗im ) = 0, unless i1 = . . . = in = 2 and j1 = . . . = jm = 2,
in which case φ1 (sn2 (s∗2 )m ) = 1. It follows easily from Nekrashevych’s description of the representation π (see [20, Section 9]) that for all g ∈ V ,
either φk (π(g)) = 0 or φk (π(g)) = 1, where k = 0, 1. Similarly, it is easily seen that, for g ∈ V , φ0 (π(g)) = 1 if and only if limx→0 g(x) = 0, and
that φ1 (π(g)) = 1 if and only if limx→1 g(x) = 1. In particular, we see that
φk (π(g)) = 1F (g), for all g ∈ T and k = 0, 1. Thus φ0 and φ1 agree on π(T ),
so clearly they also agree on Cπ∗ (T ). It follows that Cπ∗ (T ) 6= O2 , as the two
states clearly do not agree on O2 .
It turns out that the representation π is closely connected to amenability
of the Thompson group F , as we show in the following.
Proposition 4.4. The Thompson group F is amenable if and only if π is
contained in the left regular representation of T . This is further equivalent
to the fact that π is contained in the left regular representation of F .
Proof. Clearly, π|F is weakly contained in the left regular representation of F
if π is weakly contained in the left regular representation of T . Suppose now
that F is amenable. Since T acts transitively on X, all the stabilizers of
the action are isomorphic. It follows from Proposition 4.1 that π is weakly
contained in the left regular representation of T , as F is the stabilizer of 0.
Suppose that π is weakly contained in the left regular representation of F .
Let p denote the projection onto Cδ0 , which is a π(F )-invariant subspace
of ℓ2 (X). Then pπp is the trivial representation of F . This is contained
in π, so by transitivity of weak containment, we deduce that the trivial
representation of F is weakly contained in the left regular representation
of F . By [1, Theorem G.3.2], this is equivalent to amenability of F .
Theorem 4.5. If Cr∗ (T ) is simple, then F is non-amenable.
Proof. Suppose that F is amenable, and let us then prove that Cr∗ (T ) is
not simple. By Proposition 4.4, π is weakly contained in the left regular
representation of T , so there exists a ∗-homomorphism σ from Cr∗ (T ) to
Cπ∗ (T ) such that π = σ ◦ λ. Our goal is to show that the kernel of σ is a nontrivial ideal in Cr∗ (T ). Since σ is obviously not the zero map, we only need to
show that its kernel contains a non-trivial element. The ∗-homomorphism λ
is injective on the complex group algebra CT , so it suffices to find an element
x 6= 0 in CT such that π(x) = 0.
Consider the elements a and b of T (and their product) defined by
a = CDC
b = D 2 CDCD 2
ab (= ba)
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
10
Obviously, a and b commute, and it is easy to see that the values (a(x), b(x))
and (ab(x), x) agree, up to a permutation, for all x ∈ [0, 1). From this
we deduce that π(a + b)δx = π(ab + e)δx , for all x ∈ X. In particular,
π(a + b − ab − e) = 0, so that the kernel of σ is non-trivial. Thus Cr∗ (T ) is
not simple.
As mentioned in the introduction, Le Boudec and Matte Bon [19] recently
proved the converse implication to the theorem above. They also proved
therein that Cr∗ (V ) is simple, and that Cr∗ (F ) is simple if Cr∗ (T ) is simple.
The converse of this latter implication was already proved by Breuillard,
Kalantar, Kennedy and Ozawa [3].
It is still unknown whether the reduced group C ∗ -algebras of T is simple,
but it is, however, known that it has a unique tracial state. This was proved
by Dudko and Medynets [8].
Remark 4.6. We believe that is is a well-known fact that F is non-amenable if
and only if its reduced group C ∗ -algebra has a unique tracial state, although
we have not been able to find an explicit reference for this. It follows, for
example, from the fact that Cr∗ (F ) has at most two extremal tracial states,
as showed by Dudko and Medynets [8]. A different proof of this is based on
the new characterization of groups whose reduced group C ∗ -algebra has a
unique tracial state given by Breuillard, Kalantar, Kennedy and Ozawa. In
the following we briefly indicate the argument.
It was shown in the aforementioned paper that Cr∗ (F ) has a unique tracial
state if and only if the amenable radical of F is trivial. However, it is a wellknown fact (see [5, Theorem 4.3]) that every non-trivial normal subgroup of
F contains a copy of F . Hence the amenable radical is trivial if and only if
F is non-amenable.
We end this section by giving a few new equivalent characterizations of
amenability of F . We will make use of the following well-known result, for
which a proof can be found in [11, Lemma 4.1] (in the case where the action
is left translation of the group on itself).
Lemma 4.7. Let G be a discrete group acting on a set X, and let σ be the
corresponding representation on ℓ2 (X). Then kx+yk ≥ kxk, for x, y ∈ R+ G.
By scrutinizing the proof of Theorem 4.5 one sees that, with the notation
therein, the Thompson group F is non-amenable if the closed two-sided ideal
generated by the element 1 + λ(ab) − λ(a) − λ(b) inside Cr∗ (T ) is the whole
of Cr∗ (T ). This element is not unique with this property. Indeed, the same
holds true for λ(x), where x is an element of the complex group algebra CT
such that π(x) = 0. Bleak and Juschenko [2] established a partial converse
to Theorem 4.5 by proving that such x exists if the Thompson group F is
non-amenable. The equivalence of the first two statements in the following
proposition is a strengthening of their result, in the sense that we exhibit a
concrete such element x.
Proposition 4.8. With a and b as above, the following are equivalent:
(1) The Thompson group F is non-amenable.
(2) The closed two-sided ideal generated by 1 + λ(ab) − λ(a) − λ(b) in
Cr∗ (T ) is all of Cr∗ (T ).
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
11
(3) The closed convex hull of {λ(hah−1 ) + λ(hbh−1 ) : h ∈ T } contains 0.
(4) The closed two-sided ideal generated by 1 + λ(ab) − λ(a) − λ(b) in
Cr∗ (F ) is all of Cr∗ (F ).
(5) The closed convex hull of {λ(hah−1 ) + λ(hbh−1 ) : h ∈ F } contains 0.
Proof. We prove that (1) implies (4) and (5), that (4) implies (2), that (5)
implies (3), that (2) implies (1), and that (3) implies (1).
Some of these implications are straightforward. That (4) implies (2) follows from the fact that the inclusion Cr∗ (F ) ⊆ Cr∗ (T ) is unital. That is, if
(4) holds, then the closed two-sided ideal in Cr∗ (T ) generated by 1 + λ(ab) −
λ(a)− λ(b) contains Cr∗ (F ), and, in particular, the unit. That (5) implies (3)
follows directly from the fact that the inclusion Cr∗ (F ) ⊆ Cr∗ (T ) is isometric.
Let us prove that (2) implies (1), and that (3) implies (1), or, equivalently, that the negation of (1) implies the negation of (2), as well as the
negation of (3). Assume that F is amenable. Then by Proposition 4.4, π
is weakly contained in the left regular representation of T , that is, there
exists a ∗-homomorphism σ so that π = σ ◦ λ. As explained in the proof
of Theorem 4.5, 1 + λ(ab) − λ(a) − λ(b) is in the kernel of σ, which is a
non-trivial ideal. Hence the negation of (2) holds. Since σ is a contraction, to prove the negation of (3) it suffices to show that kxk ≥ 1, for every
x ∈ conv{π(hah−1 ) + π(hbh−1 ) : h ∈ T }. As explained in the proof of
Theorem 4.5, that π(a) + π(b) = 1 + π(ab). Hence
conv{π(hah−1 ) + π(hbh−1 ) : h ∈ T } = 1 + conv{π(habh−1 ) : h ∈ T }.
If x is a finite convex combination of elements of the form habh−1 with h ∈ T ,
then x ∈ R+ T . By Lemma 4.7, we conclude that k1 + xk ≥ k1k = 1.
Continuity then ensures that k1 + xk ≥ 1, for every x ∈ conv{π(hah−1 ) +
π(hbh−1 ) : h ∈ T }. Hence the negation of (3) holds.
We are left to prove that (1) implies (4) and (5), so assume that F is
non-amenable. To prove (4), it suffices to show that the closed two-sided
ideal in Cr∗ (F ) generated by 1 + λ(ab) − λ(a) − λ(b) contains 1. As it clearly
contains the set
conv λ(h) 1 + λ(ab) − λ(a) − λ(b) λ(h)∗ : h ∈ F
= 1 + conv λ(habh−1 ) − λ(hah−1 ) − λ(hbh−1 ) : h ∈ F ,
it suffices to show that the closed convex hull on the right hand side contains 0. Let ε > 0 be given. As mentioned in Remark 4.6, F has the
unique trace property. Let F1 and F2 denote the subgroups of F consisting of elements f ∈ F so that f (x) = x, for all x ∈ [0, 12 ), and f (x) = x,
for all x ∈ [ 12 , 1), respectively. Note that b ∈ F1 and a ∈ F2 . Since F1
and F2 are both isomorphic to F , they both have the unique trace property.
It follows from [11, Corollary
P 4.4] that
P there exist positive real numbers
s1 , . . . , sn , t1 , . . . , tm with nk=1 sk = m
k=1 tk = 1, and g1 , . . . , gn ∈ F1 and
h1 , . . . , hm ∈ F2 so that
n
m
X
X
and
sk λ(gk bgk−1 ) < ε
tk λ(hk ah−1
k ) < ε.
k=1
k=1
Pm
−1
Let us, for simplicity of notation, denote the elements
k=1 tk λ(hk ahk )
Pn
−1
and
k=1 sk λ(gk bgk ) by ã and b̃, respectively. As the elements of the
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
12
subgroups F1 and F2 commute, it is straightforward to check that
m
n X
X
si tj λ(gi hj ) λ(a) + λ(b) λ(gi hj )∗ = ã + b̃.
i=1 j=1
Pn
Pm
Since i=1 j=1 si tj = 1, the left hand side belongs to the convex hull of
{λ(hah−1 ) + λ(hbh−1 ) : h ∈ F }, and since kã + b̃k < 2ε, we conclude that
the convex hull of {λ(hah−1 ) + λ(hbh−1 ) : h ∈ F } contains elements of
arbitrarily small norm. Thus (5) holds.
Using similar calculations, it is straightforward to check that
m
n X
X
si tj λ(gi hj ) λ(ab) − λ(a) − λ(b) λ(gi hj )∗ = ãb̃ − ã − b̃.
i=1 j=1
Again, the left hand side belongs to the convex hull of
{λ(habh−1 ) − λ(hah−1 ) − λ(hbh−1 ) : h ∈ F }.
Since kãb̃ − ã − b̃k < ε2 − 2ε, we conclude that the closure of this convex hull
contains 0, and as mentioned above, this means that the ideal generated by
1 + λ(ab) − λ(a) − λ(b) in Cr∗ (F ) is the whole of Cr∗ (F ). This proves that
(1) implies (4) and (5), which concludes the proof.
Acknowledgement. This work is part of the second named author’s Ph.D
thesis at the University of Copenhagen under the supervision of Magdalena
Musat and Uffe Haagerup. Kristian Olesen wishes to express his sincere
gratitude for their invaluable guidance and support throughout the Ph.D
period. This manuscript was completed after the sudden passing away of
Uffe Haagerup.
References
[1] Bachir Bekka, Pierre de la Harpe, and Alain Valette. Kazhdan’s property (T), volume 11 of New Mathematical Monographs. Cambridge University Press, Cambridge,
2008.
[2] Collin Bleak and Kate Juschenko. Ideal structure of the C ∗ -algebra of Thompson
group T . preprint, 2014. arXiv:1409.8099.
[3] Emmanuel Breuillard, Mehrdad Kalantar, Matthew Kennedy, and Narutaka Ozawa.
C ∗ -simplicity and the unique trace property for discrete groups. preprint, 2014.
arXiv:1410.2518.
[4] Nathanial P. Brown and Narutaka Ozawa. C ∗ -algebras and finite-dimensional approximations, volume 88 of Graduate Studies in Mathematics. American Mathematical
Society, Providence, RI, 2008.
[5] J. W. Cannon, W. J. Floyd, and W. R. Parry. Introductory notes on Richard Thompson’s groups. Enseign. Math. (2), 42(3-4):215–256, 1996.
[6] Tullio G. Ceccherini-Silberstein and Fabio Scarabotti. Inner amenability of some
groups of piecewise-linear homeomorphisms of the real line. J. Math. Sci. (New York),
106(4):3164–3167, 2001. Pontryagin Conference, 8, Algebra (Moscow, 1998).
[7] Pierre de la Harpe. On simplicity of reduced C ∗ -algebras of groups. Bull. Lond. Math.
Soc., 39(1):1–26, 2007.
[8] Artem Dudko and Konstantin Medynets. Finite factor representations of HigmanThompson groups. preprint, 2012. arXiv:1212.1230.
[9] Edward George Effros. Property Γ and inner amenability. Proc. Amer. Math. Soc.,
47:483–486, 1975.
[10] Ariadna Fossas. PSL(2, Z) as a non-distorted subgroup of Thompson’s group T . Indiana Univ. Math. J., 60(6):1905–1925, 2011.
NON-INNER AMENABILITY OF THE THOMPSON GROUPS T AND V
13
[11] Uffe Haagerup. A new look at C ∗ -simplicity and the unique trace property of a group.
preprint, 2015. arXiv:1509.05880.
[12] Michel Imbert. Sur l’isomorphisme du groupe de Richard Thompson avec le groupe de
Ptolémée. In Geometric Galois actions, 2, volume 243 of London Math. Soc. Lecture
Note Ser., pages 313–324. Cambridge Univ. Press, Cambridge, 1997.
[13] Nikolay A. Ivanov and Tron Omland. C ∗ -simplicity of free products with amalgamation and radical classes of groups. preprint, 2016. arXiv:1605.06395.
[14] Paul Jolissaint. Moyennabilité intérieure du groupe F de Thompson. C. R. Acad. Sci.
Paris Sér. I Math., 325(1):61–64, 1997.
[15] Paul Jolissaint. Central sequences in the factor associated with Thompson’s group F .
Ann. Inst. Fourier (Grenoble), 48(4):1093–1106, 1998.
[16] Mehrdad Kalantar and Matthew Kennedy. Boundaries of reduced C ∗ -algebras of
discrete groups. preprint, 2014. arXiv:1405.4359.
[17] Svetlana Katok. Fuchsian groups. Chicago Lectures in Mathematics. University of
Chicago Press, Chicago, IL, 1992.
[18] Adrien Le Boudec. Discrete groups that are not C ∗ -simple. preprint, 2015.
arXiv:1507.03452.
[19] Adrien Le Boudec and Nicolás Matte Bon. Subgroup dynamics and C ∗ -simplicity of
groups of homeomorphisms. preprint, 2016. arXiv:1605.01651.
[20] Volodymyr V. Nekrashevych. Cuntz-Pimsner algebras of group actions. J. Operator
Theory, 52(2):223–249, 2004.
[21] Joseph Rosenblatt. Uniqueness of invariant means for measure-preserving transformations. Trans. Amer. Math. Soc., 265(2):623–636, 1981.
[22] R. Salem. On some singular monotonic functions which are strictly increasing. Trans.
Amer. Math. Soc., 53:427–439, 1943.
[23] Jean-Pierre Serre. Trees. Springer-Verlag, Berlin-New York, 1980. Translated from
the French by John Stillwell.
[24] Stefaan Vaes. An inner amenable group whose von Neumann algebra does not have
property Gamma. Acta Math., 208(2):389–394, 2012.
Department of Mathematics and Computer Science, University of Southern Denmark, Campusvej 55, DK-5230 Odense M
Department of Mathematical Sciences, University of Copenhagen, Universitetsparken 5, DK-2100 Copenhagen Ø
E-mail address: [email protected]
| 4math.GR
|
arXiv:1706.01260v2 [cs.DS] 17 Oct 2017
The Classical Complexity of Boson Sampling
Peter Clifford
Raphaël Clifford
Department of Statistics
University of Oxford
United Kingdom
Department of Computer Science
University of Bristol
United Kingdom
Abstract
We study the classical complexity of the exact Boson Sampling problem where the objective is
to produce provably correct random samples from a particular quantum mechanical distribution. The
computational framework was proposed in STOC ’11 by Aaronson and Arkhipov in 2011 as an attainable demonstration of ‘quantum supremacy’, that is a practical quantum computing experiment able
to produce output at a speed beyond the reach of classical (that is non-quantum) computer hardware.
Since its introduction Boson Sampling has been the subject of intense international research in the
world of quantum computing. On the face of it, the problem is challenging for classical computation. Aaronson and Arkhipov show that exact Boson Sampling is not efficiently solvable by a classical
computer unless P#P = BPPNP and the polynomial hierarchy collapses to the third level.
The fastest known exact classical algorithm for the standard Boson Sampling problem requires
O((m+n−1
) n2n ) time to produce samples for a system with input size n and m output modes, makn
ing it infeasible for anything but the smallest values of n and m. We give an algorithm that is much
faster, running in O(n2n + poly(m, n)) time and O(m) additional space. The algorithm is simple to
implement and has low constant factor overheads. As a consequence our classical algorithm is able to
solve the exact Boson Sampling problem for system sizes far beyond current photonic quantum computing experimentation, thereby significantly reducing the likelihood of achieving near-term quantum
supremacy in the context of Boson Sampling.
1 Introduction
The promise of significantly faster quantum algorithms for problems of practical interest is one of the most
exciting prospects in computer science. Most famously, a quantum computer would allow us to factorise
integers in polynomial time (Shor, 1997) and perform unstructured search on an input of n elements in
√
O( n) time (Grover, 1996). In the short term however, despite impressive progress in recent years, there
are still considerable challenges to overcome before we can build a fully universal quantum computer. Until
this time, the question of how to design quantum experiments which fall short of fully universal computation
but which still show significant quantum speed-up over known classical (that is non-quantum) algorithms
remains of great interest. In a breakthrough contribution in STOC ’11 Aaronson and Arkhipov (2011) proposed an experimental set-up in linear optics known as Boson Sampling as a potentially practical way to do
just this.
Since its introduction, Boson Sampling has attracted a great deal of attention with numerous experimental
efforts around the world attempting implementations for various problem sizes (see e.g. Bentivegna et al.,
2015; Spring et al., 2013; Broome et al., 2013; Tillmann et al., 2013; Crespi et al., 2013; Spagnolo et al.,
2014; Latmiral et al., 2016). The ultimate goal is to exhibit a physical quantum experiment of such a scale
that it would be hard if not impossible to simulate the output classically and thereby to establish so-called
‘quantum supremacy’. In terms of the physical Boson Sampling experiment, n corresponds to the number
of photons and m the number of output modes and increasing either of these is difficult in practice. Progress
has therefore been slow (see Lund et al., 2017, and the references therein) with the current experimental
record being n = 5, m = 9 (Wang et al., 2016).
Translated into conventional computing terms, the task is to generate independent random samples from a
specific probability distribution on multisets of size n with elements in the range 1 through m. For Boson
Sampling, the probability of each multiset is related to the permanent of an n × n matrix built from possibly
repeated rows of a larger m × n matrix where the multiset determines the rows used in the construction. The
Boson Sampling problem is to produce such samples either with a quantum photonic device or with classical
computing hardware.
On the face of it, Boson Sampling is a challenging problem for classical computation. Aaronson and Arkhipov
(2011) showed that exact Boson Sampling is not solvable in polynomial time by a classical computer unless
P#P = BPPNP and the polynomial hierarchy collapses to the third level.
Without a clear understanding of the classical complexity of the problem it is difficult to determine the
minimum experimental size needed to establish quantum supremacy. In the absence of a fast classical
algorithm for Boson Sampling, speculation about the minimum size of n (with m = n2 ) has varied a great
deal in the literature from e.g. 7 (Latmiral et al., 2016), between 20 and 30 (Aaronson and Arkhipov, 2014)
and all the way up to 50 (Lund et al., 2017). Broadly speaking the lowest estimate corresponds to the limits
of the previous fastest classical Boson Sampling algorithm and the highest with an assumption that the
polynomial hierarchy will not collapse.
Brute force evaluation of the probabilities of each multiset as a preliminary to random sampling, requires the
n
calculation of (m+n−1
n ) permanents of n × n matrices, each one of which takes O(n2 ) time with the fastest
known algorithms. If m > n2 , as considered by Aaronson and Arkhipov (2011), the total running time is
Θ(2n en (m/n)n n1/2 ). The space usage of such a naive approach can be reduced to the storage needed for
the required sample size (see e.g. Efraimidis, 2015), however the running time for anything but the smallest
values of n and m remains prohibitive. Previous to the work we present here, no faster provably correct
classical Boson Sampling algorithm was known.
Recently Neville et al. (2017) gave strong numerical evidence suggesting that a specially designed Markov
chain Monte Carlo (MCMC) implementation can sample approximately from the collision-free version of
Boson Sampling. The time cost per sample for their MCMC method for problems with size n 6 30 corresponds to that of computing around two hundred n × n permanents, making it computationally feasible for
the problem sizes they consider. This was the first practical evidence that classical Boson Sampling might
be possible faster than the naive brute force approach and is a significant motivation for the work we present
here.
We show that Boson Sampling on a classical computer can be performed both exactly and dramatically faster
than the naive algorithm, significantly raising the bar for the minimum size of quantum experiment needed
to establish quantum supremacy. Our sampling algorithm also provides the probability associated with any
given sample at no extra computational cost.
Theorem 1. The time complexity of Boson Sampling is:
O(n2n + poly(m, n)),
where poly(m, n) = O(mn2 ). The additional space complexity on top of that needed to store the input is
O(m).
2
A particularly attractive property of the running time we give is that the exponentially growing term depends
only on n, the size of the multisets we wish to sample, and not m which is potentially much larger. In
the most natural parameter regime when m is polynomial in n our algorithm is therefore approximately
(m+n−1
n ) times faster than the fastest previous method. Our algorithm is both straightforward to implement
and has small constant factor overheads, with the total computational cost to take one sample approximately
equivalent to that of calculating two n × n permanents. This gives a remarkable resolution to the previously
open question of the relationship between the classical complexity of computing the permanent of an n × n
matrix and Boson Sampling. For an implementation of the algorithm, see Clifford and Clifford (2017).
Based on the time complexity we demonstrate in Theorem 1 and bearing in mind recent feasibility studies
for the calculation of large permanents (Wu et al., 2016) we find support for n = 50 as the threshold for
quantum supremacy for the exact Boson Sampling problem.
Our sampling algorithm also applies directly to the closely related problem of scattershot Boson Sampling
(Bentivegna et al., 2015). From a mathematical perspective this produces no additional complications beyond the specification of a different m × n matrix. Once the matrix is specified we can apply our new
sampling algorithm directly.
Our techniques draw heavily on the highly developed methods of hierarchical sampling commonly used in
Bayesian computation (see e.g. Liu, 2008; Green et al., 2015); specially the use of auxiliary variables and
conditional simulation. We believe this cross-fertilisation of ideas is novel and opens up the development
of classical algorithms for sampling problems in quantum computation. It raises the intriguing question
of whether similar techniques can be applied to get significant improvements in time and space for other
proposed quantum supremacy projects such as, for example, the problem of classically simulating quantum
circuits (Boixo et al., 2016).
The algorithmic speed-up we achieve relies on a number of key innovations. As a first step we expand the
sample space of multisets into a product set [m]n , i.e. the space of all arrays r = (r1 , . . . , rn ), with each
element in the range 1 to m. This simple transformation allows us to expose combinatorial properties of the
sampling problem which were otherwise opaque. We then develop explicit expressions for the probabilities
of subsequences of arrays in this set. Application of the chain rule for conditional probabilities then leads to
our first speed-up with an O(mn3n ) time algorithm for Boson Sampling. For our main result in Theorem
1, we introduce two further innovations. First we exploit the Laplace expansion of a permanent to give an
efficient amortised algorithm for the computation of the permanent of a number of closely related matrices.
We then expand the sample space further with an auxiliary array representing column indices and show how
a conditioning argument allows Boson Sampling to be viewed as sampling from a succession of coupled
conditional distributions. The final result is our improved time complexity of O(n2n + mn2 ).
A consequence of our analysis is that we are also able to provide an efficiently computable bound on the
probability that a sampled multiset drawn from the Boson Sampling distribution will contain duplicated
elements. This enables us to bound the performance of rejection sampling variants of our sampling algorithm
when applied to the so-called ‘collision-free’ setting described by Aaronson and Arkhipov.
One further application of our new sampling algorithm lies in tackling the question of whether the output of
physical Boson Sampling devices correspond to theoretical predictions. There has been considerable interest
in this area (Aaronson and Arkhipov, 2014; Tichy et al., 2014; Wang and Duan, 2016; Walschaers et al.,
2016) however previously developed methods have had to work under the assumption that we cannot sample
from the Boson Sampling distribution for even moderate sizes of n and m. Armed with our new sampling
algorithm, we can now sample directly from the Boson Sampling distribution for a much larger range of
parameters, allowing a far greater range of statistical tests to be applied.
3
Related work and computation of the permanent
Terhal and DiVincenzo (2004) were the first to recognise that studying the complexity of sampling from lowdepth quantum circuits could be useful for demonstrating a separation between quantum and classical computation. Since that time the goal of finding complexity separations for quantum sampling problems has received a great deal of interest. We refer the interested reader to Lund et al. (2017) and Harrow and Montanaro
(2017) for recent surveys on the topic.
There is an extensive literature on the problems of sampling from unwieldy distributions with applications
in all the main scientific disciplines. Much of this work has relied on MCMC methods where the target
distribution can be identified as the equilibrium distribution of a Markov chain (see e.g. Hastings, 1970).
For Boson Sampling the most closely related work to ours is that of Neville et al. (2017) mentioned earlier.
In their paper the authors also set out the hurdles that need to be overcome for a physical Boson Sampling
experiment to reach higher values of n. Their MCMC approach does not however permit them to estimate
how well their sampling method approximates the true Boson Sampling distribution, other than through
numerical experimentation. Our new exact sampling algorithm is therefore the first provably correct algorithmic speed-up for Boson Sampling. In fact our algorithm also has smaller constant factor overheads,
costing approximately two permanent calculations of n×n matrices for all values of n (assuming m remains
polynomial in n).
There are MCMC algorithms which permit exact sampling, for example via ‘coupling from the past’, where
guarantees of performance can be provided (Propp and Wilson, 1996). However, the range of problems for
which this approach has proven applicable is necessarily more limited and there has yet to be a successful
demonstration for the Boson Sampling problem.
Permanents: Computation of the permanent of a matrix was shown to be #P-hard by Valiant (1979)
and hence is unlikely to have a polynomial time solution. Previously Ryser (1963) had proposed an
O(n2 2n ) time algorithm to compute the permanent of an n × n matrix. This was sped up to O(n2n )
time by Nijenhuis and Wilf (1978) and many years later a related formula with the same time complexity
was given by Glynn (2010). It is Glynn’s computational formula for the permanent which we will use as the
basis of our sampling algorithms.
Let M = (mi,j ) be an n × n matrix with mi,j ∈ C, and let Per M denote its permanent, then
! n n
n
YX
1 X Y
δk
δi mi,j
Per M = n−1
2
δ
k=1
(1)
j=1 i=1
where δ ∈ {−1, 1}n with δ1 = 1.
A naive implementation of this formula would require O(n2 2n ) time however by considering the δ arrays
in Gray code order this is reduced to O(n2n ) time. Note that further reduction in the time complexity may
be possible when M has repeated rows or columns; see Tichy (2011, Appendix B) and Shchesnovich (2013,
Appendix D). In the special √
case of computation of the permanent over poly(n) bit integers, Björklund
n−Ω( n/ log n)
(2016) has a faster O(n2
) time algorithm for computing the permanent.
2 Problem specification
In mathematical terms the Boson Sampling problem is as follows.
4
Let m and n be positive integers and consider all multisets of size n with elements in [m], where [m] =
{1, . . . , m}. Each multiset can be represented by an array z = (z1 , . . . , zn ) consisting of elements of
the multiset in non-decreasing order: z1 6 z2 6 · · · 6 zn . We denote the set of distinct values of z
by Φm,nQ
. The cardinality of Φm,n is known to be (m+n−1
n ) – see Feller (1968), for example. We define
m
µ(z) = j=1 sj ! where sj is the multiplicity of the value j in z.
Now let A = (aij ) be a complex valued m × n matrix consisting of the first n columns of a given m × m
Haar random unitary matrix. For each z, build an n × n matrix Az where the k-th row of Az is row zk in A
for k = 1, . . . , n and define a probability mass function (pmf) on Φm,n as
n
XY
1
defn 1
|Per Az |2 =
azk σk
q(z) =
µ(z)
µ(z) σ
2
k=1
,
z ∈ Φm,n ,
(2)
where Per Az is the permanent of Az and in the definition the summation is for all σ ∈ π[n], the set of
permutations of [n].
The Cauchy-Binet formula for permanents (see Marcus and Minc, 1965, page 579) can be used to show the
function q is indeed a pmf. There is a demonstration from first principles in the proof of Lemma 1. See also
Aaronson and Arkhipov (2011, Theorem 3.10) for a demonstration from physical principles.
The computing task is to simulate random samples from the pmf q(z).
3 Exact Boson Sampling Algorithms
We give two algorithms for Boson Sampling. Algorithm A runs in O(mn3n ) time to create a single sample.
n time complexity of the fastest solution in
This is already a significant speed-up over the Θ (m+n−1
n )n2
the literature. Algorithm A prepares the way for Algorithm B our main result with running time specified in
Theorem 1.
We will repeatedly need to sample from an unnormalised discrete probability mass function (weight function). To establish the complexity of our Boson Sampling algorithms we need only employ the most naive
linear time method although faster methods do exist. In particular, if the probability mass function is described by an array of length m then it is possible to sample in O(1) time after O(m) preprocessing time,
using Walker’s alias method (Walker, 1974; Kronmal and Peterson Jr, 1979).
Expanding the sample space: We start by expanding the sample space Φm,n and consider a distribution
on the product space [m]n , i.e. the space of all arrays r = (r1 , . . . , rn ), with each element in the range 1 to
m. Since the permanent of a matrix is invariant to row reordering, we have Per Ar = Per Az when r is any
permutation of z. Furthermore the number of distinguishable rearrangements of z is n!/µ(z).
We claim that in order to sample from q(z) we can equivalently sample from the pmf
n
1 XY
1
ari σi
|Per Ar |2 =
p(r) =
n!
n! σ
i=1
2
,
r ∈ [m]n .
(3)
In other words sample r from p(r) and then rearrange the elements of r in non-decreasing order to give z.
5
This follows since for any z there are n!/µ(z) equally likely values of r in the expanded sample space, i.e.
p(r) = p(z) for all such values of r. So by the addition rule for probabilities
q(z) =
n! 1
1
n!
p(z) =
|Per Az |2 =
|Per Az |2 ,
µ(z)
µ(z) n!
µ(z)
z ∈ Φm,n ,
as claimed.
This straightforward reformulation will make the task of efficient Boson Sampling significantly simpler.
Our approach focuses on the pmf p(r).
We begin by deriving the marginal pmfs of the leading subsequences of (r1 , . . . , rn ) in the following lemma.
This will then provide the pmfs of successive elements of (r1 , . . . , rn ) conditional on previous values. In that
way we are able to rewrite p(r1 , . . . , rn ) as a product of conditional pmfs, using the chain rule of probability,
and exploit the chain for progressively simulating r1 , . . . , rn .
Lemma 1. The joint pmf of the subsequence (r1 , . . . , rk ) is given by
p(r1 , . . . , rk ) =
(n − k)! X
Per Acr1 ,...,rk
n!
2
,
k = 1, . . . , n,
c∈Ck
where Ck is the set of k-combinations taken without replacement from [n] and Acr1 ,...,rk is the matrix formed
from rows (r1 , . . . , rk ) of the columns c of A.
Proof. The convention will be that c is a set and (r1 , . . . , rk ) is an array with possibly multiple instances of
its elements. The invariance of permanents under column and row permutations removes any ambiguity in
the interpretation of Per Acr1 ,...,rk .
P
For completeness we start by confirming that p(r) is a pmf, in other words r p(r) = 1. From the definition
of p(r) and multiplying by n!, we have
n!
X
p(r) =
r
n
X XY
X
=
n
XX XY
σ
=
ari σi
σ i=1
r
τ
n
XY
!
τ
τ
āri τi
i=1
ari σi āri τi
r i=1
m
n X
XXY
σ
, σ ∈ π[n]
ari σi
σ i=1
n
XY
r
=
2
!
, τ ∈ π[n]
!
ak,σi ākτi .
i=1 k=1
The product is 0 when σi 6= τi for any i, by the orthonormal
property of A. Otherwise when σ = τ the
P
product is 1 so the final expression reduces to n! and r p(r) = 1 as expected.
Now looking at the case k = 1 we have similarly
n!p(r1 ) =
X
r2 ...rn
=
XX
σ
n
XY
ari σi
σ i=1
ar1 σ1 ār1 τ1
τ
!
n
XY
τ
m
n X
Y
i=2 k=1
6
āri τi
i=1
ak,σi ākτi .
!
Again for this to be non-zero, each of the terms in the product must be 1. This means σi = τi , i = 2, . . . , n
which implies σ = τ .
It follows that
n!p(r1 ) =
X
σ
as claimed, so that
2
|ar1 σ1 | = (n − 1)!
n
X
k=1
|ar1 ,k |2 ,
1
|ar1 ,1 |2 + . . . , +|ar1 ,n |2 , r1 ∈ [m],
n
the sum of the squared moduli of elements of row r1 divided by n.
p(r1 ) =
In the same way,
n!p(r1 , r2 ) =
XX
σ
m
n X
Y
ar1 σ1 ār1 τ1 ar2 σ2 ār2 τ2
τ
ak,σi ākτi
i=3 k=1
!
.
Again the only non-zero case is when σi = τi , i = 3, . . . , n. However it does not imply that σ = τ , since
we can also have σ1 = τ2 , σ2 = τ1 .
Consequently,
n!p(r1 , r2 ) =
XX
ar1 σ1 ār1 τ1 ar2 σ2 ār2 τ2
σ
τ
σ3 ,...,σn =τ3 ,...,τn
= (n − 2)!
= (n − 2)!
X
c∈C2
X
X
σ∈π(c)
2
ar1 σ1 ār2 σ2
Per Acr1 ,r2
2
,
c∈C2
where C2 is the set of 2-combinations (without replacement) from [n], π(c) is the set of permutations of c
and Acr1 ,r2 is a matrix formed from rows (r1 , r2 ) and columns c of A.
In the same way
n!p(r1 , . . . , rk ) = (n − k)!
so that
p(r1 , . . . , rk ) =
X
Per Acr1 ,...,rk
2
,
c∈Ck
(n − k)! X
Per Acr1 ,...,rk
n!
2
,
k = 1, . . . , n.
c∈Ck
Our first Boson Sampling algorithm which we term Algorithm A samples from p(r) by expressing this as a
chain of conditional pmfs,
p(r) = p(r1 )p(r2 |r1 )p(r3 |r1 , r2 ) . . . p(rn |r1 , r2 , . . . , rn−1 ).
Sample r1 from p(r1 ), r1 ∈ [m]. Subsequently for stages k = 2, . . . , n, sample rk from the conditional pmf
p(rk |r1 , . . . , rk−1 ), rk ∈ [m] with (r1 , . . . , rk−1 ) fixed. At completion, the array (r1 , . . . , rn ) at stage n will
have been sampled from the pmf p(r). Sorting (r1 , . . . , rn ) in non-decreasing order, the resulting multiset
z is sampled from the Boson Sampling distribution q(z).
7
Algorithm A Boson Sampler: single sample z from q(z) in O(mn3n ) time
Require: m and n positive integers; A first n columns of m × m Haar random unitary matrix
1: r ← ∅
⊲ E MPTY ARRAY
2: FOR k ← 1 TO n DO
P
⊲ M AKE INDEXED WEIGHT ARRAY w
3:
wi ← c∈Ck | Per Ac(r,i) |2 , i ∈ [m]
4:
x ← S AMPLE (w)
⊲ S AMPLE INDEX x FROM w
5:
r ← (r, x)
⊲ A PPEND x TO r
6: END FOR
7:
z ← I NC S ORT (r)
z
⊲ S ORT r IN
NON - DECREASING ORDER
8: RETURN
Theorem 2. Algorithm A samples from the Boson Sampling distribution with time complexity O(mn3n ) per
sample. The additional space complexity of Algorithm A on top of that needed to store the input is O(m)
words.
Proof. The correctness follows from the chain rule for conditional probabilities and so we now analyse
the running time. From Lemma 1, evaluation of the pmf p(r1 ) for r1 ∈ [m] involves the sum of squared
moduli in each row of A, a total of O(mn) operations. Sampling from this pmf takes O(m) time. More
generally we can write p(rk |r1 , . . . , rk−1 ) = p(r1 , . . . , rk )/p(r1 , . . . , rk−1 ). Note that rk does not appear
in the denominator, so to sample rk from this conditional pmf we can equivalently sample from the pmf that
is proportional to the numerator, considered purely as function of rk with r1 , . . . , rk−1 fixed. The numerator
involves the calculating of several k × k permanents. With the best known algorithms for computing the
permanent, each requires O(k2k ) operations. There are (nk ) terms in Ck for each value of k, making a
combined operation count at stage k of O(m(nk )k2k ). Using
n
X
k=1
n
2
k2
= n3n ,
3
k
k
the total operation count is then O(mn3n ). We only need to store the result of one permanent calculation at
a time and also one list of m probabilities at a time. This gives a total space usage of O(m) excluding the
space needed to store the original matrix A.
We now prepare to prove Theorem 1.
Laplace expansion: We make extensive use of the Laplace expansion for permanents (see Marcus and Minc,
1965, page 578), namely that for any k × k matrix B = (bi,j ),
Per B =
k
X
⋄
bk,ℓ Per Bk,ℓ
,
ℓ=1
⋄ is the submatrix with row k and column ℓ removed. Note that B ⋄ only depends on B ⋄ the subwhere Bk,ℓ
k,ℓ
k
matrix of B with the k-th row removed. An important consequence is that when B is modified by changing
⋄ } are
its k-th row, the modified permanent can be calculated in O(k) steps, provided the values {Per Bk,ℓ
available.
⋄ , ℓ ∈ [k]} has the same time complexity as computing
We show that computation of all of the values {Per Bk,ℓ
Per B, the permanent of a single k × k matrix. We will later take advantage of this fast amortised algorithm
8
in the proof of Theorem 1 to quickly compute a set of permanents of matrices each with one row differing
from the other.
⋄ } be submatrices of B with row k and column
Lemma 2. Let B be a k × k complex matrix and let {Bk,ℓ
⋄ , ℓ ∈ [k]} can be evaluated jointly in O(k2k ) time and O(k)
ℓ removed, ℓ ∈ [k]. The collection {Per Bk,ℓ
additional space.
Proof of Lemma. By applying Glynn’s formula (1) for a given value of ℓ we have:
!
X k−1
Y
Y
1
⋄
vj (δ), ℓ ∈ [k],
δi
= k−2
Per Bk,ℓ
2
δ
where δ ∈ {−1, 1}k−1 with δ1 = 1 and vj (δ) =
i=1
Pk−1
i=1
j∈[k]\ℓ
δi bij .
Exploiting the usual trick of working through values of δ in Gray code order, the terms {vj (δ), j ∈ [k]} can
be evaluated in O(k) combined time for every new δ. This is because successive δ arrays will differ in one
element. The product of the vj (δ) terms can therefore also be computed in O(k) time giving O(k2k ) time
⋄ for a single value of ℓ, but of course this has to be replicated k times to cover all values
to compute Per Bk,ℓ
of ℓ.
Q
⋄ ℓ ∈ [k]} more efficiently we observe that each product
To compute {Per Bk,ℓ
j∈[k]\ℓ vj (δ) can be exQℓ−1
Qk
pressed as fℓ bℓ where fℓ = j=1 vj (δ), ℓ = 2, . . . , k and bℓ = j=ℓ+1 vj (δ), ℓ = 1, . . . , k − 1 are forward
and backward
cumulative products, with f1 = bk = 1. We can therefore compute all of the partial prodQ
ucts j∈[k]\ℓ vj (δ) in O(k) time, giving an overall total time complexity of O(k2k ) for jointly computing
⋄ , ℓ ∈ [k]}. Other than the original matrix, space used is dominated by the array of k accumulating
{Per Bk,ℓ
partial products.
Furthermore the computation time has constant factor overheads similar to that of computing Per B (see
Appendix).
Auxiliary column indices: We now expand the sample space even further with an auxiliary array α =
(α1 , . . . , αn ), where α ∈ π[n], the set of permutations of [n].
As with Lemma 1, the purpose is to create a succession of pmfs for leading subsequences of the array
(r1 , . . . , rn ) and thereby provide successive conditional pmfs so that r1 , . . . , rn can be simulated progressively. The novelty here is to introduce an overall conditioning variable α so that sampling can be carried out
more rapidly under that condition while demonstrating that the marginal distribution still correctly provides
the pmf p(r) in (3).
Define
1
[n]\{α
,...,αn } 2
, k = 1, . . . , n − 1.
Per Ar1 ,...,rkk+1
k!
P
For notational convenience let ek = φ(r1 , . . . , rk |α) and dk = rk φ(r1 , . . . , rk |α) for k = 1, . . . , n − 1
with en = p(r1 , . . . , rn ) and dn = p(r1 , . . . , rn−1 ). Note that φ(r1 , . . . , rk |α) is a pmf on [m]k , equivalent
to (3) using the subset {α1 , . . . , αk } of columns of A, so that in particular d1 = 1.
Q
Lemma 3. With the preceding notation, let φ(r|α) = nk=1 ek /dk then p(r) = Eα {φ(r|α)} where the
expectation is taken over α, uniformly distributed on π[n] for fixed r.
φ(r1 , . . . , rk |α) =
9
Proof. For the particular case k = n − 1, Lemma 1 shows that p(r1 , . . . , rn−1 ) can be written as a mixture
of component pmfs and expressed as an expectation over αn , i.e.
n
p(r1 , . . . , rn−1 ) =
1
1X
j
Per Ar[n]\
1 ,...,rn−1
n
(n − 1)!
j=1
2
= Eαn {φ(r1 , . . . , rn−1 |α)},
where αn is uniformly distributed on [n]. In other words dn is the expectation of en−1 over αn . More
generally when the leading subsequence is (r1 , . . . , rk ) and the active columns of A are [n]\{αk+1 , . . . , αn },
we have
X
rk
φ(r1 , . . . , rk |α) = Eαk |αk+1 ,...,αn {φ(r1 , . . . , rk−1 |α)},
k = 2, . . . , n − 1,
where the expectation is taken over αk uniformly distributed on [n] \ {αk+1 , . . . , αn } with (αk+1 , . . . , αn )
and r fixed. This means dk is the conditional expectation of ek−1 .
Q
We now rewrite φ(r|α) as en nk=2 ek−1 /dk since d1 = 1 and start by considering the expectations of the
terms ek−1 /dk , k = 2, . . . , n with respect to α2 given α3 , . . . , αn . With the exception of e1 these terms are
specified by subsets of {α3 , . . . , αn }, so they remain fixed
Since the conditional
Qunder the conditioning.
Q
expectation of e1 is d2 , the conditional expectation of en nk=2 ek−1 /dk is en nk=3 ek−1 /dk , i.e. there is
one fewer term in the product. Proceeding in this way and successively taking conditional expectations
with respect to αk−1 uniformly distributed on [n] \ {αk , . . . , αn } conditional on (αk , . . . , αn ), the product
telescopes to en (en−1 /dn ) = p(r1 , . . . , rn )en−1 /dn . Finally taking expectations over αn , this reduces to
p(r1 , . . . , rn ), since dn is the expected value of en−1 . Note a random permutation (α1 , . . . , αn ) is generated
starting from αn in reverse order by successively selecting elements for inclusion at random from what
remains. By the chain rule of expectation, we have then shown that Eα {φ(r|α)} = p(r) as claimed.
We can now describe our second and faster Boson Sampling algorithm. This algorithm is the basis of
our main result, Theorem 1. We start by sampling a random permutation α = (α1 , . . . , αn ) uniformly
distributed on π[n]. As in Algorithm A we consider a chain of conditional pmfs,
φ(r|α) =
n
Y
k=1
ek /dk = φ(r1 |α)φ(r2 |r1 , α)φ(r3 |r1 , r2 , α) . . . φ(rn |r1 , r2 , . . . , rn−1 , α),
where now φ(rk |r1 , . . . , rk−1 , α) = ek /dk , with the notation defined prior to Lemma 3.
At stage 1, we sample r1 from the pmf φ(r1 |α). Subsequently for stages k = 2, . . . , n, we sample rk
from φ(rk |r1 , . . . , rk−1 , α) with (r1 , . . . , rk−1 ) fixed, exploiting the Laplace expansion for permanents. At
completion, the array (r1 , . . . , rn ) at stage n will have been sampled from the pmf p(r). Sorting (r1 , . . . , rn )
in non-decreasing order, the resulting multiset z is sampled from the Boson Sampling distribution q(z).
Proof of Theorem 1. It follows from the chain rule for conditional probabilities that for a given fixed α
Algorithm B samples from φ(r|α). As α is uniformly distributed on π[n] the algorithm therefore samples
from the marginal distribution Eα {φ(r|α)} which by Lemma 3 is the pmf p(r) in (3). The complexity is
established as follows. At stage 1, the pmf to be sampled is
2 ,...,αn } 2
φ(r1 |α) = | Per A[n]\{α
| = | Per Aαr11 |2 = |ar1 ,α1 |2 ,
r1
r1 ∈ [m].
Sampling takes O(m) operations. At stage 2 for fixed r1 , we need to sample r2 from the pmf φ(r2 |r1 , α).
But since φ(r2 |r1 , α) = e2 /d2 and d2 does not involve r2 , we can sample from the pmf that is proportional
10
Algorithm B Boson sampler: single sample z from q(z) in O(n2n + poly(m, n)) time
Require: m and n positive integers; A first n columns of m × m Haar random unitary matrix
1: r ← ∅
⊲ E MPTY ARRAY
2: A ← P ERMUTE (A)
⊲ R ANDOMLY PERMUTE COLUMNS OF A
3: wi ← |ai,1 |2 , i ∈ [m]
⊲ M AKE INDEXED WEIGHT ARRAY w
4: x ← S AMPLE (w)
⊲ S AMPLE INDEX x FROM w
5: r ← (r, x)
⊲ A PPEND x TO r
6: FOR k ← 2 TO n DO
[k]
7:
Bk⋄ ← Ar
⋄ , ℓ ∈ [k]}
8:
C OMPUTE {Per Bk,ℓ
⊲ A S L EMMA 2
Pk
⋄
2
9:
wi ← | ℓ=1 ai,ℓ Per Bk,ℓ | , i ∈ [m]
⊲ U SING L APLACE EXPANSION
10:
x ← S AMPLE (w)
11:
r ← (r, x)
12: END FOR
13:
z ← I NC S ORT (r)
z
⊲ S ORT r IN
NON - DECREASING ORDER
14: RETURN
to e2 considered simply as a function of r2 with r1 fixed. Since e2 is proportional to | Per Aαr11,r,α22 |2 this
involves calculating m permanents of 2 × 2 matrices and then sampling; a further O(m) operations.
At stage k with r1 , . . . , rk−1 already determined, we need to sample rk from the pmf proportional to
α1 ,...,αk
,...,αk 2
| Per Aαr11,...,r
k | considered simply as a function of rk . Let B = Ar1 ,...,rk for some arbitrary value of
rk , i.e. bi,j = ari ,αj for i, j = 1, . . . , k. Using the Laplace expansion and taking squared moduli
,...,αk 2
Per Aαr11,...,r
k
=
k
X
2
⋄
ark ,αℓ Per Bk,ℓ
ℓ=1
, rk ∈ [m],
(4)
⋄ is defined in the statement of Lemma 2. From Lemma 2 all the values {Per B ⋄ } can be
where Bk,ℓ
k,ℓ
computed in O(k2k ) time. An array of length m is then obtained from (4) by summing k terms for each
rk ∈ [m]. Taking a single sample from the pmf proportional to the array takes O(m) time. This gives a total
time complexity for stage k of O(k2k ) + O(mk).
The total time for all stages k = 1, . . . , n is therefore O(n2n ) + O(mn2 ). In practical terms, the exponentially growing term is approximately equivalent to the cost of evaluating two n × n permanents. The space
usage is dominated by storing a single unnormalised pmf of size O(m) prior to sampling.
2
,...,αn
2
Remarks: At the final stage when rn is selected the value of |Per Aαr11,...,r
n | = |Per Ar | will have been
computed as in (4). Hence q(r) the pmf of the sample r is available at no extra computational cost. Finally
note that when only a single value is to be sampled from q(r) and since the columns of a Haar random
unitary matrix are already in a random order, the random permutation (α1 , . . . , αn ) in Algorithm B can be
taken to be (1, 2, . . . , n)
4 Acknowledgements
RC was funded by EPSRC Fellowship EP/J019283/1. Thank you to Ashley Montanaro, Alex Neville,
Anthony Laing and members of the Bristol QET Labs for introducing us to the Boson Sampling problem
11
and for a number of detailed and helpful conversations.
Appendix
Constant factor overheads in Lemma 2: To compare the constant factor overheads in Lemma 2, we start
by looking at Per B. From Glynn’s formula we have
!
k
Y
1 X Y
vj∗ (δ),
δi
Per B = k−1
2
i=1
δ
j∈[k]
P
where δ ∈ {−1, 1}k with δ1 = 1 and vj∗ (δ) = ki=1 δi bij . As we work through the values of δ in Gray code
order, updating vj∗ (δ), j ∈ [k] requires k additions. Evaluating the product requires k −1 multiplications and
there is one further addition as the sum is accumulated. There are 2k−1 steps making a total of 2k2k−1 = k2k
operations.
⋄ , ℓ ∈ [k]} in Lemma 2, at each stage there are k additions for updating
For the calculation of {Per Bk,ℓ
{vj (δ), j ∈ [k]}, then k − 2 multiplications
Q for calculating each of the sets {fℓ }, {bℓ }, and {fℓ bℓ }. Having
thereby obtained all the partial products j∈[k] vj (δ)/vℓ (δ), a further k additions are required as the k partial
sums are accumulated. So each of the 2k−2 steps requires 5k − 6 operations. Consequently the operation
count is approximately 25% larger than that of Per B. In practice addition and multiplication with complex
arithmetic will have different costs in terms of floating point operations.
Collision probability: Suppose that z is sampled from q(z) in (2) and that s is the associated array of
counts. The probability of duplicate elements in z is the probability that one of the elements of s is larger
than 1. In general this probability will depend on the underlying Haar random unitary A.
From Lemma 1 we have
E(si ) =
n
X
k=1
E(si (si − 1)) =
|ai,k |2
X
Perm Aci,i
c∈C2
2
=4
X
k<ℓ
|ai,k ai,ℓ |2
(5)
Let us denote the A-specific duplication probability, i.e. the chance of seeing duplicated values in z, by PD .
We have from (5)
PD 6
m
X
i=1
m
m
P(si > 2) 6
XX
1X
E si (si − 1) = 2
|ai,k ai,ℓ |2 .
2
i=1 k<ℓ
i=1
Note that the expected value of the upper bound is n(n−1)/(m+1) ∼ n2 /m when averaged over A from the
Haar random unitary class. See Aaronson and Arkhipov (2011, Theorem C.4) and Arkhipov and Kuperberg
(2012) for comparison.
12
References
Aaronson, S. and Arkhipov, A. (2011). The computational complexity of linear optics. In STOC ’11: Proc.
43rd Annual ACM Symp. Theory of Computing, pages 333–342.
Aaronson, S. and Arkhipov, A. (2014). Bosonsampling is far from uniform. Quantum Information &
Computation, 14(15-16):1383–1423.
Arkhipov, A. and Kuperberg, G. (2012). The bosonic birthday paradox. Geometry & Topology Monographs,
18:1–7.
Bentivegna, M., Spagnolo, N., Vitelli, C., Flamini, F., Viggianiello, N., Latmiral, L., Mataloni, P., Brod,
D. J., Galvão, E. F., Crespi, A., et al. (2015). Experimental scattershot boson sampling. Science advances,
1(3):e1400255.
Björklund, A. (2016). Below all subsets for some permutational counting problems. In 15th Scandinavian
Symposium and Workshops on Algorithm Theory (SWAT 2016), pages 1–17.
Boixo, S., Isakov, S. V., Smelyanskiy, V. N., Babbush, R., Ding, N., Jiang, Z., Bremner, M. J., Martinis,
J. M., and Neven, H. (2016). Characterizing quantum supremacy in near-term devices. arXiv:1608.00263.
Broome, M. A., Fedrizzi, A., Rahimi-Keshari, S., Dove, J., Aaronson, S., Ralph, T. C., and White, A. G.
(2013). Photonic boson sampling in a tunable circuit. Science, 339(6121):794–798.
Clifford,
P.
and
Clifford,
R.
(2017).
R:
Classical
https://cran.r-project.org/package=BosonSampling.
Boson
Sampling.
Crespi, A., Osellame, R., Ramponi, R., Brod, D. J., Galvão, E. F., Spagnolo, N., Vitelli, C., Maiorino, E.,
Mataloni, P., and Sciarrino, F. (2013). Integrated multimode interferometers with arbitrary designs for
photonic boson sampling. Nature Photonics, 7(7):545–549.
Efraimidis, P. S. (2015). Weighted random sampling over data streams. In Algorithms, Probability, Networks, and Games, pages 183–195. Springer.
Feller, W. (1968). An Introduction to Probability Theory and its Applications. - Vol. 1. Wiley.
Glynn, D. G. (2010). The permanent of a square matrix. European Journal of Combinatorics, 31(7):1887–
1891.
Green, P. J., Latuszyński, K., Pereyra, M., and Robert, C. P. (2015). Bayesian computation: a summary of
the current state, and samples backwards and forwards. Statistics and Computing, 25(4):835–862.
Grover, L. K. (1996). A fast quantum mechanical algorithm for database search. In STOC ’96: Proc. 28th
Annual ACM Symp. Theory of Computing, pages 212–219.
Harrow, A. W. and Montanaro, A. (2017). Quantum computational supremacy. Nature, 549(7671):203–209.
Hastings, W. K. (1970). Monte Carlo sampling methods using Markov chains and their applications.
Biometrika, 57(1):97–109.
Kronmal, R. A. and Peterson Jr, A. V. (1979). On the alias method for generating random variables from a
discrete distribution. The American Statistician, 33(4):214–218.
Latmiral, L., Spagnolo, N., and Sciarrino, F. (2016). Towards quantum supremacy with lossy scattershot
boson sampling. New Journal of Physics, 18(11).
Liu, J. S. (2008). Monte Carlo Strategies in Scientific Computing. Springer Science & Business Media.
13
Lund, A. P., Bremner, M. J., and Ralph, T. C. (2017). Quantum sampling problems, bosonsampling and
quantum supremacy. NPJ Quantum Information, 3(1):15.
Marcus, M. and Minc, H. (1965). Permanents. The American Mathematical Monthly, 72(6):577–591.
Neville, A., Sparrow, C., Clifford, R., Johnston, E., Birchall, P. M., Montanaro, A., and Laing, A. (2017).
Classical boson sampling algorithms with superior performance to near-term experiments. Nature
Physics. Advance online publication. doi:10.1038/nphys4270.
Nijenhuis, A. and Wilf, H. S. (1978). Combinatorial Algorithms: for Computers and Calculators. Academic
press.
Propp, J. G. and Wilson, D. B. (1996). Exact sampling with coupled Markov chains and applications to
statistical mechanics. Random Structures and Algorithms, 9(1&2):223–252.
Ryser, H. J. (1963). Combinatorial Mathematics, volume 14 of Carus Mathematical Monographs. Mathematical Association of America.
Shchesnovich, V. (2013). Asymptotic evaluation of bosonic probability amplitudes in linear unitary
networks in the case of large number of bosons. International Journal of Quantum Information,
11(05):1350045.
Shor, P. W. (1997). Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum
computer. SIAM Journal on Computing, 26(5):1484–1509.
Spagnolo, N., Vitelli, C., Bentivegna, M., Brod, D. J., Crespi, A., Flamini, F., Giacomini, S., Milani, G.,
Ramponi, R., Mataloni, P., et al. (2014). Experimental validation of photonic boson sampling. Nature
Photonics, 8(8):615–620.
Spring, J. B., Metcalf, B. J., Humphreys, P. C., Kolthammer, W. S., Jin, X.-M., Barbieri, M., Datta, A.,
Thomas-Peter, N., Langford, N. K., Kundys, D., et al. (2013). Boson sampling on a photonic chip.
Science, 339(6121):798–801.
Terhal, B. M. and DiVincenzo, D. P. (2004). Adaptive quantum computation, constant depth quantum
circuits and Arthur-Merlin games. Quantum Information & Computation, 4(2):134–145.
Tichy, M. C. (2011). Entanglement and Interference of Identical Particles. PhD thesis, Freiburg University.
Tichy, M. C., Mayer, K., Buchleitner, A., and Mølmer, K. (2014). Stringent and efficient assessment of
boson-sampling devices. Physical Review Letters, 113(2):020502.
Tillmann, M., Dakić, B., Heilmann, R., Nolte, S., Szameit, A., and Walther, P. (2013). Experimental boson
sampling. Nature Photonics, 7(7):540–544.
Valiant, L. (1979). The complexity of computing the permanent. Theoretical Computer Science, 8(2):189 –
201.
Walker, A. J. (1974). New fast method for generating discrete random numbers with arbitrary frequency
distributions. Electronics Letters, 10(8):127–128.
Walschaers, M., Kuipers, J., Urbina, J.-D., Mayer, K., Tichy, M. C., Richter, K., and Buchleitner, A. (2016).
Statistical benchmark for bosonsampling. New Journal of Physics, 18(3):032001.
Wang, H., He, Y., Li, Y.-H., Su, Z.-E., Li, B., Huang, H.-L., Ding, X., Chen, M.-C., Liu, C., Qin, J., Li,
J.-P., He, Y.-M., Schneider, C., Kamp, M., Peng, C.-Z., Hoefling, S., Lu, C.-Y., and Pan, J.-W. (2016).
Multi-photon boson-sampling machines beating early classical computers. arXiv:1612.06956.
14
Wang, S.-T. and Duan, L. (2016). Certification of boson sampling devices with coarse-grained measurements. Bulletin of the American Physical Society.
Wu, J., Liu, Y., Zhang, B., Jin, X., Wang, Y., Wang, H., and Yang, X. (2016). Computing permanents for
boson sampling on Tianhe-2 supercomputer. arXiv:1606.05836.
15
| 8cs.DS
|
arXiv:1702.01692v1 [cs.NE] 6 Feb 2017
Distributed Evolutionary k-way Node Separators∗
Peter Sanders
Christian Schulz
Karlsruhe Institute of Technology
Karlsruhe, Germany
[email protected]
Karlsruhe Institute of Technology
Karlsruhe, Germany and University of Vienna
Vienna, Austria
[email protected]
Darren Strash
Robert Williger
Colgate University
Hamilton, USA
[email protected]
Karlsruhe Institute of Technology
Karlsruhe, Germany
[email protected]
ABSTRACT
Computing high quality node separators in large graphs is
necessary for a variety of applications, ranging from divideand-conquer algorithms to VLSI design. In this work, we
present a novel distributed evolutionary algorithm tackling
the k-way node separator problem. A key component of
our contribution includes new k-way local search algorithms
based on maximum flows. We combine our local search with
a multilevel approach to compute an initial population for
our evolutionary algorithm, and further show how to modify
the coarsening stage of our multilevel algorithm to create effective combine and mutation operations. Lastly, we combine
these techniques with a scalable communication protocol,
producing a system that is able to compute high quality
solutions in a short amount of time. Our experiments against
competing algorithms show that our advanced evolutionary
algorithm computes the best result on 94% of the chosen
benchmark instances.
KEYWORDS
graph partitioning, node separators, max-flow min-cut
1
INTRODUCTION
Given a graph G = (V, E), the k-way node separator problem
is to find a small separator S ⊂ V , and k disjoint subsets of
V , V1 , . . . , Vk called blocks, such that no edgesSexist between
two different blocks Vi and Vj (i =
6 j) and V = i Vi ∪S. The
objective is to minimize the size of the separator S or, depending on the application, the cumulative weight of its nodes,
while the blocks Vi are balanced. Note that removing the set
S from the graph results in at least k connected components.
Many algorithms rely on small node separators. For example, small balanced separators are a popular tool in divideand-conquer strategies [2, 18, 20], are useful to speed up the
computations of shortest paths [7, 8, 28], are necessary in
scientific computing to compute fill reducing orderings with
nested dissection algorithms [13] or in VLSI design [2, 18].
∗
This work was partially supported by DFG grants SA 933/11-1.
Finding a balanced node separator is NP-hard for general
graphs even if the maximum node degree is three [4, 12].
Therefore, heuristic and approximation algorithms are used
in practice to find small node separators. The most commonly
used method to solve the node separator problem on large
graphs in practice is the multilevel approach. During a
coarsening phase, a multilevel algorithm reduces the graph
size by iteratively contracting the nodes and edges of G until
the graph is small enough to compute a node separator by
some other (presumably time consuming) algorithm. A node
separator of the input graph is then constructed by iteratively
uncontracting the graph, transferring the solution to this finer
graph, and then applying local search algorithms to improve
the solution.
Although current solvers are typically fast enough for most
applications, they unfortunately produce separators of low
solution quality. This may be acceptable in applications that
use a separator just once, however, many applications first
compute a separator as a preprocessing step, and then rely
on a high-quality separator for speed in later stages. This is
true in VLSI design [2, 18], where even small improvements
in separator size can have a large impact on computation
time and production costs. High-quality node separators can
also speed up shortest path queries in road networks, for
example, in customizable contraction hierarchies [8], where
smaller node separators yield better node orderings that are
repeatedly used to answer shortest path queries. The cost for
computing one high quality node separator is then amortized
over a many shortest path queries. Hence, our focus is on
solution quality in this work.
1.1
Our Results
The main contribution of this paper is a technique that
integrates an evolutionary search algorithm with a novel
multilevel k-node separator algorithm and its scalable parallelization. We present novel mutation and combine operators
for the problem which are based on the multilevel scheme.
Due to the coarse-grained parallelization, our system is able
to compute separators that have high quality within a few
minutes for graphs of moderate size.
Peter Sanders, Christian Schulz, Darren Strash, and Robert Williger
2 PRELIMINARIES
2.1 Basic Concepts
Throughout this paper, we consider an undirected graph
G = (V = {0, . . . , n − 1}, E) with n = |V |, and m = |E|.
Γ(v) := {u : {v, u} ∈ E} denotes the neighborhood of a node
v. A graph S = (V 0 , E 0 ) is said to be a subgraph of G = (V, E)
if V 0 ⊆ V and E 0 ⊆ E ∩ (V 0 × V 0 ). We call S an induced
subgraph when E 0 = E ∩(V 0 ×V 0 ). For a set of nodes U ⊆ V ,
G[U ] denotes the subgraph induced by U .
The graph partitioning problem, which is closely related to
the node separator problem, asks for blocks of nodes V1 ,. . . ,Vk
that partition V (i.e., V1 ∪· · ·∪Vk = V and Vi ∩Vj = ∅ for i =
6
j). A balancing constraint demands that ∀i ∈ {1..k} : |Vi | ≤
Lmax := (1+)d|V |/ke for some parameter .PIn this case, the
objective is often to minimize the total cut
i<j |Eij | where
Eij := {{u, v} ∈ E : u ∈ Vi , v ∈ Vj }. The set of cut edges is
also called edge separator. An abstract view of the partitioned
graph is the so called quotient graph, where nodes represent
blocks and edges are induced by connectivity between blocks.
The node separator problem asks to find blocks, V1 , . . . , Vk
and a separator S that partition V such that there are no
edges between the blocks. Again, a balancing constraint
demands |Vi | ≤ (1+)d|V |/ke. However, there is no balancing
constraint on the separator S. The objective is to minimize
the size of the separator |S|. Note that removing the set S
from the graph results in at least k connected components
and that the blocks Vi itself do not need to be connected
components. Two blocks of Vi and Vj are adjoint if there
exists a separator node s ∈ S that connects both blocks.
Note that s can separate more than two blocks. For the
separator case, edges in the quotient graph are induced by
adjoint blocks. By default, our initial inputs will have unit
edge and node weights. However, the results in this paper
are easily transferable to node and edge weighted problems.
A matching M ⊆ E is a set of edges that do not share any
common nodes, i.e. the graph (V, M ) has maximum degree
one. Contracting an edge {u, v} replaces the nodes u and v
by a new node x connected to the former neighbors of u and
v. We set c(x) = c(u) + c(v). If replacing edges of the form
{u, w} , {v, w} would generate two parallel edges {x, w}, we
insert a single edge with ω({x, w}) = ω({u, w}) + ω({v, w}).
Uncontracting an edge e “undoes” its contraction. In order to
avoid tedious notation, G will denote the current state of the
graph before and after a (un)contraction unless we explicitly
want to refer to different states of the graph.
The multilevel approach consists of three main phases. In
the contraction (coarsening) phase, we iteratively identify
matchings M ⊆ E and contract the edges in M . Contraction
should quickly reduce the size of the input and each computed
level should reflect the global structure of the input graph.
Contraction is stopped when the graph is small enough so
that the problem can be solved by some other potentially
more expensive algorithm. In the local search (or uncoarsening) phase, matchings are iteratively uncontracted. After
uncontracting a matching, a local search algorithm moves
nodes to decrease the size of the separator or to to improve
balance of the block while keeping the size of the separator.
The intuition behind the approach is that a good solution at
one level of the hierarchy will also be a good solution on the
next finer level so that local search will quickly find a good
solution.
2.2
Related Work
Here, we focus on results closely related to our main contributions, as well as previous work on the node separator
problem. However, we briefly mention that there has been
a huge amount of research on graph partitioning, which is
closely related to the node separator problem. We refer the
reader to [3, 5] for thorough reviews of the results in this
area.
2-way Node Separators. In contrast to the NP-hardness of
the problem in general, Lipton and Tarjan [19] showed that
small 2-way balanced separators can always be found in linear
time for planar graphs. Their planar separator theorem states
that, for planar graphs, one can always
p find a 2-way separator
S in linear time such that |S| = O( |V |) and |Vi | ≤ 2|V |/3.
Note that, to achieve better balance, the problem remains
NP-hard [11] even for planar graphs.
For general graphs there are several heuristics to compute
small node separators. A common and simple method is
to first compute an edge separator using a multilevel graph
partitioning algorithm, and then compute a node separator
by selecting nodes incident to the edge separator [24, 27]. A
major drawback to this method is that the graph partitioning
objective—to minimize the number of cut edges—differs from
the objective of the node separator problem. This difference
in combinatorial structure, unfortunately, means that graph
partitioning approaches are unlikely to find high quality solutions.
The Metis [16] and Scotch [23] graph partitioners use a
multilevel approach to obtain a 2-way node separator. After
contraction, both tools compute a node separator on the
coarsest graph using a greedy algorithm. This separator is
then transferred level-by-level, dropping non-needed nodes
on each level and applying Fiduccia-Mattheyses (FM) style
local search. LaSalle and Karypis [17] further gave a sharedmemory parallel algorithm and showed that a multilevel
approach combining greedy local search with a segmented
FM algorithm can outperform serial FM algorithms.
Other recent approaches look at variations of the node
separator problem, such as Pareto solutions for edge cut
versus balance [15], and enforcing both upper and lower
bounds on block sizes [14].
k-way Node Separators. In the theoretical algorithms literature, a multiway separator is the generalization of separator
to a higher number of blocks. Frederickson [10] first gave
this generalization, and showed that planar graphs can be
partitioned into O(|V |/k)√subsets of size at most k, where
each subgraph shares O( k) boundary vertices with other
subgraphs. Frederickson’s generalization of a separator is
different than what we consider here: we do not attempt to
Distributed Evolutionary k-way Node Separators
minimize the number of nodes that separate pairs of blocks,
instead we minimize the total number of nodes that collectively separate all blocks. In practice, the balance constraint
on the block size may also implicitly minimize the number
of separator nodes between blocks, but this is not strictly
enforced.
We are currently unaware of any existing algorithm to
compute a k-way node separator. However, we note that,
similar to 2-way node separators, a k-way node separator
may be computed by first computing a k-way edge separator,
and then keeping some or all of the nodes incident to the
edge separator as a node separator.
2.3
Detailed Related Work
Recently, we presented a new multilevel algorithm for the
2-way node separator problem [26]. We outline the details
of the multilevel algorithm here since we modify this algorithm to compute k-way node separators and then use the
modified algorithm to compute the initial population of the
evolutionary algorithm.
During coarsening, they use a two-phase approach, which
makes contraction more systematic by separating two issues:
A rating function and a matching algorithm. Hence, the
coarsening algorithm captures both local information and
the global structure of the graph. While the rating function
allows a flexible characterization of what a “good” contracted
graph is, the simple, standard definition of the matching
problem allows to reuse previously developed algorithms for
weighted matching. Their method further uses the Global
Path Algorithm (GPA) [21] to compute matchings, which
runs in near-linear time. GPA scans the edges in order of decreasing weight (rating) but rather than immediately building
a matching, it first constructs a collection of paths and evenlength cycles. Afterwards, optimal solutions are computed for
each of these paths and cycles using dynamic programming.
Coarsening is stopped as soon as the graph has less than
ten thousand nodes. To compute an initial separator, Sanders
and Schulz compute an edge separator, from which they derive
a node separator. Conversion is done by using all boundary
nodes as initial separator S and using the flow-based technique
described below to select the smallest separator contained in
the induced bipartite subgraph.
Two-way local search is based on localized local search
and flow-based techniques. Localized local search algorithms
for the node separator problem are initialized only with a
subset of given separator instead of the whole separator set.
In each iteration a separator node with the highest priority
not violating the balance constraint is moved. The priority is
based on the gain concept, i.e. the decrease in the objective
when the separator node is moved into a block. More precisely,
the gain of the node is the weight of v minus the weight of
the nodes that have to be added to the separator once it is
moved. Each node is moved at most once out of the separator
within a single local search. After a node is moved, newly
added separator nodes become eligible for movement.
The other local search algorithm contributed is based on
maximum flows. More precisely, the authors solve a nodecapacitated flow problem F = (VF , EF ) to improve a given
node separator. We shortly outline the details. Given a set
of nodes A ⊂ V , its border is defined as ∂A := {u ∈ A |
∃(u, v) ∈ E : v 6∈ A}. The set ∂1 A := ∂A ∩ V1 is called
left border of A and the set ∂2 A := ∂A ∩ V2 is called right
border of A. An A induced flow problem F is the node
induced subgraph G[A] using ∞ as edge-capacities and the
node weights of the graph as node-capacities. Additionally
there are two nodes s, t that are connected to the border of
A. More precisely, s is connected to all left border nodes ∂1 A
and all right border nodes ∂2 A are connected to t. These new
edges get capacity ∞. Note that the additional edges are
directed. F has the balance property if each (s,t)-flow induces
a balanced node separator in G, i.e. the blocks Vi fulfill the
balancing constraint. The basic idea is to construct a flow
problem F having the balance property. Such a subgraph is
found by performing breadth first searches (BFS) initialized
with the separator nodes.
The algorithms starts by setting A to S and extending
it by performing two BFS. The first BFS is initialized with
the current separator nodes S and only looks at nodes in
block V1 . The same is done during the second BFS with
the difference that now looks at nodes of block V2 . Each
node touched by any of the BFS is added to A. The BFSs
are stopped in such a way that the final solution of the flow
problem (which can be transformed into a separator of the
original problem) yields a balanced separator in the original
graph. To obtain even better solutions, larger flow-problems
can be defined by dropping the requirement that each cut in
the flow problem corresponds to a balanced node separator
in the original graph. If the resulting node separator is not
balanced, then the algorithms starts again with a smaller flow
problems, i.e. stopping the BFS to define the flow problem
earlier.
To improve solution quality, the notion of iterated multilevel
schemes has been introduced to the node separator problem.
Here, one transfers a solution of a previous multilevel cycle
down the hierarchy and uses it as initial solution. More
precisely, this can be done by not contracting any cut edge,
i.e. an edge running between a block and the separator. This
is achieved by modifying the matching algorithm to not match
any edge that runs between Vi and S (i = 1, 2, . . . , k). Hence,
when contraction is done, every edge leaving the separator
will remain and one can transfer the node separator down
in the hierarchy. Thus a given node separator can be used
as initial node separator of the coarsest graph (having the
same balance and size as the node separator of the finest
graph). This ensures non-decreasing quality, if the local
search algorithm guarantees no worsening.
3
LOCAL SEARCH AND BALANCING
Local search is used to improve a given solution to the node
separator problem on each level of the multilevel hierarchy.
As the term implies we locally search for a better separator
Peter Sanders, Christian Schulz, Darren Strash, and Robert Williger
around the current separator in order to find a new locally
optimal solution. In this section, we show how we can use
the flow-based local improvement method described above
to locally improve a k-way separator for k > 2 and describe
algorithms that are able to balance solutions, e.g. solutions
that contain blocks with too many vertices. These algorithms
are used to create the initial population of our evolutionary
algorithm as well as to provide the combine and mutation
operations. Our algorithm is called Adv and the evolutionary
algorithm that is introduced later AdvEvo.
3.1
k-way Local Search
Our k-way local search builds on top of the flow-based search
which is intended for improving a separator with k = 2. The
main idea is to find pairs of adjoint blocks and then perform
local search on the subgraph induced by adjoint block pairs.
Preprocessing. In order to find pairs of adjoint blocks, we
look at separator nodes which directly separate two different blocks, meaning these separator nodes are adjacent to
nodes from at least two different blocks not including the
separator. In general directly separating nodes do not have
to exist (see Figure 1). In other words, it may be that a
separator disconnects two blocks, but the shortest path distance between the blocks is greater or equal to two. Using
a preprocessing step, we first make sure that each separator
node is adjacent to at least two blocks, i.e. each separator
node is directly separating.
The preprocessing step works as follows: we iterate over all
separator nodes and try to remove them from the separator
if they do not directly separate two blocks. The order in
which we look at the separator nodes is given by the number of adjacent blocks (highest first). Let s be the current
separator node under consideration. If it has two or more
non-separator neighbors in different blocks, it already directly
separates at least two blocks and we continue. If s only has
neighbors in a single block in addition to the separator, we
move it into that block. Lastly, if s only has other separator
nodes as neighbors, we put it into a block having smallest
overall weight. In each step, we update the priorities of
adjacent neighboring separator nodes. Note that nodes are
only removed from the separator and never added. Moreover,
removing a node from the separator can increase the priority
Figure 1: Illustration of the preprocessing for a
graph with blocks V1 (green), V2 (blue) and separator
S (red).
of an adjacent separator node only by one. As soon as the
priority of a node is larger than one, it is directly separating
and we do not have to look at the vertex again. After the
algorithm is done, each separator node is directly separating
at least two blocks and we can build the quotient graph in
order to find adjoint blocks. Our preprocessing can introduce
imbalance to the solution. Hence, we run the balance routine
defined below after preprocessing.
Pair-wise Local Search. Subsequent to the preprocessing
step, we identify the set of all adjoint block pairs P by
iterating through all separator nodes and their adjacent nodes.
We iterate through all pairs p = (A, B) ∈ P and build the
subgraph Gp . Gp is induced by the set of nodes consisting of all
nodes in A and B as well as all separator nodes that directly
separate the blocks. After building Gp , we run local search
designed for 2-way separators on this subgraph. Note that an
improvement in the separator size between A and B directly
corresponds to an improvement to the overall separator size.
To gain even smaller separators and because the solution
is potentially modified by local search, we repeat local search
multiple times in the following way. The algorithm is organized in rounds. In each round, we iterate over the elements
in P and perform local search on each induced subgraph. If
local search has not been successful, we remove p from P .
Otherwise, we keep p for the next round.
3.2
Balancing
To guarantee the balance constraint, we use a balance operation. Given an imbalanced separator of a graph, the algorithm
returns a balanced node separator. Roughly speaking, we
move nodes from imbalanced blocks towards the blocks that
can take nodes without becoming overloaded (see Figure 2
for an illustrating example). As long as there are imbalanced
blocks, we iteratively repeat the following steps:
First, we find a path p from the heaviest block H to
the lightest block L in the quotient graph. If there is no
such path, we directly move a node to the lightest block
and make its neighbors separator nodes. Next, we iterate
through the edges (A, B) ∈ p and move nodes from A into
B. In general, we move min(Lmax − |L|, |H| − Lmax ) nodes
along the path, i.e. as many nodes as the lightest block can
take without getting overloaded and as little nodes necessary
so that the heaviest block is balanced. Moving nodes is
based on gain of the movement as defined in Section 2.3.
Basically, we use a priority queue of separator nodes that
directly separate A and B with the key being set to the
gain. Note that these movements create new separator nodes
and can potentially worsen solution quality. We use the
gain definition because our primary target is to minimize the
increase of the separator size.
Then we dequeue nodes from the our priority queue until
A is balanced. We move each dequeued node s to B and
move its neighbors being in A into the separator and the
priority queue. Also the priorities of the nodes in the queue
are updated. After moving the nodes A will be balanced. If
B is imbalanced, we continue with the next pair in the path,
Distributed Evolutionary k-way Node Separators
Figure 2: Illustration of the quotient graph and
the balancing path (blue) from an imbalanced block
(red) to the lightest block (green).
i.e. sending the same amount of nodes. If B is also balanced,
we are done with this path and do not move any more nodes.
Our algorithm continues with the next imbalanced block.
4
EVOLUTIONARY K-WAY NODE
SEPARATORS
Our EA starts with a population of individuals (in our case
the node separator of the graph) and evolves the population into different populations over several rounds. In each
round, the EA uses a selection rule based on the fitness of the
individuals (in our case the size of the separator) of the population to select good individuals and combine them to obtain
improved offspring. Note that we can use the size/weight of
the separator as a fitness function since our algorithm always
generates separators fulfilling the given balance constraint,
i.e. there is no need to use a penalty function to ensure that
the final separator is feasible. When an offspring is generated
an eviction rule is used to select a member of the population
and replace it with the new offspring. In general one has to
take both into consideration, the fitness of an individual and
the distance between individuals in the population [1]. Our
algorithm generates only one offspring per generation.
4.1
Combine Operation
We now describe the combine operator. Our combine operator assures that the offspring has an objective at least as good
as the best of both parents. Roughly speaking, the combine operator combines an individual/separator P = V1P , ..., VkP , S P
(which has to fulfill a balance constraint) with a second individual/separator C = V1C , ..., VkC , S C . Let P be the individual
with better fitness.
The algorithm begins with selecting two individuals from
the population. The selection process is based on the tournament selection rule [22], i.e. P is the fittest out of two
random individuals R1 , R2 from the population and the same
is done to select C. Both node separators are used as input
for our multi-level algorithm in the following sense. Let E
be the set of edges that are cut edges, i.e. edges that run
between blocks and the separator, in either P or C. All edges
in E are blocked during the coarsening phase, i.e. they are
not contracted during the coarsening phase. In other words
these edges are not eligible for the matching algorithm used
during the coarsening phase and therefore are not part of
any matching.
The stopping criterion of the multi-level algorithm is modified such that it stops when no contractable edge is left.
As soon as the coarsening phase is stopped, we apply the
separator P to the coarsest graph and use this as initial
separator. This is possible since we did not contract any
edge running between the blocks and the separator in P.
Note that due to the specialized coarsening phase and this
specialized initial phase we obtain a high quality initial solution on a very coarse graph which is usually not discovered
by conventional algorithms that compute an initial solution.
Since our local search algorithms guarantee no worsening
of the input solution and we use random tie breaking we
can assure non-decreasing quality. Note that the local search
algorithms can effectively exchange good parts of the solution
on the coarse levels by moving only a few vertices.
Also note that this combine operator can be extended to
be a multi-point combine operator, i.e. the operator would
use ` instead of two parents. However, during the course
of the algorithm a sequence of two point combine steps is
executed which somehow "emulates" a multi-point combine
step. Therefore, we restrict ourselves to the case ` = 2.
When the offspring is generated we have to decide which
solution should be evicted from the current population. We
evict the solution that is most similar to the offspring among
those individuals in the population that have an objective
worse or equal than the offspring itself. Here, we define
the similarity σ of two node separators S1 and S2 as the
cardinality of the symmetric difference of both separators:
σ = |(S1 \ S2 ) ∪ (S2 \ S1 )|. Therefore σ denotes the number
of nodes contained in one separator but not in the other. This
ensures some diversity in the population and hence makes
the evolutionary algorithm more effective.
4.2
Mutation Operation
The mutation operation works similar to the combine operation. The main difference is that there is only one input
individual to the multi-level algorithm and that the offspring
can be less fit compared to the input individual. Hence, only
edges that run between the blocks and the separator of that
individual are not eligible for the matching algorithm. This
way the input individual can be transferred downwards in
the hierarchy. Additionally, the solution is not used as initial
separator but the initial algorithm is performed to find an
initial separator. Note however due to the way the coarsening
process is defined the input separator is still contained in the
coarsest graph.
5
PUTTING THINGS TOGETHER AND
PARALLELIZATION
We now explain parallelization and describe how everything
is put together to be our full evolutionary algorithm AdvEvo.
We use a parallelization scheme that has been successfully
used in graph partitioning [25]. Each processing element
Peter Sanders, Christian Schulz, Darren Strash, and Robert Williger
(PE) basically performs the same operations using different
random seeds (see Algorithm 1). First we estimate the population size S: each PE creates an individuum and measures
the time t spend. We then choose S such that the time for
creating S node separators is approximately ttotal /f where
the fraction f is a tuning parameter and ttotal is the total
running time that the algorithm is given to produce a node
separator of the graph. Each PE then builds its own population, i.e. our multi-level algorithm is called several times to
create S individuals/separators. Afterwards the algorithm
proceeds in rounds as long as time is left. With corresponding
probabilities, mutation or combine operations are performed
and the new offspring is inserted into the population.
We choose a parallelization/communication protocol that
is quite similar to randomized rumor spreading [9] which
has shown to be scalable in an evolutionary algorithm for
graph partitioning [25]. We follow their description closely.
Let p denote the number of PEs used. A communication
step is organized in rounds. In each round, a PE chooses
a communication partner and sends her the currently best
node separator P of the local population. The selection of
the communication partner is done uniformly at random
among those PEs to which P not already has been send to.
Afterwards, a PE checks if there are incoming individuals
and if so inserts them into the local population using the
eviction strategy described above. If P is improved, all PEs
are again eligible. This is repeated log p times. The algorithm
is implemented completely asynchronously, i.e. there is no
need for a global synchronization.
approach, but keep the rest of the algorithm. More precisely,
this means that the population of the algorithm are still graph
partitions instead of separators, but for example selection is
based on the size of the derivable separator. Additionally,
the combine operations in KaFFPaE still optimize for cuts
instead of separators. This algorithm is called SimpleEvo.
6
EXPERIMENTS
Methodology. We have implemented the algorithm described
above within the KaHIP framework using C++ and compiled all algorithms using gcc 4.8.3 with full optimization’s
turned on (-O3 flag).
We integrated our algorithms in
KaHIP v1.00. Our new codes will also be included into the
KaHIP graph partitioning framework. Each run was made
on a machine that has four Octa-Core Intel Xeon E5-2670
processors running at 2.6 GHz with 64 GB local memory.
Our main objective is the cardinality of node separators on
the input graph. In our experiments, we use the imbalance
parameter = 3% since this is one of the default values in
the Metis graph partitioning framework. Our full algorithm
is not too sensitive about the precise choice with most of the
parameters. However, we performed a number of experiments
to evaluate the influence and choose the parameters of our
algorithms. We omit details here and refer the reader to [30].
We mark the instances that have been used for the parameter
tuning in Table 1 with a * and exclude these graphs from
our experiments.
We present multiple views on the data: average values (geometric mean) as well as convergence plots that show quality
achieved by the algorithms over time and performance plots.
Algorithm 1 locallyEvolve
estimate population size S
while time left
if elapsed time < ttotal /f then
create individual, insert into local population
else
flip coin c with corresponding probabilities
if c shows head then perform mutation
else perform combine
insert offspring into population if possible
communicate according to comm. protocol
5.1
Miscellaneous
Besides Adv and AdvEvo, we also use two more algorithms
to compare solution quality. The first one is a sequential
algorithm that starts by computing a k-way partition using
KaFFPa-Strong and derives a k-way separator by pair-wise
decoupling by using the method of Pothen and Fan [24] on
each adjacent pair of blocks. The main idea of Pothen and
Fan is to compute a minimum vertex cover in the bipartite
subgraph induced by the set of cut edges between two pairs
of blocks. The union of the computed separators nodes is
a k-way separator. In our experiments, the algorithm is
called Simple. The second algorithm, is a modification of
KaFFPaE [25] which is an evolutionary algorithm to compute
graph partitions. We modify the fitness function to be the
size of the separator that can be derived using the Simple
Graph
n
Walshaw Graphs
add20
2 395
data
2 851
3elt
4 720
uk
4 824
add32
4 960
bcsstk33*
8 738
whitaker3
9 800
crack
10 240
wing_nodal 10 937
fe_4elt2
11 143
vibrobox*
12 328
bcsstk29
13 992
4elt
15 606
fe_sphere
16 386
cti
16 840
memplus
17 758
cs4
22 499
bcsstk30
28 924
bcsstk31
35 588
fe_pwt
36 519
Graph
n
Walshaw Graphs
bcsstk32
44 609
fe_body*
45 087
t60k
60 005
wing
62 032
brack2
62 631
finan512
74 752
fe_tooth*
78 136
fe_rotor
99 617
UF Graphs
cop20k_A
99 843
2cubes_sphere*
101 492
thermomech_TC 102 158
cfd2
123 440
boneS01
127 224
Dubcova3
146 689
bmwcra_1
148 770
G2_circuit*
150 102
c-73
169 422
shipsec5
179 860
cont-300
180 895
Table 1: Walshaw graphs and florida sparse matrix
graphs from [26]. Basic properties of the instances.
Graphs with a * have been used for parameter tuning and are excluded from the evaluation.
2100
k=32
SimReps
SimEvo
1900
AdvReps
AdvEvo
2
10 50
500
normalized time tn
440
410
380
k=16
SimReps
SimEvo
1200
1300
AdvReps
AdvEvo
1100
660
5
50 500 10000
normalized time tn
SimReps
SimEvo
5
50 500 10000
normalized time tn
2
10 50
500 5000
normalized time tn
k=64
3200
SimReps
SimEvo
720
780
AdvReps
AdvEvo
AdvReps
AdvEvo
AdvReps
AdvEvo
SimReps
SimEvo
2900
k=8
k=4
2600
5
50 500 10000
normalized time tn
mean min separator size
SimReps
SimEvo
mean min separator size
AdvReps
AdvEvo
mean min separator size
160 165 170 175
k=2
1700
mean min separator size
mean min separator size
mean min separator size
Distributed Evolutionary k-way Node Separators
2
10 50
500
normalized time tn
Figure 3: Convergence plots for different values of k for different algorithms.
Peter Sanders, Christian Schulz, Darren Strash, and Robert Williger
We now explain how we compute the convergence plots. We
start explaining how we compute them for a single instance I:
whenever a PE creates a separator it reports a pair (t, separator size), where the timestamp t is the currently elapsed time
on the particular PE and separator size refers to the size of
the separator that has been created. When performing multiple repetitions, we report average values (t, avg. separator
size) instead. After the completion of algorithm we are left
with P sequences of pairs (t, separator size) which we now
merge into one sequence. The merged sequence is sorted by
the timestamp t. The resulting sequence is called T I . Since
we are interested in the evolution of the solution quality, we
I
compute another sequence Tmin
. For each entry (in sorted
order) in T I , we insert the entry (t, mint0 ≤t separator size(t0 ))
I
into Tmin
. Here, mint0 ≤t separator size(t0 ) is the minimum
I
separator size that occurred until time t. Nmin
refers to the
normalized sequence, i.e. each entry (t, separator size) in
I
Tmin
is replaced by (tn , separator size) where tn = t/tI and
tI is the average time that the sequential algorithm needs to
compute a separator for the instance I. To obtain average
values over multiple instances we do the following: for each
I
instance we label all entries in Nmin
, i.e. (tn , separator size)
is replaced by (tn , separator size, I). We then merge all
I
sequences Nmin
and sort by tn . The resulting sequence is
called S. The final sequence Sg presents event based geometric averages values. We start by computing the geometric
I
mean value G using the first value of all Nmin
(over I). To
obtain Sg , we basically sweep through S: for each entry (in
sorted order) (tn , c, I) in S, we update G, i.e. the separator
size of I that took part in the computation of G is replaced
by the new value c, and insert (tn , G) into Sg . Note, c can
be only smaller than or equal to the old value of I.
Instances. We use the small and Florida Sparse Matrix
graphs from [26] which are from various sources to test our
algorithm. Small graphs have been obtained from Chris
Walshaw’s benchmark archive [29]. Graphs derived from
sparse matrices have been taken from the Florida Sparse
Matrix Collection [6]. Basic properties of the instances can
be found in Table 1.
6.1
Separator Quality
In this section we compare our algorithms in a setting where
each one gets the same (fairly large) amount of time to
compute a separator. We do this on the graphs from our
benchmark set. We use all 16 cores per run of our machine
(basically one node of the cluster) and two hours of time
per instance when we use the evolutionary algorithm to
create separators. We parallelized repeated executions of
the sequential algorithms (embarrassingly parallel, different
seeds) and also gave them 16 PEs and two hours of time
to compute a separator. We look at k ∈ {2, 4, 8, 16, 32, 64}
and performed three repetitions per instance. To see how
the solution quality of the different algorithms evolves over
time, we use convergence plots. Figure 3 shows convergence
plots for k ∈ {2, 4, 8, 16, 32, 64}. Additionally, Tables 2 and
k
2
4
8
16
32
64
overall
AdvEvo
159.4
373.9
664.4
1097.9
1694.2
2601.8
+0%
AdvReps
0.0
+1.7%
+2.7%
+5.1%
+6.8%
+8.1%
4.1%
SimEvo
+3.8%
+5.6%
+8.4%
+10.4%
+11.8%
+15.3%
9.2%
SimReps
+5.0%
+7.7%
+11.1%
+13.4%
+14.4%
+17.8%
11.6%
Table 2: Average of AdvEvo and average increase in
separator size for different algorithms.
Algorithm
AdvEvo
AdvReps
SimEvo
SimReps
#≤
181
65
33
29
#<
122
7
3
0
Table 3: Number of instances where algorithm X is
best w.r.t to ≤ and <. The total number of instances
is 192.
3 summarize final results. Tables in the Appendix present
detailed per instance results.
First of all, the improvements of the evolutionary algorithms relative to repeated executions increase with growing
k. This is due to the fact that the problems become more difficult when increasing the number of blocks k. For larger values
of k, the quality gap between the evolutionary algorithm AdvEvo and SimEvo as well as the other algorithms increases
with more time invested. On the other hand, for k = 2 there
is almost no difference between the results produced by the
evolutionary algorithm AdvEvo and the non-evolutionary
version AdvReps. Overall, the experimental data indicates
the AdvEvo is the best algorithm. Separators produced by
AdvEvo are 4.1%, 9.2% and 11.6% smaller compared to AdvReps, SimEvo, and SimReps on average. Additionally, our
advanced evolutionary algorithm computes the best result
on 181 out of 192 instances.
Note that single executions of the simple algorithms are
much faster. However, the results of our experiments performed in this section emphasize that one cannot simply use
the best result out of multiple repetitions of a faster algorithm
to obtain the same solution quality. Yet it is interesting to
see that SimpleEvo, where only the fitness function of the
evolutionary algorithm is modified and the combine operation
still optimizes for edge cuts of partitions, computes better
solutions than its non-evolutionary counter part SimReps.
7
CONCLUSION
In this work, we derived a new approach to find small node
separators in large graphs which combines an evolutionary
search algorithm with a multilevel method. We combined
these techniques with a scalable communication protocol
and obtain a system that is able to compute high quality
solutions in a short amount of time. Experiments show that
our advanced evolutionary algorithm computes the best result
on 94% of the benchmark instances. In future work, we aim
to look at different types of applications, in particular those
applications in which the running time may not be considered
Distributed Evolutionary k-way Node Separators
a drawback when the node separator has the highest quality.
Acknowledgements. The authors acknowledge support by
the state of Baden-Württemberg through bwHPC.
REFERENCES
[1] T. Bäck. 1996. Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary Programming, Genetic
Algorithms. Ph.D. Dissertation.
[2] S. N. Bhatt and F. T. Leighton. 1984. A framework for solving
VLSI graph layout problems. J. Comput. System Sci. 28, 2 (1984),
300 – 343. DOI:http://dx.doi.org/10.1016/0022-0000(84)90071-0
[3] C. Bichot and P. Siarry (Eds.). 2011. Graph Partitioning. Wiley.
[4] T. N. Bui and C. Jones. 1992. Finding Good Approximate Vertex
and Edge Partitions is NP-hard. Inform. Process. Lett. 42, 3
(1992), 153–159.
[5] A. Buluç, H. Meyerhenke, I. Safro, P. Sanders, and C. Schulz.
2016. Recent Advances in Graph Partitioning. In Algorithm
Engineering – Selected Results (LNCS), Vol. 9920. Springer,
117–158.
[6] T. Davis. 2017. The University of Florida Sparse Matrix Collection. (2017).
[7] D. Delling, M. Holzer, K. Müller, F. Schulz, and D. Wagner.
2009. High-performance multi-level routing. The Shortest Path
Problem: Ninth DIMACS Implementation Challenge 74 (2009),
73–92.
[8] J. Dibbelt, B. Strasser, and D. Wagner. 2014. Customizable
contraction hierarchies. In 13th Int. Symp. on Exp. Algorithms
(SEA’14). Springer, 271–282.
[9] B. Doerr and M. Fouz. 2011. Asymptotically Optimal Randomized Rumor Spreading. In Proceedings of the 38th International
Colloquium on Automata, Languages and Programming, Proceedings, Part II (LNCS), Vol. 6756. Springer, 502–513.
[10] G. N. Federickson. 1987. Fast Algorithms for Shortest Paths
in Planar Graphs, with Applications. SIAM J. Comput. 16,
6 (1987), 1004–1022. DOI:http://dx.doi.org/10.1137/0216064
arXiv:http://dx.doi.org/10.1137/0216064
[11] J. Fukuyama. 2006. NP-Completeness of the Planar Separator
Problems. Journal of Graph Algorithms and Applications 10, 2
(2006), 317–328.
[12] M. R. Garey and D. S. Johnson. 2002. Computers and Intractability. Vol. 29. WH Freeman & Co., San Francisco.
[13] A. George. 1973. Nested Dissection of a Regular Finite Element
Mesh. SIAM J. Numer. Anal. 10, 2 (1973), 345–363.
[14] W. W. Hager, J. T. Hungerford, and I. Safro. 2014. A Multilevel Bilinear Programming Algorithm For the Vertex Separator
Problem. Technical Report.
[15] M. Hamann and B. Strasser. 2016. Graph Bisection with ParetoOptimization. In Proc. of the Eighteenth Workshop on Algorithm
Engineering and Experiments, ALENEX’16. SIAM, 90–102. DOI:
http://dx.doi.org/10.1137/1.9781611974317.8
[16] G. Karypis and V. Kumar. 1998. A Fast and High Quality Multilevel Scheme for Partitioning Irregular Graphs. SIAM Journal
on Scientific Computing 20, 1 (1998), 359–392.
[17] D. LaSalle and G. Karypis. 2015. Efficient Nested Dissection for
Multicore Architectures. In Euro-Par 2015: Parallel Processing.
Springer, 467–478.
[18] C. E. Leiserson. 1980. Area-Efficient Graph Layouts. In 21st
Symposium on Foundations of Computer Science. IEEE, 270–
281.
[19] R. J. Lipton and R. E. Tarjan. 1979. A Separator Theorem for
Planar Graphs. SIAM J. Appl. Math. 36, 2 (1979), 177–189.
[20] R. J. Lipton and R. E. Tarjan. 1980. Applications of a Planar
Separator Theorem. SIAM Journal On Computing 9, 3 (1980),
615–627.
[21] J. Maue and P. Sanders. 2007. Engineering Algorithms for
Approximate Weighted Matching. In Proceedings of the 6th
Workshop on Experimental Algorithms (WEA’07) (LNCS),
Vol. 4525. Springer, 242–255.
http://dx.doi.org/10.1007/
978-3-540-72845-0_19
[22] B. L Miller and D. E Goldberg. 1996. Genetic Algorithms, Tournament Selection, and the Effects of Noise. Evolutionary Computation 4, 2 (1996), 113–131.
[23] F. Pellegrini. 2017. Scotch Home Page. (2017). http://www.labri.
fr/pelegrin/scotch.
[24] A. Pothen, H. D. Simon, and K. P. Liou. 1990. Partitioning
Sparse Matrices with Eigenvectors of Graphs. SIAM J. Matrix
Anal. Appl. 11, 3 (1990), 430–452.
[25] P. Sanders and C. Schulz. 2012. Distributed Evolutionary Graph
Partitioning. In Proc. of the 12th Workshop on Algorithm Engineering and Experimentation (ALENEX’12). 16–29.
[26] P. Sanders and C. Schulz. 2016. Advanced Multilevel Node Separator Algorithms. In 15th International Symposium on Experimental Algorithms, SEA 2016, Vol. 9685. Springer, 294–309.
[27] C. Schulz. 2013. High Quality Graph Partititioning. Ph.D.
Dissertation. Karlsruhe Institute of Technology.
[28] F. Schulz, D. Wagner, and C. Zaroliagis. 2002. Using multi-level
graphs for timetable information in railway systems. In Proceedings of Algorithm Engineering and Experiments (ALENEX).
Springer, 43–59.
[29] A. J. Soper, C. Walshaw, and M. Cross. 2004. A Combined
Evolutionary Search and Multilevel Optimisation Approach to
Graph-Partitioning. Journal of Global Optimization 29, 2 (2004),
225–241.
[30] R. Williger. 2016. Evolutionary k-way Node Separators. Bachelor’s Thesis. Karlsruhe Institute of Technologie.
Peter Sanders, Christian Schulz, Darren Strash, and Robert Williger
graph
3elt
3elt
3elt
3elt
3elt
3elt
4elt
4elt
4elt
4elt
4elt
4elt
add20
add20
add20
add20
add20
add20
add32
add32
add32
add32
add32
add32
bcsstk29
bcsstk29
bcsstk29
bcsstk29
bcsstk29
bcsstk29
bcsstk30
bcsstk30
bcsstk30
bcsstk30
bcsstk30
bcsstk30
bcsstk31
bcsstk31
bcsstk31
bcsstk31
bcsstk31
bcsstk31
bcsstk32
bcsstk32
bcsstk32
bcsstk32
bcsstk32
bcsstk32
bmwcra_1
bmwcra_1
bmwcra_1
bmwcra_1
bmwcra_1
bmwcra_1
k
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
AdvEvo
AdvReps
SimpleEvo
min
avg
min
avg
min
avg
43
43
43
43
43
43
97
97
97
97
97
97
158
158
159
159
159
159
269
269
270
270
271
271
452
452
455
456
458
459
706
710
720
721
733
734
68
68
68
68
68
68
157
157
157
157
157
157
253
253
254
254
256
256
438
439
443
445
443
445
737
743
755
757
744
746
1 221
1 227
1 257
1 259
1 234
1 238
26
26
25
25
28
28
37
37
37
37
45
47
67
67
69
70
82
86
95
96
106
108
116
119
110
111
140
140
166
169
138
138
161
164
219
225
2
2
2
2
2
2
6
6
6
6
6
6
11
11
11
11
12
12
20
20
20
20
20
20
33
33
33
33
33
33
114
115
118
119
127
128
180
180
180
180
180
180
528
528
528
528
534
536
948
954
966
968
1 173
1 202
1 512
1 530
1 578
1 578
2 120
2 127
2 231
2 250
2 316
2 326
2 891
2 898
3 130
3 157
3 361
3 371
4 073
4 089
206
206
206
206
206
206
549
549
549
549
573
573
1 121
1 121
1 123
1 123
1 138
1 145
2 128
2 146
2 183
2 201
2 446
2 455
3 195
3 249
3 321
3 338
3 985
3 987
4 709
4 836
5 045
5 111
5 846
5 940
308
308
308
308
317
321
767
767
767
767
802
803
1 433
1 434
1 441
1 442
1 529
1 538
2 353
2 399
2 437
2 446
2 592
2 630
3 635
3 695
3 837
3 874
4 338
4 361
5 102
5 167
5 323
5 394
6 090
6 136
297
297
297
297
322
328
569
569
569
569
633
637
1 145
1 152
1 177
1 183
1 312
1 336
2 080
2 102
2 122
2 131
2 391
2 434
3 422
3 449
3 498
3 524
4 102
4 118
5 386
5 469
5 621
5 677
6 293
6 321
657
657
657
657
684
684
1 668
1 673
1 656
1 659
1 683
1 685
3 918
3 970
4 002
4 013
4 080
4 126
10 011 10 037
9 846
9 921 10 099 10 199
16 089 16 300 16 798 16 979 16 863 16 947
23 586 24 279 25 707 25 809 24 979 25 087
Table 4: Detailed per instance results.
SimpleReps
min
avg
43
43
97
97
160
160
272
272
466
466
746
747
68
68
157
157
254
255
442
444
749
750
1 253
1 253
28
28
49
49
94
94
133
133
170
174
218
221
2
2
6
6
12
12
20
20
33
33
131
131
180
180
534
534
1 086
1 094
2 019
2 062
2 892
2 899
4 065
4 096
206
206
573
573
1 138
1 140
2 430
2 452
3 956
4 008
5 994
6 013
317
319
798
800
1 534
1 545
2 624
2 633
4 361
4 376
6 169
6 205
321
321
627
627
1 315
1 326
2 443
2 468
4 114
4 142
6 348
6 412
683
683
1 659
1 666
4 112
4 133
10 190 10 250
16 725 16 775
24 885 25 042
Distributed Evolutionary k-way Node Separators
graph
boneS01
boneS01
boneS01
boneS01
boneS01
boneS01
brack2
brack2
brack2
brack2
brack2
brack2
cfd2
cfd2
cfd2
cfd2
cfd2
cfd2
cont-300
cont-300
cont-300
cont-300
cont-300
cont-300
cop20k_A
cop20k_A
cop20k_A
cop20k_A
cop20k_A
cop20k_A
crack
crack
crack
crack
crack
crack
cs4
cs4
cs4
cs4
cs4
cs4
cti
cti
cti
cti
cti
cti
data
data
data
data
data
data
k
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
AdvEvo
AdvReps
SimpleEvo
min
avg
min
avg
min
avg
1 524
1 524
1 524
1 524
1 563
1 563
3 357
3 357
3 357
3 357
3 465
3 469
5 112
5 128
5 139
5 148
5 316
5 351
7 728
7 781
7 776
7 826
8 139
8 179
11 082 11 127 11 400 11 440 11 619 11 670
15 264 15 271 16 495 16 683 16 173 16 219
183
183
183
183
206
208
796
796
796
796
829
831
1 740
1 741
1 759
1 761
1 896
1 899
2 794
2 806
2 868
2 874
3 108
3 119
4 160
4 184
4 282
4 289
4 632
4 647
6 053
6 084
6 361
6 381
6 808
6 841
1 030
1 030
1 030
1 030
1 036
1 036
2 543
2 546
2 548
2 551
2 684
2 688
4 304
4 313
4 304
4 312
4 569
4 581
7 068
7 095
7 018
7 036
7 416
7 440
10 723 10 873 11 165 11 272 12 066 12 088
16 521 17 138 17 829 18 021 18 093 18 179
598
598
598
598
598
598
1 041
1 042
1 063
1 065
1 184
1 184
1 786
1 814
1 807
1 825
2 188
2 192
2 863
2 874
2 893
2 916
3 526
3 528
4 299
4 340
4 413
4 450
5 466
5 483
6 452
6 474
6 667
6 679
8 094
8 105
620
620
620
620
620
620
1 673
1 675
1 676
1 676
1 733
1 741
2 919
2 934
2 939
2 942
2 997
2 999
4 721
4 744
4 765
4 780
4 842
4 864
7 241
7 333
7 525
7 645
7 481
7 502
10 757 10 837 11 721 12 107 11 135 11 147
73
73
73
73
75
75
145
145
145
145
152
152
257
257
258
258
280
282
420
420
421
422
461
465
636
639
648
648
722
724
939
943
958
959
1 098
1 100
287
287
287
287
319
319
727
729
738
740
824
825
1 108
1 109
1 133
1 134
1 244
1 245
1 548
1 558
1 623
1 630
1 812
1 818
2 132
2 148
2 273
2 286
2 512
2 516
2 864
2 909
3 179
3 202
3 433
3 435
266
266
266
266
266
266
756
758
761
761
808
808
1 243
1 270
1 311
1 315
1 537
1 539
1 821
1 845
1 897
1 925
2 287
2 298
2 426
2 457
2 646
2 647
3 257
3 267
3 234
3 242
3 581
3 596
4 489
4 491
51
51
51
51
51
51
96
96
96
96
96
96
165
165
165
165
172
172
275
275
276
276
298
301
448
450
454
454
489
492
688
690
695
697
757
760
Table 5: Detailed per instance results.
SimpleReps
min
avg
1 563
1 565
3 471
3 477
5 358
5 371
8 142
8 189
11 758 11 770
16 296 16 314
205
205
829
829
1 906
1 910
3 117
3 120
4 646
4 651
6 809
6 820
1 036
1 036
2 645
2 667
4 551
4 591
7 374
7 392
11 924 11 956
18 014 18 067
598
598
1 184
1 184
2 188
2 194
3 534
3 537
5 504
5 506
8 168
8 170
620
620
1 716
1 724
2 993
2 996
4 849
4 858
7 423
7 465
11 102 11 155
74
74
152
152
285
286
474
474
735
738
1 123
1 124
322
322
832
836
1 267
1 267
1 827
1 830
2 537
2 541
3 461
3 463
266
266
807
807
1 539
1 539
2 319
2 324
3 258
3 267
4 481
4 491
51
51
97
97
173
173
300
300
499
502
764
764
Peter Sanders, Christian Schulz, Darren Strash, and Robert Williger
graph
Dubcova3
Dubcova3
Dubcova3
Dubcova3
Dubcova3
Dubcova3
fe_4elt2
fe_4elt2
fe_4elt2
fe_4elt2
fe_4elt2
fe_4elt2
fe_pwt
fe_pwt
fe_pwt
fe_pwt
fe_pwt
fe_pwt
fe_rotor
fe_rotor
fe_rotor
fe_rotor
fe_rotor
fe_rotor
fe_sphere
fe_sphere
fe_sphere
fe_sphere
fe_sphere
fe_sphere
finan512
finan512
finan512
finan512
finan512
finan512
memplus
memplus
memplus
memplus
memplus
memplus
shipsec5
shipsec5
shipsec5
shipsec5
shipsec5
shipsec5
t60k
t60k
t60k
t60k
t60k
t60k
k
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
2
4
8
16
32
64
AdvEvo
AdvReps
SimpleEvo
min
avg
min
avg
min
avg
383
383
383
383
383
383
765
765
765
765
765
765
1 433
1 437
1 436
1 437
1 463
1 475
2 295
2 309
2 319
2 322
2 373
2 398
3 581
3 621
3 707
3 711
3 887
3 893
5 448
5 492
5 706
5 732
5 754
5 772
66
66
66
66
66
66
163
163
162
162
168
169
283
285
288
288
290
291
478
479
482
483
482
484
759
762
780
781
773
774
1 149
1 153
1 185
1 190
1 189
1 191
116
116
116
116
116
116
236
236
236
236
236
236
473
473
473
473
474
474
925
925
929
929
930
930
1 834
1 839
1 909
1 910
1 862
1 864
2 846
2 919
3 025
3 043
3 458
3 472
460
460
460
460
464
464
1 540
1 545
1 543
1 554
1 575
1 593
2 833
2 838
2 844
2 848
2 891
2 891
4 404
4 448
4 483
4 489
4 605
4 632
6 809
6 898
6 943
7 013
7 024
7 065
10 196 10 289 10 534 10 615 10 249 10 293
192
192
192
192
192
192
379
379
380
380
380
380
570
570
575
577
570
570
804
804
852
854
835
835
1 177
1 184
1 262
1 264
1 208
1 213
1 657
1 667
1 803
1 805
1 722
1 723
50
50
50
50
50
50
100
100
100
100
100
100
200
200
200
200
200
200
400
400
400
400
400
400
800
800
800
800
800
800
3 210
3 216
3 259
3 263
3 200
3 200
70
70
70
70
90
103
90
90
91
91
127
131
106
106
106
106
154
158
132
132
151
153
234
240
178
179
194
195
265
268
181
182
205
206
392
412
1 203
1 203
1 203
1 203
1 227
1 231
3 681
3 681
3 681
3 681
3 783
3 793
6 078
6 112
6 198
6 216
6 486
6 509
8 826
8 881
8 850
8 903
9 612
9 650
12 861 12 983 13 521 13 601 14 208 14 236
17 304 17 398 18 114 18 301 21 482 21 594
70
70
70
70
71
71
202
202
202
202
203
203
447
447
449
449
448
448
800
803
807
810
792
793
1 330
1 331
1 339
1 343
1 307
1 308
2 040
2 042
2 104
2 114
2 031
2 043
Table 6: Detailed per instance results.
SimpleReps
min
avg
383
383
765
765
1 463
1 464
2 346
2 355
3 821
3 856
5 763
5 766
66
66
167
167
293
293
486
487
783
785
1 206
1 208
116
116
236
236
476
476
929
929
1 872
1 873
3 447
3 462
464
464
1 573
1 580
2 898
2 899
4 550
4 589
7 037
7 062
10 242 10 261
192
192
380
380
570
570
839
839
1 216
1 218
1 745
1 750
50
50
100
100
200
200
400
400
800
800
3 200
3 200
107
107
123
126
142
144
239
242
264
268
428
436
1 227
1 233
3 801
3 803
6 531
6 549
9 570
9 651
14 316 14 340
21 549 21 611
71
71
203
203
448
448
802
804
1 335
1 336
2 098
2 103
Distributed Evolutionary k-way Node Separators
graph
thermomech_TC
thermomech_TC
thermomech_TC
thermomech_TC
thermomech_TC
thermomech_TC
uk
uk
uk
uk
uk
uk
whitaker3
whitaker3
whitaker3
whitaker3
whitaker3
whitaker3
wing
wing
wing
wing
wing
wing
wing_nodal
wing_nodal
wing_nodal
wing_nodal
wing_nodal
wing_nodal
k
AdvEvo
AdvReps
min
avg
min
avg
2
128
128
128
128
4
390
390
389
389
8
837
839
838
838
16 1 439 1 440 1 447 1 450
32 2 316 2 323 2 345 2 349
64 3 620 3 627 3 672 3 680
2
18
18
18
18
4
37
37
37
37
8
73
73
73
73
16
133
133
134
134
32
223
224
225
226
64
360
362
364
366
2
63
63
63
63
4
186
186
187
187
8
320
320
321
321
16
526
528
531
531
32
803
803
811
811
64 1 183 1 185 1 198 1 199
2
626
626
628
628
4 1 288 1 290 1 296 1 297
8 1 943 1 949 1 977 1 980
16 2 963 2 995 3 085 3 095
32 4 300 4 315 4 492 4 513
64 5 736 5 778 6 267 6 287
2
410
410
410
410
4
811
814
836
837
8 1 177 1 184 1 211 1 216
16 1 793 1 800 1 841 1 850
32 2 494 2 500 2 602 2 625
64 3 243 3 252 3 460 3 467
Table 7: Detailed per instance
SimpleEvo
min
avg
129
129
394
394
843
844
1 445 1 447
2 340 2 347
3 636 3 645
18
18
38
38
77
77
136
136
233
236
387
388
63
63
189
189
321
321
531
532
812
813
1 213 1 216
700
702
1 432 1 433
2 157 2 170
3 378 3 388
4 906 4 917
6 622 6 644
415
415
858
859
1 286 1 287
1 965 1 974
2 791 2 800
3 683 3 685
results.
SimpleReps
min
avg
129
129
394
394
848
848
1 453 1 456
2 345 2 353
3 659 3 663
18
18
38
38
76
76
137
138
240
241
402
403
63
63
189
189
322
322
535
536
821
822
1 233 1 233
704
704
1 453 1 456
2 191 2 191
3 398 3 405
4 911 4 921
6 673 6 677
415
415
863
864
1 294 1 294
1 985 1 987
2 806 2 807
3 696 3 702
| 8cs.DS
|
arXiv:1507.01122v3 [cs.AI] 13 May 2016
Modeling the Mind: A brief review
Gabriel Makdah
[email protected]
Université Claude Bernard Lyon 1, France
I - Computational Foundations
Contents
1
Introduction
5
2
The foundations of computers
6
2.1
Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2.2
Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
3
4
5
6
7
Representations
7
3.1
Dimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
3.2
High-dimensional representations . . . . . . . . . . . . . . . . . . . . . .
8
3.3
High-dimensional memory . . . . . . . . . . . . . . . . . . . . . . . . . . 10
The brain and diversity
11
4.1
Mixed selectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2
Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.3
Randomness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Computationalism, and Connectionism
13
5.1
Computationalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.2
Connectionism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.3
The necessity of variable binding . . . . . . . . . . . . . . . . . . . . . . . 15
Artificial Neural Networks
17
6.1
Parallel Distributed Representations . . . . . . . . . . . . . . . . . . . . . 17
6.2
Restricted Boltzmann Machines . . . . . . . . . . . . . . . . . . . . . . . 18
6.3
Recurrent Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 19
6.4
Using an external stack memory . . . . . . . . . . . . . . . . . . . . . . . 20
6.5
Long Short-Term Memory . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.6
Long Term memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Vector Symbolic Architectures
26
2
8
9
7.1
Vector components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
7.2
Vector operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
7.3
Holographic Reduced Representations: . . . . . . . . . . . . . . . . . . . . 28
Working memory
32
8.1
Biological architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
8.2
Memory functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
8.3
Memory operations in the pre-frontal cortex and basal ganglia . . . . . . . 34
8.4
Modeling the PFC-BG system . . . . . . . . . . . . . . . . . . . . . . . . 36
Modeling the cortical architecture
38
9.1
Multiple layers of representation . . . . . . . . . . . . . . . . . . . . . . . 38
9.2
Training unlabeled data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
9.3
Parallel constraint satisfaction processes . . . . . . . . . . . . . . . . . . . 40
9.4
Part-whole hierarchies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
9.5
Hierarchical Temporal Memory . . . . . . . . . . . . . . . . . . . . . . . 42
10 Constructing the Code: A recap
43
3
Preface
Modeling the Mind: a brief review is an annual review, available for free on arXiv.org,
whose aim is to help students and researchers unfamiliar with the field of neuroscience and
computational neuroscience gain insight into the fundamentals of this domain of study.
Creating an accurate simulation of the mind is no easy task, and while it took brilliant
minds decades to advance us to where we’re at right now, we are still ways off our final goal.
It is therefore imperative to have more research carried out in this multidisciplinary field,
taking in help from researchers in biology, neuroscience, computer science, but also
mathematics, physics, chemistry and imaging, in order to speed up this process and tip the
scales in our favor for the upcoming decades. This annual review hopes to provide the
required information for anyone who is considering this domain as his future endeavor.
The reviews will be tackling relatively global characteristics at first in order to familiarize
the reader with the basic foundations, and will be getting progressively more specific and in
tune with current research in the upcoming parts. This is Part I. It will contain basic
information about the computational aspect of this field, and will attempt to explain why
certain concepts are generally agreed upon, and the intuition behind them, going through
the essential founding works.
4
Abstract
The brain is a powerful tool used to achieve amazing feats. There have been several
significant advances in neuroscience and artificial brain research in the past two decades.
This article is a review of such advances, ranging from the concepts of connectionism, to
neural network architectures and high-dimensional representations. There have also been
advances in biologically inspired cognitive architectures, of which we will cite a few. We
will be positioning relatively specific models in much broader perspectives, while also
comparing and contrasting their advantages and weaknesses. The projects presented are
targeted to model the brain at different levels, utilizing different methodologies.
1
Introduction
Computational neuroscience has now existed for several decades. It started off as an attempt
to model individual neurons’ firing mechanics, with the Hodgkin-Huxley model[30] in 1952,
as well as neurons’ learning models, with Hebbian learning[24] in 1949. Today, scientists
use computer simulations not only to model some of the brain’s traits, but also to predicate
some of its behavior, and study new and emerging aspects of it. The results produced using
simulations often lead to testable predictions in the real world.
Neural networks were created by researchers who wanted to model architectures that
were similar to that of the brain. From constructing nodes, to learning algorithms such as
backpropagation, the brain has greatly influenced neural network development. Ever since
their conception, neural networks exhibited features that were remarkable for computers.
To this day, we are still surprised by their expanding catalog of capabilities. They are used
in tasks such as speech recognition, image recognition, learning grammars and languages,
robotics, and a wide array of subjects, that require a human-like thought process to conduct.
All of this makes artificial neural networks the prime contenders to modeling the mind.
In the upcoming sections, we will discuss some of the theories behind neural networks,
their foundations, the most widely used algorithms, as well as some theories about their
application to the field of neuroscience. Section 3 will be dedicated to examining computer
patterns, and representations in order to take maximal advantage of the computable space.
The brain, and its whole specters of variety will be presented in section 4 in order to have a
better understanding of it, before progressing any further. The concepts of computationalism,
and connectionism, two theories for modeling the brain, will be discussed in section 5. In
section 6, we will present artificial neural networks, and talk about the most widely used
architectures and algorithms, as well as the advantages of each one of them. Section 7 will
go over vector symbolic architectures and some basic operations and models to optimize
storage. Section 8 will serve as an overview to the brain’s working memory, as well as the
modeling of the pre-frontal cortex’s memory system, a key component to working memory.
We will be talking about the main problems that artificial brain architectures should be
5
tackling, as well as give examples on how research has been aiming to solve them in section
9. The final section, section 10, will serve as a conclusion to the article. We will discuss what
an optimal model of the brain would look like following the concepts that were discussed in
this review.
But first, let us give an overview of what made digital computers possible, and talk
briefly about some of the basic principles behind them in section 2.
2
2.1
The foundations of computers
Hardware
The grounds of today’s digital computers lie in the Von Neumann architecture.
The Von Neumann[84] architecture is a computer model described 60 years ago,
whose success has made computers a ubiquitous part of our lives. The architecture depicts
an electronic digital computer with different parts. The first, and arguably the most important
of these, is the processing unit, which contains an Arithmetic Logic Unit, ALU for short,
as well as processor registers. The control unit houses instruction registers and a program
counter. It helps directs operations and instructions to the processor. Memory is used to
store data and instructions, either in rapidly accessible Random-Access Memory, RAM for
short, or slower, but more capacitive external mass storage, such as hard disk drives, and
more recently, solid state drives. The machine also needs input hardware in order to take
queries from the operator, equipment such as keyboards, mice, or pressure sensors are used.
In order to display the results of its computation, the machine expresses itself through output
hardware, such as a monitor, speakers, or a robotic arm.
Data, and the instructions that manipulate data, are entities of the same kind. They are
stored in the same set of addresses and data buses. The Von Neumann architecture has only
one memory space.
A lot of attempts have been made to program computers for the kind of flexible intelligence
that characterizes human behavior, however with no remarkable success. This has lead many
to wonder whether a different computing architecture is needed, and whether the problem
was stemming from a lack of software optimization, or an inherently limiting bottleneck in
the Von Neumann architecture. Several new architectures were presented throughout the
years, each tackling different issues. Some are found widely in use today, which prove that
the Von Neumann model is not the be all and end all of architectures.
An example of such models would be the Modified Harvard architecture. This model
has 2 separate memory spaces, compared to only one in the Von Neumann architecture. The
first memory space is used to read data from, and write data to the memory. The second
set of addresses and data buses is used to fetch instructions. This results in an increase of
throughput, seeing as the memory is no longer being limited by memory spaces. While the
system is executing one instruction, it could be fetching the next one at the same time.
6
In our case, the burden falls on representation, the kinds of entities that the computer uses in
order to process its computations. If our goal was to model the mind, then we ought to be
using similar representations to those used by the brain. Instead of using low-dimensional
binary vectors as conventional computers would do, the basic idea here is to compute with
large vector patterns, called high dimensional random vectors. We will be tackling this
subject next, but first, let us give a general definition of computing.
2.2
Computing
Computing is the transformation of patterns by algorithms which are governed by rules.
Patterns, also known as representations, are the configuration of ON and OFF switches in
a sequence. In other words, they are the sequencing of bits. An algorithm helps compute
a new pattern based on the previous strings of patterns. Patterns can be meaningful when
they correspond to either abstract entities, or things in the world, such as numbers, words,
images.
To transform patterns, computers have circuits, such as the adder circuit for summing
two numbers, the OR circuit, the AND circuit...etc, whose interpretations are the same as
their uses in ordinary language. The materials that computers and circuits are made of are
subordinate. That is to say they are not a major contributor to making the logic possible.
Materials can be changed as cost of production and development varies, leading to better
constituents. The logical design can be separated from the materials and the hardware. This
intelligent design can be seen not only in man-made computers, but in humans themselves.
Although we are all capable of comparable computing power, our underlying architectures
are very different. No two people have the same neural hardware, yet the logical design still
holds from one person to the other.
3
Representations
Computers use patterns formed by strings of 0s and 1s. These are called binary vectors,
and are used in binary representations. The goal of a certain representation is to be able
to discriminate between any two objects. In other words, the bit patterns representing two
different values must differ from one another.
The meanings which patterns hold are not apparent in their standalone binary vectors.
Instead, it’s the relations between different patterns that determines their possible use for
computing. This is also true for the brain. A standalone neuron being activated does not
yield much, if any, work done. It’s the complete structure of inter-relational patterns that
generate computation.
Choosing the correct representation is subject to trade-offs and compromises. There
are no perfect representations that would be suitable for all tasks. The base-2, binary, system
is extremely efficient in arithmetic operations and is therefore used in computers. Our brains
7
very probably use a different kind of representation, with different trade-offs, better suited
for vital functions, compromising the speed and precision of arithmetic tasks.
However, brain-like abilities and functions can still be demonstrated in computer
simulations using nothing but binary patterns and operations. This proves that, up to a
certain point, the precise nature of the individual units don’t matter as much as the properties
deriving from them through the specified operations.
Representations can be done through a variety of ways, depending on the type of tasks, and
the computer architecture that is at hand. Scalars are one-dimensional vectors. They are
represented using higher dimensional vectors in computers: the scalar 46260 is represented
using a 16bit integer, which is a 16-dimensional binary vector. In the next sections, we will
talk briefly about the basic features underlying the brain’s representation models, and then
generalize to what our model should incorporate.
3.1
Dimensionality
A huge number of neurons and synapses form the brain’s circuitry. Each neuron is connected
to tens of thousands of other neurons. There are hundreds of thousands of neurons that fire
each second. The amount of information that is being exchanged inside our brains, between
neurons, and between regions, is extremely large. To match such high content and rate of
exchange, we will shy away from scalars and low-dimensional vectors. The information in
the brain is encoded on larger elements than these. This leads us to explore the possibility of
high-dimensional vector use, vectors that are thousands of bits long[37].
A high-dimensional vector is represented in a high-dimensional space, which groups
all the vectors of one representation together. Just like a 2-dimensional vector is represented
in a 2 dimensional space, an N-dimensional vector is represented in an N- dimensional space.
High dimensional modeling started off several decades back under the forms of artificial
neural networks, parallel distributed processing and connectionism. We will be exploring all
these areas later in this review. First however, let us expand a bit on these high-dimensional
representations.
3.2
High-dimensional representations
The space of representations make up the sets of units with which a computer computes.
This space is low-dimensional in conventional computers, where the memory is usually
addressed in 8-bit bytes, and the ALU applies its arithmetic operations on 32-bit vectors.
The high-dimensional representational vectors and space aren’t necessarily binary. They
can be ternary, real, or complex. They can also be further modulated using probabilistic
distributions, limited ranges of values, and sparse modeling. Sparse modeling corresponds
to loosely coupled systems. Many components are either zero, or negligibly small. This
saves a large amounts of memory, and speeds up data processing, while maintaining the
8
advantages of high-dimensionality[35]. Before advancing any further, let us take an example
of a low-dimensional vector.
Let us consider a 3 bit string:
(1, 0, 1)
This string is also considered to b e a 3 − dimensional vector.
It can b e represented in a 3 − dimensional space.
Such a three-dimensional space can contain 23 = 8 patterns.
To showcase the difference in orders of magnitude between a low, and high-dimensional
representation, let us now take an example of a high-dimensional binary vector.
Let us consider a 20,000 bit string:
(1, 0, 0, 1, 0, 1, 1, 0, 0, 1...etc)
This string is also considered to b e a 20, 000 − dimensional vector.
It can b e represented in a 20, 000 − dimensional space.
A binary 20, 000 − dimensional space consists of:
220,000 = 3.98 ∗ 106020 such patterns.
Because of the discrete nature of vector representations, all vectors are equally spaced.
Distances can be measured between points using the Euclidean metric, or the Hamming
distance. The Hamming distance is the number of places at which two binary vectors differ.
Between two 20,000-dimensional vectors, the Hamming distance cannot exceed 20,000 bits.
This can be normalized to 1, in Hamming distance, by dividing it by the dimensionality
of the vector. Because of the nature of the representational space, the large majority of
distances between two vectors are concentrated at 0.5 (or 10,000 bits in our example). In
statistical terms, picking two random vectors, there is over 99,99% chance that the distance
separating them is between 0.47 and 0.53. If we take a third vector, it will also differ from
the first two by around 10,000 bits. These vectors are considered as unrelated, and are said
to have a dot product between them of mean zero.
Because of these characteristics, if we were to take a large number of vectors, such
9
as 10 vectors that are 20,000-dimensional, as an example, this will still be a very small
percentage of the total number of vectors in vector space. Noisy vectors would still be
identifiable among the large number of chosen vectors, as long as the chosen vectors are all
randomly spaced between each others. Related, or similar, concepts can be coded in vectors
that have smaller Hamming distances separating them. This would make it so if the noise
was too great, and we were unable to recuperate the original vector, we would at the very
least obtain a closely matching, or similar, one.
9
A cognitive system such as the brain includes several representational spaces, of different
dimensions each, which are used for different tasks. As we will see later in this review, some
neurons exhibit the characteristics of low-dimensional vectors, while others demonstrate
properties of high-dimensional representations.
3.3
High-dimensional memory
The memory stores the data, and set of instructions necessary to run, and manipulate a
program. A classical memory architecture is made up of memory locations, which are
an array of addressable registers, each of which hold a string of bits. The contents of the
memory locations is made available for use by probing the memory using the location’s
address, which is also a string of bits. In our case, a memory that would store 20,000 bit
vectors would have to be addressed by 20,000 bit vectors, and would have a total of 220,000
memory locations in order to store vector space as a whole.
This overflow of memory locations would make it possible to retrieve a particular set
of data using it’s address, or even an approximation of its address through noise and error.
There are two storage modes, each with distinct properties and advantages.
3.3.1
Heteroassociative and Autoassociative memory
Heteroassociative memory is based on the mechanism of storing a memory µ using the
pattern of α as an address. µ can thereafter be retrieved using either α, or a noisy approximation of α, such as α0 , α00 ...etc. This type of memory makes it possible to store sequences
within memories and their addresses. By making α the address for the memory µ, and µ the
address for the memory M , we can sequentially recall memories µ and M using only α as
an initial address.
First coined as Correlation Matrix memories[39], this type of architecture was not
able to correctly perform a recall task using noisy addresses. Hetero-associative memories
saw a big leap with Bidirectional Associative Memories[40]. In this model, subtle noise
in memory addresses could be overcome, and memory recall worked with forwards, and
backwards passes, by storing the memory in vector pairs.
The other type of memory storage is autoassociative memory. Autoassociative memory is
achieved by storing a memory µ using µ itself as the address. Just as previously described, µ
can be recalled by µ, or by a noisy version of µ, such as µ0 , µ00 ...etc. This type of addressing
therefore works as a noise filtering technique. It is used in several content-addressable
memory models, such as the Hopfield network[32], which guarantee a convergence to a
pure, non-noisy, memory, through several iterations.
Going from µ000 , to a less noisy µ00 , to µ0 , we can then finally obtain the stored memory
µ. However, depending on the address used, its proximity with the initial value, and how
noisy a particular vector is, convergence to a false pattern could still occur.
10
4
The brain and diversity
There are two big types of neurons that are present in the brain, each of which has their own
representations. The first type of neurons perform extremely simple and specific tasks, such
as color recognition or space orientation. These neurons behave as bits, or low-dimensional
vectors. Their firing is straight forward and predictable using what we know today as basis
for their computation algorithms.
The other type of neurons is mixed selectivity neurons. These neurons are most notably
present in the thinking, planning and higher order regions of the brain. Unlike classical
selectivity neurons, these don’t respond exclusively to one stimulus or task. Instead, they
react in different ways when confronted with a wide variety of stimuli[3]. They are essential
for complex cognitive tasks and give us a computational boost and a cognitive flexibility.
Up until very recently, it was thought that these neurons fired in random, stochastic waves.
Their behavior did not conform to what we knew about neurons.
4.1
Mixed selectivity
Mixed selectivity neurons are different from classical selectivity neurons. They are nonlinear,
highly heterogeneous, seemingly disordered, and difficult to interpret. However various
teams are trying to reverse engineer their structures, by collecting data through various
neurophysiological experiments. In order to observe these features, single neuron recordings
must be performed in situ. Because of the invasive nature of the procedure, experiments
have been restricted only to behaving animals, particularly to the prefrontal cortical region,
a region known for its implications in thought, behavior and decision making.
The results seem to indicate that the populations of mixed selectivity neurons encode
distributed information. This feature was not observed in the classical neurons, which are
confined to single tasks.
The neural representations are also high-dimensional, where the dimensionality represents
the firing rate of the neurons. This makes it possible for these neurons to have a large set of
input to output relations. Through these features, neurons can generate a rich set of dynamics
and task solving capabilities[66]. The nonlinear mixed selectivity is sufficiently diverse
across neurons that the information about task types could be extracted from the covariance
between neural responses. Even when the classical individual neurons were experimentally
devoid of any useful information, the knowledge and task solving capabilities of the primate
were conserved. In order to remove classical selectivity, noise was added to every classical
neuron, equalizing their average responses, and nullifying their effects. After removing this
classical selectivity, a large drop in accuracy was observed in the early stages of the trial.
However the accuracy increased later on as the trial progressed, and more information was
gathered[67].
11
It is interesting to emphasize that these same properties were observable using highdimensional representations, simulated on a classical Von Neumann machine using binary
vectors. It’s essential to get a close look into the brain’s circuitry and representations
through neuroscience. But it is equally important to then learn to abstract away from the
neurotransmitters, membrane potentials, spike trains, and other physical attributes of the
brain, and instead tackle the resulting behaviors. This will help us understand the underlying
mathematical principles based on them. In other words, the logical design must be separated
from neural realization.
4.2
Robustness
It is widely known that humans lose gray matter as they age. In fact, it’s been observed that
there’s a physiological 10% to 30% reduction of the cortical gray matter density, in normal
individuals, between ages 7 and 60[78]. While some individuals may lose as much as 1/3rd
of their brain, this goes on with little to no perceivable change. With all these variations,
and the significant day-to-day loss of neurons, our every day life goes by uninterrupted. The
neural architecture is therefore very tolerant to component failure. That is, unlike classical
binary representations in computers, where even a single bit of difference could either result
in system errors, or lead to significantly different results.
Redundant representations are a hallmark of high-dimensional vectors. The higher the
dimensionality, the higher will the proportion of allowable errors be. As discussed earlier,
patterns can differ by a certain number of units, yet they could still be considered equivalent
if the representational-space was large enough.
For maximal robustness, and an efficient use of redundancy, the information should be
distributed equally among all units. This would lead to a steady degradation of the stored
information. This degradation is proportional to the number of failing bits, irrespective of
their types or positions. While we would like to employ a robust architecture, it is also
essential to take into account the plasticity that the brain exhibits. Modeling is therefore
done in a holistic manner, which is a perfect trade off between robustness and plasticity,
optimizing neural resources and redundancy[85].
4.3
Randomness
Brains are highly structured organs, but many details are left to randomness. Network
formation is not very constrained, with a strong component of randomness reigning on
it ever since its genesis. The distance between nodes, and their connections follow an
independent random distribution, making no two brains structurally identical[71]. Any two
brains are therefore incompatible at the level of hardware and internal patterns.
Not only is there a structural randomness, but also constant spontaneous neuronal
quantal release. This is due to the random probability associated with transmitter release[91].
12
Random potentials are therefore stochastically created, and are known as synaptic noise.
Synaptic noise can aid, or impair, various features such as signal detection, neural performance, and may shape certain patterns.
For a correct neural simulation of the brain, the system should have randomness ingrained in
it. The model should be built using stochastic patterns and connections, by drawing vectors
in a pseudo-random fashion. The actual patterns themselves aren’t of any significance. What
does matter for computation and consciousness are the relations between patterns within
the system. Patterns for two similar concepts should be similar in one system, whereas
two similar, or identical, concepts should theoretically have different patterns in different
systems.
5
Computationalism, and Connectionism
Research on intelligent systems has split into two different systems, the classical symbolic
artificial intelligence paradigm, also known as computationalism, and the connectionist AI
paradigm.
5.1
Computationalism
Symbolic AI is at the forefront of computationalism. It maintains that the correct level at
which we should model intelligent systems, such as behaving animals, or the human mind,
is through symbols. A symbol is an entity that can have arbitrary designations, such as
objects, events, and relations between objects, events, or both. Each symbol corresponds to
one, and only one, entity at any given time.[55] Symbols may be combined together to form
symbol structures, such as expressions or sequences. In a conventional computer, a symbol
is represented by a variable, which can be instantiated, by making it refer to a specific entity,
or uninstantiated, where it doesn’t refer to any specific entity, and can therefore be rebound.
5.2
Connectionism
Connectionism is split into two parts: localist representations, and distributed representations.
Each of them has their respective advantages and short-comings. We will be detailing them
in the upcoming section.
5.2.1
Localist representations
In localist representations, a single unit, or neuron, represents a single concept. It is relatively
close to the idea of symbolic AI, except instead of binding objects to units, we are binding
13
features of objects to units. This gives much more freedom in representing entities with
similar properties, but comes at the cost of harder variable binding, and a longer processing
time.
Variable binding can be achieved in several ways, most notably using signature propagation[88]. A separate node is allocated for each variable associated with each entity,
and a value is allocated to represent a particular object, it is the signature of the object it
expresses. The signature can then be propagated from one node to the others in feed-forward,
or backward chaining processes depending on the task at hand. Signatures will also need
to be manipulated during state transitions, certain actions or orders, and variable rebinding.
This can be done through vector addition, or multiplication.
Another way to bind variables in localist representations is through phase synchronization, which utilizes the temporal aspects of activations. An activation cycle is divided
into multiple phases. Each node fires at a particular phase, and can be either an object
node, which holds a certain value, or an argument node, which holds a variable. Each phase
represents a different object involved in reasoning. Objects have their own assigned phases,
and fire in a consistent manner. When an argument node fires at the same phase as an
object node, the variable is bound to its value. After that dynamic binding is accomplished,
which forces the variable node to fire synchronously in phase with the constant node it is
binding[43]. Reasoning is therefore a propagation of temporal patterns of activation.
5.2.2
Distributed representations
In distributed representations, each entity is represented by a pattern of activity that is
distributed over many units. Each unit may also be used to represent many different
entities[75].
One way to implement distributed representations is through the use of microfeatures.
Microfeatures[26] are a more detailed approach to the items of the domain. Unlike the first
order features used in localist representation, these are composing elements of the features
described above. Different entities can be represented through a distinctive pattern over
these elements, making it possible for two different features, or two different objects, to
share some microfeatures.
Another way to realize distributed representations, that doesn’t involve the use of
semantic microfeatures, is coarse coding[69]. Coarse coding imposes two conditions on the
features that are to be assigned to units. The features must be coarse, that is, they must be
general enough to be shared by more than one entity, and have an extended definition. The
assignments of the units must overlap and superpose, in such a way that one unit is involved
in representing several different objects. This redundancy leads to a highly fault-tolerant
network. The coarse code is a robust representation; the failure of one, or even several
components will not make the system unusable. These redundancies also help the system
provide a higher than normal degree of accuracy. Similarities between objects will therefore
be reflected by similarities among their representations.
14
As is obvious, systems here work at sub-conceptual levels, and entities don’t have a
precise, complete, or formal description. Instead, features are represented by a pattern of
activity over a number of different units, scattered in the network in a non-organized fashion.
Because of this, the system has a complex internal structure that is crucial for its correct
operation. This may be the most optimal way for the system to go through processing,
however it will be very hard for a human operator to decipher and interpret these structures.
As a result, variable binding has proved to be exceptionally difficult. A lot of attempts have
been made, of which we will cite a few now, while other more successful attempts will be
explained in more detail, later in this review.
The Distributed Connectionist Production System[79], DCPS for short, manipulates symbolic expressions using conditional rules. Rules act as rails, and specify which expressions
should be moved in, or out, of the working memory of the DCPS. The distributed representation is done by means of coarse coding. However, the main limiting factor is the presence
of only one variable in the whole system. Variable binding is therefore very limited, and
can only be performed on a special set of terms. By using a memory that is large enough,
we may avoid immediate limitations by temporarily storing our variables there, at the cost
of computational time. However this is a non optimal solution that does not encompass all
use case scenarios. It comes at the cost of a severe bottleneck originating from the system,
trading computational time for simplicity, with the lack of variables provided.
Tensor product representations[77] have also been used to implement variable binding.
Vectors are combined using vector addition and aperiodic convolutions. This creates a
more complex object out of 2 or more basic features. Each object is therefore made up
of a set of variables, which are the arguments or roles, and each role has a value, or filler.
A vector can represent either a variable, or a value. Variable binding is accomplished by
forming the tensor product using aperiodic convolution. In order to represent a whole
entity or item, tensors can be added together. Pattern matching can be carried out using
these tensor products, but full variable to value unification cannot be achieved through this
method. However, this concept is the origin of much of the modern view on distributed
representations, as we will be seeing later.
In order to achieve an efficient level of binding variables to their values in connectionist
representations, rapid variable binding is essential. This implies creating variables on the
fly, and binding them to their correct values in a virtually instantaneous manner, before
deploying them, and using them according to their context. However, that comes with a
wide range of problem of its own[21].
5.3
The necessity of variable binding
As discussed, in order to meet the requirements of synaptic plasticity and speed of the
brain, rapid variable creation and binding should be performed in distributed representations.
15
There have been numerous attempts that were made in order to solve this problem, each
with its own trade-offs.
Eliminative connectionism was introduced in order to contour the problem of rapid
variable creation, without tackling it head on. The objective of this approach was to
completely avoid using localized or explicit variables. However, in order to match the
brain’s capacity to induce and represent identity relations, as well as concatenation functions,
variables are needed for the manipulation of data. Explicit variables are also needed to
achieve a decent level of generalization, and functional mapping[44]. The brain has a
remarkable ability when it comes to representing relations between variables that are novel
in our experience with variables that are familiar to us. Not only that, but humans are able to
induce a general pattern from a novel test case in a remarkably fast time. Let us demonstrate
this through an example:
Consider the following set of inputs and outputs:
Input: A car is a //
Output: Car
Input: A house is a //
Output: House
Input: A capacitor is a //
Output: Capacitor
What would the Output b e for:
Input: A juitrekfluit is a //
The solution to this test case is apparent. In less than a second, a Human subject can figure
out the answer to this novel case, a word which he had never seen before. One may argue that
the semantic meaning associated with these words might bias the human operator, making
it unfair to compare him with a machine that doesn’t have any capacity at understanding
phrases or words. We will therefore present another more involved, and random, example:
Consider the following set of inputs and outputs:
Input: jutrika setek b olur hiuna //
Output: Bolur
Input: jutrika setek nyata hiuna //
Output: Nyata
Input: jutrika setek jinski hiuna //
Output: Jinski
What would the Output b e for:
Input: jutrika setek kouyte hiuna //
This generalization capacity, induced in hundredths of a second when faced with even
completely novel structures, does not arise from training a network the moment the sample
data was presented. There is simply not enough time to train a neural network this quickly,
especially at the level of complexity exhibited by the brain. If synaptic weights were indeed
being changed, this can only be done through Long Term Potentiation, or Short Term Potentiation. Long Term Potentiation requires dendritic spines growth, and therefore takes hours
to develop. Short Term Potentiation is a transitory alteration of the synaptic vicinity, and
takes around 10 seconds to develop. If what was observed here was not due to new changes
in synaptic weight, yet was novel enough to surely not have a neural network specifically
16
trained to suit it, then it very probably relies on some pre-existing skill, acquired through the
training of subnetworks within the brain at a past date. These networks are being used with
their pre-existing variables, to which novel values are being bound rapidly, on the fly.
A practical solution to this problem may reside in VSAs, or Vector Symbolic Architectures, which permit effective variable binding, and can lead to generalization in the presence
of suitable algorithms[54]. This group of architectures assumes that each value, and each
variable is represented by a distributed activation pattern within a vector. An additional
vector can be used to encode the result of binding a value, to its variable. However even
in such cases, it is impossible to predict how a network would assign and operate through
novel tasks on its own, as weight vectors gained through regular training may not suffice for
generalization to new novel cases[21]. We will be detailing VSAs later in this review.
6
Artificial Neural Networks
Artificial neural networks, ANN for short, are an adaptive computational method that learn
a particular task through a series of examples. They can be used to model various cases
of nonlinear data. Just so, it happens to be that most physical phenomena are inherently
nonlinear. This makes artificial neural networks an excellent candidate to model real world
occurrences. The test samples are multivariate data represented by several features and
components, that is to say, they are represented by high-dimensional vectors[83].
ANNs were once deemed a curse brought onto computer science. They produced
astonishingly accurate results, but using high-dimensional vectors came at the cost of an
exponentially growing number of function evaluations. This occurrence was due to the
lack of optimization of older algorithms, which used exhaustive enumeration strategies[6]
that were sub-optimal for such tasks. Currently, we are better equipped to deal with highdimensional neural networks through dynamic programming[33], conditional distribution of
features among hidden layers[8], and generally better performing and more optimized neural
network structures such as recurrent neural networks and restricted Boltzmann machines[1].
Old ANN algorithms were developed mostly for low-dimensional computing. Nowadays, with the era of massive automatic data collection systematically helping both researchers and programmers obtaining a large number of measurements, new interest has
been kindled in researching high-dimensional modeling[13]. This motivation was also
supported by parallel discoveries in neuroscience, which helped generate computer models
based off of the efficient computational abilities of our brains. The first of such models were
PDPs.
6.1
Parallel Distributed Representations
Parallel distributed processing[46], PDP for short, is a type of artificial neural network that
took its inspiration from the processing model of the brain. In situ, neural connections are
17
distributed in parallel arrays, in addition to serial pathways. Functions are not computed
sequentially in series, but are rather processed concurrently and parallelly to each others.
More than a single neuron is sending signals at a time, and up until the mid-1980s, this was
a rarely used feature in the digital computer.
As was discussed earlier, in the Von Neumann architecture, each processor is able to
carry out only one instruction at a time. In PDPs however, information is represented in a
distributive manner over different units, and more than one process can occur at any given
time. Memory is not stored explicitly, but rather is expressed in the connections between
units, similar to how neural synapses work. Learning can occur through gradual changes in
connection strengths through experience. This system also exhibits all the advantages of
distributed representations.
6.2
Restricted Boltzmann Machines
Restricted Boltzmann machines, RBM for short, are a type of artificial neural networks
that specialize in deep-learning[76]. An RBM has only 2 node layers: A visible layer, that
receives input, and a hidden layer, which allows the occurrence of probabilistic states.
An RBM uses a generative probabilistic distribution over a set of inputs, in order to
extract relevant features. More than one RBMs can be stacked on top of each others, in order
to gradually aggregate groups of features. Each RBM within the stack deals with features
that are progressively more complex. Through these numerous layers, several groups of
features are conceived, which help decode the presented concept into smaller models of
representations. The more stacks there are, the better they handle complex tasks. This,
however, comes at the cost of computational time. A network composed of aggregated RBM
stacks is called a deep-belief networks, and is considered to be one of the more prominent
breakthroughs in artificial intelligence research and deep learning.
Each layer communicates with both the previous and the subsequent layer. Nodes
do not communicate laterally with other nodes of the same layer. While each hidden layer
serves as an effective hidden layer for the higher nodes, it also acts as a visible, input layer
for the lower nodes. An RBM is fed one test trial at a time, without any prior knowledge of
its nature. Learning can be supervised, where the system infers functions and features from
labeled data, or unsupervised, where the network attempts to find a structure on its own,
among unlabeled data. The connection weights between two nodes that lead to a correct
conclusion become stronger, while the weights that lead to wrong answers become weaker.
However, restricted Boltzmann machines are not deterministic. Given a particular set of
units, initial conditions, inputs, and connections, the network won’t always arrive to the
same conclusion, due to its stochastic nature. The network operates through probabilistic
distributions over patterns, by doing several passes over the original test set. States are
updated accordingly, with a high probability of the network to converge, and a low probability
for it to diverge[70]. This is unlike deterministic networks, such as Hopfield Networks,
18
where the network is programmed to always, and consistently, attempt to converge through
each step. What this leads to is a higher operation time, at the cost of lower chance of getting
stuck at local minimums. That is, there is a higher chance to find the ”absolute truth” in
stochastic models, because their randomness is enough to, almost, force the network to settle
at the lowest state possible.
The most widely adopted algorithm for use in restricted Boltzmann machines is
Markov Chains[12], which are logical circuits, that connect two or more states using guided
and learned probabilities. Probabilities exiting one state add up to 1. This algorithm is
sequential, and given one state, it can dissect what the next state would be, and predict
subsequent states and sequences.
6.3
6.3.1
Recurrent Neural Networks
General Overview
Recurrent neural networks, RNN for short, are a powerful class of artificial neural networks
where connections between units form a directed cycle. The initial inspiration for this model
came from neurophysiological observations. Neurons were not only feedforward, they
were also interconnected. Signals between neurons carried over by all or none principles.
Thus was born the concept of RNNs. Nodes are not only linked using forward chaining
connections, there are also lateral connections, as well as connections from current layers
into past layers. This binary stream[47] had a remarkable short-term memory for its time, and
had influences that stretched far beyond neural networks. It even influenced the development
of the digital computer by Von Neumann.
The basic architecture of recurrent neural networks is the fully recurrent neural
network. Each node has directed connections to every other node. There are many models
and algorithms that seek to better the computation and learning capabilities of RNNs. Apart
from their use in the computational and mathematical fields, RNNs have experienced a large
adoption rate in brain modeling. They are not only a way for researchers to reproduce the
mind’s capabilities digitally, but also a way for them to theorize features which have not been
experimentally observed yet through the traditional means of neuroscience or psychology.
6.3.2
Turing Completeness
RNNs’ inherent representational power made them an attractive candidate for universality,
otherwise known as Turing Completeness. Turing Completeness can be achieved by proving
that the system at hand can compute each and every function computable by a Turing
machine. In other words, this means the system should be able to calculate any recursive
function, use any recursive language, and solve any problem solvable using any effective
method of computation or algorithm[81][82].
At first, computational universality was achieved using recurrent neural networks
19
with a finite number of neurons, and high order connections, by combining their activations
through multiplicative means[63]. However, this is an unattractive candidate for practical
use because of the leaps of data that must be calculated. Such a model would consume a
large amount of memory space and its convergence time would grow exponentially with
the complexity of a given problem. Such a model is sub-optimal on many different levels.
Computational universality was also achieved in linear RNNs using additive connections,
assuming an unbounded, and infinite number of neurons[90]. This is also not practical, as
having an infinite set of nodes is physically not possible.
In order to assume Turing completeness, one can simply prove that the system can simulate
a Turing machine. In the case of finite, low-order, recurrent neural networks, this can be
done by proving that there exists a network, started from an inactive initial state, that can
lead to the correct output from an input sequence, using a recursively computable function.
The input could also be removed entirely, to be then encoded into the initial state of the
network. This should lead to a valid output, and correct results. Computational universality
was therefore proved to stand true in low order connection, finite neural networks[74]. An
estimated total of 886 neurons (870 processor nodes, and 16 input and output nodes) is
needed to reach this universality.
6.4
Using an external stack memory
A recurrent neural network with no hidden layers is capable of learning state machines, and
is at least as powerful as any multilayer perceptron[19]. It contains an inherent internal
memory that’s relatively small. This internal memory, stored in weights between the nodes
is not enough to store all the details that would be necessary for intermediate or long-term
use. In order to boost memory storage and long term performance, one could theoretically
use an internal memory stack. However, this would require a large amount of resources,
and would come in the form of additional nodes and connections, which would in turn
exponentially increase computational time[64]. An optimal memory stack would therefore
be external. It would allow for easy modulation, as well as continuous communication with
the network, which permits the use of classical and continuous optimization methods, such
as gradient descent.
An example of this is the Neural Network push-down automaton (NNPDA, [11]).
This architecture consists of a set of fully recurrent neurons representing the network’s state,
and permitting classification and training. State neurons receive their knowledge from three
distinct sources. First off, they receive their information from input neurons, which register
input fed to the network by sources external to the system. Secondly, from read neurons,
that interact with the external memory stack, and keep track of the symbols that are on
top of the stack. And thirdly, they receive feedback from their own recurrent connections,
seeing as this is a recurrent neural network. The network is connected to the stack through a
non-linear error function, which approximates results that hold with high probability.
20
Although it is possible to learn simple tasks using first-order networks, higher order nets
are a necessity for learning more complex features. As we’ve grown accustomed to when it
comes to complexifying computational algorithms, this comes at the expense of a longer
processing time, and slowed convergence. First-order networks are by no means complete,
they lack several features required to correctly order stored memory, as well as isolators, to
help fend off any unwanted interference. We will see how this problem could be surpassed
using LSTMs, and NTMs.
6.5
Long Short-Term Memory
As we stated before, recurrent networks can use their feedback connections in order to store
representations of recent input events in the form of activations. This is a very useful feature
when dealing with tasks requiring multiple inquiries over a short-term memory.
With conventional algorithms such as back-propagation through time[86], the error
signal is scaled proportionally to the number of steps it is propagated back. This leads the
error to blow up, growing exponentially and leading to oscillating weights, bifurcations and
unstable learning when the error is positive[60]. If the error is negative, this in turn leads to
vanishing weights, which induces excessive learning times when bridging events through
long time lags, rendering the modification of existing weights a time consuming challenge
to be dealt with[7]. The conventional backpropagation method is simply too sensitive to
recent distractions.
6.5.1
The basic architecture
To avoid the problems that come with back-propagation, long short-term memory, LSTM
for short, was introduced[29]. It’s a second-order recurrent neural network that enforces
constant error flow, by propagating it through dedicated constant error carrousels(CEC).
LSTM uses a linear constant activation unit to store the error, as opposed to the non-linear
sigmoid activation functions used for the other computing units.
Because the constant unit follows a different scaling path than the rest of the units
in the network, units connecting to it will be receiving conflicting weight updates. This
makes efficient learning difficult to achieve, and unstable to maintain. To counteract this
phenomena, we add multiplicative input and output gates. The input gates protect the
memory content stored in the constant unit from outside perturbation; the error is therefore
only edited when it needs to be updated. Output gates protect the rest of the units in the
network from the irrelevant information stored inside the constant error carrousel; the error
therefore only affects the network when it is recalled and needed. The gates are context
sensitive nodes that can be trained, using gradient descent, to decide when to overwrite or
access the values in the CEC.
A ”memory cell” is built around the CEC, using a central linear unit, with a fixed
self-connection, and delimited by the input and output gates discussed earlier. Several
21
memory cells can share the same gates, and form a memory cell block, which is a more
efficient architecture that facilitates storage. Error inside the memory cell blocks do not
get scaled when propagated back further in time because of their linear constant activation
functions. This eliminates exploding or disappearing error values, and stabilizes long-term
memory.
This architecture can handle noise, distributed representations, and continuous values, while
increasing the ability to generalize over untrained values better than older implementations
of neural networks, including BPTT[86] and RTRL (real-time recurrent learning[68]).
6.5.2
Extensions to the basic LSTM model
LSTM, being one of the more promising RNN architectures of its time, received several
improvements along the years. Forget gates[17] were added along side the input and output
gates. Forget gates modulate the amount of activation a memory cell keeps from its previous
time-step. They are able to force a quick memory dump in memory cells that have completed
their tasks. This helps in rapidly keeping units effective, and augments the proportion of
efficiently used and recruited nodes during later computations.
Another improvement to LSTM which quickly became a staple in computation are peephole
connections. Peephole connections[18] are links between the three types of gates, and the
memory cells. In the basic LSTM model, when the output gates are closed, the memory
cell’s visible activity is null. The information contained inside the CEC is therefore hidden
from its associated gates, and surrounding nodes, making it impossible to efficiently control
the information flow into, and out of the memory cell block. These peephole connections enable the cell’s unmodulated state to be visible to outside units. This improves the network’s
performance significantly by reducing random instances of gating and probabilistic editing,
which used to happen due to insufficient information.
In order to expand LSTM beyond second-order RNNs, Generalized LSTM
(LSTM-g, [51]) was introduced. In this architecture, all the units in the network have
an identical set of operating instructions. A unit relies only on its local environment to
determine what role it will be fulfilling. Units can be input gates, output gates, forget
gates, memory cells, or any combination of these. Every unit is also trained using the same
algorithms. That is in contrast with conventional LSTM, where each type of unit had its
specific set of algorithms. A number of other changes were also made in order to improve
flexibility. This helps improve performance, and broadens the applicability of the LSTM
networks in machine learning. LSTM-g can be shown compared to the regular LSTM
architecture in Figure 1.
22
Figure 1: Comparison between an LSTM memory cell block, and its equivalent in LSTM-g. Black
lines: Weighted connections. Grey lines: Gated connections. In the LSTM model(a), the input is first
summed, then gated as a whole by the input gate. The output is also computed as a whole, all subsequent
units receive the same modulated output. Peephole connections project from an internal stage in the memory
cell to the controlling gate units, only the final value of the memory cell is visible to other units. In the LSTMg model, the inputs to the memory cell are gated individually before being summed. The output leaves the
memory cell unmodulated, and is then individually modulated by the output gate to other cells. Peephole
connections connect to the recurrent units of the memory cell, and are able to see its intermediate state. This
allows for greater flexibility. Image courtesy of [51].
6.6
Long Term memory
Recurrent neural networks stand out from other machine learning methods due to their ability
to learn and carry-out complicated transformations of data over extended periods of time.
But one aspect that was largely neglected in computer science and research was the use of
23
logical flow control and external memory. As we’ve discussed, RNNs are extremely effective
when it comes to storing recent memories. With the advent of LSTM, intermediate-term
memory storage became a lot more efficient. However, this did not significantly improve
networks’ storage capacities. During long, and complicated tasks, networks should have a
way to store large quantities of information, while efficiently maintaining control over their
flow.
6.6.1
Neural Turing Machines
Neural Turing machines[20], NTM for short, is a neural network architecture that attempts
to deal with the problems showcased earlier. NTMs are enriched with a large addressable
memory, that can be trained using gradient descent, to yield a practical mechanism for
machine learning. For maximum effectiveness, the network will have to operate on rapidlycreated variables, which are data that are quickly bound to memory slots. Their architecture
is illustrated in Figure 2.
Neural Turing Machines interact with the external world using input and output
vectors, just like all neural networks. However they also interact with a memory matrix,
using an attentional process to selectively read and write operations. Every component of the
architecture is differentiable, and instead of addressing a single chosen element per cycle, the
network could learn to interact with a variable number of elements in the external memory
bank, as it sees fit. This is done through blurry read and write operations, determined by an
attentional focus mechanism. This constrains them, and makes it possible to attend sharply
to the memory at a single location, or weakly to the memory at many locations.
Read vectors are stored in the memory matrix’s column, while the number of rows is
equal to the total number of memory locations. Writing is done in two parts. First using an
Erase Vector, a full memory location can be entirely, or partially emptied to make room for
subsequent data. After that, the network uses an Add Vector to write to the memory. Both
Erase and Add vectors are differentiable, computing their current state from their previous
state, weight vectors and input orders. The presence of two separate vectors, both being
independently subject to learning, allows a fine-grained control over which elements in each
memory location are modified and how.
The weightings arise from the combination of content-based addressing and location-based
addressing. Content-based addressing focuses attention based on the similarity between the
calculated, noisy values, and the noiseless values present in the memory space. This makes
it possible to use approximations to speed up calculations, leaving it to the autoassociative
memory to retrieve the exact value and clean up the noise[32]. Location-based addressing is
used when values with arbitrary content are employed. Since it is not possible to predict
these values, yet they still need a recognizable name and address, the values are addressed
by location, and not by content. This is designed to facilitate both simple iterations across
the locations, and random-access jumps[31]. The content of a memory location could also
include information about the location of another value inside it. This allows the focus to
24
Figure 2: Neural Turing Machine architecture. The controller, or neural network, receives the input from
an external environment, and emits an output in response. The controller is also able to read, and write from
a memory matrix, using read, and write heads. Image courtesy of [20].
jump, for example, to a location next to an address accessed by content.
Either a feedforward, or recurrent network can be used here, each with its own
advantages. While the RNN allows the controller to mix information across several timesteps of operation and permits a larger number of read and write heads, a feedforward
controller confers greater transparency of the networks operation, because its patterns are
easier to interpret by a human operator. This transparency comes at the cost of an imposed
bottleneck due to the reduced number of read and write heads.
This network architecture leaves several free parameters for the operator to decide upon,
such as the size of the memory matrix, the range of location shifts in the location-based
addressing, the number of read and write heads, as well as the type of the controller network.
All of these parameters serve to make the Neural Turing Machines[20] as adaptable as
possible for a number of different tasks.
25
6.6.2
Memory Networks
Another solution to the lack of long-term memory in neural networks are Memory Networks[87]. This architecture was conceived around the same time as the Neural Turing
Machines described earlier, by a different, and independent research team, indicating a
recent surge in interest in the use of external memories as an extension to neural networks.
The model operates in similar ways to the NTM, by combining machine learning inference
with an external memory component that can be read and written to. This is done through
four learned components, them being an input feature map, an output feature map, a response
element, that converts the output into the desired format, and a generalization component,
which updates old memories given the new input and the current state of the network.
7
Vector Symbolic Architectures
Vector symbolic architectures, VSA for short, are high-dimensional vector representations
of objects, relations and sequences, used in machine learning. A good representation helps
systems compute orders of magnitude faster, making possible what would otherwise have
taken a near infinite amount of time. An example of a VSA that was briefly discussed
earlier is the tensor product model[77]. We will also be discussing holographic reduced
representations[62] in an upcoming section, one of the more notable and widely used VSA
architectures.
7.1
Vector components
Components of vectors can be binary, sparse, continuous or complex. Vector operations
such as addition, binding and permutation can be used to group desired elements together.
Since we are looking for a representation that is specifically suited for optimizing machine
learning, a number of constraints will have to be employed. With the goal to represent
potentially hundreds of thousands of objects at a time, a distributed representation will
have to be used, otherwise storing each component in its independent vector would not be
sufficient.
Similar objects and structures will have to be mapped to similar vectors, having a
small euclidean (or Hamming) distance separating them. This makes it so that similar
vectors will have a significantly greater dot product than 2 randomly chosen vectors. Such
a design decision would help when filtering out the error and noise generated during the
approximation and computation phases, ensuring that a word’s meaning, or a pixel’s color,
doesn’t vary too much across iterations and during recovery. This also gives the network a
better ability to generalize, by finding approximate matches to novel concepts, and using its
past knowledge to determine the implications of new entities.
Another important constraint to consider is the use of fixed length vectors. The goal
26
here is to incorporate VSAs with standard machine learning algorithms, such as Structured
Classification, where inputs and outputs are most conveniently expressed as vectors with a
pre-specified number of components. Sequences, such as sentences, however don’t have a
fixed number of components, or a fixed structure. In the specific example of sentences, a
potential solution to that problem would be using part-of-speech tagging, chunking, named
entity recognition and semantic role labeling[10]. This helps us convert sentences into
unambiguous data structures using the syntactic, and semantic information contained in
them. Vectors’ dimensions should be tuned for maximum efficiency. Larger vectors have
more storage capacities, while smaller vectors decrease the system’s computation time and
make for faster network convergence.
7.2
Vector operations
All vectors being of the same size and dimension enable us to use the wide array of preexisting vector operations and algorithms, in order to combine and structuralize objects.
One such operation is the classical vector addition operator. Addition is commutative, and
encodes an unordered set of vectors into a single vector of the same dimension. A structure,
or sentence, could therefore be encoded in a single normalized vector, consisting of the sum
of the individual word vectors. A different sentence formed using the same word vectors
would also have the same resulting sum. Unlike scalar addition, vector addition preserves
constituent recognition, which enables information retrieval[9]. If the dot product of the
resulting sum vector with the constituent we are looking for, is close to the dimension of
the vector, then that element is present in the vector. In order to better demonstrate this
operation, let us consider a brief example:
Let us consider the set of n-dimensional vectors:
a, b, c, d, e, and F.
F =a+b+c+d+e
We would like to check if the vector V is one of the vectors
comp osing F.
For the purp ose of demonstration, let us consider:
V =b
V ∈F
This is done by computing the dot pro duct of the vectors
V and F :
F.V = (a + b + c + d + e) ∗ V
27
Multiplication is distributive over addition:
F.V = (a + c + d + e).V + b.V
Similar vectors have a dot pro duct that is almost equal to
their dimension n.
F.V =< mean 0 noise > +n
If the dot product of the sum vector with the tested vector is equal to, or greater, than the
vector’s dimension, this would indicate that the tested vector is one of the vectors that was
used for the sum. However it is not possible to know exactly which vector that is, or where
its position was.
Since vector addition stores only unordered sets of vectors, relying on it is not sufficient
to encode information structure and order. We need a way to bind objects in a structure
together, while retaining a certain degree of flexibility when it comes to moving them around.
We want to bind words in a sentence together, while retaining the inherent meaning that
comes with them if we change the order of the words in certain ways[80].
For that, we use the binding operator, to bind a sum of vectors as part of a structure
description, by multiplying the sum vector by a matrix. We use the features described
previously[10] to identify chunks of data that maintain coherence even when shuffled, then
bind them to a matrix, or sub-matrix, depending on their relevance in the general order of
things. Component recognition is done the same way in vector binding as it was using the
addition operator.
By using multiple binding matrices, we can bind several batches, chunks, or objects
into a single vector. We can then use the preserved recognition feature to determine if a
certain element appears in any of them, if 2 elements appear together in the same matrix, or
if they appear in the same global structure, but in different chunks. This helps subsequent
learning and alleviates the ambiguity issues faced when using only vector addition.
In order to increase robustness, we can encode structure descriptions several times
in a vector, by adding them. Structures can be represented by sequentially binding their
components to matrices, in order to fix their precise order. This gives the network the
ability to not only generalize to a novel structure that is similar to a the inputs, but also the
ability to know how similar certain structures are, or if they were exact replicas, element
and order-wise[16].
7.3
Holographic Reduced Representations:
A problem with representing concepts using connectionist networks is that items and
associations are represented in different spaces. In order for part-whole hierarchies, which
28
will be detailed later, to work as intended, we should employ a vector that acts as a reduced
description of a set of vectors, and combines their characteristics into one, for easier
manipulation. For this, and several other reasons stated earlier, Holographic Reduced
Representations, HRR for short are used. They are constructs of the association of vectors
in a reduced and compact fashion. The result of an association, of two or more vectors of a
given dimension, is a vector with the same dimension[62].
7.3.1
Matrix memory
Traditional matrix memories, are simple, and have high capacities[89]. They use three basic
operations for their storage needs. First off, they use encoding, where two item vectors
result in a memory trace, or matrix. Secondly, they can use memory trace composition,
through addition, or superposition of several memory traces, in order to compress more than
two matrices, or more than two vectors, into a single matrix. Thirdly, they use a decoding
process, where a memory trace and an item vector are used to give the other item vector. To
better illustrate these ideas, let us showcase some vector operations.
Basic vector op erations. Let us consider:
V : Space of Vectors representing items.
M : Space of Matrices representing memory traces.
: Enco ding op eration, used to transform two vectors into a matrix:
V V =M
: Trace comp osition op eration, comp oses two memory traces into
one:
M M =M
B: Deco ding op eration, used to retrieve a Vector from a Matrix,
using a cue vector:
V BM =V
Memory traces are very flexible. They can represent a number of associations, and compositions through pairs. Any item from any pair can be recovered, using the other vector it was
associated to as cue.
29
Let a, b, c, d, e and f, b e Vectors representing items.
The vectors a and b can b e enco ded into a matrix:
ab=M
Let us demonstrate memory trace comp osition:
(a b) (c d) (e f ) = C
Using any vector as a deco ding cue will recover the other vector of
the pair:
eBC=f
Noise increases as more vectors are stored in a single memory trace. This leads us towards
using autoassociative memory for noise and error filtering. However, the main issue with
using such a representation is the exponential loss of space as dimensionality increases. To
store N-dimensional vectors, O(N 2 ) units will be needed to represent the N × N matrix.
This calls for new ways, to represent vectors in memory cells using more constrict algorithms,
for better space efficiency.
7.3.2
Aperiodic convolution
Past attempts at convolutions used aperiodic convolution for associations[77]. Two vectors,
of N elements each, are convoluted into a vector of 2N − 1 units. This is done by summing
the trans-diagonals of the outer product of the two vectors. This vector grows with recursive
convolution, as more vectors are added to it. For n vectors of N elements each being
aperiodically convoluted, the resulting vector will be [n ∗ (N − 1) + 1]-dimensional, which
comes back to using O(N 2 ) units, if n and N are equal and/or infinite. While non optimal
compromises can be made, such as limiting the depth of the composition, or discarding
elements outside the central ones[14], the growth problem can be avoided entirely by using
circular convolution.
7.3.3
Circular convolution
The circular convolution[62] of two N -dimensional vectors, is an N -dimensional
vector, and can be considered as a compressed outer product.
The circular
convolution of n N -dimensional vectors is one N -dimensional vector, occupying
O(N ) units. A representation of circular convolution is displayed in Figure 3.
30
Circular correlation is the approximate inverse of circular con- Figure 3: Circular Convolution
volution, and will be able to decode a memory trace out of a represented as a compressed outer
vector cue, in order to give back the second item vector. How- product. Image courtesy of [61].
ever correlation only gives an approximate, noisy, version of
the original item vector, which necessitates the use of autoassociative memory to filter out the noise. Circular convolution
can also be approximated using Fast Fourier Transforms, for
faster vector rendering.
In all of these methods, to successfully store a vector, we
only need to store enough information to discriminate it from
the other vectors. This property makes such compressions possible, up to a certain limit. The capacity of the memory model
is the number of associations that can be presented usefully in
a single memory trace.
Convolutions can also be used for variable binding, where the variable and the value are
bound together, or representing sequences. Sequences can be represented in one memory
trace, or divided into several memory traces depending on their lengths, and the desired
accuracy[53]. The decoding process is through chaining, where the cue vector at hand will
be used to start a cascade of retrievals, correlating the trace with the current item, until the
end vector is reached[52]. In some cases, where elements could be shared among different
structures using only one past element of the sequence may not be enough. It is therefore
also possible to encode several units into the past for a more robust sequencing. To better
demonstrate these concepts, let us provide an example of them:
~: Circular convolution op erator.
a~b=V
}: Circular correlation op erator.
b}V =a
Sequences can b e stored in several ways.
Non redundant storage:
X1 ~ a + X 2 ~ b + X3 ~ c = S
31
Fully redundant storage, retrieval cue must b e built from the ground
up using correlations.
a+a~b+a~b~c=S
Semi-redundant storage, conserves a bit of redundancy, and gives
much more freedom for retrieval op erations.
X1 ~ a + a ~ b + X2 ~ b + b ~ c + X3 ~ c = S
We can restore the memory c from S using X3 as a cue:
X3 } S = c
Using correlation, and the relevant cue vectors, we can then retrieve our stored information
from the memory network. The retrieved vectors will be noisy, so an autoassociative memory
would go best with this storage method.
8
Working memory
Working memory is a concept of human cognition, which aims at explaining our performance
in tasks involving the manipulation of short-term information. The limitations of working
memory have been thoroughly studied in psychology. Conclusive results found that memory
was stored in chunks of information. The number of chunks that can be recalled at any one
time was determined to be 7 ± 2 [49]. Further explorations found that the recall process
depends on several other factors, including word length, syllables and prior familiarity
with the words[34]. Training, however, did not seem to affect these chunks. The observed
performance boosts were rather due to how chunks of information were manipulated. The
biological cognitive terrain for working memory is a tripartite architecture, formed of three
functionally complementary systems[4].
8.1
Biological architecture
The pre-frontal cortex, posterior parietal cortex, and hippocampal system are the three main
components of working memory. Let us explore them briefly.
The pre-frontal cortex, PFC for short, specializes in the active maintenance of the
internal contextual information. It is dynamically and actively updated by the basal ganglia.
Basal ganglia, BG for short, are a group of subcortical gray matter, interconnected with
32
various structures in the brain, such as the cortex, thalamus and brain stem. The PFC-BG
system relies on dopamine based learning, and is able to bias on-going processing throughout
the cerebral cortex. This is done in order to maintain information that is currently relevant,
as well as fend off interference, and steer the attentional processes. This helps in performing
task relevant processes in the face of multicentric waves of knowledge, while reducing the
influence of other cortical regions which may not be relevant to the specific task at hand[48].
The posterior parietal cortex system performs automatic sensory and motor processing
in the brain by relying on statistically accumulated knowledge. This system exhibits slow
and integrative learning, through skills that are gained over long-term training. This long
term processing power is believed to reside in the angular gyrus. It incorporates memory
tasks involving extended periods of training and use, such as language processing, reading,
writing and arithmetic operations[73].
The hippocampal system is a compound structure in the medial temporal lobe of the
brain. It is composed of several structures, the most important of which are the hippocampus,
and the entorhinal cortex. The entorhinal cortex is both the main input and output to the
hippocampus. This system is specialized for rapid learning after short, or non-recurrent trials,
as well as spacial localization and navigation. It binds together arbitrary information, which
can be recalled by other neurological structures[42]. This structure also helps consolidate
information from short-term memory, to long-term memory, through a process called Long
Term Potentiation, LTP for short.
8.2
Memory functions
These regions support basic memory functions associated with working memory. Among
the most important ones are partial recalls of on going processes, as well as controlling
processing functions. They work as a central executive unit by supervising and controlling
the flow of information into, and out of the system. The second of these emerges from the
biasing influence by the pre-frontal cortex, which is actively maintained and updated by the
basal ganglia, on the rest of the system. These two different functions are performed through
the same neurological path, and distributed in a stable configuration throughout the cortex.
We have previously showcased two different learning mechanisms, rapid learning, and
integrative learning. These operations require different types of neurons, some being rapid
learners, while others being integrative statistical learners. There have been researcher
that have attempted to model these neural structures, including replicating their functions
and performance in basic memory recall tests such as the AX-CPT. Seeing as the PFC-BG
system is the most important of these in both short lag memory and controlling other systems,
we will be exploring some of its models in later sections.
There are several demands to creating a functional pre-frontal cortex, and basal ganglia
model[23]. The system should be rapidly updating, and able to encode and maintain new
information as it occurs. It must also have the ability to selectively update information, by
33
knowing which elements to maintain in the face of interference, and which ones to update
from its previous iteration. The system should have the power to condition responses in other
parts of the network, by learning when to gate appropriately. As you have come to notice,
these are all points that have been discussed previously in their own context, through BPTT,
LSTM, NTM and PDP, but we will be seeing them combined here, in light of cognitive
memory processing.
But first, let us tackle the details of the PFC-BG operation, as well as its neurological
structures and connections, in order to better understand how one goes about implementing
such a system computationally.
8.3
Memory operations in the pre-frontal cortex and basal ganglia
The main feature of neurons in the pre-frontal cortex is their rapid update cycles. This
occurs when Go neurons, in the dorsal striatum of the basal ganglia, fire. These Go units are
medium spiny GABAergic neurons, which inhibit the substantia nigra pars reticulata. This
leads to it releasing its tonic inhibition of the thalamus. The thalamus, being disinhibited,
enables a loop of excitation into the pre-frontal cortex, which toggles the state of the bistable
neurons contained within. The Go neurons are in direct competition with NoGo neurons.
The NoGo neurons are also present in the dorsal striatum, and counteract the effect of
Go neurons by inhibiting the external global pallidus. This excites the substantia nigra
pars reticulata, which promotes the inhibition of the thalamus, therefore protecting the
information stored, and contributing towards robustness in the face of interference[50]. The
PFC-BG architecture is visible in Figure 4.
The basal ganglia is connected with the pre-frontal cortex through a series of parallel
loops. There are multiple separate representations, working as inner and outer loops. These
loops give the PFC-BG complex the ability to process several layers of information in
parallel. An outer loop will have to be maintained, serving as the framework for the
inner loop’s conditional statements. The outer loop is stable and has a slow rate of variance,
whereas the inner loop goes through rapid iterations of changes. In order to better understand
this mechanism, and its potential characteristics, let us illustrate its application through the
use of a Continuous Performance Test, the AX-CPT:
The AX-CPT task:
Subjects are presented with a random sequential stimuli of
flashing letters:
AorB, followed byXorY.
34
The test sequence is of the following nature:
A − Y − B − X − A − X − A − Y...etc
The prior stimulus should b e maintained over a delay, until the
next stimulus is presented.
The target sequence is A − X . A dedicated button should
b e pressed by the subject whenever he encounters it.
When any other sequence is detected, such as: A − Y , B − X or
B − Y , another button should b e pressed.
The first letter is therefore held in an outer lo op, around
an inner lo op which houses the second letter. This is required to
detect the target sequence.
These results are the fruit of several deductive studies performed on behaving animals that exhibit human-like memory performances, as well as humans. However, multiple
inner loops can co-exist in one outer loop. An inner loop could also play the role of an
outer loop, and house an inner loop of its own, resulting in a stack of inner loops. In order
to demonstrate that, let us present a more complicated and recent version of this test, the
1-2-AX-CPT.
The 1-2-AX-CPT task:
A 1 or 2 stimuli is added upstream of the regular A/B − X/Y
stimuli.
The test sequence therefore b ecomes:
1 − A − Y − B − X − 2 − A − X − B − Y − A − Y...etc
The target sequence here is variable:
• If the subject last saw a 1: A − X is the target sequence.
• If the subject last saw a 2: B − Y is the target sequence.
The
task
defines
an
outer
35
lo op,
1
or
2,
of
active
maintenance, within which resides a stack of inner lo ops.
Each of these inner lo ops is formed by an outer lo op, A or B,
and an inner lo op, X or Y .
A representation of the inner and outer loops is shown in Figure 5. In order to prevent
interference between different loops, this task is achieved through the unique architecture of
the PFC. This is done through isolated, self-connected, stripes of interconnected neurons,
that are capable of sustained fire, and active maintenance of the working memory. Selective
updating of these stripes occurs through parallel loops that can be overwritten independently,
in different areas of the basal ganglia and the pre-frontal cortex[2].
Using a dopamine-based learning mechanism, the Go and NoGo neurons are trained to
appropriately gate the activity of other neurons in the striatum. Dopamine acts on these cells
using different types of receptors, D1 and D2, respectively for Go and NoGo neurons. The
D1-like family (D1 and D5) activate the adenyl cyclase, increasing intra-cellular cAMP levels
during positive reinforcement. The D2-like family (D2, D3, and D4) inhibit the formation
of cAMP during negative learning reinforcement[25]. Each neuron develops its own unique
pattern of weights. cAMP in turn regulates neuronal growth and development, and mediates
some behavioral responses. This dynamic gating mechanism helps switch between updating
existing information relative to incoming stimulus, and protecting information in the face of
interference.
8.4
Modeling the PFC-BG system
To encode these features in a computational system, we will need a basal ganglia model,
with two types of neurons, Go, and NoGo, controlling a pre-frontal cortex model. The
learning problem boils down to teaching these neurons when to fire, according to the sensory
input they receive from an equivalent to the posterior cortex, and the maintained pre-frontal
cortex activations.
However, once information is encoded into the pre-frontal cortex, we would not be
able to know how beneficial it was. Feedback on whether that memory was helpful or
inappropriate will only come when that memory is recalled, later in time. Knowing which
prior events were performed for the subsequent good, or bad performance will be hard to
extrapolate, yet is critical for training our network. The pre-frontal cortex model will have
to be divided into stripes, containing independent loops. The network should then decide
what memories to encode in each stripe, reinforcing those who contribute to its success
during memory recall. This can be done through classical neural network algorithms, such
as backpropagation through time[86], although here, the problem is considerably more
36
Figure 4: A view on the memory architecture of the brain. Go and NoGo represent the medium spiny
neurons in the dorsal striatum. STN represents the subthalamic nucleus. The 3 segments in the thalamus
are VA for ventral anterior nucleus, VL for ventral lateral nucleus, and MD for the dorso-medial nucleus.
GPe is the external segment of the globus pallidus. SNr is the substantia nigra pars reticulata. The thalamus
is bidirectionally excitatory with the frontal cortex. The SNr is tonically active, and inhibates the thalamus.
When Go neurons activate, they inhibit the SNr, and stop the inhibition of the thalamus. NoGo neurons
indirectly inhibit the SNr, by inhibiting the GPe, which usually inhibits the SNr, this in turn activates the
thalamus. Image courtesy of [23].
complex because of the different types of nodes used, the latent feedback, and the unique
stripe-loop architecture.
The brain solves the latent feedback problem by predicting the subsequent rewards of
a certain stimulus[72]. Neurons connected to the basal ganglia, in the ventral tegmental area,
and the substantia nigra pars compacta predict the dopamine stimulus, thereby reinforcing
Go firing, when such maintenance usually leads to a positive reward. To model these
features, we can use the ”Primary Value Learned Value”[57], or PVLV, Pavlovian learning
algorithm. In this algorithm, the Primary Value controls performance and learning during
only the primary rewards or unconditioned stimuli. On the other hand, the learned value
37
Figure 5: Inner and outer loops represented through the 1-2-AX-CPT task. The stimuli is presented in
a sequence, and the participants respond by pressing one of two buttons. The right button, R, is pressed when
the subject is presented with the target sequence. The left button, L, is pressed when they are presented with
any other sequence. Image courtesy of [23].
processes the received stimulus, and learns about the conditioned stimuli that are reliably
associated with rewards.
The unique stripe-loop architecture can be explicitly observed in the pre-frontal cortex
during working memory recall tests. This indicates that for the brain, as well as the models
that should be used to model it, information that must be updated at different points in time,
must also be represented in different parts of the pre-frontal cortex[38]. Anterior areas of the
pre-frontal cortex are selectively activated for outer-loop information, while the dorsa-lateral
parts are active for inner-loop information. The model used should therefore employ parallel
loops, that are updated independently by external structures. A dynamic, context sensitive,
gating mechanism should be used to update, or protect, relevant information in the face of
respectively relevant, or irrelevant inputs.
9
9.1
Modeling the cortical architecture
Multiple layers of representation
In order to make the fine distinctions required to control behavior and other types of
processing, the cerebral cortex is composed of 6 layers, each of which has a distinct role.
38
The cortex needs an efficient way of adapting the synaptic weights across these multiple
layers of feature-detecting neurons. BPTT is a possible model of how neural network could
learn multiple layers of representations, however it is only suited to classify labeled training
data, and supervised learning.
For this reason, a neural network that contains both bottom-up recognition connections
and top down generative connections is needed. Bottom-up connections are used to recognize
the input data, whereas top down generative connections fulfill their roles by generating
a whole distribution of data vectors over recurrent passes. In other words, bottom-up
connections help in determining the activations for certain present features, and places
them in layers. Top-down connections try to generate the training data using the activations
learned from the bottom-up pass. The learning algorithm used for the top-down connections
is an inverted learning model using negative feedback loops[56]. These passes are done
while adjusting the weightings on the connections in order to maximize performance and
optimization.
This model is in line with several neuroscientific observations, including single cell
recordings, and the theory of reciprocal connectivity between cortical areas, which indicates
to a hierarchical progression of complexity in the cortex.
9.2
Training unlabeled data
The external data we receive as humans are random. Most of the time, there are no labeled
data being presented to us. We aren’t able to manually structure our brains to perform a
certain task. In terms of modeling, the goal is translated as finding hidden structures in
unlabeled data. This problem boils down to probabilistic density estimation. A child holds
an apple, and hears it being referred to as ”Apple” several times, making the connection
that this object is indeed called an ”Apple”. A probabilistic distribution is inferred over
the various possible settings of the hidden variables, by doing several recurrent passes[27].
However, a neural network with only one hidden layer is too simplistic, and would not
be able to handle such rich and complex structures, nor would it be able to correctly deal
with the high-dimensional operations at task. One way to solve this problem is by stacking
restricted Boltzmann machines, in order to form a deep belief net. However this is not
optimal. The hidden variables use discrete values, making exact inference possible, but very
limited to the domain and learned scenarios, causing reduced generalization capacities.
Going down the probabilistic road, and simplifying our way to the lowest common
denominator, our problem boils down to using either generative models, or a discriminative
models. Generative models are full probabilistic models. They can be used to generate the
probabilistic prediction of any, and all variables in the model. They express complex, and
complete, relationships between the observed and target variable, and are able to categorize
data statistically, to determine the mostly likely input that would generate such a signal. The
other type of probabilistic modeling, discriminative models, only model the target variables
conditionally, in views of the observed data-sets. This is a limiting probabilistic distribution,
39
and does not give the operator insight into how the data is being generated. Its role is to
simply categorize a signal.
For a long time, the general consensus was that generative models were inferior to
their discriminative counterparts[36], however that turned out to be false. Generative models,
while they may take more time to resolve, have superior computation powers, due to their
transparency towards the full probabilistic distribution. A generative model can infer a
discriminative model from its data, whereas the opposite is not true.
Modeling the mind should therefore rely on unsupervised learning, using many hidden
layers in a non-linear distributive neural network. Networks build a progressively more
complex hierarchy using sensory data, by passing it through a generative model. In order to
solve the problem of reciprocal connectivity through top-down and bottom-up connections,
a good candidate to that are PCSPs. We will talk briefly about them, before moving onto a
larger concept which makes use of them.
9.3
Parallel constraint satisfaction processes
Parallel constraint satisfaction processes[65], PCSP for short, are a type of network used to
model cognitive dissonance, and other contemporary issues in social, and Gestalt psychology.
Activations pass around symmetrically connected nodes, until the activation of all the nodes
relaxes into a state. What we mean by all nodes relax is that all constraints are met when the
change percentage between the previous, and current state of each node asymptotes. This
final state satisfies all the constraints set among the nodes, which are connected with each
otehrs, and represent hypotheses detailing the presence or absence of certain features. These
connections serve as a ground to prove hypotheses as being either consistent within each
others, or inconsistent and contradictory.
Activations spread in parallel among the nodes. Nodes with positive links, or connections, will activate each others reciprocally. Nodes with negative links inhibit each others.
This results in a global and viable solution to the constraints among the entire set of nodes.
9.4
Part-whole hierarchies
Representing simple structures in a distributive representation network is complex. Objects
are not symbols, or singular entities, they are instead represented as patterns across many
units. It is even harder to represent complex structures, such as sentences, where the
meaning is composed of several constituents, with relations interweaving across several
of it sub-elements. Attentional focus must therefore be brought on the constituents of that
structure, while at the same time maintaining the whole meaning in check. To deal with that
problem, one potential solution is using part-whole hierarchies, where objects at one level
are composed of inter-related objects at the next level down.
In the classical digital computer, hierarchical data structures are composed of a set of
40
fields, or variables, containing pointers to the content of the fields, their values. Addresses
are a small representation of the symbol they point to, and many symbols can be put together
to create a fully-articulated representation. This simple representation comes at the cost of
being limited by the Von Neumann bottleneck[5], in which the system runs out of hardware
space.
A famous account of disagreement between connectionists and symbolic architecture
advocates is that of Fodor and Pylyshyn[15]. They believed that the mind operates around
rule governed formulations, and the manipulation of sentences in an inner linguistic code.
Connectionism was seen as a step backwards, devoid of any real computation, resembling
associationism, where an input is associated with its output through trial and error. However,
as more progress was being made in both human cognition, and connectionism, many of
their points didn’t hold through. Neural networks can generalize to non test-case scenarios,
create an internal representation to express regularities in the domain, and use isomorphisms
in order to encode structures more efficiently. This indicates that distributed representations
have much more power than previously thought.
In a PDP network, patterns also allow remote access to fuller representations. Most
processing can be done using parallel constraint satisfaction processes, in which an activation
passes around symmetrically connected nodes, until the states satisfy the constraints set
among the nodes.
PCSP results in an inflexible inner loop, leading to two different ways of performing
inference[65].
9.4.1
Intuitive inferences, and rational inferences
Inferences are performed efficiently by allowing the neural network to settle down into a
stable state. This allows the network to reach a conclusive result. Inferences are influenced
by the network’s weightings, past knowledge, and connection strength. Some operations
require a single settling of the network to resolve, such cases are referred to as intuitive
inferences.
Other operations require a more serial approach, where several intuitive inferences
are performed in sequence, this is known as rational inferences. Rational inferences involve
sequences of changes in the way in which parts of the task are mapped into the network. It
is important to note that a single task can sometimes be performed in different ways, either
requiring intuitive inference, or rational inference, depending on the degree of complexity
of its iterations.
9.4.2
Internal network structure
In order to build a part-whole hierarchy, it is possible to map out the internal representation
of the network by hand. The algorithm then sorts the inputs, applies the necessary steps, and
derives conclusions using our base framework. It is also possible to give the network the
41
freedom to use its experience of a set of propositions to construct its own internal representation of concepts. Propositions are presented in a neutral way, and the network translates the
input into active units, using its own associations. Similar patterns of activity will therefore
represent similar relations, and the representation will be built around correlations between
the given input. This leads to better generalization due to the network searching for features
that make it simple to express the regularities of the domain.
Instead of using a separate bit for each possible object that could be stored in the
memory, each bit is shared between many objects. A bit is applied to one position of the
object, the combination of its identity and position within the object activates a role specific
unit that contributes to the recognition of the object. The recognition model is shared across
the entities, within one level. This model is called within-level timesharing[45].
In order to share knowledge in the connections between different levels in the partwhole hierarchy, it is necessary to implement flexible mapping of pieces of the task into a
module of the network. This is done by choosing a node as the current whole, the other units
are devoted to either describing its global properties, or describing its major constituents.
The pattern of activity that represents a same object is different depending on the whole, or
context, it was presented in[28]. This implementation readily violates the assumption that
each entity in the world has a unique corresponding representation in the network, which
was an essential part of classical symbolic architectures. One system that hits a lot of these
points, and provides a good biologically plausible architecture for computing is HTMs.
9.5
Hierarchical Temporal Memory
Hierarchical Temporal Memory[22], HTM for short, is a bio-inspired machine learning
model, that makes use of distributed representations to store information. Its main goal is
modeling the neocortex, using nodes that are arranged in columns, layers, regions, and in a
hierarchy.
It is compromised of two separate stages, a spatial pooler, and a temporal pooler. The spatial
pooler creates a sparse binary representation from the network’s input. The temporal pooler
makes predictions, that are not provided in the input, in response to the vector sequences.
The predictions can range from next step predictions, to predictions into the future several
steps ahead.
The model represents dendrites using non-linear segments, and mimics them using a
sum and threshold function. These structures connect with other neurons, or nodes, forming
an equivalent of a synapse, leading a node to fire when enough activations have been received
by any segments, through an OR operator. The segments, and their weights can be learned,
and changed adaptively with the received inputs. Their number for a given node, as well as
their total number in the structure can change with time to accommodate machine learning.
HTMs are trained on large sets of patterns and sequences, where information is stored
in a distributed fashion, following what was described previously. The network has the
42
freedom to choose how, and where the information is stored, making for better generalization.
HTMs have VSA features[58], and are able to achieve bundling, by storing several objects
in one vector. They are also able to achieve variable binding, where a role can be used to
obtain the values of the filler.
10
Constructing the Code: A recap
Most of the work that has been carried out in the field of artificial neural networks, and
presented here, has targeted very specific features. From language processing, to visual
imagery, and models of memory, none of these models encompass the entirety of the mind.
Using the limited knowledge presented in this review, we will attempt to make our own
model of a general brain architecture. As we have previously seen, the brain is limited by
several factors, working memory is one example of them. In the future we may want to
exceed them, but for now, to insure a one to one scale, we will work within these boundaries.
Symbolic and localist architectures are not flexible enough, and would not be able to
deal with the large amounts of data that our brains are presented with on a daily basis. For
this reason, our model will be using a distributed representation. Entities will be represented
by patterns, and spread across many nodes. Variable binding is an essential part of human
cognition, but it used to be a challenge to deal with in distributed connectionist models.
However vector symbolic architectures have proved that they can handle variable binding
to a fairly decent degree. In order to encode enough features, and have enough robustness
and plasticity in our networks, we will opt to use high-dimensional vectors. The highdimensional vectors used should all be of the same dimension, to enable us to use the wide
array of pre-existing, and thoroughly tested mathematical knowledge on them. Our brains
are interconnected, with constant signal propagation. We will therefore model them using a
recurrent system of artificial neural networks.
Although there is a large number of neuron types in the brain, most of them are due
to their anatomical, morphological and chemical differences. The two types of neurons
that absolutely have to be accounted for, are the classical specificity neurons, and mixed
selectivity neurons. Classical specificity neurons are simple, and integrate activity coming
from different regions in a linear way, in order to activate, or inhibit a certain structure. They
respond specifically to one stimulus or task. Unfortunately, currently, there is not enough
data to determine exactly where the classical neurons reside in the brain, and exactly what
are their advantages over mixed selectivity neurons. For this purpose, and the purpose of
simplification, we will assume that classical neurons are used exclusively for the input, and
output of data. They will be modeled using linear nodes, and will help transfer data into,
and out of the neural architecture.
Mixed selectivity neurons are overwhelmingly present in higher-order structures. They
have complex, and non-linear activation patterns. These are the neurons responsible for most
of the logical computation done in our brain. Therefore, they will be modeled as hidden
43
nodes, with non-linear activation functions. Our input and output layers will be classical
selectivity neurons, represented by linear nodes. Our hidden layers will be using mixed
selectivity neurons, represented by non-linear nodes. Nodes will use a sum and threshold
function, through a leaky integrate and fire model. They will be connected with each others
by links, playing the role of synapses. The number of links will be variable, and subject to
machine learning. New connections can be created or dismantled after long-term training.
Inserting gating nodes seems an essential part of any neural network. We do not want
oscillating weights. We do not want any gradients to vanish, or explode. And most
importantly, we do not want the network to be heavily influenced by recent inputs and
modifications. However, we doubt there are gating neurons in the brain. As a result, all
nodes, whether they are gates, memory cells, or regular neurons, will have to be identical. It
is therefore necessary to use flexible mappings between the entities. Each node will play the
role of the whole, and all the units connecting to it will be used to gate information in, and
out of it. Each node will have its own set of representations, and a representation of a certain
object will be different between two nodes. Using these characteristics, it is theoretically
possible to have each node behave differently in response to the signal it is receiving. A
node will know if it is being used as a gate node, or computing node, or both at the same
time in response to different clusters of nodes, based on the input pattern it has received,
and based on its previous activations. However, how this will be achieved mathematically
remains to be seen.
The cortex is divided into separate areas, and the concept that each area controls a
more or less specific functions has been tested several time, and is agreed upon. In all
humans, the visual cortex is located in the Brodmann areas 17, 18, 19, while the auditory
cortex is located in Brodmann areas 41 and 42. However we believe that this is not due to a
special internal structure of the brain. This is simply due to how exterior stimuli is fed into
the brain following its specific paths. Although not discussed here, these areas can change if
the signal was rerouted from one structure to others[59].
But one characteristic that is constant, for a seemingly unwarranted reason, is the distinction
between the cortex, and the central gray nuclei(CGNs). CGNs are an interconnected
set of neurons that compile activations from different areas of the brain, perform their
computations, then reroute them to other structures. It is hard to believe that such unique,
yet consistent elements would arise through training. We are led to believe that there is an
internal predisposition to such a structure, and we should therefore enforce it into our model.
The cerebral cortex has a laminar structure, and is composed of 6 layers forming
microcircuits. These microcircuits are grouped into cortical minicolumns. Each layer
has special characteristic connections with the others. This can be seen rather constantly
through the cortex. For this purpose, we will be stacking unrestricted Boltzmann machines
to model this architecture. Unrestricted Boltzmann Machines are identical to RBMs, with
the exception of the presence of connections between hidden units. Central grey nuclei will
be modeled using fully interconnected, or highly interconnected recurrent neural networks.
44
The connections should be flexible and differentiable end to end. The network will be
trained through unsupervised learning, using unlabeled data, in ways comparable to how
the brain would receive its data. Because there are generally no restrictions on the network,
the network will construct its own internal representation. Similar entities will end up with
similar vectors.
In order to effectively deal with the extremely large amounts of data that a model such
as the brain is bombarded with on a daily basis, we will be using Holographic Reduced
Representations to conserve memory space. A variable, and its value will be bound together
through circular convolution, in order to form one vector. Vectors can be compressed
together to sequence them, or indicate similar entities. This will also allow us to maintain
the sequential recall order for complex tasks. In order to save time, and be able to increment
steps at a higher rate, we will be using approximate operations. This includes using FFT
instead of circular convolution on HRRs. These approximations will require a clean-up
memory, so an autoassociative storage method will be employed.
As we have seen throughout this review, tremendous ground has been covered, and astonishing work has been done so far in the field of computational neuroscience. How far we will
be able to go in the following years remains to be seen. The path to the future is still very
much obscure, but with the recent surge of interest in the concept of mind modeling, and
the combined efforts of researchers from several subspecialities, this goal will definitely be
within our reach in the near future.
45
References
[1]: Ackley, David H., Geoffrey E. Hinton, and Terrence J. Sejnowski. ”A learning algorithm
for boltzmann machines.” Cognitive science 9.1 (1985): 147-169.
[2]: Alexander, Garrett E., Mahlon R. DeLong, and Peter L. Strick. ”Parallel organization
of functionally segregated circuits linking basal ganglia and cortex.” Annual review of
neuroscience 9.1 (1986): 357-381.
[3]: Asaad, Wael F., Gregor Rainer, and Earl K. Miller. ”Neural activity in the primate
prefrontal cortex during associative learning.” Neuron 21.6 (1998): 1399-1407.
[4]: Atallah, Hisham E., Michael J. Frank, and Randall C. O’Reilly. ”Hippocampus, cortex,
and basal ganglia: Insights from computational models of complementary learning systems.”
Neurobiology of learning and memory 82.3 (2004): 253-267.
[5]: Bacus, J. ”Can programming be liberated from the von Neuman style.” Comm. ACM
21 (1978): 899.
[6]: Bellman, Richard, et al. Adaptive control processes: a guided tour. Vol. 4. Princeton:
Princeton university press, 1961.
[7]: Bengio, Yoshua, Patrice Simard, and Paolo Frasconi. ”Learning long-term dependencies
with gradient descent is difficult.” Neural Networks, IEEE Transactions on 5.2 (1994): 157166.
[8]: Bengio, Yoshua, and Samy Bengio. ”Modeling High-Dimensional Discrete Data with
Multi-Layer Neural Networks.” NIPS. Vol. 99. 1999.
[9]: Caid, William R., Susan T. Dumais, and Stephen I. Gallant. ”Learned vector-space
models for document retrieval.” Information processing and management31.3 (1995): 419429.
[10]: Collobert, Ronan, et al. ”Natural language processing (almost) from scratch.”The
Journal of Machine Learning Research 12 (2011): 2493-2537.
[11]: Das, Sreerupa, C. Lee Giles, and Guo-Zheng Sun. ”Learning context-free grammars:
Capabilities and limitations of a recurrent neural network with an external stack memory.”
Proceedings of The Fourteenth Annual Conference of Cognitive Science Society. Indiana
University. 1992.
[12]: Desjardins, Guillaume, et al. ”Tempered Markov chain Monte Carlo for training of
restricted Boltzmann machines.” International Conference on Artificial Intelligence and
Statistics. 2010.
[13]: Donoho, David L. ”High-dimensional data analysis: The curses and blessings of
dimensionality.” AMS Math Challenges Lecture (2000): 1-32.
[14]: Eich, Janet M. ”A composite holographic associative recall model.”Psychological
Review 89.6 (1982): 627.
46
[15]: Fodor Jerry A., and Zenon W. Pylyshyn. ”Connectionism and cognitive architecture:
A critical analysis.” Cognition 28.1 (1988): 3-71.
[16]: Gallant, Stephen I., and T. Wendy Okaywe. ”Representing objects, relations, and
sequences.” Neural computation 25.8 (2013): 2038-2078.
[17]: Gers, Felix A., Jrgen Schmidhuber, and Fred Cummins. ”Learning to forget: Continual
prediction with LSTM.” Neural computation 12.10 (2000): 2451-2471.
[18]: Gers, Felix, and Jrgen Schmidhuber. ”Recurrent nets that time and count.”Neural
Networks, 2000. IJCNN 2000, Proceedings of the IEEE-INNS-ENNS International Joint
Conference on. Vol. 3. IEEE, 2000.
[19]: Giles, C. Lee, et al. ”Learning and extracting finite state automata with second-order
recurrent neural networks.” Neural Computation 4.3 (1992): 393-405.
[20]: Graves Alex, Greg Wayne, and Ivo Danihelka. ”Neural Turing Machines.” arXiv
preprint arXiv:1410.5401 (2014).
[21]: Hadley, Robert F. ”The problem of rapid variable creation.” Neural computation21.2
(2009): 510-532.
[22]: Hawkins, Jeff, Subutai Ahmad, and Donna Dubinsky. ”Hierarchical temporal memory including HTM cortical learning algorithms.” Techical report, Numenta, Inc, Palto
Alto(2010).
[23]: Hazy, Thomas E., Michael J. Frank, and Randall C. OReilly. ”Banishing the homunculus: making working memory work.” Neuroscience 139.1 (2006): 105-118.
[24]: Hebb, Donald Olding. The organization of behavior: A neuropsychological theory.
Psychology Press, 2005.
[25]: Hernndez-Lpez, Salvador, et al. ”D1 receptor activation enhances evoked discharge in
neostriatal medium spiny neurons by modulating an L-type Ca2+ conductance.” The Journal
of neuroscience 17.9 (1997): 3334-3342.
[26]: Hinton, Geoffrey E. ”Learning distributed representations of concepts.”Proceedings of
the eighth annual conference of the cognitive science society. Vol. 1. 1986.
[27]: Hinton, Geoffrey E. ”Learning multiple layers of representation.” Trends in cognitive
sciences 11.10 (2007): 428-434.
[28]: Hinton, Geoffrey E. ”Mapping part-whole hierarchies into connectionist networks.”
Artificial Intelligence 46.1 (1990): 47-75.
[29]: Hochreiter, Sepp, and Jrgen Schmidhuber. ”Long short-term memory.” Neural computation 9.8 (1997): 1735-1780.
47
[30]: Hodgkin, Alan L., and Andrew F. Huxley. ”A quantitative description of membrane
current and its application to conduction and excitation in nerve.”The Journal of physiology
117.4 (1952): 500-544.
[31]: Holzer, Adrian, Patrick Eugster, and Benot Garbinato. ”Evaluating implementation
strategies for location-based multicast addressing.” Mobile Computing, IEEE Transactions
on 12.5 (2013): 855-867.
[32]: Hopfield, John J. ”Neural networks and physical systems with emergent collective
computational abilities.” Proceedings of the national academy of sciences 79.8 (1982):
2554-2558.
[33]: Hsu, Yuan-Yih, and Chien-Chuen Yang. ”A hybrid artificial neural network-dynamic
programming approach for feeder capacitor scheduling.” Power Systems, IEEE Transactions
on 9.2 (1994): 1069-1075.
[34]: Hulme, Charles, et al. ”The role of long-term memory mechanisms in memory span.”
British Journal of Psychology 86 (1995): 527.
[35]: Ingster, Yuri I., Christophe Pouet, and Alexandre B. Tsybakov. ”Classification of sparse
high-dimensional vectors.” Philosophical Transactions of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences367.1906 (2009): 4427-4448.
[36]: Jordan, A. ”On discriminative vs. generative classifiers: A comparison of logistic
regression and naive bayes.” Advances in neural information processing systems 14 (2002):
841.
[37]: Kanerva, Pentti. ”Hyperdimensional computing: An introduction to computing in
distributed representation with high-dimensional random vectors.” Cognitive Computation
1.2 (2009): 139-159.
[38]: Koechlin, Etienne, et al. ”Dissociating the role of the medial and lateral anterior
prefrontal cortex in human planning.” Proceedings of the National Academy of Sciences
97.13 (2000): 7651-7656.
[39]: Kohonen, Teuvo. ”Correlation matrix memories.” Computers, IEEE Transactions on
100.4 (1972): 353-359.
[40]: Kosko, Bart. ”Bidirectional associative memories.” Systems, Man and Cybernetics,
IEEE Transactions on 18.1 (1988): 49-60.
[41]: Levitt, Jonathan B., et al. ”Topography of pyramidal neuron intrinsic connections in
macaque monkey prefrontal cortex (areas 9 and 46).” Journal of Comparative Neurology
338.3 (1993): 360-376.
[42]: Maass, Anne, et al. ”Laminar activity in the hippocampus and entorhinal cortex related
to novelty and episodic encoding.” Nature communications 5 (2014).
48
[43]: Malsburg, Cvd. ”The correlation theory of brain function.” Internal Report MPI fr
biophysikalische Chemie 81 (1981).
[44]: Marcus, G. F. ”The algebraic mind: Reections on connectionism and cognitive science.”
(2000). Cambridge, MA: MIT Press.
[45]: McClelland, James L., and David E. Rumelhart. ”An interactive activation model of
context effects in letter perception: I. An account of basic findings.”Psychological review
88.5 (1981): 375.
[46]: McClelland, James L., David E. Rumelhart, and Geoffrey E. Hinton. ”The appeal of
parallel distributed processing.” Cambridge, MA: MIT Press, 1986.
[47]: McCulloch, Warren S., and Walter Pitts. ”A logical calculus of the ideas immanent in
nervous activity.” The bulletin of mathematical biophysics 5.4 (1943): 115-133.
[48]: McNab Fiona, and Torkel Klingberg. ”Prefrontal cortex and basal ganglia control
access to working memory.” Nature neuroscience 11.1 (2008): 103-107.
[49]: Miller, George A. ”The magical number seven, plus or minus two: some limits on our
capacity for processing information.” Psychological review 63.2 (1956): 81.
[50]: Mink, Jonathan W. ”The basal ganglia: focused selection and inhibition of competing
motor programs.” Progress in neurobiology 50.4 (1996): 381-425.
[51]: Monner, Derek, and James A. Reggia. ”A generalized LSTM-like training algorithm
for second-order recurrent neural networks.” Neural Networks 25 (2012): 70-83.
[52]: Murdock, Bennet B. ”A distributed memory model for serial-order information.”Psychological Review 90.4 (1983): 316.
[53]: Murdock, Bennet B. ”Serial-order effects in a distributed-memory model.” (1987). In
David S. Gorfein and Robert R. Ho man, editors, memory and learninig. The Ebbinghaus
Centennial Conference, pages 277310. Lawrence Erlbaum Associates
[54]: Neumann, Jane. ”Learning the systematic transformation of holographic reduced
representations.” Cognitive Systems Research 3.2 (2002): 227-235.
[55]: Newell, Allen. ”Physical Symbol Systems*.” Cognitive science 4.2 (1980): 135-183.
[56]: Oh Jong-Hoon, and H. Sebastian Seung. ”Learning Generative Models with the Up
Propagation Algorithm.” Advances in Neural Information Processing Systems. 1998.
[57]: O’Reilly, Randall C., et al. ”PVLV: the primary value and learned value Pavlovian
learning algorithm.” Behavioral neuroscience 121.1 (2007): 31.
[58]: Padilla, Daniel E., and Mark D. McDonnell. ”A neurobiologically plausible vector
symbolic architecture.” Semantic Computing (ICSC), 2014 IEEE International Conference
on. IEEE, 2014.
49
[59]: Pallas, Sarah L., Anna W. Roe, and Mriganka Sur. ”Visual projections induced into
the auditory pathway of ferrets. I. Novel inputs to primary auditory cortex (AI) from the
LP/pulvinar complex and the topography of the MGNAI projection.” Journal of Comparative
Neurology 298.1 (1990): 50-68.
[60]: Pineda, Fernando J. ”Dynamics and architecture for neural computation.”Journal of
Complexity 4.3 (1988): 216-245.
[61]: Plate, Tony ”Estimating analogical similarity by dot-products of Holographic Reduced
Representations.” Advances in neural information processing systems (1994): 1109-1109.
[62]: Plate, Tony. ”Holographic Reduced Representations: Convolution Algebra for Compositional Distributed Representations.” IJCAI. 1991.
[63]: Pollack, Jordan B. ”On connectionist models of natural language processing.” Computing Research Laboratory, New Mexico State University, 1987.
[64]: Pollack, Jordan B. ”Recursive distributed representations.” Artificial Intelligence46.1
(1990): 77-105.
[65]: Read Stephen J., Eric J. Vanman, and Lynn C. Miller. ”Connectionism, parallel
constraint satisfaction processes, and gestalt principles:(Re) introducing cognitive dynamics
to social psychology.” Personality and Social Psychology Review 1.1 (1997): 26-53.
[66]: Rigotti, Mattia, et al. ”Internal representation of task rules by recurrent dynamics: the
importance of the diversity of neural responses.” Frontiers in computational neuroscience 4
(2010).
[67]: Rigotti, Mattia, et al. ”The importance of mixed selectivity in complex cognitive tasks.”
Nature 497.7451 (2013): 585-590.
[68]: Robinson, A. J., and Frank Fallside. The utility driven dynamic error propagation
network. University of Cambridge Department of Engineering, 1987.
[69]: Rosenfeld, Ronald, and David S. Touretzky. ”Coarse-coded symbol memories and
their properties.” Complex Systems 2.4 (1988): 463-484.
[70]: Roweis, Sam. ”Boltzmann machines.” lecture notes (1995).
[71]: Rudolph-Lilith, Michelle, and Lyle E. Muller. ”Aspects of randomness in biological
neural graph structures.” BMC Neuroscience 14.Suppl 1 (2013): P284.
[72]: Schultz, Wolfram. ”Predictive reward signal of dopamine neurons.” Journal of
neurophysiology 80.1 (1998): 1-27.
[73]: Seghier, Mohamed L. ”The angular gyrus multiple functions and multiple subdivisions.”
The Neuroscientist 19.1 (2013): 43-61.
[74]: Siegelmann, Hava T., and Eduardo D. Sontag. ”On the computational power of neural
nets.” Journal of computer and system sciences 50.1 (1995): 132-150.
50
[75]: Smolensky, Paul. ”On the proper treatment of connectionism.” Behavioral and brain
sciences 11.01 (1988): 1-23.
[76]: Smolensky, Paul. ”Parallel distributed processing: explorations in the microstructure
of cognition, vol. 1. chapter Information processing in dynamical systems: foundations of
harmony theory.” MIT Press, Cambridge, MA, USA 15 (1986): 18.
[77]: Smolensky, Paul. ”Tensor product variable binding and the representation of symbolic
structures in connectionist systems.” Artificial intelligence 46.1 (1990): 159-216.
[78]: Sowell, Elizabeth R., et al. ”Mapping cortical change across the human life span.”
Nature neuroscience 6.3 (2003): 309-315.
[79]: Touretzky, David S., and Geoffrey E. Hinton. ”A distributed connectionist production
system.” Cognitive Science 12.3 (1988): 423-466.
[80]: Treisman, Anne. ”Solutions to the binding problem: progress through controversy and
convergence.” Neuron 24.1 (1999): 105-125.
[81]: Turing, Alan Mathison. ”On computable numbers, with an application to the Entscheidungsproblem.” J. of Math 58.345-363 (1936): 5.
[82]: Turing, Alan Mathison. ”On computable numbers, with an application to the Entscheidungsproblem. A correction.” Proceedings of the London Mathematical Society 2.1 (1938):
544-546.
[83]: Verleysen, Michel. ”Learning high-dimensional data.” Nato Science Series Sub Series
III Computer And Systems Sciences 186 (2003): 141-162.
[84]: Von Neumann, John. ”First Draft of a Report on the EDVAC.” IEEE Annals of the
History of Computing 4 (1993): 27-75.
[85]: Wagemans, Johan, et al. ”A century of Gestalt psychology in visual perception: II.
Conceptual and theoretical foundations.” Psychological bulletin 138.6 (2012): 1218.
[86]: Werbos, Paul J. ”Generalization of backpropagation with application to a recurrent gas
market model.” Neural Networks 1.4 (1988): 339-356.
[87]: Weston, Jason, Sumit Chopra, and Antoine Bordes. ”Memory networks.” arXiv
preprint arXiv:1410.3916 (2014).
[88]: Williams, Ronald J., and David Zipser. ”A learning algorithm for continually running
fully recurrent neural networks.” Neural computation 1.2 (1989): 270-280.
[89]: Willshaw, David. ”Holography, associative memory, and inductive generalization.”
(1985).
[90]: Wolpert, D. ”A computationally universal field computer wich is purely linear.”
Relatrio Tcnico. LA-UR-91-2937, Los Alamos National Laboratory, 1991.
51
[91]: Yarom, Yosef, and Jorn Hounsgaard. ”Voltage fluctuations in neurons: signal or
noise?.” Physiological Reviews 91.3 (2011): 917-929.
52
| 9cs.NE
|
Optimal Hybrid Full-Duplex/Half-Duplex Scheme
for Buffer Aided Relay Systems
Cheng Li∗ , Bin Xia∗ , Pihe Hu∗ Yao Yao‡
∗ Department
arXiv:1802.06001v2 [cs.IT] 5 Mar 2018
of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, China
‡ Huawei Technologies Co. Ltd
Emails: {lichengg, bxia, hupihe}@sjtu.edu.cn, [email protected]
Abstract—Full-duplex (FD) communication has received great
interest in recent years due to the potential of doubling the
spectral efficiency. However, how to alleviate the detrimental
effects of the residual self-interference (RSI) incurred by the FD
mode is still a challenging problem. In this paper, focusing on
the statistical throughput maximization, we propose an optimal
hybrid FD/half-duplex (HD) scheme for the one-way FD buffer
aided relay system. To solve this problem, we divide the system
into four different transmission modes and formulate the problem
as a binary integer programming problem. By relaxing the binary
variables to be continuous ones, we solve the problem using the
Karush-Kuhn-Tucker (KKT) optimal conditions. We obtain the
selection probability of each mode based on the instantaneous
channel outage states. The proposed scheme not only achieves
the optimal FD or HD mode selection, but also realizes adaptive
source-to-relay or relay-to-destination link selection. Simulation
results show that the proposed scheme offers 95% maximum gain
over the HD counterparts.
I. I NTRODUCTION
Deploying relay nodes in networks is an efficient way to
improve the quality of service of cell edge users [1]. Field
measurements have shown that relay nodes can improve the
system throughput and enhance the outdoor to indoor coverage
[2]. Owing to the various advantages, relay communications
have received great interest. For instance, in [3], the outage
performance was studied for the bidirectional half-duplex
(HD) decode-and-forward (DF) relay systems. Recently, the
full-duplex (FD) mode, which receives and transmits in the
same frequency band simultaneously, was proposed [4]. In [5],
the authors showed that the FD mode could efficiently improve
the performance of the relay systems over the HD counterparts.
However, the performance gain was still greatly limited by the
residual self-interference (RSI) after self-interference cancellation. In [6], the authors analyzed the outage performance of
the two-way FD relay system with RSI. The results revealed
that the HD counterparts outperformed the FD system in the
high RSI regime.
To overcome the detrimental effects of the RSI, the authors
in [7] have proposed a hybrid HD/FD scheme based on the
instantaneous transmission rate. However, the authors did not
consider the buffer at the relay node. The received packets
needed to be forwarded immediately. Hence, the end-to-end
performance was limited by the worse one of the S-R and
R-D links. In [8], the authors have proposed an adaptive link
selection protocol for the traditional HD relay system with a
buffer. The results revealed that the buffer could efficiently
improve the end-to-end throughput. Intuitively, to deploy FD
mode in the aforementioned systems will improve the system
performance as it bears the potential to double the spectral
efficiency. However, we cannot simply extend the previous
schemes to the FD cases due to the simultaneous transmission
and reception. In addition, the RSI, which is a critical factor
that limits the system performance, should be carefully treated
as well. The relaying scheme and the hybrid FD/HD mode
need to be re-designed to exploit the maximum potential of
the FD mode. To the best of the authors’ knowledge, few
works has been done on this topic.
To utilize the advantages of the buffer, in [9], [10], the
authors have designed the hybrid FD/HD protocols to improve
the system throughput. In these protocol the FD mode is
always preferred whenever it is available. In fact, we find that
the FD mode is not always preferred. In addition, in [10],
the HD and FD mode may be both possible. In this case, we
design the selection probabilities to select different modes.
In this paper, we are dedicated to design a hybrid FD/HD
protocol to maximize the system statistic throughput over the
infinite time horizon for the one-way FD DF relay system.
To solve this problem, we first divide the system into four
different transmission modes and then formulate the problem
as a binary integer programming problem. By relaxing the
binary variables to be continuous ones, we solve the problem
using the Karush-Kuhn-Tucker (KKT) optimal conditions. The
proposed scheme not only achieves the optimal FD or HD
mode selection but also realizes adaptive S-R or R-D link
selection. In addition, the general expressions of the system
throughput is derived. Simulation results validate that the
proposed scheme effectively overcomes the detrimental effects
of the RSI and offers 95% maximum throughput gain over the
HD counterparts.
II. S YSTEM M ODEL
In this section, we describe the system model, including the
channel mode, basic transmission modes and different channel
gain regions.
A. Channel Model
The system model is depicted in Fig. 1. The direct link
between S and D does not exist. The relay node has the
capability to work in the full-duplex (FD) mode and is
equipped with a buffer B. In addition, we denote the channel
K
J
K
6
5
K
'
J
Fig. 1. The three-node one-way FD DF relay system model.
coefficients of the S-R and R-D links as h1 (i) and h2 (i) in
the i-th time slot, respectively. The block fading channels are
considered. We assume that |h1 (i)| and |h2 (i)| are subject to
the independent, stationary and ergodic random distributions.
All the channels are contaminated by the thermal noise. When
the relay node works in the FD mode, the S to R link
will be impaired by the RSI as well after self-interference
cancellation. We use γ1F (i), γ1H (i) and γ2 (i) to denote the
signal-to-interference-plus-noise ratios (SINRs) of the S-R
link in the FD mode, HD mode and the R-D link, respectively,
which are given by
γ1F (i) =
P1 g1 (i)
P1 g1 (i)
P2 g2 (i)
, γ H (i) =
, γ2 (i) =
, (1)
IR + σr2 1
σr2
σd2
where P1 and P2 denote the transmit powers of S and R,
respectively. g1 (i) = |h1 (i)|2 and g2 (i) = |h2 (i)|2 denote the
channel gains of the S-R link and R-D link, respectively.
σr2 and σd2 denote the variances of the noise at S and D,
respectively. IR = KR P2 , where KR is the indicator of the
RSI level. In this paper, we consider the fixed rate transmission
scenario with rate R0 , which has been considered in various
applications, such as packet transmission. In addition, we
assume that the relay node knows the global CSI.
B. Different Transmission Modes
In the system, there exists four different transmission modes,
denoted by M1 , M2 , M3 , M4 . To be specific, in the mode M1 ,
the source S sends messages to the relay R, but R keeps silent
and the messages are stored in the buffer B. In the mode M2 ,
the source S keeps silent, but the relay R extracts messages
from the buffer B and forwards them to the destination D. In
the mode M3 , the system works in the FD mode, i.e., S and
R transmit signals simultaneously.
In order to specify which mode is selected, we let dj (i) ∈
{0, 1} to serve as the mode indicator. For instance, when the
mode Mj is selected in the i-th time slot, we set dj (i) to 1,
otherwise we set dj (i) to 0. In each
P4 time slot, only one mode
can be selected, we obtain that j=1 dj (i) = 1. In order to
decide whether the information can be successfully decoded
in the mode Mj , we adopt Oj (i) ∈ {0, 1}, to serve as the
transmission indicator. If the transmission in the mode Mj is
successful, we set Oj (i) to 1, otherwise we set Oj (i) to 0.
For the mode M4 , since all the nodes keep silent, we always
set O4 = 0.
C. Different Channel Gain Regions
For better illustration, we divided the whole channel gain
plain into six un-overlapped regions, which are denoted by
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 005
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 05
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
5
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 005
0 0 0 00 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
5
5
J
J
J
Fig. 2. Different channel gain regions.
Rk , k ∈ {1, ..., 6} as shown in the Fig. 2, where g1H =
γ (I +σ2 )
γ0 σr2
P1
r
and g1F = 0 R
denote the outage threshold of the SP1
R link under the HD mode and FD mode, respectively. g2H =
γ0 σd2
P2 denotes the outage threshold of the R-D link. For clarity,
the relationships between the transmission indicators and the
different channel gain regions are summarized in the following
table.
TABLE I
V IABILITY OF D IFFERENT M ODES
Channel Gain Regions
O1
O2
O3
R1
1
1
1
R2
1
1
0
R3
0
1
0
R4
1
0
0
R5
1
0
0
R6
0
0
0
III. O PTIMAL H YBRID FD/HD M ODE S ELECTION S CHEME
In this section, we first formulate the system statistical
throughput maximization problem. Then, the optimal mode
selection scheme is presented.
A. Problem Formulation
In this paper, we desire to devise an optimal hybrid FD/HD
scheme to maximize the long-term system throughput. First,
the rate of the S-R link is given by
R1 =
N
1 X
[d1 (i)O2 (i) + d3 (i)O3 (i)] × R0 .
N →+∞ N
i=1
lim
(2)
Similarly, the maximum average rate of the R-D link is
given by
R2 =
N
1 X
[d2 (i)O2 (i) + d3 (i)O3 (i)] × R0 .
N →+∞ N
i=1
lim
(3)
In this paper, we consider a stationary buffer state at the
relay node. Hence, we have the following lemma.
Lemma 1. For a stationary stochastic system, the average
arrival rate needs to be smaller than the departure rate, i.e.,
R1 ≤ R2 .
(4)
Proof. According to the queuing theory, if R1 > R2 , the
queue in the buffer will statistically increase to infinity. Hence,
we obtain this lemma.
Lemma 2. The optimal hybrid FD/HD scheme maximizing the
system throughput is achieved when the queue at the buffer is
at the edge of non-absorb state, which can be expressed as
R1 = R2 .
(5)
Proof. According to the Lemma 1, R1 ≤ R2 . If R1 < R2 , the
system throughput will be limited by R1 . However, we can
choose the mode M1 in more time slots to increase the rate
R1 . Once R1 = R2 , R1 will cannot be increased furthermore,
otherwise, the system cannot keep stationary.
The system throughput can be quantified by the average
received bits of information at the destination node D in each
time slot. Hence, we have
T = R2 .
(6)
It is noted that we abandon the limitation min{Q(i −
1), R0 }. This can be interpreted as that there are only countable number of time slots that the queue length of the buffer
is less than R0 when the buffer is at the edge of non-absorb
state. In other words, the relay node always has enough data
stored in the buffer. Hence, when averaged on the infinite time
horizon, the effects of Q(i) < R0 can be neglected.
Now, based on Lemma 1 and Lemma 2, the considered
throughput maximization problem can be formulated as P1 ,
which is given by
max
dj (i)∈{0,1}
T
C1: R1 = R2 ,
s.t.
C2:
4
X
dj (i) = 1, ∀i
j=1
C3: dj (i) ∈ {0, 1}, ∀i, k
(7)
However, since the variable dj (i) can only be 1 or 0,
the original optimization problem P1 is a binary integer
programming problem, which is hard to solve. Alternatively,
we observe that if we relax the binary variable dj (i) to be
continuous one in the interval [0, 1], the original problem
will transfer to a linear programming problem. For the linear
programming problem, the optimal value always obtained at
the vertexes of the feasible set, i.e., dj (0) = 1 or 0, which is
consistent with the original problem [11]. The standard form
of the relaxed problem P2 is given by
min
dj (i)∈[0,1]
s.t.
−T
C1: R1 − R2 = 0,
C2:
4
X
dj (i) − 1 = 0, ∀ i
Due to that the optimal scheme relies on the statistic CSI,
we first present the following different statistic CSI cases
Ψ1 : PR3 ≤ PR4 + PR5 − PR1 − PR2 ,
Ψ2 : PR4 + PR5 − PR1 − PR2 < PR3 ≤ PR4 + PR5 − PR2 ,
Ψ3 : PR4 + PR5 − PR2 < PR3 ≤ PR4 + PR5 + PR2 ,
Ψ4 : PR4 + PR5 + PR2 < PR3 ≤ PR4 + PR5 + PR2 + PR1 ,
(9)
Ψ5 : PR3 > PR4 + PR5 + PR2 + PR1 ,
where PRk denotes the probability of the channel gain region
Rk . As the relay node has known the statistic information of
the channel variation, thus the probability of the regions from
R1 to R6 can be easily derived.
Theorem 1. The optimal hybrid FD/HD mode selection
scheme maximizing the system throughput of the considered
buffer aided FD relay system with RSI is given in the Table
II.
where Pkj , k ∈ {1, 2, 3, 4, 5, 6}, j ∈ {1, 2, 3, 4} denotes the
selection probability of the mode Mj in the region Rk . In
the region R6 , all the active modes are unaccessible, only the
inactive mode M4 can be selected, i.e., P64 = 1. It is noted that
the selection probabilities, which equals to 0, are not given.
Proof. See Appendix.
Remark 1. The five non-overlapping statistic CSI cases is
proposed in the design of the optimal scheme (please see
the Appendix). The five cases represent five different relative
qualities of the S-R and R-D links. In the case Ψ1 , the S-R
link is far stronger than the R-D link. Hence, all the time
slots in the region R1 need to be allocated to the mode M2
to balance the constraint C1. In the case Ψ2 , the S-R link is
slightly stronger than the R-D link. We only need to allocate
a part of the time slots in the region R1 to the mode M2 . In
the case Ψ3 , the S-R link and the R-D link are comparable.
By allocating the time slots in the R2 to the mode M1 or
M2 could make the constraint C1 satisfied. Thus, all the time
slots in the region R1 could be allocated to the mode M3 . In
the case Ψ4 and Ψ5 , the situation will be contrary, more time
slots need to be allocated to the mode M1 .
Remark 2. Theorem 1 validates that the FD mode is not
always preferred even it is available. In addition, in the case of
Ψ1 and Ψ5 , the FD mode M3 has never been selected, which
means that the system always works in the pure HD mode. In
other words, although the relay node has the capacity to work
in the FD mode, the system cannot achieve any performance
gain due to the strong disparities of the qualities of the S-R
link and the R-D link.
C. General Expressions of the System Throughput
j=1
C3: dj (i) − 1 ≤ 0, ∀ i, k
C4: − dj (i) ≤ 0, ∀ i, k
B. Optimal Hybrid FD/HD Scheme
(8)
In this subsection, we present the general expressions of the
system throughput of the proposed hybrid FD/HD scheme.
TABLE II
V IABILITY OF D IFFERENT M ODES WITH D IFFERENT SINR REGIONS
R1
P12
Ψ1
P12 =
Ψ2
R2
P22
=1
PR4 +PR5 −PR2 −PR3
PR 1
P13 = 1 − P12
P11 =
Ψ4
P21 = 1
P11 = 1
1
R4 ∪ R5
PR1 +PR2 +PR3
PR4 +PR5
P44 = 1 − P41
P41 =
=1
P32 = 1
P41 = 1
P32 = 1
P41 = 1
P31 = 1
P41 = 1
PR1 +PR2 +PR4 +R5
PR 3
P34 = 1 − P32
1
10
Ψ
0
1
−2
10
−3
10
Pro_Reg_R1
Pro_Reg_R2
Pro_Reg_R3
Pro_Reg_R4R5
Pro_Reg_R6
−4
10
−5
1
2
3
4
5
6
7
8
Fixed transmission rate R (bps/Hz)
0
10
−1
10
−2
10
−3
10
Pro_Reg_R1
Pro_Reg_R2
Pro_Reg_R3
Pro_Reg_R4R5
Pro_Reg_R6
−4
10
−5
9
10
Probabilities of different regions Rk
k
Probabilities of different regions R
Ψ
10
10
Ψ5
4
3
3
0
Ψ
Ψ
Ψ
1
10
−1
P41 = 1
1
10
Ψ2
Ψ3
Probabilities of different regions Rk
P32 =
P21 = 1
10
10
PR2 +PR4 +PR5 −PR3
2PR2
P23 = 1 − P22
PR3 −PR2 −PR4 −R5
PR 1
P13 = 1 − P11
Ψ5
=1
P22 = 1
P22 =
P13 = 1
Ψ3
R3
P32
1
0
2
10
−1
10
−2
10
−3
10
Pro_Reg_R1
Pro_Reg_R2
Pro_Reg_R3
Pro_Reg_R4R5
Pro_Reg_R6
−4
10
−5
3
4
5
6
7
8
Fixed transmission rate R (bps/Hz)
9
10
10
1
0
2
3
4
5
6
7
8
Fixed transmission rate R (bps/Hz)
9
10
0
(a)
(b)
(c)
Fig. 3. Probabilities vs. the fixed transmission rate R0 under different system setups, in which 1) S-R is stronger than R-D link; 2) S-R link is equivalently
equal to the R to D link; 3) S-R link is weaker than R-D link.
Proof. Theorem 2 can be easily proved using the different
selection probabilities shown in the Theorem 1. The system
average throughput can be generally obtained by
T =
[PR1 (P12
+
P13 )
+
PR2 P22
+
PR3 P32 ]R0 .
(11)
6
System throughput T (bps/Hz)
Theorem 2. The general expressions of the maximum system
throughput of the proposed hybrid FD/HD scheme corresponding to different statistic CSI cases are given by
Ψ1 , Ψ2
(PR1 + PR2 + PR3 )R0 ,
P +P +P +P
T =
Ψ3
(PR1 + R2 R3 2 R4 R5 )R0 ,
Ψ4 , Ψ5
(PR1 + PR2 + PR4 + PR5 )R0 ,
(10)
5
3
I = 10 , 20, 10, 4, 0
Maximum
Gain
r
4
3
I =0
r
2
I =4
r
I =10
r
1
I =20
r
3
I =10
r
0
1
2
3
4
5
6
7
8
9
10
Fixed transmission rate R0 (bps/Hz)
Fig. 4. System throughput vs. the fixed transmission rate R0
Substituting Pjk in the Theorem 1 into the above expression,
the expressions of the system throughput can be obtained.
IV. S IMULATION R ESULTS
In this section, we conduct numerical simulations to verify
the effectiveness of the proposed scheme. In the simulation,
the specific Rayleigh fading is considered with channel gain
expectations Ω1 = E{g1 } and Ω2 = E{g2 }.
In the Fig. 3, we plot the probabilities of different instantaneous CSI regions versus the fixed transmission rate R0 under
different system parameters. In particular, in the subfigure (a),
σr2 = σd2 = 1, Ir = 5, Ω1 = 0.8 and Ω2 = 0.6, P1 = P2 = 25
dB. In this case, we note that although P1 = P2 , the S-R
link is stronger than the R-D link, which leads to the statistic
CSI cases Ψ1 , Ψ2 and Ψ3 . In the subfigure (b), we set the
transmit power P1 = 23.75 dB and P2 = 25 dB. Although
Ω1 > Ω2 , but P1 < P2 , which offsets the disparity of the
channel variances. Due to the symmetry, there is only one case
Ψ3 . In the subfigure (c), we set P1 = 25 dB and P2 = 30 dB.
In this case, although P1 << P2 , which results to the cases
Ψ3 , Ψ4 and Ψ5 . In this figure, we reveal that the statistic cases
from Ψ1 to Ψ5 exist under different system setups.
5
4
Ψ5
Ψ3
3.5
4
System throughput T (bps/Hz)
System throughput T (bps/Hz)
4.5
3.5
3
2.5
Ir=0
2
Ir=4
Ir=10
1.5
Ir=20
2
1
1.995
0.5
1.985
3
Ir=10
1.99
15
0
0
20
25
3
2.5
2
Ψ4
Ψ5
Ψ3
1.5
I =0
r
Ir = 4
1
Ir = 10
Ir = 20
0.5
30
I =P
r
10
20
30
40
50
0
0
60
Transmit power P1 (dB)
20
30
40
50
2
60
Transmit Power P2 (dB)
Fig. 5. System throughput vs. the transmit power P1 .
In Fig. 4, we plot the system throughput versus the fixed
transmission rate R0 . In particular, the system parameters
are set as follows: P1 = P2 = 30 dB, σd2 = σr2 = 1,
Ω1 = 0.8, Ω2 = 0.6. We note that the fixed transmission
rate R0 remarkably affects the system throughput. In the
low and strong RSI cases, the system throughput will firstly
increase with R0 , after reaching the maximum value, the
system throughput will decrease with the continuing increase
of R0 . Thus, proper design the value of R0 can optimize the
system performance. However, in the case Ir = 20, there are
two values that both maximize the system throughput, which
indicates the joint effects of the RSI and fixed rate R0 on the
system performance. In addition, in the high R0 regime, nonRSI achieves the same performance compared to the strong
RSI. The reason is that all the time slots in the region R1
and R2 are allocated to the mode M2 . However, the sum
probability of the region R1 and R2 has no relationship with
the variance of the RSI.
Fig. 5 plots the system throughput versus the transmit power
P1 . We set P2 = 30 dB, Ω1 = Ω2 = 0.8, σr2 = σd2 = 1, R0 =
4 bps/Hz. We note that the whole figure can be categorized
into two cases: Ψ5 in the low transmit power regime and Ψ3
in the high transmit power regime. In the case Ψ5 , the S-R
link is strengthened with the increase of P1 and the system
throughput increases significantly. Due to PR1 is very small,
the system will almost directly convert to the case Ψ3 . In the
case Ψ3 , the system throughput first increases very slowly,
and then fastly. The reason is that, the increase of P1 mainly
leads to the increase of PR2 . However, the time slots in the
region R2 need to be allocated to the modes M1 or M2 . Thus,
the system throughput increases slowly. In the high transmit
power regime, for example, P1 ∈ (40, 60) dB, the increase of
P1 mainly leads to the increase of PR1 . In the region R1 , all
the time slots are allocated to the mode M3 . Thus, the system
throughput increases rapidly.
Fig. 6 plots the system throughput versus the transmit power
P2 . We note that the system throughput keeps the same in the
low P2 regime under different RSI cases. The reason is that in
the low P2 regime, the system performance is restricted by the
10
Fig. 6. System Throughput vs. the transmit power P2 .
R−D link, which is none relative to the RSI. With the increase
of P2 , the system throughput will increase proportionally.
However, it is observed that for the strong RSI case, i.e.,
Ir = P2 , the system throughput reach the maximum value
when P2 ≈ 18 dB. Then, the throughput will decrease with the
increase of P2 and finally converge to a constant. The reason is
that the RSI increases with the transmit power, which reduces
the SINR performance at the relay node.
V. C ONCLUSION
In this paper, we designed an optimal hybrid FD/HD scheme
for the one-way DF FD relay system with a buffer. We focused
on the throughput maximization by the proposed scheme. To
solve this problem, we divided the system into four different
transmission modes and formulated the original problem as a
binary integer problem. By relaxing the problem, we obtained
the optimal mode selection probabilities based on the KKT
optimal conditions. The proposed scheme not only achieved
the optimal FD mode or HD mode selection, but also realized
the adaptive S-R or R-D link selection. Simulation results
validated that the proposed scheme could fully exploit the
potential of the FD mode.
A PPENDIX
We note that the relaxed problem P2 is a linear programming problem. Since the Karush-Kuhn-Tucker (KKT) conditions are necessary for the optimal solution, we first examine
the KKT conditions of the problem P2 . The Lagrangian
function for the problem P2 is given by
L(dj (i), α, β(i), µj (i), νj (i)) =
N
1 X
[d2 (i)O2 (i) + d3 (i)O3 (i)] × R0
−
N i=1
+
N
α0 X
[d2 (i)O2 (i) − d1 (i)O1 (i)] × R0
N i=1
N X
4
N X
4
X
X
µj (i)(dj (i) − 1)
dj (i) − 1) +
+ β(i)(
i=1 j=1
i=1 j=1
−
N
X
νj (i)dj (i)
(12)
i=1
select the silent mode M4 in the region R4 +R5 to balance the
arrival rate and departure rate. We have the following equation
where α0 , β(i), µj (i), νj (i) are the non-negative Lagrange
multipliers of the constraints from C1 to C4. For the optimal
solution, the derivatives of (12) equal to zero, i.e.,
(PR4 + PR5 )P41 = PR1 + PR2 + PR3
PR1 + PR2 + PR3
,
⇒ P41 =
PR4 + PR5
∂L
= 0, ∀ i, j
∂dj (i)
since that there are only two modes M1 and M4 available in
the region R4 and R5 , the selection probability of the mode
M4 is P44 = 1 − P41 .
Next, if α0 ∈ (0, 1), V3 (i) > max{V1 (i), V2 (i)}, thus, the
mode M3 will be selected with the probability 1 in the region
R1 . This case implies that by the adjustment of the allocation
of the time slots in the region R2 to the mode M1 and M2
can realize the equality of the constraint C1. Hence, we have
PR4 + PR5 − PR2 < PR3 ≤ PR4 + PR5 + PR2 , i.e., Ψ3 . For
the region R2 , we have
(13)
for j = 1, 2, 3, 4, we have
V1 (i) = N [β(i) + µ1 (i) − ν1 (i)] = α0 O1 (i)R0
V2 (i) = N [β(i) + µ2 (i) − ν2 (i)] = (1 − α0 )O2 (i)R0
V3 (i) = N [β(i) + µ3 (i) − ν3 (i)] = O3 (i)R0
V4 (i) = N [β(i) + µ4 (i) − ν4 (i)] = 0
(14)
where Vj (i) is defined as the selection function. For simplicity,
we first examine that the mode M1 is selected. In this case,
d1 (i) = 1, we get dj (i) = 0, j ∈ {2, 3, 4}, thus, the constraints
C3 for j = 2, 3, 4 and C4 for j = 1 are inactive. According to
the complementary slackness condition, we get that µj (i) =
0, j = 2, 3, 4 and ν1 (i) = 0. Here, we find that
V1 (i) − Vj (i) = N [µ1 (i) + νj (i)], j = 2, 3, 4
(15)
Since µ1 (i) ≥ 0 and νj (i) ≥ 0, V1 (i) − Vj (i) ≥ 0, j = 2, 3, 4.
This implies that when M1 is selected, V1 (i) ≥ Vj (i), ∀j, j =
2, 3, 4. The results will be similar if we set dj (i) = 1, j =
2, 3, 4. Now we can conclude that the mode with the largest
selection function should be selected in each time slot.
Based on the property of the selection function, next we
will discuss the different values of α0 in the right hand side
of (14). We consider the region R1 , in this region, all of
O1 (i), O2 (i) and O3 (i) equal to 1. According to the KKT
conditions, we know that α0 ≥ 0. First, we consider that
α0 = 0, V2 (i) = V3 (i) > V1 (i), indicating mode M2 or
mode M3 should be selected. We assume that the selection
probability of the mode Mj is P1j , thus the constraint C1 leads
to the following equation
PR4 + PR5 = PR1 P12 + PR2 + PR3
P12 + P13 = 1
PR4 + PR5 − PR2 − PR3
⇒ P12 =
PR1
(16)
where PR2 is based on the preliminary 2, i.e., the time slots
in the region R2 are preferred to be utilized to balance the
constraint C1. The selection of the mode M2 in the region
R1 means all the time slots in the region R2 are allocated to
the mode M2 . From (16), we note that if P12 < 1, we have
PR4 + PR5 − PR2 − PR1 < PR3 ≤ PR4 + PR5 − PR2 , i.e.,
Ψ2 .
However, if P12 = 1, we get PR3 ≤ PR4 +PR5 −PR2 −PR1 ,
i.e., Ψ1 . In this case, even if all the time slots in the region R1
are allocated to the mode M2 , R1 ≥ R2 . Hence, we need to
(17)
PR4 + PR5 + PR2 P21 = PR2 P22
P21 + P22 = 1
PR2 + PR4 + PR5
⇒ P21 =
,
2PR3
P22 = 1 − P21 ,
(18)
If α0 = 1, V1 (i) = V3 (i) > V2 (i), mode M1 and M3
should be selected in the region R1 . This case corresponds to
that PR2 + PR4 + PR5 < PR3 ≤ PR5 + PR4 + PR2 + PR1 ,
i.e., Ψ4 . The selection probability can be expressed as the
following
PR4 + PR5 + PR2 + PR1 P11 = PR3
P11 + P13 = 1
PR3 − PR2 − PR4 − PR5
⇒ P11 =
PR1
3
1
P1 = 1 − P1 ,
(19)
In this case, the mode M1 is selected in the region R1 ,
which means that all the time slots in the region R2 are
allocated to the mode M1 , i.e., P21 = 1 in the region R2 .
Last, we consider that α0 > 1, we note that in this case,
V1 (i) = max{Vj (i)}, j = 1, 2, 3, 4, meaning that only the
mode M1 is selected in the region R1 , which corresponds to
the case PR3 > PR1 +PR2 +PR4 +PR5 , i.e., case Ψ5 . Hence,
to satisfy the equality in the constraint C1, the inactive mode
M4 should be selected in the time slots of the region R3 . The
selection probability is given by
PR1 + PR2 + PR4 + PR5 = PR3 P12
PR1 + PR2 + PR4 + PR5
⇒ P32 =
,
PR3
(20)
and P34 = 1−P32 . Hitherto, all the cases from Ψ1 and Ψ5 have
been discussed. Based on the property of the KKT conditions,
the optimal solution of the hybrid FD/HD mode selection
problem is obtained. The proof is completed.
R EFERENCES
[1] C. Hoymann, W. Chen, J. Montojo, A. Golitschek, C. Koutsimanis, and
X. Shen, “Relaying operation in 3gpp lte: challenges and solutions,”
IEEE Communications Magazine, vol. 50, no. 2, pp. 156–162, Feb. 2012.
[2] Y. Qian, Z. Guo, R. Fan, H. Wang, J. Liu, Y. Yan, X. Shen, and Z. Hu,
“Improving outdoor to indoor coverage by use of td-lte in-band relay,”
in Proc. IEEE PIMRC, Sept. 2013, pp. 2658–2662.
[3] P. Liu and I. M. Kim, “Performance analysis of bidirectional communication protocols based on decode-and-forward relaying,” IEEE Trans.
Commun., vol. 58, no. 9, pp. 2683–2696, Sept. 2010.
[4] A. Sabharwal, P. Schniter, D. Guo, D. W. Bliss, S. Rangarajan, and
R. Wichman, “In-band full-duplex wireless: Challenges and opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1637–1652, Sept.
2014.
[5] Z. Zhang, Z. Ma, Z. Ding, M. Xiao, and G. K. Karagiannidis, “Fullduplex two-way and one-way relaying: Average rate, outage probability,
and tradeoffs,” IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 3920–
3933, Jun. 2016.
[6] C. Li, Z. Chen, Y. Wang, Y. Yao, and B. Xia, “Outage analysis of
the full-duplex decode-and-forward two-way relay system,” IEEE Trans.
Veh. Technol., vol. 66, no. 5, pp. 4073–4086, May 2017.
[7] T. Riihonen, S. Werner, and R. Wichman, “Hybrid full-duplex/halfduplex relaying with transmit power adaptation,” IEEE Trans. Wireless
Commun., vol. 10, no. 9, pp. 3074–3085, Sept. 2011.
[8] N. Zlatanov, R. Schober, and P. Popovski, “Buffer-aided relaying with
adaptive link selection,” IEEE J. Sel. Areas Commun., vol. 31, no. 8,
pp. 1530–1542, Aug. 2013.
[9] M. G. Khafagy, A. E. Shafie, A. Sultan, and M. S. Alouini, “Throughput
maximization for buffer-aided hybrid half-/full-duplex relaying with selfinterference,” in Proc. IEEE ICC, Jun. 2015, pp. 1926–1931.
[10] M. M. Razlighi and N. Zlatanov, “Buffer-aided relaying for the two-hop
full-duplex relay channel with self-interference,” IEEE Trans. Wireless
Commun., vol. 17, no. 1, pp. 477–491, Jan. 2018.
[11] S. Dang, J. P. Coon, and G. Chen, “Resource allocation for full-duplex
relay-assisted device-to-device multicarrier systems,” IEEE Wireless
Commun. Lett., vol. 6, no. 2, pp. 166–169, Apr. 2017.
| 7cs.IT
|
1
Transforming Cooling Optimization for Green Data
Center via Deep Reinforcement Learning
arXiv:1709.05077v1 [cs.AI] 15 Sep 2017
Yuanlong Li, Yonggang Wen, Kyle Guan, and Dacheng Tao
Abstract—Cooling system plays a key role in modern data
center. Developing an optimal control policy for data center
cooling system is a challenging task. The prevailing approaches
often rely on approximated system models that are built upon
the knowledge of mechanical cooling, electrical and thermal
management, which is difficult to design and may lead to suboptimal or unstable performances. In this paper we propose
to utilize the large amount of monitoring data in data center
to optimize the control policy. To do so, we cast the cooling
control policy design into an energy cost minimization problem
with temperature constraints, and tab it into the emerging
deep reinforcement learning (DRL) framework. Specifically, we
propose an end-to-end neural control algorithm that is based
on the actor-critic framework and the deep deterministic policy
gradient (DDPG) technique. To improve the robustness of the
control algorithm, we test various DRL related optimization
techniques, such as recurrent decision making, discounted return,
different neural network architectures, and different stochastic
gradient descent algorithms, and adding additional constraints
on the output of the policy network. We evaluate the proposed
algorithms on the EnergyPlus simulation platform and on a real
data trace collected from the National Super Computing Centre
(NSCC) of Singapore. Our results show that the proposed endto-end cooling control algorithm can achieve about 10% cooling
cost saving on the simulation platform compared with a canonical
two stage optimization algorithm; and it can achieve about 13.6%
cooling energy saving on the NSCC data trace. Furthermore,
it shows high accuracy in predicting the temperature of the
racks (with mean absolute error 0.1 degree) and can control
the temperature of the data center zone close to the predefined
threshold with variation lower to 0.2 degree.
I. I NTRODUCTION
With the emergence and proliferation of services and applications such as cloud computing and social networks, data
centers play an ever important role. It is predicted that global
data center IP traffic will grow 3-fold from 2014 to 2019 with
a compound annual growth rate (CAGR) of 25 percent [1]. At
the same time, the high energy consumption of data centers is
drawing more and more attention due to economic, social, and
environmental concerns. Data center electricity consumption
in the U.S. alone is projected to increase to roughly 140
billion kilowatt-hours annually by 2020, costing $13 billion in
electricity bills and emitting nearly 100 million tons of carbon
pollution per year [2]. In this paper, we focus on one of the
Yuanlong Li and Yonggang Wen are with School of Computer Science and
Engineering, Nanyang Technological University, Nanyang Avenue, Singapore
639798. Email: {liyuanl, ygwen}@ntu.edu.sg.
K. Guan is with Bell Labs, Nokia, Holmdel, NJ 07733 USA.
Dacheng Tao is with the Centre for Quantum Computation and Intelligent
Systems and the Faculty of Engineering and Information Technology, University of Technology, Sydney, 81 Broadway Street, Ultimo, NSW, Australia.
major source of energy consumption in the data center (about
37% [3]), the cooling energy.
Cooling energy optimization involves the control of a complex cooling system, which consists of multiple components,
such as cooling tower, chiller, and ventilation system, etc. A
common practice of data center cooling system control is to
adjust the set-points, i.e., the target values of different control
variables. For example, by setting the temperature control
variable at the outlet of an air conditioner to a desired value,
the air conditioner can adjust its internal state to meet the
set-point by consuming certain amount of energy. An optimal
selection of these set-points can be challenging, as the process
relies on the knowledge of the cooling system, from thermal
dynamics to mechanics. Many existing approaches are based
on an approximated system model that often incorporates
the first-order effects of thermal, electrical, and mechanical
principles [4]–[8]. These approximated models are sometimes
either inadequate or inaccurate to capture the intricacies of
various interacting processes of data center cooling operations,
thus could lead to sub-optimal or unstable cooling controls.
Recently, the “model-free” approach has emerged as an attractive alternative. A model-free approach does not assume
any specific model of the underlying system. Instead, the
control policies are learned and derived from the massive data
collected on the system status and energy consumption [9].
This approach is especially advantageous when the complexity
of the underlying system makes an accurate system modeling
a daunting task.
Currently, the prevailing control optimization approach,
which includes a model building stage and a solving stage, can
be referred to as the two-stage (TS) approach. In comparison,
the model-free approach (also referred to as the end-to-end
or the one-stage approach) uses directly the unprocessed
and often high-dimensional input, thus omitting the model
building or feature extraction stage. One particular control
optimization frameworks to our interest is reinforcement learning (RL) [10], which enables an artificial intelligence (AI)
agent to learn better decision making by directly interacting
with environment. Algorithms that combine RL with deep
learning (DL), such as deep Q-network (DQN), have been
successfully applied to the task of training AI agent to play
video games at the human performance level with only raw
pixel inputs, thus showing the potential of the end-to-end
approach [11]. The continuous domain extension of DQN,
called deep deterministic policy gradient (DDPG), has also
shown promising results on simulated physical control tasks
[12]. However, DDPG has not been widely studied in the
context of more practical and complex control optimization,
2
such as cooling system optimization for data centers (the
subject of this paper). It is yet to be demonstrated if an end-toend approach can achieve similar or better control performance
compared with the TS approaches. In addition, DDPG is
a simulation based algorithm. Like many RL algorithms, it
requires an excessive amount of computation and possibly very
long training time. Thus it is also of interest to see how the
challenges are met. Note that Google claims that they use AI
method [13] to reduce the PUE of their data center; yet no
detailed methodology or performance evaluation results are
disclosed. In our previous work, we reviewed the data center
energy cost models [14] and the existed cooling optimization
approaches [15], and conducted several data center power
analysis and control studies [16] [17] [18]; based on which we
believe that a data driven learning based optimization method
is needed for data center energy optimization, which can be
used to achieve optimization effects with minimum human
innervations and reduce the data center management difficulty.
In this paper, we investigate the viability of applying the
end-to-end neural control algorithm for data center cooling
control optimization, and evaluate the algorithm from various
aspects. We develop a neural control optimization algorithm
adapted from the DDPG and the actor-critic architecture [12]
[19]. Our proposed algorithm is off-policy, as it uses a single
pre-collected trace to learn and improve the control policy.
Beside the standard version of the algorithm which makes
control decisions based on the current states, we develop also a
recurrent version of the algorithm which is based on the recent
states and control settings. In addition, we propose to impose a
consecutive constraint to the output of the policy network. We
also evaluate the effects of discounted return, different neural
network architectures, and different weight parameters update
algorithms in training the neural networks.
To test the proposed algorithm, we use the EnergyPlus
[20] platform to build a test case; beside the simulation
case, we also collected a real data trace from the National
Super Computing Centre (NSCC) of Singapore and tested our
algorithm on it. For the simulation test case, we control five
different set-points to achieve minimum PUE and to maintain
the temperature of the data center zone within a pre-defined
range. The results of the proposed algorithms are compared
with those generated by a standard two-stage optimal control
algorithm and the default data center zone temperature setpoint control algorithm (embedded in the simulation software).
The results indicate that the proposed algorithm not only
can successfully maintain the temperature of the data center
zone within a pre-defined range under varying workloads and
weather conditions, but also achieve lower PUE compared with
a general TS algorithm. For the real data from NSCC test case,
we focus on optimizing the airflow rate setting of the three precision cooling units (PCUs) which are used to cool 26 racks.
Our results show that the proposed algorithm can approximate
the real data with high accuracy and can output energy saving
control settings according to the cooling requirements. The
main contributions of the paper are summarized as follows.
First, we propose an end-to-end and DRL-based framework
that can be used for data center cooling control optimization.
We also propose algorithms that train the neural network with
a pre-collected data trace. This approach is well suited for
a practical data center equipped with a monitoring/sensing
system that collects data in real time.
Second, we build a test-bed with the EnergyPlus and also
collect a real data trace from NSCC Singapore to evaluate the
proposed control algorithms. Our simulation results indicate
that the proposed control algorithm can accomplish the cooling
control tasks with about 10% cooling cost saving compared
with a canonical two stage optimization algorithm. Our results
on the real data trace from NSCC show that we can achieve
about 13.6% energy saving to their major air cooling units.
Third, we gain new insights into some DRL-related optimization techniques that are unique to data center cooling
control. For instance, we show that the recurrent decision
making, which is generally ignored in existing TS approaches,
can significantly improve the performance in cooling control
optimization. We also show that adding consecutive constraints
to the output of the policy network can further enhance the
performance of the algorithm. In regard to the widely used
discounted return technique in RL research, we show that
its impact is very limited as it cannot lead to significant improvement in performance stability. Our results also show that
for the cooling control optimization, a modest neural network
architecture coupled with proper weight-update algorithm can
provide competitive and robust performance. Also our results
on the real data from NSCC show that the proposed algorithm
can approximate the real energy data and temperature readings
with high accuracy (with normalized mean absolute error less
than 10%) and can output control settings that can reduce the
energy cost while satisfying the cooling requirements. Also we
show that the trade-off between the energy cost and cooling
effects can be adjusted by varying the temperature constraint
penalty weight parameter.
In summary, we demonstrate the feasibility and effectiveness of applying end-to-end neural control algorithms to
the data center cooling optimization. The evaluation of the
performance of the proposed algorithms serves as the first
step to build intelligent data center management system that
requires minimal manually intervention. Though this work is
simulation based, it does shed new light on the application
of deep reinforcement learning (DRL) to practical data center
control optimization.
II. P ROBLEM F ORMULATION
To study the cooling control optimization problem, in this
section we utilize a simulation model to present a cooling control optimization problem formulation. The simulation model
is based on the widely adopted building energy simulation
platform EnergyPlus [20]. Although the model is largely
simplified, it does capture the major cooling dynamics and
is thus adequate for studying the cooling control optimization.
A. Simulation System Model
The model is based on a simulation example provided by
EnergyPlus. As illustrated in Fig. 1, the model consists of
two data center zones (z1 and z2 ) and the associated cooling
systems. The two data center zones are only different in size
3
•}
'DWD &HQWH
U
•}
v
v
î
í
Dy Z
^W>/dd Z
Dy Z
'LUHFW ([SDQVLRQ ';
&RROLQJ
&KLOOHU &RROLQJ
^W>/dd Z
Fig. 1. System architecture: a data center consists of two server zones, z1
and z2 , and their associated and independently operated cooling facilities. A
zone is modeled as a single point, i.e., we only model the total load of each
zone and neglect the detailed thermal distribution inside the zone. The cooling
facility consists of direct expansion (DX) system and chiller cooling system.
Fig. 2. The main components of a cooling system. The temperatures at five
outlet ports (marked as open circles) are served as the temperature set-points:
TDEC , TIEC , TCW ater , TCCoil , and Tsup .
and location. In addition, their corresponding cooling systems
are independently operated. In the following, we describe the
cooling system with a focus on identifying the state space
(parameters that characterize the system), the action space
(control variables), and the reward (optimization objectives),
while omitting the details of the facility structures and operation processes that are unrelated to problem setup (yet might
be critical parts of the overall system).
1) Data Center Model – State Space and Reward: The data
center has two server zones placed side by side, z1 and z2 ,
with each zone being a standalone server room. The two zones
are different in size ( 15.24*15.24m2 and 15.24*17.00m2 for
z1 and z2 , respectively) but similar in other aspects. The heat
in each zone is generated by IT equipment (ITE) and other
sources (such as illuminations), with ITE as the dominant heat
source. The load of the ITE is defined as α · L, where L is
the designed load per square meter and α is a load factor that
varies at different time slots. In our simulation, we use a public
trace collected from Wikimedia [21] to set α to be the same for
z1 and z2 . The heat generated by the illumination is assumed
to be a constant (per square meter) as the lights inside the data
center are on all the time. Also there is heat generated by the
human workers in the data center. This part varies according
to the work time schedule. For this work, each zone is simply
modeled as a single point heat source to be cooled. This is
less accurate compared to the computational fluid dynamics
(CFD) based thermal analysis. We leave a finer-grained model
to future work. We note that even with a finer-grained model,
the proposed framework and algorithms will remain the same,
albeit with much larger state space and action space.
In the context of RL framework, we use a tuple of workload
level and the ambient temperature to represent the state, since
both of them affect the cooling load. We use the tuple of PUE
and the temperature of the data center zone to represent the
reward. In the context of the data center cooling, PUE needs
to be minimized and the data center temperature needs to be
kept within certain range.
2) Cooling System Model: Action Space: z1 and z2 are
equipped with the same cooling facility which consists of two
different types cooling systems: direct expansion (DX) cooling
system and chilled water (chiller) cooling system. Both cooling
systems are supplied with cool water from a cooling tower, but
they use the cool water in different ways. In the DX system,
the cool water passes through coils and cool down the air flow
passing over the coils. In the chiller system, the cool water is
used first to refrigerate another water stream (chilled water),
which in turn cools down the air flow.
In our model, both systems are included as shown in Fig.
2. The intake ambient air flow is first cooled by two types of
evaporative coolers: directive (DEC) and indirect (IEC), and
then passes over the DX cooling coils and is further fed to the
data center. The cold airflow to cool the data zones can also
come from the chiller.
For the action space, the underlying control algorithm in EnergyPlus (referred to as the Kbest control algorithm) uses the
following five set-points to control the zone temperature: DEC
outlet temperature (airflow) TDEC , IEC outlet temperature
(airflow) TIEC , chilled water loop outlet temperature (water
flow) TCW ater , DX cooling coil outlet temperature (airflow)
TCCoil , and air loop outlet temperature (airflow) Tsup .
In the KBest algorithm, the adjustment of these five setpoints is based on the knowledge of the underlying system
dynamics. In our proposed learning algorithm, the same five
set-points are used as control variables, but, on the contrary,
they are learned from the pre-collected data trace. Neither
physical content of these variables nor the relationship information among them will be used in the training.
B. Problem Statement
We formulate the cooling control optimization problem as
follows. We are given a time-varying tuple of the ambient
air temperature Tamb and the load factor α. The problem is
to determine the values of five control set-points, so as to
minimize the objective function as stated in Eq. (1).
The objective function aims to strike a balance between
minimizing the PUE and preventing overheating in the server
zone. In particular, the objective function consists of two
parts: the first part is PUE (denoted as P U E ), which is to
be minimized; the second part accounts for the penalty of the
overheating (for both z1 and z2 ). The penalty function has
the form of λ ln·(1 + exp(Tzi − φ)), for i ∈ [1, 2], with λ,
Tzi , and φ denoting the scaling factor of the penalty, average
temperature of zone i, and overheating threshold, respectively.
As shown in Fig. 3, here we choose the softplus [22] function
ln(1 + exp(.)), in lieu of a much simpler but non-smooth
single-sided penalty max(0, Tz1 − φ) (denoted as SimCost,
4
P U E + λ · ln(1 + exp(Tz1 − φ)) + λ · ln(1 + exp(Tz2 − φ)),
min
TCW ater ,TCCoil ,TDEC ,TIEC ,Tsup
s.t.
(1)
LTCW ater ≤ TCW ater ≤ UTCW ater ,
LTCCoil ≤ TCCoil ≤ UTCCoil ,
LTDEC ≤ TDEC ≤ UTDEC ,
LTIEC ≤ TIEC ≤ UTIEC ,
LTsup ≤ Tsup ≤ UTsup .
5. 0
max(0, Tz1 − φ)
4. 0
ln(1 + exp(Tz1 − φ)
3. 0
2. 0
1. 0
0. 0
−6. 0
−4. 0
−2. 0
0. 0
Tz1 − φ
2. 0
4. 0
6. 0
Fig. 3. The temperature penalty functions: a straightforward penalty function
could be max(0, Tz1 − φ), which is non-smooth and is difficult to optimize;
the softplus function (ln(1 + exp(Tz1 − φ)), which is smooth and is better
for the cost function optimization.
which will be tested in Section IV), to avoid the difficulty
in the neural network training phase. We reverse the sign of
this very cost function to calculate the reward for a stateaction pair, following the convention of RL framework in
which reward functions are meant to be maximized. During
the training, we simply minimize this cost function, since
commonly the training optimization algorithms are designed
for minimization. In this formulation, we also apply the
same control setting for both z1 and z2 , which is reasonable
considering that the same workload trace is used for both z1
and z2 .
III. T HE P ROPOSED A PPROACH : N EURAL E ND - TO -E ND
C OOLING C ONTROL A LGORITHMS
In this section, we present the end-to-end cooling control
algorithms (CCA), adapted from the DDPG, which combines
the key RL techniques and methods such as deep Q-network
(DQN), deterministic policy gradient (DPG), and actor-critic
algorithm. In the following, we first provide an overview of
the related RL concepts and techniques. We then describe a
complete algorithm flow and the design of the neural networks.
A. Overview of Q-learning and Policy Gradient
Our goal is to enable an AI agent to learn an optimal
control policy from a data set that records a sequence of states,
actions taken, and rewards at discrete time steps. Within the
RL framework, this goal is achieved by using either valuebased or policy-based approaches. Central to the value-based
approaches is the Q-learning technique, which uses a Qfunction Q(s, a) to represent the maximum discounted future
reward for an action a taken at a state s. As Q(s, a) indicates
the utility (value) of choosing an action a at state s, an optimal
policy π simply chooses the action that has the highest Q value
for a state s, i.e., π = argmaxa Q(s, a). Though for discrete
state and action space (especially when the space is small), Qfunction can be represented as a table computed by Bellman
equation iterative updating, in practice it is often estimated
by a function approximator such as a neural network, like the
Deep Q-network (DQN) [11]. With policy-based approaches,
policy-gradient (PG) is an important algorithm that optimizes a
policy end-to-end by computing noisy estimates of the gradient
of the expected reward and then updating the policy in the
gradient direction. In practice, policy iterations and updates
are often carried out by a neural network, referred to as
policy network. Traditionally, PG methods have assumed a
stochastic policy, which gives a probability distribution over
actions. Recent research has proved that the stochastic policy
gradient, which is the gradient of the policy’s performance, is
equivalent to the deterministic policy gradient, thus allowing
all the machinery for stochastic policy gradients to be applied
to deterministic policy gradients [23].
When state or action space is represented by continuous
variables, a naive adaptation of DQN or PG via discretization
of state or action space often results in intractability or very
slow learning convergence (even divergence). We use DDPG
algorithm, which is essentially a hybrid method combining the
policy gradient method and the value function [12]. Specifically, DDPG is an actor-critic algorithm: it uses two neural
networks, referred to as actor network and critic network,
to approximate the policy function and the value function,
respectively. These networks compute action predictions for
the current state and generate a temporal-difference (TD) error
signal at each time step. The input of the actor network is
the current state, and the outputs are real values representing
actions chosen from a continuous action space. The critic’s
output is simply the estimated Q-value of the current state
and of the action given by the actor. The deterministic policy
gradient theorem provides the update rule for the weights of
the actor network. The critic network is updated from the
gradients obtained from the TD error signal.
5
B. Online learning vs. batch learning and Off-Policy vs. OnPolicy
We note that RL algorithms can be directly used as online
learning algorithms. This means that the control algorithm
can learn in an online manner, e.g., starting from an initial
state and adjusting itself with the input it received from the
ongoing process, either the real operation or the simulation.
However, this will be problematic for the data center cooling
task, which cannot risk erroneous settings. In this work, we
focus on the control algorithms that are pre-trained by the
offline data first, which is referred to as “batch learning”. Batch
algorithms can be further divided into two categories based on
how the training data is generated: off-policy and on-policy.
Off-policy algorithms generally employ a separate behavior
policy, which is independent of the policy being estimated,
to generate the training trace; while on-policy directly uses
control policy being estimated (in the real control practice or
more likely in a simulator) to generate training data traces.
For the case of data center simulation, the cost for simulation
time is high. Thus off-policy algorithms are easier to apply
and more suitable for our case.
In summary, we propose an off-policy control algorithm
adapted from canonical DDPG. The algorithm employs only
a single offline trace for batch learning. In the following we
introduce the details of the proposed algorithm.
C. Cooling Control Algorithm with Single Offline Trace (CCASOT)
The flowchart of the proposed Cooling Control Algorithm
with Single Offline Trace (CCASOT) is shown in Algorithm
1. For the training task, a data trace is collected (line 1),
which contains entries: state (Tamb (t), α(t)), action (TDEC (t),
TIEC (t),TCW ater (t), TCCoil (t) and Tsup (t)), and reward data
y(t) is computed by the objective function (1) based on the
observed PUE and temperature data (note here we do not
incorporate future reward, which will be discussed in Section
III-E2). All these data are prepared as time series of N
time steps, which are further divided into training data and
validation data.
Before the training starts, we first initialize two neural networks (line 2). The critic network Q(XQ |θQ ) (parameterized
by θQ ) approximates the Q-value of a state-action pairs: It
takes both state and action (combined into a single tuple XQ )
as the input, and outputs a scalar value which represents the
cost of an action a taken at a state s. The actor network
µ(Xµ |θµ ) (parameterized by θµ ) is policy network: it takes
the current state (Xµ ) and outputs a five dimensional vector
corresponding to the five control factors shown in the cooling
system model.
The training procedure is shown in lines 3-21. Here we
use standard neural network training procedure with multitraining-epochs. Within each epoch (line 5-9), each batch of
training data is used to update the weights of the neural
networks using gradient descent. The critic network is updated
by minimizing the mean square error between the output of
Q and the objective y; while the policy network is updated by
minimizing the output of Q when taking action at current state
according to the output of µ. To avoid over-fitting, we also
compute the validation error to keep track of the best weight
parameter settings for the two neural networks respectively,
as shown in lines 11-19. On important note is that for the µ
network, the validation error can be very small at the beginning
due to that at that time the Q network is not well learned. For
µ
safety, one can use a periodical re-initialization of the Eval
to
solve this problem.
Algorithm 1 CCASOT: Cooling Control Algorithm with
Single Offline Trace
1: Data preparation: Collect a single offline trace with
historical data series of the following variables: Ambient
environment temperature Tamb (t), workload level r(t),
and settings of the control factors TCW ater (t), TCCoil (t),
TDEC (t), TIEC (t), and Tsup (t), where t = 1, ..., N ,
N is the total number of time steps. Denote XQ =
{Tamb (t), r(t), TCW ater (t), TCCoil (t), TDEC (t), TIEC (t)}
and Xµ = {Tamb (t), r(t)}, where XQ is the input of
the Q-neural network Q and Xµ is the input of the
policy neural network µ. Collect also the PUE and outlet
temperature data series and compute the cooling reward
y(t) at each time step t according to objective function
Eq. (1). Divide the data into training data (used for
training the neural network) and validation data (used for
validate the neural network to find the best parameter
settings). Denote IV as the index of sample data in the
validation set.
2: Initialization: Create neural networks Q(XQ |θ Q ) and
µ(Xµ |θµ ), and randomly set the values of the weight
parameters θQ and θµ for the neural networks. Create
0
0
target network Q0 and µ0 with θQ = θQ and θµ = θµ .
Q
µ
Set initial validate error Eval
= 1e100, Eval
= 1e100 and
Qbest = N ull, µbest = N ull.
3: for Epoch = 1, ..., M axEpoch do
4:
Randomly divide the training data into batches of size
M and denote the index set of the data in the ith batch
as Ii .
5:
for i = 1, ..., N/M do
6:
Update
θQ by minimizing:
P
Q 2
7:
j∈Ii (y(j) − Q(XQ (j)|θ ) /M .
µ
8:
Update
θ by minimizing:
P
µ
Q
9:
j∈Ii Q([Xµ (j), µ(Xµ (j)|θ )]|θ )/M .
10:
end for P
0Q
11:
Eval
← j∈IV (y(j) − Q(XQ (j)|θQ )2 .
0Q
Q
12:
if Eval < Eval
then
Q
0Q
13:
Eval ← Eval
14:
Qbest = θQ
15:
end if P
0µ
16:
Eval
← j∈IV Q([Xµ (j), µ(Xµ (j)|θµ )]|θQ ).
0µ
µ
17:
if Eval < Eval
then
µ
0µ
18:
Eval ← Eval
19:
µbest = θµ
20:
end if
21: end for
22: Return: The optimal Q and µ neural network with weight
parameter settings Qbest and µbest respectively.
6
D. Neural Network Design
We setup the Q-network as follows. Denote the number of
states as ns and the number of control factors as nc . We
use a three-layer neural network (input, hidden, and output
layer) to implement Q. Specifically, with weight matrices
(1)
(2)
(1)
(2)
W(ns +nc )×nhid1 , Wnhid1 ×1 , bias vectors b1×nhid1 , b1×1 , and
input XQ , the output of Q-network is given by:
Q(XQ ) = σ((σ(XQ W (1) + b(1) ))W (2) + b(2) ),
where σ(·) is the sigmoid function; nhid1 is the number of
hidden nodes in the hidden layer of the Q-network. Note that
as the output range of the sigmoid functions is [0,1], in the
training the ground-truth of the output of Q must be scaled
into range [0, 1].
The set-up of the µ-network is as follows. Similar to Q, we
use a three-layer neural network to implement µ. With weight
(3)
(4)
(3)
(4)
matrices Wns ×nhid2 , Wnhid2 ×nc , bias vectors b1×nhid2 , b1×nc ,
and input Xµ , the output of µ is given by:
µ(Xµ ) = σ((σ(Xµ W (3) + b(3) ))W (4) + b(4) ),
where nhid2 is the number of hidden nodes in the hidden layer
of µ. Note that for each state, the output of µ-network is a
nc -dimensional vector, in which each element is distributed in
range [0,1].
During the training procedure, all input values are scaled
into range [0, 1] in the following manner:
xmin
xrange
x
= min(x);
= max(x) − min(x);
← (x − xmin )/xrange ,
where x stands for any state or control factor. The output of
the µ-network will be transformed in the following manner to
recover the real control setting xreal (used in simulation or
real practice):
xreal = xµ · xrange + xmin .
Consequently, the output of µ and the recovered true control
setting values are in valid ranges (the specific range for each
control variable after recovering), ensuring that the output
of the policy network µ automatically meets the control
constraints due to the neural network design.
E. Further Improvement of the Algorithm
In this section, we discuss two important issues in designing
the training data (input, output) pair for the Q-network and an
additional cost for the policy network. For the Q-network, in
the above listed algorithm, the training data can be presented as
a pair of traces (x(t), y(t)): x(t) represents the input to the Qnetwork, which consists of the state and actions taken at time
slot t; y(t) represents the output of the Q-network, which is
equivalent to the cost. In this case, the control setting at time
slot t depends only on the state at time slot t. In addition, the
cost to be minimized is also only the cost incurred at time
slot t. If we expand the time span for both decision making
and cost minimization, we have two different design variants –
recurrent decision making and discounted return, with respect
to the input and output of the Q-network respectively. For the
policy network, as the control settings should not vary abruptly
between two consecutive time steps, we propose an additional
cost item to improve the robustness of the policy network.
Details are shown below.
1) Recurrent Decision Making: With respect to the input
x(t) = {s(t), a(t)} of the Q-network, we can consider to
incorporate previous states and control actions into the inputs;
These additional information can affect the control decision
setting. Such approach is usually called recurrent approach
[24], which has been extensively studied in RL. We introduce
the recurrent algorithm with the CCASOT as an example. In
the standard CCASOT, the decisions are made solely based
on the current state s(t). In the recurrent version, the decision
are made based on the recent states and actions s(t − τ ), a(t −
τ ), s(t − τ + 1), a(t − τ + 1), ..., s(t − 1), a(t − 1), s(t), a(t),
where τ is the maximum number of time steps to trace back.
Making decisions with the recent states and actions, which can
be considered as a short time series, actually concedes certain
degree of the prediction of the future states of the short series.
Also, using the recent historical series as the input data can
concede certain degree of inner state inference. We denote this
recurrent version of CCASOT as CCASOT/his, which can be
trained in the same manner as CCASOT, with modified input
vector.
2) Discounted Return: With respect to the output y(t) of
the Q-network, one can use the expected return of the future
as the cost. In RL, as a general practice, the expected return is
usually defined as a discounted summation of the cost in the
future as the expected return, i.e., y(t)+ηy(t+1)+η 2 y(t+2)+
η 3 y(t + 3)..., where 0 < η < 1 is the discounted factor. In the
above proposed algorithm, we simply set y(t) as the training
cost, instead of the discounted return. The reason is that in the
general RL problems, the system state transition is determined
by the actions taken (standard Markov decision process); while
in our problem, the state transition is largely determined by
the weather and the workload fluctuations (although some
inner states are affected by the actions, which makes it a
semi-Markov process). To evaluate the effectiveness of the
discounted return, we carry out tests with discounted cost. That
is, we replace the training cost function for the Q-network, in
which the reward y is replaced by y + ηQ(Xµ0 , µ(Xµ0 )), where
Xµ0 is the next state to the current state Xµ . This version of
algorithm is denoted with a postfix “/DR”.
3) Constraints on Consecutive Set-points: In cooling control optimization, the control settings at consecutive time steps
shall not change abruptly. This is due to the fact that the
set-point is not directly controlled and each of setting values
takes certain time delay to be in effect. In such case, we can
add a constraint to limit the variability of the outputs of the
control policy in consecutive time steps. In particular, we add
the following consecutive constraint (CC) cost penalty when
updating the policy network:
νCC =
X
j∈Ii &j+1∈Ii
(xµj − xµj+1 )2 ,
7
where Ii is the i-th training batch and xµ is the output of the
policy network. This version of algorithm is denoted with a
postfix “/CC”. We carry out tests to verify its effects in the
next section.
IV. S IMULATION BASED N UMERICAL E VALUATION
R ESULTS AND A NALYSIS
In this section we present numerical evaluation results of the
proposed CCASOT. Simulations on the EnergyPlus platform
are carried out to collect the training data and evaluate the proposed algorithms. Two baseline algorithms are compared with
our proposed solutions: one is the default control algorithm
KBest from EnergyPlus, which controls the simulation with
the knowledge of the system dynamics; another is a general
TS control optimization algorithm which is trained with the
same data for the proposed approach.
A. Simulation Configurations
We use EnergyPlus to collect the training data and evaluate
different control algorithms for the following reasons. First, it
is impractical to directly test control algorithms on a real data
center due to the potential risk and the long running time.
Second, EnergyPlus, whose development is initiated by the
U.S. Department of Energy Building Technologies Office, is a
widely recognized and reliable simulation platform to model
building cooling energy consumption. Third, EnergyPlus provides the flexibility that allows simulations with user defined
algorithms, control actions, and schedules.
The simulation is configured as follows. We adopt the
original data center model provided by the EnergyPlus platform to make this simulation based study tractable, as shown
in Section II-A. We choose Singapore as the location and
select the corresponding weather file to revise the simulation
configuration file accordingly. We use a CPU loading trace
collected from the monitoring system of Wikimedia as the
workload trace for the data center model in the simulation.
The whole simulation period is one year and simulation data
are collected every 6 minutes. To avoid the potential issues of
time delay between an action and the resulting state change,
we “synchronize” the traces by aligning the state at the i-th
time step with the action and the reward at the (i + 1)-th time
step. For the simplicity of notation, the time slot index t used
in this paper is already after time slot realignment.
We use two different traces to test the algorithms. The
first trace, referred to as “D” trace, is generated by the
default KBest control algorithm of EnergyPlus, with the zone
temperature set-point set to 29o C. The second trace, referred
to as “r” trace, is collected when the control variables are
randomly selected from valid ranges for the simulation model.
That is, we select set-points according to TCW ater (t) ∈ [5, 9],
TCCoil (t) ∈ [18, 23], TDEC (t) ∈ [21, 31], TIEC (t) ∈ [21, 31],
and Tsup (t) ∈ [18, 23]. These two traces are used to test
different control algorithms for four test cases corresponding
to four different test periods. For the whole 1-year simulation
period, we select the last l% as the test period of the year,
where we set l to 15%, 25%, 35%, and 45% in four test cases.
We set λ = 1.0 in the objective function to compute the cost.
B. Algorithm Configurations
For the proposed CCASOT, the hidden layer sizes of Q and
µ network are set to 50 and 10, respectively. The optimization
algorithm to update the weight parameters is Adadelta [25].
The maximum training epoch M axEpoch is set to 100. The
training batch size is set to 1. For the CC version of the
proposed algorithm, we set the batch size to 5. For DR version
of the proposed algorithm, η is set to 0.1.
For the TS optimization algorithm, we use the following
settings. In the first stage, we use a three-layer neural network
to approximate the system dynamics (which is the same as the
Q-network of the CCASOT). The hidden layer size is set to
50. The maximum training epoch to train this neural network
is also set to 100. In the second stage, an iterative differential
evolutionary algorithm provided by Scipy [26] is used to find
the optimal solution for each test state.
For the proposed CCASOT and the TS optimization algorithms, the optimal control settings generated by these two
algorithms are further tested by simulations on the EnergyPlus
platform. That is, for each state at the testing phase, we use
the settings provided by the CCASOT or the TS algorithm
and then record the resulting state changes and rewards for
performance evaluation.
C. Comparing the proposed CCASOT to Baseline Algorithms
In this section we present the results of the average PUE and
maximum zone temperatures (during the test period), obtained
by using CCASOT, and also TS and KBest algorithms. Based
on results we further compare and evaluate the underlying
control algorithms.
Table I shows the first and second-order statistics of the
PUE and maximum zone temperatures in 10 independent
runs. For better examining these results, we also plot PUE
and the maximum zone temperature (for z1 ) distribution in
Fig. 4. We can observe the following. a) With KBest control
algorithm, the average PUE is the smallest. However, the
monitored maximum zone temperature is much higher than
that of the other algorithms (above 30o C vs. the pre-defined
set-point 29o C). This result shows that the KBest algorithm
produces a relatively coarse control granularity. b) CCASOT
algorithms can control the temperature within the constraint.
Its better performance proves that the algorithm indeed can
learn the underlying system dynamics and make reliable
decisions. c) Compared with the CCASOT, the general TS
algorithm under-performs. Since we utilize an iterative solver
at each time step, the TS algorithm approach is not only time
consuming, but also leads to slow convergence and instability.
While with the CCASOT, the trained actor network works
as a deterministic non-linear function that maps the states to
actions, thus generating much more stable results. Moreover,
the comparison between CCASOT and TS proves that the
actor network can indeed attain optimal or close-to optimal
control settings. d) The recurrent version of CCASOT has
better performance. This indicates that it is beneficial to also
take recent states and actions into consideration for decision
making. e) The proposed CCASOT/his performs the best
8
TABLE I
C OMPARISON OF THE PROPOSED CCASOT WITH THE BASELINE ALGORITHMS : KB EST AND TS. T HE RESULTS SHOW THAT THE CCASOT/ HIS
PERFORMS WELL ON BOTH THE R AND D TRACES , WITH SMALLER PUE OR BETTER MAXIMUM ZONE TEMPERATURE ( o C) ACHIEVED .
15%
25%
35%
45%
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
KBest
mean±std
1.3794±0.0000
30.0572±0.0000
30.0519±0.0000
1.3771±0.0000
30.0611±0.0000
30.0614±0.0000
1.3807±0.0000
30.0697±0.0000
30.0821±0.0000
1.3819±0.0000
30.0697±0.0000
30.0821±0.0000
CCASOT/r
mean±std
1.4400±0.0470
29.5385±1.0346
29.5390±1.0354
1.4222±0.0215
29.0543±0.0057
29.0544±0.0057
1.4342±0.0398
29.3834±0.9914
29.3806±0.9834
1.4294±0.0424
29.7146±1.3242
29.7101±1.3157
CCASOT/D
mean±std
1.4428±0.0418
29.0538±0.0067
29.0540±0.0066
1.4230±0.0306
29.0571±0.0044
29.0572±0.0044
1.4360±0.0436
29.0579±0.0045
29.0578±0.0045
1.4297±0.0289
29.0589±0.0025
29.0588±0.0024
CCASOT/his/r
mean±std
1.4209±0.0016
29.0561±0.0002
29.0562±0.0002
1.4112±0.0019
29.0562±0.0002
29.0563±0.0003
1.4284±0.0283
29.0573±0.0005
29.0572±0.0003
1.4170±0.0016
29.0571±0.0005
29.0571±0.0003
TABLE II
C OMPARING THE CCASOT INCORPORATED WITH CONSECUTIVE
CONSTRAINT (CC) AND DISCOUNTED RETURN (DR). A LL
TEMPERATURES ARE IN o C.
15%
25%
35%
45%
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
CCASOT/his/r
mean±std
1.4209±0.0016
29.0561±0.0002
29.0562±0.0002
1.4112±0.0019
29.0562±0.0002
29.0563±0.0003
1.4284±0.0283
29.0573±0.0005
29.0572±0.0003
1.4170±0.0016
29.0571±0.0005
29.0571±0.0003
CCASOT/his/D
mean±std
1.4206±0.0061
29.0585±0.0000
29.0586±0.0000
1.4158±0.0070
29.0585±0.0001
29.0586±0.0001
1.4181±0.0093
29.0597±0.0003
29.0596±0.0003
1.4200±0.0084
29.0596±0.0004
29.0595±0.0004
CCASOT/his/r/CC
mean±std
1.4196±0.0006
29.0560±0.0003
29.0561±0.0003
1.4096±0.0004
29.0561±0.0002
29.0562±0.0002
1.4166±0.0010
29.0572±0.0005
29.0572±0.0004
1.4159±0.0002
29.0574±0.0005
29.0573±0.0004
CCASOT/his/D/CC
mean±std
1.4155±0.0034
29.0585±0.0000
29.0586±0.0000
1.4081±0.0058
29.0585±0.0000
29.0586±0.0000
1.4137±0.0045
29.0599±0.0000
29.0598±0.0000
1.4159±0.0140
29.0599±0.0000
29.0598±0.0000
CCASOT/his/r/DR
mean±std
1.6181±0.1006
29.0682±0.1622
29.0546±0.1547
1.6307±0.0685
29.0223±0.0127
29.0068±0.0088
1.6001±0.0799
29.0844±0.1940
29.0753±0.2064
1.6630±0.0195
29.0314±0.0040
29.0063±0.0048
CCASOT/his/D/DR
mean±std
1.4258±0.0364
29.0585±0.0000
29.0586±0.0000
1.4288±0.0449
29.0585±0.0000
29.0586±0.0000
1.4181±0.0140
29.0599±0.0000
29.0598±0.0000
1.4214±0.0131
29.0599±0.0000
29.0598±0.0000
CCASOT/his/D
mean±std
1.4206±0.0061
29.0585±0.0000
29.0586±0.0000
1.4158±0.0070
29.0585±0.0001
29.0586±0.0001
1.4181±0.0093
29.0597±0.0003
29.0596±0.0003
1.4200±0.0084
29.0596±0.0004
29.0595±0.0004
D. Additional Methods for Further Improving the Performance of CCASOT
In this subsection we show results of CCASOT/his, incorporated with CC or DR, to verify whether these techniques
can be helpful. The results are shown in Table II.
We observe in Table II that with CC, the CCASOT performs
more stably and produces better PUE. Compared with a
general TD/r, the performance of PUE drops from 1.45 to 1.41,
equivalent to about 10% reduction of cooling cost. It should
be mentioned that although CC is useful in our problem (as
it limits the range of the control setting change), its benefit
does not necessarily extends to other problem set-ups. For the
DR, we observe that the algorithm produces unstable results,
especially for the r trace. This can be explained by the fact
that for data center temperature control the effect of a control
setting at a particular time step does not propagate too far into
the future. In other words, the effect drastically diminishes
after a certain duration. Using DR may over-emphasize the
effect of one action on the states and actions in the far future.
As such, the general practice of RL does not necessarily hold
for our solution approaches.
E. Comparing Different Neural Network Designs
In this subsection, we compare different ways to implement
the neural network. We compare the basic network design
with the four other different approaches: 1) SimCost: with
the simple definition of cost function analyzed presented in
Section II-B; 2) Drop: we add a drop-out layer with a total
number of nodes that is five times that of hidden layer;
3) Target: with the target network of DDPG; 4) NewArc:
with a different architecture in reference [27]. The results are
TS/D
mean±std
1.4555±0.0421
29.2108±0.4615
29.2088±0.4552
1.4219±0.0241
29.1985±0.3421
29.1950±0.3375
1.4426±0.0363
29.4921±0.7694
29.4876±0.7647
1.4573±0.0432
29.7679±0.7193
29.7602±0.7093
TS/his/r
mean±std
1.4537±0.0338
29.2009±0.1211
29.1907±0.1185
1.4434±0.0288
29.9581±0.7474
29.9555±0.7357
1.4484±0.0293
29.6196±0.1655
29.6024±0.1686
1.4409±0.0351
29.3090±0.2759
29.3175±0.2940
TS/his/D
mean±std
1.4295±0.0191
29.2360±0.1518
29.2310±0.1539
1.4289±0.0248
29.9856±0.5003
29.9830±0.4931
1.4293±0.0210
30.0536±0.3854
30.0383±0.3810
1.4276±0.0221
29.7497±0.4973
29.7494±0.4850
TABLE III
C OMPARING DIFFERENT NEUAL NETWORK IMPLEMENTATIONS . THE
BASIC , S IM C OST, D ROP, TARGET, AND N EWA RC . A LL TEMPERATURES
ARE IN o C.
15%
25%
35%
45%
among all the learning algorithms studied here, and it shows
competitive performance on both training traces.
TS/r
mean±std
1.4189±0.0204
31.6394±0.9517
31.6336±0.9528
1.4625±0.0414
31.5420±1.1069
31.5319±1.1014
1.4745±0.0507
31.6859±1.1578
31.6968±1.1376
1.4595±0.0552
30.9061±1.4223
30.8943±1.4151
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
CCASOT/r
mean±std
1.4400±0.0470
29.5385±1.0346
29.5390±1.0354
1.4222±0.0215
29.0543±0.0057
29.0544±0.0057
1.4342±0.0398
29.3834±0.9914
29.3806±0.9834
1.4294±0.0424
29.7146±1.3242
29.7101±1.3157
CCASOT/r/SimCost
mean±std
1.4603±0.0484
29.6380±1.1911
29.6394±1.1936
1.4042±0.0373
30.0447±1.5143
30.0381±1.5040
1.4461±0.0486
29.6481±1.1728
29.6434±1.1683
1.4459±0.0614
30.0390±1.5210
30.0316±1.5100
CCASOT/r/Drop
mean±std
1.4649±0.0384
30.8055±1.0116
30.7851±1.0020
1.4537±0.0589
30.5054±0.8537
30.5028±0.8470
1.4700±0.0514
30.1120±1.0948
30.0964±1.0874
1.4759±0.0219
30.7437±1.0035
30.7337±1.0039
CCASOT/r/Target
mean±std
1.4546±0.0690
30.0622±1.1306
30.0582±1.1293
1.4442±0.0514
29.1408±0.2515
29.1415±0.2520
1.4674±0.0543
29.0559±0.0060
29.0558±0.0059
1.4732±0.0473
29.3066±0.5221
29.3082±0.5205
CCASOT/r/NewArc
mean±std
1.4625±0.0551
29.0562±0.0002
29.0563±0.0002
1.4948±0.0000
30.0622±0.0000
30.0803±0.0000
1.4665±0.0000
30.0762±0.0000
30.0839±0.0000
1.4502±0.0000
30.0762±0.0000
30.0839±0.0000
shown in Table III, all with r trace. The results show that the
simple cost function causes performance degeneration, which
proves that the cost function we choose at the first place can
successfully generate optimal or near-optimal control policies.
For the other network architectures, they do not show any
significant performance improvement. The comparisons show
that for cooling control problem, implementing CCASOT with
a simple neural network design that is easy to train can produce
better and more robust performance, which is critical for
designing a practical controller. However, it should be noticed
that for problems with high-dimension of state and action
space, we may need to use a deeper (more hidden layers)
neural network to extract useful features from the input.
F. Comparing Different Training Algorithms
The above results are so far generated with the training
algorithm AdaDelta. It is also interesting to evaluate the
performance of other training algorithms. In Table IV, we
show the results for different training algorithms: AdaDelta
(Ada), rmsprop [28] (Rms), and stochastic gradient descent
(Sgd). The results show that the best training algorithm is
AdaDelta, while Rmsprop performs almost as well. This is
expected, since both algorithms adapt step size in updating
the weight parameters of the neural networks. These results
indicate that choosing a proper training algorithm is important,
even for the simple cooling control instance studied in this
paper.
G. Examples of Simulation Results
In this subsection, we present the examples of the PUE and
temperature traces obtained from our simulation. As shown
in Fig. 5, we compare the best TS variant TS/his/D, and the
9
1.60
1.55
1.50
1.45
1.40
1.35
32.5
32.0
31.5
31.0
30.5
30.0
29.5
29.0
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
(a)
1.52
1.50
1.48
1.46
1.44
1.42
1.40
1.38
1.36
32.5
32.0
31.5
31.0
30.5
30.0
29.5
29.0
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
(b)
1.55
1.50
1.45
1.40
1.35
32.5
32.0
31.5
31.0
30.5
30.0
29.5
29.0
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
(c)
1.54
1.52
1.50
1.48
1.46
1.44
1.42
1.40
1.38
1.36
32.5
32.0
31.5
31.0
30.5
30.0
29.5
29.0
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
KBest
CCASOT/r
CCASOT/D CCASOT/his CCASOT/his
/r
/D
TS/r
TS/D
TS/his/r
TS/his/D
(d)
Fig. 4. Average PUE (over the test period) and maximum zone temperature distribution achieved by different control algorithms for the test case l = 15%,
25%, 35%, and 45%, shown in (a), (b), (c), and (d), respectively. In each sub-figure, the upper part shows PUE while the lower part shows maximum zone
temperature (o C). The average values are based on 10 independent runs. The red line in each box is the median and the red square is the mean value.
10
TABLE IV
C OMPARING DIFFERENT TRAINING ALGORITHMS : A DADELTA (A DA ),
RMSPROP (R MS ), AND STOCHASTIC GRADIENT DESCENT (S GD ).
15%
25%
35%
45%
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
PUE
Tmax (z1 )
Tmax (z2 )
CCASOT/his/r/Ada
mean±std
1.4209±0.0016
29.0561±0.0002
29.0562±0.0002
1.4112±0.0019
29.0562±0.0002
29.0563±0.0003
1.4284±0.0283
29.0573±0.0005
29.0572±0.0003
1.4170±0.0016
29.0571±0.0005
29.0571±0.0003
CCASOT/his/D/Ada
mean±std
1.4206±0.0061
29.0585±0.0000
29.0586±0.0000
1.4158±0.0070
29.0585±0.0001
29.0586±0.0001
1.4181±0.0093
29.0597±0.0003
29.0596±0.0003
1.4200±0.0084
29.0596±0.0004
29.0595±0.0004
CCASOT/his/r/Rms
mean±std
1.4205±0.0011
29.1451±0.0000
29.1373±0.0000
1.4129±0.0051
29.7210±0.0000
29.7259±0.0000
1.4257±0.0155
29.1967±0.0000
29.1896±0.0000
1.4170±0.0017
29.0915±0.0000
29.0781±0.0000
CCASOT/his/D/Rms
mean±std
1.4330±0.0320
29.1854±0.0000
29.1800±0.0049
1.4119±0.0077
29.7210±0.0000
29.7259±0.0000
1.4154±0.0070
29.8650±0.0000
29.8516±0.0000
1.4160±0.0084
29.5037±0.0000
29.5047±0.0000
CCASOT/his/r/Sgd
mean±std
1.4711±0.0713
29.5891±0.8971
29.5803±0.8913
1.4921±0.0653
29.7210±0.0000
29.7259±0.0000
1.4637±0.0633
29.1967±0.0000
29.1896±0.0000
1.4454±0.0289
29.5597±0.9586
29.5415±0.9493
CCASOT/his/D/Sgd
mean±std
1.4499±0.0477
29.1853±0.0003
29.1792±0.0047
1.4091±0.0051
29.7210±0.0000
29.7259±0.0000
1.4281±0.0398
29.8650±0.0000
29.8516±0.0000
1.4517±0.0521
29.5037±0.0000
29.5047±0.0000
TABLE V
T HE DATA ENTRIES AND THEIR USAGE IN OPTIMIZATION OF THE AIRFLOW
RATE OF THE TARGET PCU S .
State
Action
best CCA variants CCASOT/his/D/CC with the KBest control
algorithm in EnergyPlus.
The PUE curves show that: a) The PUE curve of the KBest
algorithm can drop to a very low level (about 1.1) in the
simulation. This implies that the cooling system is almost
not running. Such a greedy cooling setting shows the advantage of the KBest algorithm, utilizing the underlying system
dynamics. b) In comparison, learning-based algorithm cannot
achieve such low PUE. Among learning-based algorithm, the
PUE curve of the CCASOT/his/D/CC is lower than that of
TS/his/D most of the time, thus shows the better performance
of CCASOT. c) From the temperature traces, we observe that
with the learning algorithms, the temperature is kept very close
to the predefined threshold of 29o C. Note that at the beginning
of the test period, the temperature of the data center zone is
high for both TS and CCASOT, due to transient caused by
the transition from the KBest algorithm and to the learning
algorithm (the settings before the test period are provided by
the KBest algorithm). However, for both TS and CCASOT,
the zone temperature soon stabilizes to predefined temperature
threshold and only fluctuates within a small range (less than
0.2o C), much smaller than that of KBest. As such, in tallying
the statistics of PUE and maximum zone temperature, we omit
the first 100 time steps of the test period. We can also note
that at certain time steps, TS fails to control the temperature
with in the pre-set range, as shown in Fig. 5 (h).
V. T EST ON R EAL DATA T RACE FROM NSCC
To further investigate the proposed algorithm, we test the
best version CCASOT/his on a data trace collected from the
National Super Computing Centre (NSCC) of Singapore and
show its performance on optimizing the energy cost while
satisfying the cooling requirements (rack intake temperature).
We focus on optimization of the air cooling system for the
computing nodes in NSCC. The 3-D model of the research
target is shown in Fig. 6. There are 26 racks in the target
system. Three precision cooling units (PCUs) supply cold air
for these racks. The PCUs supply cold air at about 20 degrees.
Cold air enters into the cold aisle and then goes through
the racks and at last returns back to the PCUs. There are
other cooling facilities installed: for racks 1-20, an additional
warm water cooling system is used to cool the CPU/GPU and
memory chips; for racks 21-26, an additional rear-door cooling
system is used. The warm water cooling system and rear-door
cooling system will not be studied here thus we omit further
details.
We try to optimize the total supply flow rate of the target
PCUs shown in Fig. 6, aiming to minimize their power
Reward
P1: total power consumption of racks 1-26;
P2: total power consumption of the other racks;
P3: total heat load recorded in warm water cooling system;
F1: water flow rate; T1: water flow supply temperature;
F2: air flow rate of other PCUs;
T2: average supply temperature of other PCUs;
F3: rear door cooling distribution units chilled water flow rate; and
T3: the average supply temperature of the target PCUs.
F4: The airflow supply rate of the target PCUs
P4: Power consumption of the target PCUs; and
T4: average rack intake temperature.
consumption while maintain the average intake temperature
of the racks (measurement point at the height of the 36U of
each rack).
To apply the proposed algorithm, we collected the related
data for the optimization goal, as shown in Table V. Note that
several measurements of the warm water cooling system and
rear-door cooling system are also used. Our experiments show
that including these readings can increase the approximation
accuracy of the Q-network. We collected these data entries for
every 3 minutes during March 1 to 15 of 2017. For these data,
we use the first 85% as the training data, and the last 15% as
the test data. With the above data, we utilize the proposed
algorithm to train the Q and µ network. In the Q network, we
tried a special design that will approximate the energy cost
and rack temperature respectively.
Experimental results show that the proposed algorithm
works well as expected. First, the results show that the
normalized mean absolute error of the Q network is smaller
than 10% (0.1 degree), as shown in Fig 7. This proves that
the network can successfully capture the system dynamics.
Second, we study how the control settings output by the
policy network change when different settings of the penalty
parameter λ (used in the objective function (1)) are used, as
shown in Fig. 8. In figure 8 (a), we show when λ changes from
0.0, 0.1, to 10.0, the control settings output by the proposed
algorithm change as the curves’ color changes from dark to
light yellow, while the real setting is shown as the blue curve.
At the same time, we show how the predicted rack temperature
changes correspondingly. We can observe that 1) With the
control setting of our proposed algorithm, we can achieve a
lower PCU supply rate (means lower power consumption),
and achieve a rack temperature curve which will not surpass
the maximum temperature of the observed real data; 2) With
higher value of λ, the predicted flow rate is increasing and the
predicted temperature curve is lower. This is reasonable as
the larger λ is, the higher “price” the temperature constraint
gains, which then causes the control setting optimization to
evolve towards even lower rack temperature direction. These
results show that our algorithm can be utilized our to generate
different control settings with different cooling requirements,
while optimize the corresponding energy cost. For example,
when we set λ = 1, we can control the rack temperature under
26.5 degree, while achieving cooling energy saving 13.6%.
11
30.2
Default
TS/D
CCASOT/His/D/CC
2.2
PUE
2.0
1.8
1.6
1.4
Default
TS/his/D
CCASOT/His/D
30.0
Temperature
2.4
29.8
29.6
29.4
29.2
29.0
1.2
0
2000
4000
6000
Time (unit: 6mins)
8000
28.8
10000
0
2000
4000
(a)
1.8
1.6
1.4
Default
TS/his/D
CCASOT/His/D
30.0
Temperature
PUE
2.0
29.8
29.6
29.4
29.2
29.0
1.2
0
5000
10000
Time (unit: 6mins)
15000
28.8
20000
0
5000
10000
Time (unit: 6mins)
(c)
15000
20000
(d)
30.2
Default
TS/D
CCASOT/His/D/CC
2.0
1.8
1.6
1.4
Default
TS/his/D
CCASOT/His/D
30.0
Temperature
2.2
PUE
12000
30.2
Default
TS/D
CCASOT/His/D/CC
2.2
1.2
29.8
29.6
29.4
29.2
29.0
0
5000
10000
15000
20000
Time (unit: 6mins)
25000
28.8
30000
0
5000
10000
(e)
15000
20000
Time (unit: 6mins)
25000
30000
(f)
31.5
Default
TS/D
CCASOT/His/D/CC
2.0
1.8
1.6
1.4
Default
TS/his/D
CCASOT/His/D
31.0
Temperature
2.2
PUE
10000
(b)
2.4
1.0
6000
8000
Time (unit: 6mins)
30.5
30.0
29.5
29.0
1.2
0
5000
10000
15000
20000
25000
Time (unit: 6mins)
30000
35000
(g)
28.5
0
5000
10000
15000
20000
25000
Time (unit: 6mins)
30000
35000
(h)
Fig. 5. Simulation result comparison of the Default algorithm, TS/D and CCASOT/his/r. For different test cases, PUE of the test period is shown in the
left and the data center temperature (for z1 ) is in the right; (a)(b): test case 15%; (c)(d): test case 25%; (e)(f): test case 35%; (g)(h): test case 45%. All
temperatures are in o C.
VI. R ELATED W ORKS
Cooling system control optimization has been extensively
studied. A significant portion of these works [4] [5] [6] [7]
[8] [9] focus on using a two stage optimization procedure to
find the optimal control setting. For the first stage, various
kinds of models are built to evaluate the policy quality, such
as approximate mathematical model [4] [5] [6] [7] [8] and
neural network model [9]. For the second stage, the control
variables are then optimized with the built model. Another area
of research for cooling control optimization is to optimize the
ice-storage system [29] [30], so that it can be used for cooling
when the electricity price is high.
In addition to the cooling system optimization surveyed
above, there exists an extensive body of works that focuses on
the ICT of the data center. For example, in [31] [32] [33], the
workload dispatch problem is studied to optimize the thermal
map in the data center to improve cooling efficiency. In [34],
the workload dispatch problem is extended to geographically
distributed data centers. Moreover, renewable energy is incorporated in [35] to reduce the cooling cost.
Reinforcement Learning deals with agents that learn to take
better actions directly from experience of interacting with the
environment. With the recent rapid advancement of RL, deep
reinforcement learning (DRL) [11] has shown its strength in
various fields. The deep Q-network (DQN) proposed in [11]
applies a neural network approximation to the Q table in
Q-learning [36]. The subsequent studies on DQN have been
focusing on improving the training stability of the framework
such as in [37] and extending the framework to solve problems
with continuous control variables [12]. Various applications of
12
19500
19000
Air flow rate (CFM)
3UHFLVLRQ FRROLQJ
XQLW 3&8
18500
18000
17500
17000
16500
0DQDJHPHQW
UDFNV
6XSHUFRPSXWLQJ
UDFNV
&ROG DLVOH
0
200
400
600
Time (step size: 3mins)
800
1000
800
1000
(a)
27.5
Power
Fig. 6. 3-D model of the target system in NSCC. There are 26 racks, which
are cooled by the PCUs. The cold air supplied by the PCUs first goes through
the raised floor to the cold aisle, and then cools the racks before returning to
the PCUs.
7.4
7.2
7.0
6.8
6.6
6.4
6.2
6.0
5.8
1e12
0
Validation Predicted Results
Validation Real Data
200
400
600
Time (step size: 3mins)
800
1000
(a)
Temperature (C)
27.0
Validation Predicted Results
Validation Real Data
26.5
26.0
25.5
0
200
400
600
Time (step size: 3mins)
800
1000
(b)
Fig. 7. The Q-network can predict the energy and rack temperature with
high accuracy. We use part of the training data to validate the accuracy of
the Q-network, the results are shown in (a) power consumption and (b) rack
temperature.
deep reinforcement learning have been proposed such as video
processing [38] and text-based game [39].
VII. C ONCLUSION
Data center powers the modern society as the infrastructure
for information storage, processing, and dissemination. At the
same time, data center consumes a formidable amount of
electricity, among which a large portion is used in cooling.
To develop an optimal control policy for the complex cooling
system for a data center is a difficult task. We propose and
Temperature (C)
27.0
26.5
26.0
25.5
25.0
0
200
400
600
Time (step size: 3mins)
(b)
Fig. 8. Test results on the NSCC data trace when λ changed from 0.0 to
10: (a) control setting (airflow rate) predicted by the proposed algorithm; (b)
the correspondingly predicted rack temperature. Darker color curves are with
smaller λ settings, and the blue curve (with ”+” marker) is the real data
trace collected in the test period. Note lower airflow rate means lower energy
consumption. When we set λ=1, we can achieve energy saving 13.6% while
the maximum rack temperature is lower than 26.5 degree.
verify an end-to-end approach for the control optimization of
the cooling system for a data center. Compared to the existed
TS optimization method, the end-to-end method can directly
optimize a policy network based on the observed historical
data, while the policy can output the optimized control settings
for any given state. Adapted from DDPG and the actor-critic
framework, our algorithm is a batch off-policy algorithm, with
which we tested various different DRL settings like recurrent
decision making, discounted return, etc.
We evaluate the proposed algorithm on the simulation
platform EnergyPlus. The simulation results show that our
method can maintain the data center temperature within the
predefined threshold while achieving low PUE, and saving
about 10% cooling energy compared to a canonical two
stage optimization approach. We also apply our method to
a real data trace collected from NSCC of Singapore, and
the results show the high approximation accuracy and the
predicted control setting can clearly reduce the cooling cost by
about 14% while maintain the rack intake temperature under
predefined threshold. The results prove that our algorithm can
successfully learn the system dynamics from the monitoring
data and can help to improve the cooling efficiency.
R EFERENCES
[1] “Cisco Global Cloud Index: Forecast and Methodology, 2014-2019,”
http://www.cisco.com/c/en/us/solutions/collateral/service-provider/
global-cloud-index-gci/Cloud Index White Paper.html,
accessed:
2015-07-01.
13
[2] “America’s data centers consuming and wasting growing amounts of
energy,” http://www.nrdc.org/energy/data-center-efficiency-assessment.
asp, accessed: 2015-07-01.
[3] “Data Centre Energy Efficiency Benchmarking - E2 Singapore,”
http://www.e2singapore.gov.sg/DATA/0/docs/Resources/
NEADCEnergyBenchmarkingSummary-FinalReport(3).pdf, accessed:
2015-07-01.
[4] J. Sun and A. Reddy, “Optimal control of building HVAC systems using complete simulation-based sequential quadratic programming (csbsqp),” Building and environment, vol. 40, no. 5, pp. 657–669, 2005.
[5] B. Ahn and J. Mitchell, “Optimal control development for chilled water
plants using a quadratic representation,” Energy and Buildings, vol. 33,
no. 4, pp. 371–378, 2001.
[6] L. Lu, W. Cai, L. Xie, S. Li, and Y. C. Soh, “Hvac system optimizationinbuilding section,” Energy and Buildings, vol. 37, no. 1, pp. 11–22, 2005.
[7] Y.-C. Chang, “A novel energy conservation methodoptimal chiller loading,” Electric Power Systems Research, vol. 69, no. 2, pp. 221–226,
2004.
[8] Z. Ma and S. Wang, “An optimal control strategy for complex building
central chilled water systems for practical and real-time applications,”
Building and Environment, vol. 44, no. 6, pp. 1188–1198, 2009.
[9] T. Chow, G. Zhang, Z. Lin, and C. Song, “Global optimization of
absorption chiller system by genetic algorithm and neural network,”
Energy and buildings, vol. 34, no. 1, pp. 103–109, 2002.
[10] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.
MIT press Cambridge, 1998, vol. 1, no. 1.
[11] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller, “Playing atari with deep reinforcement learning,” arXiv preprint arXiv:1312.5602, 2013.
[12] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa,
D. Silver, and D. Wierstra, “Continuous control with deep reinforcement
learning,” arXiv preprint arXiv:1509.02971, 2015.
[13] “DeepMind
AI
Reduces
Google
Data
Centre
Cooling
Bill
by
40%,”
https://deepmind.com/blog/
deepmind-ai-reduces-google-data-centre-cooling-bill-40/,
accessed:
2017-03-01.
[14] M. Dayarathna, Y. Wen, and R. Fan, “Data center energy consumption modeling: A survey,” IEEE Communications Surveys & Tutorials,
vol. 18, no. 1, pp. 732–794, 2016.
[15] W. Zhang, Y. Wen, Y. W. Wong, K. C. Toh, and C.-H. Chen, “Towards
joint optimization over ict and cooling systems in data centre: A survey,”
IEEE Communications Surveys & Tutorials, vol. 18, no. 3, pp. 1596–
1616, 2016.
[16] W. Xia, Y. Wen, K.-C. Toh, and Y.-W. Wong, “Toward green data
centers as an interruptible load for grid stabilization in singapore,” IEEE
Communications Magazine, vol. 53, no. 11, pp. 192–198, 2015.
[17] Y. Li, H. Hu, Y. Wen, and J. Zhang, “Learning-based power prediction
for data centre operations via deep neural networks,” in Proceedings
of the 5th International Workshop on Energy Efficient Data Centres.
ACM, 2016, p. 6.
[18] J. Yin, P. Sun, Y. Wen, H. Gong, M. Liu, X. Li, H. You, J. Gao,
and C. Lin, “Cloud3dview: an interactive tool for cloud data center
operations,” in ACM SIGCOMM Computer Communication Review,
vol. 43, no. 4. ACM, 2013, pp. 499–500.
[19] I. Grondman, L. Busoniu, G. A. Lopes, and R. Babuska, “A survey of
actor-critic reinforcement learning: Standard and natural policy gradients,” IEEE Transactions on Systems, Man, and Cybernetics, Part C
(Applications and Reviews), vol. 42, no. 6, pp. 1291–1307, 2012.
[20] D. B. Crawley, L. K. Lawrie, F. C. Winkelmann, W. F. Buhl, Y. J.
Huang, C. O. Pedersen, R. K. Strand, R. J. Liesen, D. E. Fisher, M. J.
Witte et al., “Energyplus: creating a new-generation building energy
simulation program,” Energy and buildings, vol. 33, no. 4, pp. 319–331,
2001.
[21] “Wikimedia Grid Report,” https://ganglia.wikimedia.org/latest/, accessed: 2015-07-01.
[22] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier neural
networks.” in Aistats, vol. 15, no. 106, 2011, p. 275.
[23] D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, and M. Riedmiller,
“Deterministic policy gradient algorithms,” in Proceedings of the 31st
International Conference on Machine Learning (ICML-14), 2014, pp.
387–395.
[24] B. Bakker, “Reinforcement learning memory,” 2002.
[25] M. D. Zeiler, “Adadelta: an adaptive learning rate method,” arXiv
preprint arXiv:1212.5701, 2012.
[26] E. Jones, T. Oliphant, P. Peterson et al., “Open source scientific tools
for python,” 2001.
[27] “Deep
Deterministic
Policy
Gradients
in
TensorFlow,”
http://pemami4911.github.io/blog/2016/08/21/ddpg-rl.html,
accessed:
2017-03-01.
[28] T. Tieleman and G. Hinton, “Lecture 6.5-rmsprop: Divide the gradient
by a running average of its recent magnitude,” COURSERA: Neural
networks for machine learning, vol. 4, no. 2, 2012.
[29] J. E. Braun, “Reducing energy costs and peak electrical demand through
optimal control of building thermal storage,” ASHRAE transactions,
vol. 96, no. 2, pp. 876–888, 1990.
[30] C.-C. Lo, S.-H. Tsai, and B.-S. Lin, “Ice storage air-conditioning system
simulation with dynamic electricity pricing: A demand response study,”
Energies, vol. 9, no. 2, p. 113, 2016.
[31] E. Pakbaznia and M. Pedram, “Minimizing data center cooling and
server power costs,” in Proceedings of the 2009 ACM/IEEE international
symposium on Low power electronics and design. ACM, 2009, pp.
145–150.
[32] L. Li et al., “Coordinating liquid and free air cooling with workload
allocation for data center power minimization,” in 11th International
Conference on Autonomic Computing (ICAC 14). USENIX Association,
2014.
[33] A. Banerjee, T. Mukherjee, G. Varsamopoulos, and S. K. Gupta,
“Integrating cooling awareness with thermal aware workload placement
for hpc data centers,” Sustainable Computing: Informatics and Systems,
vol. 1, no. 2, pp. 134–150, 2011.
[34] C. Chen, B. He, and X. Tang, “Green-aware workload scheduling in
geographically distributed data centers,” in 2012 IEEE 4th International
Conference on Cloud Computing Technology and Science (CloudCom).
IEEE, 2012, pp. 82–89.
[35] Z. Liu, Y. Chen, C. Bash, A. Wierman, D. Gmach, Z. Wang, M. Marwah,
and C. Hyser, “Renewable and cooling aware workload management for
sustainable data centers,” in ACM SIGMETRICS Performance Evaluation Review, vol. 40, no. 1. ACM, 2012, pp. 175–186.
[36] C. J. Watkins and P. Dayan, “Q-learning,” Machine learning, vol. 8, no.
3-4, pp. 279–292, 1992.
[37] H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning
with double q-learning,” CoRR, abs/1509.06461, 2015.
[38] J. Koutnı́k, J. Schmidhuber, and F. Gomez, “Evolving deep unsupervised
convolutional networks for vision-based reinforcement learning,” in Proceedings of the 2014 Annual Conference on Genetic and Evolutionary
Computation. ACM, 2014, pp. 541–548.
[39] K. Narasimhan, T. Kulkarni, and R. Barzilay, “Language understanding
for text-based games using deep reinforcement learning,” arXiv preprint
arXiv:1506.08941, 2015.
| 3cs.SY
|
arXiv:1510.06207v3 [math.ST] 4 May 2016
Functional delta-method for the
bootstrap of quasi-Hadamard
differentiable functionals
Eric Beutner∗
Henryk Zähle†
Abstract
The functional delta-method provides a convenient tool for deriving the asymptotic distribution of a plug-in estimator of a statistical functional from the asymptotic distribution of the respective empirical process. Moreover, it provides a tool
to derive bootstrap consistency for plug-in estimators from bootstrap consistency
of empirical processes. It has recently been shown that the range of applications
of the functional delta-method for the asymptotic distribution can be considerably enlarged by employing the notion of quasi-Hadamard differentiability. Here
we show in a general setting that this enlargement carries over to the bootstrap.
That is, for quasi-Hadamard differentiable functionals bootstrap consistency of
the plug-in estimator follows from bootstrap consistency of the respective empirical process. This enlargement often requires convergence in distribution of the
bootstrapped empirical process w.r.t. a nonuniform sup-norm. The latter is not
problematic as will be illustrated by means of examples.
Keywords: Bootstrap; Functional delta-method; Quasi-Hadamard differentiability;
Statistical functional; Weak convergence for the open-ball σ-algebra.
∗
†
Department of Quantitative Economics, Maastricht Univ., [email protected]
Department of Mathematics, Saarland University, [email protected]
1
1. Introduction
The bootstrap is a widely used technique to approximate the unknown error distribution of estimators. Since the seminal paper by Efron (1979) many variants of his
bootstrap procedure have been introduced in the literature. Furthermore, the bootstrap
has quickly been extended to other data than a sample of independent and identically
distributed random variables. For general accounts on the bootstrap one may refer to
Efron and Tibshirani (1994), Shao and Tu (1995), Davison and Hinkely (1997), Lahiri
(2003), among others.
For a (tangentially) Hadamard differentiable map f the functional delta-method leads
to the asymptotic distribution of an (f (Tbn ) − f (θ)) whenever the asymptotic distribution
of an (Tbn − θ) is known. Here Tbn is an estimator for a (possibly infinite-dimensional) parameter θ, and (an ) is a sequence of real numbers tending to infinity such that an (Tbn −θ)
has a non-degenerate limiting distribution. This extends to the bootstrap, i.e. bootstrap
consistency of an (f (Tbn∗ ) − f (Tbn )) follows from bootstrap consistency of an (Tbn∗ − Tbn ) for
(tangentially) Hadamard differentiable f ; see, for instance, van der Vaart and Wellner
(1996, Theorems 3.9.11 and 3.9.13). Here Tbn∗ is a bootstrapped version of Tbn based on
some random mechanism. For a recent partial generalization of these results, see also
Volgushev and Shao (2014). Parr (1985) established a functional delta-method for the
bootstrap of Fréchet differentiable maps f , and Cuevas and Romo (1997) obtained a
corresponding result for the so-called smoothed bootstrap.
A drawback of the classical functional delta-method is its restricted range of applications. For many statistical functionals f (including classical L-, V- and M-functionals)
the condition of (tangential) Hadamard differentiability is simply too strong. For this
reason Beutner and Zähle (2010) introduced the notion of quasi-Hadamard differentiability, which is weaker than (tangential) Hadamard differentiability but still strong
enough to obtain a generalized version of the classical functional delta-method; see also
the Appendix C. Combined with results for weak convergence of empirical processes
w.r.t. nonuniform sup-norms the concept of quasi-Hadamard differentiability led to some
new weak convergence results for plug-in estimators of statistical functionals based on
dependent data; see Beutner and Zähle (2010, 2012), Ahn and Shyamalkumar (2011),
Beutner et al. (2012), Krätschmer et al. (2015), and Krätschmer and Zähle (2016). See
also Beutner and Zähle (2014) and Buchsteiner (2015) for some recent results on weak
convergence of empirical processes w.r.t. nonuniform sup-norms.
In this article, we will show that the notion of quasi-Hadamard differentiability admits
even a functional delta-method for the bootstrap. This enlarges the set of functionals
f for which bootstrap consistency of an (f (Tbn∗ ) − f (Tbn )) follows immediately from bootstrap consistency of an (Tbn∗ − Tbn ). To illustrate this, let us briefly discuss distortion risk
functionals as examples for f where the parameter θ is a distribution function F on the
real line, Tbn represents the empirical distribution function Fbn of n real-valued random
2
variables with distribution function F , and Tbn∗ corresponds to a bootstrapped version
Fbn∗ of Fbn .
Given a continuous concave distortion function g, i.e. a concave function g : [0, 1] →
[0, 1] being continuous at 0 and satisfying g(0) = 0 = 1 − g(1), the corresponding
distortion risk functional fg : D(fg ) → R is defined by
fg (F ) :=
ˆ
0
−∞
g(F (t)) dt −
ˆ
∞
0
1 − g(F (t)) dt,
(1)
where D(fg ) is a suitable subset of the set of all distribution functions F for which
both integrals on the right-hand side are finite. Note that distortion risk functionals
associated with continuous concave distortion functions correspond to coherent distortion risk measures (cf. Example 4.5) which are of special interest in mathematical finance and actuarial mathematics. It was discussed in Beutner and Zähle (2010) and
Krätschmer et al. (2015) that these functionals are typically not Hadamard differentiable w.r.t. the usual sup-norm k · k∞ but only quasi-Hadamard differentiable w.r.t.
suitable nonuniform sup-norms kvkφ := kvφk∞ stronger than k · k∞ (i.e. with continuous weight functions φ : R → [1, ∞) satisfying lim|x|→∞ φ(x) = ∞). The functional
delta-method in the form of Corollary 4.2 below then shows that an (fg (Fbn∗ ) − fg (Fbn ))
has the same limiting distribution as an (fg (Fbn ) − fg (F )) whenever the bootstrapped empirical process an (Fbn∗ − Fbn ) converges in distribution to the same limit as the empirical
process an (Fbn − F ). As “differentiability” can be obtained only for certain nonuniform
sup-norms k · kφ , the latter convergence in distribution has to be guaranteed for exactly
these nonuniform sup-norms k · kφ . Fortunately, such results can be easily obtained
from Donsker results for appropriate classes of functions; see Sections 5.1–5.2 for examples. So the notion of quasi-Hadamard differentiability together with the functional
delta-method based on it provides an interesting field of applications for the bootstrap of
Donsker classes. We emphasize that our approach leads in particular to new bootstrap
results for empirical distortion risk measures based on β-mixing data; for details and
other examples see Section 5.3.
It is worth recalling that the empirical process an (Fbn − F ), regarded as a mapping
from Ω to the nonseparable space of all bounded càdlàg functions equipped with the
sup-norm, is not measurable w.r.t. the Borel σ-algebra. This problem was first observed by Chibisov (1965) and carries over to nonuniform sup-norms. There are different ways to deal with this fact; for a respective discussion see, for instance, Section
1.1 in van der Vaart and Wellner (1996). One possibility is to use the concept of weak
convergence (or convergence in distribution) in the Hoffmann-Jørgensen sense; see, for
instance, van der Vaart and Wellner (1996), Dudley (1999), Lahiri (2003), and Kosorok
(2010). Another possibility is to use the open-ball σ-algebra w.r.t. which the empirical
process is measurable. Here we work throughout with the open-ball σ-algebra and weak
convergence (and convergence in distribution) as defined in Billingsley (1999, Section
3
6); see also Dudley (1966, 1967), Pollard (1984), and Shorack and Wellner (1986). This
implies in particular that we have to take proper care of the measurability of the maps
an (Tbn − θ) and an (Tbn∗ − Tbn ) for every n ∈ N.
The rest of the article is organized as follows. In Section 2 we briefly explain the
setting chosen here and give some definitions that will be used throughout. The main
result and its proof are presented in Sections 3 and 6, respectively. Applications of our
main result are given in Section 4 and illustrated in Section 5. Additional definitions and
results that are needed for our main result are given in the Appendix. The Appendix
is organized as follows. In Sections A and B we give some results on weak convergence,
convergence in distribution, and convergence in probability for the open-ball σ-algebra
which are needed in Section C. In Section C we first present an extended Continuous
Mapping theorem for convergence in distribution for the open-ball σ-algebra. This
complements the extended Continuous Mapping theorems for weak convergence for the
Borel σ-algebra and for convergence in distribution in the Hoffmann-Jørgensen sense
which are already known from the literature. In the second part of Section C we use the
extended Continuous Mapping theorem to prove an extension (compared to Theorem
4.1 in Beutner and Zähle (2010)) of the functional delta-method based on the notion
of quasi-Hadamard differentiability. This extension is needed for the proof of our main
result, i.e. for the proof of a functional delta-method for the bootstrap. Two results that
ensure measurability of maps involved in our approach are given in Section D.
2. Basic definitions
In this section we introduce some notation and basic definitions. As mentioned in the
introduction, weak convergence and convergence in distribution will always be considered
for the open-ball σ-algebra. Borrowed from Billingsley (1999, Section 6) we will use
the terminology weak◦ convergence (symbolically ⇒◦ ) and convergence in distribution◦
(symbolically ❀◦ ). For details see the Appendices A and B. In a separable metric
space the notions of weak◦ convergence and convergence in distribution◦ boil down to
the conventional notions of weak convergence and convergence in distribution for the
Borel σ-algebra. In this case we also use the symbols ⇒ and ❀ instead of ⇒◦ and ❀◦ ,
respectively.
Let V be a vector space and E be a subspace of V. Let k · kE be a norm on E and
◦
B be the corresponding open-ball σ-algebra on E. Let (Ω, F , P) be a probability space,
and (Tbn ) be a sequence of maps
Tbn : Ω −→ V.
Regard ω ∈ Ω as a sample drawn from P, and Tbn (ω) as a statistic derived from ω. Let
θ ∈ V, and (an ) be a sequence of positive real numbers tending to ∞. Assume that
4
an (Tbn − θ) takes values only in E and is (F , B◦ )-measurable for every n ∈ N, and that
an (Tbn − θ) ❀◦ ξ
in (E, B◦ , k · kE )
(2)
for some (E, B◦ )-valued random variable ξ.
Now, let (Ω′ , F ′ , P′ ) be another probability space and set
(Ω, F, P) := (Ω × Ω′ , F ⊗ F ′ , P ⊗ P′ ).
The probability measure P′ represents a random experiment that is run independently
of the random sample mechanism P. In the sequel, Tbn will frequently be regarded as a
map defined on the extension Ω of Ω. Let
Tbn∗ : Ω −→ V
be any map and assume that an (Tbn∗ −Tbn ) takes values only in E and is (F, B◦ )-measurable
for every n ∈ N. Since Tbn∗ (ω, ω ′) depends on both the original sample ω and the outcome ω ′ of the additional independent random experiment, we may regard Tbn∗ as a
bootstrapped version of Tbn . For the formula display (3) in the following Definition 2.1,
note that the mapping ω ′ 7→ an (Tbn∗ (ω, ω ′) − Tbn (ω)) is (F ′, B◦ )-measurable for every
fixed ω ∈ Ω, because an (Tbn∗ − Tbn ) is (F , B◦ )-measurable with F = F ⊗ F ′ . That is,
an (Tbn∗ (ω, ·) − Tbn (ω)) can be seen as an (E, B◦ )-valued random variable on (Ω′ , F ′ , P′ ) for
every fixed ω ∈ Ω.
Definition 2.1 (Bootstrap version almost surely) We say that (Tbn∗ ) is almost
surely a bootstrap version of (Tbn ) w.r.t. the convergence in (2) if
an (Tbn∗ (ω, ·) − Tbn (ω)) ❀◦ ξ
in (E, B◦ , k · kE ),
P-a.e. ω.
(3)
Next we intend to introduce the notion of bootstrap version in (outer) probability. To
this end let the map Pn : Ω × B◦ → [0, 1] be defined by
Pn ((ω, ω ′), A) := Pn (ω, A) := P′ ◦ {an (Tbn∗ (ω, ·) − Tbn (ω))}−1[A],
(ω, ω ′) ∈ Ω, A ∈ B◦ .
(4)
∗ b
b
It provides a conditional distribution of an (Tn −Tn ) given Π, where the (F, F )-measurable
map Π : Ω → Ω is defined by
Π(ω, ω ′) := ω.
(5)
This follows from Lemma D.2 (with X(ω, ω ′) = g(ω, ω ′) = an (Tbn∗ (ω, ω ′) − Tbn (ω)) and
Y = Π). Informally, Π(ω, ω ′) specifies that part of the realization (ω, ω ′) of the extended
random mechanism P⊗P′ that represents the “observed data”; see also Remark 2.5 below
and the discussion preceding it. By definition Pn is a probability kernel from (Ω, σ(Π))
to (E, B◦ ). However, it is directly clear from (4) that Pn can also be seen as a probability
kernel from (Ω, F ) to (E, B◦ ).
5
Let d◦BL denote the bounded Lipschitz distance (defined in (44) in the Appendix A)
on the set M◦1 of all probability measures on (E, B◦ ). Note that a sequence (µn ) ⊆ M◦1
converges weak◦ ly to some µ0 ∈ M◦1 which concentrates on a separable set, if and only
if d◦BL (µn , µ0 ) → 0; cf. Theorem A.3. In general the mapping ω 7→ d◦BL (Pn (ω, ·), law{ξ})
is not necessarily (F , B(R+ ))-measurable. For this reason we have to use the outer
probability in (6). Recall that the outer probability Pout [S] of an arbitrary subset S ⊆ Ω
is defined to be the infimum of P[S] over all S ∈ F with S ⊇ S.
Definition 2.2 (Bootstrap version in (outer) probability) We say that (Tbn∗ ) is a
bootstrap version in outer probability of (Tbn ) w.r.t. the convergence in (2) if
lim Pout
n→∞
ω ∈ Ω : d◦BL(Pn (ω, ·), law{ξ}) ≥ δ = 0
for all δ > 0.
(6)
When (E, k · kE ) is separable, we may replace in (6) the outer probability Pout by the
ordinary probability P and we will say that (Tbn∗ ) is a bootstrap version in probability of
(Tbn ) w.r.t. the convergence in (2).
The second part of Definition 2.2 can be justified as follows. The assumed separability
of (E, k · kE ) implies that M◦1 is just the set M1 of all Borel probability measures on E
and that ω 7→ Pn (ω, ·) can be seen as an (F , σ(Ow ))-measurable mapping from Ω to M1
(cf. Lemma D.1); here Ow refers to the weak topology on M1 (cf. Remark A.1). By the
reverse triangle inequality for metrics we also have that the mapping µ 7→ dBL (µ, law{ξ})
is continuous (recall that dBL := d◦BL is a metric when (E, k · kE ) is separable) and thus
(σ(Ow ), B(R+ ))-measurable. It follows that the mapping ω 7→ dBL (Pn (ω, ·), law{ξ}) is
(F , B(R+ ))-measurable.
As our interest lies in deriving bootstrap results for functionals f of Tbn∗ from bootstrap
results for Tbn∗ itself, we introduce some more notation and restate Definition 2.2 for f (Tbn∗ ).
e k · k e ) be another normed vector space and assume that k · k e is separable. In
Let (E,
E
E
e Denote
particular, the open-ball σ-algebra coincides with the Borel σ-algebra Be on E.
e B).
f1 the set of all probability measures on (E,
e Let
by M
e
f : Vf −→ E
be any map defined on some subset Vf ⊆ V. Assume that Tbn and Tbn∗ take values
e
only in Vf and that an (f (Tbn∗ ) − f (Tbn )) is (F, B)-measurable.
Moreover let the map
e
e
Pn : Ω × B → [0, 1] be defined by
e
Pen ((ω, ω ′), A) := Pen (ω, A) := P′ ◦{an (f (Tbn∗ (ω, ·))−f (Tbn (ω)))}−1[A], (ω, ω ′) ∈ Ω, A ∈ B.
(7)
∗
It provides a conditional distribution of an (f (Tbn ) − f (Tbn )) given Π, where Π is as in (5).
This follows from Lemma D.2 (with X(ω, ω ′) = g(ω, ω ′) = an (f (Tbn∗ (ω, ω ′)) − f (Tbn (ω)))
e B).
e However,
and Y = Π). By definition Pen is a probability kernel from (Ω, σ(Πn )) to (E,
6
it is directly clear from (7) that Pen can also be seen as a probability kernel from (Ω, F )
e B).
e Finally assume that
to (E,
an (f (Tbn ) − f (θ)) ❀ ξe
e B,
e k · ke )
in (E,
E
(8)
e B)-valued
e
for some (E,
random variable ξe and let deBL denote the bounded Lipschitz
f1 as defined in (44).
distance on M
Definition 2.3 (Bootstrap version in probability) We say that (f (Tbn∗ )) is a bootstrap version in probability of (f (Tbn )) w.r.t. the convergence in (8) if
e ≥δ = 0
lim P ω ∈ Ω : deBL (Pen (ω, ·), law{ξ})
n→∞
for all δ > 0.
(9)
e is (F , B(R+ ))-measurable. Indeed,
Note that the mapping ω 7→ deBL (Pen (ω, ·), law{ξ})
e k · k e ) is
one can argue as subsequent to Definition 2.2, because we assumed that (E,
E
separable.
Remark 2.4 Note that (9) implies that (9) still holds when the bounded Lipschitz
f1 which generates the weak topology.
distance deBL is replaced by any other metric on M
When (E, k · kE ) is separable, then the same is true for (6).
✸
We conclude this section with some comments on the probability kernel Pn defined in
(4). As mentioned above, it is a conditional distribution of an (Tbn∗ − Tbn ) given Π, where
to some extent Π(ω, ω ′) = ω can be seen as the “observable” sample. On the other hand,
for technical reasons the sample space Ω is often so complex so that only a portion Πn (ω)
of an element ω ∈ Ω can indeed be “observed”. For instance, when the sample space is
an infinite product space, i.e. (Ω, F ) = (S N , S ⊗N ) for some measurable space (S, S), then
de facto one can only observe a finite-dimensional sample, say the first n coordinates
(ω1 , . . . , ωn ) of the infinite-dimensional sample ω = (ω1 , ω2 , . . .) ∈ S N . In this case it is
obviously appealing to interpret Pn as a conditional distribution of an (Tbn∗ − Tbn ) given
Πn , where Πn : S N × Ω′ → S n is given by
Πn ((ω1 , ω2 , . . .), ω ′) := (ω1 , . . . , ωn ).
(10)
Under additional mild assumptions this is indeed possible. This follows from the next
Remark 2.5 if we take there Πn as given in (10) and (Ω(n) , F (n) ) equal to (S n , S ⊗n ).
Analogously one can regard Pen defined in (7) as a conditional distribution of an (f (Tbn∗ ) −
f (Tbn )) given Πn .
Remark 2.5 Let (Ω(n) , F (n) ) be a measurable space and Πn : Ω → Ω(n) be an (F, F (n) )measurable map for every n ∈ N. Assume that for every n ∈ N the value Πn (ω, ω ′)
7
depends only on ω and that there exist maps τn : Ω(n) → V and τn∗ : Ω(n) × Ω′ → V
such that
τn (Πn (ω, ω ′)) = Tbn (ω)
and
and
τn∗ (Πn (ω, ω ′), ω ′) = Tbn∗ (ω, ω ′)
gn (ω (n) , ω ′) := an τn∗ (ω (n) , ω ′) − τn (ω (n) ) ,
for all ω ∈ Ω, ω ′ ∈ Ω′
(11)
(ω (n) , ω ′) ∈ Ω(n) × Ω′
(12)
provides an (F (n) ⊗ F ′ , B◦ )-measurable map gn : Ω(n) × Ω′ → E. (This implies in
particular that an (Tbn∗ − Tbn ) takes values only in E and is (F , B◦ )-measurable). Then the
map Pn : Ω×B◦ → [0, 1] defined by (4) provides a conditional distribution of an (Tbn∗ − Tbn )
given Πn . This follows again from Lemma D.2 (with X(ω, ω ′) = an (Tbn∗ (ω, ω ′) − Tbn (ω)),
Y = Πn , and g = gn ).
✸
3. Abstract delta-method for the bootstrap
Theorem 3.1 below establishes an abstract delta-method for the bootstrap for quasiHadamard differentiable maps. It uses the notation and definitions introduced in Section
2. More precisely, let V, (E, k · kE ), (Ω, F , P), (Ω′ , F ′, P′ ), (Ω, F , P), Tbn , Tbn∗ , Pn , f , Vf ,
e k · k e ), and Pen be as in Section 2. As before assume that (E,
e k · k e ) is separable, and
(E,
E
E
∗
that Tbn and Tbn take values only in Vf .
As already discussed in the introduction, in statistical applications the role of Tbn is often played by the empirical distribution function of n identically distributed random variables (sample), so that the plug-in estimator f (Tbn ) can be represented as a function of the
sample. This special case will be studied in detail in Section 4. Due to the measurability
problems discussed in the introduction we work with the open-ball σ-algebra B◦ in our
general setting. This is different from the conventional functional delta-method for the
bootstrap in the form of van der Vaart and Wellner (1996, Theorems 3.9.11 and 3.9.13)
where the measurability problem is overcome by using the concept of convergence in
distribution in the Hoffmann-Jørgensen sense. Moreover, compared to the conventional functional delta-method we work with a weaker notion of differentiability, namely
with quasi-Hadamard differentiability. This kind of differentiability was introduced by
Beutner and Zähle (2010) and is recalled in Definition C.3 in the Appendix.
Theorem 3.1 (Delta-method for the bootstrap) Let θ ∈ Vf . Let E0 ⊆ E be a
separable subspace and assume that E0 ∈ B◦ . Let (an ) be a sequence of positive real
numbers tending to ∞, and consider the following conditions:
(a) an (Tbn − θ) takes values only in E, is (F , B◦ )-measurable, and satisfies
an (Tbn − θ) ❀◦ ξ
8
in (E, B◦ , k · kE )
(13)
for some (E, B◦ )-valued random variable ξ on some probability space (Ω0 , F0 , P0 )
with ξ(Ω0 ) ⊆ E0 .
e is (F , B)-measurable.
e
(b) The map f (Tbn ) : Ω → E
(c) The map f is quasi-Hadamard differentiable at θ tangentially to E0 hEi with quasiHadamard derivative f˙θ in the sense of Definition C.3.
(d) The quasi-Hadamard derivative f˙θ can be extended from E0 to E such that the
e is linear and (B◦ , B)-measurable.
e
extension f˙θ : E → E
Moreover, the extension
˙
e
fθ : E → E is continuous at every point of E0 .
e is (F , B)-measurable.
e
(e) The map f (Tbn∗ ) : Ω → E
(f ) an (Tbn∗ − θ) and an (Tbn∗ − Tbn ) take values only in E and are (F , B◦ )-measurable, and
(Tbn∗ ) is almost surely a bootstrap version of (Tbn ) w.r.t. the convergence in (13) in
the sense of Definition 2.1. The latter means that
an (Tbn∗ (ω, ·) − Tbn (ω)) ❀◦ ξ
in (E, B◦ , k · kE ),
P-a.e. ω.
(14)
(f ’) an (Tbn∗ − θ) and an (Tbn∗ − Tbn ) take values only in E and are (F , B◦ )-measurable, and
(Tbn∗ ) is a bootstrap version in outer probability of (Tbn ) w.r.t. the convergence in
(13) in the sense of Definition 2.2. The latter means that
lim Pout
n→∞
ω ∈ Ω : d◦BL (Pn (ω, ·), law{ξ}) ≥ δ = 0
for all δ > 0.
(15)
Then the following assertions hold:
e
(i) If conditions (a)–(c) hold, then an (f (Tbn ) − f (θ)) and f˙θ (ξ) are respectively (F , B)e
and (F0 , B)-measurable,
and
an (f (Tbn ) − f (θ)) ❀ f˙θ (ξ)
e B,
e k · k e ).
in (E,
E
(16)
e
(ii) If conditions (a)–(f ) hold, then an (f (Tbn∗ )−f (Tbn )) and f˙θ (ξ) are respectively (F , B)∗
e
b
b
and (F0 , B)-measurable, and (f (Tn )) is a bootstrap version in probability of (f (Tn ))
w.r.t. the convergence in (16) in the sense of Definition 2.3. The latter means that
lim P ω ∈ Ω : deBL Pen (ω, ·), law{f˙θ (ξ)} ≥ δ = 0
n→∞
for all δ > 0.
(iii) Assertion (ii) still holds when assumption (f ) is replaced by (f ’).
9
(17)
Recall that (E, k · kE ) was not assumed to be separable, so that the mapping ω 7→
is not necessarily (F , B(R+ ))-measurable. Further note that the
Counterexample 1.9.4 in van der Vaart and Wellner (1996) (where Pout [|ξn − 0| ≥ δ] =
P[1out
Bn ≥ δ] = 1 obviously holds for every n ∈ N and δ ∈ (0, 1), with ξn := 1Bn ) shows that
in general P-a.s. convergence of a sequence (ξn ) of non-(F , B(R))-measurable functions
ξn : Ω → R does not imply convergence in outer probability of (ξn ). In particular it is
not clear to us whether or not condition (f) implies condition (f’). For that reason we
consider both conditions separately.
Note that in contrast to the conventional functional delta-method in the form of
van der Vaart and Wellner (1996, Theorems 3.9.11 and 3.9.13) condition (a) of Theorem 3.1 does not involve convergence in distribution in the Hoffmann-Jørgensen sense
(based on outer integrals) and condition (f) of Theorem 3.1 does not involve the concept
of convergence in outer probability. Thus assertion (ii) of Theorem 3.1 shows in particular that a comprehensive version of the functional delta-method for the bootstrap can
be stated without using the concepts of outer integrals and outer probabilities. Indeed,
(part (ii) of) Theorem 3.1 in the form of (part (ii) of) Corollary 4.2 below (together with
Lemmas 5.1 and 5.3) covers plenty of classical plug-in estimators.
d◦BL (Pn (ω, ·), law{ξ})
4. Application to plug-in estimators of statistical
functionals
Let D be the space of all càdlàg functions v on R with finite sup-norm kvk∞ :=
supt∈R |v(t)|, and D be the σ-algebra on D generated by the one-dimensional coordinate
projections πt , t ∈ R, given by πt (v) := v(t). Let φ : R → [1, ∞) be a weight function,
i.e. a continuous function being non-increasing on (−∞, 0] and non-decreasing on [0, ∞).
Let Dφ be the subspace of D consisting of all x ∈ D satisfying kxkφ := kxφk∞ < ∞ and
lim|t|→∞ |x(t)| = 0. The latter condition automatically holds when lim|t|→∞ φ(t) = ∞.
Let Dφ := D ∩ Dφ be the trace σ-algebra on Dφ . The σ-algebra on Dφ generated by
the k · kφ -open balls will be denoted by Bφ◦ . The following lemma shows that it coincides
with Dφ .
Lemma 4.1 Dφ = Bφ◦ .
Proof Without of loss of generality we assume lim|t|→∞ φ(t) = ∞. We denote by Br (x)
the k · kφ -open ball around x ∈ Dφ with radius r, that is, Br (x) := {y ∈ Dφ : kx − ykφ <
r}. On the one hand, for every t ∈ R and a ∈ R we have
[
πt−1 ((a/φ(t), ∞)) = {x ∈ Dφ : x(t) > a/φ(t)} =
Bn (xn ),
(18)
n∈N
where xn = xn,t,a is defined by xn (s) := (a + (n + 1/n)1[t,t+1/n) (s))/φ(s). Thus,
πt−1 ((b, ∞)) lies in Bφ◦ for every t ∈ R and b ∈ R. That is, πt is (Bφ◦ , B(R))-measurable.
10
Hence, Dφ ⊆ Bφ◦ . On the other hand, any open ball Br (x) can be represented as
Br (x) =
\
t∈Q
{y ∈ Dφ : |x(t) − y(t)|φ(t) < r} =
\
t∈Q
πt−1 (x(t) − r/φ(t), x(t) + r/φ(t)) ,
and so it lies in Dφ . Hence, Bφ◦ ⊆ Dφ .
✷
For any given distribution function F on the real line, let Cφ,F ⊆ Dφ be a k · kφ separable subspace and assume Cφ,F ∈ Dφ . Moreover let f : D(f ) → R be a map
defined on a set D(f ) of distribution functions of finite (not necessarily probability)
Borel measures on R. In particular, D(f ) ⊂ D. In the following, D, (Dφ , Dφ , k · kφ ),
Cφ,F , f , D(f ), and (R, B(R), | · |) will play the roles of V, (E, B◦ , k · kE), E0 , f , Vf , and
e B,
e k · k e ), respectively.
(E,
E
Let (Ω, F , P) be a probability space and F ∈ D(f ) be the distribution function of
a Borel probability measure on R. Let (Xi ) be a sequence of identically distributed
real-valued random variables on (Ω, F , P) with distribution function F . Let Fbn : Ω → D
be the empirical distribution function of X1 , . . . , Xn , which will play the role of Tbn . It
is defined by
n
1X
b
1[Xi ,∞) .
(19)
Fn :=
n i=1
Assume that Fbn takes values only in D(f ). Let (Ω′ , F ′ , P′ ) be another probability space
and set (Ω, F, P) := (Ω × Ω′ , F ⊗ F ′ , P ⊗ P′ ). Moreover let Fbn∗ : Ω → D be any map; see
Section 5 for an illustration. Assume that Fbn∗ take values only in D(f ). In the present
setting Theorem 3.1 can be reformulated as follows.
Corollary 4.2 Let F ∈ D(f ). Let (an ) be a sequence of positive real numbers tending
to ∞, and consider the following conditions:
(a) an (Fbn − F ) takes values only in Dφ and satisfies
an (Fbn − F ) ❀◦ B
in (Dφ , Dφ , k · kφ )
(20)
for some (Dφ , Dφ )-valued random variable B on some probability space (Ω0 , F0 , P0 )
with B(Ω0 ) ⊆ Cφ,F .
(b) The map f (Fbn ) : Ω → R is (F , B(R))-measurable.
(c) The map f is quasi-Hadamard differentiable at F tangentially to Cφ,F hDφ i with
quasi-Hadamard derivative f˙F in the sense of Definition C.3.
(d) The quasi-Hadamard derivative f˙F can be extended from Cφ,F to Dφ such that
the extension f˙F : Dφ → R is linear and (Dφ , B(R))-measurable. Moreover, the
extension f˙F : Dφ → R is continuous at every point of Cφ,F .
11
(e) The map f (Fbn∗ ) : Ω → R is (F , B(R))-measurable.
(f ) an (Fbn∗ − Fbn ) takes values only in Dφ and is (F , Dφ )-measurable, and (Fbn∗ ) is almost
surely a bootstrap version of (Fbn ) w.r.t. the convergence in (20) in the sense of
Definition 2.1. The latter means that
an (Fbn∗ (ω, ·) − Fbn (ω)) ❀◦ B
in (Dφ , Dφ , k · kφ ),
P-a.e. ω.
(21)
(f ’) an (Fbn∗ − Fbn ) takes values only in Dφ and is (F, Dφ )-measurable, and (Fbn∗ ) is a
bootstrap version in outer probability of (Fbn ) w.r.t. the convergence in (20) in the
sense of Definition 2.2. The latter means that (with Pn defined as in (4))
lim Pout
n→∞
ω ∈ Ω : d◦BL (Pn (ω, ·), law{B}) ≥ δ = 0
for all δ > 0,
(22)
Then the following assertions hold:
(i) If conditions (a)–(c) hold, then an (f (Fbn )−f (θ)) and f˙F (B) are respectively (F , B(R))and (F0 , B(R))-measurable, and
an (f (Fbn ) − f (F )) ❀ f˙F (B)
in (R, B(R)).
(23)
(ii) If conditions (a)–(f ) hold, then an (f (Fbn∗) − f (Fbn )) and f˙F (B) are respectively
(F, B(R))- and (F0 , B(R))-measurable, and (f (Fbn∗ )) is a bootstrap version in probability of (f (Fbn )) w.r.t. the convergence in (23) in the sense of Definition 2.3. The
latter means that (with Pen defined as in (7))
lim P ω ∈ Ω : deBL Pen (ω, ·), law{f˙F (B)} ≥ δ = 0
n→∞
for all δ > 0.
(iii) Assertion (ii) still holds when assumption (f ) is replaced by (f ’).
Proof Corollary 4.2 is a consequence of Theorem 3.1, because the measurability assumption in condition (a) and the first measurability assumption of condition (f) (respectively
(f’)) of Theorem 3.1 are automatically satisfied in the present setting. Indeed, an (Fbn −F )
is easily seen to be (F , Dφ )-measurable, and the sum of two (F , Dφ )-measurable maps
is clearly (F , Dφ )-measurable and we assumed here (through(f) (respectively (f’))) that
an (Fbn∗ − Fbn ) is (F, Dφ )-measurable.
✷
Conditions (e)–(f’) of Corollary 4.2 will be illustrated in Sections 5.1–5.2. The following examples illustrate conditions (a)–(d) of Corollary 4.2. See also Section 5.3 for
specific applications.
12
Example 4.3 (for condition (a)) Assume that X1 , X2 , . . . are i.i.d. with distribution
´
function F , and let φ be a weight function. If φ2 dF < ∞, then Theorem 6.2.1 in
Shorack and Wellner (1986) shows that
√
n(Fbn − F ) ❀◦ BF
in (Dφ , Dφ , k · kφ ),
where BF is an F -Brownian bridge, i.e. a centered Gaussian process with covariance
function Γ(t0 , t1 ) = F (t0 ∧ t1 )(1 − F (t0 ∨ t1 )). Note that BF jumps where F jumps and
that lim|t|→∞ BF (t) = 0. Thus, BF takes values only in the set Cφ,F ⊂ Dφ consisting
of all x ∈ Dφ whose discontinuities are also discontinuities of F . It was shown in
Krätschmer et al. (2015, Corollary B.4) that the set Cφ,F is k·kφ -separable and contained
in Dφ .
✸
Example 4.4 (for condition (a)) Let φ be any weight function, (Xi ) be strictly
stationary and β-mixing with distribution function F , and assume that E[φ(X1 )p ] < ∞
for some p > 2 and that the mixing coefficients satisfy βn = o(n−p/(p−2) (log n)2(p−1)/(p−2) ).
Then
√
eF
n(Fbn − F ) ❀◦ B
in (Dφ , Dφ , k · kφ ),
eF is a centered Gaussian process with covariance function Γ(t0 , t1 ) = F (t0 ∧
where B
P P
t1 )(1 − F (t0 ∨ t1 )) + 1i=0 ∞
k=2 Cov(1{X1 ≤ti } , 1{Xk ≤t1−i } ). The result follows by verifying the assumptions of Theorem 2.1 in Arcones and Yu (1994). We will verify these
eF jumps where F jumps
assumptions in the proof of Theorem 5.4 below. Note that B
eF (x) = 0. Thus, B
eF takes values only in the k · kφ -separable and
and that lim|x|→∞ B
Dφ -measurable set Cφ,F introduced in Example 4.3. For illustration, note that many
GARCH processes are strictly stationary and β-mixing; see, for instance, Francq and Zakoı̈an
(2010, Chapter 3) and Boussama et al. (2011).
✸
Further examples for condition (a) can be found in Beutner and Zähle (2010, 2012,
2014), Beutner et al. (2012), and Buchsteiner (2015).
Example 4.5 (for condition (b)) Let g be a continuous concave distortion function
as introduced before (1). For every real-valued random variable X (on some given atom´∞
less probability space) satisfying 0 g(1 − F|X| (x) dx < ∞ the distortion risk measure
associated with g is defined by ρg (X) := fg (FX ) with fg as in (1). Here FX and F|X|
denote the distribution functions of X and |X|, respectively. The set Xg of all random
variables X satisfying the above integrability condition provides a linear subspace of
L1 ; this follows from Denneberg (1994, Proposition 9.5) and Föllmer and Schied (2011,
Proposition 4.75). It is known that ρg is a law-invariant coherent risk measure; see,
for instance, Wang and Dhaene (1998). If specifically g(s) = (s/α) ∧ 1 for any fixed
13
α ∈ (0, 1), then we have Xg = L1 and ρg is nothing but the Average Value at Risk at
level α.
The risk functional fg : D(fg ) → R corresponding to ρg was already introduced in
(1), where D(fg ) is the set of all distribution functions of the random variables of Xg .
P
Now, the mapping ω 7→ Fbn (ω, t) = n1 ni=1 1[Xi (ω),∞) (t) is (F , B(R))-measurable for every
t ∈ R. Due to the monotonicity of g also the mapping ω 7→ g(Fbn (ω, t)) is (F , B(R))measurable for every t ∈ R. By the right-continuity of the mapping t 7→ g(Fbn (ω, t))
for every fixed ω ∈ Ω we obtain in particular that the mapping (ω, t) 7→ g(Fbn (ω, t)) is
(F ⊗ B(R), B(R))-measurable. Fubini’s theorem then implies that the mapping ω 7→
fg (Fbn (ω, ·)) is (F , B(R))-measurable. So we have in particular that condition (b) of
Corollary 4.2 holds.
✸
Example 4.6 (for conditions (c)–(d)) Let fg : D(fg ) → R be as in Example 4.5.
Let F ∈ D(fg ) with 0 < F (·) < 1, and φ be a weight function satisfying the integrability
condition
ˆ ∞
g(γF (t))
dt < ∞
for some γ ∈ (0, 1).
(24)
−∞ F (t) φ(t)
Assume that the set of points t ∈ R for which g is not differentiable at F (t) has Lebesgue
measure zero. Then Theorem 2.7 in Krätschmer et al. (2015) shows that the functional
fg is quasi-Hadamard differentiable at F tangentially to Cφ,F hDφ i with quasi-Hadamard
derivative f˙g;F : Cφ,F → R given by
ˆ ∞
˙
fg;F (x) :=
g ′ (F (t)) x(t) dt,
x ∈ Cφ,F ,
−∞
where g ′ denotes the right-sided derivative of g and Cφ,F is as in Example 4.3. Recall
that Cφ,F is k · kφ -separable and contained in Dφ ; cf. Corollary B.4 in Krätschmer et al.
(2015). The derivative f˙g;F can be extended to Dφ through
ˆ ∞
f˙g;F (x) :=
g ′(F (t)) x(t) dt,
x ∈ Dφ ,
−∞
and the extension is linear and continuous on Dφ . The linearity is obvious and the
continuity is ensured by part (ii) of Lemma 4.1 in Krätschmer et al. (2015). Thus,
condition (c) of Corollary 4.2 holds. Moreover, using arguments as in Example 4.5, one
can easily show that the extension f˙g;F : Dφ → R is also (Dφ , B(R))-measurable. That
is, condition (d) of Corollary 4.2 holds too.
✸
5. Bootstrap results for empirical processes
In the following two subsections, we will give examples for bootstrap versions (Tbn∗ ) of
(Tbn ) in the sense of Definitions 2.1 and 2.2 in the context of Section 4, i.e. in the case
14
where Tbn is given by an empirical distribution function Fbn of real-valued random variables. As mentioned in the introduction these examples can be combined with the
quasi-Hadamard differentiability of statistical functionals to lead to bootstrap consistency for the corresponding plug-in estimators. Examples include empirical distortion
risk measures as well as U- and V-statistics which will be discussed in Section 5.3.
5.1. I.i.d. observations
We will adopt the notation introduced in Section 4. In particular, (Xi ) will be a sequence
of identically distributed real-valued random variables on (Ω, F , P) with distribution
function F , and Fbn will be given by (19). Let (Wni ) be a triangular array of nonnegative
real-valued random variables on (Ω′ , F ′ , P′) such that (Wn1 , . . . , Wnn ) is an exchangeable
random vector for every n ∈ N, and define the map Fbn∗ : Ω → D by
n
1X
Wni (ω ′) 1[Xi (ω),∞) .
Fbn∗ (ω, ω ′) :=
n i=1
(25)
Note that the sequence (Xi ) and the triangular array (Wni ) regarded as families of random variables on the product space (Ω, F, P) := (Ω×Ω′ , F ⊗F ′ , P⊗P′ ) are independent.
Of course, we will tacitly assume that (Ω′ , F ′ , P′ ) is rich enough to host all of the random
variables described in (a)–(b) in Theorem 5.2.
Lemma 5.1 an (Fbn∗ − Fbn ) takes values only in Dφ and is (F, Dφ )-measurable. That is,
the first part of condition (f ) (respectively (f ’)) of Corollary 4.2 holds true.
Proof First of all note that an (Fbn∗ ((ω, ω ′), t) − Fbn (ω, t)) can be written as
an
n
1 X
n
i=1
Wni (ω ′ )1[Xi (ω),∞) (t) −
1X
1[Xi (ω),∞) (t) =: Ξn ((ω, ω ′), t)
n i=1
n
for all t ∈ R and (ω, ω ′) ∈ Ω. The mapping (ω, ω ′) 7→ Ξn ((ω, ω ′), t) is (F, B(R))measurable for every t ∈ R, and the mapping t 7→ Ξn ((ω, ω ′), t) is right-continuous for
every (ω, ω ′) ∈ Ω. It follows that the mapping (ω, ω ′) 7→ Ξn ((ω, ω ′), ·) form Ω to D is
(F , D)-measurable. Further, Ξn ((ω, ω ′), ·) obviously takes values only in Dφ for every
(ω, ω ′) ∈ Ω. Thus Ξn can indeed be seen as an (F, Dφ )-measurable map from Ω to Dφ
(⊆ D).
✷
The proof of the following Theorem 5.2 strongly relies on Section 3.6.2 in van der Vaart and Wellner
(1996). In fact, the elaborations in Section 3.6.2 in van der Vaart and Wellner (1996)
yield slightly stronger results compared to those of Theorem 5.2, because van der Vaart
and Wellner work in a more general framework. More precisely, they establish outer
almost sure bootstrap results for the empirical process w.r.t. convergence in distribution
15
in the Hoffmann-Jørgensen sense. The first result on Efron’s bootstrap for the empirical
process of i.i.d. random variables was given by Bickel and Freedman (1981, Theorem 4.1)
for the uniform sup-norm, that is, for φ ≡ 1. Gaenssler (1986) extended this result to
Vapnik–Červonenkis classes. For a version of Efron’s bootstrap in a very general set-up,
see also Giné and Zinn (1990, Theorem 2.4).
Theorem 5.2 Assume that the random variables X1 , X2 , . . . are i.i.d., their distribution
´
function F satisfies φ2 dF < ∞, and one of the following two settings is met.
(a) (Efron’s bootstrap) The random vector (Wn1 , . . . , Wnn ) is multinomially distributed according to the parameters n and p1 = · · · = pn = n1 for every n ∈ N.
(b) (Bayesian bootstrap) Wni = Yi /Y n for every n ∈ N and i = 1, . . . , n, where
P
Y n := n1 nj=1 Yj and (Yj ) is any sequence of nonnegative i.i.d. random variables on
´∞
(Ω′ , F ′, P′ ) with distribution µ which satisfies 0 µ[(x, ∞)]1/2 dx < ∞ and whose
standard deviation coincides with its mean and is strictly positive.
Then (condition (a) and) the second part of condition (f ) of Corollary 4.2 hold for
√
an = n, B = BF and Fbn∗ defined in (25), where BF is an F -Brownian bridge, i.e. a
centered Gaussian process with covariance function Γ(t0 , t1 ) = F (t0 ∧ t1 )F (t0 ∨ t1 ).
Proof The claim of Theorem 5.2 would follow from the second assertion of Theorem
3.6.13 in van der Vaart and Wellner (1996) with F = Fφ := {φ(x)1(−∞,x] : x ∈ R} if
we could show that the assumptions of Theorem 3.6.13 in van der Vaart and Wellner
(1996) are fulfilled in each of the settings (a)–(b). At this point we stress the facts
that convergence in distribution in the Hoffmann-Jørgensen sense implies convergence
in distribution◦ for the open-ball σ-algebra and that outer almost sure convergence (as
defined in part (iii) of Definition 1.9.1 in van der Vaart and Wellner (1996)) implies
almost sure convergence (i.e. convergence almost everywhere) in the classical sense. The
latter follows from Proposition 1.1 in Dudley (2010).
In Theorem 3.6.13 in van der Vaart and Wellner (1996) it is assumed that the following three assertions hold:
P
1) Fφ is a Donsker class w.r.t. P, and (t1 , . . . , tn ) 7→ supf ∈Fφ,δ | ni=1 λi f (ti )| is a
measurable mapping on the completion of (Rn , B(Rn ), P⊗n
X1 ) for every δ > 0,
λ1 , . . . , λn ∈ R and n ∈ N. Here we set Fφ,δ := {f1 − f2 : f1 , f2 ∈ Fφ , ρP (f1 − f2 ) <
δ} with ρP (f ) := VarP [f (X1 )]1/2 , where VarP refers to the variance w.r.t. P.
2
2) Eout
P [f (X1 ) ] < ∞ for the envelope function f (t) := supf ∈Fφ (f (t) − EP [f (X1 )]),
where Eout
P refers to the outer expectation w.r.t. P.
3) (Wn1 , . . . , Wn,n ) is an exchangeable nonnegative random vector for every n ∈ N,
and the triangular array (Wni ) satisfies condition (3.6.8) in van der Vaart and Wellner
(1996).
16
We will now verify 1)–3).
´
1): The assumption φ2 dF < ∞ ensures that Fφ is a Donsker class w.r.t. P; cf.
Example 4.3. To verify the second part of assertion 1), let δ > 0 arbitrary but fixed and
f ∈ Fφ,δ with ρP (f ) < δ. Now, f has the representation f = φ(x1 )1(−∞,x1 ] −φ(x2 )1(−∞,x2 ]
for some x1 , x2 ∈ R, and
ρP (f ) = VarP φ(x1 )1(−∞,x1 ] (X1 ) − φ(x2 )1(−∞,x2 ] (X1 )
= φ(x1 )2 F (x1 )(1 − F (x1 )) + φ(x2 )2 F (x2 )(1 − F (x2 ))
−φ(x1 )φ(x2 )F (x1 ∧ x2 )(1 − F (x1 ∧ x2 ))
depends (right) continuously on (x1 , x2 ). So we can find a sequence (gm ) in the countable
subclass Gφ,δ := {gq1,q2 = φ(q1 )1(−∞,q1] − φ(q2 )1(−∞,q2] : q1 , q2 ∈ Q, ρP (gq1 ,q2 ) < δ} of Fφ,δ
such that gm (t) → f (t) for every t ∈ R. For instance, gm := gq1,m ,q2,m for any sequences
(q1,m ) and (q2,m ) in Q such that q1,m ց x1 , q2,m ց x2 and ρP (gq1,m ,q2,m ) < δ. As discussed
in Example 2.3.4 in van der Vaart and Wellner (1996) this implies that the second part
of assertion 1) holds.
2): We first of all note that in the present setting the outer expectation Eout
P can be
replaced by the classical expectation EP w.r.t. P. Indeed, the envelope function f can
be written as
f (t) = sup(1(−∞,x] (t) − F (x))φ(x) = sup(1(−∞,q] (t) − F (q))φ(q)
x∈R
q∈Q
and is thus Borel measurable. So it remains to show E[f (X1 )2 ] < ∞. To this end, we
´
note that the assumption φ2 dF < ∞ implies
M1 := sup F (t)2 φ(t)2 < ∞
M2 := sup(1 − F (t))2 φ(t)2 < ∞.
and
t>0
t≤0
Furthermore, for t ≤ 0 we have
(1(−∞,x] (t) − F (x)) φ(x) =
2
2
(1 − F (x))2 φ(x)2 , t ≤ x
F (x)2 φ(x)2 , t > x
and so, since the mapping x 7→ (1 − F (x))2 φ(x)2 is non-increasing on [t, 0],
f (t)2 = sup(1(−∞,x] (t) − F (x))2 φ(x)2 ≤ max{M1 , (1 − F (t))2 φ(t)2 , M2 } =: g(t).
x∈R
For t > 0 we obtain similarly
f (t)2 = sup(1(−∞,x] (t) − F (x))2 φ(x)2 ≤ max{M1 , F (t)2 φ(t)2 , M2 } =: g(t),
x∈R
because the mapping x 7→ (F (x)φ(x))2 is non-decreasing on (0, t]. Hence, E[f (X1 )2 ] ≤
´
E[g(X1 )2 ] < ∞ due to our assumption φ2 dF < ∞.
3): Examples 3.6.10 and 3.6.12 in van der Vaart and Wellner (1996) show that assertion 3) holds in each of the settings (a)–(b).
✷
17
5.2. Stationary, β-mixing observations
As in Section 5.1, we will adopt the notation introduced in Section 4. In particular, (Xi )
will be a sequence of identically distributed real-valued random variables on (Ω, F , P)
with distribution function F , and Fbn will be given by (19). Let (ℓn ) be a sequence of
integers such that ℓn ր ∞ as n → ∞, and ℓn < n for all n ∈ N. Set kn := ⌊n/ℓn ⌋ for
all n ∈ N. Let (Inj )n∈N, 1≤j≤kn be a triangular array of random variables on (Ω′ , F ′ , P′ )
such that In1 , . . . , Inkn are i.i.d. according to the uniform distribution on {1, . . . , n} for
every n ∈ N. Define the map Fbn∗ : Ω → D by
n
1X
Wni (ω ′)1[Xi (ω),∞)
Fbn∗ (ω, ω ′) :=
n i=1
(26)
with
′
Wni (ω ) :=
kn
X
j=1
1{Inj ≤i≤(Inj +ℓn −1)∧n} (ω ) + 1{Inj +ℓn −1>n, 1≤i≤Inj +ℓn −1−n} (ω ) .
′
′
(27)
Note that, as before, the sequence (Xi ) and the triangular array (Wni ) regarded as
families of random variables on the product space (Ω, F, P) := (Ω × Ω′ , F ⊗ F ′ , P ⊗ P′ )
are independent.
At an informal level this means that given a sample X1 , . . . , Xn , we pick kn blocks
of length ℓn in the (artificially) extended sample X1 , . . . , Xn , Xn+1 , . . . , Xn+ℓn−1 (with
Xn+i := Xi , i = 1, . . . , ℓn − 1) where the start indices In1 , In2 , . . . , Inkn are chosen
independently and uniformly in the set of all indices {1, . . . , n}:
block 1:
block 2:
XIn1 , XIn1 +1 , . . . , XIn1 +ℓn −1
XIn2 , XIn2 +1 , . . . , XIn2 +ℓn −1
..
.
block kn :
XInkn , XInkn +1 , . . . , XInkn +ℓn −1
The bootstrapped empirical distribution function Fbn∗ is then defined to be the distribution function of the discrete finite (not necessarily probability) measure with atoms
X1 , . . . , Xn carrying masses Wn1 , . . . , Wnn respectively, where Wni specifies the number
of blocks which contain Xi .
Lemma 5.3 an (Fbn∗ − Fbn ) takes values only in Dφ and is (F, Dφ )-measurable. That is,
the first part of condition (f ) (respectively (f ’)) of Corollary 4.2 holds true.
Proof The proof of Lemma 5.1 with the obvious modifications also applies to Lemma
5.3.
✷
The bootstrap method induced by the bootstrapped empirical distribution function Fbn∗
defined in (26)–(27) is the so-called circular bootstrap; see, for instance, Politis and Romano
18
(1992) and Radulovic (1996). The circular bootstrap is only a slight modification of
the moving blocks bootstrap that was independently introduced by Künsch (1989) in
the context of the sample mean and by Liu and Singh (1992). Bühlmann (1994, 1995),
Naik-Nimbalkar and Rajarshi (1994), and Radulovic (1996) extended Künsch’s approach
to empirical processes of strictly stationary, mixing observations. Doukhan et al. (2015)
extended Shao’s so-called dependent wild bootstrap for smooth functions of the sample
mean (cf. Shao (2010)) to the empirical process of strictly stationary and β-mixing observations. For an application of the delta-method based on the notion of quasi-Hadamard
differentiability the most interesting results are those that allow for weight functions φ
with lim|x|→∞ φ(x) → ∞. The following result is derived from Theorem 1 in Radulovic
(1996).
Theorem 5.4 (Circular bootstrap) Denote by F the distribution function of X1 and
assume that the following conditions hold:
´
(a) φp dF < ∞ for some p > 2.
(b) The sequence of random variables (Xi ) is strictly stationary and β-mixing with
mixing coefficients (βi ) satisfying βi = O(i−b ) for some b > p/(p − 2).
p−2
(c) The block length ℓn satisfies ℓn = O(nγ ) for some γ ∈ (0, 2(p−1)
).
Then (condition (a) and) the second part of condition (f ’) of Corollary 4.2 hold for an =
√
eF and Fb∗ defined in (26), where B
eF is a centered Gaussian process with con, B = B
n
P P
variance function Γ(t0 , t1 ) = F (t0 ∧t1 )(1−F (t0 ∨t1 ))+ 1i=0 ∞
k=2 Cov(1{X1 ≤ti } , 1{Xk ≤t1−i } ).
A similar result that allows to verify condition (f) of Corollary 4.2 (where in (21) the
empirical distribution function Fbn is replaced by the conditional expectation of Fbn∗ ) can
be found in Bühlmann (1995, Theorem 1).
Proof of Theorem 5.4 It was shown in Arcones and Yu (1994, Theorem 2.1) that
under conditions (a)–(b) of Theorem 5.4 the condition (a) of Corollary 4.2 is satisfied;
see also Example 4.4. In the following we will show that under assumption (a) of
Theorem 5.4 the following two assumptions of Theorem 1 in Radulovic (1996) are met
for the class of functions Fφ := {fx : x ∈ R} with fx (·) := φ(x)1(−∞,x] (·) for x ≤ 0 and
fx (·) := −φ(x)1(x,∞) (·) for x > 0:
1) Fφ is a VC-subgraph class.
´ p
2) f dF < ∞ for the envelope function f (t) := supx∈R |fx (t)|.
The other assumptions of Theorem 1 in Radulovic (1996) are just our assumptions
√
(b) and (c). Then, since we may identify the maps x 7→ n(Fbn (x) − F (x))φ(x) and
´
√ ´
√
x 7→ n(Fbn∗ (x) − Fbn (x))φ(x) with respectively fx 7→ n( fx dFbn − fx dF ) and fx 7→
19
´
√ ´
n( fx dFbn∗ − fx dFbn ), Theorem 1 in Radulovic (1996) implies that condition (f’) of
Corollary 4.2 is satisfied too.
Before verifying 1), let us recall the definition of VC-subgraph class; cf., for instance,
van der Vaart and Wellner (1996, Section 2.6). First recall that the VC-index of a collection C of subsets of a nonempty set Y is defined by V (C) := inf{n : mC (n) < 2n }
with the convention inf ∅ := ∞, where
mC (n) :=
max
y1 ,...,yn ∈Y
#{C ∩ {y1 , . . . , yn } : C ∈ C}.
(28)
A collection C is said to be a VC-class if V (C) < ∞. A class F of functions f : R → R is
said to be a VC-subgraph class if the collection CF := {{(x, t) ∈ R2 : t < f (x)} : f ∈ F}
is a VC-class of sets in Y := R2 .
1): We will show that Fφ is a VC-subgraph class with V (CFφ ) ≤ 3. For V (CFφ ) ≤ 3 it
suffices to show that mCFφ (3) < 23 . Note that that mCFφ (3) < 23 means that for every
choice of y1 , y2 , y3 ∈ R2 there exists at least one of the 23 subsets of {y1, y2 , y3 } which
cannot be represented as C ∩ {y1 , y2, y3 } for any C ∈ CFφ . By way of contradiction
assume that there exist y1 = (x1 , t1 ), y2 = (x2 , t2 ), y3 = (x3 , t3 ) in R2 such that every
subset of {y1 , y2 , y3 } has the representation C ∩ {y1 , y2 , y3} for some C ∈ CFφ . Then, in
particular, there exist C12 , C13 , C23 ∈ CFφ such that
C12 ∩ {(x1 , t1 ), (x2 , t2 ), (x3 , t3 )} = {(x1 , t1 ), (x2 , t2 )},
C13 ∩ {(x1 , t1 ), (x2 , t2 ), (x3 , t3 )} = {(x1 , t1 ), (x3 , t3 )},
C23 ∩ {(x1 , t1 ), (x2 , t2 ), (x3 , t3 )} = {(x2 , t2 ), (x3 , t3 )}.
(29)
We may and do assume without loss of generality that x1 ≤ x2 ≤ x3 . Then, if (29) held
true, there would exist x12 , x13 , x23 ∈ R such that
t1 < fx12 (x1 ),
t2 < fx12 (x2 ),
t1 < fx13 (x1 ),
t2 ≥ fx13 (x2 ),
t1 ≥ fx23 (x1 ),
t2 < fx23 (x2 ),
t3 ≥ fx12 (x3 ),
t3 < fx13 (x3 ),
t3 < fx23 (x3 ).
(30)
First assume x12 ≤ 0. In this case we have fx12 (·) = 1(−∞,x12 ] (·)φ(x12 ) and thus
t3 ≥ 0 (due to t3 ≥ fx12 (x3 )). But then fx13 and fx23 are also of the form fx13 (·) =
1(−∞,x13 ](·)φ(x13 ) and fx23 (·) = 1(−∞,x23 ] (·)φ(x23 ), because t3 < fx13 (x3 ), t3 < fx23 (x3 ),
and functions of the form fx (·) = −1(x,∞) φ(x) take values only in (−∞, −1] ∪ {0}. From
the second and the third line of (30) we can now conclude that fx13 (x1 ) = fx13 (x3 ),
x3 ≤ x13 , and fx23 (x2 ) = fx23 (x3 ), x3 ≤ x23 , respectively. It follows that
fx13 (x1 ) = fx13 (x2 )
and
fx23 (x1 ) = fx23 (x2 ),
(31)
because x2 ≤ x3 (which implies x2 ∈ (−∞, x13 ]) and x1 ≤ x2 (which implies x1 ∈
(−∞, x23 ]). On the other hand, by (30) we obviously have
fx13 (x1 ) > fx23 (x1 )
and
20
fx23 (x2 ) > fx13 (x2 ).
(32)
But (31) and (32) contradict each other.
Now assume x12 > 0. This implies that fx12 takes values only in (−∞, −1] ∪ {0},
and therefore fx12 (x1 ) ≤ 0 and fx12 (x2 ) ≤ 0. It follows that t1 < 0 and t2 < 0. The
latter two inequalities imply fx23 (x1 ) < 0 and fx13 (x2 ) < 0, respectively. It follows
that x23 > 0 and x13 > 0, because otherwise fx23 or fx13 would take values only in
{0}∪[1, ∞). In particular, t3 < 0 (since t3 < fx23 (x3 )). That is, we have t1 , t2 , t3 < 0 and
fx12 (·) = −1(x12 ,∞) (·)φ(x12 ), fx13 (·) = −1(x13 ,∞) (·)φ(x13 ), fx23 (·) = −1(x23 ,∞) (·)φ(x23 ).
From the third line of (30) we first conclude that x1 > x23 , because t1 < 0 (so that
t1 ≥ fx23 (x1 ) is only possible if x1 > x23 ). Then we also have x2 > x23 and x3 > x23 ,
because x3 ≥ x2 ≥ x1 . This implies fx23 (x1 ) = fx23 (x2 ) = fx23 (x3 ), and we conclude
from the third line of (30) that t1 > t2 . Similarly, from the second line of (30) we obtain
t2 > t3 . Summarizing we must have
0 > t1 > t2 > t3 .
(33)
Recall that we assumed (by way of contradiction) that y1 = (x1 , t1 ), y2 = (x2 , t2 ), y3 =
(x3 , t3 ) are such that every subset of {y1 , y2, y3 } has the representation C ∩ {y1, y2 , y3 }
for some C ∈ CFφ . In particular, there exists a set C2|1,3 ∈ CFφ with
C2|1,3 ∩ {(x1 , t1 ), (x2 , t2 ), (x3 , t3 )} = {(x2 , t2 )}.
That is, there exists some x2|1,3 ∈ R such that
t1 ≥ fx2|1,3 (x1 ),
t2 < fx2|1,3 (x2 ),
t3 ≥ fx2|1,3 (x3 ).
(34)
Since t1 < 0, we must have x2|1,3 > 0 (i.e. fx2|1,3 (·) = −1(x2|1,3 ,∞) (·)φ(x2|1,3 )) and x1 >
x2|1,3 . The latter inequality implies in particular x2 > x2|1,3 and x3 > x2|1,3 , because
x3 ≥ x2 ≥ x1 . Hence fx2|1,3 (x1 ) = fx2|1,3 (x2 ) = fx2|1,3 (x3 ). In view of (34), this gives
t2 < t3 . But this contradicts (33).
2): The envelope function f is given by f (t) = φ(t) for t ≤ 0 and by f (t) = φ(t−) =
φ(t) (recall that φ is continuous) for t > 0. Then under assumption (a) the integrability
condition 2) holds.
✷
5.3. Some applications
In this section we discuss two specific examples. First we rigorously treat the case of
empirical distortion risk measures. Thereafter we informally discuss bootstrap results
for U- and V-statistics.
1) Let fg : D(fg ) → R be the distortion risk functional associated with a continuous
concave distortion function as in (1) and Example 4.5, and let φ : R → [1, ∞) be any
continuous function. Let F ∈ D(fg ) satisfy the integrability condition (24). Let (Xi ) be
21
a strictly stationary sequence of real-valued random variables on some probability space
(Ω, F , P) with distribution function F . Let Fbn be the empirical distribution function of
´
X1 , . . . , Xn defined by (19). If X1 , X2 , . . . are independent, φ2 dF < ∞, and Fbn∗ is as
in Theorem 5.2 (on some extension (Ω, F, P) = (Ω × Ω′ , F ⊗ F ′ , P ⊗ P′ ) of the original
probability space), then Corollary 4.2, Example 4.3, Examples 4.5–4.6, and Theorem
5.2 show that (fg (Fbn∗ )) is a bootstrap version in probability of (fg (Fbn )). This bootstrap
consistency can also be obtained by results on L-statistics by Helmers et al. (1990) and
Gribkova (2002). However, the latter results rely on the independence of X1 , X2 , . . ..
To the best of our knowledge so far there do not exit general results on bootstrap
consistency for empirical distortion risk measures associated with continuous concave
distortion functions when the data X1 , X2 , . . . are dependent. On the other hand, our
theory admits such results. Indeed, if the sequence (Xi ) is β-mixing with mixing rate as
´
in condition (b) of Theorem 5.4, φp dF < ∞ for some p > 2, and Fbn∗ is as in Theorem
5.4, then Corollary 4.2, Example 4.4, Examples 4.5–4.6, and Theorem 5.4 show that
(fg (Fbn∗ )) is a bootstrap version in probability of (fg (Fbn )). We emphasize that the results
by Lahiri (2003, Chapter 4.4) for α-mixing data do not cover this bootstrap consistency,
because Lahiri assumes Fréchet differentiable for fg which fails for continuous concave
distortion functions g.
2) Let fh : D(fh ) → R be the V-functional corresponding to a given Borel measurable
function h : R2 → R (sometimes referred to as kernel) which is given by
¨
fh (F ) :=
h(x1 , x2 ) dF (x1 )dF (x2 ),
(35)
where D(fh ) denotes the set of all distribution functions on the real line for which
the double integral in (35) exists. It was shown in Theorem 4.1 in Beutner and Zähle
(2012) that subject to some regularity conditions on h and F the V-functional fh is
quasi-Hadamard differentiable at F w.r.t. a suitable nonuniform sup-norm. Similar as
in Example 4.6 it can be shown that condition (d) of Corollary 4.2 holds for the quasiHadamard derivative of fh . Then again, if (Xi ) is a stationary β-mixing sequence of
random variables with distribution function F and mixing rate as in condition (b) of
´
Theorem 5.4, φp dF < ∞ for some p > 2, and Fbn∗ is as in Theorem 5.4, Corollary 4.2
shows that (fh (Fbn∗ )) is a bootstrap version in probability of (fh (Fbn )). Other approaches
to show bootstrap consistency for non-degenerate U- and V-statistics can be found,
for example, in Arcones and Giné (1992), Janssen (1994), and Dehling and Wendler
(2010) (yet another approach was exemplified for the variance by Dudley (1990)); see
also Bücher and Kojadinovic (2015) who use results of Dehling and Wendler (2010).
Among other things Dehling and Wendler (2010, Theorem 2.1) also establish bootstrap
consistency for non-degenerate U- and V-statistics for β-mixing sequences. Whereas
their approach requires an additional integrability condition on (X1 , Xk ), our approach
(based on Corollary 4.2 that we just outlined) requires stronger regularity conditions on
22
the kernel h. Looking at condition (b) in Theorem 5.4 and the condition on the mixing
coefficient in Dehling and Wendler (2010, Theorem 2.1), it seems that both approaches
impose the same condition on the mixing coefficient. Thus, the approach based on
Corollary 4.2 may supplement the results in Dehling and Wendler (2010).
6. Proof of Theorem 3.1
We start with a convention and a general remark. We will equip the product space E :=
E × E with the metric d((x1 , x2 ), (y1, y2 )) := max{kx1 − y1 kE ; kx2 − y2 kE }, and denote
◦
◦
the corresponding open-ball σ-algebra on E by B . Note that B ⊆ B◦ ⊗ B◦ , because any
d-open ball in E is the product of two k · kE -open balls in E. Analogously the product
e := E
e ×E
e will be equipped with the metric d((e
e x1 , x
space E
e2 ), (e
y1 , ye2 )) := max{ke
x1 −
e
ye1 k e ; ke
x2 − ye2 k e }. By the separability of (E, k · k e ) the corresponding Borel σ-algebra
E
E
E
e cf. Dudley (2002, Proposition 4.1.7).
Be coincides with the product σ-algebra Be ⊗ B;
e B)-valued
e B)e
e
So the couple (ξ1 , ξ2) is an (E,
random variable when ξ1 and ξ2 are (E,
e B)-valued
e
valued random variables. In particular, h(ξ1 , ξ2 ) is an (E,
random variable
e →E
e is continuous. Since the addition and the multiplication by constants
when h : E
in normed vector spaces are continuous, we have in particular that a linear combination
e B)-valued
e
e B)-valued
e
of two (E,
random variables is again an (E,
random variable. This
fact will be used frequently in the sequel without further mentioning.
e
(i): By assumption (b) we have that f (Tbn ) is (F , B)-measurable.
This implies that
e k·k e )
b
e
an (f (Tn )−f (θ)) is (F , B)-measurable for every n ∈ N, because we assumed that (E,
E
is separable. Now, assertion (i) directly follows from the functional delta-method in the
form of Theorem C.4.
(ii): Recall that Tbn will frequently be seen as a map defined on the extension Ω of
e
Ω. From the above we therefore have that f (Tbn ) is (F, B)-measurable.
Moreover, f (Tbn∗ )
e
due assumption (e). In particular, the map an (f (Tbn∗ ) − f (Tbn )) is
is (F, B)-measurable
e
e k · k e ) is separable. By assumptions
(F , B)-measurable,
because we assumed that (E,
E
e
(a) and (d) we also have that the map f˙θ (ξ) is (F0 , B)-measurable,
and by assumptions
∗
˙
b
b
e
(d) and (f) we have that the map fθ (an (Tn − Tn )) is (F, B)-measurable.
To verify (17), we will adapt the arguments of Section 3.9.3 in van der Vaart and Wellner
en defined by
(1996). First note that Q
en (ω, A)
e := P′ ◦ f˙θ an (Tbn∗ (ω, ·) − Tbn (ω)) −1 [A],
e
e ∈ Be
Q
ω ∈ Ω, A
provides a conditional distribution of f˙θ (an (Tbn∗ − Tbn )) given Π. This follows from Lemma
D.2 (with X(ω, ω ′) = g(ω, ω ′) = f˙θ (an (Tbn∗ (ω, ω ′) − Tbn (ω))) and Y = Π). Now, let δ > 0
be arbitrary but fixed. For (17) it suffices to show that
hn
δ oi
e
e
e
= 0
(36)
lim P ω ∈ Ω : dBL Pn (ω, ·), Qn (ω, ·) ≥
n→∞
2
23
and
oi
en (ω, ·), law{f˙θ (ξ)} ≥ δ
= 0.
(37)
ω ∈ Ω : deBL Q
n→∞
2
e n(ω, ·)) and ω 7→ deBL (Q
en (ω, ·), law{f˙θ (ξ)}) are
Note that the maps ω 7→ deBL (Pen (ω, ·), Q
e k · k e ) was assumed to be separable. For the latter
(F , B(R+ ))-measurable, because (E,
lim P
hn
E
map one can argue as subsequent to Definition 2.2. For the former map one can argue
f1 , deBL ) is separable (cf. Remark A.2 and Theorem A.4)
in the same way, noting that (M
and that the metric distance of two random variables in a separable metric space is also
measurable (cf. Klenke (2014, Lemma 6.1)). In particular, the events in (36) and (37)
are F -measurable.
We first show (37). By (14) in assumption (f), the Continuous Mapping theorem in the
form of Billingsley (1999, Theorem 6.4) (along with P0 ◦ ξ −1 [E0 ] = 1 and the continuity
of f˙θ ), and the implication (a)⇒(g) in the Portmanteau theorem A.3, we have
en (ω, ·), law{f˙θ (ξ)} = 0
lim deBL Q
P-a.e. ω.
n→∞
Since almost sure convergence of real-valued random variables implies convergence in
probability, we arrive at (37).
To verify (36), we set
ηn (ω, ω ′) := an f (Tbn∗ (ω, ω ′)) − f (Tbn (ω)) − f˙θ an (Tbn∗ (ω, ω ′) − Tbn (ω))
and
∗
′
∗
′
e
e
˙
b
b
b
b
ηn,eh (ω, ω ) := h an f (Tn (ω, ω )) − f (Tn (ω)) − h fθ an (Tn (ω, ω ) − Tn (ω))
′
f 1 with BL
f 1 as defined before (44). We then obtain
for every e
h ∈ BL
P
oi
e n(ω, ·) ≥ δ
ω ∈ Ω : deBL Pen (ω, ·), Q
2
ˆ
ˆ
hn
δ oi
e
en (ω, de
= P ω ∈ Ω : sup
h(e
x) Pen (ω, de
x) − e
h(e
x) Q
x) ≥
2
e
f1
h∈BL
ˆ
hn
e
= P ω ∈ Ω : sup
h an f (Tbn∗ (ω, ω ′)) − f (Tbn (ω)) P′ [dω ′ ]
hn
e
f1
h∈BL
δ oi
e
h f˙θ an (Tbn∗ (ω, ω ′) − Tbn (ω)) P′ [dω ′ ] ≥
2
hn
oi
δ
= P ω ∈ Ω : sup
ηn,eh (ω, ω ′) P′[dω ′ ] ≥
2
e
f1
h∈BL
ˆ
hn
δ oi
= Pout ω ∈ Ω : sup
ηn,eh (ω, ω ′) P′ [dω ′] ≥
2
e
f1
h∈BL
ˆ
hn
δ oi
ηn,eh (ω, ω ′) P′ [dω ′ ] ≥
≤ Pout ω ∈ Ω : sup
2
e
f1
h∈BL
−
ˆ
ˆ
24
= P
out
≤ Pout
≤ Pout
≤ P
out
hn
hn
hn
hn
ω ∈ Ω : sup
e
f1
h∈BL
ˆ
δ oi
+
ηn,eh (ω, ω ) 1{|ηn,eh |≥δ/4} (ω, ω ) P [dω ] ≥
2
ˆ
δ
δ oi
∈ Ω : + sup
ηn,eh (ω, ω ′) 1{|ηn,eh |≥δ/4} (ω, ω ′) P′ [dω ′] ≥
4 eh∈BL
2
f
ˆ 1
δ oi
∈ Ω : sup
2 1{|ηn,eh |≥δ/4} (ω, ω ′) P′[dω ′ ] ≥
4
e
f
h∈BL
ˆ 1
oi
δ
∈ Ω : 2 1{kηn kEe ≥δ/4} (ω, ω ′) P[dω ′] ≥
,
(38)
4
ˆ
ω
ω
ω
ηn,eh (ω, ω ′) 1{|ηn,eh |<δ/4} (ω, ω ′) P′ [dω ′]
′
′
′
′
where the second last and the last step are ensured by ke
hk∞ ≤ 1 and the Lipschitz
continuity of h (with Lipschitz constant 1), respectively. We have seen above that the
e k · ke )
e
maps an (f (Tbn ) − f (Tbn∗ )) and f˙θ (an (Tbn∗ − Tbn )) are (F, B)-measurable.
Since (E,
E
e
is separable, we can conclude that the map ηn is (F, B)-measurable.
Since the map
e → R+ is continuous and thus (B,
e B(R+ ))-measurable, we obtain that the map
k · kEe : E
kηn kEe : Ω → R+ is (F, B(R+ ))-measurable. By Fubini’s theorem we can conclude that
the map
ˆ
ω 7−→
1{kηn kEe ≥δ/4} (ω, ω ′) P′ [dω ′]
is (F , B(R+ ))-measurable. Therefore, we may replace the outer probability Pout by the
ordinary probability P in the last line of (38). So we obtain
hn
oi
e n(ω, ·) ≥ δ
P ω ∈ Ω : deBL Pen (ω, ·), Q
2
ˆ
hn
δ oi
≤ P ω ∈ Ω : 2 1{kηn kEe ≥δ/4} (ω, ω ′) P′[dω ′ ] ≥
4
oi
hn
hn
δ oi
δ
′
′
′
′
≥
= P ω ∈ Ω : P ω ∈ Ω : kηn (ω, ω )kEe ≥
4
8
ˆ hn
oi
8
δ
≤
P′ ω ′ ∈ Ω′ : kηn (ω, ω ′)kEe ≥
P[dω]
δ
4
δ oi
8 hn
P (ω, ω ′) ∈ Ω : kηn (ω, ω ′)kEe ≥
=
δ
4
δi
8 h
∗
∗
˙
b
b
b
b
P an f (Tn ) − f (Tn ) − fθ an (Tn − Tn ) Ee ≥ ,
=
δ
4
where for the third and the fourth step we used respectively Markov’s inequality and
the representation of the product measure P = P ⊗ P′ as given in Bauer (2001, Formula
(23.3)). Thus, it remains to show that
e k · k e ) w.r.t. P,
an f (Tbn∗) − f (Tbn ) − f˙θ an (Tbn∗ − Tbn ) →p 0Ee
(39)
in (E,
E
e To prove
where →p refers to convergence in probability and 0Ee denotes the null in E.
(39), we note that by assumption (b) we have that an (Tbn − θ) converges in distribution◦
25
to some separable random variable, ξ. So we may apply part (ii) of Theorem C.4 to
obtain
an f (Tbn ) − f (θ) − f˙θ an (Tbn − θ) →p 0Ee
e k · k e ) w.r.t. P,
in (E,
E
(40)
an f (Tbn∗ ) − f (θ) − f˙θ an (Tbn∗ − θ) →p 0Ee
e k · k e ) w.r.t. P.
in (E,
E
(41)
e k · k e ) was assumed to be separable
where condition (g) of Theorem C.4 holds since (E,
E
(cf. the discussion at the beginning of the proof). Further, in the following we will show
that an (Tbn∗ − θ) converges in distribution◦ to some separable random variable too. So
we may apply part (ii) of Theorem C.4 once more to obtain
Now, (40), (41) and the linearity of f˙θ imply (39).
It remains to show that an (Tbn∗ − θ) converges in distribution◦ to some separable random variable. For this it suffices to show that (an (Tbn − θ), an (Tbn∗ − Tbn )) converges in
◦
◦
distribution◦ to (ξ1 , ξ2 ) in (E, B , d), where (ξ1 , ξ2 ) is an (E, B )-valued random variable
(on some probability space) which takes values only in E0 := E0 × E0 . In fact, the
extended Continuous Mapping theorem C.1 applied to the functions hn : E → E and
h0 : E0 → E0 ⊆ E given by respectively hn (x, y) := x + y and h0 (x, y) := x + y then
implies that an (Tbn∗ − θ) = an (Tbn∗ − Tbn ) + an (Tbn − θ) converges in distribution◦ to the
separable random variable ξ1 + ξ2 . For the application of the extended Continuous Mapping theorem note that hn (an (Tbn∗ − Tbn ), an (Tbn − θ)) = an (Tbn∗ − θ) is (F, B◦ )-measurable
by the first part of condition (g) and that the map h0 : E0 → E is continuous and
◦
(B0 , B◦ )-measurable for B0 := B ∩ E0 = B ∩ E0 . For the latter measurability take into
account that E0 is separable w.r.t. d and argue as at the beginning of the proof. Also
◦
note that (an (Tbn − θ), an (Tbn∗ − Tbn )) can be seen as an (E, B )-valued random variable,
◦
because it is obviously (F , B◦ ⊗ B◦ )-measurable and B ⊆ B◦ ⊗ B◦ .
To show that (an (Tbn − θ), an (Tbn∗ − Tbn )) converges in distribution◦ to some separable
random element (ξ1 , ξ2), we will adapt some of the arguments of the proof of Theorem
2.2 in Kosorok (2008) where weak convergence is understood in the Hoffmann-Jørgensen
sense. Let (Ω1 ×Ω2 , F1 ⊗F2 , P1 ⊗P2 ) := (E, B◦ ⊗B◦ , (P0 ◦ξ −1 )⊗(P0 ◦ξ −1 )) (with ξ and P0 as
in condition (b)) and ξi be the i-th coordinate projection on Ω1 × Ω2 = E, i = 1, 2. Then
◦
(ξ1 , ξ2 ) can be seen as an (E, B )-valued random variable on (Ω1 × Ω2 , F1 ⊗ F2 , P1 ⊗ P2 ),
◦
◦
because by B ⊆ B◦ ⊗B◦ it is clearly (B◦ ⊗B◦ , B )-measurable. In view of the implication
(f)⇒(a) in the Portmanteau theorem A.3, for the convergence in distribution◦ of the pair
(an (Tbn − θ), an (Tbn∗ − Tbn )) to the random variable (ξ1 , ξ2 ) it suffices to show that
ˆ
ˆ
∗
′
b
b
b
h an (Tn − θ), an (Tn − Tn ) d(P ⊗ P ) −→
h(ξ1 , ξ2 ) d(P1 ⊗ P2 )
◦
◦
for every h ∈ BL1 , where BL1 denotes the set of all real-valued functions on E = E × E
◦
that are (B , B(R))-measurable, bounded by 1 and Lipschitz continuous with Lipschitz
26
◦
constant 1 (as defined before (44)). So, let h ∈ BL1 . We have
ˆ
ˆ
∗
′
b
b
b
h(an (Tn − θ), an (Tn − Tn )) d(P ⊗ P ) − h(ξ1 , ξ2 ) d(P1 ⊗ P2 )
ˆ
ˆ
∗
′
≤
h(an (Tbn − θ), an (Tbn − Tbn )) d(P ⊗ P ) − h(an (Tbn − θ), ξ2 ) d(P ⊗ P2 )
ˆ
ˆ
b
h(an (Tn − θ), ξ2 ) d(P ⊗ P2 ) − h(ξ1 , ξ2 ) d(P1 ⊗ P2 )
+
=: S1 (n) + S2 (n).
For every x2 ∈ E, define the function hx2 : E → R by hx2 (x1 ) := h(x1 , x2 ) and note that
hx2 is bounded, continuous, and (B◦ , B(R))-measurable. The latter measurability means
−1
−1
◦
that h−1
x2 (B) = (h (B))x2 := {x1 ∈ E : (x1 , x2 ) ∈ h (B)} lies in B for every B ∈ B(R).
−1
◦
◦
By the (B , B(R))-measurability of h the set h (B) lies in B . By Lemma 23.1 of Bauer
(2001) the set Ax2 := {x1 ∈ E : (x1 , x2 ) ∈ A} lies in B◦ for every A ∈ B◦ ⊗ B◦ . Thus, in
◦
view of B ⊆ B◦ ⊗ B◦ , it follows that the functions hx2 is indeed (B◦ , B(R))-measurable.
Now, with the help of Fubini’s theorem we obtain
S2 (n)
≤
ˆ
=
ˆ
ˆ
b
h(an (Tn (ω) − θ), ξ2 (ω2 )) P[dω] − h(ξ1 (ω1 ), ξ2 (ω2 )) P1 (dω1 ) P2 [dω2 ]
ˆ
ˆ
b
hξ2 (ω2 ) (an (Tn (ω) − θ)) P[dω] − hξ2 (ω2 ) (ξ1 (ω1 )) P1 (dω1 ) P2 [dω2 ].
ˆ
In view of assumption (a), the integrand of the outer integral converges to 0 for every
ω2 . So, since khx2 (·)k∞ ≤ 1 for every x2 ∈ E, the Dominated Convergence theorem
implies that the summand S2 (n) converges to 0. For every x1 ∈ E, define the function
hx1 : E → R by hx1 (x2 ) := h(x1 , x2 ) and note that hx1 ∈ BL◦1 for every x1 ∈ E (for the
measurability of hx1 one can argue as for hx2 above). With the help of Fubini’s theorem
we obtain
ˆ ˆ
h(an (Tbn (ω) − θ), an (Tbn∗ (ω, ω ′) − Tbn (ω))) P′[dω ′]
S1 (n) ≤
ˆ
− h(an (Tbn (ω) − θ), ξ2 (ω2 )) P2 [dω2 ] P[dω]
ˆ ˆ
=
han (Tbn (ω)−θ) (an (Tbn∗ (ω, ω ′) − Tbn (ω))) P′[dω ′ ]
ˆ
− han (Tbn (ω)−θ) (ξ2 (ω2 )) P2 [dω2 ] P[dω]
ˆ
ˆ
≤
sup
ham (Tbm (ω)−θ) (an (Tbn∗ (ω, ω ′) − Tbn (ω))) P′[dω ′]
m∈N
ˆ
− ham (Tbm (ω)−θ) (ξ2 (ω2 )) P2[dω2 ] P[dω].
(42)
27
The integrand of the outer integral is bounded above by d◦BL (Pn (ω, ·), law{ξ2 }). So
it follows by the second part of assumption (f) and the implication (a)⇒(g) in the
Portmanteau theorem A.3 that the integrand of the outer integral converges to 0 for Pa.e. ω. In view of kham (Tbm (ω)−θ) (·)k∞ ≤ 1 for every m ∈ N, the Dominated Convergence
theorem implies that the summand S1 (n) converges to 0 too. This completes the proof
of part (ii).
(iii): One can proceed as for the proof of part (ii). It again suffices to show (36)
and (37). The proof of (36) can be transferred nearly verbatim. The convergence of
the upper bound in (42) to zero was justified by the classical Dominated Convergence
theorem. This time one has to use slightly different arguments. The upper bound in
(42) is bounded above by
ˆ out
ˆ
ˆ
∗
′
′
′
b
b
sup
h(an (Tn (ω, ω ) − Tn (ω))) P [dω ] − h(ξ2 (ω2 )) P2[dω2 ] P[dω],
h∈BL◦1
which equals
out
ˆ
d◦BL (Pn (ω, ·), law{ξ}) P[dω].
´ out
refers to the outer integral (outer expectation). By (15) in assumption (f’),
Here
the integrand of the latter integral converges to 0 in outer probability. Lemma 3.3.4 in
Dudley (1999) then implies
ˆ out
lim sup
d◦BL (Pn (ω, ·), law{ξ}) P[dω] ≤ 0.
n→∞
It follows that the summand S1 (n) again converges to 0. This gives (36).
It remains to show that (37) can also be derived from assumption (f’). We have
P
oi
en (ω, ·), law{f˙θ (ξ)} ≥ δ
ω ∈ Ω : deBL Q
2 ˆ
ˆ
hn
δ oi
e
e
˙
e
= P ω ∈ Ω : sup
h(e
x) Qn (ω, de
x) − h(e
x) law{fθ (ξ)}[de
x] ≥
2
e
f1
h∈BL
ˆ
hn
e
= P ω ∈ Ω : sup
h f˙θ an (Tbn∗ (ω, ω ′) − Tbn (ω)) P′ [dω ′]
hn
e
f1
h∈BL
hn
= P ω ∈ Ω : sup
e
f1
h∈BL
= P
hn
ω ∈ Ω : sup
e
f1
h∈BL
−
ˆ
−
ˆ
δ oi
e
h f˙θ (ξ(ω0 )) P0 [dω0 ] ≥
2
e
h ◦ f˙θ an (Tbn∗ (ω, ω ′) − Tbn (ω)) P′ [dω ′ ]
ˆ
ˆ
δ oi
e
h ◦ f˙θ ξ(ω0 ) P0 [dω0 ] ≥
2
e
h ◦ f˙θ (x) Pn (ω, dx)
28
≤ P
out
= Pout
hn
hn
−
ω ∈ Ω : sup
h∈BL◦1
δ oi
e
h ◦ f˙θ (x) law{ξ}[dx] ≥
2
ˆ
ˆ
h(x) Pn (ω, dx) −
h(x) law{ξ}[dx] ≥
ˆ
ω ∈ Ω : d◦BL Pn (ω, ·), law{ξ} ≥
δ
2(Lf,θ ∨ 1)
oi
,
δ
2(Lf,θ ∨ 1)
oi
(43)
where Lf,θ > 0 denotes the Lipschitz constant of the linear and continuous (thus Lipschitz continuous) map f˙θ . The last line in (43) converges to 0 as n → ∞ by (15) in
assumption (f’). This gives (37), and the proof is complete. ✷
A. Weak topology and weak convergence for the
open-ball σ-algebra
Let (E, d) be a metric space and B◦ be the σ-algebra on E generated by the open balls
Br (x) := {y ∈ E : d(x, y) < r}, x ∈ E, r > 0. We will refer to B◦ as open-ball σ-algebra.
If (E, d) is separable, then B◦ coincides with the Borel σ-algebra B. If (E, d) is not
separable, then B◦ might be strictly smaller than B and thus a continuous real-valued
function on E is not necessarily (B◦ , B(R))-measurable. Let Cb◦ be the set of all bounded,
continuous and (B◦ , B(R))-measurable real-valued functions on E, and M◦1 be the set of
all probability measures on (E, B◦ ). For every f ∈ Cb◦ we consider the mapping
ˆ
◦
πf : M1 −→ R,
µ 7−→ f dµ.
The weak◦ topology Ow◦ on M◦1 is defined to be the topology O(F) generated by the class
T
of functions F := {πf : f ∈ Cb◦ }. That is, Ow◦ := O(SF ) := O topology on M◦ with O⊇SF O
1
for the system SF := {πf−1 (G′ ) : f ∈ Cb◦ , G′ ∈ OR }, where OR is the usual topology of
open sets in R. In other words, the weak◦ topology is the coarsest topology on M◦1 w.r.t.
which each of the maps πf , f ∈ Cb◦ , is continuous. A sequence (µn ) in M◦1 converges to
some µ0 ∈ M◦1 in the weak◦ topology Ow◦ if and only if
ˆ
ˆ
f dµn −→
f dµ0
for all f ∈ Cb◦ ;
see, for instance, Lemma 2.52 in Aliprantis and Border (2006) (take into account that
every sequence is a net). In this case, we also say that (µn ) converges weak◦ ly to µ0
and write µn ⇒◦ µ0 . It is worth mentioning that two probability measures µ0 , ν0 ∈ M◦1
´
´
coincide if µ0 [E0 ] = ν0 [E0 ] = 1 for some separable E0 ∈ B◦ and f dµ0 = f dν0 for all
uniformly continuous f ∈ Cb◦ ; see, for instance, Billingsley (1999, Theorem 6.2).
29
Remark A.1 Recall that B◦ = B when (E, d) is separable. In this case we suppress the
superscript ◦ and write simply Cb , M1 , weak, Ow , and ⇒ instead of Cb◦ , M◦1 , weak◦ ,
Ow◦ , and ⇒◦ , respectively.
✸
Denote by BL◦1 the set of all (B◦ , B(R))-measurable functions f : E → R satisfying
|f (x)−f (y)| ≤ d(x, y) for all x, y ∈ E and supx∈E |f (x)| ≤ 1. Note that BL◦1 is contained
in the set of all uniformly continuous functions in Cb◦ . The bounded Lipschitz distance
on M◦1 is defined by
ˆ
ˆ
◦
(44)
f dµ − f dν .
dBL (µ, ν) := sup
f ∈BL◦1
It is easily seen that the mapping d◦BL : M◦1 × M◦1 → R+ satisfies the axioms of a
pseudo-metric on M◦1 , i.e. that it is symmetric and satisfies d◦BL (µ, µ) = 0 as well as the
triangle inequality.
Remark A.2 If (E, d) is separable, then we again suppress the superscript ◦ and write
simply BL1 and dBL instead of BL◦1 and d◦BL , respectively. In this case the bounded
Lipschitz distance dBL provides even a metric on M1 , because BL1 is separating in M1 ;
the latter follows from the proof of Theorem 1.2 in Billingsley (1999).
✸
Theorem A.3 (Portmanteau theorem) Let µn ∈ M◦1 , n ∈ N0 , and assume that
µ0 [E0 ] = 1 for some separable E0 ∈ B◦ . Then the following conditions are equivalent:
(a) µn ⇒◦ µ0 .
´
´
(b) f dµn → f dµ0 for all uniformly continuous f ∈ Cb◦ .
(c) lim supn→∞ µn [F ] ≤ µ0 [F ] for all closed F ∈ B◦ .
(d) lim inf n→∞ µn [G] ≥ µ0 [G] for all open G ∈ B◦ .
(e) µn [A] → µ0 [A] for every A ∈ B◦ for which B◦ contains an open set G and a closed
set F such that G ⊆ A ⊆ F and µ0 [F \ G] = 0.
´
´
(f ) f dµn → f dµ0 for all f ∈ BL◦1 .
(g) d◦BL (µn , µ0 ) → 0.
Proof The equivalence of the conditions (a), (b), (c), (d), and (e) is known from
Theorem 6.3 of Billingsley (1999), and the implications (b)⇒(f) is trivial. The arguments
in the proof of (b)⇒(c) in Theorem 6.3 of Billingsley (1999) also prove the implication
(f)⇒(c). Indeed, the function f defined in (6.1) in Billingsley (1999) is bounded by 1 and
Lipschitz continuous with Lipschitz constant ε−1 , εf is an element of BL◦1 for ε ∈ (0, 1],
30
´
´
´
´
and f dµn → f dµ if and only if εf dµn → εf dµ. Finally, the equivalence of (a)
and (g) was discussed in Example IV.3.22 of Pollard (1984).
✷
The following Theorem A.4 is a special case of Theorem 15.12 in Aliprantis and Border
(2006). Recall that a topological space is separable if it contains a countable dense subset;
a subset is dense in a topological space if its closure coincides with the whole space.
Theorem A.4 The topological space (M1 , Ow ) is metrizable and separable if (E, d) is
separable.
The bounded Lipschitz distance dBL provides a metric on M1 when (E, d) is separable;
cf. Remark A.2. Also recall that the topology generated by a metric consists of all d-open
subsets of the underlying space. As a consequence of Theorem A.4 and the Portmanteau
theorem A.3 we obtain the following well known result.
Corollary A.5 If (E, d) is separable, then the bounded Lipschitz distance dBL generates
the weak topology Ow on M1 .
Proof First, two topologies O and O′ on a nonempty set coincide if and only if the
identity is a homeomorphism w.r.t. O and O′ . Second, a topology is first countable if it
is metrizable; cf. Aliprantis and Border (2006, p. 27). Thus it follows by the second part
of Theorem 2.40 in Aliprantis and Border (2006) that two metrizable topologies coincide
if and only if convergence of any sequence in O implies convergence of the sequence in
O′ and vice versa. By Theorem A.4 the topology Ow is metrizable, and the topology
O(dBL ) generated by the metric dBL is metrizable anyway. Thus the equivalence of (a)
and (g) in Theorem A.3 implies Ow = O(dBL ), i.e. the metric dBL indeed generates the
weak topology Ow .
✷
B. Convergence in distribution and convergence in
probability for the open-ball σ-algebra
Let (E, d) be a metric space and B◦ the open-ball σ-algebra on E. A sequence (Xn ) of
(E, B◦ )-valued random variables is said to converge in distribution◦ to an (E, B◦ )-valued
random variable X0 if the sequence (law{Xn }) weak◦ ly converges to law{X0 }. In this
case, we write Xn ❀◦ X0 . In the case where the random variables Xn , n ∈ N0 , are all
defined on the same probability space (Ω, F , P) the sequence (Xn ) is said to converge
in probability◦ to X0 if the mappings ω 7→ d(Xn (ω), X0(ω)), n ∈ N, are (F , B(R+ ))measurable and satisfy
lim P[d(Xn , X0 ) ≥ ε] = 0
n→∞
31
for all ε > 0.
(45)
In this case, we write Xn →p,◦ X0 . As usual, by P-almost sure convergence of the
sequence (Xn ) to X0 , abbreviated by Xn → X0 P-a.s., we will mean that there exists a
set N ∈ F with that P[N] = 0 and d(Xn (ω), X0(ω)) → 0 for all ω ∈ Ω \ N.
Proposition B.1 Let Xn , n ∈ N0 , be (E, B◦ )-valued random variables on a common
probability space (Ω, F , P), and assume that the mappings ω 7→ d(Xn (ω), X0(ω)), n ∈ N,
are (F , B(R+ ))-measurable. Then Xn → X0 P-a.s. implies Xn →p,◦ X0 .
Proof By assumption the variables d(Xn , X0 ), n ∈ N, are (F , B(R+ ))-measurable, and
therefore the variable lim supn→∞ d(Xn , X0 ) is (F , B(R+ ))-measurable. Since Xn → X0
P-a.s., we obtain P[lim supn→∞ d(Xn , X0 ) = 0] = 1. This implies
P lim sup{d(Xn , X0 ) ≥ ε} ≤ P lim sup d(Xn , X0 ) ≥ ε = 0
n→∞
for all ε > 0
n→∞
which together with the reverse of Fatou’s lemma gives lim supn→∞ P[d(Xn , X0 ) ≥ ε] = 0
for every ε > 0.
✷
When X0 takes almost surely values in a separable measurable set, then convergence
in probability◦ implies convergence in distribution◦ of Xn to X0 :
Proposition B.2 Let Xn , n ∈ N0 , be (E, B◦ )-valued random variables on a common
probability space (Ω, F , P), and assume that P[X0 ∈ E0 ] = 1 for some separable E0 ∈ B◦ .
Then Xn →p,◦ X0 implies Xn ❀◦ X0 .
´
´
Proof For any f ∈ BL◦1 we have | f dPXn − f dPX0 | ≤ 2 P[d(Xn , X0 ) ≥ ε/2] + ε/2 for
´
´
all ε > 0, i.e. f dPXn → f dPX0 . The claim then follows by the implication (f)⇒(a)
in the Portmanteau theorem A.3.
✷
The following lemma implies that the measurability condition in the definition of
convergence in probability◦ is automatically satisfied when X0 is constant, i.e. when
X0 (·) = x for some x ∈ E.
Lemma B.3 For every x ∈ E, the mapping y 7→ d(x, y) is continuous and (B◦ , B(R))measurable.
Proof The continuity is obvious, and the (B◦ , B(R))-measurability follows by
{d(x, ·) < a} = {y ∈ E : d(x, y) < a} = Ba (x) ∈ B◦
and {d(x, ·) < a} = ∅ ∈ B◦ for every a ≤ 0.
for every a > 0
✷
For constant X0 we also have that convergence in probability◦ of Xn to X0 is equivalent
to convergence in distribution◦ of Xn to X0 :
32
Proposition B.4 Let Xn , n ∈ N, be (E, B◦ )-valued random variables on a common
probability space (Ω, F , P), and x0 ∈ E be a constant. Then:
(i) Xn → x0 P-a.s. implies Xn →p,◦ x0 .
(ii) Xn →p,◦ x0 if and only if Xn ❀◦ x0 .
Proof Part (i) follows from Proposition B.1 and Lemma B.3. To prove part (ii), first
assume Xn ❀◦ x0 . Set f (x) := min{d(x, x0 ); 1}, x ∈ E, and note that f ∈ Cb◦ . By
Markov’s inequality and Lemma B.3 we obtain
ˆ
ˆ
1
1
P[d(Xn , x0 ) ≥ ε] ≤
f (Xn (ω)) P[dω] −→
f (x0 ) P[dω] = 0, n → ∞
ε
ε
for every ε > 0. That is, Xn →◦ x0 . The other direction in part (ii) follows from
T
Proposition B.2, because the set {x0 } = n∈N B1/n (x0 ) is separable and lies in B◦ . ✷
Recall that B◦ = B when (E, d) is separable. In this case we suppress the superscript
◦
and write simply ❀, →p , convergence in distribution, and convergence in probability instead of ❀◦ , →p,◦ , convergence in distribution◦ , and convergence in probability◦ ,
respectively.
C. An extended Continuous Mapping theorem and a
delta-method for the open-ball σ-algebra
As mentioned in the introduction, Theorem 3.1 is based on a generalization of Theorem 4.1 in Beutner and Zähle (2010), which in turn is a generalization of the classical
functional delta-method in the form of Theorem 3 of Gill (1989). The proof of the
generalization of Theorem 4.1 in Beutner and Zähle (2010) is based on the extended
Continuous Mapping theorem C.1 below. An extended Continuous Mapping theorem
for convergence in distribution for the Borel σ-algebra can be found in Kallenberg (2002,
Theorem 4.27). A corresponding result for convergence in distribution in the HoffmannJørgensen is given, for example, in van der Vaart and Wellner (1996, Theorem 1.11.1).
However, we could not find a version of this result for convergence in distribution◦ for
the open-ball σ-algebra. So we include a proof for Theorem C.1. Note that Theorem
C.1 is a generalization of the “ordinary” Continuous Mapping theorem for convergence
in distribution◦ for the open-ball σ-algebra as given by Billingsley (1999, Theorem 6.4).
e d e ) be metric spaces and B◦ and Be◦ be the open-ball σ-algebras on
Let (E, d) and (E,
E
e respectively.
E and E,
Theorem C.1 (Extended CMT for random variables) Let En ⊆ E and ξn be an
(E, B◦ )-valued random variable on some probability space (Ωn , Fn , Pn ) such that ξn (Ωn ) ⊆
33
En , n ∈ N. Let ξ0 be an (E, B◦ )-valued random variable on some probability space
e be a
(Ω0 , F0 , P0 ) such that ξ0 (Ω0 ) ⊆ E0 for some separable E0 ∈ B◦ . Let hn : En → E
e is (Fn , Be◦ )-measurable, n ∈ N. Let h0 : E0 → E
e
map such that the map hn (ξn ) : Ωn → E
be a (B0◦ , Be◦ )-measurable map, where B0◦ := B◦ ∩ E0 (⊆ B◦ ). Moreover, assume that the
following two assertions hold:
(a) ξn ❀◦ ξ0 .
e n (xn ), h0 (x0 )) → 0 when d(xn , x0 ) → 0.
(b) For every xn ∈ En , n ∈ N0 , we have d(h
Then hn (ξn ) ❀◦ h0 (ξ0 ).
Remark C.2 Note that we do not assume in Theorem C.1 that the maps hn , n ∈ N,
are (B◦ , Be◦ )-measurable. This implies that for n ∈ N the law Pn ◦ (hn (ξn ))−1 of hn (ξn )
can not necessarily be represented as the image law of ξn ’s law Pn ◦ ξn−1 w.r.t. hn .
✸
Proof of Theorem C.1 According to the implication (d)⇒(a) in the Portmanteau
e ≥ P0 ◦ h0 (ξ0 )−1 [G]
e for
theorem A.3, it suffices to show that lim inf n→∞ Pn ◦ hn (ξn )−1 [G]
e ∈ Be◦ . So, let G
e ∈ Be◦ be open. First we note that
every open set G
e
h−1
0 (G)
∩ E0 ⊆
∞ n \
∞
[
m=1
k=m
e
h−1
k (G)
oint \
E0 ,
(46)
e
where the superscript int refers to the interior of a set. Indeed: For every x0 ∈ h−1
0 (G)∩E0
e for all k ≥ m
there exists an m ∈ N and a neighborhood U of x0 such that hk (x) ∈ G
and x ∈ U. Otherwise we could find for every m ∈ N some km ≥ m and xm ∈ B1/m (x0 )
e km (xm ), h0 (x0 )) 6→ 0
e But then we had d(xm , x0 ) → 0 and d(h
such that hkm (xm ) 6∈ G.
e and G
e is open), which contradicts assumption (b).
(take into account that h0 (x0 ) ∈ G
T
T∞
∞
−1 e int
−1 e
e
Hence U ⊆ k=m h−1
k (G) and thus x0 ∈ { k=m hk (G)} . In particular, h0 (G) ∩ E0 ⊆
S∞ T∞
−1 e int
m=1 { k=m hk (G)} . Now (46) is obvious.
Further, for every m ∈ N we can find a union Gm of countably many open balls such
that
∞
∞
n\
oint \
n\
oint
−1 e
−1 e
h (G)
E 0 ⊆ Gm ⊆
h (G)
,
(47)
k
k
k=m
k=m
and we may and do assume G1 ⊆ G2 ⊆ · · · . To prove this one can proceed by an
T
e int we can find an open ball
induction on m. First let m = 1. For every x ∈ { ∞
h−1
k (G)}
T∞ −1 k=1 int
e
, because the latter set is open.
Brx (x) around x which is contained in { k=1 hk (G)}
T
−1 e int
The system which consists of the open balls Brx (x), x ∈ { ∞
k=1 hk (G)} , provides an
T
T∞ −1
e int E0 . Since the latter set is separable (recall that E0 was
open cover of { k=1 hk (G)}
assumed to be separable), Lindelöf’s theorem ensures that there is a countable subcover.
The set G1 can now be defined as the union of the elements of this subcover. Next assume
that G1 , . . . , GM are unions of countably many open balls such that G1 ⊆ · · · ⊆ GM and
34
T
e int we can find an open
(47) holds for m = 1, . . . , M. For every x ∈ { ∞
h−1
k (G)}
T∞ k=M +1
e int , because the latter set is
(G)}
ball Brx (x) around x which is contained in { k=M +1 h−1
T∞ k
−1 e int
e int T∞
open. The system which consists of Brx (x), x ∈ { k=M +1 h−1
k (G)} \ { k=M hk (G)}
and of the countably many open balls which unify to GM provides an open cover of
T
−1 e int T
E0 . Since the latter set is separable, Lindelöf’s theorem ensures
{ ∞
k=M +1 hk (G)}
that there is a countable subcover. Without loss of generality we may and do assume that
the countably many open balls which unify to GM belong to this countable subcover.
Defining GM +1 as the union of the elements of this subcover we obtain GM ⊆ GM +1 and
(47) for m = M + 1.
As countable unions of open balls the sets Gm , m ∈ N, are open and lie in B◦ . Then,
using (46), the first “⊆” in (47), and the inclusions G1 ⊆ G2 ⊆ · · · (along with the
continuity from below of P0 ◦ ξ0−1 ),
e = P0 ◦ ξ −1 h−1 (G)
e
P0 ◦ h0 (ξ0 )−1 G
0
0
e
= P0 ◦ ξ0−1 h−1
0 (G) ∩ E0
∞ n \
∞
h
oint \ i
[
out
−1 e
≤ P0 ξ 0 ∈
hk (G)
E0
m=1
k=m
∞
h
i
[
≤ P0 ξ 0 ∈
Gm
m=1
≤ P0 ◦ ξ0−1
∞
h[
= sup P0 ◦
m∈N
Gm
i
m=1
ξ0−1[Gm ]
≤ sup lim inf Pn ◦ ξn−1 [Gm ],
m∈N
(48)
n→∞
where the last step follows from assumption (a) and the implication (a)⇒(d) in the
Portmanteau theorem A.3. Now, (48) and the second “⊆” in (47) yield
∞
i
h
\
−1 e
out
e
hk (G)
P0 ◦ h0 (ξ0 ) G ≤ sup lim inf Pn ξn ∈
−1
n→∞
m∈N
k=m
e
≤ lim inf Pn ξn ∈ h−1
n (G)
n→∞
e .
= lim inf Pn ◦ hn (ξn )−1 G
n→∞
This completes the proof.
✷
Before giving the generalization of Theorem 4.1 in Beutner and Zähle (2010) we recall
e be vector spaces,
the definition of quasi-Hadamard differentiability. For this let V and E
e respectively.
and E ⊆ V be a subspace of V. Let k · kE and k · kEe be norms on E and E,
35
e be a map
Definition C.3 (Quasi-Hadamard differentiability) Let H : VH → E
defined on some VH ⊆ V, and E0 be a subset of E. Then H is said to be quasiHadamard differentiable at x ∈ VH tangentially to E0 hEi if there is some continuous
e such that
map Ḣx : E0 → E
lim Ḣx (x0 ) −
n→∞
H(x + εn xn ) − H(x)
εn
e
E
= 0
(49)
holds for each triplet (x0 , (xn ), (εn )), with x0 ∈ E0 , (xn ) ⊆ E satisfying kxn − x0 kE → 0
as well as (x + εn xn ) ⊆ VH , and (εn ) ⊂ (0, ∞) satisfying εn → 0. In this case the map
Ḣx is called quasi-Hadamard derivative of H at x tangentially to E0 hEi.
e is a vector space equipped with a norm k · k e , and let 0 e denote the
Recall that E
E
E
e := E
e ×E
e and let Be◦ be the σ-algebra on E
e generated by the open
e Set E
null in E.
e x1 , x
balls w.r.t. the metric d((e
e2 ), (e
y1 , ye2)) := max{ke
x1 − ye1 kEe ; ke
x2 − ye2 kEe }. Recall that
◦
◦
e
e is the product of two k · k e -open balls in E.
e
ball in E
Be◦ ⊆ Be ⊗ Be , because any d-open
E
Let (Ωn , Fn , Pn ) be a probability space and Xn : Ωn → E be any map, n ∈ N. Recall
that ❀◦ and →◦ refer to convergence in distribution◦ and convergence in probabilityp,◦ ,
respectively.
e be a map defined on some VH ⊆ E,
Theorem C.4 (Delta-method) Let H : VH → E
◦
and x ∈ VH . Let E0 ∈ B be some k · kE -separable subset of E. Let (an ) be a sequence
of positive real numbers tending to ∞, and consider the following conditions:
(a) Xn takes values only in VH .
(b) an (Xn − x) takes values only in E, is (Fn , B◦ )-measurable and satisfies
an (Xn − x) ❀◦ X0
in (E, B◦ , k · kE )
(50)
for some (E, B◦ )-valued random variable X0 on some probability space (Ω0 , F0 , P0 )
with X0 (Ω0 ) ⊆ E0 .
(c) an (H(Xn ) − H(x)) is (Fn , Be◦ )-measurable.
(d) The map H is quasi-Hadamard differentiable at x tangentially to E0 hEi with quasie
Hadamard derivative Ḣx : E0 → E.
(e) (Ωn , Fn , Pn ) = (Ω, F , P) for all n ∈ N.
(f ) The quasi-Hadamard derivative Ḣx can be extended to E such that the extension
e is continuous at every point of E0 and (B◦ , Be◦ )-measurable.
Ḣx : E → E
e→E
e defined by h(e
(g) The map h : E
x1 , x
e2 ) := x
e1 − x
e2 is (Be◦ , Be◦ )-measurable.
36
Then the following two assertions hold:
(i) If conditions (a)–(d) hold true, then Ḣx (X0 ) is (F0 , Be◦ )-measurable and
e Be◦ , k · k e ).
an H(Xn ) − H(x) ❀◦ Ḣx (X0 )
in (E,
E
(ii) If conditions (a)–(g) hold true, then
an H(Xn ) − H(x) − Ḣx an (Xn − x) →p,◦ 0Ee
e k · k e ).
in (E,
E
(51)
Remark C.5 It is apparent from the following proof that for part (i) of Theorem C.4
it is not necessary to assume (as in Definition C.3) that the quasi-Hadamard derivative
Ḣx is continuous. It would suffice to require in Definition C.3 that the map Ḣx is
(B0◦ , Be◦ )-measurable for the trace σ-algebra B0◦ := B◦ ∩ E0 (⊆ B◦ ).
✸
Proof of Theorem C.4 For the proof of part (i) we adapt the arguments in the proof
of Theorem 3.9.4 in van der Vaart and Wellner (1996), which then allow for an easy
proof of part (ii).
(i): For every n ∈ N, let En := {xn ∈ E : a−1
n xn + x ∈ VH } and define the map
e by
hn : En → E
H(x + a−1
n xn ) − H(x)
.
hn (xn ) :=
a−1
n
e by
Moreover, define the map h0 : E0 → E
h0 (x0 ) := Ḣx (x0 ).
Now, the claim would follow by the extended Continuous Mapping theorem C.1 applied
to the functions hn , n ∈ N0 , and the random variables ξn := an (Xn − x), n ∈ N, and
ξ0 := X0 if we can show that the assumptions of Theorem C.1 are satisfied. First,
ξn (Ωn ) ⊆ En and ξ0 (Ω0 ) ⊆ E0 clearly hold. Second, by assumption (c) we have that
hn (ξn ) = an (H(Xn ) − H(x)) is (Fn , Be◦ )-measurable. Third, the map h0 is continuous
by assumption (on the quasi-Hadamard derivative). Thus h0 is (B0◦ , Be◦ )-measurable,
because the trace σ-algebra B0◦ := B◦ ∩ E0 coincides with the Borel σ-algebra on E0
(recall that E0 is separable). In particular, Ḣx (X0 ) is (F0 , Be◦ )-measurable. Fourth,
condition (a) of Theorem C.1 holds by assumption (b). Fifth, condition (b) of Theorem
C.1 is ensured by assumption (d) (note that (d) implies (49)).
e by
(ii): For every n ∈ N, let En and hn be as above and define the map hn : En → E
hn (xn ) := (hn (xn ), Ḣx (xn )).
e by
Moreover, define the map h0 : E0 → E
h0 (x0 ) := (h0 (x0 ), Ḣx (x0 )) = (Ḣx (x0 ), Ḣx (x0 )).
37
We will first show that
hn (an (Xn − x)) ❀◦ h0 (X0 )
e
e Be◦ , d).
in (E,
(52)
For (52) it suffices to show that the assumption of the extended Continuous Mapping
theorem C.1 applied to the functions hn and ξn (as defined above) are satisfied. The
claim then follows by Theorem C.1. First, we have already observed that ξn (Ωn ) ⊆
En and ξ0 (Ω0 ) ⊆ E0 . Second, we have seen in the proof of part (i) that hn (ξn ) is
e is
(Fn , Be◦ )-measurable, n ∈ N. By assumption (f) the extended map Ḣx : E → E
◦ e◦
◦
(B , B )-measurable, which implies that Ḣx (ξn ) is (Fn , Be )-measurable. Thus, hn (ξn ) =
(hn (ξn ), Ḣx (ξn )) is (Fn , Be◦ ⊗ Be◦ )-measurable (to see this note that, in view of Be◦ ⊗ Be◦ =
e = E
e × E,
e Theorem 7.4 of Bauer
σ(π1 , π2 ) for the coordinate projections π1 , π2 on E
(2001) shows that the map (hn (ξn ), Ḣx (ξn )) is (Fn , Be◦ ⊗ Be◦ )-measurable if and only if
the maps hn (ξn ) = π1 ◦ (hn (ξn ), Ḣx (ξn )) and Ḣx (ξn ) = π2 ◦ (hn (ξn ), Ḣx (ξn )) are (Fn , Be◦ )-
measurable). In particular, the map hn (ξn ) = (hn (ξn ), Ḣx (ξn )) is (Fn , Be◦ )-measurable,
n ∈ N. Third, we have seen in the proof of part (i) that the map h0 = Ḣx is (B0◦ , Be◦ )measurable. Thus the map h0 is (B0◦ , Be◦ ⊗ Be◦ )-measurable (one can argue as above)
and in particular (B0◦ , Be◦ )-measurable. Fourth, condition (a) of Theorem C.1 holds by
assumption (b). Fifth, condition (b) of Theorem C.1 is ensured by assumption (d) and
the continuity of the extended map Ḣx at every point of E0 (recall assumption (f)).
Hence, (52) holds.
By assumption (g) and the ordinary Continuous Mapping theorem (cf. Billingsley
e → E,
e (e
(1999, Theorem 6.4)) applied to (52) and the map h : E
x1 , x
e2 ) 7→ x
e1 − x
e2 , we
now have
hn (an (Xn − x)) − Ḣx (an (Xn − x)) ❀◦ Ḣx (X0 ) − Ḣx (X0 ),
i.e.
an H(Xn ) − H(x) − Ḣx an (Xn − x) ❀◦ 0Ee .
By Proposition B.4 we can conclude (51).
✷
D. Probability kernels and conditional distributions
Let (Ω, F ) be a measurable space. Let (E, d) be a metric space and B◦ be the open-ball
σ-algebra on E. A map P : Ω × B◦ → [0, 1] is said to be a probability kernel from (Ω, F )
to (E, B◦ ) if P ( · , A) is (F , B([0, 1]))-measurable for every A ∈ B◦ , and P (ω, · ) is a
probability measure on (E, B◦ ) for every ω ∈ Ω. Of course, we may regard P as a map
from Ω to M◦1 . Recall that M◦1 = M1 when (E, d) is separable. If in this case the set
M1 is equipped with the weak topology Ow , then a probability kernel can be regarded
as an M1 -valued random variable (w.r.t. any probability measure on (Ω, F )):
38
Lemma D.1 Let (E, d) be separable and P be a probability kernel from (Ω, F ) to (E, B).
Then the mapping ω 7→ P (ω, •) is (F , σ(Ow ))-measurable.
Proof Since (E, d) was assumed to be separable, the proof of the implication (4)⇒(1)
in Theorem 19.7 in Aliprantis and Border (2006) shows that σ(Ow ) equals the σ-algebra
generated by the system {πf−1 (A) : f ∈ Cb , A ⊆ R open}. So it suffices to show that
the set
ˆ
−1
−1 −1
−1
P ( · , •) (πf (A)) = πf (P ( · , •)) (A) =
f (x) P ( · , dx)
(A)
is contained in F for every open A ⊆ R and f ∈ Cb . But this follows from the well known
´
fact (see e.g. Lemma 1.41 in Kallenberg (2002)) that the mapping ω 7→ f (x)P (ω, dx)
is (F , B(R))-measurable for every f ∈ Cb . This finishes the proof.
✷
Now, let (Ω′ , F ′) and (D, D) be further measurable spaces. Let P and P′ be probability
measures on respectively Ω and Ω′ , and set (Ω, F, P) := (Ω × Ω′ , F ⊗ F ′ , P ⊗ P′). Let Y :
Ω → D be an (F , D)-measurable map and X : Ω → E be an (F, B◦ )-measurable map.
Note that Y can also be regarded as a (D, D)-valued random variable on (Ω, F , P), and
we are doing that in Lemma D.2. The following lemma shows that under an additional
assumption, the conditional distribution of X given Y can be specified explicitly.
Lemma D.2 Assume that X(ω, ω ′) = g(Y (ω), ω ′) holds for all (ω, ω ′) ∈ Ω and some
(D ⊗ F ′ , B◦ )-measurable map g : D × Ω′ → E. Then the map P : Ω × B◦ → [0, 1] defined
by
P ((ω, ω ′), A) := P (ω, A) := P′ ◦ X(ω, ·)−1[A],
(ω, ω ′) ∈ Ω, A ∈ B◦
provides a conditional distribution of X given Y .
Proof First, P provides a probability kernel from (Ω, σ(Y )) to (E, B◦ ). Indeed: The
mapping ω
e ′ 7→ X(ω, ω
e ′) is (F ′, B◦ )-measurable for every fixed ω ∈ Ω, because X is
(F , B◦ )-measurable. So it immediately follows that the mapping A′ 7→ P (ω, A′) is
a probability measure on (E, B◦ ) for every ω ∈ Ω. Further, the mapping (ω, ω
e ′) 7→
(Y (ω), ω
e ′) is clearly (σ(Y ) ⊗ F ′ , D ⊗ F ′)-measurable, which implies that the mapping
(ω, ω
e ′) 7→ X(ω, ω
e ′) = g(Y (ω), ω
e ′) is (σ(Y ) ⊗ F ′ , B◦ )-measurable. By Tonelli’s part of
´
Fubini’s theorem it follows that the mapping ω 7→ 1A (X(ω, ω
e ′)) P′ [de
ω ′] = P (ω, A) is
(σ(Y ), B([0, 1]))-measurable for every A ∈ B◦ . In particular, the mapping (ω, ω ′) 7→
P ((ω, ω ′), A) = P (ω, A) is (σ(Y ), B([0, 1]))-measurable for every A ∈ B◦ .
Second, by Fubini’s theorem we obtain for every B ∈ D and A ∈ B◦ ,
ˆ
ˆ
′
′
P ((ω, ω ), A) P[d(ω, ω )] =
P′ ◦ X(ω, ·)−1[A] P[d(ω, ω ′)]
{Y ∈B}
{Y ∈B}
ˆ
=
P′ ◦ X(ω, ·)−1[A] P[dω]
{Y ∈B}
39
=
¨
=
ˆ
1{Y ∈B} (ω) 1{X(ω,·)∈A} (ω ′ ) P′ [dω ′] P[dω]
1{Y ∈B} (ω) 1{X(ω,·)∈A} (ω ′ ) P[d(ω, ω ′)]
= P {Y ∈ B} ∩ {X ∈ A} .
This completes the proof.
✷
Acknowledgement
The second author gratefully acknowledges support by BMBF through the project HYPERMATH under grant 05M13TSC.
References
Ahn, J.Y. and Shyamalkumar, N.D. (2011). Large sample behavior of the CTE and
VaR estimators under importance sampling. North American Actuarial Journal, 15,
393–416.
Aliprantis, C.D. and Border, K.C. (2006). Infinite dimensional analysis. A hitchhiker’s
guide. Springer-Verlag, Berlin.
Arcones, M. and Giné, E. (1992). The bootstrap of U- and V-statistics. Annals of Statistics, 20, 655–674.
Arcones, M.A. and Yu, B. (1994). Central limit theorems for empirical and U-processes
of stationary mixing sequences. Journal of Theoretical Probability, 7, 47–71.
Bauer, H. (2001). Measure and integration theory. de Gruyter, Berlin.
Beutner, E., Wu, W.B. and Zähle, H. (2012). Asymptotics for statistical functionals of
long-memory sequences. Stochastic Processes and their Applications, 122, 910–929.
Beutner, E. and Zähle, H. (2010). A modified functional delta method and its application
to the estimation of risk functionals. Journal of Multivariate Analysis, 101, 2452–2463.
Beutner, E. and Zähle, H. (2012). Deriving the asymptotic distribution of U- and Vstatistics of dependent data using weighted empirical processes. Bernoulli, 18, 803–
822.
Beutner, E. and Zähle, H. (2014). Continuous mapping approach to the asymptotics of
U- and V-statistics. Bernoulli, 20, 846–877.
40
Bickel, P.J. and Freedman D.A. (1981). Some asymptotic theory for the bootstrap.
Annals of Statistics, 9, 1196–1217.
Billingsley, P. (1999). Convergence of probability measures. Wiley, New York.
Boussama, F., Fuchs, F. and Stelzer, R. (2011). Stationarity and geometric ergodicity
of BEKK multivariate GARCH models, Stochastic Processes and their Applications,
121, 2331–2360.
Buchsteiner, J. (2015). Weak convergence of the sequential empirical process of some
long-range dependent sequences with respect to a weighted norm. Statistics & Probability Letters, 96, 170–179.
Bücher, A. and Kojadinovic, I. (2015). Dependent multiplier bootstraps for nondegenerate U-statistics under mixing conditions with applications, Journal of Statistical Planning and Inference, to appear, doi:10.1016/j.jspi.2015.09.006.
Bühlmann, P. (1994). Blockwise bootstrapped empirical process for stationary sequences.
Annals of Statistics, 22, 995–1012.
Bühlmann, P. (1995). The blockwise bootstrap for general empirical processes of stationary sequences. Stochastic Processes and their Applications, 58, 247–265.
Chibisov, D.M. (1965). An investigation of the asymptotic power of tests of fit. Theory
of Probability and its Applications, 10, 421–437.
Cuevas, A. and Romo, J. (1997). Differentiable functionals and smoothed bootstrap,
Annals of the Institute of Statistical Mathematics, 49, 355–370.
Davison, A.C. and Hinkley, D.V. (1997). Bootstrap methods and their application. Cambridge University Press, Cambridge.
Dehling, H. and Wendler, M. (2010). Central limit theorem and the bootstrap for Ustatistics of strongly mixing data. Journal of Multivariate Analysis, 101, 126–137.
Denneberg, D. (1994). Non-additive measure and integral. Kluwer, Dordrecht.
Doukhan, P., Lang, G., Leucht, A. and Neumann, M.H. (2015). Dependent wild bootstrap for the empirical process. Journal of Time Series Analysis, 36, 290–314.
Dudley, R.M. (1966). Weak convergence of probabilities on nonseparable metric spaces
and empirical measures on Euclidean spaces. Illinois Journal of Mathematics, 10,
109–126.
Dudley, R.M. (1967). Measures on non-separable metric spaces. Illinois Journal of Mathematics, 11, 449–453.
41
Dudley, R.M. (1990). Nonlinear functionals of empirical measures and the bootstrap.
In: E. Eberlein, J. Kuelbs, M.B. Marcus (Eds.), Probability in Banach Spaces, Vol. 7,
Birkhäuser, Boston (1990), 63–82.
Dudley, R.M. (1999). Uniform central limit theorems. Cambridge University Press, Cambridge.
Dudley, R.M. (2002). Real analysis and probability. Cambridge University Press, Cambridge.
Dudley, R.M. (2010). An extended Wichura theorem, definitions of Donsker class, and
weighted empirical distributions, In: E. Giné, V. Koltchinskii, R. Norvaisa (Eds.),
Selected Works of R.M. Dudley, Springer, New York, 38–75.
Efron, B. (1979). Bootstrap methods: Another look at the jackknife. Annals of Statistics,
7, 1–26.
Efron, B. and Tibshirani, R. (1994). An introduction to the bootstrap. Chapman & Hall,
New York.
Föllmer, H. and Schied, A. (2011). Stochastic finance. An introduction in discrete time.
de Gruyter, Berlin.
Francq, C. and Zakoı̈an, J.-M. (2010). GARCH Models. Structure, statistical inference
and financial applications. Wiley, Chichester.
Gänssler, P. (1986). Bootstrapping empirical measures indexed by Vapnik–Červonenkis
class of sets. In: Proceedings of the IV Vilnius Conference (1985), Probability Theory
and Mathematical Statistics, Vol. 1 (VNU Science Press, Utrecht), 467–481.
Gill, R.D. (1989). Non- and semi-parametric maximum likelihood estimators and the
von Mises method (Part I). Scandinavian Journal of Statistics, 16, 97–128.
Giné, E. and Zinn, J. (1990). Bootstrapping general empirical measures. Annals of Probability, 18, 851–869.
Gribkova, N. (2002). Bootstrap approximation of distributions of the L-statistics. Journal of Mathematical Sciences, 109, 2088–2102.
Helmers, R., Janssen, P. and Serfling, R. (1990). Berry–Esséen and bootstrap results for
generalized L-statistics, Scandinavian Journal of Statistics, 17, 65–77.
Hoffmann-Jørgensen, J. (1984). Stochastic processes in Polish spaces. Unpublished
manuscript.
42
Janssen, P. (1994). Weighted bootstrapping of U-statistics. Jornal of Statistical Planning
and Inference, 38, 31–41.
Kallenberg, O. (2002). Foundations of modern probability. Springer-Verlag, New York.
Klenke, A. (2014). Probability theory. Springer-Verlag, Berlin.
Kosorok, M.R. (2008). Bootstrapping the Grenander estimator, In: Balakrishnan, N.,
Peña, E.A., Silvapulle, M.J. (Eds.). Beyond Parametrics in Interdisciplinary Research:
Festschrift in Honor of Professor Pranab K. Sen., Institute of Mathematical Statistics:
Hayward, CA., 282–292.
Kosorok, M.R. (2010). Introduction to empirical processes and semiparametric inference.
Springer Science+Business Media, LLC, New York.
Krätschmer, V., Schied, A. and Zähle, H. (2015). Quasi-Hadamard differentiability of
general risk functionals and its application. Statistics and Risk Modeling, 32, 25–47.
Krätschmer, V. and Zähle, H. (2016). Statistical inference for expectile-based risk measures. Submitted for publication (arXiv:1601.05261).
Künsch, H.R. (1989). The jackknife and the bootstrap for general stationary observations. Annals of Statistics, 17, 1217–1241.
Lahiri, S.N. (2003). Resampling methods for dependent data. Springer-Verlag, New York.
Liu, R.Y. and Singh, K. (1992). Moving blocks jackknife and bootstrap capture weak
dependence. In: R. LePage and L. Billard (Eds.), Exploring the limits of bootstrap.
Wiley, New York, 225–248.
Naik-Nimbalkar, U.V. and Rajarshi, M.B. (1994). Validity of blockwise bootstrap for
empirical processes with stationary observations. Annals of Statistics, 22, 980–994.
Parr, W.C. (1985). The bootstrap: Some large sample theory and connections with
robustness, Statistics and Probability Letters, 3, 97–100.
Politis, D.N. and Romano, J.P. (1992). A circular block-resampling procedure for stationary data, In: R. LePage and L. Billard (Eds.), Exploring the Limits of Bootstrap.
Wiley, New York, 263–270.
Pollard, D. (1984). Convergence of stochastic processes. Springer-Verlag, New York.
Radulovic, D. (1996). The bootstrap for empirical processes based on stationary observations. Stochastic Processes and their Applications, 65, 259–279.
Shao, J. and Tu, D. (1995). The jackknife and bootstrap. Springer-Verlag, New York.
43
Shao, X. (2010). The dependent wild bootstrap. Journal of the American Statistical
Association, 105, 218–235.
Shorack, G.R. and Wellner, J.A. (1986). Empirical processes with applications to statistics. Wiley, New York.
van der Vaart, A.W. and Wellner, J.A. (1996). Weak convergence and empirical processes. Springer-Verlag, New York.
Volgushev, S. and Shao, X. (2014). A general approach to the joint asymptotic analysis
of statistics from sub-samples. Electronic Journal of Statistics, 8, 390–431.
Wang, S. and Dhaene, J. (1998). Comontonicity, correlation order and premium principles. Insurance: Mathematics and Economics 22, 235–242.
44
| 10math.ST
|
MedTQ: Dynamic Topic Discovery and Query Generation for Medical
Ontologies
Feichen Shen, PhD, Yugyung Lee, PhD
School of Computing and Engineering
University of Missouri-Kansas City, MO, USA
Abstract
Background
Biomedical ontology refers to a shared conceptualization for a biomedical domain of interest that has vastly
improved data management and data sharing through the open data movement. The rapid growth and availability of
biomedical data make it impractical and computationally expensive to perform manual analysis and query
processing with the large scale ontologies. The lack of ability in analyzing ontologies from such a variety of sources,
and supporting knowledge discovery for clinical practice and biomedical research should be overcome with new
technologies.
Methods
In this study, we developed a Medical Topic discovery and Query generation framework (MedTQ), which was
composed by a series of approaches and algorithms. A predicate neighborhood pattern-based approach introduced
has the ability to compute the similarity of predicates (relations) in ontologies. Given a predicate similarity metric,
machine learning algorithms have been developed for automatic topic discovery and query generation. The topic
discovery algorithm, called the hierarchical K-Means algorithm was designed by extending an existing supervised
algorithm (K-means clustering) for the construction of a topic hierarchy. In the hierarchical K-Means algorithm, a
level-by-level optimization strategy was selected for consistent with the strongly association between elements
within a topic. Automatic query generation was facilitated for discovered topic that could be guided users for
interactive query design and processing.
Results
A topic hierarchy was constructed using the DrugBank ontology as a case study. In this hierarchy, 8 specific topics
were generated. Ranking of predicates and concepts of these topics were also computed. An experiment has been
conducted to find an optimal number of the topics using the four different clustering algorithms, K-means,
Clustering Large Application (Clara), Partition Around Medoids (Pam), and Hierarchical Clustering. A number of
SPARQL queries generated were automatically generated from the discovered topics to demonstrate the ability to
retrieve information from the DrugBank ontology.
Conclusions
The paper addresses knowledge discovery through analysis of ontologies for clinical practice and biomedical
research. The model of predicate-oriented neighborhood pattern is explained in the context of topic discovery and
query generation for ontologies. The MedTQ framework enhances knowledge discovery by capturing underlying
structures from domain specific data and ontologies.
Keywords—knowledge discovery, query generation and processing, topic hierarchy
1
I. INTRODUCTION
In recent years, a large number of ontologies have been introduced for clinical practice and biomedical research. As
a result, there have been increasing demands on knowledge discovery and sharing of large scale biomedical data.
The first notable effort shown by the biomedical and scientific community toward connecting scattered medical data
is to materialize them through the open data movement (i.e., the Linked Data, bio2RDF, OBO, LinkedCT) 1.
However, many of these ontologies are still not fully annotated nor connected with other ontologies. The rapid
growth and availability of biomedical data make it impractical and computationally expensive to perform manual
analysis and query processing with the large scale ontologies2.
In reality, there is a limited capacity to carry out dynamic analysis and query processing with these large-scale
datasets. The current medical ontologies and services are not sufficient to be combined together due to the lack of
underlying cohesive structure and semantics3. The previous work on inferring ontology structure4 have mainly
focused on determining whether or not there was a relationship between a given pair of concepts irrespective of the
connections between them or the strength of the association. In particular, the techniques have yet to be fully
implemented on dynamic analysis and query processing with the ontologies. There are endpoint services (e.g.,
BioPortal) for biomedical research5, but most of them are not functioning properly or even if they are working,
specific query content and formats are not at a practical level.
Knowledge discovery through analysis of ontologies for clinical practice and biomedical research has become a
challenging task6, 7. We need an advanced approach to thoroughly understand ontologies instead of simply getting a
slice of reference ontology and applying them for a query process or decision support8. Subsequently, it is essential
to know what information exists and what meaningful relationships are present among associated domains (e.g.,
identification of genes responsible for a disease9, 10, development of drugs for their treatment11 or detect associations
between diseases and phenotypes12, 13). Once the structure of an ontology has been defined, it is useful to identify
and differentiate the context and strength of influence in domains and extract cohesive structure and semantics from
ontologies.
In this paper, we presented a semantic framework, called the MedTQ framework. The MedTQ performs dynamic
topic discovery (relationships) and automatic query generation through the analysis of predicates among concepts
and role names, called the Predicate Neighborhood Patterns (PNP) in biomedical ontologies. Furthermore, a new
clustering technique, called the Hierarchical Predicate-based K-Means clustering (HPKM) was proposed to
dynamically identify latent topics and automatically generate queries based on the discovered patterns. We have also
implemented an interactive tool that allows researchers to explore ontologies and generate queries by combining
interesting contents, and then retrieve relevant information in a logical way. In addition, topics were further
evaluated based on prioritized information of medical ontologies for biomedical research.
The contributions of this paper are fourfold.
Formal definition of predicate neighborhood patterns (PNP)
Hierarchical Predicate-based K-means clustering (HPKM)
Automatic query generation based on the discovered topics (clusters)
Interactive tool for dynamic query generation and an endpoint for query processing
A case study was designed with a major medical ontology (i.e., DrugBank 14) to demonstrate the dynamic topic
discovery and query generation by the MedTQ framework. We have implemented a prototype of the MedTQ system
and evaluated the statistical significance of our model in discovery of topics. In addition, we successfully validated
the clustering results, thereby providing a solid evidence for automatic query generation.
The major content of this paper is organized as follows: We first present the MedTQ framework in Section II. We
then describe the implementation of the MedTQ system in Section III. We present the main results and discussion in
Section IV and Section V. The conclusion is discussed in Section VI.
II.
METHODS
In this paper, we proposed a semantic framework, called the MedTQ that identifies the relationships present among
concepts and discovers knowledge through the construction of a hierarchy of topics (called the topic hierarchy) from
biomedical ontologies. In the topic hierarchy, the abstractions of topics are analyzed for preserving information that
is relevant in a given context (topic) without revealing the details of an underlying ontology structure. The topic
models based on the relationships and their neighborhood patterns are defined as a graph in different levels of
abstraction.
2
We first rationalized a predicate-centric model ‘Predicate Neighboring Patterns (PNP)’ that specifies high
connectivity on the RDF/OWL graph for information sharing. Second, we presented a Hierarchical Predicate-based
K-Means clustering (HPKM) algorithm to cluster the graph based on the PNP patterns. Finally, we presented a query
generation model for automatic query generation from the discovered topics.
A. Predicate Neighboring Patterns
The predicate neighboring patterns (PNP) defines the patterns of predicates playing an important role in sharing
information and connecting the concepts in the ontology. The RDF/OWL data model15 specifies resources
(information on the entities and their relationship in the given ontologies) in the form of triples <subject (S),
predicate (P), object (O)>, where S denotes the resource, and P denotes aspects of the resource and expresses a
relationship between S and O. Multiple S can be connected to multiple O through a single predicate. A predicate P is
representing a binary relation between two concepts (S and O) in ontologies. In RDF/OWL, P is represented as a
property to express a kind of relationship (e.g., rdfs:subClassOf) between domain (subject) and range (object). The
subject and object can be either from the same ontology or from different ontologies. From the basic unit of <S, P,
O>, a specific context of a predicate P can be discovered from the associated concepts (S and O). Interestingly, the
neighbors of predicates P will also provide additional information through the association context.
In this paper, two types of predicate patterns are defined as follows:
Share Pattern: As shown in Figure 1, this pattern describes the resources sharing relationships (P) between
interacting concepts such as shared subjects (S) or shared objects (O) through the given relationship. Assume that
two predicates are given as follows: P1 <Si, Oi> and P2 <Sj, Oj> where Si, Sj are a set of subjects and Oi, Oj are a set
of objects in given ontologies. The pattern describes that the same subject and object are shared by two predicates P1
and P2, the same subject shared, and the object shared.
Figure 1 Predicate Sharing Patterns
Figure 2 Predicate Connection Patterns
Connection Pattern: As shown in Figure 2, this pattern describes the relationships based mainly on the connectivity
of concept(s) through the respective predicates. This pattern is a frequently recurring pattern with predicates
observed during query processing as the basis for joining one query pattern to another. This type of pattern describes
the comprehension of the connectivity relationships between interacting predicates. Assume that two predicates are
given as follows: P1 <Si, Oi> and P2 <Sj, Oj> where Oi is equal to Sj and P1 is directly connected to P2 in the given
ontologies Oi, Oj. Since the connection pattern at level 1 will be modelled in Shared Patterns, Connection Patterns
are restricted to any patterns whose levels are greater than or equal to 2.
3
B. Predicate Neighboring Measurements
We define the measurement for the predicate neighboring patterns (PNP) in terms of sets of concepts and relations
(predicates) over the ontologies. For this purpose, we now describe how to quantify similarities between different
predicates based on the PNP pattern describing the relationships between predicates 𝑷𝒊 and 𝑷𝒋 through a concept 𝑪.
We formally define the similarity between predicates based on the shared patterns and connection patterns.
Definition 1: Given a directed graph 𝐺(𝐶, 𝑃), concepts 𝐶 denote subject S and object O and 𝑃 predicate in a RDF
schema graph, respectively. Let 𝑑(𝑃𝑖 , 𝑃𝑗 ) represent the number of concepts 𝐶 between 𝑃𝑖 and 𝑃𝑗 . 𝑟(𝑃𝑖 , 𝑃𝑗 ) determines
if a predicate 𝑃𝑖 is reachable from another predicate 𝑃𝑗 . 𝑙(𝑃𝑖 , 𝑃𝑗 ) indicates the shortest distance between 𝑃𝑖 and 𝑃𝑗 .
𝑙(𝑃𝑖 , 𝑃𝑗 ) =
0,
𝑃𝑖 = 𝑃𝑗
1,
𝑑(𝑃𝑖 , 𝑃𝑗 ) = 1
𝐿1 + 𝐿2 ,
𝐿1 = 𝑑(𝑃𝑖 , 𝑃𝑘 ) 𝐿2 = 𝑑(𝑃𝑘 , 𝑃𝑗 )
𝑟(𝑃𝑖 , 𝑃𝑘 ) = 𝑡𝑟𝑢𝑒, 𝑟(𝑃𝑘 , 𝑃𝑗 ) = 𝑡𝑟𝑢𝑒,
𝑟(𝑃𝑖 , 𝑃𝑗 ) = 𝑡𝑟𝑢𝑒
{
The similarity measurement for the PNP patterns varies based on different neighboring levels for each pair of
predicates. Basically, we gave a higher shared score to predicates with more shared concepts and lower scores to
predicates with less shared ones. Similarly, we gave a higher connection similarity score to closer predicates and
lower scores to further predicates. We now define these two probability based similarity scores: i) 𝑷𝑺𝒔 (𝑷𝒊 , 𝑷𝒋 ) is
defined as a shared pattern of any two predicates 𝑷𝒊 and 𝑷𝒋 ii) 𝑷𝑺𝒄 (𝑷𝒊 , 𝑷𝒋 ) for a connection pattern of any two
predicates.
Definition 2: Given predicates 𝑃𝑖 and 𝑃𝑗 in a directed RDF schema Graph 𝐺(𝐶, 𝑃). Let 𝐶(𝑃𝑖 ) and 𝐶(𝑃𝑗 ) denote the
entities (subjects or objects) that are directed connected to 𝑃𝑖 and 𝑃𝑗 regardless of the direction respectively.
𝑃𝑆𝑠 (𝑃𝑖 , 𝑃𝑗 ) indicates the probability-based similarity for a shared pattern between 𝑃𝑖 and 𝑃𝑗 .
1,
𝑙(𝑃𝑖 , 𝑃𝑗 ) = 0
𝑃𝑆𝑠 (𝑃𝑖 , 𝑃𝑗 ) =
0,
(|𝐶(𝑃𝑖 ) ∩ 𝐶(𝑃𝑗 )|)2
,
{ |𝐶(𝑃𝑖 )| ∗ |𝐶(𝑃𝑗 )|
𝑙(𝑃𝑖 , 𝑃𝑗 ) → ∞
𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
Definition 3: For a connection pattern of any two predicates 𝑃𝑖 and 𝑃𝑗 , 𝑃𝑆𝑐 (𝑃𝑖 , 𝑃𝑗 ) defines the probability-based
similarity for a connection pattern between 𝑃𝑖 and 𝑃𝑗 as follows:
𝑃𝑆𝑐 (𝑃𝑖 , 𝑃𝑗 ) = {
𝑃𝑆𝑠 (𝑃𝑖 , 𝑃𝑘 ) ∗ 𝑃𝑆𝑠 (𝑃𝑘 , 𝑃𝑗 ),
𝑙(𝑃𝑖 , 𝑃𝑗 ) = 2
max { 𝑃𝑆𝑐 (𝑃𝑖 , 𝑃𝑘 ) ∗ 𝑃𝑆𝑐 (𝑃𝑘 , 𝑃𝑗 )},
𝑖≤𝑘<𝑗
𝑙(𝑃𝑖 , 𝑃𝑗 ) > 2
The definition is influenced by the chain matrix multiplication problem (a kind of dynamic programming) that
involves the question of determining the optimal sequence for performing a series of operations. After we got the
similarity score for all pairs of predicates, we used formula in Definition 4 and 5 to generate a similarity matrix for
clustering.
Definition 4: Given the total number of predicate n and the probability-based similarity score for shared patterns
𝑃𝑆𝑐 (𝑃𝑖 , 𝑃𝑗 ) and connection patterns 𝑃𝑆𝑠 (𝑃𝑖 , 𝑃𝑗 ) between predicates 𝑃𝑖 and 𝑃𝑗 , SM[𝑃𝑖 , 𝑃𝑗 ] indicates a similarity matrix
for all pairs of predicates 𝑃𝑖 and 𝑃𝑗
SM[𝑃𝑖 , 𝑃𝑗 ] = {
𝑃𝑆𝑐 (𝑃𝑖 , 𝑃𝑗 ),
𝑙(𝑃𝑖 , 𝑃𝑗 ) ≥ 2
𝑃𝑆𝑠 (𝑃𝑖 , 𝑃𝑗 ),
𝑂𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
As shown in Figure 3, an example of the predicate similarity computation for shared patterns and connection
patterns was presented. In this example, a shared pattern is identified between predicates 𝑃1 and 𝑃2 and connection
patterns are identified between 𝑃1 and 𝑃3 , 𝑃1 and 𝑃4 , 𝑃1 and 𝑃5. Based on the PNP patterns, SM[𝑃𝑖 , 𝑃𝑗 ] is computed.
4
Figure 3 Predicate Neighboring Patterns (PNP) and Similarity Matrix
C. Hierarcical Predicate-based K-Means Clustering
The clustering approach we proposed here is based on the similarity measurement for the predicate neighboring
patterns (PNP) inherent in the ontologies. We posited that predicate-based clustering is a required step for efficient
query processing involving the alignment and integration of ontologies. Given that predicates are more closely related
to some predicates than others, predicates can be clustered for efficient query processing - the task of classifying a
collection of predicates into clusters (or topics). The guiding principle is to minimize inter-cluster (inter-topic)
similarity and maximize intra-cluster (intra-topic) similarity, based on the similarity measure for the PNP patterns.
We now present our clustering algorithm, called the Hierarchical Predicate-based K-Means clustering (HPKM)
that is designed by combining the divisive hierarchical clustering algorithm16 and K-Means algorithms17 for
generating K topics level-by-level in an optimal manner. Similar to the K-Means algorithm, the HPKM is an
unsupervised learning approach partitioning ontologies into K topics by clustering each predicate in the ontologies
with the nearest mean. Similar to the divisive hierarchical clustering algorithm16, the HPKM clusters ontologies into
smaller topics in a hierarchical manner. The PNP pattern-based similarity and the silhouette width (SW)18 were
computed for achieving the objective of the clustering which is maximizing intra-cluster similarities and minimizing
inter-cluster similarities19. If the SW of a topic is higher than α, this topic will be clustered into K smaller topics. The
value of silhouette 𝑠𝑤(𝑝𝑖 ) (i.e., silhouette width) can be ranged between -1 and 1. For each predicate 𝑝𝑖 , we computed
the following two similarity: inter-cluster similarity and intra-cluster similarity.
5
Intra-cluster similarity 𝒂(𝒑𝒊 ): This measure refers to the similarity of data in a single cluster. Let 𝒂(𝒑𝒊 ) be the
average dissimilarity of 𝒑𝒊 (taking the inverse of the SM matrix computed from the PNP algorithm) with all other
data within the same cluster. It can be validated how well 𝒑𝒊 is assigned to its cluster according to 𝒂(𝒑𝒊 ) such as the
smaller the value, the better the assignment. We then define the average dissimilarity of predicate 𝒑𝒊 to a cluster 𝑪 as
the average of the distance from 𝒑𝒊 to predicates in 𝑪𝒊 .
Inter-cluster similarity 𝒃(𝒑𝒊 ): This measure refers to the similarity between clusters. Let 𝒃(𝒑𝒊 ) be the lowest
average similarity of 𝒑𝒊 to the sibling clusters 𝑪𝒋 that has the same parent cluster with 𝑪𝒊 of which 𝒑𝒊 is not a
member. The cluster with this lowest average similarity is said to be the "sibling (neighboring) cluster", 𝑪𝒋 , of
𝒑𝒊 because it is the next best fit cluster for predicate 𝒑𝒊 .
A silhouette width can be computed as follows:
𝑠𝑤(𝑝𝑖 ) =
𝑏(𝑝𝑖 ) − 𝑎(𝑝𝑖 )
max{𝑎(𝑝𝑖 ), 𝑏(𝑝𝑖 )}
More specifically, it can be defined as follows: There are three possible cases about the silhouette width: (i) if the
silhouette width 𝑠𝑤(𝑝𝑖 ) is close to one, this means that the predicate 𝑝𝑖 is appropriately clustered. (ii) If 𝑠𝑤(𝑝𝑖 ) is
close to a negative one, then the predicate p would be not appropriate here but would be more appropriate if it was
clustered in its neighboring cluster 𝐶𝑗 . (iii) If 𝑠𝑤(𝑝𝑖 ) is near zero then this means that the predicate 𝑝𝑖 is on the
border of two natural clusters, namely 𝐶𝑖 𝑎𝑛𝑑 𝐶𝑗 .
1−
𝑠𝑤(𝑝𝑖 ) =
𝑎(𝑝𝑖 )
,
𝑏(𝑝𝑖 )
𝑖𝑓 𝑎(𝑝𝑖 ) < 𝑏(𝑝𝑖 )
0,
𝑖𝑓 𝑎(𝑝𝑖 ) = 𝑏(𝑝𝑖 )
)
𝑏(𝑝𝑖
− 1, 𝑖𝑓 𝑎(𝑝𝑖 ) > 𝑏(𝑝𝑖 )
{ 𝑎(𝑝𝑖 )
For each topic, we computed the average 𝑠𝑤(𝑝𝑖 ) over all data of a topic as a measure of how tightly grouped all
the predicates in the topic are. Thus the average 𝑠𝑤(𝑝𝑖 ) over all predicates of the entire dataset is a measure of how
appropriately the predicates have been clustered.
The average 𝑠𝑤(𝑝𝑖 ) over all predicates of each topic was computed at each level. For example, at level 1, K= 2
was computed using the SW. Furthermore, after partitioning into two topics, the silhouette widths, 0.89 (for 20
predicates) and 0.71 (for 43 predicates) are computed for each topic. At level 2, for the left topic, K= 5 and for the
right topic, K = 2 were computed, respectively. After clustering, silhouette widths, 0.52 (for 4 predicates) and 0.7
(for 6 predicates), 0.59 (for 3 predicates), 0.92 (for 4 predicates), and 0.38 (for 3 predicates) and two silhouette
widths, 0.76 (for 35 predicates) and 0.66 (for 8 predicates) were computed for each topic. At level 3, one of the
topics were partitioned into two (K= 2). Two silhouette widths, 0.77 (for 20 predicates) and 0.65 (for 15 predicates)
were computed for each topic. If there are too many or too few topics, as may occur when a poor choice of k in each
level is used in the hierarchical K-means algorithm, some of the topics will typically display much narrower
silhouettes than the rest. Thus silhouette averages are used to determine the number of topics within a dataset. We
increased the likelihood of the silhouette (α = 0.5) being maximized at the correct number of topics by re-scaling the
data using feature weights that are topic specific.
6
Figure 4 Silhouette Width and Number of Topics in Topic Hierarchy
In the HPKM, a topic of interest was further clustered into K subtopics (the optimal K subtopics) using a heuristic
algorithm, Neighborhood Silhouette Width (NSW). NSW is similar to the silhouette method that validates the
consistency checking by examining how well each predicate fits some uniformity criterion in its cluster, whereas
Neighborhood Silhouette Width (NSW) is the average of the weighted SW for the (neighbored) topics at a specific
level that have the same parents. The Neighborhood Silhouette Width (NSW) is computed by the sum of the
multiplication of silhouette width and the number of predicates in a particular topic, 𝑵𝒖𝒎𝑷(𝑻𝒊 ), divided by the total
number of predicates in the neighboring topics. The optimal k for a topic 𝑻𝒍 at level l will be determined based on
the highest Neighborhood Silhouette Width 𝒏𝒔𝒘(𝑻𝒍 )
∑𝒌𝒊=𝟏 𝒔𝒘(𝑻𝒊 ) ∗ 𝑵𝒖𝒎𝑷(𝑻𝒊 ))
𝒏𝒔𝒘(𝑻𝒍 ) =
∑𝒌𝒊=𝟏 𝑵𝒖𝒎𝑷(𝑻𝒊 ))
For example, as shown in Figure 4, for the given 𝒔𝒘(𝑻𝟏_𝟏 ) is 0.89 and 𝑵𝒖𝒎𝑷(𝑻𝟏_𝟏) is 20 and 𝒔𝒘(𝑻𝟏_𝟐 ) is 0.71
and 𝑵𝒖𝒎𝑷(𝑻𝟏𝟏 ) is 43, the first level’s Neighborhood Silhouette Width 𝒏𝒔𝒘(𝑻𝟏_𝟏 ) is computed as follows
𝒏𝒔𝒘(𝑻𝟏_𝟏 ) =
(𝟎.𝟖𝟗∗𝟐𝟎+𝟎.𝟕𝟏∗𝟒𝟑)
𝟐𝟎+𝟒𝟑
= 𝟎. 𝟕𝟕. Therefore, at level 1, the highest NSW value is 0.77 and the optimal K is
determined as 2. Similarly, the second level’s Neighborhood Silhouette Width 𝒏𝒔𝒘(𝑻𝟐_𝟏 ) is computed as followed:
(𝟎.𝟓𝟐∗𝟒+𝟎.𝟕∗𝟔+𝟎.𝟓𝟗∗𝟑+𝟎.𝟗𝟐∗𝟒+𝟎.𝟑𝟖∗𝟑)
(𝟎.𝟕𝟔∗𝟑𝟓+𝟎.𝟔𝟔∗𝟖)
𝒏𝒔𝒘(𝑻𝟐_𝟏 ) =
= 𝟎. 𝟔𝟒. and 𝒏𝒔𝒘(𝑻𝟐_𝟐 ) =
= 𝟎. 𝟕𝟒. Therefore, the
𝟒+𝟔+𝟑+𝟒+𝟑
𝟑𝟓+𝟖
highest NSW for 𝑻𝟐_𝟏 𝐚𝐧𝐝 𝑻𝟐_𝟐 at level 2 is (𝑻𝟐_𝟏 ) =0.64, (𝑻𝟐_𝟏 ) =0.74 and the optimal K is determined as 5 and 2,
respectively.
According to the optimal k determined by 𝒏𝒔𝒘(𝑻𝒍 ), the level of the hierarchy that can represent topics at
multiple tasks will be constructed at different levels until there is no further change in the hierarchy. If the silhouette
width of each
Algorithm 1 Hierarchical Predicate-based K-Means Clustering (HPKM)
// P is an n × n predicate similarity matrix, n is the number of predicates in ontologies
// δ is the threshold of silhouette width 𝐶𝑖𝑗
Input: P, δ
// a hierarchy with a set of clusters 𝐶𝑖𝑗 𝑡ℎ𝑒 𝑗𝑡ℎ 𝑐𝑙𝑢𝑠𝑡𝑒𝑟 𝑎𝑡 𝑖 𝑡ℎ level
Output: 𝐶 = {𝐶11 , 𝐶12 … 𝐶𝑖𝑗 }
1. i=1
2. repeat
3.
// n is the number of input predicates, m is the number of predicates of cluster j at level i,
4.
// optimal k (𝑘 ≤ m ≤ n) from m predicates of the cluster at level i (𝐶𝑖 ) using 𝑛𝑠𝑤(𝑐𝑖 ) function
5.
𝑠𝑤1 = 𝑛𝑠𝑤(𝑐𝑖 )
// compute Neighborhood silhouette width
7
6.
k= OptimalK(𝐶𝑖 , 𝑠𝑤1 )
// fine the optimal k based on 𝑛𝑠𝑤(𝐶𝑖 )
7.
𝐶ℎ𝑎𝑛𝑔𝑒1 = 𝑓𝑎𝑙𝑠𝑒
8.
If (k > 1) then
9.
𝐶ℎ𝑎𝑛𝑔𝑒1 = 𝑡𝑟𝑢𝑒
10.
for j = 1 to k
11.
𝜇𝑖𝑗 = RM(𝑝𝑗1 , 𝑝𝑗2 … 𝑝𝑗𝑚 )
// random mean for predicates in 𝐶𝑖𝑗 (cluster j at level i)
12.
end
13.
foreach 𝑝𝑖𝑗 ∈ 𝑃𝑖 𝑑𝑜
14.
𝜇𝑖𝑗 = Argmin(𝑝𝑖𝑗 , 𝜇𝑖𝑗 ) 𝑗 ∈ {1 … 𝑘}
15.
end
16.
𝐶ℎ𝑎𝑛𝑔𝑒2 = 𝑓𝑎𝑙𝑠𝑒
17.
sw = 0
18.
repeat
19.
foreach 𝜇𝑖𝑗 ∈ 𝑈𝑖 𝑑𝑜
20.
UpdateCluster(𝜇𝑖𝑗 )
21.
end
22.
foreach 𝑝𝑖𝑗 ∈ 𝑃𝑖 𝑑𝑜
23.
𝑁𝐶𝑒𝑛 = 𝐴𝑟𝑔𝑚𝑖𝑛(𝑝𝑖𝑗 , 𝜇𝑖𝑗 ) 𝑗 ∈ {1 … 𝑘}
24.
if 𝑁𝐶𝑒𝑛 ≠ 𝜇𝑖𝑗 then
25.
𝜇𝑖𝑗 = 𝑁𝐶𝑒𝑛
26.
𝐶𝑖𝑗 = 𝐶𝑖𝑗 ∪ 𝑝𝑖𝑗
27.
𝑐ℎ𝑎𝑛𝑔𝑒𝑑2 = 𝑡𝑟𝑢𝑒
28.
endif
29.
𝑠𝑤2 = SilhouetteWidth (𝐶𝑖𝑗 )
// compute silhouette width
30.
while 𝐶ℎ𝑎𝑛𝑔𝑒𝑑2 == 𝑡𝑟𝑢𝑒
31. while 𝐶ℎ𝑎𝑛𝑔𝑒1 == 𝑡𝑟𝑢𝑒 and 𝑠𝑤2 ≥ 𝛿
topic is lower than the threshold α, the clustering will be terminated. Thus, the maximum overall average silhouette
width will be taken as the optimal clustering algorithms for the topic hierarchy. In this way, we can achieve the
HPKM objective of maximizing intra-cluster similarities and minimizing inter-cluster similarities. The algorithm of
Hierarchical Predicate-based K-Means clustering (HPKM) is given as shown in Algorithm 1.
D. Topic Ranking in Topic Hierarchy
To characterize each topic in the hierarchy by an integrated rank, we computed the average value of the following
five classifications: i) Top 20 Predicates, ii) Top 20 Concepts, iii) Similarity, iv) Silhouette Width, and v) Density.
The first two rankings measure how popular they are relative to the rest of predicates and concepts. In determining
the rankings for Top 20 Predicates and Top 20 Concepts, the weight of a predicate or concept that occurs in
ontologies is simply proportional to the term frequency (about 20% and 30% were considered, respectively). For
Top 20 Concepts, as there may be some duplicates among topics, the duplicates are eliminated before deciding the
ranking. Similarity and Silhouette Width are measures for local (intra-relation of topics) and global (inter-relation of
topics) similarity, respectively. Both measurements seem to be equally important in reflecting the importance of
topics in a topic hierarchy. Similarity measurement was specified in Section II.B and silhouette width in Section
II.C.
In order to measure the Density, we used network concepts, such as in-degree and out-degree of concepts (C)
and predicates (P) in an ontology; in our model, C and P are vertexes, whereas V and the links between C and P are
edges, E in a graph. For a predicate (P), the number of incoming edges (E) adjacent to a concept (C) is called the indegree of the predicate (P) and the number of outgoing edges adjacent to a concept (C) is its out-degree of the
predicate (P). The density (D) is computed as a ratio of the number of edges |E| to the number of possible edges
𝟐|𝑬|
between nodes (|𝑽| = |𝑷| + |𝑪|) as follows: 𝑫 = (|𝑪|+|𝑷|)(|𝑪|−𝟏)(|𝑷|−𝟏). The results of the topic ranking were used in
query generation and query processing.
E. Query Generation
From the HPKM, a topic hierarchy is generated. The Query Generation algorithm will start crawling the leaf nodes
(the topics at the bottom level) in a given topic hierarchy and generate a query that is a part of a particular topic TG
8
(a RDF graph) in the topic hierarchy. The algorithm will crawl the topic graph TG to generate a query graph QG;
QG is a subset of the TG. Many variation of queries can be generated from this process. The query generation
algorithm automatically generates queries by traversing topic graphs. The topic graph has three predicates, namely
drug from the Sider domain (in pink), affected-organism from the DrugBank domain (in red) and x-pubchemsubstance from the Pharmacogenomics Knowledge Base (PharmGKB)20 (in green).
We generated a query by traversing the predicate that has the highest rank δ (the highest sum of the in-degree
and out-degree of the predicate) and traverse its neighbors level-by-level (Breadth-first Search) in the descending
order of the similarity in the SM computed by the PNP algorithm. For this traversal, we consider the neighbors
whose similarity scores are higher than threshold β. In this example, we started with the best predicate drug and then
visit its neighbors whose similarity scores are higher than the threshold (β = 0.2) in a descending order. For
example, drug, its nearest neighbor, x-pubchem-substance with the similarity score 0.5, thus we expand drug with an
additional predicate, x-pubchem-substance. And then drug’s next nearest neighbor is affected-organism with the
similarity score 0.1. Since the similarity score is less than the threshold β, (i.e., 0.1≤ 0.2), we terminated the
navigation. The algorithm runs until there is no more neighboring predicates to be considered. The generated query
includes triples with two predicates, drug and x-pubchem-substance, and their subject variables (?E and ?D) and
object variables (?D and ?R). The type of variable ?E is known as Drug Effect, ?D as Drug, and ?R as PharmGKB
Resource. This can be converted to a triplet form such as <?D typeof Drug>. As shown in Figure 5, an example of
the automatically was presented to demonstrate the generation of a SPARQL query21 for a given topic graph.
Figure 5 Automatic Query Generation
III.
IMPLEMENTATION
The MedTQ system was implemented using Java in Eclipse Juno Integrated Development Environment22. Apache
Jena API23 was used to parse OWL/RDF datasets and retrieve triple information. We used R computing
environment24 for our experimental validation. We implemented a software plugin for query and schema graph
visualization using CytoScape 3.0.225. In addition, we have built a SPARQL query endpoint on a single machine that
is hosted by the UMKC Distributed Intelligent Computing (UDIC) lab. (Figure 6) The OPEN LINK Virtuoso server
version 6.1.326 was installed and the five Bio2RDF datasets (Bio2RDF ClinicalTrial27, Bio2RDF DrugBank28,
Bio2RDF OMIM29, Bio2RDF PharmGKB30, and Bio2RDF Sider31) were imported into the graph domain
http://Bio2RDF.com#.
9
Figure 6 SPARQL Endpoint
Figure 7 MedTQ Interactive Query Tool
The MedTQ tool can be used for browsing the generated topics and performing interactive design and processing
of queries. As shown in Figure 7, step 1 shows the list of topics for a given ontology (DrugBank). Step 2 shows the
list of NLP questions for a selected topic (Topic 7). Step 3 shows the automatically generated SPARQL query and
the query results. Step 4 shows the topic and query graphs for the selected query.
The steps for the query generation and processing using MedTQ tool are explained as follows:
Step1: A user first selects a dataset (e.g., DrugBank) to be analyzed, then choose an algorithm to generate a topic
hierarchy (e.g., three level hierarchy). A clustering algorithm (e.g., Hierarchical K-Means Clustering) button is
selected for the construction of a topic hierarchy (DrugBank). Topics generated from the topic hierarchy
construction are listed in the top left box. In this example, the eight topics are shown with the detailed description
including a list of the highest ranked predicates and their concepts (with high in-degree/out-degree).
Step2: The user selects a topic (e.g., 7th topic) to view, then this allows users to explore top ten natural language
queries automatically generated by the proposed query generation algorithm.
Step3: A query can be selected and modified through the interactive query editor based on the topics or predicates
shown in Step 2. Once the design of a query is complete, the corresponding natural language query expressions and
the corresponding SPARQL query will be generated.
Step4: After choosing the natural language query expressions (e.g., what are the enzyme, target and transporterrelation of a drug?), the add query button can be clicked to select its corresponding SPARQL query into the bottom
left box.
10
Step5: When the query button is clicked, the SPARQL query will be executed and the query output will be shown in
the bottom right box.
Step6: When the show query cluster button is clicked, the corresponding cluster graph will be displayed on the
canvas in the right panel. Moreover, by clicking the show query graph button, the relevant concepts and predicates
in the SPARQL query will also be highlighted.
Figure 8 DrugBank Topic Hierarchy
IV.
RESULTS
In this paper, we conducted a case study with a biomedical ontology (DrugBank). DrugBank is a public database
with drug information on properties, structure, and biology of small molecule and biotech drugs. Important topics
are drugs’ targets, enzymes, transporters, and carriers. These may switch roles depending on the drug to which they
bind so that some drugs specifically target transporters i.e., a transporter can also be the target. It is a key resource
for bioinformatics and cheminformatics research. Based on DrugBank database and other life science data,
Bio2RDF project32 normalized them into a uniform format in a distributed network to support biomedical
knowledge translation and discovery.
A. Topic Hierarchy Generated using HPKM Approach
In this case study, we demonstrated the details of knowledge discovery as well as query generation in the proposed
framework. We are particularly interested in generating interesting queries using the proposed PNP model and
HPKM algorithms. In addition, the experiments have been conducted to validate the correctness of our approach.
Table 1 shows the details of the DrugBank Ontology. In this case study, the unique concepts (C) of DrugBank
ontology, excluding the duplicates, are considered. Only the domain specific predicates (P) excluding built-in
11
predicates were considered. The number of edges in the graph (|E|) was computed as the sum of in-degree and outdegree. The overall density was computed based on the vertices (P+C) and the edges (E).
Table 1. DrugBank Ontology
Features
# Total Concepts
#Unique Concepts in DrugBank (C)
# Total Predicates
#Unique Domain Specific Predicate (P)
Num
116
93
68
63
Features
Sum of Indegree and Outdegree (|E|)
#Triples
# Domain Specific Triplets (T)
Density (D)
Num
519
737
401
0.043
The base URL of predicates is http://bio2rdf.org/drugbank_vocabulary. However, the concepts are from 25
different domains as shown in Table 2. Interestingly, all predicates are from the same domain and that gives us a
good basis for linking concepts together either from same or different domains. This is one of the reasons we
proposed a predicate-oriented approach. The concepts’ domain URLs and their short notations are shown in Table
2.
Table 2. Short Notation for DrugBank and Related Domains
prefix
ahv:
av:
bv:
cv:
dv:
dpv:
gv:
gav:
gcv:
giv:
gtv:
hv:
iv:
Domain URL
http://bio2rdf.org/ahfs_vocabulary:
http://bio2rdf.org/atc_vocabulary:
http://bio2rdf.org/bindingdb_vocabulary:
http://bio2rdf.org/chemspider_vocabulary:
http://bio2rdf.org/drugbank_vocabulary:
http://bio2rdf.org/dpd_vocabulary:
http://bio2rdf.org/genbank_vocabulary:
http://bio2rdf.org/genatlas_vocabulary:
http://bio2rdf.org/genecards_vocabulary:
http://bio2rdf.org/gi_vocabulary:
http://bio2rdf.org/gtp_vocabulary:
http://bio2rdf.org/hgnc_vocabulary:
http://bio2rdf.org/iuphar_vocabulary:
prefix
kv:
owl:
pcv:
pdv:
psv:
pv:
uv:
chv:
nv:
phv:
unv:
wv:
Domain URL
http://bio2rdf.org/kegg_vocabulary:
http://www.w3.org/2002/07/owl#
http://bio2rdf.org/pubchem.compound_vocabulary:
http://bio2rdf.org/pdb_vocabulary:
http://bio2rdf.org/pubchem.substance_vocabulary:
http://bio2rdf.org/pubmed_vocabulary:
http://bio2rdf.org/uspto_vocabulary:
http://bio2rdf.org/chebi_vocabulary:
http://bio2rdf.org/ndc_vocabulary:
http://bio2rdf.org/pharmgkb_vocabulary:
http://bio2rdf.org/uniprot_vocabulary:
http://bio2rdf.org/wikipedia_vocabulary:
From the HPKM algorithm for each domain ontology, the topic hierarchy was generated. A topic hierarchy was
generated for a single domain ontology, DrugBank. The topic hierarchy for DrugBank has the number of topics
<2:7:8> with 2 topics at the first level, 7 topics at the second level, and 8 topics at the third level. K-means
clustering was performed in a top-down manner until the average of clusters’ silhouette width is higher than a
certain threshold (> 0.5). The number on each edge in the topic hierarchy represents the percentage of predicates that
the upper level topic graph contributes to the lower level graph. For example, for the two topics in the first level of
DrugBank, 66% of predicates of the DrugBank ontology are contributed to Topic 1 (T1_1) while 34% to Topic 2
(T1_2). The contribution rate is ranged between 0 and 1. Interestingly, predicates are unique to their topic graph,
however, some concepts in a topic may appear in more than one topics. Moreover, for each topic at 3rd level, top 2
ranked predicates (computed based on in-degree/out-degree) were selected as a representative term for each topic.
B. Top Predicates and Concepts of DrugBank
Table 3 shows the ranks for Top 20 predicates and Top 20 concepts that were computed in terms of the sum of their
in-degree and out-degree. These predicates and concepts are shown in terms of Predicate Rank (PR), Predicates,
Predicate IO (PIO) and Predicate Topic ID (PIO), corresponding Concepts, Concept Rank (CR), and together with
the description of predicates specified by DrugBank. From this list, many top predicates are from Topic 3_6, Topic
3_7, and Topic 3_1. Many top concepts are from Topic 3_7, Topic 3_2, and Topic 3_3. The prefix dv: of these
concepts indicates the domain http://bio2rdf.org/drugbank_vocabulary. Some of the Top 20 Concepts were not
directly mapped with the predicates in the Top 20 Predicates. These concepts are dv:Enzyme-Relation, dv:TargetRelation, dv:Carrier-Relation, dv:LogP, dv:LogS, dv:Molecular-Formula, dv:Molecular-Weight, dv:TransporterRelation, dv:Water-Solubility, dv:Bioavailability, dv:Boiling-Point, dv:Caco2-Permeability. These results show that
the predicates rankings are not always the same with the concept rankings.
Table 3. Top 20 Predicates and Concepts in DrugBank Ontology
PR
Predicates
PIO
TID
CR
Concepts
Description
12
1
source
66
T3_6
2
calculatedproperties
experimental
-properties
56
T3_6
dv:Source
N/A
28
T3_6
N/A
4
transporter
17
T3_7
7
dv:Transporter
5
target
16
T3_7
6
dv:Target
6
drug
14
T3_4
2
dv:Drug
7
enzyme
13
T3_7
5
dv:Enzyme
8
carrier
12
T3_7
4
dv:Carrier
9
action
11
T3_4
10
synonym
10
T3_1
21
dv:Synonym
11
brand
9
T3_1
12
category
8
T3_1
51
21
dv:Brand
dv:Category
13
form
8
T3_8
N/A
14
ingredient
8
T3_8
N/A
Dosage forms are essentially pharmaceutical drug
products in the form in which they are marketed for use.
An ingredient is a substance that forms part of a mixture
15
x-genbank
7
T3_7
N/A
Link between DrugBank and GenBank
16
x-uniprot
7
T3_7
N/A
Link between DrugBank and UniProt
17
manufacturer
6
T3_1
18
mixture
6
T3_1
19
toxicity
6
T3_1
20
absorption
6
T3_2
3
46
N/A
51
dv:Manufacturer
45
dv:Mixture
51
dv:Toxicity
51
dv:Absorption
Data sources indicate the source of the information
present in each drug properties.
The calculation from the structure for the drug.
A medicinal product may be approved for use in one
disease or condition but still be considered experimental
for other diseases or conditions.
A membrane bound protein which shuttles ions, small
molecules or macromolecules across membranes, into
cells or out of cells.
A protein, macromolecule, nucleic acid, or small
molecule to which a given drug binds, resulting in an
alteration of the normal function of the bound molecule
and a desirable therapeutic effect.
A drug is a chemical substance that, when ingested, has a
high biological response to quantity ratio compared to
regular foods.
A protein which catalyzes chemical reactions involving a
given drug (substrate).
Drug carriers may be used in drug design to increase the
effectiveness of drug delivery to the target sites of
pharmacological actions.
The drugs that enter the human tend to stimulate certain
receptors, ion channels, act on enzymes or transporter
proteins. As a result, they cause the human body to react
in a specific way.
Alternate names of the drug
Brand names from different manufacturers
Category of drugs
Companies known to manufacture the given drug
A mixture refers to the physical combination of two or
more substances on which the identities are retained and
are mixed in the form of solutions, suspensions, and
colloids.
Lethal dose (LD50) values from test animals, description
of side effects and toxic effects seen in humans
Description of how much of the drug or how readily the
drug is taken up by the body
Table 4 shows the duplicated concepts among topics. The total number of instances is 40 and the number of
duplicates is 23. dv:Resource and dv:Drug appear in almost all the topics. According to this analysis, the sets of the
topic groups {T3_1 and T3_8}, {T3_4 and T3_7}, and {T3_5 and T3_7} are similar. However, these are quite
different from the outcomes from the predicated-oriented clustering algorithm.
Table 4. Duplicated Concepts and their Topic ID in DrugBank Ontology
Concepts
Freq
Topics
Concepts
Freq
Topics
dv:TransporterRelation
2
T3_4, T3_7
dv:Enzyme-Relation
2
T3_4, T3_7
dv:Carrier
2
T3_5, T3_7
dv:Resource
8
dv:Drug
6
T3_1, T3_2, T3_3, T3_4, T3_5,
T3_6, T3_7, T3_8
T3_1, T3_2, T3_3, T3_4, T3_6,
T3_7
uv:Resource
2
T3_1, T3_8
13
dv:Mixture
2
T3_1, T3_8
dv:Target
2
T3_5, T3_7
dv:Patent
2
T3_1, T3_8
dv:Enzyme
2
T3_5, T3_7
dv:Pharmaceutical
2
T3_2, T3_8
dv:Transporter
2
T3_5, T3_7
dv:Carrier-Relation
2
T3_4, T3_7
uv:Resource
2
T3_1, T3_8
dv:Target-Relation
2
T3_4, T3_7
C. Validation for Hierarchical K-Means Clustering
An experiment has been conducted to find an optimal number of the clusters using the four different clustering
algorithms, K-means17, Clustering Large Application (Clara)33, Partition Around Medoids (Pam)34, and Hierarchical
Clustering16. The results of the optimal K validation algorithm presented in Section II.C based on the clustering
outcomes by the four different algorithms. As shown in Figure 9, Clara, Pam and Hierarchical clustering algorithms
are not a good approach to find an optimal cluster number since they show a relative stable silhouette width for
varying the number of clusters. The proposed HPKM algorithm determines the most significant number of clusters
at each level such as K = 2 with SW = 0.77 at level 1 and K = 5 with SW = 0.64 and K=2 with SW = 0.74 at level 2
and K = 2 with SW = 0.72 at level 3. The HPKM algorithm was validated and compared against other algorithms in
terms of the cluster number and the silhouette width.
Figure 9 Optimal K Branching Factors using Multiple Clustering Techniques
D. Results for Topic Generation
For the DrugBank ontology, we’ve considered 63 concepts, 116 predicates. The relevance scales of five different
rankings and an overall ranking are evaluated. As shown in Figure 10, overall rank was computed in terms of the
following 5 criteria: i) Top 20 Concepts, ii) Top 20 Predicates, iii) Similarity, iv) Silhouette Width, v) 𝐷𝑒𝑛𝑠𝑖𝑡𝑦.
Eight topics ranked from best to worst as follows: Topic 3_4, Topic 3_7, Topic 3_6, Topic 3_2, Topic 3_1, Topic
3_3, Topic 3_8, and Topic 3_5. Specifically, Topic 3_4 shows the best ranking for all three criteria such as
Similarity, Silhouette Width and Density. However, Topic 3_7’s Top 20 Concept Ranking, Top 20 Property Ranking,
and Similarity Ranking are relatively good. From the ranking results, we have observed that the proposed ranking
system correctly captured Topic 3_4 and Topic 3_7 as the core topics of DrugBank. Topic 3_5 was ranked the worst
among the eight topics. Since Topic 3_5 is a connector topic whose predicates are mainly used to connect DrugBank
14
with other domains. It is relatively less important from a single domain (DrugBank) perspective. However, Topic
3_5 would be very useful from a cross domain perspective.
Figure 10 Topic Rankings for DrugBank
E. Results for Topic Discovery & Query Generation
We now show the four best topic graphs at level 3 of the DrugBank topic hierarchy as follows: Topic 3_4, Topic
3_7, Topic 3_6, and Topic 3_2. In addition, the automatically generated query and query results of each topic are
also shown.
Rank 1: Topic 3_4 (T3_4): T3_4’s overall rank is 1st among 8 topics. This topic graph consists of 6 predicates and
12 concepts with 72 in-degree and out-degree. As shown in Figure 11, among 12 concepts, five concepts
(dv:Resource; dv:Drug, dv:Carrier, dv:Enzyme, dv:Target) are ranked among Top 20 Concepts and all 6 predicates
of this topic are ranked among Top 20 Predicates. In particular, there are two groups of predicates; one is with four
predicates such as transporter, target, enzyme, carrier with concepts dv:Target-Relation, dv:Target-Relation,
dv:Enzyme-Relation, dv:Carrier-Relation, respectively. Another group of predicates such as x-genbank and xuniprot is a connector predicate group that is mainly used to connect between internal concepts (e.g., dv:Drug,
dv:Enzyme) and external concepts (e.g., gv:Resource, unv:Resource). Specifically, T3_4 shows very high rankings
for Similarity, Silhouette Width, and Density while showing a relatively low ranking for Top 20 Concepts. In this
graph, concepts are represented as a circle, predicates as a triangle, and links as an arrow. In addition, the dark red
items are the predicates and concepts mentioned in Query-1.
15
Figure 11 Topic_3_4 (Topic 4 at Level 3) Graph in DrugBank
Query-1: The query is automatically generated from Topic 3_4 (one of the top ranked topics) by our query
generation algorithm. This query allows users to find the most relevant drugs in terms of their target, enzyme,
enzyme relation, and target relation. The SPARQL format of Query-1 was automatically generated by the Query
Generation algorithm (described in Section II-E) considering the top predicates and their concepts (described in
Section IV-B) below.
Q1. For any two drugs which share the same target and transporter enzyme, what are all the possible drugs,
enzyme, target, enzyme relations, target relations?
select distinct ?druglabel, ?targetlabel, ?erlabel, ?trlabel, ?drug2label, ?enzymelabel where {
?drug <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Drug> .
?target <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Target> .
?drug <http://bio2rdf.org/drugbank_vocabulary:target> ?target .
?drug <http://bio2rdf.org/drugbank_vocabulary:transporter> ?enzyme .
?er <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Enzyme-Relation> .
?tr <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Target-Relation> .
?er <http://bio2rdf.org/drugbank_vocabulary:enzyme> ?enzyme .
?tr <http://bio2rdf.org/drugbank_vocabulary:enzyme> ?target.
?drug2 <http://bio2rdf.org/drugbank_vocabulary:target> ?target .
?drug2 <http://bio2rdf.org/drugbank_vocabulary:transporter> ?enzyme .
?drug <http://www.w3.org/2000/01/rdf-schema#label> ?druglabel.
?target <http://www.w3.org/2000/01/rdf-schema#label> ?targetlabel.
?er <http://www.w3.org/2000/01/rdf-schema#label> ?erlabel.
?tr <http://www.w3.org/2000/01/rdf-schema#label> ?trlabel.
?drug2 <http://www.w3.org/2000/01/rdf-schema#label> ?drug2label.
?enzyme <http://www.w3.org/2000/01/rdf-schema#label> ?enzymelabel. }
The Query-1 results include Gemcitabine, Fluorouracil, Ribavirin as the relevant drugs, Thymidylate synthase
and Adenosine kinaseas as the target and Equilibrative nucleoside transporter 1 as the enzyme. Specifically, as
shown in Figure 12, partial outputs of Query-1 include information on drug target and transporter enzyme of some
drugs (e.g., Gemcitabine [drugbank:DB00441], Fluorouracil [drugbank:DB00544], Ribavirin [drugbank:DB00811])
and their enzyme relations, target relations for any drugs that share the same target and transporter enzyme.
16
Figure 12 Partial Results of Query-1 in DrugBank
Rank 2: Topic 3_7 (T3_7): T3_7’s overall rank is 2nd among 8 topics (together with T3_4). As shown in Figure 13,
this topic graph is composed of 7 concepts represented as a circle and 3 predicates as a triangle with 31 in-degree
and out-degree. In T3_7, three predicates, drug, action, reference, whose in-degree and out-degree are 14, 11, and 6,
respectively, are all nicely connected with 7 concepts. The predicates drug and action are ranked at 6th, 9th and many
of the concepts in this topic are ranked among Top 20 Concepts. For T3_7, the rankings for Top 20 Concepts, Top
20 Predicates, Similarity, Silhouette Width, and Density are very good. In this graph, concepts are represented as a
circle, predicates as a triangle, and links as an arrow. In addition, the dark red items like drug and reference are the
predicates mentioned in Query-2.
Figure 13 Topic 3_7 (Topic 7 at Level 3) Graph in DrugBank
Query-2: The query graph was automatically generated from Topic 3_4 (one of the top ranked topics) to depict the
query information. This query allows users to find the relevant drugs that have common Target-Relation, CarrierRelation and Transporter-Relation and also provide their PubMed references for relations of target, transporter, and
carrier with these drugs. The SPARQL format of Query-2 was automatically generated by the Query Generation
algorithm (described in Section II-E) considering the top predicates and their concepts (described in Section IV-B)
below.
Q2. For any two drugs which share the common target-relation, carrier-relation and transporter-relation, what
are all the possible combinations? What are the pubmed references for these target-relations, carrier-relation and
transporter-relation?
17
select distinct ?druglabel, ?drug2label, ?targetlabel,?carrierlabel,?transporterlabel,?reference1,?reference2,?reference3 where
{
?drug <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Drug> .
?drug2 <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Drug> .
?target <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Target-Relation> .
?carrier <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Carrier-Relation> .
?transporter <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:TransporterRelation> .
?target <http://bio2rdf.org/drugbank_vocabulary:drug> ?drug .
Optional {?target <http://bio2rdf.org/drugbank_vocabulary:drug> ?drug2} .
?carrier <http://bio2rdf.org/drugbank_vocabulary:drug> ?drug .
Optional { ?carrier <http://bio2rdf.org/drugbank_vocabulary:drug> ?drug2 } .
?transporter <http://bio2rdf.org/drugbank_vocabulary:drug> ?drug .
Optional { ?transporter <http://bio2rdf.org/drugbank_vocabulary:drug> ?drug2 }.
?target <http://bio2rdf.org/drugbank_vocabulary:reference> ?reference1.
?carrier <http://bio2rdf.org/drugbank_vocabulary:reference> ?reference2.
?transporter <http://bio2rdf.org/drugbank_vocabulary:reference> ?reference3.
?drug <http://www.w3.org/2000/01/rdf-schema#label> ?druglabel.
?drug2 <http://www.w3.org/2000/01/rdf-schema#label> ?drug2label.
?target <http://www.w3.org/2000/01/rdf-schema#label> ?targetlabel.
?carrier <http://www.w3.org/2000/01/rdf-schema#label> ?carrierlabel.
?transporter <http://www.w3.org/2000/01/rdf-schema#label> ?transporterlabel.
}
As shown in Figure 14, partial results from the query include information about some drugs such as Phenytoin
(DrugBank:DB00252), Lepirudin (DrugBank:DB00001) and Deferasirox (DrugBank:DB01609). The results show
all the possible combinations and their PubMed references for Target-relation, Carrier-relation, and Transporterrelation of drugs sharing the common information of Target-relation, Carrier-relation, and Transporter-relation.
Figure 14 Results of Query-2 in DrugBank
Rank 3: Topic 3_6 (T3_6): As shown in Figure 15, this topic consists of 3 predicates and 34 concepts with a high
sum of in-degree and out-degree, 150. In particular, there are two subgraphs; one is with two predicates such as
source and calculated-properties with concepts dv:Boiling-Point and dv:Bioavailability, respectively. Another
predicate experimental-properties is connected with concepts such as dv:Water-Solubility. Specifically, T3_6 highly
ranked in Top 20 Predicates and Silhouette Width while being lowly ranked in Similarity. This means each
predicate has their own concepts while having the least common concepts with other predicates. Since the similarity
ranking of this topic is low, the shared information is limited. Interestingly, this graph shows a connection pattern
from dv:experimental-properties to dv: source. The overall rank is 3rd among the eight topics. In this graph, the dark
red items like source and calculated-properties are the predicates mentioned in Query-3.
18
Figure 15 Graph of Query-3 in DrugBank
Query-3. The query graph was automatically generated from Topic 3_6 to depict the query information. This query
allows users to find drugs and all their experimental properties and calculated properties which have LogP
experimental properties (octanol-water partition coefficient). The SPARQL format of Query-3 was automatically
generated by the Query Generation algorithm (described in Section II-E) considering the top predicates and their
concepts (described in Section IV-B) below.
Q3. For any drug, what are all its experimental properties and calculated properties which contain octanol-water
partition coefficient?
select distinct ?druglabel,?logp1label,?logp2label,?source1label,?source2label where {
?drug <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Drug> .
?logp1 <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:LogP> .
?logp2 <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:LogP> .
?drug <http://bio2rdf.org/drugbank_vocabulary:experimental-properties> ?logp1 .
?drug <http://bio2rdf.org/drugbank_vocabulary:calculated-properties> ?logp2 .
?logp1 <http://bio2rdf.org/drugbank_vocabulary:source> ?source1 .
?logp2 <http://bio2rdf.org/drugbank_vocabulary:source> ?source2 .
?drug <http://www.w3.org/2000/01/rdf-schema#label> ?druglabel.
?logp1 <http://www.w3.org/2000/01/rdf-schema#label> ?logp1label.
?logp2 <http://www.w3.org/2000/01/rdf-schema#label> ?logp2label.
?source1 <http://www.w3.org/2000/01/rdf-schema#label> ?source1label.
?source2 <http://www.w3.org/2000/01/rdf-schema#label> ?source2label.
}
As the Query-3 results shown in Figure 16, the relevant drug and their experimental and calculated-properties are
reported as {L-Histidine, logP: -3.32 from CHMELIK,J ET AL. (1991), logP: -3.1 from ALOGPS} and LPhenylalanine, logP: -1.38 from AVDEEF,A (1997), logP: -1.4 from ALOGPS}. The partial query results on Query3 include the information on some drugs (e.g., L-Histidine [drugbank:DB00117], L-Phenylalanine
[drugbank:DB00120], L-Arginine [drugbank:DB00125]) and all their experimental properties and calculated
properties that have LogP experimental properties (octanol-water partition coefficient).
19
Figure 16 Results of Query-3 in DrugBank
Rank 4: Topic 3_2 (T3_2): As shown in Figure 17, unique pattern in T3_2 is two dominant concepts, Resource and
Drug, whose in-degree and out-degree are 46 and 33, respectively, are fully connected to the remaining 17 concepts
via 20 different predicates such as absorption, protein-binding. The rankings for Top 20 Concepts and Silhouette
Width are relatively good while the rankings for Top 20 Predicates and Density are poor. The overall rank is 4th
among the eight topics. In this graph, concepts are represented as a circle, predicates as a triangle, and links as an
arrow. The dark red items like abortion and product are the predicates and dv:Drug and dv:Pharmaceutical are the
concepts mentioned in Query-4.
Figure 17 Topic 3_2 (Topic 2 at Level 3) Graph in DrugBank
Query-4: The query graph was automatically generated from Topic 3_2 to depict the topic and query information.
This query allows users to find drugs and their absorption, affected-organism, clearance pharmacokinetic
measurement, pharmaceutical information, and protein binding information. The SPARQL format of Query-4 was
automatically generated by the Query Generation algorithm (described in Section II-E) considering the top
predicates and their concepts (described in Section IV-B) below.
Q4. For any two drugs which share the same absorption, affected-organism, clearance and pharmaceutical, what
are all the possible combinations?
20
select distinct ?druglabel, ?drug2label, ?absorptionlabel, ?aolabel, ?clearancelabel, ?pharmaceuticallabel, ?pblabel where {
?drug <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Drug> .
?drug2 <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://bio2rdf.org/drugbank_vocabulary:Drug> .
?drug <http://bio2rdf.org/drugbank_vocabulary:absorption> ?absorption .
Optional {?drug2 <http://bio2rdf.org/drugbank_vocabulary:absorption> ?absorption} .
?drug <http://bio2rdf.org/drugbank_vocabulary:affected-organism> ?ao .
?drug <http://bio2rdf.org/drugbank_vocabulary:clearance> ?clearance .
Optional {?drug2 <http://bio2rdf.org/drugbank_vocabulary:clearance> ?clearance} .
?drug <http://bio2rdf.org/drugbank_vocabulary:product> ?pharmaceutical .
Optional {?drug2 <http://bio2rdf.org/drugbank_vocabulary:product> ?pharmaceutical } .
?drug <http://bio2rdf.org/drugbank_vocabulary:protein-binding> ?pb .
Optional {?drug2 <http://bio2rdf.org/drugbank_vocabulary:protein-binding> ?pb}
?drug <http://www.w3.org/2000/01/rdf-schema#label> ?druglabel.
?drug2 <http://www.w3.org/2000/01/rdf-schema#label> ?drug2label.
?absorption <http://www.w3.org/2000/01/rdf-schema#label> ?absorptionlabel.
?ao <http://www.w3.org/2000/01/rdf-schema#label> ?aolabel.
?clearance <http://www.w3.org/2000/01/rdf-schema#label> ?clearancelabel.
?pharmaceutical <http://www.w3.org/2000/01/rdf-schema#label> ?pharmaceuticallabel.
?pb <http://www.w3.org/2000/01/rdf-schema#label> ?pblabel.
}
The Query-4 results on any relevant drugs and their pharmaceutical information include {Gemcitabine, Lepirudin,
Gemzar 1 gm Solution Vial}, {Tiotropium, Lepirudin, Spiriva 18 mcg Capsule}. As shown in Figure 18, partial
query results on Query-4 include the information on some drugs (e.g., Gemcitabine[drugbank:DB00441],
Tiotropium[drugbank:DB01409], Carvedilol[drugbank:DB1136]) having the common information of their
absorption, affected-organism, clearance pharmacokinetic measurement, pharmaceutical information, and protein
binding information.
Figure 18 Partial Results of Query-4 in DrugBank
V.
DISCUSSION
A. Knowledge Disocvery from General and Medical Ontologies
In general settings, many efforts have been made to perform knowledge discovery with semantic web techniques.
For instance, the SMARTSPACE proposed a distributed platform for semantic knowledge discovery from services
using multi-agent approach35. The PEMAR introduced a smart phone middleware for activity recognition discovery
based on semantic models36. A mobile-cloud computing framework was established to discover infrastructure
condition based on a back-end semantic knowledge discovery engine37. In our previous work, we have built a
situation aware mobile application framework38, 39 to discovery users’ activities in a dynamic way based on the
semantic web rule language (SWRL)40.
In biomedical settings, the advance in Linked Data technologies, such as the standard graph data model (RDF)
and distributed SPARQL capability, allows us to easily access distributed data. Most of the work has mainly focused
on building or using ontologies for data normalization, bridging and reasoning. Widely used medical ontologies are
Bio2RDF32, TMO (Translational Medicine Ontology)41, Chem2Bio2RDF42, SIO (Semanticscience Integrated
Ontology)43, ATC (Anatomical Therapeutic Chemical) and DrugBank integration44, Linked Life Data45.
21
Drug discovery research heavily relies on multiple information sources to validate potential drug candidates as
shown in the Open PHACTS project46. In complicated domains, it takes time to develop and maintain ontologies47-50.
There have been various studies on using semantic techniques to improve data integration and share information.
DrugBank is one of the key resources which provide bioinformatics and cheminformatics studies with complete
information on drug and drug targets. However, these efforts merely support physical integration of multiple
biomedical ontologies without considering semantic integration of data. In particular, human intervention is strongly
required so that these are not suitable for comprehensive and accurate knowledge discovery especially from a large
amount of data. Furthermore, semantic interoperability is difficult to achieve in these systems as the conceptual
models underlying datasets are not fully exploited.
Pattern based knowledge analysis has been conducted in many aspects of biomedical research. Warrender and
Lord proposed an axiom based generalized and localized pattern driven approach in biomedical ontology
engineering51. Want et al., designed a biomedical pattern discovery algorithm based on a supervised learning
approach52. Rafiq et al., developed an algorithm to discover temporal patterns in genomic databases53. However, our
approach is different in that, first of all, we focused on a more general approach for graph structural pattern analysis
and discovery. In addition, we have combined an unsupervised learning algorithm with a pattern discovery
technique to provide a more dynamic way of knowledge discovery from large amount of ontologies.
Our work is motivated by previous work on highlighting the importance of ontological relations54-56. Tartir et al.,
pointed out that there are numerous meaningful relations other than class-subclass relations that would be useful for
understanding the ontologies57. Sabou et al., considered ontological relations to be the primary criterion for the
summary extraction of ontologies, in which a relatively small number of concepts typically have a high degree of
connectivity through hops58. In our study, we hypothesized that a similarity measurement based on predicate
neighborhood patterns would be more effective in finding relevant information than a concept-based measurement.
Our approach defined a new concept of predicate based patterns and neighboring closeness for an automatic
knowledge discovery and query generation system.
In this paper, we applied k-means on predicate similarity matrix to group n predicates into k different clusters in
order to form various topics. This approach is unique in terms of three aspects: i) PNP-based similarity measure, ii) a
hierarchical approach, iii) level-by-level optimality using a silhouette heuristic function and density function. We
utilize the intrinsic property of predicates’ neighborhoods, the strong dependence of predicates and their resources
with high similarity inside a topic cluster is reflected in their co-occurrence in both their own neighborhoods and the
neighborhoods of predicates close-by. It was found that the outcomes from the HPKM algorithm preserve the local
neighborhoods of predicates inside topics and shows a high similarity inside a topic cluster (as seen the topic
hierarchy in Section IV).
B. Query Generation and Query Processing
There are several works on query generation and processing for ontologies, such as59-61. Queries can often be
difficult to formulate across these datasets62. In particular, the work Lorey et al., proposed59 has a similar approach
to our work in terms of detecting recurring query patterns based on the distance among RDF graph patterns and
identifying query templates from the analysis of the RDF graph structure. However, this work focuses more on
concepts of the instance level of RDF graphs for the pattern identification and template extraction. Unlike this work,
we focused on a new paradigm, such as predicate based similarity patterns, at the schema level for topic discovery
and query suggestion.
Biomedical data contributors have provided public SPARQL endpoints to query the datasets. However, studies
done by Quilitz et al.,63 and work developed by Alexander et al.,64 merely provided the statistical information on the
datasets instead of conceptual analysis for knowledge discovery from biomedical datasets. There is little effort for
the schema level analysis of the concepts and their relationships in these datasets with respect to systematic and
semantic querying. Seaborne and Prudhommeaux pointed out the difficulty with the SPARQL syntax and
expression21, because the precise details of the structure of the graph should be specified for queries in the triple
pattern through the various heterogeneous schemas. In reality, users may not be familiar with the details of datasets,
and it is hard to express the precise relationships between concepts in the SPARQL syntax and expressions. Thus,
this can be a bottleneck for users to query through the endpoints of medical ontologies.
Callahan et al. provided a SPARQLed web application for SPARQL query generation by suggesting context
sensitive IRI60. However, they could not provide strong associated queries as we do. Unlike this work, we can
provide not only valid but also meaningful query suggestions in a dynamic manner according to users’ interesting
topics. Godoy et al. presented a collaborative environment to allow user to register queries manually through wiki
pages and share and execute the queries for linked data61. A series of desired queries might be generated using large
22
ontologies like the NCI thesaurus by extracting relevant information 8. The GLEEN project aims to develop a useful
service for simplified, materialized views of complex ontologies 65. However, these works lack the comprehensive
semantic analysis of large sources and the usage of the knowledge for query processing. Unlike these works, our
approach is to automate query generation through predicate neighborhood pattern-based topic discovery without any
human intervention.
C. Future Work on Cross-domain Knowledge Discovery and Query Generation
In the future, we will apply the proposed approach to heterogeneous biomedical ontologies among multiple domains
(e.g., Drug to Gene, Drug to Disease) for knowledge discovery with appropriate semantic granularity66 and query
generation. Specifically, we have investigated the combination of the Human Phenotype Ontology (HPO) 67 with a
collaborative filtering algorithm68 to accelerate rare disease diagnosis69, 70. In the future, we plan to incorporate
heterogeneous knowledge bases to facilitate the implementation of rare disease diagnosis into clinical practice. The
accuracy of the pattern analysis and dynamic similarity computation in cross domain analysis is highly depending on
how well datasets are normalized. Therefore, to achieve this goal, the first essential step is to develop an effective
normalization technique for heterogeneous biomedical datasets. Famous work in this area include Bio2RDF32, SIO43.
In addition to lexicon and semantic based data normalization, some other techniques such as natural language
processing and graph pattern based analysis can be used. Word2vec71 proposed a way to calculate distance between
elements based on bag-of-words. By using this approach, we can find synonym terms from heterogeneous domains
and normalize terms with highest similarity. LIMES72 applied triangle inequality on graph data to find the path
between a source and target nodes. For a source and target nodes in different datasets, we can find predicates along
the path as RDF built-ins <rdfs:sameAs> and <rdfs:seeAlso> and also normalize the synonym terms. A framework
was designed to infer the links between entities across multiple heterogeneous social networks73. It would be another
inspiration for us to better normalize cross domain datasets.
The second task in our future work is to discover more interesting connection patterns and query from integrated
datasets since cross domain connection patterns usually provide more valuable heterogeneous information and
relationships. There are many data integration works like BioPortal5, Bio2RDF32 and OBO74 which are able to
integrate all potential cross domain connection pattern knowledge into one normalized repository. However, none of
them has the ability to discover those potential connection patterns in a dynamic way. However, MedTQ has a
capacity to handle such dynamic cross domain knowledge discovery by exploring new strategies to connect them
together and retrieve information from integrated ontologies and data.
VI.
CONCLUSION
In this paper, we proposed the MedTQ framework for topic discovery and query generation through the analysis of
ontologies. For the MedTQ framework, we have newly designed the Predicate Neighboring Pattern (PNP) model and
performed similarity measurements, the Hierarchical Predicate-based K-Means clustering (HPKM) algorithm and
dynamic query generation algorithm. The proposed MedTQ framework was evaluated using a case study with
Bio2RDF ontologies (DrugBank). In this case, we demonstrated that MedTQ framework can dynamically discover
cohesive topics for a given ontology as well as generate interesting queries for the discovered topics. In addition, we
successfully validated the optimal clustering results, thereby providing a solid evidence for automatic topic
discovery and query generation. In particular, we have implemented and deployed a tool for topic discovery and
interactive query generation as well as the SPARQL endpoint for query processing with multiple medical ontologies
and datasets from Bio2RDF.
Figure Legends
Figure 1. Predicate Share Patterns: In a RDF triplet <S, P, O>, there is a direction from a subject S to an object O
through a predicate P. From the basic unit of RDF triple <S, P, O>, a specific context of a predicate P can be
discovered from the associated concepts (S and O). Interestingly, the neighbors of predicates P will also provide
additional information through the association context. In this figure, a circle represents a concept (S stands for a
subject and O stands for an object) and a square represents a predicate (relations). The Share pattern describes the
resources sharing relationships (P) between interacting concepts such as shared subjects (S) or shared objects (O)
through the given relationship. This pattern describes that the same subject and object are shared by two predicates
P1 and P2 (the leftmost one), the same subject shared (the rightmost one), and the object shared (the middle one).
23
Figure 2. Predicate Connection Patterns: In a RDF triplet <S, P, O>, there is a direction from a subject S to an
object O through a predicate P. From the basic unit of RDF triple <S, P, O>, a specific context of a predicate P can
be discovered from the associated concepts (S and O). Interestingly, the neighbors of predicates P will also provide
additional information through the association context. In this figure, a circle represents a concept (S stands for a
subject and O stands for an object) and a square represents a predicate (relations). The Connection pattern describes
the resources are connected through multiple relationships (P) between interacting concepts such as shared subjects
(S) or shared objects (O). The pattern in this figure describes three types of connection patterns: Level 1: given two
RDF triples <S1, P1, O1> and <S2, P2 , O2>, O1 is equal to S2 and then there is a predicate connection pattern. Level 2:
given three RDF triples <S1, P1, O1>, <S2, P2 , O2>, and <S3, P3, O3>, O1 is equal to S2, O2 is equal to S3, and then
there is a predicate connection pattern between P 1 and P3. Level 2: given four RDF triples <S1, P1, O1>, <S2, P2 , O2>,
<S3, P3, O3>, and <S3, P3, O3>, O1 is equal to S2, O2 is equal to S3, and O3 is equal to S4, ,then there is a predicate
connection pattern between P1 and P4.
Figure 3. Predicate Neighbouring Patterns (PNP) and Similarity Matrix: This matrix represents the predicate
similarity computation for the Share patterns and Connection patterns. A Share pattern is identified between
predicates 𝑃1 and 𝑃2 and the Connection patterns are identified between 𝑃1 and 𝑃3 , 𝑃1 and 𝑃4 , 𝑃1 and 𝑃5. Based on
the PNP patterns, SM[𝑃𝑖 , 𝑃𝑗 ] is computed based on the formulas given in Definitions 1 – 4.
Figure 4. Sublette Width and Number of Topics in Topic Hierarchy: This figure shows a topic hierarchy that
was constructed by the HPKM algorithm. This topic hierarchy is composed of topics in a tree structure. Each node
in each topic is representing concepts (circles) and predicates (triangles) of the topic. The average 𝑠𝑤(𝑝𝑖 ) is
computed based on the predicates of each topics (The sw value for each topic is shown together with the size of
predicates in the parentheses) level-by-level. The branching factor K is defined by the branching optimization
approach.
Figure 5. Automatic Query Generation: This figure illustrates how to generate a SPARQL query from topics
generated. The query generation is based on two step process: for a given topic, i) transforming from a topic graph
to a query graph and ii) transforming from the query graph to a SPARQL query. In this example, a topic was
discovered from the integration of three ontologies, DrugBank, Sider, PharmGKB ontologies. The query generated
from a topic graph includes triples with two predicates, drug and x-pubchem-substance, and their subject variables
(?E and ?D) and object variables (?D and ?R). The type of variable ?E is known as Drug Effect, ?D as Drug, and ?R
as PharmGKB Resource. These can be converted to a triplet form such as <?D typeof Drug>. The SPARQL query
shown in the top-left textbox is automatically generated for the given topic graph (right-side).
Figure 6. SPARQL Endpoint: A web-based query endpoint was developed and deployed. In this example, a
SPARQL query was generated from Topic 3_4 of the DrugBank ontology (Topic 4 at level 3) based on the top
ranked concepts (such as dv:Resource; dv:Drug, dv:Carrier, dv:Enzyme, dv:Target) and the top ranked predicates
(such as transporter, target, enzyme, carrier) with their concepts dv:Target-Relation, dv:Target-Relation,
dv:Enzyme-Relation, dv:Carrier-Relation. Through this endpoint, a SPARQL query can be designed in an
interactive manner and the query results can be retrieved.
Figure 7. MedTQ Interactive Query Tool: This figure shows how to use the MedTQ tool for browsing the
generated topics and performing interactive design and processing of queries. Step 1 shows the list of topics for a
given ontology (DrugBank). Step 2 shows the list of NLP questions for a selected topic (Topic 7). Step 3 shows the
automatically generated SPARQL query and the query results. Step 4 shows the topic and query graphs for the
selected query.
Figure 8. DrugBank Topic Hierarchy: The topic hierarchy was generated for a single domain ontology,
DrugBank. As seen in this figure, DrugBank has the number of topics <2:7:8> with 2 topics at the first level, 7
topics at the second level, and 8 topics at the third level (T3_1, T3_2, T3_3, T3_4, T3_5, T3_6, T3_7, T3_8). Kmeans clustering was performed in a top-down manner until the average of topics’ Silhouette Width is higher than a
24
certain threshold (> 0.5). The number on each edge in the topic hierarchy represents the percentage of predicates that
the upper level topic graph contributes to the lower level graph. The eight topics in the bottom level (i.e., level 3) are
shown with their high ranked two unique predicates.
Figure 9. Optimal K Branching Factors using Multiple Clustering Techniques. Compared to existing clustering
algorithms (such as Clara, PAM, and Hierarchical Clustering), K-Means Clustering showed the best optimal cluster
number showing with a stable silhouette width for varying the number of clusters. The proposed HPKM algorithm,
which is an extension of K-Means clustering, determines the most significant number of branching factors for
clustering at each level (shown as a red circle) such as (a) K = 2 with SW = 0.77 at level 1; (b) K = 5 with SW =
0.64 at level 2; (c) K=2 with SW = 0.74 at level 2; (e) K = 2 with SW = 0.72 at level 3.
Figure 10. Topic Rankings for DrugBank. This shows relevance scales of five different rankings and an overall
ranking. The overall rank was computed in terms of the five criteria: i) Top 20 Concepts, ii) Top 20 Predicates, iii)
Similarity, iv) Silhouette Width, v) 𝐷𝑒𝑛𝑠𝑖𝑡𝑦. The numbers shown in this figure are rankings. Topic 4 at level 3
(T3_4) and Topic 7 at level 3 (T3_7) are ranked 1 st and 2nd (of 8 topics), respectively.
Figure 11. Topic 3_4 (Topic 4 at Level 3) Graph in DrugBank. This topic graph includes 6 predicates and 12
concepts with 72 in-degree and out-degree. Specifically, two groups of predicates are shown; one is with four
predicates such as transporter, target, enzyme, carrier with concepts dv:Target-Relation, dv:Target-Relation,
dv:Enzyme-Relation, dv:Carrier-Relation. Another group of predicates such as x-genbank and x-uniprot is a
connector predicate group that connects between internal concepts (e.g., dv:Drug, dv:Enzyme) and external
concepts (e.g., gv:Resource, unv:Resource). In this graph, concepts are represented as a circle, predicates as a
triangle, and links as an arrow. Specifically, the dark red items are the predicates and concepts mentioned in Query-1.
This topic describes the relations between drugs with their target, enzyme, enzyme relation, and target relation. The
prefixes in this graph describe the domain of the concepts as follows: dv:http://bio2rdf.org/drugbankvocabulary and
gv: http://bio2rdf.org/genbankvocabulary
Figure 12. Partial Results of Query-1 in DrugBank. This figure shows the partial results from Query-1 about drug
target and transporter enzyme of some drugs (e.g., Gemcitabine [drugbank:DB00441], Fluorouracil
[drugbank:DB00544], Ribavirin [drugbank:DB00811]). All the possible drugs, enzyme, target, enzyme relations,
target relations for any drugs that share the same target and transporter enzyme.
Figure 13. Topic 3_7 (Topic 7 at Level 3) Graph in DrugBank. Topic 3_7 is highly ranked. In this topic graph,
three predicates, (i.e., drug, action, reference) are all nicely connected with 7 concepts (dv:Transporter-Relation,
dv:Target-Relation, dv:Drug, dv:Carrier-Relation, dv:Enzyme-Relation, pv:Resource, dv:Resource). Concepts are
represented as a circle, predicates as a triangle, and links as an arrow. In addition, the dark red triangle items (drug
and reference) are the predicates and the dark red circle items (dv:Transporter-Relation, dv:Target-Relation,
dv:Drug, dv:Carrier-Relation) are the concepts mentioned in Query-2. The prefixes in this graph describe the
domain of the concepts as follows: dv: http://bio2rdf.org/drugbankvocabulary and
pv:
http://bio2rdf.org/pubmedvocabulary
Figure 14. Results of Query-2 in DrugBank. This figure shows the partial results from Query-2 about some drugs
such as Phenytoin (DrugBank:DB00252), Lepirudin (DrugBank:DB00001) and Deferasirox (DrugBank:DB01609).
All the possible combinations and their PubMed references for Target-relation, Carrier-relation and Transporterrelation of drugs sharing the common information of Target-relation, Carrier-relation and Transporter-relation are
retrieved.
Figure 15. Graph of Query-3 in DrugBank. This shows T3_6 (Topic 6 at level 3) topic graph in which concepts
are represented as a circle, predicates as a triangle, and links as an arrow. In particular, there are two subgraphs; one
is with two predicates such as source and calculated-properties with concepts dv:Boiling-Point and
dv:Bioavailability, respectively. Another predicate experimental-properties is connected with concepts such as
25
dv:Water-Solubility. The prefix in this graph describes the domain of the concepts and predicates as follows: dv:
http://bio2rdf.org/drugbankvocabulary.
Figure 16. Results of Query-3 in DrugBank. Query-3 is designed to retrieve information on some drugs’
experimental properties and calculated properties of drugs that contain octanol-water partition coefficient. The
partial query results on Query-3 include the information on some drugs (e.g., L-Histidine [drugbank:DB00117], LPhenylalanine [drugbank:DB00120], L-Arginine [drugbank:DB00125]) and all their experimental properties and
calculated properties that have LogP experimental properties (octanol-water partition coefficient).
Figure 17. Topic 3_2 (T3_2) Graph in DrugBank. In this graph, concepts are represented as a circle, predicates as
a triangle, and links as an arrow. In this graph, two dominant concepts, dv:Resource and dv:Drug, are fully
connected to the remaining concepts via predicates such as absorption, protein-binding. In this graph, concepts are
represented as a circle, predicates as a triangle, and links as an arrow. The dark red items like abortion and product
are the predicates and dv:Drug and dv:Pharmaceutical are the concepts mentioned in Query-4. The dark red triangle
items like abortion and product are the predicates and the dark red circle items like dv:Drug and dv:Pharmaceutical
are the concepts mentioned in Query-4. The prefixes in this graph describe the domain of the concepts as follows:
dv:
http://bio2rdf.org/drugbankvocabulary,
chv:
http://bio2rdf.org/chebivocabulary,
dpv:
http://bio2rdf.org/dpdvocabulary, kv: http://bio2rdf.org/keggvocabulary, pdv: http://bio2rdf.org/pdbvocabulary,
phv:
http://bio2rdf.org/pharmgkbvocabulary,
psv:
http://bio2rdf.org/pubchem.substancevocabulary,
and
wv:http://bio2rdf.org/wikipediavocabulary
Figure 18. Partial Results of Query-4 in DrugBank. Query-4 is designed to retrieve information on some drugs’
absorption, affected-organism, clearance pharmacokinetic measurement, pharmaceutical information, and protein
binding information. The partial query results on Query-4 include the information on some drugs (e.g.,
Gemcitabine[drugbank:DB00441], Tiotropium[drugbank:DB01409], Carvedilol[drugbank:DB1136]) having the
common information of their absorption, affected-organism, clearance pharmacokinetic measurement,
pharmaceutical information, and protein binding information.
Reference
1.
Tilahun BC. Linked Data based Health Information Representation, Visualization and Retrieval System on the
Semantic Web; 2013.
2. Cheung K-H, et al. A journey to Semantic Web query federation in the life sciences. BMC bioinformatics.
2009;10(10):S10.
3. Simperl E. Reusing ontologies on the Semantic Web: A feasibility study. Data & Knowledge Engineering.
2009;68(10):905-25.
4. An Y, et al. Inferring complex semantic mappings between relational tables and ontologies from simple
correspondences. OTM Confederated International Conferences" On the Move to Meaningful Internet
Systems"; 2005: Springer.
5. Whetzel PL, et al. BioPortal: enhanced functionality via new Web services from the National Center for
Biomedical Ontology to access and use ontologies in software applications. Nucleic acids research.
2011;39(suppl_2):W541-W5.
6. Shen F, et al. Knowledge discovery from biomedical ontologies in cross domains. PloS one.
2016;11(8):e0160005.
7. Shen F, et al. BmQGen: Biomedical query generator for knowledge discovery. Bioinformatics and Biomedicine
(BIBM), 2015 IEEE International Conference on; 2015: IEEE.
8. Nekrutenko A, et al. Next-generation sequencing data interpretation: enhancing reproducibility and
accessibility. Nature Reviews Genetics. 2012;13(9):667.
9. Zhang Y, et al. An integrative computational approach to identify disease-specific networks from PubMed
literature information. Bioinformatics and Biomedicine (BIBM), 2013 IEEE International Conference on; 2013:
IEEE.
10. Zhang Y, et al. Systematic identification of latent disease-gene associations from PubMed articles. PloS one.
2018;13(1):e0191568.
26
11. Zhu Q, et al. Exploring the pharmacogenomics knowledge base (pharmgkb) for repositioning breast cancer
drugs by leveraging web ontology language (OWL) and cheminformatics approaches. Biocomputing 2014:
World Scientific; 2014. p. 172-82.
12. Shen F, et al. Using Human Phenotype Ontology for Phenotypic Analysis of Clinical Notes. Studies in health
technology and informatics. 2017;245:1285-.
13. Shen F, et al. Phenotypic Analysis of Clinical Narratives Using Human Phenotype Ontology. Studies in health
technology and informatics. 2017;245:581-5.
14. Wishart DS, et al. DrugBank: a comprehensive resource for in silico drug discovery and exploration. Nucleic
acids research. 2006;34(suppl_1):D668-D72.
15. Klyne G, et al. Resource description framework (RDF): Concepts and abstract syntax. 2006.
16. Johnson SC. Hierarchical clustering schemes. Psychometrika. 1967;32(3):241-54.
17. Hartigan JA, et al. Algorithm AS 136: A k-means clustering algorithm. Journal of the Royal Statistical Society
Series C (Applied Statistics). 1979;28(1):100-8.
18. Rousseeuw PJ. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of
computational and applied mathematics. 1987;20:53-65.
19. Chen M-S, et al. Data mining: an overview from a database perspective. IEEE Transactions on Knowledge and
data Engineering. 1996;8(6):866-83.
20. Hewett M, et al. PharmGKB: the pharmacogenetics knowledge base. Nucleic acids research. 2002;30(1):163-5.
21. Prud E, et al. SPARQL query language for RDF. 2006.
22. Eclipse Juno Integrated Development Environment. (Accessed 2018, at:https://www.eclipse.org/juno/).
23. Carroll JJ, et al. Jena: implementing the semantic web recommendations. Proceedings of the 13th international
World Wide Web conference on Alternate track papers & posters; 2004: ACM.
24. The R Project for Statistical Computing. (Accessed 2018, at: http://www.r-project.org/).
25. Shannon P, et al. Cytoscape: a software environment for integrated models of biomolecular interaction
networks. Genome research. 2003;13(11):2498-504.
26. Erling O, et al. RDF Support in the Virtuoso DBMS. Networked Knowledge-Networked Media: Springer;
2009. p. 7-24.
27. ClnicalTrial. (Accessed 2016, at: http://cu.clinicaltrials.bio2rdf.org/sparql).
28. DrugBank. (Accessed 2016, at: http://cu.drugbank.bio2rdf.org/sparql).
29. OMIM. (Accessed 2016, at: http://cu.omim.bio2rdf.org/sparql).
30. PharmGKB. (Accessed 2016, at: http://cu.pharmgkb.bio2rdf.org/sparql).
31. SIDER. (Accessed 2016, at: http://cu.sider.bio2rdf.org/sparql).
32. Belleau F, et al. Bio2RDF: towards a mashup to build bioinformatics knowledge systems. Journal of biomedical
informatics. 2008;41(5):706-16.
33. Ng RT, et al. CLARANS: A method for clustering objects for spatial data mining. IEEE transactions on
knowledge and data engineering. 2002;14(5):1003-16.
34. Kaufman L, et al. Partitioning around medoids (program pam). Finding groups in data: an introduction to
cluster analysis. 1990:68-125.
35. Dasgupta S, et al. SMARTSPACE: Multiagent based distributed platform for semantic service discovery. IEEE
Transactions on Systems, Man, and Cybernetics: Systems. 2014;44(7):805-21.
36. Vaka P, et al. PEMAR: A pervasive middleware for activity recognition with smart phones. Pervasive
Computing and Communication Workshops (PerCom Workshops), 2015 IEEE International Conference on;
2015: IEEE.
37. Chen Z, et al. Collaborative mobile-cloud computing for civil infrastructure condition inspection. Journal of
Computing in Civil Engineering. 2013;29(5):04014066.
38. Shen F. Situation Aware Mobile Apps Framework: University of Missouri-Kansas City; 2012.
39. Shen F, et al. SAMAF: Situation aware mobile apps framework. Pervasive Computing and Communication
Workshops (PerCom Workshops), 2015 IEEE International Conference on; 2015: IEEE.
40. Horrocks I, et al. SWRL: A semantic web rule language combining OWL and RuleML. W3C Member
submission. 2004;21:79.
41. Luciano JS, et al. The Translational Medicine Ontology and Knowledge Base: driving personalized medicine by
bridging the gap between bench and bedside. Journal of biomedical semantics; 2011: BioMed Central.
42. Chen B, et al. Chem2Bio2RDF: a semantic framework for linking and data mining chemogenomic and systems
chemical biology data. BMC bioinformatics. 2010;11(1):255.
43. Dumontier M, et al. The Semanticscience Integrated Ontology (SIO) for biomedical research and knowledge
discovery. Journal of biomedical semantics. 2014;5(1):14.
27
44. Croset S, et al. Integration of the anatomical therapeutic chemical classification system and drugbank using owl
and text-mining. GI Workgroup Ontologies in Biomedicine and Life Sciences (OBML). 2012.
45. Momtchev V, et al. Expanding the pathway and interaction knowledge in linked life data. Proc of International
Semantic Web Challenge. 2009.
46. Williams AJ, et al. Open PHACTS: semantic interoperability for drug discovery. Drug discovery today.
2012;17(21-22):1188-98.
47. Baorto D, et al. Practical experience with the maintenance and auditing of a large medical ontology. Journal of
biomedical informatics. 2009;42(3):494-503.
48. Ziegler P, et al. Three Decades of Data Intecration—all Problems Solved? Building the Information Society:
Springer; 2004. p. 3-12.
49. Shen F, et al. Using semantic web technologies for quality measure phenotyping algorithm representation and
automatic execution on EHR data. Biomedical and Health Informatics (BHI), 2014 IEEE-EMBS International
Conference on; 2014: IEEE.
50. Tao C, et al. Phenotyping on EHR data using OWL and semantic web technologies. International Conference
on Smart Health; 2013: Springer.
51. Warrender JD, et al. A pattern-driven approach to biomedical ontology engineering. arXiv preprint
arXiv:13120465. 2013.
52. Wang H, et al. An integrative and interactive framework for improving biomedical pattern discovery and
visualization. IEEE Transactions on Information Technology in Biomedicine. 2004;8(1):16-27.
53. Rafiq MI, et al. Computational method for temporal pattern discovery in biomedical genomic databases.
Computational Systems Bioinformatics Conference, 2005 Proceedings 2005 IEEE; 2005: IEEE.
54. Shen F. A pervasive framework for real-time activity patterns of mobile users. Pervasive Computing and
Communication Workshops (PerCom Workshops), 2015 IEEE International Conference on; 2015: IEEE.
55. Shen F. A graph analytics framework for knowledge discovery: University of Missouri-Kansas City; 2016.
56. Shen F, et al. Predicate Oriented Pattern Analysis for Biomedical Knowledge Discovery. Intelligent information
management. 2016;8(3):66.
57. Tartir S, et al. Ontological evaluation and validation. Theory and applications of ontology: Computer
applications: Springer; 2010. p. 115-30.
58. Sabou M, et al. Exploring the semantic web as background knowledge for ontology matching. Journal on data
semantics XI: Springer; 2008. p. 156-90.
59. Lorey J, et al. Detecting SPARQL query templates for data prefetching. Extended Semantic Web Conference;
2013: Springer.
60. Callahan A, et al. Improved Dataset Coverage and Interoperability with Bio2RDF Release 2. SWAT4LS; 2012.
61. García Godoy MJ, et al. Sharing and executing linked data queries in a collaborative environment.
Bioinformatics. 2013;29(13):1663-70.
62. Loizou A, et al. On the formulation of performant sparql queries. Web Semantics: Science, Services and Agents
on the World Wide Web. 2015;31:1-26.
63. Quilitz B, et al. Querying distributed RDF data sources with SPARQL. European Semantic Web Conference;
2008: Springer.
64. Alexander K, et al. Describing Linked Datasets. LDOW; 2009.
65. Detwiler LT, et al. Regular paths in sparql: Querying the nci thesaurus. AMIA Annual Symposium Proceedings;
2008: American Medical Informatics Association.
66. Shen F, et al. Populating Physician Biographical Pages Based on EMR Data. AMIA Summits on Translational
Science Proceedings. 2017;2017:522.
67. Robinson PN, et al. The Human Phenotype Ontology: a tool for annotating and analyzing human hereditary
disease. The American Journal of Human Genetics. 2008;83(5):610-5.
68. Breese JS, et al. Empirical analysis of predictive algorithms for collaborative filtering. Proceedings of the
Fourteenth conference on Uncertainty in artificial intelligence; 1998: Morgan Kaufmann Publishers Inc.
69. Shen F, et al. Leveraging Collaborative Filtering to Accelerate Rare Disease Diagnosis. American Medical
Informatics Association. 2017.
70. Shen F, et al. Accelerating Rare Disease Diagnosis with Collaborative Filtering. American Medical Informatics
Association. 2017.
71. Mikolov T, et al. Distributed representations of words and phrases and their compositionality. Advances in
neural information processing systems; 2013.
72. Ngomo A-CN, et al. Limes-a time-efficient approach for large-scale link discovery on the web of data. IJCAI;
2011.
28
73. Kong X, et al. Inferring anchor links across multiple heterogeneous social networks. Proceedings of the 22nd
ACM international conference on Information & Knowledge Management; 2013: ACM.
74. Smith B, et al. The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration.
Nature biotechnology. 2007;25(11):1251.
29
| 2cs.AI
|
On Equivariant Poincaré Duality,
Gysin Morphisms and Euler Classes
Alberto Arabia∗
arXiv:1702.03889v6 [math.AT] 10 Nov 2017
July 2017
Abstract. Given a connected compact Lie group G, we set up the formalism
of the G-equivariant Poincaré duality for oriented G-manifolds, following the
work of J.-L. Brylinski leading to the spectral sequence
ExtHG (HG,c (M ),HG ) ⇒ HG (M )[dM ].
The equivariant Gysin functors
(−)! : G-Manπ
resp. (−)∗ : G-Man
D+ (DGM(HG )),
+
D (DGM(HG )),
M
ΩG (M )[dM ],
f
f!
M
ΩG,c (M )[dM ],
f
f∗
are then the covariant functors from the category of oriented G-manifolds and
proper (resp. unrestricted) maps, to the derived category of the category of
differential graded modules over HG , defined as the composition of the Cartan complex of equivariant differential forms functor ΩG,c (−) (resp. ΩG (−))
with the duality functor IRHom•HG ((−),HG ) and the equivariant Poincaré ad
•
0
junction IDG (M ) : ΩG (M )[dM ] → IRHom
HG ΩG,c (M ),HG (resp. IDG (M ) :
•
ΩG,c (M )[dM ] → IRHomHG ΩG (M ),HG ). Equivariant Euler classes
are next
introduced for any closed embedding i : N ⊆ M as EuG N ,M := i∗ i! (1)
where i∗ i! : HG (N ) → HG (N ) is the push-pull operator. We end recalling
localization and fixed point theorems.
About this work. These notes were originally intended as an appendix to a
book on the foundations of equivariant cohomology of G-manifolds. The idea of
introducing Gysin morphisms through an equivariant Poincaré duality formalism à la Grothendieck-Verdier has many theoretical advantages and is somewhat
uncommon in the equivariant setting, warranting publication of these notes.
Contents
1 Nonequivariant Background
1.1 Category of Cochain Complexes . . . . .
1.1.1 Fields in Use. . . . . . . . . . . .
1.1.2 Vector Spaces Pairings. . . . . . .
1.1.3 Graded Vector Spaces. . . . . . .
1.1.5 Differential Graded Vector Space.
1.1.11 The Dual Complex. . . . . . . . .
1.2 Some Categories of Manifolds . . . . . .
1.2.1 Manifolds. . . . . . . . . . . . . .
1.2.4 G-Manifolds. . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
4
4
4
5
6
6
6
∗ Université Paris Diderot-Paris 7, IMJ-PRG, CNRS, Bâtiment Sophie Germain, bureau
608, Case 7012, 75205. Paris Cedex 13, France. Contact: [email protected].
1
2
Equivariant Poincaré Duality and Gysin Morphisms
1.3 Poincaré Pairing . . . . . . . . . . . . . . . . . . .
1.4 Manifolds and maps of Finite de Rham Type . . .
1.4.1 Definitions. . . . . . . . . . . . . . . . . . .
1.4.3 Ascending Chain Property. . . . . . . . . .
1.4.5 Existence of Proper Functions. . . . . . . .
1.5 Manifolds With Boundary . . . . . . . . . . . . .
1.6 Proof of Proposition 1.4.4 . . . . . . . . . . . . .
1.7 The Gysin Functor . . . . . . . . . . . . . . . . .
1.7.1 The Right Adjoint Map. . . . . . . . . . .
1.7.3 The Gysin Morphism. . . . . . . . . . . . .
1.7.7 The Image of D 0 (M ). . . . . . . . . . . . .
1.8 The Gysin Functor for Proper Maps . . . . . . . .
1.9 Principal Examples of Gysin Morphisms . . . . .
1.9.1 Universal Property of the Gysin Morphism.
1.9.2 Constant Map. . . . . . . . . . . . . . . . .
1.9.4 Open Embedding. . . . . . . . . . . . . . .
1.9.5 Locally Trivial Fibration. . . . . . . . . . .
1.9.7 Zero Section of a Vector Bundle. . . . . . .
1.10 Constructions of Gysin Morphisms . . . . . . . .
1.10.1 The Proper Case. . . . . . . . . . . . . . .
1.10.2 The General Case. . . . . . . . . . . . . . .
1.11 Exercises . . . . . . . . . . . . . . . . . . . . . . .
1.11.1 Gysin Long Exact Sequence. . . . . . . . .
1.11.2 Lefschetz Fixed Point Theorem. . . . . . .
1.12 Conclusion. . . . . . . . . . . . . . . . . . . . . .
2 Equivariant Cohomology Background
2.1 Category of Cochain g-Complexes . . . .
2.1.1 Fields in Use. . . . . . . . . . . .
2.1.2 g-modules. . . . . . . . . . . . . .
2.1.5 Differential Graded g-Complexes.
2.1.7 Morphisms of g-Complexes. . . . .
2.1.8 Category of g-Complexes. . . . . .
2.1.9 Split g-Complexes. . . . . . . . . .
2.2 Equivariant Cohomology of g-Complexes
2.2.1 The symmetric Algebra of g ∨ . . .
2.2.2 Cartan Complexes. . . . . . . . .
2.2.7 Split G-Complexes. . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
9
9
9
9
10
11
12
12
12
15
16
18
18
18
18
18
19
20
20
20
21
21
22
23
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
29
29
29
30
30
30
30
32
32
32
34
3 Equivariant Cohomology of G-Manifolds
3.1 Equivariant Differential Forms . . . . . . .
3.1.1 Fields in Use. . . . . . . . . . . . .
3.1.2 G-Derivations and Contractions. . .
3.2 The Borel Construction . . . . . . . . . . .
3.2.1 The Classifying Space. . . . . . . .
3.2.6 Serre Spectral Sequences. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
35
35
35
39
39
40
4 Equivariant Poincaré Duality
4.1 Differential Graded Modules over a Graded Algebra
4.1.1 Graded Algebras. . . . . . . . . . . . . . . .
4.1.3 Graded Modules. . . . . . . . . . . . . . . .
4.1.9 Differential Graded Modules. . . . . . . . . .
4.1.10 The Hom• (−, −) and (− ⊗ −)• Bi-functors.
4.1.13 The Dual Complex. . . . . . . . . . . . . . .
4.1.14 The Forgetful Functor. . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
41
41
41
43
43
44
44
3
Alberto Arabia
4.2 Deriving Functors . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Deriving Functors Defined on the Category GM(HG ). . .
4.2.2 Simple Complex Associated with a Double Complex. . .
4.2.3 Spectral Sequences Associated with Double Complexes. .
L?
4.2.4 The IR? Hom•HG (−, −) and (−) ⊗IH
( ) Bi-functors. . .
G −
•
•
4.2.6 The Ext and Tor Bi-functors. . . . . . . . . . . . . . .
4.2.7 Defining IR? Hom• (−,HG ) on DGM(HG ). . . . . . . . .
4.2.8 Spectral Sequences Associated with IR? Hom•HG (−,HG ).
4.3 Equivariant Integration . . . . . . . . . . . . . . . . . . . . . . .
4.3.2 Equivariant Integration vs. Integration Along Fibers. . .
4.4 Equivariant Poincaré Pairing . . . . . . . . . . . . . . . . . . . .
4.5 G-Equivariant Poincaré Duality Theorem . . . . . . . . . . . . .
4.5.2 Torsion-freeness, Freeness and Reflexivity. . . . . . . . . .
4.6 T-Equivariant Poincaré Duality Theorem . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
44
44
45
46
46
47
47
48
50
51
51
52
54
55
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
56
56
58
59
60
61
62
62
62
62
63
63
64
6 The field of fractions of HG
6.1 The Localization Functor . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Localized Equivariant Poincaré Duality. . . . . . . . . . . . . . .
6.1.4 Localized Equivariant Gysin Morphisms. . . . . . . . . . . . . .
65
65
66
7 Equivariant Euler Classes
7.1 The Nonequivariant Euler Class . . . . . . . . . . . . . . . . .
7.2 G-Equivariant Euler Class . . . . . . . . . . . . . . . . . . . .
7.2.2 G-Equivariant Euler Class of Discrete Fixed Point Sets
7.2.4 G- and T-Equivariant Euler Classes of a Fixed Point. .
7.3 Torsions in Equivariant Cohomology Modules . . . . . . . . .
7.3.1 Torsions. . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3.4 The slice theorem. . . . . . . . . . . . . . . . . . . . . .
7.3.7 Orbit Type of T-Manifolds. . . . . . . . . . . . . . . .
7.4 Localization Theorems . . . . . . . . . . . . . . . . . . . . . .
66
67
67
67
70
70
71
71
72
5 Equivariant Gysin Morphism
5.1 G-Equivariant Gysin Morphism in the General Case .
5.1.1 Equivariant Finite de Rham Type Coverings. . .
5.2 G-Equivariant Gysin Morphism for Proper Maps . . .
5.3 Comparison Theorems . . . . . . . . . . . . . . . . . .
5.4 Universal Property of the equivariant Gysin Morphism
5.5 Group Restriction . . . . . . . . . . . . . . . . . . . . .
5.6 Explicit Constructions of Equivariant Gysin Morphisms
5.6.1 Constant Map. . . . . . . . . . . . . . . . . . . .
5.6.2 Equivariant Open Embedding. . . . . . . . . . .
5.6.3 Equivariant Projection. . . . . . . . . . . . . . .
5.6.4 Equivariant Fiber Bundle. . . . . . . . . . . . .
5.6.5 Zero Section of an Equivariant Vector Bundle .
5.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
8 Miscellany
9 Appendix
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
74
74
Acknowledgements. To Matthias Franz for his helpful criticism on the equivalence between equivariant duality and reflexivity, and to Frances Cowell for her
wise comments which allowed to significantly improve this work.
4
Equivariant Poincaré Duality and Gysin Morphisms
1. Nonequivariant Background
1.1. Category of Cochain Complexes
1.1.1. Fields in Use. Unless otherwise specified, K denotes either the field of
real numbers R, or the field of complex numbers C.
1.1.2. Vector Spaces Pairings. Whenever K is understood the expression
“vector space” means vector space over K. If V is a vector space, we denote by
V ∨ := HomK (V,K) its dual .
Given a bilinear map β : V × W → K, also called a pairing, consider the
two linear maps γβ : V → W ∨ and ρβ : W → V ∨ , defined by γβ (v)(w) =
ρβ (w)(v) := β(v,w) and respectively called the left and right adjoint maps associated with β. One says that β is a nondegenerate pairing whenever the adjoint
maps are injective, and one says that β is a perfect pairing whenever they are
bijective. For example, the canonical pairing V ∨ × V → K, (λ,v) 7→ λ(v), is
always nondegenerate and it is perfect if and only if V is finite dimensional.
1.1.3. Graded Vector Spaces. A graded space is a family V := {V m }m∈Z
of vector spaces. A graded homomorphism α : V → W of degree d = deg(α)
is a family of linear maps {αm : V m → W m+d }m∈Z , composition of such is
defined degree by degree, i.e. β ◦ α = {βm+d ◦ αm }m∈Z . One has deg(α ◦ β) =
degα + degβ.
We denote by HomgrdK (V ,W ) the space of graded homomorphisms of degree
d and by Hom•K (V ,W ) the graded space of all graded homomorphisms, i.e. the
family
Hom•K (V ,W ) = HomgrdK (V ,W ) d∈Z .
When d = 0, we may write HomgrK (V ,W ) for Homgr0K (V ,W ).
1.1.4 The category GV(K) of graded vector spaces is the category whose objects
are graded spaces and whose morphisms are the graded homomorphisms of
degree 0. We denote equivalently MorGV(K) (V ,W ) := HomgrK (V ,W ) the set
of morphisms from V to W .
1.1.5. Differential Graded Vector Space. A differential graded vector space
(V ,d), a complex in short, is a graded vector space V together a differential
or coboundary d : Endgr1 (V ) such that d 2 = 0. A morphism of complexes
α : (V ,d) → (V 0 ,d 0 ) is a morphism α ∈ HomgrK (V ,V 0 ) commuting with
differentials, i.e. α ◦ d = d 0 ◦ α. The complexes and their morphisms constitute
the category DGM(K) of differential graded vector spaces.
1.1.6 A morphism of complexes α : (V ,d) → (V 0 ,d 0 ) induces a morphism
between the graded spaces of cohomologies H(α) : H(V ,d) → H(V 0 ,d 0 ). The
morphism α is a quasi-isomorphism, quasi-injection, quasi-surjection, whenever
H(α) is respectively, an isomorphism, injection, surjection.
1.1.7 Let m ∈ Z. If L is a vector space, we denote by L[m] the graded space
defined by L[m]−m = L and L[m]n = 0 if n + m 6= 0. If α : V → W is a linear
map, we denote by α[m] : V [m] → W [m] the morphism of graded spaces equal
to α in degree −m and 0 otherwise. The correspondence L
(L[m],0) and
α
α[m] is a functor
[m] : Vec(K) → DGM(K).
5
Alberto Arabia
More generally, If V is a graded space we denote by V [m] the graded space
(V [m])i = V m+i , and if α : V → W is a graded homomorphism we denote
by α[m] : V [m] → W [m] the graded homomorphism α[m]i = αm+i . Next, if
(V ,d) is a complex, (V ,d)[m] is the complex V [m],(−1)m d[m] . The correspondence (V ,d)
(V ,d)[m], α
α[m] is the m-th shift functor
[m] : DGM(K)
DGM(K).
1.1.8 Given two complexes (V ,d) and (V 0 ,d 0 ), we recall the definition of the
complexes
Hom•K (V ,V 0 ),D
and (V ⊗K V 0 )• ,∆ .
As graded vector spaces they are
(
m
0
0
Homm
K (V ,V ) = HomgrK (V ,V )
m ∈ Z 7→
Q
(V ⊗K V 0 )m = b+a=m V a ⊗K V 0b
and their differentials are
(
Dm (f ) = d 0 ◦ f − (−1)m f ◦ d
∆m (v ⊗ v 0 ) = d(v) ⊗ v 0 + (−1)|v| v ⊗ d 0 (v 0 )
0
where v ⊗ v 0 ∈ V |v| ⊗ V 0|v | . (1 )
1.1.9. Exercise. Verify that the following complexes coincide as graded vector
spaces but not as complexes even though they are naturally isomorphic.
Hom•K (V [n],W [m]) ' Hom•K (V ,W )[m − n]
V [n] ⊗ W [m] ' (V ⊗ W )[m + n]
1.1.10 Given a morphism of complexes ϕ : (V ,d) → (W ,d) the map
m
0
0
Homm
K (W ,V ) → HomK (V ,V ),
α 7→ α ◦ ϕ,
is well defined for all m ∈ Z and commutes with differentials so that one has a
morphism of complexes
Hom•K (α,V 0 ) : Hom•K (W ,V 0 ),D → Hom•K (V ,V 0 ),D .
The correspondence (V ,d)
contravariant functor
Hom•K (V ,V 0 ), α
Hom•K (−,V 0 ) : DGM(K)
Hom•K (α,V 0 ) is then a
DGM(K).
1.1.11. The Dual Complex. The functor Hom•K (−,K[0]) is the duality functor , simply denoted by (−)∨ := Hom•K (−,K[0])
(−)∨ : DGM(K)
DGM(K).
The complex (V ,d)∨ is called the dual complex associated with (V ,d). One has
(V ∨ )m = HomK (V −m ,K),
1 It
Dm = (−1)m+1 t d−(m+1)
is worth noting that these formulae are inspired by the super Lie bracket equalities
[[d,f ]] = df − (−1)|d||f | f d
and
[[d,ab]] = [[d,a]] + (−1)|a||d| a[[d,b]]
where [[d,[[d, −]]]] = 0 is an immediate consequence of |d| = 1 and d2 = 0.
6
Equivariant Poincaré Duality and Gysin Morphisms
1.1.12. Remark. One must take care that the natural embedding of vector
spaces V ⊆ V ∨∨ gives only an inclusion of complexes (V ,−d) ⊆ (V ,d)∨∨
where the sign of the differential has changed ! The canonical isomorphism
: (V ,d) → (V ,−d),
m = (−1)m idV m
()
is then necessary to get a canonical embedding (V ,d) ,→ (V ,d)∨∨ .
The next statement will be used without mention, it is left as an exercise.
1.1.13. Proposition
a) A morphism of complexes α : (V ,d) → (V 0 ,d 0 ) is a quasi-isomorphism if
and only if α∨ is so.
b) There exists a canonical isomorphism between the cohomology of the dual
and the dual of the cohomology, i.e.
∨
h (V ,d)∨ −−
.
'→ h(V ,d)
where h denotes the graded vector space of the cohomologies of a complex.
1.2. Some Categories of Manifolds
1.2.1. Manifolds. The names manifold and map (when applied to manifolds)
will be shortcuts for real differentiable manifold and smooth map. Manifolds are
equidimensionnal, i.e. all their connected components have the same dimension,
unless otherwise specified. The notation “dM ”, unless otherwise indicated, will
always denote the dimension on M .
1.2.2 Man (resp. Manor ) denotes the category of (equidimensional) manifolds
(resp. oriented) and smooth maps. Over Man one has the (real) de Rham complex contravariant functor
Ω(−) : Man
DGM(R)
and the de Rham cohomology contravariant functor
H(−) : Man
ModN (R).
or
1.2.3 Manπ (resp. Manor
π ) denotes the subcategory of Man (resp. Man )
with the same objects but with only proper maps. Over Manπ one has, in
addition to the previous functors, the (real) de Rham complex with compact
supports contravariant functor:
Ωc (−) : Manπ
DGM(R)
and the de Rham cohomology with compact support contravariant functor
Hc (−) : Manπ
DGM(R).
The inclusion Ωc (−) ⊆ Ω(−) induces a morphism of functors Hc (−) → H(−).
1.2.4. G-Manifolds. Let G denote a Lie group. A manifold endowed with
a smooth action of G is called a G-manifold . A map f : M → N between
G-manifolds is called a G-equivariant if it commutes with the action of G. The
class of G-manifolds and G-equivariant maps constitutes the category G-Man.
7
Alberto Arabia
The categories G-Manor , G-Manπ , G-Manor
π are the analogues of those already introduced in this section.
1.3. Poincaré Pairing
The reference for this section is [BT] (I §5 p. 44). Let M be an oriented manifold.
The composition of theRbilinear map ΩdM −i (M ) ⊗ Ωic (M ) → Ωdc M (M ), α ⊗ β 7→
α ∧ β, with integration M : Ωdc M (M ) → R, gives rise to a nondegenerate pairing
IP (M ) : ΩdM −i (M ) ⊗ Ωic (M ) −→ R
Z
α⊗β
7−→
α∧β
(IP )
M
inducing the Poincaré pairing in cohomology
P(M ) : H dM −i (M ) ⊗ Hci (M ) −→ R
Z
[α] ⊗ [β]
7−→
[α] ∪ [β]
(P)
M
The left adjoint map associated with IP is
ID(M ) : ΩdM −i (M ) −→ Ωic (M )∨
α
Z
7−→ ID(α) := β 7→
α∧β
and one has
(ID)
M
ID(M ) (−1)dM dα (β) =
Z
ZM
=
(−1)dM dα ∧ β
(−1)dM + |α|+1 α ∧ dβ = (−1)|β| ID(M )(α)(dβ),
M
Hence, following the conventions introduced in 1.1.7 and 1.1.8, ID(M ) is a morphism of complexes from Ω(M )[dM ] to Ωc (M )∨ .
1.3.1. Exercise. Show that (IP ) is a nondegenerate pairing.
1.3.2. Theorem Poincaré duality theorem. Let M be an oriented manifold.
a) The morphism of complexes, called the Poincaré morphism,
ID(M ) : Ω(M )[dM ] ,−→ Ωc (M )∨
(∗)
is a quasi-isomorphism, i.e. the morphism of graded spaces it induces in
cohomology
D(M ) : H(M )[dM ] −−
→ Hc (M )∨ ,
(∗∗)
'
is an isomorphism.
b) The Poincaré pairing in cohomology
P(M ) : H(M ) ⊗ Hc (M ) −−→ R
(∗∗∗)
is always nondegenerate. It is perfect (see 1.1.2) if and only if H(M ) is finite
dimensional.
Proof. The (a) part (cf. [BT] p. 44–, for details) states the bijectivity of the left
adjoint map associated with P. Then, for each fixed i, one obtains by duality
the bijectivity of Di∨ : (HcdM −i (M ))∨∨ → H i (M )∨ and the composition of
this map with the canonical embedding i : HcdM −i ,→ (HcdM −i )∨∨ is the right
8
Equivariant Poincaré Duality and Gysin Morphisms
adjoint map ρP : Hc (M )[dM ] → H(M )∨ (see 1.1.2). The “finite dimensional”
condition then ensures the bijectivity of i , hence of ρP .
1.3.3. Exercise. Let M and N be oriented manifolds. We denote by
Z
ID(−) : Ω(−)[d− ] → Ω(−)∨
,
I
D(α)(β)
=
α∧β
c
the left adjoint map of the Poincaré pairing, and by ID0 (−) : Ωc (−)[d− ] →
R
Ω(−)∨ , ID0 (β)(α) = α∧β the right adjoint map. A pair (L,R) of morphisms of
complexes L : Ω(N ) → Ω(M ) and R : Ωc (M )[dM ] → Ωc (N )[dN ] is a (Poincaré)
adjoint pair whenever
Z
Z
L(α) ∧ β =
α ∧ R(β)
M
N
for all α ∈ Ω(N ) and β ∈ Ωc (M ). Show that
a) If (L,R1 ) and (L,R2 ) are adjoint pairs, then R1 = R2 . One says that R = R1
is the (Poincaré) right adjoint of L.
b) If (L1 ,R) and (L2 ,R) are adjoint pairs, then L1 = L2 . One says that L = L1
is the (Poincaré) left adjoint of R.
c) If (L,R) is an adjoint pair, then
ID ◦ L = R∨ ◦ ID,
ID0 ◦ R = L∨ ◦ ID0 ,
i.e. the following diagrams are commutative
ID 0 (M )
ID(M )
Ω(M ) ,−−−
−−→ Ωc (M )∨
(')
x
x
∨
L
R
Ωc (M ) ,−−−−−→ Ω(M )∨
∨
Ry
yL
Ω(N ) ,−−−
−−→ Ωc (N )∨
(')
Ωc (N ) ,−−−−−→ Ω(N )∨
ID 0 (N )
ID(N )
d) Do the exercise in the cohomological framework, i.e. replace Poincaré pairing
(IP ) by (P), ID by D : H[d] → Hc∨ , ID0 by D 0 : Hc [d] → H ∨ , and define the
notion of (Poincaré) adjoint pair in cohomology.
Show that if (L,R) is an adjoint pair of morphisms of complexes, then
(H(L),Hc (R)) is an adjoint pair in cohomology so that one has
D ◦ H(L) = Hc (R)∨ ◦ D ,
D 0 ◦ Hc (R) = H(L)∨ ◦ D 0 .
In particular, H(L) identifies with the dual of Hc (R) via Poincaré duality.
1.3.4. Remark. We shall see that, given f : M → N , the pullback morphism
f ∗ : Ω(N ) → Ω(M ) may or may not admit a right Poincaré adjoint at the
complexes level, but that it will always do so at the cohomology level, this right
adjoint is the Gysin morphism f∗ : Hc (M ) → Hc (N ), so that (H(f ∗ ),f∗ ) is a
Poincaré adjoint pair in cohomology. On the other hand, when f is a proper
map, the pullback f ∗ : Ωc (N ) → Ωc (M ) is also well defined and one may look
for a left Poincaré adjoint to f ∗ , i.e. some morphism L : Ω(M )[dM ] → Ω(N )[dN ]
Z
Z
L(α) ∧ β =
α ∧ f ∗ (β).
N
M
Again, this will sometimes be possible at the complex level and will always be
possible at the cohomology level leading to the notion of the Gysin morphism
Alberto Arabia
9
for proper maps f! : H(M ) → H(N ), so that (f! ,Hc (f ∗ )) is a Poincaré adjoint
pair in cohomology.
1.4. Manifolds and maps of Finite de Rham Type
1.4.1. Definitions. A manifold M is said to be of finite (de Rham) type whenever its de Rham cohomology ring H(M ) is finite dimensional. A map between
manifolds f : M → N si said to be of finite (de Rham) type if N is the union of
a countable ascending chain U := {U0 ⊆ U1 ⊆ ···} of open subspaces of finite
type such that each subspace f −1 (Um ) ⊆ M is of finite type.
1.4.2. Remarks
a) By Poincaré duality (1.3.2), M is of finite type if and only if its de Rham
cohomology with compact support Hc (M ) is finite dimensional.
b) A compact manifold is of finite type ([BT] 5.3 pp. 42-43). An oriented manifold is of finite type if and only if its Poincaré pairing in cohomology is perfect
(1.3.2-(b)), which will be used in our discussion of the Gysin morphism.
c) Since any manifold is the union of a countable ascending chain {↑ Um }
of open submanifolds of finite type (cf. 1.4.4), any locally trivial fibration
f : M → N is of finite type (exercise).
1.4.3. Ascending Chain Property. Although general manifolds need not be
of finite type, they are always the inductive limit of such. More precisely, any
manifold M is the union of an ascending chain {U0 ⊆ U1 ⊆ ···} of open subsets
of finite type of M . This weaker finiteness property, sufficient for our needs, is
generally proved by a riemannian geometry argument (2 ). When a manifold is
endowed with the action of a Lie group G, we will require in addition that each
Un be G-stable.
1.4.4. Proposition. Let G be a compact Lie Group. A G-manifold M is the
union of a countable ascending chain U := {U0 ⊆ U1 ⊆ ···} of G-stable open
subsets of M of finite type.
The next sections recall some facts needed in the proof of this proposition.
1.4.5. Existence of Proper Functions. Recall that a map between manifolds
f : M → N is said to be proper whenever f −1 (F) is compact for any compact
subset F of N . The aim of this section is to show that on a G-manifold there
are always proper G-invariant functions.
Since the existence of proper functions over compact manifolds is clear, let
M be a noncompact manifold. Fix a countable, locally finite covering U :=
{Un }n∈N of M , where each Un is a relatively compact open subset of M , and note
that the noncompactness of M implies that the family is necessarily infinite.
Next, fix a smooth partition of unity {φn }n∈N subordinate to U . This means in
particular that for each n ∈ N, the equality φn (x) = 0 holds whenever x 6∈ Un .
Then one has for every N ∈ N
X
1=
φn (x), ∀x 6∈ U0 ∪ ··· ∪ UN .
()
n>N
2 In these notes, a good cover of M (also known as Leray cover ) is a finite open covering
U = {Ui | i = 1,...,r} of M such that all intersections Ui1 ∩ ... ∩ Uik are either vacuous or
acyclic ([BT], p. 5). The existence of good covers is established in loc.cit. §5, p. 42.
10
Equivariant Poincaré Duality and Gysin Morphisms
P
Now, for every x ∈ M , the infinite sum Φ(x) := n∈N n·φn (x) is actually finite
and smooth with respect to x ∈ M , as it is a locally finite sum of smooth
functions.
1.4.6. Lemma. The positive function Φ : M → R is unbounded and proper.
Proof. By property () one has
X
X
Φ(x) >
n·φn (x) > N
n>N
n>N
φn (x) = N ,
∀x 6∈ U0 ∪ ··· ∪ UN , ()
which clearly implies that Φ is an unbounded function on M . Now, to see that
Φ is proper, remark that if F ⊆ R is compact, then F ⊆ [−N,N ] for some
N ∈ N and Φ−1 (F) ⊆ U0 ∪ ··· ∪ UN by (). But the closure U0 ∪ ··· ∪ UN is a
compact subset of M because each U i is assumed compact. As a closed subset
of a compact set, Φ−1 (F) is compact.
As a corollary of the preceding lemma one has:
1.4.7. Proposition. Manifolds M endowed with a smooth action of a compact
Lie group G admit proper G-invariant positive functions Φ : M → R.
Proof. If M is compact, any positive constant map Φ will do. If M is not
compact, let φ : M → R denote any unbounded proper positive function (see
1.4.6), and set:
Z
Φ(x) :=
φ(g·x) dg,
G
R
where dg is a G-invariant form of top degree on G, such that 1 = G dg. The
correspondence x 7→ Φ(x) is clearly a well-defined nonnegative unbounded Ginvariant function of M into R. Now, for each N ∈ N, the set
MN := G·φ−1 ([−N,N ])
is compact and G-stable, and if y 6∈ MN , one has φ(g·y) > N for all g ∈ G, so
that
Z
Φ(y) =
φ(g·y) dg > N ,
()
G
and properness of Φ follows by the same argument as in lemma 1.4.6: If F is a
compact subset of R, then F ⊆ [−N,N ] for some N ∈ N, and Φ−1 (F) ⊆ MN
by (). Then Φ−1 (F) is compact since it is closed in the compact set MN .
1.5. Manifolds With Boundary
The following is well known.
1.5.1. Proposition. The interior of a compact manifold with boundary is of
finite type.
Proof. Let B be a compact manifold with boundary and let M be its interior.
Gluing B with itself along its boundary ∂B, one gets the “double” B t∂B B,
which is a compact manifold without boundary. Then, from the long exact
sequence of de Rham cohomology with compact support (see 1.11.1-(a))
··· −→ Hci (M ) × Hci (M ) −→ H i (B t∂B B) −→ H 0 i(∂B) −→ ··· ,
11
Alberto Arabia
where H ∗ (B t∂B B) and H ∗ (∂B) are finite-dimensional, the finiteness of
Hc∗ (M ) follows immediately. The finiteness of H ∗ (M ) results from Poincaré
duality H ∗ (M ) ∼
= Hc∗ (M )∨ .
1.6. Proof of Proposition 1.4.4
Recall that the connected components of a manifold
M are always open and
`
closed submanifolds of M . In particular, if M = a∈A Ca denotes the decomposition of M in connected components, the indexing set A is finite or countable,
and the restriction of a proper function Φ : M → R to each Ci remains proper.
If all the connected components of M are compact, we may index them by
natural numbers C0 ,C1 ,... and define Un := C0 ∪ C1 ∪ ··· ∪ Cn . Each Un is
then open in M and is also a compact manifold, hence it is of finite type. The
ascending chain {U0 ⊆ U1 ⊆ ···} satisfies the conditions of the proposition.
If M contains a noncompact connected component C, fix any proper positive
G-invariant function Φ : M → R, which is possible due to 1.4.7, and note that
Φ(C) is necessarily unbounded, since otherwise C ⊆ Φ−1 ([0,T ]) for some T ∈ R,
and C would be compact as Φ is proper over C. Moreover, there exists N ∈ N
such that Φ(M ) ⊇ Φ(C) ⊇ (]N,+∞[), since Φ(C) is unbounded and connected.
Now, by Sard’s theorem, the interior of the set of critical values of Φ : M → R
is empty so that there exists an unbounded increasing sequence of positive real
numbers {N < t0 < ··· < tn < ···}n∈N which are regular values of Φ. Each subset
Mn := Φ−1 (tn ) is then a submanifold of codimension 1 in M and, moreover,
it is compact and G-stable since Φ is proper and G-invariant. Similarly, the
sets Un := Φ−1 (]−∞,tn [) and Wn := Φ−1 (]tn ,+∞[), clearly nonempty, are Gstable open subsets of M . One then easily checks that U n = Un t Mn and
W n = Mn t Wn are in fact G-manifolds with boundary Mn embedded in M .
Furthermore, U n is compact as one has U n := Φ−1 (]−∞,tn ]) = Φ−1 ([0,tn ])
since Φ is positive.
R
M
Φ
Wn
Wn
Wn
tn
n
n
n
Un
Un
M
compact
n
0
We can then apply proposition 1.5.1 and state that Un is a G-stable open subset
of finite type of M . The ascending chain {U0 ⊆ U1 ⊆ ···} satisfies the conditions
of the proposition.
12
Equivariant Poincaré Duality and Gysin Morphisms
1.7. The Gysin Functor
In this section we dualize ID(M ) : Ω(M )[dM ] → Ωc (M )∨ , getting an injection
D 0 (M ) : Hc (M )[dM ] ,→ H(M )∨ whose image will be shown to be functorial
on the category Manor of oriented manifolds.
1.7.1. The Right Adjoint Map. In 1.3 we introduced the left adjoint map
associated with Poincaré pairing, i.e. the quasi-isomorphism
ID(M ) : Ω(M )[dM ] −(')
−→ Ωc (M )∨ .
By duality, this map yields ID(M )∨ : Ωc (M )∨∨ → Ω(M )[dM ]∨ which is also
a quasi-isomorphism and, composed with the embedding Ωc (M ) ⊆ Ωc (M )∨∨ ,
gives rise to the injection and quasi-injection (1.3.1, 1.1.6)
⊆
D∨
Ωc (M )[dM ],d ,−−−−→ Ωc (M )∨∨ [dM ],−d −−I(')
−−→ Ω(M )∨ ,−d
x
0
ID (M )
(See 1.1.11 for the sign of differentials.) One has (cf. 1.3.3)
Z
ID0 (M )(β) = α 7→
α∧β ,
M
which clearly it is the right adjoint map associated with the Poincaré pairing IP .
The following proposition paraphrases the statement 1.3.2-(b).
1.7.2. Proposition. Let M be an oriented manifold.
a) The morphism of complexes
ID0 (M ) : Ωc (M )[dM ],d) ,−→ Ω(M )∨ ,−d
is always an injection and a quasi-injection. We will denote by
D 0 (M ) : Hc (M )[dM ] ,−→ H(M )∨
the induced injection in cohomology.
b) The morphism D 0 (M ) is an isomorphism if and only if M is of finite type.
1.7.3. The Gysin Morphism. The last statement shows that in the oriented
finite type case, compact support cohomology canonically coincides with the
dual of arbitrary support cohomology so that if M and N are such, the diagram
D 0 (M )
∨
Hc (M)[dM ] ,−−−−
'−−→ H(M )
f∗ y
⊕
y H(f ∗ )∨
(D 0 )
0
D (N )
∨
Hc (N )[dN ] ,−−−−
'−−→ H(N )
can be closed in one and only one way with the morphism of graded spaces
f∗ : Hc (M )[dM ] → Hc (N )[dN ]. It is then clear that the correspondence defined
over the category of oriented finite type manifolds, that assigns M
M∗ :=
Hc (M )[dM ] and f
f∗ , is a covariant functor.
When the manifold N in (D 0 ) is oriented but not of finite type, D 0 (N ) is
still an injection but it is no longer surjective so that it is not obvious that
the diagram can be closed. Statement (b) in the following theorem gives a
positive answer to this question showing that the image of D 0 (−) is “stable
under pullbacks”. Hence, it will always be possible to induce f∗ : M∗ → N∗ ,
13
Alberto Arabia
the Gysin morphism associated with f . After that, the correspondence M
M∗ := Hc (M )[dM ] and f
f∗ will appear to be a covariant functor defined
over the whole category Manor , the Gysin functor .
1.7.4. Theorem and definitions
a) Let M be oriented and endow its open subsets with induced orientations.
For any inclusion of open subsets i : V ⊆ W , denote by i∗ : Ωc (V ) → Ωc (W )
the map that assigns to β ∈ Ωc (V ) its extension by zero to W , called the
pushforward of β. Then, the following diagrams
D 0 (V )
ID 0 (V )
∨
Ωc (V)[dM ] ,−−−−−−→ Ω(V
)
∗∨
i∗ y
y (i )
∨
Hc (V)[dM ] ,−−−−−−→ H(V
)
∗∨
Hc (i∗ )y
y H(i )
Ωc (W )[dM ] ,−−−−−−→ Ω(W )∨
Hc (W )[dM ] ,−−−−−−→ H(W )∨
D 0 (W )
ID 0 (W )
are commutative, i.e. (i∗ ,i∗ ) is a Poincaré adjoint pair (1.3.3).
b) For any map f : M → N between oriented manifolds, one has the diagram
D 0 (M )
∨
Hc (M )[dM ] ,−−−−−−→ H(M
)
f∗ y
y H(f ∗ )∨
(D 0 )
0
D (N )
Hc (N )[dN ] ,−−−−−−→ H(N )∨
where H(f ∗ )∨ Im(D 0 (M )) ⊆ Im(D 0 (N )), so that there exists one and only
one morphism of graded spaces
f∗ : Hc (M )[dM ] −→ Hc (N )[dN ]
()
called the Gysin morphism associated with f , making (D 0 ) commutative,
i.e. (H(f ∗ ),f∗ ) is a Poincaré adjoint pair in cohomology, which means that,
for any [α] ∈ H(N ) and [β] ∈ Hc (M ), the equation in X,
Z
Z
f ∗ [α] ∪ [β] =
[α] ∪ X ,
()
M
N
admits one and only one solution in Hc (N ), namely X = f∗ [β].
Furthermore, f∗ in () is a morphism of H(N )-modules, i.e. the equality,
called the projection formula,
f∗ f ∗ [α] ∪ [β] = [α] ∪ f∗ ([β])
()
holds for all [α] ∈ H(N ) and [β] ∈ Hc (M ).
c) The correspondence
or
(−)∗ : Man
GV(R)
with
M
f
M∗ := Hc (M )[dM ]
f∗
is a covariant functor. It will be called the Gysin functor.
d) If M and N are oriented of finite type, then f ∗ : H(N ) → H(M ) is an isomorphism if and only if the Gysin morphism f∗ : Hc (M )[dM ] → Hc (N )[dN ]
is also an isomorphism.
Proof. (a) The commutativity results from the equality
Z
Z
αV ∧β =
α ∧ i∗ β
V
W
14
Equivariant Poincaré Duality and Gysin Morphisms
for α ∈ Ω(W ) and β ∈ Ωc (V ), which is evident since the support of α ∧ i∗ β is
contained in V .
(b) One must verify that, given [β] ∈ Hc (M ), the linear form
Z
[α] ∈ H(N ) 7→
f ∗ [α] ∪ [β]
M
is of the form
Z
[α] ∈ H(N ) 7→
[α] ∪ [β 0 ]
N
for some [β 0 ] ∈ Hc (N ). Now, thanks to proposition 1.4.4, there exists an open
subset W ∈ N of finite type such that f −1 W contains the support of β, denoted
β := β f −1 W . One then has the following commutative diagram:
D 0 (f −1 W )
[β] ∈ Hc (f −1 W )[dM ]
Hc (i∗ )
"
[β] ∈ Hc (M )[dM ]
/ H(f −1 W )∨
(i∗ )∨
(I)
D 0 (M )
"
/ H(M )∨
(f ∗ )∨
(II)
/ H(W )∨
0
D (W )
0
[β ] ∈ Hc (W )[dN ]
Hc (i∗ )
'
"
Hc (N )[dM ]
(i∗ )∨
(I)
0
D (N )
(f ∗ )∨
"
/ H(N )∨
where subdiagrams (I) are commutative after (c) and the commutativity of (II)
is just functoriality of pullbacks.
Following the arrows, we see that
(f ∗ )∨ ◦ D 0 (M )([β]) = (i∗ )∨ ◦ (f ∗ )∨ ◦ D 0 (f −1 W )([β])
= (i∗ )∨ ◦ D 0 (W )([β 0 ])
= D 0 (N ) ◦ Hc (i∗ )([β 0 ])
where [β 0 ] ∈ Hc (W )[dN ] verifies
D 0 (W )([β 0 ]) = (f ∗ )∨ ◦ D 0 (f −1 W )([β])
which is possible since D 0 (W ) is surjective as W is of finite type !
The statement about the equation () is clear and implies formally the projection formula since
Z
Z
[ω] ∪ f∗ f ∗ [α] ∪ [β] =
f ∗ [ω] ∪ f ∗ [α] ∪ [β]
N
ZM
Z
∗
=
f ([ω] ∪ [α]) ∪ [β] =
[ω] ∪ [α] ∪ f∗ [β].
M
N
Finally, (c) is trivial since D is bijective over its image, and (d) is clear.
0
1.7.5. Remark. It is important to note that the main ingredients in the proof
are (i) the Poincaré pairings, (ii) Poincaré duality, (iii) the ascending chain
property (1.4.3). In later sections of these notes we will show that all these
ingredients exist also in the equivariant setting so that the last theorem and its
proof will extend verbatim to G-manifolds and G-equivariant cohomology.
15
Alberto Arabia
1.7.6. Exercise. Let f : M → N be a map between oriented manifolds. Show
that the dual of the Gysin morphism f∗ : Hc (M )[dM ] → Hc (N )[dN ] coincides,
via Poincaré duality, with the pullback morphism f ∗ : H(N ) → H(M ).
1.7.7. The Image of D 0 (M ). The next proposition will be used when extending the Gysin functor to the equivariant context. It gives a description of
the image of D 0 (M ) in terms of ascending chains of open finite type subsets of
M , which was the main reason why we proved that such coverings always exist
(see 1.4.4).
1.7.8. Proposition. Let U be a filtrant open covering (3 ) of a manifold M .
a) Let i : V ⊆ W denote an inclusion of open subsets of M .
The map i∗ : Ωc (V ) ⊆ Ωc (W ), that assigns to β ∈ Ωc (V ) the differential
form i∗ (β) ∈ Ωc (W ) equal to β over V and 0 otherwise, is a well-defined
morphism of complexes inducing in cohomology the morphism of graded
spaces Hc (i∗ ) : Hc (V ) → Hc (W ). One has also the morphism of complexes
i∗ : Ω(W ) → Ω(V ) that restricts a differential form of W to V , and the
corresponding morphism of graded spaces H(i∗ ) : H(W ) → H(V ).
These constructions, applied to the elements of U , give rise to the inductive systems {Ωc (U )}U ∈U and {Hc (U )}U ∈U , and to the projective systems
{Ω(U )}U ∈U and {H(U )}U ∈U , whence the canonical maps
ν : lim
Ω (U ) → Ωc (M ) and H(ν) : lim
H (U ) → Hc (M ),
−→U ∈U c
−→U ∈U c
µ : Ω(M ) → lim
Ω(U ) and H(µ) : H(M ) → lim
H(U ).
←−U ∈U
←−U ∈U
All these maps are bijective.
b) Suppose M is oriented, then the map
ID0 (U ) : Ωc (M )[dM ],d −→ lim
Ω(U )∨ ,−d
−→U ∈U Z
β
7→
α 7−→
α∧β
M
is a well-defined morphism of complexes inducing in cohomology the map
D 0 (U ) : Hc (M )[dM ] → lim
H(U )∨
−→U ∈U
c) Suppose further that each U ∈ U is of finite type. Then ID0 (U ) is a quasiisomorphism, and one has
Im(D 0 (M )) = lim
H(U )∨ ⊆ H(M )∨
−→U ∈U
()
Moreover, the adjoint D 0 (U )∨ canonically identifies with D(M ); more precisely, the following diagram is commutative:
D 0 (U )∨
lim H(U )[dM ] = ( lim H(U )∨ )∨ [dM ] −−−−−−→ Hc (M )∨
←−
−→
x
U ∈U
U ∈U
'
D(M )
H(M )[dM ] −−−−−−−−−−−−−−−−−−−−−−−−→ Hc (M )∨
3 We recall that U = {U }
a a∈A is said filtrant whenever for all U1 ,U2 ∈ U there exists
U3 ∈ U such that (U1 ∪ U2 ) ⊆ U3 .
16
Equivariant Poincaré Duality and Gysin Morphisms
Proof. (a) The map ν : lim
Ω∗ (U ) → Ω∗c (M ) is injective since it’s the limit
−→U ∈U c
of a filtrant inductive system of injective maps. The image of ν is the union of
Ω∗c (U ) for the same reason. Now, if ω ∈ Ω∗c (M ), its support, being compact, is
contained in some U ∈ U so that
pushforward of ω U ∈ Ω∗c (U ). This
S ω is the
∗
∗
justifies the equality Ωc (M ) = U ∈U Ωc (U ) and proves that ν is surjective.
Standard arguments on the homology of filtrant inductive systems of complexes
prove that H(ν) is bijective.
The map µ : Ω(M ) → lim U ∈U Ω(U ) is injective, since a differential form
←−
is null if and only if it is locally null. To see it is also surjective, let {αU ∈
Ω(U )}U ∈U be a given projective system of differential forms, and note that for
any x ∈ M , the element α̃(x) := αU (x) is well defined since if x ∈ U1 ∈ U
and x ∈ U2 ∈ U , one chooses U3 ∈ U s.t. U1 ∪ U2 ⊆ U3 , in which case
αU1 (x) = αU3 (x) = αU2 (x). Likewise, one verifies the differentiability of α̃. It is
clear that α̃ U = αU , which ends the proof that µ is surjective.
It remains only to justify why H(µ) is bijective. This is immediate when M is
orientable, since H(µ) is then just the Poincaré dual of Hc (ν) which has already
been shown to be bijective. Otherwise, when M is not orientable, we lift U to
the orientation manifold M̃ associated with M through the canonical Z/2Zcovering p : M̃ M , setting therefore U˜ :={Ũ :=p−1 (U )|U ∈U }. As M̃ is
orientable, the map H(M̃ ) → lim U ∈U H(Ũ ) is now bijective, and because this
←−
map is also compatible with the reversing-orientation action of Z/2Z, it induces a
bijection between invariants subspaces H(M̃ )Z/2Z −−
lim U ∈U H(Ũ )Z/2Z , and
'→ ←−
one concludes since H(U ) = H(Ũ )Z/2Z .
(b) Endow each U ∈ U with the induced orientation. Taking the inductive
limit of the maps ID0 (U ) : Hc (U )[dM ] → H(U )∨ and applying (a) one sees
immediately that ID(U ) = lim
ID0 (U ).
−→U ∈U
(c) By 1.7.2 the maps ID0 (U ) : Hc (U )[dM ] → H(U )∨ are quasi-isomorphisms
for each U ∈ U , hence ID(U ) = lim
ID0 (U ) is also a quasi-isomorphism
−→U ∈U
since U is filtrant. The rest of the statement is then clear by duality.
1.8. The Gysin Functor for Proper Maps
In this section, the Gysin morphism for compact supports f∗ : Hc (M )[dM ] →
Hc (N )[dN ] will be extended to arbitrary supports f! : H(M )[dM ] → H(N )[dN ]
when f : M → N is a proper map. As we will see this case is much simpler
than the general one as it results immediately from Poincaré duality.
When f : M → N is proper, the pullback f ∗ : Ω(N ) → Ω(M ) respects
compact supports and induces a morphism of complexes f ∗ : Ωc (N ) → Ωc (M ),
giving rise to the covariant functor from Manπ to Vec(K)
M
Hc (M )∨ ,
f
Hc (f ∗ )∨ .
When M is oriented, ID0 (M ) may be extended from Ωc (M ) to Ω(M ) by
setting (see 1.7.1)
Z
ID0 (M )(α) = β 7→
β ∧ α , ∀α ∈ Ω(M ), ∀β ∈ Ωc (M ),
M
Alberto Arabia
so that the diagram
17
ID 0 (M )
Ω(M ) −−−
−−→ Ωc (M )∨
(')
x
x
⊆
ID 0 (M )
Ωc (M ) −−−−−→ Ω(M )∨
is commutative, and, moreover, with its first line a quasi-isomorphism as it is
simply the Poincaré duality map ID(M ) up to ±1.
1.8.1. Definition. If f : M → N is a proper map between oriented manifolds,
the Gysin morphism associated with f is the map f! : H(M )[dM ] → H(N [dN ])
making commutative the diagram
D 0 (M )
∨
H(M)[dM ] −−−−
'−−→ Hc (M )
f! y
y Hc (f ∗ )∨
D 0 (N )
∨
H(N )[dN ] −−−−
'−−→ Hc (N )
The next theorem, analog to 1.7.4 and almost immediate, is left as an exercise.
1.8.2. Theorem and definitions
a) For β ∈ Hc (N ) and α ∈ H(M ), the equation in X,
Z
Z
∗
f [β] ∪ [α] =
[β] ∪ X ,
M
(??)
N
admits one and only one solution in H(N ), namely X = f! [α].
Furthermore, f! is a morphism of Hc (N )-modules, i.e. the equality, called
the projection formula for proper maps,
f! f ∗ [β] ∪ [α] = [β] ∪ f! [α]
(???)
holds for all [β] ∈ Hc (N ) and [α] ∈ H(M ).
b) The following correspondence is a covariant functor:
M
M! := H(M )[dM ]
or
f! : Manπ
GV(R)
with
f
f!
We will refer to it as the Gysin functor for proper maps.
c) The pullback f ∗ : Hc (N ) → Hc (M ) is an isomorphism if and only if the
Gysin morphism f! : H(M )[dM ] → H(N )[dN ] is also an isomorphism.
d) The natural map φ(−) : Hc (−)[d− ] → H(−)[d− ] (1.2.3) is a homomorphism
of Gysin functors (−)∗ → (−)! over the category Manor
π , i.e. the diagrams
φ(M )
Hc (M
)[dM ] −−−−→ H(M)[dM ]
f∗ y
y f!
φ(N )
Hc (N )[dN ] −−−−→ H(N )[dN ]
are natural and commutative.
18
Equivariant Poincaré Duality and Gysin Morphisms
1.9. Principal Examples of Gysin Morphisms
1.9.1. Universal Property of the Gysin Morphism. This property is the
statement (b) in theorem 1.7.4, which says that if f : M → N is a map between
oriented manifolds, then for each [β] ∈ Hc (M ), the element f∗ ([β]) ∈ Hc (N ) is
determined by the equality, for all [α] ∈ H(N ),
Z
Z
∗
f [α] ∪ [β] =
M
[α] ∪ f∗ [β]
()
N
The pair (f∗ ,f ∗ ) is a Poincaré adjoint pair in cohomology (1.3.3).
1.9.2. Constant Map. Let M be oriented and denote by cM : M → {•} the
constant map. One applies () taking α = 1:
Z
Z
cM ∗ ([β]) =
1 ∪ cM ∗ [β] =
β.
{•}
M
so that the Gysin morphism cM ∗ : Hc (M )[dM ] → Hc ({•}) = R is the integration
map, Poincaré dual of the graded algebra homomorphism c∗M : R → H(M ).
1.9.3. Exercise. Show that c∗M : Ω({•}) → Ω(M ) admits a right Poincaré
adjoint at the complex level, i.e. cM ∗ : Ωc (M )[dM ] → Ω({•}).
1.9.4. Open Embedding. Let M be oriented. Given an open embedding
i : U ⊆ M , endow U with the induced orientation. For any β ∈ Ωc (U ) one has
the tautological equality:
Z
Z
α ∧ i∗ β
(∗)
αU ∧β =
M
U
where i∗ β ∈ Ωc (M ) denotes the extension by zero of β. The Gysin morphism
i∗ : Hc (U )[dU ] → Hc (M )[dM ] is therefore the pushforward i∗ = Hc (i∗ )[dM ]
(see 1.7.8-(a)). Note also that the equality (∗) shows that the pair (i∗ ,i∗ ) is a
Poincaré adjoint pair (1.3.3).
1.9.5. Locally Trivial Fibration. Let π : E → B be a locally trivial fibration
with base space B (connected for simplicity) and total space E both assumed
oriented, with fiber F of dimension dF endowed with the induced orientation.
Under these assumptions one has the operation of integration along F (see [BT]
I§6 pp. 61-63) which is a morphism of complexes
Z
: Ωc (E)[dF ] → Ωc (B)
F
satisfying
Z
E
π∗ α ∧ β =
Z
B
Z
α∧
β ,
(∗)
F
R
so that after the adjunction property (), one has π∗ = F [dB ] and the Gysin
morphism is Rthe shift of integration along fibers. Note again that (∗) shows that
the pair (π ∗ , F [dB ]) is a Poincaré adjoint pair.
19
Alberto Arabia
1.9.6. Proposition. Let (π,V ,B) and (π,V 0 ,B 0 ) be two oriented locally trivial
fibrations. Let g : B 0 → B be a proper map and assume the following diagram
g
cartesian:
0
V
−−→ V
yπ
πy
g
B 0 −−→ B
i.e. V 0 = (b0 ,v) ∈ B 0 × V g(b0 ) = π(v) . Then
∗
g ◦ π∗ = π∗ ◦ g ∗ : Hc (V ) → Hc (B 0 )
π ∗ ◦ g! = g! ◦ π ∗ : H(B 0 ) → H(V )
Hint. By adjointness, the first equality is equivalent to the
R The first
R second.
equality follows from the equality for differential forms g ∗ F ω = F g ∗ (ω) for
all ω ∈ Ωc (V ), that may be verified locally in B 0 (loc.cit.).
1.9.7. Zero Section of a Vector Bundle. Let (π,V ,B) be a vector bundle
and assume B and V oriented. The zero section map σ : B → V is a closed
embedding, hence proper, so that we have the Gysin morphism for proper maps
σ! : H(B) → H(V ). By the adjunction property (??) (see 1.8.2-(a)), one has
for all β ∈ Hc (V ) and α ∈ H(B)
Z
Z
Z
∗
[β] ∪ σ! ([α]) =
σ [β] ∪ [α] =
σ ∗ [β] ∪ σ ∗ (π ∗ [α])
V
B
B
Z
Z
()
=
σ ∗ ([β] ∪ π ∗ [α]) ∪ 1 =
[β] ∪ π ∗ [α] ∪ σ! (1)
B
V
where Φ := σ! (1) is the Thom class of the pair (B,V ). The Gysin morphism
associated with the zero section of a fiber bundle
σ! : H(B)[dB ] → H(V )[dV ]
(!)
is then the multiplication by the Thom class
σ! ([α]) = π ∗ [α] ∪ Φ.
(!!)
Finally, note that σ! is generally not an isomorphism, since it identifies, via
Poincaré duality, with the dual of the proper pullback σ ∗ : Hc (V ) → Hc (B)
(see 1.8.1) which is generally not an isomorphism (4 ).
It can be seen ([BT] §I.6 p. 64) that if α ∈ Hc (B), then π ∗ [α] ∪ Φ naturally
belongs to Hc (V ) so that the Gysin morphism
σ∗ : Hc (B)[dB ] → Hc (V )[dV ]
(?)
is given by the same equality (!!),
σ∗ ([β]) = π ∗ [β] ∧ Φ.
(??)
On the other hand, the Poincaré lemma for vector bundles asserts that the
pullback π ∗ : H(B) → H(V ) is an isomorphism and this implies, via Poincaré
duality (see 1.7.6), that π∗ : Hc (V )[dV ] → Hc (B)[dB ] is also an isomorphism.
Now, by functoriality, one has π∗ ◦ σ∗ = id, so that σ∗ is also an isomorphism.
This isomorphism is the Thom isomorphism.
4 For example, if B is compact, H (B) = H(B) = H(V ) and σ ∗ would give a graded
c
isomorphism Hc (V ) ' H(V ), and by Poincaré duality H 0 (V ) ' H dV (V ), which impossible
if V is a vector bundle of positive dimension over B.
20
Equivariant Poincaré Duality and Gysin Morphisms
1.9.8. Proposition. Let (π,V ,B) and (π,V 0 ,B 0 ) be two oriented vector bundles and assume the cartesian diagram in 1.9.6 with g : B 0 → B proper. Denote
by σ : B → V and σ : B 0 → V 0 the zero section maps. The diagram
g
B 0 −−→ B
σy
yσ
g
V 0 −−→ V
is cartesian and the equalities
g ∗ ◦ σ∗ = σ∗ ◦ g ∗ : Hc (B) → Hc (V 0 )
σ ∗ ◦ g! = g! ◦ σ ∗ : H(V 0 ) → H(B)
hold.
Hint. It is a corollary of 1.9.6 since σ∗ is the inverse of π∗ .
1.10. Constructions of Gysin Morphisms
In this last preliminary section we summarize the steps in the construction of
the Gysin morphisms.
1.10.1. The Proper Case. Let f : M → N be a proper map of oriented manifolds. ToR α ∈ Ω(M ) we assign the linear form on Ωc (N ) defined
by ID0 (f )(α) : β 7→ M f ∗ β ∧ α. In this way we obtain diagram
f!
Ω(M )[dM ]
/ Ω(N )[dN ]
L
ID 0 (f )
)
ID 0 (N ) (quasi-iso)
Ωc (N )∨
which may be closed in cohomology, since ID0 (N ) is a quasi-isomorphism. Note
that the closing arrow f! , the Gysin morphism for proper maps, in general exists
only at the cohomology level.
1.10.2. The General Case. Let f : M → N be a map of oriented manifolds.
0
To
R β∗∈ Ωc (M ) we assign the linear form on Ω(N ) defined by ID (f )(β) : α 7→
f α ∧ β. In this way we obtain the diagram
M
f∗
Ωc (M )[dM ]
/ Ωc (N )[dN ]
L
0
ID (f )
)
ID 0 (N )
Ω(N )∨
quasi-iso if N is of
finite type
which may be closed in cohomology (as in the proper case), when N is of finite
type, since then ID0 (N ) is a quasi-isomorphism (1.7.2-(b)).
When N is not of finite type, one fixes any filtrant covering U of N made up
of open finite type subsets of N (see 1.4.4), and replaces ID0 (N ) with ID0 (U ).
In this way, we get (see 1.7.8-(b,c)), the following diagram:
Ωc (M )[dM ]
f∗
/ Ωc (N )[dN ]
=
Ωc (N )[dN ]
L
ID 0 (f,U )
ID 0 (U ) (quasi-iso)
*
lim
Ω(U )∨
−→U ∈U
⊆
ID 0 (N )
/ Ω(N )∨
21
Alberto Arabia
where ID0 (f,U ) is defined as follows. For β ∈ Ωc (M ) denote by |β| its support
and by Uβ ⊆ U the system consisting of U ∈ U s.t. |β| ⊆ f −1 U . One has a
natural map lim Ω(U )∨ → lim Ω(U )∨ (which is in fact is bijective). Now,
−→Uβ
−→U R
for every U ∈ Uβ the linear map M f ∗ (−) ∧ β : Ω(U ) → R, is well defined
and is compatible with restriction, so that it defines an element of lim Ω(U )∨ ,
−→Uβ
and then of lim Ω(U )∨ . This element is ID0 (f,U )(β) by definition.
−→U
The closing arrow f∗ , the Gysin morphism associated with a general map f ,
is then defined in cohomology as the composition D 0 (U )−1 ◦ H(ID(f,U )).
1.10.3. Remark. In all cases, the Gysin morphism appears as the composition of a morphism of complexes with the “inverse” of a quasi-isomorphism,
which obviously is possible in cohomology but also in the derived category of
complexes since this is its main property, i.e. a morphism in derived category is
an isomorphism if and only if it induces an isomorphism in cohomology. Gysin
morphisms are well defined morphisms of the derived category of complexes of
vector spaces.
1.11. Exercises
1.11.1. Gysin Long Exact Sequence. Let i : F ⊆ M be a closed embedding
of oriented manifolds. Assume F compact, for simplicity. Put U := M \F and
j : U ⊆ M the canonical injection.
a) i) Let F denote the set of open neighborhood of F. Restriction morphisms
W
RV
: Ω(W ) → Ω(V ) for all W ⊇ V ⊇ F, give rise to a filtrant inductive
W
system {RV
| W ⊇ V in F } and a canonical morphism of complexes
M
RF : Ω(M ) → lim
Ω(V ). Show that the short sequence
−→V ∈F
RM
j∗
0 → Ωc (U ) −−→ Ωc (M ) −−F
→ lim Ω(V ) → 0
−→F
where j∗ is the pushforward morphism, is exact.
ii) Restrictions RFV : Ω(V ) → Ω(F) for V ⊇ F, define a morphism of the
W
inductive system {RV
| W ⊇ V in F } into Ω(F). Denote by RFF :=
V
lim R . Show that
−→F F
RFF : lim Ω(V ) → Ω(F)
−→F
is a quasi-isomorphism.
iii) Derive the existence of the long exact sequence of compact support cohomology
i
∗
c
i
··· → Hck (U ) −−∗→ Hck (M ) −−
→ H k (F) −−k→ Hck+1 (U ) → ···
()
iv) Endow M with a Riemannian metric d : M × M → R. For each ∈ R,
denote
(
IB (F) := {m ∈ M | d(m,F) < }
S (F) := {m ∈ M | d(m,F) = }
If is small enough, IB2 (F) is a fiber bundle with fiber RdM −dF over F
via the geodesic projection π : IB2 → F. By restriction, π : S → F is
22
Equivariant Poincaré Duality and Gysin Morphisms
a fiber bundle with compact fiber SdM −dF −1 . Denote by ` : S ,→ IB2 \F
the canonical injection. We have the following maps
π
π
B2
π
S
π
F
j
S −−`→ IB2 \F −−
→U
⊆
jy ⊆
πy
i
F ,−−−−−−⊆
−
−−−−→ M
Show that the connecting morphism c : H(F) → Hc (U )[1] is given by the
composition of the following morphisms
∗
`∗ [−dS ]
j,∗
π
H(F) −−
→ Hc (S ) −−−−−−−→ Hc (IB2 \F)[1] −−−→ Hc (U
x )[1]
c
where `∗ : Hc (S )[dS ] → Hc (IB2 \F)[dIB2 ] denotes the Gysin morphism
associated with `.
b) i) Dualizing and shifting the long exact sequence of compact support (),
justify the exactness of the Gysin long exact sequence
δ[−1]
i∗ [−dM ]
j∗
−−→ H(F)[dF − dM ] −−−−−−→ H(M ) −−→ H(U ) −−δ→
()
where i : F → N and j : U → N are the canonical injections and δ is
adjoint to the shift of the connecting morphism c in ().
ii) Show that the connecting morphism δ : H(U ) → H(F)[−(dM − dF − 1)]
is simply the restriction to S followed by integration along fibers of π
Z
δ(α) =
α S .
SdM −dF −1
1.11.2. Lefschetz Fixed Point Theorem. Let M be a oriented manifold.
Denote by δ : M → M × M the diagonal embedding x 7→ (x,x) and let
∆ := Im(δ). Given f : M → M , denote Gr(f ) : M → M × M the graph map
x 7→ (f (x),x). The Lefschetz class of f is by definition
L(f ) := Gr(f )∗ (δ! (1)) ∈ H dM (M ),
R
and its Lefschetz number is Λf := M L(f ).
a) Explain de following the equalities
Z
Z
Z
Λf := Gr(f )∗ (δ! (1))= δ! (1)∪Gr(f )! (1)=(−1)dM δ ∗ (Gr(f )! (1)).
M
M ×M
()
M
b) Assuming that f has no fixed points, show that the Gysin morphism
Gr(f )! : H(M )[dM ] → H(M × M )[2dM ]
factorizes through Hc (M × M \∆) so that Λf = 0.
c) From now on suppose that M is orientable. Let B := {ei }i∈I be a graded
basis of H(M ) and let B 0 := {e0i }i∈I denote the Poincaré dual basis of B,
Alberto Arabia
23
i.e. such that ei ∪ e0j = δi,j [ζ], where [ζ] denotes the fundamental class of M .
Using the projection formula for δ : M → M × M show that
X
δ∗ (1) =
(−1)deg(ei ) ei ⊗ e0i ,
i∈I
Z
Prove the equality:
M
δ! (1) ∆ =
X
(−1)k dim H k (M ) .
k∈N
d) Combining () with the last result, show the Lefschetz fixed point formula
X
Λf =
(−1)k Tr f ∗ : H k (M ) → H k (M )
k∈N
In particular, if this alternating sum doesn’t vanish, f has fixed points !
1.12. Conclusion. We have reached the end of the preliminaries on Poincaré
duality and Gysin morphism in the nonequivariant setting. As shown, the key
ingredient is Poincaré duality so that, in order to extend the constructions to
G-manifolds, we propose ourselves to follow the same approach. It will therefore
be necessary first to introduce Poincaré pairings and Poincaré duality in the Gequivariant framework. We devote section 4 entirely to this subject. In section 5,
the G-equivariant Gysin morphisms associated with equivariant maps will then
be defined following the same procedures described in 1.10.
2. Equivariant Cohomology Background
1950 Cartan’s ENS Seminar. In May/June 1950, Henri Cartan gave lectures
n◦ 19/20 of the Séminaire Cartan at the Ecole Normale Supérieure de Paris (5 ).
The talks, which concerned a principal fiber bundle p : E → B of base space
a manifold B, and fiber space a compact connected Lie group G (of Lie algebra g := Lie(G)), focused on setting an algebraic framework for the study
of the relationship between the cohomologies of E , B and G, and incorporating characteristic classes through an algebraized approach of the Chern-Weil
homomorphism ch : S(g)G → H(B).
In the first lecture, Cartan introduces the algebraic analog to the universal
fiber bundle IEG of G, the Weil algebra W (g), as an object of the category of
g-differential graded algebras representing the functor “algebraic connections”,
in the same way the classifying space IBG co-represents the G-principal fiber
bundles functor.
Morg -adg (W (g)),Ω(E ))
Weil
restriction to basic subcomplexes
ch : S(g)G → H(B) o
O
Hot(B,IBG)
5 Lecture
Chern
Steenrod
algebraic connexions on Ω(E )
O
?
infinitesimal connections on Ω(E )
G-principal fiber bundle p : E → B
19 on May 15, and lecture 20 in two sessions: May 23 and June 19. The contents
of these lectures where also presented at the Colloque de topologie (espaces fibrés), held at
Brussels on June 5 to 8.
24
Equivariant Poincaré Duality and Gysin Morphisms
It is in his second lecture that Cartan studies different ways to relate the
cohomologies of E , B and G. Of these, the most interesting to us is the construction of a differential graded algebra whose cohomology is that of B, taking
as its main ingredients the de Rham complex of the total space (Ω(E ),d) and
something else related to the Lie group G. For that, Cartan recalls that, through
the pullback p∗ : Ω(B) → Ω(E ), the complex Ω(B) is identified with the subcomplex Ω(E )bas of basic elements of Ω(E ) viewed as a g-differential graded
algebra. But this wouldn’t really help, as Ω(E ) is lost. Instead, Cartan pursues
his previous idea and introduces the complex W (g)⊗Ω(E ) as a candidate for the
‘de Rham complex’ of the topological space IEG × E , and replaces the previous
pullback with the map Ω(B) → W (g) ⊗ Ω(E ), ω 7→ 1 ⊗ p∗ (ω), the image of
which lies in the subcomplex of basic elements of W (g) ⊗ Ω(E ), denoted by
Ωg (E ) := (W (g) ⊗ Ω(E ))bas .
Cartan then states that the resulting map Ω(B) → Ωg (E ), which is a homomorphism of differential graded algebras, induces an isomorphism in cohomology,
and, moreover, that one has
Ωg (E ) = ((S(g) ⊗ Ω(E ))g ,dg )
(∗)
with
dg (P ⊗ ω) = P ⊗ dω +
X
i
P ei ⊗ c(ei )ω,
(∗∗)
where d is the differential in Ω(E ), {ei } is a basis of g of dual basis {ei }, and
c(ei ) is the contraction operator associated with the vector field generated by
the infinitesimal action of ei on E (6 ). The differential graded algebra Ωg (E ),
nowadays commonly known as the Cartan’s complex , met Cartan’s requirements
perfectly.
At this point, it is worth emphasizing that the construction of Ωg (M ) made
sense whether or not the action of G on M is free. Although clear, this was out of
focus at the time of Cartan’s lectures, where research was mostly concentrated
on manifolds and principal fiber bundles rather than on general G-manifolds,
and still less on general topological G-spaces.
1960 Borel’s IAS Seminar. Some years later, in 1958-59, Armand Borel, who
had been an active participant in the Cartan Seminar and in the Leray’s courses
at the Collège de France since his arrival in Paris in 1949, held his Seminar on
transformation groups at the Institute for Advanced Study in Princeton ([B2 ]).
There, Borel drew attention to the advantages of considering for any locally
compact G-space X, the orbit space of IEG × X under the diagonal action of G:
XG := IEG ×G X ,
as the homotopically best-suited substitute for the orbit space X/G in whatever
way the group G (continuously) acts on X ! It was best-suited, mainly because
the space XG bundles together not only the space X and the group G, but also
orbits and classifying spaces.
6 All
these statements, that are not difficult to prove, were given without any justification
by Cartan. Later Michel André in his Ph.D. thesis ([A], 1962) directed by Claude Chevalley
gave complete proofs for Cartan’s lectures statements.
25
Alberto Arabia
Indeed, XG appears as the total space of the following two maps
7 7 IBG
[X]
XG := IEG ×G X
p1
p2
[IBGx ]
''
X/G,
• p1 : XG IBG , (w,x) 7→ w, a locally trivial fibration of fiber space X, and
• p2 : XG X/G, (w,x) 7→ x, where the fibers are the classifying spaces
IBGx of the different stability groups Gx for x ∈ X.
As Borel says in its introduction: It allows us to tie together the cohomology
groups of X, X/G, and the fixed point set F , with those of the classifying spaces
of the stability groups and of G.
The space XG , which Borel called twisted product, is known today as the
homotopy quotient, the homotopy orbit space or more frequently the Borel construction.
Beyond its immediate aim of the homological study of the set of fixed points
F := X G , the seminar laid most of the foundations of what would later be
known as the equivariant cohomology of locally compact G-spaces. Orbit types,
slice theorems, spectral sequences, fixed points theorems, were already present,
if not yet in their final form, at least at a level that would appeal to other
mathematicians for further development.
1968 Atiyah-Segal: Equivariant K-theory. The restriction map
H(XG ) → H(FG )
(†)
appears in almost every section of applications of the Borel Seminar. Conditions
are often given to ensure it is an isomorphism, but they are quite restricting. In
addition, in the case of the circle group G := T 1 action, although it is clear that
Borel was aware of the fact that (†) is an isomorphism modulo H(IBG )-torsion,
he never states it in these terms. It is only ten years later, in 1968, when Atiyah
and Segal introduce the equivariant K-theory for locally compact G-spaces that
the following enhancement of (∗) appears for the first time.
Localization theorem ([AS, S], 1968). The localized restriction map
KG (X)P → KG (G.X S )P
(‡)
where P is a prime ideal of KG (•), S is minimal among the subgroups of G
such that P is the inverse image of a prime ideal of KS (•), and X S is the set of
S-fixed points, is an isomorphism.
1971 Quillen: Equivariant cohomology. As K-theory has a cohomological
behavior, the equivariant cohomology theory soon came to light. Daniel Quillen
did this in [Q] (1971), when he explicitly merged the ideas of Atiyah-Segal and
Borel. The equivariant cohomology of a G-space X (with coefficients in a ring
A), denoted by HG (X), is thus defined as the ordinary cohomology of Borel’s
construction XG , i.e.
HG (X) := H(XG ;A).
26
Equivariant Poincaré Duality and Gysin Morphisms
Quillen proves the localization theorem for the case where G is an elementary
p-group, and for the case where G is a torus T. The second is stated as:
Theorem. Assume either X is compact or paracompact with dimch (X) < +∞
and that the set of identity components of the isotropy groups of points of X is
finite. Then the inclusion of X T in X induces an isomorphism
HT (X) HT1 (•) − 0 −1 → HT (XT ) HT1 (•) − 0 −1 .
(‡‡)
1975. In this year, Wu-yi Hsiang’s book “Cohomology theory of topological
transformation groups” ([H]) appeared, in which the third chapter promptly
introduces the reader to the foundations of equivariant cohomology for locally
compact G-spaces. It includes a version of the localization theorem more in
the vein of Atiyah-Segal that Hsiang calls the Borel-Atiyah-Segal localization
theorem. It is stated as follows:
Theorem. Assume either X is compact or paracompact with dimch (X) < +∞
and that the set of identity components of the isotropy groups of points of X is
finite. For a multiplicative system S ⊆ HG (•) ( = H(IBG )), put
X S = x ∈ X no element of S maps to zero in H(IBG ) → H(IBGx ) .
Then, the localized restriction map
S −1 HG (X) → S −1 HG (X S ),
is an isomorphism.
For example, if G is a torus T, and S = HT (•)\{0}, one has X S = X F so
that we again see that the kernel and cokernel of the restriction map
HT (X) → HT (X T )
are torsion HT (•)-modules as stated in Quillen’s (‡‡).
Comment. At this point it is worth noting that Hsiang’s introduction to equivariant cohomology suffices for our purpose, which is to introduce equivariant
Poincaré duality, equivariant Gysin morphisms and localisation theorems. In
this regard, we could have chosen to work with singular or sheaf cohomology
with coefficients in arbitrary field and characteristic, and, in many cases, even
in the ring of integers (7 ). But we chose to work with cohomology with real coefficients and to use the Cartan’s complex because those were the choices already
set for the book in which this text was intended as an appendix.
That said, it is worth recalling other significant moments in the development
of equivariant cohomology theory and its applications.
7 To be complete, to work with coefficients in an arbitrary abelian group A, would need
two more ingredients: the equivariant cohomology with compact supports HG,c (X;A) and,
if X is an oriented manifold, the equivariant integration map
Z
: HG,c (X ;A)[dX ] → HG (• ;A).
X
Both are standard concepts associated with the locally trivial fibration p1 : XG IBG . The
first is the well-known cohomology with compact vertical support :
HG,c (X ;A) := H(IRc∗ IRp! (AXG )),
where c : IBG → {•} is the constant map, and the second is the pushforward map, also known
27
Alberto Arabia
1980. Atiyah-Bott, Berline -Vergne: The equivariant differential forms.
The reader may have noticed that the Cartan’s complex played no role in the previous paragraphs. This is because, at the time, Borel, Quillen, Hsiang, . . . where
mostly interested in applying equivariant cohomology (often with coefficients
in fields of positive characteristic) to find conditions for the existence of fixed
points in locally compact G-spaces and to infer cohomological properties of
the fixed point sets from those of the ambient space X (for example being a
cohomological manifold when X is such).
In the early 1980s the whole theory underwent an unexpected development
when N. Berline and M. Vergne succeeded in proving the Duistermaat-Heckman
formula on the pushforward of the Liouville measure on a symplectic manifold
under the moment map ([DH], 1982) by a new fixed point theorem for Gmanifolds inspired by an old paper old paper of Bott ([Bo], 1966).
Let G be a compact Lie group and M a G-manifold. For X ∈ g := Lie(G),
let X ∗ be the vector field over M generated by the infinitesimal action of X, and
denote by c(X) and L(X) respectively the contraction and the Lie derivative
operators associated with X ∗ , acting on the differential algebra of complex de
Rham differential forms (C ⊗ Ω(M ),d). The vector X is called nondegenerate
if, for m ∈ M fixed by the one-parameter group exp(tX), the linear operator
Lm (X) on the tangent space Tm (M ) induced by the Lie derivative L(X), is
invertible. (Notice that this condition implies that Tm (M ) is even-dimensional.)
In their joint work, Berline and Vergne introduce the following linear operator
on C ⊗ Ω(M ):
dX := d − 2πic(X ∗ ).
()
2
It verifies dX
= −2πiL(X), so that if one denotes by Ω(M )X the sub-algebra of
de Rham differential forms on M invariant under exp(tX), the pair (Ω(M )X ,dX )
is a (Z/2Z-graded) differential algebra. Let us denote by HX (M ) its cohomology.
The following fixed point theorem is then proved.
Theorem ([BV1 , BV2 , BV3 ]). Let G be a compact Lie group and M an oriented
compact G-manifold (of even dimension). Then, if X ∈ g is nondegenerate and
µ ∈ HX (M ), one has:
Z
X
µ(m)
µ=
Pf(L
m (X))
M
X
m∈M
where µ(m) is the restriction of µ to the singleton {m}, M X is the fixed point
set (necessarily finite) of exp(tX) and Pf(Lm (X)) is the Pfaffian of Lm (X).
At about the same time the Atiyah-Bott paper ([AB], 1984) appeared. Motivated by the same work of Duistermaat-Heckman, as well as a recent work
of Witten ([W]), it introduced a de Rham model for the equivariant cohomolas Gysin map associated with p1 , and also integration along fibers (in this case X) :
Z
= p1! : HG,c (X ;A)[dX ] → HG (• ;A).
X
This map, by the way, is compatible with the Leray-Serre spectral sequence for p1 . For example, at the IE2 page it’s just the map:
IE2 = H(IBG ;A) ⊗ Hc (X ;A) −→ H(IBG ;A)
Z
α⊗β
7−→
β α.
X
28
Equivariant Poincaré Duality and Gysin Morphisms
ogy of manifolds and states the corresponding localization theorems. In loc.cit.
(th. 4.13) Atiyah-Bott, taking finite dimensional approximations of IEG , shows
that the cohomology of the Cartan’s complex (Ωg (M ),dg ) is the ordinary cohomology of the topological space MG . In this way the original, and somehow
neglected, Cartan’s complex (Ωg (M ),dg ) turned out to have been an excellent
model for the equivariant cohomology of manifolds. The elements of Ωg (M )
have since become known as the G-equivariant (de Rham) differential forms.
When one compares the Berline -Vergne operator dX () to Cartan’s operator dg (p. 24), one immediately understands that the map
evX : Ωg (M ) −→ Ω(M )X
P ⊗ µ 7−→ P (−2πiX)µ
commutes with differentials inducing the map between cohomologies
C ⊗ HG (M )
evX
/ HX (M ).
If T ⊆ G is the torus topologically generated by X ∈ g, we have M X = M T
and the commutative diagram of restrictions to fixed point sets:
C ⊗ HT (M )
∼
·
⊕
C ⊗ HT (M )
/ HX (M )
evX
T
evX
'
/ / HX (M X )
(∗)
where the left vertical arrow is an isomorphism modulo HT -torsion after Quillen.
Now, the proof of the Berline-Vergne fixed-point theorem proves also that the
right vertical arrow in (∗) is a true isomorphism (8 ). As a consequence, the map
evX : C ⊗ HT (M ) → HX (M ) is surjective and the Berline -Vergne fixed point
theorem could also be justified through the Atiyah-Bott’s de Rham
R version of
the localization theorem. Indeed, the equivariant integration map M gives rise
to the commutative diagram
evX
/ HX (M )
C ⊗ HT (M )
R
R
⊕
M
M
/C
HT (•)
evX
with, in the second line, evX (P ) = P (−2πiX). Then, by the localization theorem for HT (M ), we see that for all µ ∈ HX (M ) and every µ̃ ∈ HT (M ) such
that evX (µ̃) = µ, one has:
Z
Z
X
µ(m)
µ=
µ̃ (−2πiX) =
,
Eu
(m,M
)(−2πiX)
T
M
M
T
m∈M
where EuT (m,M ) is the equivariant Euler class of m ∈ M , as introduced by
Atiyah-Bott in (2.19)-loc.cit.
8 This results from the fact that, thanks to the Poincaré lemma for Berline -Vergne cohomology stating that the pullback map HX (M ) → HX (R × M ) is an isomorphism, it is easy
to check that one has a long exact sequence of Berline -Vergne cohomologies:
→ HX,c (M \M X ) → HX (M ) → HX (M X ) →
where HX,c
(M \M X )
= 0, after the original proof of the Berline -Vergne fixed point theorem.
Alberto Arabia
29
The Berline -Vergne and Atiyah-Bott works stimulated renewed interest in
equivariant cohomology, in particular because of its applications to Lie group
representation theory. What happened next goes well beyond the scope of this
work. For interested readers, an excellent account of equivariant cohomology theory for manifolds can be found in chapters 6 and 7 of the book [BGV] (1992), and
for singular spaces in the [GKM] (1998) article, which also reviews the equivariant intersection cohomology following R. Joshua ([Jo], 1987) as well as the
Poincaré duality in equivariant intersection cohomology following J-.L. Brylinski ([Br], 1992). The latter of these happens to have been the original reason for
these notes.
2.1. Category of Cochain g-Complexes
2.1.1. Fields in Use. Unless otherwise stated, Lie groups and Lie algebras,
vector spaces, complexes of vector spaces, linear maps, tensor products and
related stuff, will be defined over the field of real numbers R.
2.1.2. g-modules. Let g be a real Lie algebra. A representation of g, also
called a g-module, will be a real vector space V together with a Lie algebra
homomorphism ρV : g → EndR (V ). For simplicity, the notation “Y ·v” will
frequently replace “ρV (v)” when the representation is understood.
The trivial representation of g on a vector space V , is the one where ρV = 0.
Given g-modules V and W , a g-module morphism from V to W is a linear
map λ : V → W s.t. λ ◦ ρV (Y ) = ρW (Y ) ◦ λ for all Y ∈ g. We denote by
Homg (V,W ) the subspace of HomR (V,W ) of such maps.
A g-module V is said to be:
• simple or irreducible, if it is nonzero and has no nontrivial submodules;
• semisimple, if it is a direct sum of irreducible g-modules;
• reducible if it is a direct sum of two nonzero g-modules;
• completely reducible if it is a direct sum of irreducible modules;
The g-modules and their morphisms constitute a category, the category of
g-modules denoted by Mod(g).
2.1.3. Exercise. Let V be a g-module. Show the equivalence of:
a) V is completely reducible.
b) V is a sum of irreducible modules.
c) If W is a submodule of V then V = V 0 ⊕ W for some submodule V 0 .
2.1.4. Exercise. Given a g-module V , denote by V g the subspace of ginvariant vectors of V , i.e. of v ∈ V , such that Y · v = 0 for all Y ∈ g.
a) Show that for all ϕ ∈ Homg (V,W ), ϕ(V g ) ⊆ W g . Derive the fact that the
correspondence V
V g, ϕ
ϕ V g is fonctorial from Mod(g) intito Vec(R).
b) Endow R with the trivial action of g. Show that the map
Homg (R,V ) → V g ,
ϕ 7→ ϕ(1),
is a natural isomorphism of functors Homg (R, −) → (−)g . In particular,
(−)g is left exact but not necessarily exact.
30
Equivariant Poincaré Duality and Gysin Morphisms
2.1.5. Differential Graded g-Complexes. A differential graded g-complex ,
a g-complex in short, is a quadruple (C,d,θ,c) where:
• (C,d) is a complex in DGM(R) (cf. 1.1.5);
• θ : g → EndGV(R) (C) is a Lie algebra morphism, the g-derivation (9 );
• c : g → MorGV(R) (C,C[−1]) is a linear map, the g-contraction;
such that, for all X,Y ∈ g
i) c(X) ◦ c(Y ) + c(Y ) ◦ c(X) = 0
ii) d ◦ c(X) + c(X) ◦ d = θ(X)
iii) θ(Y ) ◦ c(X) − c(X) ◦ θ(Y ) = c([Y,X])
()
2.1.6. Remark. From ()-(ii), one immediately obtains d ◦ θ(−) = θ(−) ◦ d
which implies that θ naturally induces an action of g on the cohomology of
(C,d). However, that same condition shows that c(X) is a homotopy for θ(X),
so that this induced action is in fact trivial.
2.1.7. Morphisms of g-Complexes. A morphism of graded g-complexes, or
morphism of g-complexes in short, α : (C,d,θ,c) → (D,d,θ,c), is a morphism
of complexes α : (C,d) → (D,d) commuting with derivations and contractions,
i.e. such that α ◦ θ = θ ◦ α and α ◦ c = c ◦ α.
2.1.8. Category of g-Complexes. The g-complexes (C,d,θ,c) and their
morphisms constitute the category of g-complexes denoted by DGM(g,R).
In the sequel, a g-complex (C,d,θ,c) may be denoted by (C,d) and even
simply C, whenever the remaining data are understood.
2.1.9. Split g-Complexes. Given an inclusion of g-modules N ⊆ M , we will
use the notation “N |M ” to express that the natural map
Homg (V,M ) −→ Homg (V,M/N )
(‡)
is surjective for all finite dimensional g-module V .
Exercise. Show that the condition N |M is equivalent to the fact that for every
g-submodule M 0 ⊆ M such that N ⊆ M 0 is of finite codimension, there exists
a g-submodule H ⊆ M 0 such that M 0 = H ⊕ N .
Definition. For a g-complex (C,d), let B i := im(di−1 ) and Z i := ker(di )
respectively be the g-submodules of i-coboundaries and i-cocycles of (C,d). The
g-complex (C,d) will be called g-split whenever one has
B i |Z i |C i ,
for all i ∈ Z.
9 Recall that given two Z-graded vector spaces C and D, we denote by Mor
GV(R) (C,D) the
group of graded homomorphisms of degree zero from C into D. The terminology derivation
comes from the fact that in the main case where (C,d) is the de Rham complex of a Gmanifold, the group G acts on (C,d) by differential graded algebra automorphisms, so that
the infinitesimal action of its Lie algebra g := Lie(G) will be by differential graded algebra
derivations.
Alberto Arabia
31
2.1.10. Lemma. Keep the above notations and prove the following,
M g
Mg
a) If N |M , the natural map g −→
is an isomorphism.
()
N
N
b) The condition B i |Z i is equivalent to the fact that (Z i )g → (Z i /B i ) is surjective, and it is also equivalent to the existence of a g-submodule H i of Z i
such that Z i = B i ⊕ H i , in which case H i is a trivial g-module isomorphic
to Z i /B i .
c) A g-complex (C,d) such that each C i is completely reducible, is g-split.
Proof
a) After 2.1.4, the functor (−)g is isomorphic to Homg (R; −) and the sequence
0 → N g → M g → (M/N )g is left exact. The split condition ensures it is
also right exact.
b) Recall that Hi := Z i /B i is a trivial g-module (see 2.1.6). Following (a),
the split condition immediately gives the surjection (Z i )g (Hi )g = Hi .
Conversely, one clearly has Homg (Hi , −) = HomR (Hi ,(−)g ) and, thereafter,
the commutative diagram
Homg (Hi ,Z i ) −−→ Homg (Hi ,Hi )
HomR (Hi ,(Z i )g ) −− HomR (Hi ,Hi )
where the surjectivity of the second line implies the surjectivity of the first
one. In particular, there exists σ ∈ Homg (Hi ,Z i ) such that π ◦ σ = id
where π : Z i Hi denotes the canonical projection. Setting H i := Im(σ)
completes de proof.
c) Clear from exercise 2.1.3.
2.1.11. Proposition. Let (C,d) be a g-split g-complex.
a) The inclusion C g ⊆ C is a quasi-isomorphism
b) If V is a finite dimensional semi-simple g-module, the inclusions
Vg ⊗C
⊇
V g ⊗ Cg
⊆ (V ⊗ C)g
•
•
g
g
g
HomR (V ,C) ⊇ HomR (V ,C ) ⊆ Hom•g (V,C)
are quasi-isomorphisms.
Proof
a) Immediate from (2.1.10-(a)).
b) Let us first show that if W is a simple g-module different from R, the complexes (W ⊗ C)g and Hom•g (W,C) are acyclic.
It suffices to treat only the Hom• case, since one has
Hom•g (W,C) = Hom•R (W,C)g = (W ∨ ⊗ C)g .
An i-cocycle of Hom•g (W,C) is a g-module morphism λ : W → C i such
that d ◦ λ = 0, i.e. such that im(λ) ⊆ Z i . But the composition of λ with
the surjection Z i Z i /B i is null since g acts trivially on cohomology, so
that in fact im(λ) ⊆ B i . Now, thanks to the fact that Z i |C i , we can lift
λ : W → B i to µ : W → C i−1 and we have thus proved that λ = d ◦ µ, i.e.
that λ is a coboundary.
32
Equivariant Poincaré Duality and Gysin Morphisms
If V is a semisimple g-module, one decomposes V as V g ⊕ W , where W
is a direct sum of simple g-modules different from R. Then
Hom•g (V,C) = Hom•g (V g ,C) ⊕ Hom•g (W,C)
is quasi-isomorphic to Hom•g (V g ,C) after the previous paragraph. But
Hom•g (V g ,C) = Hom•g (V g ,C g ) = Hom•R (V g ,C g ),
so that Hom•R (V g ,C g ) ⊆ Hom•g (V,C) is clearly a quasi-isomorphism.
Finally, that the inclusion Hom•R (V g ,C g ) ⊆ Hom•R (V g ,C) is a quasig
r
isomorphism results from
Q (a) since VQ ' R and the inclusion being considered becomes simply 16i6r C g ⊆ 16i6r C.
2.2. Equivariant Cohomology of g-Complexes
2.2.1. The symmetric Algebra of g ∨ . Let S(g) be the ring of polynomial
maps from g to R, graded by twice the polynomial degree and denote by S d (g)
the subspace of elements of degree d, in particular S 2 (g) = g ∨ and S m (g) = 0
for every odd integer m. Let θ : g → DerR (S(g)) denote the Lie algebra homomorphism induce by coadjoint representation of g on g ∨ .
Fix for later use a vector space basis {ei } of g, of dual basis {ei }.
2.2.2. Cartan Complexes. Given a g-complex (C,d,θ,c), we are interested
in the polynomial maps ω : g 3 Y 7→ ω(Y ) ∈ C, i.e. the elements ω ∈ S(g) ⊗ C.
The Lie algebra g acts on each S a (g) ⊗ C b by the formula
θ(Y )(P ⊗ µ) := θ(Y )(P ) ⊗ µ + P ⊗ θ(Y )(µ),
∀Y ∈ g .
A polynomial map Y 7→ ω(Y ) is then g-invariant if and only if it satisfies the
equality
θ(X)(ω(Y )) + ω([X,Y ]) = 0,
for all X,Y ∈ g. Put
M
Cg := (S(g) ⊗ C)g =
Cgk
(Cg )
k∈Z
where Cgk :=
P
a+b=k
S a (g) ⊗ C
b g
. The S(g)-linear map dg : Cg → Cg ,
dg (1 ⊗ ω) = 1 ⊗ dω +
X
i
ei ⊗ c(ei )ω
is a morphism of graded spaces of degree +1. It verifies dg2 =
that, over Cg , one has
X
dg2 =
ei θ(ei ) ⊗ id.
(dg )
P
ie
i
⊗ θ(ei ), so
i
But Ξ := i ei θ(ei ) is the null operator on S(g). Indeed, since it acts as a
derivation on S(g), it suffices to show that it vanishes on any λ ∈ g ∨ , i.e. that
Ξ(λ)(ej ) = 0 for all j, which comes from the straightforward computation
X
X
Ξ(λ)(ej ) =
ei θ(ei )(λ) (ej ) =
ei (ej )λ([ei ,ej ]) = λ([ej ,ej ]) = 0.
P
i
Hence,
dg2
= 0 in Cg . This dg ∈
i
Endgr1S(g)g (Cg )
is the Cartan differential .
Alberto Arabia
33
2.2.3. Definition. The pair Cg ,dg is a complex. It is the Cartan (equivariant) complex
associated with the g-complex (C,d,θ,c), and the cohomology of
Cg ,dg is its g-equivariant cohomology, denoted in the sequel by
Hg (C) := h Cg ,dg
2.2.4. Important Remark. The graded space Cg is an S(g)g -graded module
(4.1.3), the differential dg is S(g)g -linear, and the cohomology Hg (C) is an
S(g)g -graded module.
2.2.5 Any morphism of g-complexes α : (C,d,θ,c) → (D,d,θ,c) induces a
canonical S(g)-linear morphism of complexes αg : Cg → Dg by the formula
αg = id ⊗ α.
2.2.6. Theorem. With the above notations one has,
a) The correspondence (C,d,θ,c)
(Cg ,d) and α
αg is a covariant functor
from DGM(g,R) into DGM(R).
b) For every g-complex (C,d,θ,c), there exists a spectral sequence converging
to Hg (C) with
g
IE0p,q = S p (g) ⊗ C q , d0 = 1 ⊗ d ⇒ Hgp+q (C).
c) Let G be a compact Lie group, g := Lie(G) and C end D two g-split
g-complexes (2.1.9).
i) The (IE2 ,d2 ) spectral sequence term in (b) is given by
P
IE2p,q = S p (g)g ⊗ H q (C), d2 = i ei ⊗ c(ei ) ⇒ Hgp+q (C).
ii) If H m (C) = 0 for all odd (or for all even) m, then
Hg (C) = S(g)g ⊗ h(C).
iii) If α : C → D is a quasi-isomorphism of g-complexes, αg : Cg → Dg is a
quasi-isomorphism.
d) Let G be a commutative compact Lie group and g := Lie(G),
i) For every g-complex (C,d,θ,c), the subcomplex (C g ,d) is stable under
θ and c, i.e. (C g ,d,θ,c) is a well defined g-complex.
ii) If j : C g ,→ C denotes the inclusion map, jg is a quasi-isomorphism.
iii) The (IE2 ,d2 ) spectral sequence term in (b) is given by
P
IE2p,q = S p (g) ⊗ H q (C g ), d2 = i ei ⊗ c(ei ) ⇒ Hgp+q (C)
iv) If H m ((C)g ) = 0 for all odd (or for all even) m, then
Hg (C) = S(g) ⊗ h(C g ).
v) If α : C g → D g is a quasi-isomorphism, αg is a quasi-isomorphism.
Proof
a) Clear.
34
Equivariant Poincaré Duality and Gysin Morphisms
g
b) For m ∈ Z, let Km = S >m (g) ⊗ C . Each Km is clearly a sub-complex
of (Cg ,dg ) and Cg = K0 ⊇ K1 ⊇ ··· is a regular decreasing filtration of
(Cg ,dg ) (see [Go] §4 pp. 76-) giving rise to the stated spectral sequence.
c) i) The assumption that G is compact ensures that each (finite dimensional)
g-module S p (g) is semisimple.
Proposition 2.1.11-(b) may be used,
and
S p (g) ⊗ C)g ,1 ⊗ d is quasi-isomorphic to (S p (g)g ⊗ C,1 ⊗ d . Consequently (IE0 ,d0 ) in (b) is quasi-isomorphic to S(g)g ⊗ C,1 ⊗ d and
IE1p,q = S p (g)g ⊗ H q (C). But the differential d1 : IE1p,q → IE1p+1,q is
null since the S(g) vanishes in odd degrees, therefore IE1 = IE2 , which
completes the proof of the claim.
ii) Since the differential dr is of total degree 1 and that IErp,q = 0 if p or q is
odd for all r > 2, one has dr = 0 for r > 2, and IE2 = IE∞ .
iii) Follows immediately from (c-i).
d) i) We must check that θ(Y )c(X)C g = 0 for all X,Y ∈ g, but, on C g one
has θ(Y )c(X) = θ(Y )c(X) + c(X)θ(Y ) = c([Y,X]) = c(0) since g is
abelian and from property (iii) of g-complexes (see 2.1.5-()).
ii,iii,iv,v) Left to the reader.
2.2.7. Split G-Complexes. It’s worth noting that the proof of 2.2.6-(c) makes
use of the split condition (2.1.9) only for the finite dimensional sub-g-modules
V ∈ S(g), whose g-module structure is obtained by differentiating its natural
G-module structure.
The split condition 2.1.9 can easily be adapted to the context of G-modules.
For any inclusion of G-modules N ⊆ M one writes “N |M ” whenever the natural
map
HomG (V,M ) −→ HomG (V,M/N )
(‡)
is surjective for all finite dimensional G-module V .
2.2.8. Definition. A complex of G-modules (C,d) is said to be G-split whenever B i |Z i |C i , for all i ∈ Z.
The proof of the following proposition is the same as 2.1.11.
2.2.9. Proposition. Let (C,d) be a G-split complex of G-modules such that
the natural action of G in cohomology is trivial. Then,
a) The inclusion C G ⊆ C is quasi-isomorphism.
b) If V is a semisimple finite dimensional G-module, the inclusions
VG⊗C
Hom•R (V G ,C)
are quasi-isomorphisms.
⊇
V G ⊗ CG
⊇
Hom•R (V G ,C G )
⊆ (V ⊗ C)G
⊆ Hom•G (V,C)
35
Alberto Arabia
3. Equivariant Cohomology of G-Manifolds
3.1. Equivariant Differential Forms
3.1.1. Fields in Use. Unless otherwise stated, manifolds, Lie groups and Lie
algebras, vector spaces, complexes of vector spaces, linear maps, tensor products
and related stuff, will be defined over the field of real numbers R.
3.1.2. G-Derivations and Contractions. Let G be a connected Lie group.
Denote by g := Lie(G) = Te G the Lie algebra of G endowed with the adjoint
action. As in 2.2.1, let S(g) be the ring of polynomial maps from g to R, graded
by twice the polynomial degree.
Let M be a G-manifold. Each Y ∈ g defines a vector field on M by setting
~ (m) := d t 7→ exp(tY ) · m
Y
dt
t=0
~ · ω denote the contraction of the differential form ω ∈ Ω(M ) by the
Let Y
~ . The map c(Y ) : Ω(M ) → Ω(M ), ω 7→ Y
~ · ω, is then an anvector field Y
tiderivation of degree −1 and the map c : g → MorGV(K) (Ω(M ),Ω(M )[−1])
verifies the condition (i) for g-complexes (see 2.1.5-()).
~ , gives a Lie algebra
The Lie derivative with respect to the vector field Y
representation θ : g → EndGV(K) (Ω(M )) by algebra derivations.
Both of the operators θ(Y ) and c(Y ), resp. the G-derivations and the Gcontractions stabilizes the subcomplex of compact support differential forms,
and (Ω(M ),d,θ,c) and (Ωc (M ),d,θc) become g-complexes in the sense of 2.1.5.
3.1.3. Definition. Let G be a compact connected Lie group. The complex
of G-equivariant differential forms, resp. with compact support, of M , are the
following Cartan complexes (2.2.3)
ΩG (M ),dG
:= Ω(M )g ,dg
=
resp. ΩG,c (M ),dG := Ωc (M )g ,dg =
S(g) ⊗ Ω(M )
G
,dg
G
S(g) ⊗ Ωc (M ) ,dg .
Their cohomology, denoted by HG (M ), resp. HG,c (M ), are the G-equivariant
cohomology, resp. with compact support, of M .
In the case where M = {•}, we have HG ({•}) = S(g)G = S(g)g . The notation
“HG ” stands for “HG ({•})”.
The Cartan complexes ΩG (M ),ΩG,c (M ) and the equivariant cohomology
spaces HG (M ) and HG,c (M ) are HG -graded modules (cf. 4.1.3).
3.1.4. Proposition. Let G be a compact connected Lie group.
a) The complexes of G-modules (Ω(M ),d) and (Ωc (M ),d) are G-split (2.2.7).
In particular, if C denotes (Ω(M ),d) or (Ωc (M ),d), the inclusions
S(g)G ⊗ C ⊇ S(g)G ⊗ C G ⊆ (S(g) ⊗ C)G
are quasi-isomorphisms.
b) The correspondence M
(Ω(M ),d,θ,c), f
f ∗ is a contravariant functor
from the category of G-manifolds into the category of G-split g-complexes.
36
Equivariant Poincaré Duality and Gysin Morphisms
c) The correspondence M
(Ωc (M ),d,θ,c), f
f ∗ is a contravariant functor from the category of G-manifolds and proper maps to the category of
G-split g-complexes.
d) There exists a functor on the category G-Man of G-manifolds and Gequivariant maps that assigns to every G-manifold M a spectral sequence
that converges to its equivariant cohomology
p+q
IE2p,q = S p (g)g ⊗ H q (M ) ⇒ HG
(M ).
e) There exists a functor on the category G-Manπ of G-manifolds and Gequivariant proper maps that assigns to every G-manifold M a canonical
spectral sequence that converges to its equivariant cohomology with compact
support
p+q
(M ).
IE2p,q = S p (g)g ⊗ Hcq (M ) ⇒ HG,c
Proof
a) For i ∈ N, the pushforward action of G on Ωi (M ) is defined as g∗ (ω) :=
(g −1 )∗ (ω) for all g ∈ G and ω ∈ Ωi , so that (g1 g2 )∗ = g1∗ ◦ g2∗ .
If V be is a (smooth) finite dimensional representation of G over C, we
make the group G act on Hom(V,Ωi (M )) by the formula
(g·λ)(v) = g∗ λ(g −1 v) , ∀λ ∈ Hom(V,Ωi (M )),
so that λ is a G-module morphism if and only if g·λ = λ. We claim that
there exists a “symmetrization” operator
Σ : Hom(V,Ωi (M )) → Hom(V,Ωi (M ))G
such that Σ2 = Σ and Σ(λ) = λ if and only if λ is a G-module morphism.
Indeed, let λ be a linear map from V to Ωi (M ). For every i-tuple of vector
fields {χ1 ,...,χi } over M and each v ∈ V , the real function
Z
g∗ λ(g −1 v) (x) χ1 (x),...,χi (x) dg ∈ R
M 3 x 7→
G
R
where dg is a G-invariant form of top degree on G, such that 1 = G dg,
is a smooth function because V is finite dimensional, and it depends
linearly on v ∈ V , and multilinearly and antisymmetrically on the χ∗ . We
therefore have an i-differential form which we denote by
Z
Σ(λ)(v) :=
g∗ λ(g −1 v) dg,
(∗)
G
and whose fundamental properties are
• Σ(d ◦ λ) = d ◦ Σ(λ);
• Σ(λ) : V → Ωi (M ) is a G-module morphism;
• Σ(λ) = λ if λ is already a G-module morphism.
We can now resume the proof that Z i (M )|Ωi (M ). Given a G-module
morphism µ ∈ HomG (V,B i+1 (M )), there always exists a linear map λ :
V → Ωi (M ) lifting µ, i.e. such that µ = d ◦ λ, but then one applies the
symmetrization operator Σ and one gets µ = Σ(µ) = Σ(d ◦ λ) = d ◦ Σ(λ),
which shows that the G-module morphism Σ(λ) lifts µ.
37
Alberto Arabia
For Zci (M )|Ωic (M ), note that, since V is finite dimensional, the supports
of the elements in λ(V ) are contained in one and the same compact subset
C ⊆ M , but then the supports of the g∗ (λ(g −1 v)) in (∗) are contained in
G·C which is obviously compact. Therefore, given λ : V → Ωc (M ), one
gets a linear map Σ(V ) : V → Ωc (M ) which is a G-module morphism, and
the preceding arguments apply to the compactly supported case.
To prove that B i (M )|Z i (M ), il suffices, from 2.1.10-(b), to show that
every cocycle is cohomologous to a G-invariant cocycle. But before doing
so, let us recall a general homotopy argument. Given a smooth map ϕ :
R×M → N , if ω ∈ Ωi (N ) the pullback ϕ∗ ω belongs to Ωi (R×M ), i.e. is a
section of the exterior algebra bundle of the cotangent bundle T ∗ (R×M ) of
R×M . Now, the canonical decomposition T ∗ (R×M ) as the direct sum of
cotangent bundles T ∗ (R) ⊕ T ∗ (M ), gives rise to a canonical decomposition
of the i-th exterior power of the cotangent bundle
Vi ∗
Vi
Vi−1 ∗
T (R×M ) = (T ∗ M ) ⊕ T ∗ (R) ⊗
(T M ) .
Consequently, the pullback ϕ∗ (ω) canonically decomposes as
ϕ∗ (ω)(t,x) = α(t,x) + dt ∧ β(t,x),
where α (resp. β) is a section of the vector bundle
Vi−1 ∗
T (M )) over the base space R×M .
Vi
T ∗ (M ) (resp.
When ω is in addition a cocycle, so is ϕ∗ (ω) and, in view of the previous
decomposition, this amounts to the following two conditions
dα(t,x) = 0,
∂
α(t,x) = dβ(t,x),
∂t
where d is the coboundary in Ω(M ) (t is then assumed constant). In particular, if ϕt : M → N denotes the map x 7→ ϕ(t,x), we get
ϕ∗t (ω) − ϕ∗0 (ω) = α(t) − α(0)
Z t
Z t
Z t
∂
=
α(t)dt =
dβ(t)dt = d
β(t)dt ,
0 ∂t
0
0
(∗∗)
and the cocycles ϕ∗t (ω) are all cohomologous to ϕ∗0 (ω).
At this point itR is worth noting that this process also gives a canonical
1
element $(x) = 0 β(t,x) dt ∈ Ωi−1 (M ), depending on ω and such that
ϕ∗1 (ω) − ϕ∗0 (ω) = d$.
Under the hypothesis of our proposition, a first consequence of these notes,
is that if ω ∈ Z i (M ) then g ∗ ω is cohomologous to ω for all g ∈ G. Indeed,
since G is connected, there is a smooth path γ : R → G such that γ(0) =
e and γ(1) = g, and then taking ϕ : R×M → M , (t,x) 7→ γ(t)·x, one
concludes that g ∗ ω = γ1∗ (ω) ∼ γ0∗ (ω) = ω.
More generally, given any diffeomorphism φ : RdG → G onto an open
subset U ⊆ G, one defines a smooth multiplicative action of R over U by
setting t ? g := φ(t · φ−1 (g)) for all t ∈ R and g ∈ U , and considers, for each
38
Equivariant Poincaré Duality and Gysin Morphisms
g ∈ U , the map ϕg : R×M → M , ϕg (t,x) = (t ? g)x. After that, if ω is a
cocycle of Ωi (M ) we will have
Z 1
g ∗ ω − g0∗ ω = d
β(t,g)dt ,
(∗∗∗)
0
with g0 := φ(0) and where β(t,g) denotes a family of elements of Ωi−1 (M )
depending smoothly on (t,g) ∈ R×U , i.e. for any (i − 1)-tuple (χ1 ,...,χi−1 )
of vector fields over M , the following map is smooth:
R×U ×M 3 (t,g,x) 7→ β(t,g,x)(χ1 (x),...,χi−1 (x)) ∈ R.
We now come to a key point. If in addition, one has a compactly supported
function ρ : U → R, then, for any top degree form dg on G, one has
Z
Z
Z
ρ(g)g ∗ ωdg =
ρ(g) g ∗ ω − g0∗ ω dg +
ρ(g)dg g0∗ ω
G
G
=d
G
Z Z
1
ρ(g)β(t,g)dg +
G 0
Z
G
ρ(g)dg g0∗ ω
R R1
where G 0 ρ(g)β(t,g)dg is a smooth differential form over M . But, as we
already show that g0∗ ω ∼ ω, since G is connected, we may conclude that
Z
Z
∗
ρ(g)dg ω,
ρ(g)g ωdg ∼
G
G
something that is satisfied by any compactly supported function ρ : G → R
whose support is contained in any open subset of M diffeomorphic to RdG .
If we now make use of the fact that G is compact (which we haven’t
done
R
so far), we can choose the form dg to be G-invariant such that G dg = 1,
and we can fix a smooth partition of unity {ρi } subordinate to a finite good
cover (cf. (2 )) of G. Then
Z
Z
Z
P
P
∗
∗
Σ(ω) :=
g ωdg =
ρi (g)g ωdg =
ρi (g)g ∗ ωdg
G
G i
∼
i
PZ
i
ρi (g)dg ω =
G
G
Z P
ρi (g)dg ω = ω
G i
where, obviously, Σ(ω) is a G-invariant cocycle, which completes the poof
that B i (M )|Z i (M ) as G-modules.
If we denote by |−| the support of a differential form, we see in what
precedes that for t ∈ [0,1] and g ∈ G one has
(
|β(t)| ⊆ γ([0,1])·|ω|
in (∗∗)
|ρ(g)β(t,g)| ⊆ ([0,1] ? |ρ|)|ω|
in (∗∗∗)
so that if |ω| is compact, the previous arguments show that Σ(ω) − ω is in
fact the differential of a compactly supported differential form, i.e. we have
also proved that Bci (M )|Zci (M ).
b,c,d) Follow by (a) and 2.2.6 by interchanging g and G, by 2.2.7 and 2.2.9.
Alberto Arabia
39
3.1.5. Exercise and remarks. Show that the conclusion in 3.1.4-(a) does
not change if we weaken the connectedness hypothesis of G to simply require
the action of G on C to be homotopically trivial. Show that this arrives in
particular when, G being connected, one is interested in HH (M ) where H
is a closed subgroup of G, connected or not. In that case, if H◦ denotes the
connected component of 1 ∈ H , one has HH (M ) = HH◦ (M )W and HH ,c (M ) =
HH◦ ,c (M )W , where W = H /H◦ .
3.2. The Borel Construction
3.2.1. The Classifying Space. Let G be a compact connected Lie group
and IEG a universal fiber bundle for G. Recall that this topological space is the
limit of an inductive system in the category of (right) G-manifolds {IEG (n) →
IEG (n + 1)}n∈N , where IEG (n) is compact, connected, oriented, n-acyclic and,
moreover, the action of G on IEG (n) is free. A classifying space of G is then the
quotient manifold IBG = IEG /G, limit of the inductive system in the category of
manifolds {IBG (n) → IBG (n + 1)} where each IBG (n) := IEG (n)/G is compact,
simply connected since G is connected, and oriented.
3.2.2 Given a G-manifold M , the quotient M /G may lack good differentiability
properties since the action of G is not, in general, a free action. A key idea to
deal with this issue, dating to the 1950s, is to replace the G-manifold M by the
product IEG × M endowed with the diagonal action of G, g·(e,x) := (eg −1 ,gx).
Now, because IEG is “contractible”, the topological space IEG × M has the same
homotopy type as M and moreover has the advantage that G acts freely on it.
The quotient space is denoted, following Armand Borel (10 ):
MG := (IEG × M )/G
The natural fibration of fiber M :
π
MG := IEG ×G M −−M IEG /G =: IBG
[e,x]
7 −−→ [e]
establishes an important link between the three spaces M ,MG ,IBG . Finally,
if f : M → N is a G-equivariant map, the induced map fG : MG → NG ,
[e,m] 7→ [e,f (m)], is well defined and the diagram
fG
M
G −−−−→ NG
y πN
πM y
IBG ====== IBG
is clearly commutative.
3.2.3. Definition. The functor M
MG , f
fG , from the category of Gmanifolds to the category of fiber spaces over the classifying space IBG , is called
the Borel construction
3.2.4 Although the topological space MG is not a manifold, it is the limit of
an inductive system of such. Indeed, for each n ∈ N, since the compact group
G acts freely on the product manifold IEG (n) × M , the topological quotient
10 Confer §3 of chapter IV in [B ], especially the remark §3.9, reproduced at the end of these
2
notes, where Borel cites previous works of Conner and of Shapiro using this construction in
some special cases.
40
Equivariant Poincaré Duality and Gysin Morphisms
MG (n) = IEG (n) ×G M has a natural manifold structure, canonically oriented
whenever M is so. One gets an inductive system in the category of manifolds
{µn : MG (n) → MG (n + 1)}n∈N with MG = lim MG (n), and even an inductive
−→
system in the category of fibrations with fiber M and compact base spaces
µn
+ 1) 99999K ···
99999K MG
(n) −−−−→ MG (n
y
πM ,n+1
πM ,n y
MG
πM y
99999K IBG (n) −−−−→ IBG (n + 1) 99999K ···
IBG
βn
giving rise to the projective system of de Rham complexes Rham
∗
µ∗
n
Ω (MG (n + 1)) −−→
Ω∗ (MG (n)) n∈N
and the projective system of the Rham cohomology
d
H d (µ∗
n)
H (MG (n + 1)) −−−−−
→ H d (MG (n))
n∈N
for each d ∈ N, which has the remarkable property that, for a given d, the
system is stationary, i.e. H d (µ∗n ) is bijective for sufficiently large n.
The same remarks hold for the compact support case since the maps µn are
proper. One then has the projective system of de Rham complexes
∗
µ∗
n
Ωc (MG (n + 1)) −−→
Ω∗c (MG (n)) n∈N
and; for each d ∈ N, the stationary projective systems of the Rham cohomology
d
H d (µ∗
n)
Hc (MG (n + 1)) −−−−−
→ Hcd (MG (n)) n∈N .
3.2.5. Remark. One can show that in both cases H d (µ∗n ) is bijective for all
n > d + 1. The projective limit of {H d (MG (n))}n∈N identifies then canonically with the d-th singular cohomology H d (MG ;R), and the projective limit
of {Hcd (MG (n))}n∈N with the d-th singular cohomology of vertical compact
d
support Hc.v.
(MG ;R). Using Cartan’s results in [Ca1 , Ca2 ] one obtains canonical isomorphisms
HG (M ) ' H(MG ;R)
and HG,c (M ) ' Hc.v. (MG ;R).
3.2.6. Serre Spectral Sequences. The fibrations πM ,n in 3.2.4 are Serre
fibrations and as such, give rise to a projective system of spectral sequences
( p,q
IE2 (MG (n)) := H p (IBG (n)) ⊗ H q (M ) ⇒ H p+q (MG (n))
p,q
IEc,2
(MG (n)) := H p (IBG (n)) ⊗ Hcq (M ) ⇒ Hcp+q (MG (n))
whose limits are the (Serre) spectral sequence associated with πM .
( p,q
IE2 (MG ) := H p (IBG ) ⊗ H q (M )
p,q
IEc,2
(MG ) := H p (IBG ) ⊗ Hcq (M )
(IE(MG ))
3.2.7. Proposition. The Serre spectral sequences (IE(MG )) associated with
the fibration πM : MG → IBG canonically identifies with the spectral sequences
already met in 3.1.4-(d).
Proof. Implicit in [Ca1 , Ca2 ].
Alberto Arabia
41
3.2.8. Exercise. Let f : M → N be a G-equivariant map between oriented
G-manifolds. For each n ∈ N, as in 3.2.4, denote by fG (n) : MG (n) → NG (n)
the corresponding induced map over IBG (n).
a) Show that the following diagrams are cartesian with µ(n) and ν(n) proper.
µ(n)
MG
+ 1)
(n) −−−→ MG (n
fG (n)y
y fG (n+1)
ν(n)
NG (n) −−−→ NG (n + 1)
b) Prove the following equalities
ν(n)∗ ◦ fG (n + 1)∗ = fG (n)∗ ◦ µ(n)∗
fG (n + 1)∗ ◦ ν(n)! = µ(n)! ◦ fG (n)∗
c) When f : M → N is moreover a closed embedding, one defines the equivariant cohomology with support in M by
HG,M (N ) := HMG (NG ).
Show that there exists a convergent spectral sequence (IEM ⊆N ,r ,dr )
p,q
q
p+q
p
IEM
⊆N ,2 := H (IBG ) ⊗ HM (N ) ⇒ HG,M (N ).
4. Equivariant Poincaré Duality
4.1. Differential Graded Modules over a Graded Algebra
4.1.1. Graded Algebras. A graded algebra is a graded vector space A ∈
GV(R) with a family of bilinear maps · : Aa × Ab → Aa+b such that the triple
(A,0,+, ·) is an R-algebra.
4.1.2. Examples
a) For a graded vector space N ∈ GV(R), the space of graded endomorphisms
(End•R (N ),0,+,id,◦) (1.1.3) is a noncommutative graded algebra.
b) HG = S(g)g is a positively and evenly graded commutative algebra.
c) Ω(M ) and ΩG (M ) are positively graded anticommutative algebras.
d) Ωc (M ) and ΩG,c (M ) are positively graded anticommutative algebras, with
no unit element whenever M is not compact.
4.1.3. Graded Modules. An HG -graded module, HG -gm in short, is a graded
space V ∈ GV(R) together with a homomorphism HG → Endgr0R (V ) of graded
algebras of degree 0. Given two HG -gm’s V and W , a graded homomorphism of
HG -gm’s of degree d from V to W is a graded homomorphism of graded spaces
α : V → W of degree d (1.1.3), which is compatible with the action of HG , i.e.
α(P ·v) = P · α(v) for all P ∈ HG and v ∈ V . We denote by HomgrdHG (V ,W )
the space of such homomorphisms and by
Hom•HG (V ,W ) = HomgrdHG (V ,W ) d∈Z
the graded space of graded homomorphisms of HG -gm’s.
When d = 0, we may write HomgrHG (V ,W ) instead of Homgr0HG (V ,W ).
42
Equivariant Poincaré Duality and Gysin Morphisms
4.1.4. Example. Examples 4.1.2-(c,d) are examples of HG -graded modules.
4.1.5 The category GM(HG ) of HG -graded modules is the category whose objects are the HG -gm and whose morphisms are the graded homomorphisms of
degree 0. We will equivalently write MorGM(HG ) (V ,W ) and HomgrHG (V ,W )
the set of morphisms from V to W .
L
4.1.6 A direct sum
HG [ma ], with ma ∈ Z, is called a free HG -graded module.
a∈A
4.1.7. Proposition
a) An object V ∈ GM(HG ) is projective (resp. injective) if and only if the
functor Hom• (V , −) : GM(HG )
GM(HG ) (resp. Hom• (−,V )) is exact.
b) The category GM(HG ) is an abelian category with enough injective and
projective objects. The cohomological dimension of GM(HG ) is finite and
equals the rank of G.
Proof. (a) is an immediate consequence of the direct decomposition of functors
M
M
Hom•HG (−, −) =
HomgrHG (−, −[m]) =
HomgrHG (−[−m], −).
m∈Z
m∈Z
(b) – For the injectivity properties, let {va }a∈A be a family of homogeneous
generators for V ∈ GM(HG ) and consider, for each a ∈ A, the map γa :
HG [−da ] → V , x 7→ xva which is clearly a morphism in GM(HG ). The sum
X
M
γa :
HG [−da ] V
()
a∈A
a∈A
represents V as the quotient in GM(HG ) of a free, and thus projective, HG -gm.
– For the injectivity properties we reproduce the proof of theorem 1.2.2 in [Go]
§1.4 in the context of graded rings.
b := Hom• (V ,(Q/Z)[0])
()
V
The correspondence V
Z
is an additive contravariant functor from the category of left (resp. right) HG -gm
to the category of right (resp. left) HG -gm (11 ), and is exact, by (a), since
HomgrZ (−,(Q/Z)[0]) = HomZ ((−)0 ,Q/Z)
and since Q/Z is an injective Z-module.
b , v 7→ (γ 7→ γ(v)) is an injective morphism.
Lemma 1. The map ν(V ) : V → V
Proof of lemma 1. Because ν(V ) is clearly a morphism of graded modules,
it is injective if and only if it doesn’t kill any homogeneous nonzero element.
If 0 6= v ∈ V d , the subgroup Z·v ⊆ V d is isomorphic to some Z/nZ for
n 6= ±1, and there exists a nonzero homomorphism γ 00 : Z·v → Q/Z (exercise),
restriction of some γ 0 : V d → Q/Z (thanks to the injectivity of Q/Z). Extend
this γ 0 to the whole of V , assigning zero on the homogeneous factors V e when
e 6= d. This last extension, denoted by γ : V → Q/Z, is a graded morphism
of degree −d and verifies ν(V )(v)(γ) = γ(v) 6= 0 by construction, so that
ν(V )(v) 6= 0, which completes the proof of lemma 1.
N is a right HG -gm, the structure of left HG -module of Hom•Z (N ,(Q/Z)[0]) is given
by (x·γ)(y) := γ(yx) for all x ∈ HG and y ∈ N . If N is a left HG -gm, the structure of right
HG -module of Hom•Z (N ,(Q/Z)[0]) is given by (γ ·x)(y) := γ(xy) for all x ∈ HG and y ∈ N .
11 If
43
Alberto Arabia
b is injective.
Lemma 2. For any free right HG -gm F, the left HG -gm F
Proof of lemma 2. We recall (cf. [Bk] Chap. II, §4, Prop. 1) that for any left
HG -dgm N , the maps
Hom•HG (N ,Hom•Z (HG ,(Q/Z)[0])) o
γ
v 7→ (x 7→ ξ(xv)) o
/ Hom• (N ,(Q/Z)[0])
Z
/ v 7→ γ(v)(1)
ξ
are isomorphisms of graded vector spaces each inverse to the other. It follows
that Hom•Z (HG ,(Q/Z)[0])) is an injective left HG -gm if and only if the functor
Hom•Z (−,(Q/Z)[0]) is exact, but this is equivalent, by (a), to the exactness of
the functor HomgrZ (−,(Q/Z)[0]) = HomZ ((−)0 ,Q/Z), which is clear since Q/Z
is an injective Z-module.
b
Now, if V is a left HG -gm, fix some epimorphism of right HG -gm π : F V
b
b is injective and composed
where F is free as in (). The morphism π
b:V →F
b , injective by lemma 1, we get an injective morphism
with ν(V ) : V → V
b of left HG -gm, where F
b is an injective left HG -gm by lemma 2. This
V ,→ F
completes the proof of the existence of enough injective objects in GM(HG ).
The statement about dimch (GM(HT )) results from the fact (Chevalley’s theorem) that HG is a polynomial algebra in rk(G) variables. One may then refer
to Hilbert’s Syzygy Theorem (cf. [J] p. 385, and Ex. 2, p. 387).
4.1.8. Exercise. Let A be a graded R-algebra which is an integral domain.
a) Show that S −1 A, where S denotes the multiplicative system of homogeneous
nonzero elements of A, is an injective object of GM(A). Also, prove that
the canonical inclusion A ,→ S −1 A is an injective envelope for A.
b) Show that when rk(G) > 0, the degrees of a non trivial injective object of
GM(HG ) cannot be bounded below (12 ).
c) Show that if V ∈ GM(HG ) is positively graded, it admits projective resolutions in GM(HG ) all of whose terms are positively graded.
The next two sections are straightforward generalizations of sections 1.1.5 and
1.1.8 from graded vector spaces to HG -graded modules.
4.1.9. Differential Graded Modules. An HG -differential graded module,
HG -dgm in short, is a pair (V ,d) with V ∈ GM(HG ) and d ∈ Endgr1HG (V ),
called differential , is such that (V ,d) is a complex, i.e. d 2 = 0. A morphism
of HG -dgm α : (V ,d) → (V 0 ,d 0 ) is a morphism of HG -gm’s which is also a
morphism of complexes, i.e. d 0 ◦ α = α ◦ d. The HG -dgm’s and their morphisms constitute the category DGM(HG ) of HG -differential graded modules.
The category DGM(HG ) is an abelian category.
4.1.10. The Hom• (−, −) and (− ⊗ −)• Bi-functors. Given two HG -dgm’s
(V ,d) and (V 0 ,d 0 ), we recall the definition of the HG -dgm’s
Hom•HG (V ,V 0 ),D•
and (V ⊗HG V 0 )• ,∆• .
12 A graded space V is said to be bounded below (resp. above), if there exists N ∈ Z such
that V i = 0 for all i < N (resp. i > N ). The graded algebra HG is bounded below by 0.
44
Equivariant Poincaré Duality and Gysin Morphisms
As HG -graded modules they are defined by
m
0
0
Homm
HG (V ,V ) := HomgrHG (V ,V )
m 7→
0 m
0 m
(V ⊗HG V ) := π(V ⊗R V )
where π : V ⊗R V 0 V ⊗HG V 0 , v ⊗ v 0 7→ [v ⊗ v 0 ], is the canonical (graded)
surjection (see remark 4.1.12). The differentials D• and ∆• are:
Dm (f ) = d 0 ◦ f − (−1)m f ◦ d
∆m ([v ⊗ v 0 ]) = [d(v) ⊗ v 0 ] + (−1)|v| [v ⊗ d 0 (v 0 )]
0
where v ⊗ v 0 ∈ V |v| ⊗ V 0|v | and |v| + |v 0 | = m.
The fact that D and ∆ are HG -linear results from the fact that HG is graded
only by even degrees (!).
4.1.11 These constructions are natural w.r.t. each side entry which means that
one has in fact defined two bi-functors
Hom•HG ((−),(−)) : DGM(HG ) × DGM(HG )
DGM(HG )
((−) ⊗HG (−))• : DGM(HG ) × DGM(HG )
DGM(HG )
which are bi-additive and have the usual variances and exactnesses. For example,
the first one is contravariant and left exact on the left entry, and covariant and
left exact on the right entry, while the second one is bi-covariant and right exact.
4.1.12. Remark. Some care must be taken with the tensor product since it
hides some subtleties. A good way to understand it is to note that V ⊗HG V 0 is
the quotient of the graded space V ⊗R V 0 by the subspace W spanned by the
tensors P v ⊗ v 0 − v ⊗ P v 0 with P ∈ HG and (v,v 0 ) ∈ V × V 0 both homogeneous.
One then shows that W is a graded subcomplex of (V ⊗R V 0 ,∆), so that the
canonical surjection π : (V ⊗R V 0 ,∆) (V ⊗R V 0 ,∆)/W is an epimorphism
of graded complexes, therefore inducing over V ⊗HG V 0 a structure of HG -dgm.
Again, a key point is that HG is graded only by even degrees.
4.1.13. The Dual Complex. In section 1.1.11 we introduced the duality
functor Hom•K (−,K) : DGM(K)
DGM(K) and noted that it was an exact
functor (1.1.13). In the framework of HG -dgm’s, the corresponding functor is
the HG -duality functor
Hom•HG (−,HG ) : DGM(HG )
DGM(HG )
which is generally not exact, nor does it respect quasi-isomorphisms.
4.1.14. The Forgetful Functor. If we disregard differentials, HG -dgm’s simply appear as HG -gm’s, and likewise for morphisms. Forgetting the complex
structure gives the forgetful functor o : DGM(HG )
GM(HG ) which is exact
and will often be implicit in some of our considerations.
4.2. Deriving Functors
4.2.1. Deriving Functors Defined on the Category GM(HG ). We have
already shown (4.1.7) that the abelian category GM(HG ) has enough projective
and injective objects (13 ). We will now recall the definition of the left and right
13 See
Grothendieck [Gr], chapter I, Thm. 1.10, p. 135.
45
Alberto Arabia
derived functors associated with an additive functor F : Ab0 → Ab between
abelian categories where Ab0 has enough projective and injective objects.
The left and right derived functors, respectively IL? F : Ab0
K? (Ab) and
IR? F : Ab0
K? (Ab) (14 ) applied to an object V ∈ Ab0 are defined by the
following steps. First, choose an injective and a projective resolution of V ,
d
d
d−1
d
d
0
1
2
0 −−→ V −−→ I 0 −−→
I 1 −−→
I 2 −−→
···
d−2
0
P 0 −−→ V −−→ 0.
··· −−→ P −2 −−→ P −1 −−→
d1
d0
··· , and
I 1 −→
Next, let I ? V stand for the truncated complex 0 → I 0 −→
d−1
d0
P 0 → 0 , and set
P ? V for ··· −→ P −1 −→
?
IL F(V ) := F(P ? V )
(∗)
IR? F(V ) := F(I ? V )
One proves that the complexes (∗) are homotopically independent of the chosen
resolutions so that each canonically defines an object of K∗ (Ab).
As the targets of the derived functors IR? F and IL? F are complexes, one is
interested in their cohomologies. Their classical notations are
(
(IRi F)(V ) := H i (IR? (−))
(ILi F)(−) := H i (IL? F(−)).
It is easily seen from the above definitions that the augmentation morphisms of
complexes : V [0] → I ? and : P? → V [0], give rise to natural morphisms of
complexes F(V [0]) → (IR? F)(V ) and (IL? F)(V ) → F(V [0]), inducing canonical morphisms
F(V ) → (IR0 F)(V )
and
(IL0 F)(V ) → V .
These are isomorphisms whenever F is respectively left and right exact.
4.2.2. Simple Complex Associated with a Double Complex. The category C \ (Ab) of (cochain) complexes of an abelian category Ab is again an
abelian category so that we can look at the category C ?,\ (Ab) := C ? (C \ (Ab))
of (cochain) complexes of C \ (Ab) also called double (cochain) complexes of Ab.
A double complex N ?,\ := (N ?,\ ,δ?,\ ,d?,\ ) ∈ C ?,\ (Ab) is generally represented
as a two dimensional ladder all of whose subdiagrams are commutative.
O
O
O
di−1,j+1
δi−2,j+1
/N
i−1,j+1
O
di,j+1
δi−1,j+1
di−1,j
δi−2,j
/ N i−1,j
O
/ N i−1,j−1
O
dj−2
di+1,j+1
δi,j+1
di,j
δi−1,j
di−1,j−1
δi−2,j−1
/N
O
i,j+1
/ N i,j
O
di+1,j
δi,j
di,j−1
δi−1,j−1
/ N i,j−1
O
dj−2
/ N i+1,j+1δi+1,j+1 /
O
/ N i+1,j
O
δi+1,j
/
di+1,j−1
δi,j−1
/ N i+1,j−1δi+1,j−1 /
O
dj−2
14 K? (Ab) (resp. K (Ab)) is the category of cochain (resp. chain) complexes of Ab whose
?
morphisms are the morphisms of complexes modulo homotopy.
46
Equivariant Poincaré Duality and Gysin Morphisms
The simple (or total) complex associated withN ?,\ is the complex (Tot◦ (N ?,\ ),D◦ ),
where, for all m ∈ Z,
(
L
N i,j+1
O
Totm (N ?,\ ) := m=a+b N i,j
Dm (ni,j ) := di,j (ni,j ) + (−1)j δi,j (ni,j )
di,j
N i,j
(−1)j δi,j
/ N i+1,j
In this way, one obtains an additive exact functor
Tot◦ := C ?,\ (Ab)
C ◦ (Ab).
4.2.3. Spectral Sequences Associated with Double Complexes. The
double complex N ?,\ is said to be of the first quadrant if {(i,j) | N i,j 6= 0} ⊆ N×
N. As explained in [Go] (§4.8, p. 86), one assigns to this kind of double complex,
two regular decreasing filtrations of (Tot◦ (N ?,\ ),D◦ ). The first is relative to the
line \-filtration Tot◦ (N ?,\ )` := Tot◦ (N ?,\≥` ), and the second to the column ?filtration Tot◦ (N ?,\ )c := Tot◦ (N ?≥c,\ ). Each filtration gives rise to a spectral
sequence converging to the cohomology of (Tot◦ N ?,\ ,D◦ ), respectively
( 0 p,q
IE2 := H\p H?q (N ?,\ ) ⇒ H◦p+q (Tot◦ N ?,\ ,D◦ ),
IE2p,q := H?p H\q (N ?,\ ) ⇒ H◦p+q (Tot◦ N ?,\ ,D◦ ),
00
where H? (resp. H\ ) is the cohomology w.r.t. δ? (resp. d\ ).
?
L
(−) Bi-functors. Given two
4.2.4. The IR? Hom•HG (−, −) and (−) ⊗IH
G
HG -graded modules V ,W ∈ GM(HG ), we may consider the four functors
Hom•HG (V , −),
Hom•HG (−,W ),
V ⊗HG (−),
(−) ⊗HG W ,
where the first two are left exact and the other two are right exact.
Now, given projective resolutions P \ (V ) → V , P \ (W ) → W and an injective
resolution W → I ? (W ) in GM(HG ), we have natural morphisms of double
complexes
Hom(P \ (V ),W [0]? ) −→ Hom(P \ (V ),I ? W ) ←− Hom(V [0]\ ,I ? W )
P \ (V ) ⊗ W [0]?
−→ P \ (V ) ⊗ P ? (W ) ←− V [0]\ ⊗ P ? (W )
(15 ) where, to avoid confusion, we omit the indication of the •-grading, giving
rise to canonical morphisms of complexes on HG -gm
Hom(P \ (V ), W ) −→ Tot Hom(P \ (V ), I ? W ) ←− Hom(V , I ? W )
P \ (V ) ⊗ W
−→ Tot P \ (V ) ⊗ P ? (W ) ←− V ⊗ P ? (W )
(‡)
The following proposition is classical (loc. cit.).
4.2.5. Proposition. The morphisms (‡) are quasi-isomorphisms. (16 )
Sketch of the proof. For the first line of (‡), one notes that the morphisms of
complexes are compatible with line and column filtrations of double complexes
of the first quadrant. In the case of
Hom(P \ (V ),W ) → TotHom(P \ (V ),I ? W ),
15 By
16 In
W [0]• we denote the complex satisfying W [0]0 = W and W [0]i = 0 for i 6= 0.
fact they are homotopic equivalences, but we won’t need to be so precise.
47
Alberto Arabia
since for each i ∈ Z the map Hom(P i (V ),W ) → TotHom(P i (V ),I ? W ) is a
quasi-isomorphism, the induced map on the 00IE terms of the associated spectral
sequences (4.2.3) is an isomorphism and we conclude. The case of
TotHom(P \ (V ),I ? W ) ←− Hom(V ,I ? W )
is almost the same except that now we must consider the line filtration and use
the 0IE spectral sequence.
The second line in (‡) is dealt with in the same way after observing that 4.2.3
also applies (dually) to double complexes of the third quadrant.
As a consequence of 4.2.5, in each line of (‡) the complexes represent the
same objet in the derived category D? (GM(HG )). They are classically denoted
L?
by IR? Hom•HG (V ,W ) and V ⊗IH
W . The constructions are natural w.r.t. each
G
entry so that we get two bi-functors
IR? Hom•HG ((−),(−)) : GM(HG ) × GM(HG )
?
L
( ))• : GM(HG ) × GM(HG )
((−) ⊗IH
G −
D? (GM(HG ))
D? (GM(HG ))
()
which are bi-additive and have the usual variances and exactnesses. They clearly
extend the bi-functors in 4.1.11 from GM(HG ) to D? (GM(HG )).
4.2.6. The Ext• and Tor• Bi-functors. Given V ,W ∈ GM(HG ), one defines for i ∈ Z (
•
i
?
Exti,•
HG (V ,W ) := H? IR HomHG (V ,W )
L?
W
Tor•HG ,i (V ,W ) := H?i V ⊗IH
G
Where H?i is the i’th cohomology functor on D? (GL(HG )).
4.2.7. Defining IR? Hom• (−,HG ) on DGM(HG ). We proceed as in 4.2.4
except that we will consider only injective resolutions of HG in GM(HG ). (17 )
Let V := (V ,d) ∈ DGM(HG ). For any N ∈ GM(HG ), we already en•
dowed the HG
-graded module Hom•HG (V ,N ) with a canonical structure of
•
HG -differential graded module (Hom•HG (V ,N ),D• ) (cf. 4.1.11) in such a way
that we obtain a left exact functor
Hom•HG (V , −) : GM(HG )
DGM(HG ).
It follows that the target of the functor IR? Hom•HG (−,HG ) := Hom•HG (−,I ? N )
is the category C ? (DGM(HG )). The functor transforms homotopies to identities,
and respects quasi-isomorphisms, it therefore induces a contravariant functor
IR? Hom•HG (−,HG ) : DGM(HG )
D? (DGM(HG )),
()
17 The good notion of projective resolution for an H -dgm (V ,d) is the one of simultaneous
G
projective resolutions. These are resolutions ··· → P 2 → P 1 → P 0 → V → 0 (∗) in DGM(HT )
where P i is a projective HG -gm’s, and such that the derived sequence ··· → hP 2 → hP 1 →
hP 0 → hV → 0 (∗∗) is a projective resolution in GM(HG ). When the graded space V is
bounded below (cf. (12 )), the complex (V ,d) always admits simultaneous projective resolutions
and the derived functor IR? Hom• ((V ,d),HG ) may as well be defined as Hom• (P \ ,HG ), as
in the case of HG -gm’s, but we won’t use this point of view in these notes.
48
Equivariant Poincaré Duality and Gysin Morphisms
which we will call the derived duality functor . This functor, composed with
the i’th cohomology functor H?i : D? (DGM(HG ))
DGM(HG ) gives the i’th
extension functor
•
i
?
Exti,•
DGM(HG ).
HG (−,HG ) := H? IR HomHG (−,HG ) : DGM(HG )
The family (indexed by ? ∈ Z) of all these functors
•
?
?
Ext?,•
HG (−,HG ) := H IR HomHG (−,HG ) : DGM(HG )
•
DGM?,• (HG
) ()
•
•
•
where DGM?,• (HG
) is the category of ?-graded HG
-dgm (the action of HG
does
?
not affect the ?-degree), constitutes a ∂-functor in K (DGM(HG )).
4.2.8. Spectral Sequences Associated with IR? Hom•HG (−,HG ). In the
last paragraph we defined the derived duality functor () for any HG -dgm (V ,d)
as the complex of HG -dgm’s
IR? Hom•HG ((V ,d),HG ) := Hom•HG ((V ,d),(I ? HG ,δ? )),
(†)
that will be represented as a double complex with lines indexed by ‘•’ and
columns by ‘’?’. The differentials d and δ? commute, d increases de •-degree
and leaves unchanged the ?-degree, while δ? does the opposite.
4.2.9. Proposition Let (V ,d) ∈ DGM(HG ).
a) There exist convergent spectral sequences
0 p,q
p+q
IE
:= H•p (Extq,•
Tot◦ IR? Hom•HG (V ,HG ),D◦
HG (V ,HG )) ⇒ H◦
00 p,q
IE := Extp,q
⇒ H◦p+q Tot◦ IR? Hom•HG (V ,HG ),D◦
HG (hV ,HG ))
b) If V is projective as HG -gm (18 ), then the following morphism of complexes
induced by the augmentation : HG → I ? is a quasi-isomorphism:
()
Hom•HG ((V ,d),HG ) −→ Tot◦ IR? Hom•HG ((V ,d),HG ).
c) If hV is projective as HG -gm, then the following natural morphisms of
HG -gm’s are isomorphisms:
()
Hom•HG (hV ,HG ) −→ Tot◦ Hom•HG (hV ,I ? ) −→ h Tot◦ Hom•HG (V ,I ? )
Proof. (a) By 4.1.7 we can fix an injective resolution HG → (I? HG ,δ? ) of HG gm of finite length, whereby the line •-filtration and the column ?-filtration are
both regular for the total order ‘• + ?’. We have
(0IE0p,? ,d0 ) = (Homgrp (V ,I ∗ HG ),δ? )
(00IE0p,• ,d0 ) = Hom• ((V ,d),I p HG )
and the proposition follows. (b,c) are straightforward consequences of (a).
4.2.10. Proposition. Let (V ,d) be an HG -dgm.
a) For any N ∈ GM(HG ), there exists a natural morphism of HG -modules
ξ(V ,N ) : h(Hom•HG ((V ,d),N )) → Hom•HG (hV ,N )
b) If V and hV are projective (free) HG -gm, then ξ(V ,HG ) is an isomorphism.
18 A
projective HG -gm is always free, cf. in [J] the corollary of theorem 6.21, p. 386.
49
Alberto Arabia
c) Let (V ,d) and (V 0 ,d 0 ) be HG -dgm’s where V and V 0 are projective (free)
HG -gm’s. If α : (V ,d) → (V 0 ,d 0 ) is a quasi-isomorphism of HG -dgm’s, the
following induced morphism of HG -dgm’s is a quasi-isomorphism:
Hom•H (α,HG ) : Hom•H ((V 0 ,d 0 ),HG ) → Hom•H ((V ,d),HG ).
Proof. In order to minimise notations we shall write ‘Hom• ’ for ‘Hom•HG ’.
Let Z ⊆ V , resp. B ⊆ V , denote the HG -graded submodules of cocycles,
resp. coboundaries, of (V ,d), and let N be any HG -graded module.
(a) Applying the functor Hom• (−,N ) to the short exact sequence:
d
0 → Z −→ V −→ B[1] → 0,
(†)
one gets the left exact sequence of HG -complexes
β
α
0 → Hom• (B,N )[−1] −→ Hom• (V ,N ) −→ Hom• (Z,N ),
and the short exact sequence of HG -complexes
β
α
0 → Hom• (B,N )[−1] −→ Hom• (V ,N ) −→ Q • (Z,N ) → 0,
(∗)
•
where Q (Z,N ) denotes the image of β. Note that the left and right complexes
in (∗) have null differentials so that they coincide with their cohomology.
The cohomology sequence associated with (∗) is the long exact sequence
ci−1
a
b
ai+1
c
i
i
i
−→ Homi (B,N ) −→
hi Hom• (V ,N ) −→
Q i (Z,N ) −→
Homi (B,N ) −→ ,
where one easily verifies that ci is the restriction to Q i (V ,N ) of the natural
map Hom• (Z,N ) → Hom• (B,N ) induced by the inclusion B ⊆ Z. In this
way we obtain the exact triangle of HG -graded modules
b
a[+1]
c
hHom• (V ,N ) −→ Q • (Z,N ) −→ Hom• (B,N ) −→ .
(∗∗)
•
On the other hand, if we apply Hom (−,N ) to the short exact sequence
0 → B ⊆ Z → hV → 0,
(††)
we obtain the left exact sequence
b0
c0
0 → Hom• (hV ,N ) −→ Hom• (Z,N ) −→ Hom• (B,N ),
which, joined to (∗∗), gives rise to the following commutative diagram with
exact horizontal lines:
b
/ Q • (V ,N ) c / Hom• (B,N )
hHom• (V ,N )
L
(D)
=
ξ(V ,N )
⊆
0
0
b
c
•
•
•
/ Hom (hV ,N )
/ Hom (Z,N )
/ Hom (B,N )
0
establishing the existence of ξ(V ,N ).
(b) If hV is projective, the connecting morphism c0 is surjective and
IRi Hom•HG (Z,N ) = IRi Hom•HG (B,N ),
•
∀i ≥ 1.
()
•
It follows that ξ(V ,N ) is bijective if and only if Q (V ,N ) = Hom (Z,N ),
which is equivalent, as V is projective, to IR1 Hom•HG (B[1],N ) = 0, and to
IR1 Hom•HG (Z[1],N ) = 0,
thanks to (). Let us prove this equality.
()
50
Equivariant Poincaré Duality and Gysin Morphisms
For each ` ∈ Z, the projectivity of V [`] and the exactness of (†), implies that
IRi Hom•HG (Z[`],N ) ' IRi+1 Hom•HG (B[`+1],N ),
∀i ≥ 1,
and, again by (),
IRi Hom•HG (Z[`],N ) ' IRi+1 Hom•HG (Z[`+1],N ),
∀i ≥ 1,
so that we have, for all m ≥ 1
IR1 Hom•HG (Z[1],N ) ' IR1+m Hom•HG (Z[1+m],N ),
and () follows from tha fact that dimch (GM(HG )) < +∞ (4.1.7-(b)).
(c) Consider the exact triangle in DGM(HG )
α
p1
p2
(V ,d) −→ (V 0 ,d 0 ) −→ (ĉ(α),∆) −→
[+1]
where (ĉ(α),∆) denotes the cone of α, i.e. the HG -gm ĉ(α) := V 0 ⊕ V [1] with
differential ∆(v 0 ,w) := (d 0 v 0 + αω,−dω). Note that h(ĉ(α)) = 0 since by the
universal property of the cone construction, ĉ(α) is acyclic if and only if α is a
quasi-isomorphism. Now, since additive functors respect exactness of triangles
and cones, the morphism Hom•HG (α,HG ) is a quasi-isomorphism if and only if
the complex ĉ(Hom•HG (α,HG )) = Hom•HG (ĉ(α),HG ) is acyclic. This is indeed
the case following (b) because, ĉ(α) and h(ĉ(α)) being both projective HG -gm,
we have h(Hom•HG (ĉ(α),HG )) = Hom•HG (h(ĉ(α)),HG ) = 0.
4.2.11. Remarks. If one disregards the morphism ξ(V ,HG ) in 4.2.10-(a), then
the fact that h(Hom•HG (V ,HG )) and Hom•HG (hV ,HG ) are isomorphic HG gm’s when V and hV are projectives is an immediate result of 4.2.9-(b,c).
The statement 4.2.10-(c) is a straightforward consequence of 4.2.9-(b). Indeed,
it claims that the restriction of the functor Hom•HG (−,HG ) to the full subcategory of HG -dgm’s (V ,d) with V projective (free) as HG -gm is a derived functor,
so that, as such, it preserves quasi-isomorphisms.
4.2.12. Exercise. Prove that V and hV are projective (free) HG -gm if and
only if Z and B are. Hint. Follow the ideas in the proof of 4.2.10-(b).
4.3. Equivariant Integration
Let G be a compact connected Lie group and M an oriented G-manifold of
dimension dM .
R
Extend the R-linear integration map M : Ωc (M ) → R by S(g)-linearity to
the map
Z
: S(g) ⊗ Ωc (M ) −−→ S(g)
()
M
R
P ⊗ω
7 −−→ P M ω
As G acts on S(g) ⊗ Ωc (M ) by g·(PR ⊗ ω) := g·P
R ⊗ g·ω, the above integration
map is G-equivariant since one has M g·ω = M ω, as a consequence of the
connectedness of G (see proof 3.1.4-(a)). Therefore, the restriction of () to the
subspace of G-equivariant differential forms with compact support
G
g
ΩG,c (M ) := S(g) ⊗ Ωc (M ) = S(g) ⊗ Ωc (M )
Alberto Arabia
takes values in HG := S(g)G (3.1.3). We denote this restriction by
Z
: ΩG,c (M ) → HG
51
()
M
and call it the equivariant integration, which is clearly a morphism of HG -graded
modules of degree −dM .
Now, the graded algebra ΩG,c (M
P ) has already been endowed with the differential dG (P ⊗ ω) = P ⊗ dω + i P ei ⊗ c(ei )ω (see 3.1.3, 2.2.2-(dg )), and a
homogeneous equivariant form ζ ∈ ΩdG,c (M ) of total degree d decomposes in a
unique way as a sum
X X
ζ=
Q ⊗ ωQ
0≤i≤d/2
Q∈B(i)
where B(i) denotes a vector space basis of S i (g) and ωQ ∈ Ωd−2degQ (M ). As a
consequence, one easily checks that if ζ is an equivariant coboundary, the terms
Q ⊗ ωQ in the above decomposition such that ωQ ∈ Ωc (M ) is of top
R degree dM
dM
are already coboundaries, i.e. ωQ ∈ Bc (M ), and consequently M ζ = 0. We
have thereby proved the following lemma.
Z
dG (ΩG,c (M )) = 0.
4.3.1. Lemma.
M
Therefore, the equivariant integration () induces a morphism of HG -graded
modules of degree [−dM ] in cohomology:
Z
: HG,c (M ) → HG
()
M
also called equivariant integration.
4.3.2. Equivariant Integration vs. Integration Along Fibers. As we already pointed out in 3.2.7, G-equivariant cohomology is canonically isomorphic to the projective limit of the de Rham cohomologies of the fibered spaces
πn : MG (n) = IEG (n) ×G M IBG (n) (3.2.4). Moreover, for each fixed d ∈ N
the projective system {H d (MG (n))}n is stationary and converges to H d (MG ).
Now, each πM ,n : MG (n) → IBG (n) is a fibration with oriented base space,
whose fiber is the oriented
manifold M . The operation of integration along M
R
is then well defined M : Hc (MG (n))[dM ] → Hc (IBG (n)) (see 1.9.5) and induces
a limit map
πM ,∗ : HG,c (M )[dM ] → H(IBG ) = HG ({•})
Proposition. The map πM ,∗ coincides with the equivariant integration.
Proof. Left to the reader.
4.4. Equivariant Poincaré Pairing
Equivariant integration is what we need to mimic the nonequivariant Poincaré
pairing (1.3) in the equivariant framework.
52
Equivariant Poincaré Duality and Gysin Morphisms
4.4.1 The composition of the HG -bilinear map
R ΩG (M )⊗ΩG,c (M ) → ΩG,c (M ),
α ⊗ β 7→ α ∧ β, with equivariant integration M : ΩG,c (M ) → HG , gives rise to
a nondegenerate pairing
IPG (M ) : ΩG (M ) ⊗ ΩG,c (M ) −→ HG
Z
α⊗β
7−→
α∧β
(IPG )
M
inducing the Poincaré pairing in equivariant cohomology
PG (M ) : HG (M ) ⊗ HG,c (M ) −→ HG
Z
[α] ⊗ [β]
7−→
[α] ∪ [β]
(PG )
M
4.4.2 The left adjoint map associated with IPG (see 1.3) is now the map
IDG (M ) : ΩG (M ) −→ Hom•HG ΩG,c (M ),HG
Z
(IDG )
α
7−→ IDG (M )(α) := β 7→
α∧β
M
and one has, following lemma 4.3.1, for α homogeneous
Z
IDG (−1)dM dG α (β) =
(−1)dM dG α ∧ β
ZM
=
(−1)dM + |α|+1 α ∧ dG β = (−1)|β| IDG (α)(dG β),
M
Hence, following the conventions in 1.1.7 and 4.1.11, IDG (M ) is a morphism of
HG -graded complexes from ΩG (M )[dM ] to Hom•HG (ΩG,c (M ),HG ).
4.4.3 The right adjoint map associated with IPG (see 1.7.1) is the map
0
IDG
(M ) : (ΩG,c (M ),dG ) −→ Hom•HG ΩG (M ),HG ,−D
Z
0
(IDG
)
0
β
7−→ IDG
(M )(β) := α 7→
α∧β
M
which is also a morphism of HG -graded complexes.
0
4.4.4. Exercise. Verify that (IPG ) is a nondegenerate pairing and that IDG
(M )
is a morphism of HG -graded complexes.
4.5. G-Equivariant Poincaré Duality Theorem
4.5.1. Theorem. Let G be a compact connected Lie group, and M an oriented
G-manifold of dimension dM . Then,
a) The HG -graded morphism of complexes
IDG (M ) : ΩG (M )[dM ] −→ Hom•HG ΩG,c (M ),HG
is a quasi-isomorphism.
b) The morphism IDG (M ) induces the “Poincaré morphism in G-equivariant
cohomology” (see 4.2.10-(a))
DG (M ) : HG (M )[dM ] −→ Hom•HG HG,c (M ),HG
53
Alberto Arabia
If HG,c (M ) is a free HG -module, DG (M ) is an isomorphism.
c) There are natural spectral sequence converging to HG (M )[dM ]
( p,q
q
p+q
IE2 (M ) = ExtpHG HG,c (M ),HG
⇒ HG
(M )[dM ]
p
p+q
⊗R Hom•R (Hcq (M ),R) ⇒ HG
IF2p,q (M ) = HG
(M )[dM ]
where, in the first one, q denotes the graded vector space degree.
d) Moreover, if M is of finite type, the HG -graded morphism of complexes
0
IDG
(M ) : ΩG,c (M )[dM ] −→ Hom•HG ΩG (M ),HG
is a quasi-isomorphism, and mutatis mutandis for (b) and (c).
Proof
a) We recall the filtration of the Cartan complex we already used in the proof
of 2.2.6-(b): An equivariant form in (ΩG (M ),dG ) is said to be of index m if
it belongs to the subspace
ΩG (M )m := S >m (g) ⊗ Ω(M ) G .
One easily checks that each ΩG (M )m is stable under the Cartan differential
dG , that ΩG (M ) = ΩG (M )m for all m 6 0 and that one has a decreasing
filtration
ΩG (M ) = ΩG (M )0 ⊇ ΩG (M )1 ⊇ ΩG (M )2 ⊇ ···
(∗)
Furthermore, ΩiG (M ) ∩ ΩG (M )m = 0 whenever m > i, so that (∗) is a
regular filtration (see [Go] §4 pp. 76-).
In a similar way, λ ∈ Hom•HG (ΩG (M ),HG ) is said to be of index m
whenever
G
>a+m
λ S a (g) ⊗ Ωc (M )
⊆ HG
,
∀a ∈ N,
and we denote Hom•HG (ΩG,c (M ),HG )m the subspace of such maps. As before, each of these spaces is is a subcomplex of (Hom•HG (ΩG (M )),D) and
the decreasing filtration
··· ⊇ Hom•HG (ΩG,c ,HG )m ⊇ Hom•HG (ΩG,c ,HG )m+1 ⊇ ···
verifies for each λ homogeneous of degree i
a + dimM + i > degλ (S a (g) ⊗ Ωc (M ))G > a + i,
(∗∗)
∀a ∈ N,
so that (∗∗) is also regular.
An immediate verification shows that IDG (M ) is a morphism of graded
filtered modules, i.e.
IDG (M ) ΩG (M )[dM ]m ⊆ Hom•HG ΩG,c (M ),HG m ,
∀m ∈ Z,
giving rise, therefore, to a morphism between the associated spectral sequences (see [Go], §4 Thm. 4.3.1, p. 80) whose IE0 terms are
(
(S(g) ⊗ Ω(M ))G ,1 ⊗ d [dM ]
Hom•HG S(g) ⊗ Ωc (M ))G ,1 ⊗ d ,HG
54
Equivariant Poincaré Duality and Gysin Morphisms
and which are respectively quasi-isomorphic to
(
HG ⊗ Ω(M ),d [dM ]
Hom•HG HG ⊗ Ω(M ),d ,HG
Indeed, the first one is just 3.1.4-(a), and the second one results from the fact
that, since G is compact, there is a canonical isomorphism S(g) = HG ⊗R H,
where H denotes the (graded) subspace of G-harmonic polynomials of S(g)
(see [Dx], thm. 7.3.5 p. 241, §8 pp. 277-), so that
S(g) ⊗R Ωc (M ) G ,1 ⊗ dM = HG ⊗R H ⊗ Ωc (M ) G ,1 ⊗ dM ,
and the quasi-isomorphisms of 3.1.4-(a)
HG ⊗ Ωc (M ),d ⊇ HG ⊗ Ωc (M )G ,d ⊆ (S(g) ⊗ Ωc (M ))G ,1 ⊗ d
are morphisms of complexes of free HG -graded modules. Consequently, the
induced morphisms on the corresponding HG -dual complexes will still be
quasi-isomorphisms (cf. 4.2.10-(c)).
Putting together these observations, the induced morphism on the IE1
terms of the concerned spectral sequences by IDG (M ), is simply
11⊗D(M )
HG ⊗ H(M )[dM ] −−−−−−→ HG ⊗ Hom•R (Hc (M ),R)
||
HomHG (HG ⊗ Hc (M ),HG )
where one recognizes in 11 ⊗ D(M ) the classical Poincaré duality 1.3.2.
b) This is a straightforward application of proposition 4.2.10 since, as we noted
in the previous paragraphs, ΩG := ΩG (M ) is a free HG -gm.
c) The first spectral sequence IE(M ) is just the 00IE spectral sequence of 4.2.9
converging to the right hand side of (?). On the other hand, the spectral
sequence, IF2p,q (M ), is the one we used in the proof of (a).
d) Left to the reader.
4.5.2. Torsion-freeness, Freeness and Reflexivity. Proposition 4.5.1-(b,d)
shows that the freeness of equivariant cohomology as HG -gm is a sufficient condition for equivariant Poincaré duality to hold. The question then arises whether
some weaker condition could be equivalent to duality. Apart from freeness, two
other properties have been thoroughly study in Allday-Franz-Puppe [AFP].
• Torsion-freeness. An HG -gm V is said to be torsion-free if, for all v ∈ V ,
Ann(v) := {P ∈ HG | P · v = 0} = 0.
The torsion-freeness of equivariant cohomology is clearly a necessary condition for duality as the modules Hom•HG (−,HG ) are torsion-free. It is also
a sufficient condition for the injectivity of the Poincaré morphism (Prop.
5.9 [AFP], see ex. 6.1.3), but it is not for duality as the explicit examples
of Franz-Puppe [FP] (2006) show.
• Reflexivity. An HG -gm V is said to be reflexive if the natural map
V → Hom•HG (Hom•HG (V ,HG ),HG )
is an isomorphism.
For a finite type manifold M , while the reflexivity of HG (M ) and
HG,c (M ) are clearly necessary conditions to duality, the converse, which is
55
Alberto Arabia
also true, is more subtle. The equivalence between duality and reflexivity
has been established in [AFP] (Prop. 5.10) for G abelian, and in Franz [F2 ]
(Cor. 5.1) for general G . (19 )
The following diagram illustrates the relationship between the different kinds
of nontorsions in equivariant cohomology and significant properties of the equivariant Poincaré pairing.
{free} ⊆ {reflexive}
⊆ {torsion-free}
m
m
o
o n
n
Nondegenerate
Perfect
⊆
Poincaré pairing
Poincaré pairing
It is worth noting that in [F1 ] (2015), Franz gives the first known examples
of compact manifolds having reflexive but nonfree equivariant cohomology.
4.6. T-Equivariant Poincaré Duality Theorem
When G is a compact connected torus T = S1 × ··· × S1 , we have:
HT = S(t)
ΩT (M ) = S(t) ⊗R Ω(M )T
ΩT,c (M ) = S(t) ⊗R Ωc (M )T
so that
HomHT (ΩT,c ,HT ) = HomS(t) S(t) ⊗ Ωc (M )T ,S(t)
= HomR Ωc (M )T ,S(t)
= S(t) ⊗R HomR Ωc (M )T ,R
The left adjoint map IDT (M ) associated with the T-equivariant Poincaré pairing IPT (see 4.4.2) identifies naturally to 11 ⊗ ID(M ),
IDT (M )
S(t) ⊗ Ω(M )T [dM ] −1−
−−−−→ S(t) ⊗ HomR Ωc (M )T ,R
1⊗ID(M )
Z
P ⊗α
7 −−−−−−→
P ⊗ β 7→
α∧β
M
and the right adjoint map (see 4.4.3) to
ID 0 (M )
T
S(t) ⊗ Ωc (M )T [dM ] −11⊗I
−−D
−−
→ S(t) ⊗ HomR Ω(M )T ,R
0 −
(M )
Z
P ⊗β
7 −−−−−−→ P ⊗ α 7→
α∧β
M
The following theorem is a particular case of 4.5.1.
4.6.1. Theorem. Let T be a compact connected torus, and M an oriented
T-manifold of dimension dM .
a) The HT -graded morphism of complexes
IDT (M ) : ΩT (M )[dM ] −→ HomHT ΩT,c (M ),HT
is a quasi-isomorphism.
19 The key point is to prove that reflexivity of H (M ) implies that Exti (H (M ),H )=0,
G
G
G
HG
for all i > 0, in which case the proof 4.2.10-(b) applies and duality follows as in 4.5.1-(b).
56
Equivariant Poincaré Duality and Gysin Morphisms
b) The morphism IDT (M ) induces the “Poincaré morphism in T-equivariant
cohomology” (see 4.2.10-(a))
DT (M ) : HT (M )[dM ] −→ HomHT HT,c (M ),HT
If HT,c (M ) is a free HT -module, DT (M ) is an isomorphism.
c) There are natural spectral sequences converging to HT (M )[dM ]
( p,q
q
IE2 (M ) = ExtpHT HT,c (M ),HT
⇒ HTp+q (M )[dM ]
IF2p,q (M ) = HTp ⊗R HomR (Hcq (M ),R) ⇒ HTp+q (M )[dM ]
where, in the first one, q denotes the graded vector space degree.
d) Moreover, if M is of finite type, the HG -graded morphism of complexes
0
IDT
(M ) : ΩT,c (M )[dM ] −→ HomHT ΩT (M ),HT
is a quasi-isomorphism, and mutatis mutandis for (b) and (c).
Proof (a) Since we have the identification IDT (M ) = 11 ⊗ ID(M ), we may
conclude using 2.2.6-(d). Statements (b,c,d) are particular cases of 4.5.1.
4.6.2. Remark. Recall that HT,c (M ) is a free HT -module whenever M has
no odd (or no even) degree cohomology with compact support (2.2.6-(d)-(iv)).
Obviously, though not very interesting, this is also the case when T acts trivially
on M , since then c(Y ) = θ(Y ) = 0, ∀Y ∈ t, and HT,c (M ) = HT ⊗R Hc (M ).
5. Equivariant Gysin Morphism
We now follow the steps in section 1.10 for the construction of the Gysin
morphisms in the equivariant framework.
5.1. G-Equivariant Gysin Morphism in the General Case
5.1.1. Equivariant Finite de Rham Type Coverings. We have already
proved in 1.6 that if G is a compact Lie Group, a G-manifold M is the union
of a countable ascending chain U := {U0 ⊆ U1 ⊆ ···} of G-stable open subsets
of M of finite type.
The following theorem, the equivariant analog of 1.7.8, is a corollary of the
G-equivariant Poincaré duality theorem 4.5.1. The proof is left to the reader.
5.1.2. Theorem. Let G be a compact connected Lie group, and M an oriented
G-manifold of dimension dM . Then,
a) For every filtrant covering U of M by G-stable open subsets, the canonical
map lim
Ω (U ) → ΩG,c (M ) is bijective, and the map
−→U ∈U G,c
0
HomHG (ΩG (U ),HG ),−D
IDG
(U ) : (ΩG,c (M ),dG )[dM ] −→ lim
−→U ∈U
analog to 1.7.8-(b) is a well defined morphism of complexes.
0
b) Moreover, if the open sets in U are of finite type, the map IDG
(U ) is a
quasi-isomorphism.
57
Alberto Arabia
5.1.3 We now closely follow the instructions of section 1.10.2 for the construction
of Gysin morphisms.
Let f : M → N be a G-equivariant map between oriented G-manifolds. To
0
β
∈
G,c (M ) we assign the linear form on ΩG (N ) defined by IDG (f )(β) : α 7→
R Ω
∗
f α ∧ β. In this way we obtain the diagram
M
f∗
ΩG,c (M )[dM ]
/ ΩG,c (N )[dN ]
L
0
IDG
(f )
ID 0 (N )
quasi-iso if N is of
finite type
*
HomHG (ΩG (N ),HG )
0
which may be closed in cohomology whenever N is of finite type, since IDG
(N )
is then a quasi-isomorphism (4.5.1-(d)).
When N is not of finite type, one fixes any equivariant covering U of N made
0
up of open finite type subsets (5.1.1), and replaces IDG
(N ) by the morphism
0
IDG (U ) of theorem 5.1.2. The diagram
f∗
ΩG,c (M )[dM ]
/ ΩG,c (N )[dN ]
L
0
IDG
(U ) (quasi-iso)
0
IDG
(f,U )
*
lim
HomHG (ΩG (U ),HG )
−→U ∈U
0
where IDG
(f,U ) is defined as in 1.10.2, may be closed in cohomology since
0
IDG (U ) is a quasi-isomorphism. The closing arrow
f∗ : HG,c (M )[dM ] → HG,c (N )[dN ]
the equivariant Gysin morphism associated with f , is therefore defined as
0
f∗ := DG
(U )−1 ◦ h(ID(f,U )).
5.1.4. Theorem and definitions. With the above notations,
a) The equality
Z
f ∗ [α] ∪ [β] =
M
Z
[α] ∪ f∗ [β]
()
N
holds for all [α] ∈ HG (N ) and [β] ∈ HG,c (M ).
b) Furthermore, f∗ is a morphism of HG (N )-modules, i.e. the equality, called
the equivariant projection formula,
f∗ f ∗ [α] ∪ [β] = [α] ∪ f∗ ([β])
()
holds for all [α] ∈ HG (N ) and [β] ∈ HG,c (M ).
c) The correspondence
or
(−)∗ : G-Man
GM(HG )
with
M
f
M∗ := HG,c (M )[dM ]
f∗
is a covariant functor. We will refer to it as the equivariant Gysin functor.
58
Equivariant Poincaré Duality and Gysin Morphisms
d) Suppose that M and N are manifolds of finite type. If the pullback morphism f ∗ : HG (N ) → HG (M ) is an isomorphism, then the Gysin morphism
f∗ : HG,c (M )[dM ] → HG,c (N )[dN ] is also.
Proof. (a) Immediate from de definition of the Gysin morphism.
(b) Unlike the proof of the nonequivariant statement 1.7.4-(b), this claim is
no longer a formal consequence of (a) because equivariant cohomology may have
torsion elements, something that doesn’t affect equivariant integration. Instead,
when N is of finite type and since then ID0 (N ) is a quasi-isomorphism, we can
check that the following equality holds at the cochain level,
0
0
IDG
(f )(f ∗ (α) ∪ β) = “ID0 (N )(α ∪ f∗ (β))” = IDG
(f )(β) ◦ µr (α),
(†)
where the central term is there for purely heuristic reasons and where we denote
µr (α) : ΩG (N ) → ΩG (N ) the right multiplication by α, i.e. µr (α)(−) = (−)∪α.
The identification of the left and right terms in (†) is then a straightforward
0
verification from the definition of IDG
(f ). When N is not of finite type, we
0
0
follow the same arguments with IDG (f,U ) instead of IDG
(f ).
∗
(c) is clear. (d) as f : ΩG (N ) → ΩG (M ) is a quasi-isomorphism, the induced
map Hom•HG (ΩG (N ),HG ) → Hom•HG (ΩG (M ),HG ) is also, following 4.2.100
0
(c), and one concludes, since IDG
(M ) and IDG
(N ) are quasi-isomorphisms.
5.1.5. Exercise. Prove the following enhancement of the statement 5.1.4-(d). If
π : V → B is a vector bundle over an oriented manifold B, the map π is of finite
type (1.4.1), and π ∗ : HG (B) → HG (V ) and π∗ : HG,c (V )[dV ] → HG,c (B)[dB ]
and both isomorphisms (cf. 1.4.2-(c)).
5.2. G-Equivariant Gysin Morphism for Proper Maps
Following 1.10.1, let f : M → N be a proper G-equivariant map between
oriented G-manifolds. To α ∈R ΩG (M ) we assign the HG -linear form on ΩG,c (N )
0
defined by IDG
(f )(α) : β 7→ M f ∗ β ∧ α. In this way we obtain the diagram
ΩG (M )[dM ]
f!
/ ΩG (N )[dN ]
L
0
IDG
(f )
)
0
IDG
(N ) (quasi-iso)
ΩG,c (N )∨
0
which may be closed in cohomology because IDG
(N ) is a quasi-isomorphism, as
shown in 4.5.1-(a). The closing arrow:
f! : HG (M )[dM ] → HG (N )[dN ] ,
the equivariant Gysin morphism associated with a proper map f , is therefore
defined as
0
0
f! := DG
(U )−1 ◦ h(IDG
(f )).
59
Alberto Arabia
5.2.1. Theorem and definitions. With the above notations,
a) The equality
Z
f ∗ [β] ∪ [α] =
Z
M
[β] ∪ f! [α]
(??)
N
holds for all [α] ∈ HG (M ) and [β] ∈ HG,c (N ).
b) Furthermore, f! is a morphism of HG,c (N )-modules, the equality, called the
equivariant projection formula for proper maps,
f! f ∗ [β] ∪ [α] = [β] ∪ f! [α]
(???)
holds for all [β] ∈ HG,c (N ) and [α] ∈ HG (M ).
c) The correspondence
f! : G-Manor
π
GM(HG )
with
M
f
M! := HG (M )[dM ]
f!
is a covariant functor. We will refer to it as the equivariant Gysin functor
for proper maps.
d) If the pullback morphism f ∗ : HG,c (N ) → HG,c (M ) is an isomorphism, the
Gysin morphism f! : HG (M )[dM ] → HG (N )[dN ] is also an isomorphism.
e) The natural map φ(−) : HG,c (−)[d− ] → HG (−)[d− ] (1.2.3) is a homomorphism between the two equivariant Gysin functors (−)∗ → (−)! over the
category G-Manor
π , i.e. the diagrams
φ(M )
HG,c (M ) −−−−→ HG (M )
f∗ y
y f!
φ(N )
HG,c (N ) −−−−→ HG (N )
are naturally commutative.
Proof. Same as the proof of 5.1.4, left to the reader.
5.3. Comparison Theorems
The next theorem establishes a close connexion between the nonequivariant
and the equivariant Gysin morphisms. It is a basic tool for the generalization of
known properties of classical Gysin morphisms into the equivariant framework.
5.3.1. Theorem. Let G be a compact connected Lie group and f : M →
N a G-equivariant map between oriented G-manifolds. There exists a natural
morphism of the spectral sequences IF of theorem 4.5.1-(c) converging to the
Gysin morphism f∗ : HG,c (M )[dM ] → HG,c (N )[dN ],
IFc,2 (M ) = HG ⊗ Hc (M )[dM ] ⇒ HG,c (M )[dM ]
1⊗f∗ y
y f∗
IFc,2 (N ) = HG ⊗ Hc (N )[dN ] ⇒ HG,c (M )[dN ]
60
Equivariant Poincaré Duality and Gysin Morphisms
and in the proper case to f! : HG (M )[dM ] → HG (N )[dN ],
IF2 (M ) = HG ⊗ H(M )[dM ] ⇒ HG (M )[dM ]
1⊗f! y
y f!
IF2 (N ) = HG ⊗ H(N )[dN ] ⇒ HG (M )[dN ]
Proof. Clear from the proof of 4.5.1 and the definition of Gysin morphisms.
5.4. Universal Property of the equivariant Gysin Morphism
Proposition Let f : M → N be a G-equivariant map between oriented Gmanifolds.
a) A morphism of complexes ϕ∗ : ΩG,c (M )[dM ] → ΩG,c (N )[dN ] induces the
equivariant Gysin morphism
f∗ : HG,c (M )[dM ] → HG,c (N )[dN ],
if and only if
Z
Z
∗
f α∧β =
α ∧ ϕ∗ β , ∀α ∈ ΩG (N ), ∀β ∈ ΩG,c (M ).
M
N
b) If f is a proper, a morphism of complexes ϕ! : ΩG (M )[dM ] → ΩG (N )[dN ]
induces the equivariant Gysin morphism
f! : HG (M )[dM ] → HG (N )[dN ],
if and only if
Z
Z
f ∗β ∧ α =
β ∧ ϕ! α, ∀α ∈ ΩG (M ), ∀β ∈ ΩG,c (N ).
M
N
5.4.1. Remark. The last proposition is a simple consequence of the definition
of the Gysin morphism. But one must beware that, unlike the nonequivariant
case (1.9.1), it is generally not true that the equivariant Gysin morphism is
characterized by the equality of cohomology classes:
Z
Z
f ∗ [α] ∪ [β] =
[α] ∪ f∗ [β], ∀[α] ∈ HG (N ), ∀[β] ∈ HG,c (M ).
()
M
N
(or (??) for f! in the proper case). For example, the uniqueness of f∗ satisfying
the relation (), results only from the injectivity of the map:
IDG (N ) : HG,c (N ) −−−→ HomHG HG (N ),HG )
Z
[β]
−−−→ [α] →
[α] ∧ [β]
N
a property that is not always satisfied. Indeed, let T be a torus and N a
compact oriented T-manifold without fixed points. We know from the localization theorem, that HT (N ) is a torsion HT -module and consequently that
HomHT (H (N ),HT ) = 0, so that IDG (N ) is null, although HT (N ) 6= 0.
Exercise. Let T = S1 × S1 act on N = S1 by (t,u)(v) = uv.
a) HT = R[X,Y ], HT (N ) = R[Y ], EndHT (HT (N )) = R[Y ].
61
Alberto Arabia
b) For any map f : N → N and any λ ∈ EndHG (HG (N )) one has
Z
Z
f ∗ [α] ∪ [β] =
[α] ∪ λ[β], ∀[α],[β] ∈ HT (N ).
N
N
c) Let N be any oriented G-manifold such that HG,c (N ) is an HG -free module.
Show that condition () (resp. (??) for proper maps) of theorem 5.1.4 (resp.
5.2.1) completely characterizes Gysin morphisms for maps f : M → N .
5.5. Group Restriction
Let G be a compact connected Lie group. For any closed subgroup H ⊆ G,
connected or not, and for any G-manifold M , the canonical projection of Borel
constructions IEG ×H M IEG ×G M which is a locally trivial fibration
with fiber G/H , induces by inverse image the restriction homomorphism of
equivariant cohomology rings
ResG
H : HG (M ) → HH (M ).
At this point, one could react against the possible lack of connectednes of H
in so far as this property has been everywhere required in these notes. However,
a careful examination shows that connectednes is only needed to ensure that the
action of G on M is homotopically trivial, a property that is clearly inherited
by any subgroup H of a connected group G, whether the subgroup is connected
or not (cf. 3.1.5). In that case if H◦ denotes the connected component of 1 in
H and WH = H /H◦ , we have
HH = S(h)H
and HH (M ) = HH◦ (M )WH .
5.5.1. Theorem. For any closed subgroup H ⊆ G and any equivariant map
f : M → N between oriented G-manifolds, the following diagrams of Gysin
morphisms are commutative:
HG (M )
f∗
HH (M )
/ HG (N )
HG,c (M )
ResG
H
ResG
H
f∗
ResG
H
/ HH (N )
f!
/ HG,c (N )
(f is proper)
HH ,c (M )
f!
ResG
H
/ HH ,c (N )
Proof. For a general map f : M → N the diagram of induced maps between
Borel constructions
IEG ×H M f / IEG ×H N
IEG ×G M
π
f
/ IEG ×G M
π
is cartesian and if we endow G/H with an orientation, the integration along
the fibers of π enters in the commutative diagram of complexes:
ΩH ,c (N )
Z
ΩG,c (N )
f∗
L
G/H
f∗
/ ΩH ,c (M )
Z
/ ΩG,c (M )
G/H
62
Equivariant Poincaré Duality and Gysin Morphisms
We may then conclude thanks to 5.4-(a) and that
R
G/H
is adjoint to ResG
H.
The case where f : M → N is proper follows in the same way.
5.6. Explicit Constructions of Equivariant Gysin Morphisms
Although we gave a universal definition for the equivariant Gysin morphism in
the last section, it is worth recalling alternative constructions for some particular
maps where there exist explicit morphisms of Cartan complexes inducing the
Gysin morphism, just as in the nonequivariant case (1.9).
5.6.1. Constant Map. Let M be an oriented G-manifold. The constant map
cM : M → {•} is G-equivariant, HG ({•}) = HT is free and IDG ({•}) is bijective. Therefore, the cohomological adjunction 5.4.1-() uniquely determines
the Gysin morphism and we have, for all β ∈ ΩT,c (M ):
Z
Z
cM ∗ (β) =
1 ∪ cM ∗ [β] =
β.
{•}
M
5.6.2. Equivariant Open Embedding. Let M be an oriented G-manifold.
If U is a G-invariant open set in M , denote by ι : U ⊆ M the injection and
endow U with the induced orientation. One has a natural inclusion of Cartan
complexes ιG : ΩG,c (U ) → ΩG,c (M ), and the elementary equality
Z
Z
ι∗G (α) ∧ β =
α ∧ ιG,∗ (β), ∀α ∈ ΩG (M ), ∀β ∈ ΩG,c (U ),
U
M
shows immediately that the following induced map is the equivariant gysin map:
H(ιG∗ ) : HG,c (U )[dU ] −→ HG,c (M )[dM ].
5.6.3. Equivariant Projection. Given two oriented G-manifolds M ,N , denote by pr : M × N N , the projection (x,y) 7→ y.
The map
ϕ∗
Ωc (M ) ⊗ Ωc (N ) −−→ Ωc (N )
R
ϕ∗
ν⊗µ
7 −−→ M ν µ
is a morphism of HG -gm’s commuting with G-derivations (G is connected), and
with G-contractions since
ϕ∗ c(X)(ν ⊗ µ) = ϕ∗ c(X)(ν) ⊗ µ + (−1)degν ϕ∗ ν ⊗ c(X)(µ)
= (−1)dM ϕ∗ ν ⊗ c(X)(µ) = (−1)dM c(X) ϕ∗ (ν ⊗ µ ,
R
as M ι(X)ν = 0. The morphism ϕ∗ may then be naturally extended to a morG
G
phism of Cartan complexes ϕG∗ : S(g)⊗Ωc (M )⊗Ωc (N ) → S(g)⊗Ωc (N )
satisfying
Z
Z
∗
pr (α)∧β =
α∧ϕG∗ (β), ∀β ∈ ΩG,c (M )×HG ΩG,c (N ), ∀α ∈ ΩG (N ).
M ×N
N
On the other hand, since the natural map Ωc (M ) ⊗ Ωc (N ) → Ωc (M × N ) is
a quasi-isomorphism (Künneth [BT] p. 50), the induced map
(S(g) ⊗ Ωc (M ) ⊗ Ωc (N ))G → (S(g) ⊗ Ωc (M × N ))G = ΩG,c (M × N )
Alberto Arabia
63
is also a quasi-isomorphism and one may conclude that
∗
∗
prG,∗ : HG,c
(M × N )[dM ] −→ HG,c
(N )
induced by ϕG,∗ is the equivariant Gysin map associated with pr.
5.6.4. Equivariant Fiber Bundle. Let (π,V ,B) be an oriented G-equivariant fiber bundle with fiber F. Integration
along fibers (see [BT] I§6 pp. 61-63)
R
gives a morphism of complexes F : Ωc (V )R → Ωc (B) such
R that if ψ : V →
R V
is an isomorphism exchanging fibres,Rthen F ◦ψ ∗ = ψ ∗ ◦ F , consequently F is
G-equivariant. On the other hand, F commutes with the contractions c(X).
Indeed, since these are local operators, it suffices (modulo unit partitions if
necessary) to verify the claim over a trivializing open subset of V , i.e. over
π −1 (U ) for U s.t. π −1 U ∼ F × U , where we are in the case of a projection
already discussed in 5.6.3.
R
R
Now,
R the map F : S(g) ⊗ ΩcR(V ) → S(g) ⊗ Ωc (B), given by F P ⊗ ω :=
P ⊗ F ω restricts naturally to F : ΩG,c (V )[dF ] → ΩG,c (B) as a morphism
R
R
R
of Cartan complexes satisfying V π∗ α ∧ β = B α ∧ F β since it is so in the
nonequivariant case 1.9.5-(∗).
5.6.5. Zero Section of an Equivariant Vector Bundle
The Equivariant Thom Class. Let (π,V ,B) be a G-equivariant oriented
vector bundle. In 5.1.5, we pointed out that the Gysin morphism for compact
supports π∗ : Hc (V )[dF ] → Hc (B) is an isomorphism, so that, in particular:
Hci (V ) = 0,
for all i < dF .
()
5.6.6. Proposition and definition. Assume G is compact and connected.
a) There exist homogeneous G-equivariant cocycles of total degree dF
ΦG = Φ[dF ] + Φ[dF −2] + Φ[dF −4] + ···
G
with Φ[i] ∈ S(g) ⊗ Ωic (V ) where Φ[dF ] ∈ Ωdc F (V )G represents the Thom
class of (B,V ) (see 1.9.7). Two such cocycles are cohomologous. The map
ϕG,∗
(S(g) ⊗ Ωc (B))G −−−−→ (S(g) ⊗ Ωc (V ))G [dF ]
ν
7 −−−−→
π ∗ ν ∧ Φ̃
is a morphism of Cartan complexes, and the same with ‘ Ω’ instead of ‘ Ωc’.
b) The zero section σ : B ,→ V of the vector bundle π : V → B is a proper
G-equivariant map. The equivariant Gysin morphisms
n σ∗ : HG,c (B))[dB ] → HG,c (V )[dV ]
σ! : HG (B))[dB ] → HG (V )[dV ]
are both induced by the morphism of complexes ϕG,∗ of (a).
Proof. (a) Let n = dF . Since G is connected and compact, there exists Φ[n] ∈
Ωnc (V )G representing the Thom class of V . We have
dG (Φ[n] ) = d(Φ[n] ) + c(X)Φ[n] = c(X)Φ[n] ,
where c(X)Φ[n] ∈ (S(g) ⊗ Ωn−1
(V ))G and d c(X)Φ[n] = L(X)Φ[n] = 0. But
c
then c(X)Φ[n] is a coboundary of compact support following () and, again
64
Equivariant Poincaré Duality and Gysin Morphisms
thanks to the connectedness of G, there exists Φ[n−2] ∈ (S(g) ⊗ Ωn−2
(V ))G s.t.
c
[n]
[n−2]
c(X)Φ = dΦ
. The iteration of this procedure, possible because of the
vanishing condition (), leads to the G-equivariant cocycle ΦG . The cohomological uniqueness is proved in a similar way. The fact that ϕG,∗ is compatible
with differentials is obvious as ΦG is a cocycle.
(b) By the universal property of the equivariant Gysin morphisms 5.4, it
suffices to verify the equality
Z
Z
∗
σ α∧β =
α ∧ ϕG,∗ (β),
∀α ∈ ΩG (V ), ∀β ∈ ΩG,c (B).
B
V
∗
Since π : H(B) → H(V ) is an isomorphism, the same is true in equivariant
cohomology following 3.1.4-(d), so that there exists α0 ∈ ΩG (B) s.t. α ∼ π ∗ α0 .
We are thus lead to verify that
Z
Z
0
α ∧β =
π ∗ (α0 ∧ β) ∧ ΦG ,
∀α0 ∈ ΩG (B), ∀β ∈ ΩG,c (B),
B
V
and this follows from the universal
R of the nonequivariant Thom class
R property
(1.9.7) that states that one has: B ω B = V ω ∧ Φ, ∀ω ∈ H(V ).
5.7. Exercises
1) Restate and solve exercise 1.11 in the equivariant framework. In particular:
• If i : B ,→ M is a closed equivariant embedding of oriented G-manifolds,
denote by j : U := M \B ,→ M the complementary open embedding,
and justify the existence of the following triangles where the left arrows
are Gysin morphisms and the right ones are restriction morphisms.
i) The equivariant compact support cohomology triangle
∗
j∗
i
HG,c (U )[dU ] −−→ HG,c (M )[dM ] −−
→ HG,c (B)[dB ] −
−→ .
[+1]
()
ii) The equivariant Gysin triangle
j∗
i!
HG (B)[dB ] −−
→ HG (M )[dM ] −−→ HG (U )[dU ] −
−→
[+1]
()
• In the equivariant version of the Lefschetz fixed point exercise (1.11.2)
you will define the G-equivariant Lefschetz class of f by
dM
LG (f ) := Gr(f )∗ (δ! (1)) ∈ HG
(M ),
R
and its equivariant Lefschetz number ΛG,f := M LG (f ). Prove that
dM
ResG
(M )
1 LG (f ) = L(f ) ∈ H
ΛG,f = Λf
and conclude that the equivariant Lefschetz number coincides with the
nonequivariant one. In particular, if HG (M ) is a torsion module (7.3.1),
the Euler caracteristic of M is zero.
2) Show that if f : B → M is G-equivariant between oriented G-manifolds,
the projective limit (see 3.2.5) of nonequivariant Gysin morphisms
lim f (n)∗ : Hc (BG (n))[dB ] → Hc (MG (n))[dM ]
←−n
is well defined and coincides with the equivariant Gysin morphism
f∗ : HG,c (B)[dB ] → HG,c (M )[dM ].
65
Alberto Arabia
And mutatis mutandis for the proper case.
3) i) Show that in 5.6.5, the restriction of the equivariant Thom class to the
complement of the zero section, is an equivariant coboundary. (Hint: remark that [ΦG ] = σ! (1) and use (1)-()).
ii) (**) Show that the multiplication by [ΦG ] defines a map from HG (B) to
the equivariant cohomology of V with supports in B:
(−) ∧ [ΦG ] : HG (B) → HG,B (V ).
Show next that this map is an isomorphism. (Hint: use the spectral sequence of exercise 3.2.8-(c) to reduce to the nonequivariant case).
iii) (**) Extend (ii) to the case of a closed embedding B ,→ M of oriented
manifolds. (Hint: show that B may be seen as the zero section of a tubular
G-stable neighborhood B and use (and justify) the fact that the restriction map HG,B (M ) = HG,B (B ) is an isomorphism).
6. The field of fractions of HG
6.1. The Localization Functor
Denote by QG the field of fractions of HG . The localization functor is the base
change functor
QG ⊗HG (−) : GM(HG )
Vec(QG )
(20 ) General result of commutative algebra state for any HG -module N , the HG module QG ⊗HG N is flat and injective (as in appendix §9). The localization
functor is exact and when applied to Cartan complexes, we obtain the localized
Cartan complexes
QG ⊗HG ΩG (M ),id ⊗ dG
and QG ⊗HG ΩG,c (M ),id ⊗ dG
whose cohomologies, the localized equivariant cohomologies, respectively denoted QG (M ) and QG,c (M ), satisfy :
QG (M ) = QG ⊗HG HG (M )
and QG,c (M ) = QG ⊗HG HG,c (M ).
6.1.1. Localized Equivariant Poincaré Duality. The localized equivariant
cohomology is very close to the non equivariant cohomology in that the Poincaré
duality pairings are perfect. The following, analog of 4.5.1, simply results from
the fact that QG is a flat and injective HG -module (details are left to the reader).
6.1.2. Theorem. Let G be a compact connected Lie group, and M an oriented
G-manifold of dimension dM . Then,
a) The morphism of (nongraded) complexes
IDG (M ) : QG ⊗HG ΩG (M )[dM ] −→ Hom•QG QG ⊗HG ΩG,c (M ),QG
20 Note that we loose grading in considering this kind of localization. From this point of
view, it would have been more clever to tensor by the ring LG := S −1 HG , where S denotes
the multiplicative system of nonzero homogeneous elements of HG . As appendix 9 explains,
if N is an HG -gm, the HG -module LG ⊗HG (−) is graded, flat and injective, which is what
we really need about localization.
66
Equivariant Poincaré Duality and Gysin Morphisms
induces an isomorphism
DG (M ) : QG (M )[dM ] −→ Hom•QG QG,c (M ),QG
b) Moreover, if M is of finite type, the morphism of complexes
0
IDG
(M ) : QG ⊗HG ΩG,c (M )[dM ] −→ Hom•QG QG ⊗HG ΩG (M ),QG
induces an isomorphism
0
DG
(M ) : QG,c (M )[dM ] −→ Hom•QG QG (M ),QG
6.1.3. Exercise. Let M be of finite type. Prove that the torsion-freeness (4.5.2)
of HG (M ) (resp. HG,c (M )) is a necessary and sufficient condition for
DG (M ) : HG (M )[dM ] → Hom•HG HG,c (M ),HG
0
(resp. DG
(M )) to be injective. Discuss the case where M is not of finite type.
Hint: Let M be HG -gm. Show that the canonical map M → QG ⊗HG M is
injective if and only if M is torsion-free.
Show that if M is also of finite type
the natural map Hom•HG M,HG → Hom•HG M,QG induces an isomorphism
QG ⊗HG Hom•HG M,HG ' Hom•QG QG ⊗QG M,QG . Apply 6.1.2.
6.1.4. Localized Equivariant Gysin Morphisms. As a consequence of theorem 6.1.2, if f : M → N is a map between oriented G-manifolds, the localized
Gysin morphisms
(
f∗ : QG,c (M ) → QG,c (N )
f! : QG (M ) → QG (N ),
if f is proper,
are catacterized, as in the nonequivariant framework, by the adjoint equalities,
Z
Z
∗
f [β] ∪ [α] =
[β] ∪ f∗ [α]
M
Z
M
N
f ∗ [β] ∪ [α] =
Z
[β] ∪ f! [α],
if f is proper.
N
7. Equivariant Euler Classes
The reference for this section is Atiyah-Bott’s paper [AB], notably §2 and §3.
7.1. The Nonequivariant Euler Class
Given a pair of oriented manifolds (N ,M ) with N ⊆ M , we denote by N a
tubular neighborhood of N in M . As the inclusion N ⊆ N has the same nature
as the inclusion of the zero section of a vector bundle σ : B ⊆ V (1.9.7), we
may define the Thom class [Φ(N ,M )] of the pair (N ,M ) following the same
principle, that is, by means of the Gysin morphism associated with the closed
embedding i : N ⊆ M . We thus set :
[Φ(N ,M )] := i! (1) ∈ H dM −dN (M )
67
Alberto Arabia
7.1.1. Definition. The Euler class Eu N ,M of the pair (N ,M ) is the restriction of the Thom class to H(N ) (21 ), i.e. :
()
Eu N ,M := i∗ i! (1) = [Φ(N ,M )] N ∈ H dM −dN (N ).
7.2. G-Equivariant Euler Class
The generalization of the concept of Euler class to the equivariant framework is
straightforward thanks to the equivariant Gysin morphism formalism: Given a
pair of oriented G-manifolds (N ,M ) with N ⊆ M , we denote by iG : N
⊆M
the inclusion map and define the G-equivariant Euler class EuG N ,M of the
pair (N ,M ) by the same formula ():
dM −dN
EuG N ,M := i∗G iG! (1) = [ΦG (N ,M )] N ∈ HG
(N ) .
where iG! : HG (N )[dN ] → HG (M )[dM ] is now the equivariant Gysin morphism.
7.2.1. Exercise. Given oriented G-manifolds L ⊆ N ⊆ M , prove the folowing
formula for nested equivariant Euler classes
EuG L,M = EuG L,N ∪ EuG N ,M L
Hint: Use the projection formula for Gysin morphisms.
7.2.2. G-Equivariant Euler Class of Discrete Fixed Point Sets
In the sequel, we denote by M G the subspace of G-fixed points of M .
When N is a discrete subspace of M G , one has
Y
dM
EuG N ,M ∈ HG
(N ) =
S dM (g)G ,
b∈N
and EuG N ,M is simply the family of invariant polynomials
EuG N ,M = EuG b,M ∈ S dM (g)G b∈N .
7.2.3. Proposition. If N is a finite subset of M G , one has
Z
Z
X
EuG b,M =
ΦG (N ,M )∪ΦG (N ,M ) and |N | =
ΦG (N ,M ).
b∈N
M
M
Proof. The constant function 11N and, a fortiori, the Thom class ΦG (N ,M ),
are both of compact supports. The equalities then immediately follow from the
adjoint property of the Gysin morphism i∗ : HG,c (N ) → HG,c (M ) which gives:
Z
X
i∗ (11N ) ∪ α, ∀α ∈ HG (M ).
αb =
b∈N
M
7.2.4. G- and T-Equivariant Euler Classes of a Fixed Point. Let T
be the maximal torus of the compact connected Lie group G and denote by
T 0 := NG (T) the normalizer of T in G. We have T ⊆ T 0 ⊆ G and if we
choose IEG as universal fiber bundle for any of these groups, we can easily
21 Cf.
formula (2.19), p. 5, in loc.cit.
68
Equivariant Poincaré Duality and Gysin Morphisms
compare the corresponding Borel constructions for a given G-manifold M . In
this way we obtain a natural commutative diagram of locally trivial fibrations:
MT := IEG ×T M
IBT
p
/ / MT 0 := IEG ×T 0 M
/ / IBT 0
p
q
/ / MG := IEG ×G M
/ / IBG
q
()
– The Weyl group of (G,T), i.e. the finite group W := NG (T)/T, acts on
the right of MT and p is the orbit map for this action. In particular, the map
p∗ : HT 0 (M ) → HT (M )W is an isomorphism.
– The fibers of q are isomorphic to G/T 0 which is acyclic in rational cohomology.
Indeed, this space is the orbit space of G/T for the right action of W and we
know from an old result of Leray (22 ) that, under this action, H(G/T) is the
regular representation of W . In particular H(G/T 0 ) = H(G/T)W = K, which
implies that
q ∗ : HG (M ) → HT 0 (M )
is an isomorphism.
This is a consequence of the general fact that if q : X → Y is a locally trivial
fibration with acyclic fiber F between manifolds, then q ∗ : H(Y ) → H(X) is
an isomorphism. Indeed, if U = {U } is an good cover of Y (cf. (2 )) such that
q : f −1 (U ) → U is a trivial fibration for all U ∈ U , then f −1 (U ) = U × F
and the cover f −1 (U ) := {f −1 (U )} will be also good for X. In that case, q ∗
establishes an isomorphism of Čech cohomologies q ∗ : Ȟ(U ;Y ) → Ȟ(f −1 U ;X)
which are known to be canonically isomorphic to de Rham cohomologies. By
this result, the maps q ∗ : H(IEG(m) ×G M ) → H(IEG(m) ×T 0 M ) are bijective
for all finite dimensional approximation IEG(m) of IEG, which suffices to our
purposes as equivariant cohomology is the projective limit of the cohomologies
of these approximations (3.2.5). (23 )
– Summing up, we have the following two canonical isomorphisms
HG (M )
q∗
/ HT 0 (M )
p∗
/ HT (M )W
(‡)
– When M = {•}, we obtain a commutative diagram of Chern-Weil homomorphisms
/ S(t)
⊆
S(g)G Chv / S(t)W
'
HG
'
q∗
'
/ HT 0 /
'
p∗
/ HT
(‡‡)
where Chv : S(g)G → S(t)W is the map that associates a symmetric polynomial function on g with its restriction to the subspace t. The diagram already
shows that Chv is an isomorphism, a claim known as the Chevalley isomorphism
following the celebrated, much more general, Chevalley’s restriction theorem.
22 The statement appears as the Lemma 27.1 in the Ph.D. thesis of A. Borel, defended at
La Sorbonne (with Leray as president) in 1952, ([B3 ], lemme 27.1, p. 193). Borel attributes
the result to J. Leray ([L]).
23 In [AB], p. 4, the interested reader will find partial indications to a different justification,
that seems to rely on [Gt].
69
Alberto Arabia
At this point it is worth noting that for each b ∈ M G the group G acts
naturally on the tangent space Tb (M ) through a linear representation. Now,
if we endow M with a G-invariant riemannian metric, the exponential map
exp : Tb (M ) → M is a G-equivariant diffeomorphism between Tb (M ) and an
open neighborhood of b in M , so that the computation of equivariant Euler
classes on fixed points may be greatly simplified by linearizing the data. The
following proposition deals with the linear case.
7.2.5. Proposition. Let V be a linear representation of a compact connected
Lie group G with maximal torus T.
a) The equivariant Euler class EuT 0,V belongs to S(t)W and the Chevalley
isomorphism Chv : S(g)G → S(t)W exchanges EuG 0,V and EuT 0,V .
b) If V := V1 ⊕ V2 as G-module, then EuG 0,V = EuG 0,V1 EuG 0,V2 .
c) Denote by C(α) the complex vector space C endowed with the representation
of T corresponding to the (nonzero) weight α ∈ t ∨ , i.e. exp(tx)(z) = eitα(x) z,
for all t ∈ R and z ∈ C. If L
the decomposition of V in irreducible representations of T is V = Rµ0 ⊕ α C(α)µ(α) , then
Y
EuT 0,V = 0µ0
αµ(α) .
α
d) EuG 0,V =
6 0 if and only if V T = {0}.
Proof. (a) After the natural isomorphism of functors HG (−) ' HT 0 (−) of (‡),
it suffices to justify the commutativity of the following diagram:
HT 0 (0)
p
∗
HT (0)
/ HT 0 (V )
i!
(I)
i!
p
∗
/ HT (V )
i∗
/ HT 0 (0)
(II)
i
∗
p∗
/ HT (0)
The commutativity of the subdiagram (II) is obvious. For (I) we check its dual,
the diagram
i∗
/ HT (0)
HT,c (V )
p!
HT 0 ,c (V )
where p! =
R
W
(I∨ )
i∗
p!
/ HT 0 (0)
, which is also clearly commutative.
(b,c,d) left to the reader. Hint for (c). Following (b), it suffices to show that
EuT 0,C(α ) = α. Taking polar coordinates (ρ,θ) ∈ R+ × [0,2π] in C, the
nonequivariant Thom class Φ(0,C) is of the form
Φ[2] = λ(ρ) ρdρ ∧ dθ,
where λ : R → R is a nonnegative differential function
R ∞ with compact support
equal to 1 in a neighborhood of 0 and such that 0 λ(ρ) ρdρ = 1/2π. As
it is clear that Φ[2] is invariant under the action of the unit circle action, it is
70
Equivariant Poincaré Duality and Gysin Morphisms
T-invariant. We can thus use this differential 2-form to construct an equivariant
Thom class following the procedure described in the proof of 5.6.6-(a). We have
(dt Φ[2] )(X) = c(X) λ(ρ) ρdρ ∧ dθ = −2πα(X) λ(ρ) ρdρ,
and Φ[0] (X) is necessarily equal to
Z
Z ρ
Φ (ρ,θ)(X) = −2πα(X)
λ(ρ) ρdρ −
+∞
[0]
0
λ(ρ) ρdρ ,
0
since it must be of compact support. In this way we obtain
EuT 0,C(α ) = ΦT (0,C(α)) 0 = Φ[0] (0)(X) = α(X).
7.2.6. Exercise. If G is the special
orthogonal group SO(3) of the euclidean
space R3 , show that EuG 0,R3 = 0. Conclude that isolated G-fixed points
may have a null equivariant Euler class when G is nonabelian, contrary to the
abelian case.
7.3. Torsions in Equivariant Cohomology Modules
7.3.1. Torsions. The annihilator of an element v of an HG -gm V , is the ideal
Ann(v) := {P ∈ HG | P · v = 0}.
One says that v is a torsion element if Ann(v) 6= 0, otherwise v is a torsion-free
element. The HG -gm V is called a torsion module if all its elements are torsion
elements, it is called a torsion-free module if zero is its the only torsion element,
otherwise, it is called a nontorsion module.
7.3.2. Exercise. Given an HG -gm V , let τ (V ) be the subset of its torsion
elements. Show that
1) τ (V ) is a torsion module and the quotient ϕ(V ) := V /τ (V ) is torsion-free.
The natural map: QG ⊗HG V → QG ⊗HG ϕ(V ) is an isomorphism.
2) QG ⊗HG V = 0 if and only if V is torsion.
3) HomgrHG (V ,QG ) = 0 if and only if V is torsion.
4) An inductive limit of torsion modules is a torsion module.
5) A projective limit of torsion modules may be a nontorsion module.
7.3.3. Exercise. The annihilator of an HG -module is the ideal
\
Ann(V ) := {P ∈ HG | P · V = 0} =
Ann(v).
v∈V
1) Show that if Ann(V ) 6= 0, then V is torsion, but the converse may fail. (24 )
2) Show that if V is an HT -algebra with unit element, then Ann(V ) = Ann(1).
3) Let {U1 ⊆ U2 ⊆ ···Un ···} be an increasing
family of G-stable open subsets of
S
a G-manifold M such that M = n Un . Suppose that HG (Un ) and HG,c (Un )
are torsion for all n ∈ N. Show that HG,c (M ) is torsion, whereas HG (M )
may fail to be torsion.
24 Hint:
For P ∈ HG , let W (P ) := HG /(HG · P ) and take V :=
L
P ∈HG W (P ).
71
Alberto Arabia
4) In (3) show that {Ann(HG (Un ))}n is a decreasing sequence of ideals and
that
\
Ann(HG (M )) =
Ann(HG,c (Un )).
n∈N
In particular, if the set {Ann(HG (Un ))} is finite, then HG (M ) is torsion.
7.3.4. The slice theorem. Given a G-manifold M , the slice theorem (25 )
claims that, through every point x ∈ M a submanifold
S(x) passes, a slice through x, such that the map G/Gx
Gx
G/
G ×Gx S(x) → M ,
[(g,x)] 7→ g · x,
Gx
x
where Gx denotes the isotropie group of x, is a diffeomorphism onto a G-stable neighborhood Vx of x. Then
HG (Vx ) = H(IEG ×G G ×Gx S(x)) = HGx (S(x)),
S(x)
M
Vx
as well as HG,c (Vx ) = HGx ,c (Vx ) (exercise). As a consequence, the HG -module structures of HG (Vx ) and HG,c (Vx ) factorise through
the natural ring homomorphism ρx : HG → HGx .
7.3.5. Proposition. Let T be a torus. For every point x in a T-manifold M ,
the following equivalences hold.
a) ρx : HT → HTx is injective if and only if x ∈ M T . The HT -modules HG (Vx )
and HG,c (Vx ) are torsion if and only if x 6∈ M T
b) If x ∈ M T , then EuT x,M 6= 0 if and only if x is an isolated point of M T .
Proof. (a) If Tx 6= T, there exist closed subtorus H ⊆ T such that T = H × Tx
+
and dim(H ) > 0, in which case ker(ρx ) = HH
⊗ HTx 6= 0. (b) is 7.2.5-(d).
7.3.6. Remark. The interesting point of this proposition is that it faithfully
translates topological properties of a point in a T-manifold into algebraic properties of HT -modules, opening the way to the algebraic study of the topology
of T-spaces. When G is no longer abelian, both claims may fail. For (a), if T is
a maximal torus for G and M = G/T, the isotropie group of x = g[T] ∈ M is
the maximal torus Gx = gTg −1 , and ρx is the inclusion HG = (HT )W ⊆ HGx .
Thus, ρx is injective although x is not a G-fixed point. Exercise 7.2.6 gives a
counterexample for (b).
7.3.7. Orbit Type of T-Manifolds. The torsions of the HT -modules
HT,c (M ) and HT (M ) play a central role in the fixed point theorem.
When M T = ∅, the slice theorem and 7.3.5-(a) show that M may be covered by a family of T-stable open subspaces Vx where HT (Vx ) is killed by the
elements of the nontrivial kernel ρx : HT → HTx . But then any finite union of
those subspaces will also have torsion equivariant cohomology thanks to MayerVietoris sequences, and, if M is compact, we can already say that HT (M ) is
torsion. When M is not compact we may not be able to conclude the same
(cf. 7.3.3-(3)) unless we have some kind of finiteness condition on the kernels of
ρx . As shown in exercise 7.3.3-(4), such condition may be the finiteness of the
25 See
Hsiang [H] §I.3, p. 11.
72
Equivariant Poincaré Duality and Gysin Morphisms
set those kernels, or, what amounts to the same, the set of the isotropy groups
OT (M ) := {Tx | x ∈ X} which is called the orbit type of the T-space M (26 ).
Definition. A T-manifold M is said of finite orbit type if OT (M ) is finite.
7.3.8. Exercise. Show that a T-manifold M is always locally of finite orbit
type. In particular, if M is compact, it is of finite orbit type.
Hint. Use the slice theorem. If x 6∈ M T , show that the slice S(x) is a strict
submanifold of M stable under Gx and that OG (G · S(x)) = OGx (S(x)), then
conclude by induction on dim(M ). Otherwise, if x ∈ M T , linearize the action as
in 7.2.5 and conclude showing that there is a one-to-one correspondence between
isotropy groups in the T-space Tx M and subsets of the set of nonzero weights
of the linear representation of T on Tx M .
7.3.9. Proposition. If M T = ∅ and M is of finite orbit type, then
HT,c (M ) ⊗HT QT = HT (M ) ⊗HT QT = 0.
Proof. – Torsion of HT,c (M ). Let (U ,⊆) be the set of G-stable open subspaces
U ⊆ M , such that HT,c (U ) is torsion, partially ordered by set inclusion. The
set U is non empty as it contains every slice neighborhood Vx (7.3.4) and it is
an inductive poset by exercise 7.3.3-(3), so that Zorn lemma can be applied. Let
U be a maximal element in U . For any y ∈ M , let Vy be a slice neighborhood
of y. By the exactness of the Mayer-Vietoris sequence for compact supports:
0
0
0
0
0
··· → HG,c
(U ∩Vy ) → HG,c
(U )⊕HG,c
(Vy ) → HG,c
(U ∪Vy ) → HG,c
(U ∩Vy )[1] →,
0
we easily conclude that HG,c
(U ∪Vy ) is torsion. Then U ⊇ Vy , by the maximality
of U , hence U = M .
– Torsion of HT (M ). We cannot use the same argument as in the compact
support case because a projective limit of torsion modules is no longer necessarily torsion. The finiteness assumption on the set of orbit types will now be
crucial.
Let I be the intersection of all the ideals ker(ρx : HT → HTx ) for x ∈ M . The
finiteness of the orbit type of M ensures that I 6= 0. Let (U ,⊆) be the set of
G-stable open subspaces U ⊆ M , such that I ⊆ Ann(HT (U )), partially ordered
by set inclusion. The set U is non empty as it contains every slice neighborhood
Vx (7.3.4) and it is an inductive poset by exercise 7.3.3-(4), so that Zorn lemma
can be applied. Let U be a maximal element in U . For any y ∈ M , let Vy be a
slice neighborhood of y. Thanks to the exactness of the first terms of the Mayer0
0
0
0
(U ) ⊕ HG
(Vy ) → HG
(U ∩ Vy ) →, we
Vietoris sequence: 0 → HG
(U ∪ Vy ) → HG
0
easily see that 1 ∈ HG (U ∪ Vy ) is killed by I. Then I ⊆ Ann(HG (U ∪ Vy )) by
7.3.3-(2) and U ⊇ Vy , by the maximality of U , hence U = M .
7.4. Localization Theorems
Given a T-manifold M and a nontrivial closed subgroup {1} 6= H ⊆ T, the
fixed point set M H := {x ∈ M | h · x = x ∀h ∈ H } is a submanifold whose
26 See
[H] chap IV §2, p. 54, for the general definition nottably for non abelian groups.
Alberto Arabia
73
connected components (not necessarily of equal dimensions) are stable under
the action of T, and furthermore they are orientable if M is (27 ).
Terminology. An homomorphism of HT -modules α : L → L0 will be called
an isomorphism modulo torsion if its kernel and cokernel are both torsion HT modules, i.e. if the induced homomorphism of QT -modules
α ⊗HT id : L ⊗HT QT → L0 ⊗HT QT
is an isomorphism.
7.4.1. Proposition. Let M be an oriented T-manifold of finite orbit type.
For any H nontrivial closed subgroup of T, denote by ιH : M H ,→ M the set
inclusion. The following morphisms of HT -gm (28 ) are isomorphisms modulo
torsion.
ιH ! : HT (M H )[dM H ] → HT (M )[dM ]
Gysin morphisms
ιH ∗ : HT,c (M H )[dM H ] → HT,c (M )[dM ]
Restriction morphisms
ι∗H : HT,c (M ) → HT,c (M H )
ι∗H : HT (M ) → HT (M H )
Proof. The kernel and cokernel of the restriction ι∗H : HT,c (M ) → HT,c (M H ) lay
within HT,c (U ), where U := M \M H . Now, as the isotropy groups of the points
of U are strict subgroups of T, there are no T-fixed points, i.e. U T = ∅, and
we can conclude that HT,c (U ) is an HT -torsion module by 7.3.9. In particular,
any submodule of HT,c (U ), viz. the kernel and the cokernel of ι∗H , is a torsion
HT -module. By duality the same is true for ιH ,! : HT (M H ) → HT (M ).
The other restriction ι∗H : HT (M ) → HT (M H ) is a little more tricky as its
kernel and cokernel lay within HT,U (X) which we have not yet proved is an
HT -torsion module. For that, recall that since one has short exact sequences of
local section functors over open subspaces
0 → ΓU1 ∩U1 (−) −→ ΓU1 (−) ⊕ ΓU2 (−) −→ ΓU1 ∪U2 (−) → 0
where ΓU (−) denotes the kernel of the restriction Γ (M , −) → Γ (M \U, −), one
may follow a Mayer-Vietoris procedure to approach HT,U (X) by successively
adding slice open sets Vx ⊆ U (7.3.4). In this way, to show that HT,U (M )
is a torsion module, it suffices to show that each HT,Vx (M ) is so. Now, this
HT -module occurs in the exact triangle
HT,Vx (M ) −→ H(M ) → HT (M \Vx ) →
where M \Vx is T-equivariantly homotopic to M \T·x since the slice S(x) is
a submanifold of M , therefore HT,Vx (M ) ' HT,T·x (M ) ' HT (T·x) = HTx ,
which proves that HT,Vx (M ) is a torsion HT -module.
27 We recall that the reason for this is that under the action of H , the tangent spaces T (M )
x
for x ∈ M H split as the direct sum of Tx (M H ) and a sum of H -irreducible two dimension
representations C(α) (cf. 7.2.5-(c)) which are besides canonically oriented by their character.
Thus, the orientation of Tx (M H ) determines that of Tx (M ) and vice versa.
28 As the submanifold M H need not be connected nor equidimensionnal the shift indication
in a notation as HT (M H )[dM H ] must be understood component-wise.
74
Equivariant Poincaré Duality and Gysin Morphisms
7.4.2. Theorem. Let M be a T-oriented manifold of finite orbit type such
that M T is a discrete subspace of M . Then
a) For all µ ∈ HT,c (M ) the following “localization formula” is satisfied:
Z
X
µx
.
µ=
Eu
T x,M
M
T
x∈M
b) If M is compact of positive dimension
X
1
.
0=
Eu
T x,M
T
x∈M
Proof. (a) From 7.4.1, the morphism iT,∗ : HT,c (M T ) → HT,c (M ) is an isomorphism modulo torsion, so that it suffices to prove the localization formula for
T
the equivariant
R Thom classes ΦT (x,M ) for all x ∈ M . But we have already
shown that M ΦT (x,M ) = 1 (7.2.3) and that ΦT (x,M ) x = EuT x,M by
definition. (b) Apply the localization formula to 1 ∈ HT0 (M ).
8. Miscellany
Excerpt from [B2 ] (A. Borel, IV-§3 , p. 55, 1960), where, for the first time, a
reference to what is nowadays known as the Borel construction appears.
The references ‘[2]’, ‘[5]’ and ‘[7]’ correspond to ours [B1 ], [Co] and [Gr].
9. Appendix
We explain the following fact mentioned in footnote (20 ).
Proposition. Let A := A0 ⊕ A1 ⊕ ··· be a graded ring. Denote by S the
multiplicative system generated by the nonzero graded elements of A. The ring
L := S −1 A is a graded A-module such that, for any A-graded module N , the
tensor product L⊗A N is flat and injective in the category of graded A-modules.
Proof. • L ⊗ N is flat. For any graded ideal I of A, one has the long exact
sequence:
0 → TorA
1 (L,A/I) −→ L ⊗ I −→ L −→ L ⊗ (A/I) → 0
(∗)
where A/I is a torsion graded A-module. The annihilators of the elements of
A/I are graded ideals, generated, as such, by invertible elements of L. Therefore
TorA
1 (L,A/I) = 0,
∀i ∈ N,
75
Alberto Arabia
and we have from (∗) the equality L ⊗ I = L from which, we deduce
L⊗I ⊗N =L⊗N
for any A-graded module N . The ideal criterion of flatness applies, and the
A-graded module L ⊗ N is flat.
• L ⊗ N is injective. Let α : M1 ⊆ M2 be a graded inclusion of graded
A-modules. We must show that any morphism λ : M1 → L ⊗ N of graded
A-modules can be extended to M2 .
M1
λ
{
L⊗N
α
⊆
/ M2
λ0
In the contrary, Zorn’s lemma will led us to assume that M2 % M1 and
that λ may not be further extended. In particular, A · m ∩ M1 6= 0 for any
homogeneous m ∈ M2 \ M1 , hence the quotient M2 /M1 is a torsion module.
One then has
L ⊗ M1 = L ⊗ M2 ,
and a contradiction arises as a consequence of the diagram
/ HomgrA (M1 ,L ⊗ N )
HomgrA (M2 ,L ⊗ N )
∼
=
HomgrL (L ⊗ M2 ,L ⊗ N )
∼
=
(=)
/ HomgrL (L ⊗ M1 ,L ⊗ N )
where the horizontal arrows are induced by the inclusion M1 ⊆ M2 and the
vertical arrows are the well-known canonical natural isomorphisms.
76
Equivariant Poincaré Duality and Gysin Morphisms
References
[AFP] C. Allday, M. Franz, V. Puppe. Equivariant cohomology, syzygies and orbit
structure. Trans. Amer. Math. Soc. 366, no. 12, 6567–6589 (2014).
[A] M. André. Cohomologie des algèbres différentielles où opère une algèbre de Lie.
Ph.D. Thesis 1962, advisor Claude C. Chevalley. Tôhoku Math. J. (2) 14 (1962),
no. 3, 263–311. http://projecteuclid.org/download/pdf 1/euclid.tmj/1178244116.
[AB] M.F. Atiyah, R. Bott. The moment map and equivariant cohomology. Topology
23, no. 1, 1–28 (1984).
[AS] M.F. Atiyah, G.B. Segal. The index of elliptic operators II. Ann. of Math. (2)
87, 531–545 (1968).
[BGV] N. Berline, E. Getzler, M. Vergne. Heat kernels and Dirac operators. Corrected
reprint of the 1992 original. Grundlehren Text Editions. Springer-Verlag, (2004).
[BV1 ] N. Berline, M. Vergne. Classes caractéristiques équivariantes. Formule de localisation en cohomologie équivariante. C. R. Acad. Sci. Paris Sér. I Math. 295,
no. 9, 539–541 (1982). http://gallica.bnf.fr/ark:/12148/bpt6k62356694/f77.
[BV2 ] N. Berline, M. Vergne. Zéros d’un champ de vecteurs et classes caractéristiques
équivariantes. Duke Math. J. 50, no. 2, 539–549, (1983).
[BV3 ] N. Berline, M. Vergne. Fourier transforms of orbits of the coadjoint representation. Representation theory of reductive groups (Park City, Utah, 1982), 53–67,
Progr. Math., 40, Birkhäuser, (1983).
[B1 ] A. Borel. Nouvelle démonstration d’un théorème de P. A. Smith. Comment.
Math. Helv. 29, 27–39 (1955).
[B2 ] A. Borel. “Seminar on transformation groups”. With contributions by G. Bredon,
E. E. Floyd, D. Montgomery, R. Palais. Annals of Mathematics Studies, No. 46.
Princeton University Press, Princeton, N.J. (1960).
[B3 ] A. Borel. Sur La Cohomologie des Espaces Fibres Principaux et des Espaces
Homogenes de Groupes de Lie Compacts Annals of Mathematics, Second Series,
Vol. 57, No. 1, pp. 115-207 (1953).
[Bo] R. Bott. Vector fields and characteristic numbers. Michigan Math. J. 14 231–244
(1967).
[BT] R. Bott, L.W. Tu. “Differential forms in algebraic topology”. Graduate Texts in
Mathematics, 82. Springer-Verlag, New York-Berlin (1982).
[Bk] N. Bourbaki. “Algebra I. Chapters 1–3”. Elements of Mathematics (Berlin).
Springer-Verlag, Berlin, 1998.
[Br] J.-L. Brylinski. Equivariant intersection cohomology. Kazhdan-Lusztig theory
and related topics (Chicago, IL, 1989), 5–32, Contemp. Math., 139, Amer. Math.
Soc., Providence, RI, 1992.
[Ca1 ] H. Cartan. Notions d’algèbre différentielle; application aux groupes de Lie et
aux variétés où opère un groupe de Lie. Colloque de topologie (espaces fibrés),
Bruxelles 1950, pp. 15–27 (1951)
[Ca2 ] H. Cartan. La transgression dans un groupe de Lie et dans un espace fibré
principal. Colloque de topologie, Bruxelles 1950, pp. 15–27 (1951)
[Co] P.E. Conner. On the action of the circle group. Michigan Math. J. 4 1957 241–247
[Dx] J. Dixmier. “Enveloping algebras”. Revised reprint of the 1977 translation.
Graduate Studies in Mathematics, 11. AMS, 1996.
[DH] J.J. Duistermaat, G.J. Heckman. On the variation in the cohomology of the
symplectic form of the reduced phase space. Invent. Math. 69, no. 2, 259–268,
(1982). Addendum to: “On the variation in the cohomology of the symplectic
form of the reduced phase space”. Invent. Math. 72, no. 1, 153–158, (1983).
[FP] M. Franz, V. Puppe. Freeness of equivariant cohomology and mutants of compactified representations. pp. 87-98 in: Megumi Harada et al. (eds.), Toric Topology (Osaka, 2006), Contemp. Math. 460, AMS, Providence (2008)
Alberto Arabia
77
[F1 ] M. Franz. Big polygon spaces. Int. Math. Res. Not. 2015, 13379-13405 (2015).
[F2 ] M. Franz. Syzygies in equivariant cohomology for non-abelian Lie groups. pp. 325360 in: F. Callegaro et al. (eds.), Configuration spaces (Cortona, 2014), Springer
INdAM Ser. 14, Springer, Cham (2016)
[FF] R. Fossum, H.-B. Foxby. The category of graded modules. Math. Scand. 35,
288–300 (1974).
[Go] R. Godement. “Topologie algébrique et théorie des faisceaux”. Troisième édition
revue et corrigée. Actualités Scientifiques et Industrielles, No. 1252. Hermann,
Paris, (1973).
[GKM] M. Goresky, R. Kottwitz, R. MacPherson. Equivariant cohomology, Koszul
duality, and the localization theorem. Invent. Math. 131 (1998), no. 1, 25–83.
[Gt] D.H. Gottlieb. Fibre bundles and the Euler characteristic. J. Differential Geometry 10, 39–48, (1975).
[Gr] A. Grothendieck. Sur quelques points d’algèbre homologique. Tôhoku Math. J.
(2) 9 1957 119–221.
[H] W. Hsiang. “Cohomology theory of topological transformation groups”. Ergebnisse
der Mathematik und ihrer Grenzgebiete, Band 85 (1975)
[J] N. Jacobson. “Basic algebra. II”. Second edition. W. H. Freeman and Company,
New York, 1989
[Jo] R. Joshua. Vanishing of odd-dimensional intersection cohomology. Math. Z. 195
(1987), no. 2, 239–253 (1987).
[L] J. Leray. Sur l’homologie des groupes de Lie, des espaces homogènes et des espaces
fibrés principaux. (French) Colloque de topologie (espace fibrés), Bruxelles, 101–
115 (1950).
[Q] D. Quillen. The spectrum of an equivariant cohomology ring. I, II. Ann. of Math.
(2) 94 (1971), 549–572; ibid. (2) 94 (1971), 573–602.
[S] G. Segal. Equivariant K-theory. Inst. Hautes Études Sci. Publ. Math. No. 34,
129–151 (1968).
[W] E. Witten Supersymmetry and Morse theory. J. Differential Geom. 17 (1982),
no. 4, 661–692 (1983).
Index
adjoint
left equivariant map, 52
maps, 4
pair, 8, 18
right equivariant map, 52
annihilator
of an element, 70
of module, 70
antiderivation, 35
ascending chain property, 9
augmentation, 45
de Rham
cohomology, 6
with compact support, 6
complex, 6
with compact supports, 6
derivation, 30, 35
derived
category, 21
duality fonctor, 48
functor (left and right), 45
diagonal action, 39
diagonal embedding, 22
differential, 4, 43
graded g-complex, 30
graded module, 43
graded vector space, 4
double (cochain) complex, 45
of the first quadrant, 46
of the third quadrant, 47
dual
of a complex, 5
vector space, 4
duality functor, 5, 48
bidual embedding, 6, 7, 12
bilinear map, 4
Borel construction, 39, 68, 74
bounded below (above), 43
Cartan
complex, 33
differential, 32
cartesian diagram, 19, 20, 41
category
of (oriented) manifolds (with proper
maps), 6
of differential graded g-complexes,
30
of differential graded modules over
HG , 43
of differential graded vector spaces,
4
of G-manifolds, 6
of g-modules, 29
of graded vector spaces, 4
of H -graded modules, 42
Chern-Weil homomorphism, 68
Chevalley
restriction theorem, 68
Theorem, 43
classifying space, 39
closed embedding, 19, 21, 41, 64, 73
coboundary, 4
column ?-filtration, 46
completely reducible module, 29
complex
double, 45
simple, 46
cone of a morphism of dgm’s, 50
constant map, 18
contraction, 30, 35
embedding
bidual, 6, 7, 12
closed, 19, 21, 41, 64, 73
diagonal, 22
open, 18, 64
equivariant
Cartan complex, 33
cogolomogy, 35
cohomology of a g-complex, 33
cohomology with support, 41
differential form, 35
Gysin functor, 57
Gysin functor for proper maps, 59
Gysin morphism, 57
Gysin morphism of a proper map,
58
integration, 51
left adjoint map, 52
map, 6
projection formula
for proper maps, 59
right adjoint map, 52
Euler
class of the pair (N ,M ), 67
extension by zero, 18
Exti,•
HG ( , ), 47
78
Alberto Arabia
fibration, 18
filtrant covering, 15
filtration
column, 46
line, 46
regular, 34, 46, 53
finite
(de Rham) type, 9
de Rham type, 53, 56, 66
length resolution, 48
orbit type, 72
type map, 58
first quadrant double complex, 46
fixed point
theorem, 71
forgetful functor, 44
free HG -graded module, 42, 48
free HG -graded module, 50
functor
duality-, 48
forgetfull, 44
i’th extension, 48
g-complex, 30
morphism, 30
G-contraction, 35
G-derivation, 35
g-invariant, 29
G-manifold, G-map, 6
g-module, 29
g-module morphism, 29
G-split complex, 34
g-split complex, 30
g-tivial representation, 29
good cover, 9, 38, 68
graded
homomorphism, 41
algebra, 41
homomorphism, 4
space, homomorphism, 4
graph map, 22
Gysin
exact long sequence, 22
functor, 13, 13
for proper maps, 17
morphism, 8, 13, 21
for proper maps, 9
h, cohomology as graded space, 6
HG -dgm, 43
HG -duality functor, 44
HG -graded module, HG -gm, 41
Hilbert’s Syzygy Theorem, 43
79
index of an equivariant form, 53
injective
HG -graded module, 42
resolution of finite length, 48
integration, 18
integration along fibers, 18
isomorphism modulo torsion, 73
isotropie group, 71
i’th extension functor, 48
Lefschetz
equivariant class, number of an
equivariant map, 64
class, number of a map, 22
fixed point formula, 23
left adjoint, 8
Leray cover, 9
Lie
algebra, 29
group, 9
line \-filtration, 46
localization
functor, 65
localized
Cartan complex, 65
equivariant cohomology, 65
locally trivial fibration, 18
manifold, 6
map
equivariant, 6
of manifolds, 6
morphism
augmentation, 45
Gysin, 13, 21
of complexes, 4
of differential graded modules, 43
of g-complexes, 30
of graded vector spaces, 4
of H -graded modules, 42
nondegenerate pairing, 4, 7, 52, 55
nontorsion module, 70
open embedding, 18, 64
orbit type, 72
pairing, 4
perfect pairing, 4, 7, 55, 65
perfect pariring, 9
Poincaré
morphism, 7
in G-equivariant cohomology, 52
in T-equivariant cohomology, 56
80
Equivariant Poincaré Duality and Gysin Morphisms
Poincaré pairing
in cohomology, 7, 9
in equivariant cohomology, 52
projection formula, 13, 57
for proper maps, 17
projective
HG -graded module, 42, 48, 49
HG -graded module, 50
proper map, 9
pullback, 12, 14–16
pushforward, 13, 16, 18, 21, 36
quadrant (first,third), 46, 47
quasi
-injection, 4, 12
-isomorphism, 4
-surjection, 4
reducible module, 29
reflexive module, 54
regular filtration, 34, 46, 53
right adjoint, 8
map, 12
semisimple module, 29
Serre spectral sequence, 40
shift functor, 5
simple
complex, 46
module, 29
slice theorem, 71
spectral sequence, 40
Syzygy, 43
Theorem
Chevalley, 43
Hilbert, 43
Thom
class of a pair (N ,M ), 66
class of a vector bundle, 19
isomorphism, 19
Tor•HG ( , ), 47
torsion element, module, 70
torsion-free module, 54, 70
trivial representation, 29
tubular neighborhood, 65, 66
universal fiber bundle, 39, 67
weight, 69
Weyl group, 68
zero section, 19, 63
| 4math.GR
|
Blocks with defect group Z2n × Z2n × Z2m
Chao Wu, Kun Zhang and Yuanyang Zhou
arXiv:1803.07262v1 [math.GR] 20 Mar 2018
School of Mathematics and Statistics, Central China Normal University, Wuhan, 430079, P.R. China
Abstract In this paper, we prove that a block with defect group Z2n × Z2n × Z2m , where n ≥ 2
and m is arbitrary, is Morita equivalent to its Brauer correspondent.
Keywords: Finite groups; blocks; Morita equivalences
1. Introduction
Let p be a prime number and O a complete discrete valuation ring with an algebraically closed
field k of characteristic p. Denote by K a fraction field of O. Assume that O has characteristic 0
and that K is big enough for groups considered below. Let G be a finite group, and b a block of G
over O — a central primitive idempotent in the group algebra OG. The product OGb of OG and
b is a block algebra. Let P be a defect group of the block b, and c the Brauer correspondent of b
in the normalizer NG (P ). Assume that P is abelian. Broué’s abelian defect group conjecture says
that the block algebras OGb and ONG (P )c are derived equivalent (see [2]).
Denote by Zn the residue class group modulo n. Recently, some authors proved the conjecture
for some 2-blocks with abelian defect groups, such as Z2 × Z2 (see [7] and [29]), Z2 × Z2 × Z2
(see [9]), and Z2n × Z2n (see [10] for n ≥ 2). The conjecture also has been verified in many other
cases. We are especially interested in the case Z2n × Z2n . In [10], the authors classified 2-blocks
of quasisimple groups with abelian defect groups, and then proved that the block algebras OGb
and ONG (P )c are Morita equivalent when P is Z2n × Z2n for some n ≥ 2. By inspecting the
classification, we observe that the block algebras OGb and ONG (P )c are Morita equivalent when
G is quasisimple and P is Z2n × · · · × Z2n for some n ≥ 2. So it may be interesting to ask whether
such an observation holds for a general finite group G. In the paper, we investigate blocks with
defect group Z2n × Z2n × Z2n , where n ≥ 2. More generally, we prove Broué’s abelian defect group
conjecture for blocks with defect groups Z2n × Z2n × Z2m , where n ≥ 2 and m is arbitrary.
Theorem. Assume that P is Z2n × Z2n × Z2m , where n ≥ 2 and m is arbitrary. Then the block
algebras OGb and ONG (P )c are Morita equivalent.
Given a finite p-group P , Donovan’s conjecture says that there are finitely many Morita equivalence classes of block algebras over k with defect group P . It is expected that Donovan’s conjecture
should hold over O too. It is easy to see that the quotient group NG (P )/CG (P ) is trivial, or Z3 ,
or Z7 , or the Frobenius group Z7 ⋊ Z3 . By the structure theorem of block algebras with normal
defect group, we conclude the following.
Corollary. Assume that P is Z2n × Z2n × Z2m , where n ≥ 2 and m is arbitrary. Then the block
algebra OGb is Morita equivalent to OP , or O(P ⋊ Z3 ), or O(P ⋊ Z7 ), or O(P ⋊ (Z7 ⋊ Z3 )). In
particular, Donovan’s conjecture holds.
We close the section with two remarks. When m 6= n, the block algebra OGb is Morita equivalent
to OP or O(P ⋊ Z3 ). When m = 0, by Proposition 6.7 below, the Morita equivalence between OGb
and ONG (P )c in Theorem is basic in the sense of [23].
1
2. Preliminaries
In this section, we collect some elementary lemmas. Let G and G′ be finite groups with common
center Z. The central product G ∗ G′ over Z is isomorphic to the subgroup G ⊗ G′ in the tensor
product OG ⊗OZ OG′ and the group algebra O(G ∗ G′ ) is isomorphic to OG ⊗OZ OG′ .
Lemma 2.1. Keep the notation as above. Then any block of G ⊗ G′ is of the form b ⊗ b′ , where
b and b′ are blocks of G and G′ respectively. If P and P ′ are defect groups of b and b′ respectively,
then P ⊗ P ′ is a defect group of b ⊗ b′ . Moreover, the order of the inertial quotient of the block
b ⊗ b′ is the product of the order of the inertial quotient of the block b and that of b′ .
Proof. The map G × G′ → G ⊗ G′ , (g, g ′ ) 7→ g ⊗ g′ is a surjective group homomorphism with
kernel {(z, z −1 )|z ∈ Z} isomorphic to Z. Denote by Z ′ the maximal p′ -subgroup of Z. Note that
the quotient group Z/Z ′ is a p-group. The homomorphism G × G′ → G ⊗ G′ factors through the
canonical homomorphisms
G × G′ → (G × G′ )/Z ′ and (G × G′ )/Z ′ → G ⊗ G′ .
˜ ′ the element of the tensor product OG ⊗O OG′
Let a ∈ OG and a′ ∈ OG′ . We denote by a⊗a
′
˜ ′ and a⊗a′ . There is an O-algebra isomorphism
determined by a and a , in order to differentiate a⊗a
˜ 2 for any x ∈ G and any x′ ∈ G′ . We identify
O(G × G′ ) ∼
= OG ⊗O OG′ sending (x1 , x2 ) onto x1 ⊗x
′
′
˜ ′ , where b and
the algebras O(G × G ) and OG ⊗O OG . Then any block of G × G′ is of form b⊗b
b′ are blocks of G and G′ . If P and P ′ are defect groups of b and b′ , then P × P ′ is a defect group
˜ ′ . The inertial quotient of b⊗b
˜ ′ is the direct product of the inertial quotients of b and b′ .
of b⊗b
1 P
Set e = |Z ′ | z∈Z ′ z. The homomorphism G×G′ → (G×G′ )/Z ′ induces an algebra isomorphism
O(G × G′ )e ∼
= O((G × G′ )/Z ′ ). We identify the two algebras O(G × G′ )e and O((G × G′ )/Z ′ ). By
˜ ′ , where b and b′ are blocks of G and
the last paragraph, any block of (G × G′ )/Z ′ is of form b⊗b
′
′
′
′
˜ ′
G ; if P and P are defect groups of b and b , then P × P is isomorphic to a defect group of b⊗b
′
˜ is isomorphic to the direct product of the
(see [16, Theorem 5.7.4]); the inertial quotient of b⊗b
′
inertial quotients of b and b (see [20, Theorem 3.6]).
The homomorphism (G × G′ )/Z ′ → G ⊗ G′ induces a surjective homomorphism O((G ×
G′ )/Z ′ ) → O(G ⊗ G′ ) with kernel the central p-subgroup Z/Z ′ , which induces a one to one correspondence between blocks of (G × G′ )/Z ′ and G ⊗ G′ (see [16, 5.8.11]). If D/Z ′ is a defect group
of a block of (G × G′ )/Z ′ , then D/Z is a defect group of the corresponding block of G ⊗ G′ (see
[16, Theorem 5.8.10]). By [21, Theorem 2.9], a block of (G × G′ )/Z ′ and the corresponding block
of G ⊗ G′ have isomorphic inertial quotients. The proof is done.
2.2. Let CF(G, K) be the K-vector space of all K-valued class functions on G, and CFp′ (G, K) the
subspace of all functions in CF(G, K) vanishing on p-singular elements of G. For any χ ∈ CF(G, K)
and any central idempotent e in OG, extending χ to a function over KG by the K-linearity, we
define a new class function e · χ on G by the equality (e · χ)(g) = χ(ge) for any g ∈ G. We set
CF(G, e, K) = {e · χ|χ ∈ CF(G, K)} and CFp′ (G, e, K) = CF(G, e, K) ∩ CFp′ (G, K). It is easy
to see that CF(G, e, K) and CFp′ (G, e, K) are subspaces of CF(G, K) and that we have direct sum
decompositions
M
M
CF(G, K) =
CF(G, b, K) and CFp′ (G, K) =
CFp′ (G, b, K),
b
b
where b runs over the set of blocks of G. Denote by Irr(G) and Irr(G, b) the set of irreducible
ordinary characters of the group G and the set of irreducible ordinary characters in a block b of G. It
2
is known that Irr(G) and Irr(G, b) are bases of CF(G, K) and CF(G, b, K) and that the orthogonality
relationship of characters determines the inner products on CF(G, K) and CF(G, b, K). We denote
by CFp′ (G, O) the O-submodule of all O-valued class functions on G vanishing on all p-singular
elements of G, and by CFp′ (G, e, O) the intersection of CFp′ (G, O) and CF(G, e, K).
2.3. Let b and b′ be blocks of G and G′ . Denote by b′◦ the inverse image of b′ through the
opposite isomorphism OG′ ∼
= OG′ . There is an O-algebra isomorphism O(G × G′ ) ∼
= OG ⊗O OG′ .
′
′
′◦
Identifying the two algebras O(G × G ) and OG ⊗O OG , b ⊗ b is a block of G × G′ . Given a
generalized character ν of G × G′ in the block b ⊗ b′◦ , we define a map Iν : ZIrr(G′ , b′ ) → ZIrr(G, b)
by the equality
1 X
ν(g, h)α(h)
Iν (α)(g) = ′
|G |
′
h∈G
ZIrr(G′ , b′ )
for any α ∈
and any g ∈ G, where ZIrr(G, b) and ZIrr(G′ , b′ ) are the free abelian
groups generated by Irr(G, b) and by Irr(G′ , b′ ). The map Iν can be extended to a K-linear map
CF(G′ , b′ , K) → CF(G, b, K), denoted by IνK . We remind that the definition of Iν is slightly different
from the definition of Iµ in [3, (F)]. We modify the definition of Iµ in [3, (F)], mainly to fill gaps,
which are produced when we use to µ the induction and extension of generalized characters. Such
a modification does not influence the use of all known theorems on perfect isometries (see [3]). If
Iν is a perfect isometry and we denote by χ∗ the dual of an ordinary character χ, then
X
ν=
Iν (χ) × χ∗ ,
χ∈Irr(G′ , b′ )
where Iν (χ) × χ∗ denotes a generalized ordinary character of G × G′ determined by χ∗ and Iν (χ).
2.4. Let u be a p-element of G. We denote by duG the surjective map CF(G, K) → CFp′ (CG (u), K)
defined by duG (χ)(s) = χ(us) for any χ ∈ CF(G, K) and any p-regular element s of CG (u), and by euG
the map CFp′ (CG (u), K) → CF(G, K) sending any ϕ onto the function euG (ϕ) which takes ϕ(s) at
us for any p-regular element s of CG (u) and 0 at g ∈ G if the p-part of g is not conjugate to u. The
composition duG ◦ euG is the identity map on CFp′ (CG (u), K). Let (u, e) be a b-Brauer element. We
(u, e)
the composition of duG and the projection CFp′ (CG (u), K) → CFp′ (CG (u), e, K).
denote by dG
(u, e)
the
By [4, Theorem A2.1] euG maps CFp′ (CG (u), e, K) into CF(G, b, K) and we denote by eG
(u, e)
(u, e)
u
is the identity map on
◦ eG
restriction of eG on CFp′ (CG (u), e, K). The composition dG
CFp′ (CG (u), e, K).
2.5. Let (P, f ) be a maximal b-Brauer pair. We denote by Br(P, f ) the category, whose objects
are all b-Brauer pairs (Q, fQ ) contained in (P, f ) and whose morphisms from (Q, fQ ) to (R, fR ) are
group homomorphisms Q → R induced by elements g ∈ G such that (Q, fQ )g ≤ (R, fR ). We assume
that the blocks b and b′ have a commom defect group P and that the inclusions P ⊂ G and P ⊂ G′
induce an isomorphism between the categories Br(P, f ) and Br(P, f ′ ), where (P, f ′ ) is a suitable
maximal b′ -Brauer pair. The map IνK : CF(G′ , b′ , K) → CF(G, b, K) is said to be compatible with
a local system {IT : CFp′ (CG′ (T ), fT′ , K) → CFp′ (CG (T ), fT , K)}{T (cyclic)⊂P } (see [3, Definition
(x, f )
(x, f ′ )
4.3]), if for any cyclic subgroup T of P and any generator x of T , we have dG T ◦ IνK = IT ◦ dG′ T .
By the isomorphism theorem of groups, the local system is uniquely determined by IνK if it exists.
Lemma 2.6. Keep the notation and the assumption in 2.5 and assume that P is abelian and
that the map IνK : CF(G′ , b′ , K) → CF(G, b, K) is compatible with a local system {IT }{T (cyclic)⊂P } .
3
(x, f ′ )
P
(x, f )
Then we have IνK = x∈I eG hxi ◦ Ihxi ◦ dG′ hxi , where hxi is the cyclic subgroup generated by x
and I is a set of representatives of orbits of the conjugation action of NG (P, f ) on P .
Proof. Obviously {(x, fhxi )|x ∈ I} is a set of representatives of orbits of the action of NG (P, f ) on
the set of b-Brauer elements contained in (P, f ), and since P is abelian, {(x, fhxi )|x ∈ I} is a set of
(x, f
)
representatives of G-conjugacy classesLof b-Brauer elements. The family {dG hxi |x ∈ I} induces a
K-linear isomorphism CF(G, b, K) ∼
= x∈I CFp′ (CG (x), fhxi , K) (see the last paragraph of [3, 4A]).
(x′ , fhx′ i )
Since dG
(x, fhxi )
◦ eG
= 0 for different x, x′ ∈ I, we have
(x, fhxi )
dG
(x, fhxi )
◦ IνK = dG
◦
X
(y, fhyi )
eG
′ )
(y, fhyi
◦ Ihyi ◦ dG′
.
y∈I
So IνK =
P
(y, fhyi )
y∈I eG
′ )
(y, fhyi
◦ Ihyi ◦ dG′
.
The main idea of the following proposition is already included in the first sentence of [28, 1.6].
Proposition 2.7. Keep the notation and the assumption in 2.5 and assume that P is abelian.
Then the map IνK : CF(G′ , b′ , K) → CF(G, b, K) is compatible with a local system {IT }{T (cyclic)⊂P }
if and only if IνK (λ ∗ χ) = λ ∗ IνK (χ) for any χ ∈ CF(G′ , b′ , K) and any NG (P, f )-stable character
λ of P , where ∗ denotes the ∗-construction of characters due to Broué and Puig.
Before the proof, we remark that any ordinary character λ of P is NG (P, f )-stable if and only
if it is NG′ (P, f ′ )-stable, since the Brauer categories Br(P, f ) and Br(P, f ′ ) are isomorphic.
Proof. We assume that the map Iν is compatible with a local system {IT }{T (cyclic)⊂P } . For any
b-Brauer element (u, fhui ) and any χ ∈ CF(G, b, K), set χ(u, fhui ) = (e(u, fhui ) ◦ d(u, fhui ) )(χ). Let I
be a set of representatives of orbits of the conjugation action of NG (P, f ) on P . Since P is abelian,
{(x, fhxi )|x ∈ I} is a set of representatives of G-conjugacy classes of b-Brauer elements. Since
′ )|x ∈ I} is a set of representatives
the categories Br(P, f ) and Br(P, f ′ ) are isomorphic, {(x, fhxi
′
′
of orbits of the action of NG′ (P, f ) on the set of b -Brauer elements contained in (P, f ′ ) and
′ )|x ∈ I} is also a set of representatives of the G′ -conjugacy classes of b′ -Brauer elements.
{(x, fhxi
P
(x, f )
(x, f ′ )
For any χ ∈ CF(G′ , b′ , K), we have χ = x∈I χ hxi . Since IνK (λ ∗ χ)(x, fhxi ) = λ(x)eG hxi ◦ Ihxi ◦
′ )
(x, fhxi
dG′
(χ
′ )
(x, fhxi
) = (λ ∗ IνK (χ))(x, fhxi ) (see Lemma 2.6), we have
X
X
IνK (λ ∗ χ) =
IνK (λ ∗ χ)(x, fhxi ) =
(λ ∗ IνK (χ))(x, fhxi ) = λ ∗ IνK (χ).
x∈I
x∈I
Conversely, we assume that IνK (λ ∗ χ) = λ ∗ IνK (χ) for any χ ∈ CF(G′ , b′ , K) and any NG (P, f )stable character λ of P . Obviously NG (P, f ) acts on Irr(P ) and the number of orbits of the action
of NG (P, f ) on Irr(P ) is equal to the cardinality of I. We list the orbit sums of the action of
NG (P, f ) on Irr(P ) as λj , where 1 ≤ j ≤ n and n = |I|. For any λi and any χ ∈ Irr(G′ , b′ ), we have
P
(u, f ′ )
λi ∗ χ = u∈I λi (u)χ hui . The matrix (λi (u)), whose entries are coefficients in the equality, is
an invertible matrix, since Irr(P ) is a basis of CF(P, K). Denoting by (aiu ) the invertible matrix
of (λi (u)), we have
X
(u, f ′ )
2.7.1
χ hui =
aiu λi ∗ χ.
i
4
λi (u)(IνK (χ))(u, fhui ) , similarly we have
Since IνK (λi ∗ χ) = λi ∗ IνK (χ) =
P
2.7.2
(IνK (χ))(u, fhui ) =
u∈I
X
aiu IνK (λi ∗ χ).
i
Combining the equalities 2.7.1 and 2.7.2, we have
(IνK (χ))(u, fhui ) = IνK (χ
2.7.3
′ )
(u, fhui
).
For any u ∈ I, we denote by V (u, fhui ) the image of the homomorphism e(u, fhui ) . Obviously the set
{χ(u, fhui ) |χ ∈ Irr(G, b)} is a generator set of the space V (u, fhui ) , the space CF(G, b, K) is equal to the
L
(u, f )
sum u∈I V (u, fhui ) since the family {dG hui |u ∈ I} induces a K-linear isomorphism CF(G, b, K) ∼
=
L
L
(u, fhui )
(v,
f
)
hvi . By
is equal to the sum v∈I−{u} V
u∈I CFp′ (CG (u), fhui , K), and the kernel of dG
the equality 2.7.3, IνK maps V
kernel of
(u, f )
dG hui .
Therefore
′ )
(u, fhui
IνK
′ )
(u, fhui
onto V (u, fhui ) and thus it maps the kernel of dG′
onto the
induces a K-linear homomorphism
′
, K) → CFp′ (CG (u), fhui , K)
Ihui : CFp′ (CG′ (u), fhui
(u, fhui )
such that dG
′ )
(u, fhui
◦ IνK = Ihui ◦ dG′
.
Lemma 2.8. Keep the notation and the assumption in 2.5 and assume that P is abelian. Assume
that the map IνK : CF(G′ , b′ , K) → CF(G, b, K) is compatible with a local system {IT }{T (cyclic)⊂P }
and that IνK maps CF(G′ , b′ , O) into CF(G, b, O). Then for any cyclic subgroup T of P , IT maps
CFp′ (CG′ (T ), fT′ , O) into CFp′ (CG (T ), fT , O).
Proof.
(x, fhxi )
Take x ∈ P . Since eG
′ )
(x, fhxi
◦ Ihxi ◦ dG′
(x, fhxi )
= eG
(x, fhxi )
◦ dG
(x, fhxi )
(the second equality is obtained by the equality 2.7.3), we have eG
Since the homomorphisms
(x, f )
eG hxi
CF(G′ , b′ , O) onto CF(G, b, O),
contained in CF(G, b, O). Since
(x, fhxi )
dG
′ )
(x, fhxi
◦ IνK = IνK ◦ eG′
′ )
(x, fhxi
◦ dG′
′ )
(x, fhxi
◦ Ihxi = IνK ◦ eG′
.
maps CFp′ (CG (hxi), fhxi , O) into CF(G, b, O) and IνK maps
′ , O) through e(x, fhxi ) ◦ I
the image of CFp′ (CG′ (hxi), fhxi
hxi is
G
(CF(G, b, O)) ⊂ CFp′ (CG (hxi), fhxi , O),
′ , O) into CF ′ (C (hxi), b , O).
Ihxi maps CFp′ (CG′ (hxi), fhxi
G
p
hxi
2.9. Now we take an (OG, OG′ )-bimodule M , which induces a Morita equivalence between OGb
and OG′ b′ . By [1, Theorem 2.1 and Exercise 4.5], the Morita equivalence induced by M induces
an algebra isomorphism
ρ : Z(OG′ b′ ) ∼
= Z(OGb)
such that for any a ∈ Z(OGb) and any a′ ∈ Z(OG′ b′ ), a and a′ correspond to each other if
and only if am = ma′ for any m ∈ M , where Z(OGb) and Z(OG′ b′ ) are the centers of the
algebras OGb and O′ G′ b′ . In order to consider the character of the bimodule M and basic Morita
equivalences in Section 3, we also regard the module M as an O(G × G′ )-bimodule by the equality
(g, g ′−1 )m = gmg ′ for any g ∈ G, any g′ ∈ G′ and any m ∈ M . Let µ be the character of M . The
5
map Iµ : ZIrr(G′ , b′ ) → ZIrr(G, b) is a perfect isometry. By [3, Theorem 1.5], the perfect isometry
Iµ induces an algebra isomorphism
ρ′ : Z(OG′ b′ ) ∼
= Z(OGb).
Lemma 2.10.
Proof.
Keep the notation as the above. The isomorphisms ρ and ρ′ coincide.
Consider the extensions of the isomorphisms ρ and ρ′ by the K-linearity
ρK : Z(KG′ b′ ) ∼
= Z(KGb) and ρ′K : Z(KG′ b′ ) ∼
= Z(KGb).
P
−1 in KG.
For any χ ∈ Irr(G), we denote by eχ the central primitive idempotent χ(1)
g∈G χ(g)g
|G|
′ , b′ ). We have e m = me ′ for any m ∈ M , and thus
Set eχ = ρK (eχ′ ) for any χ′ ∈ Irr(GP
χ
χ
µ(eχ , 1) = µ(1, eχ′∗ ) 6= 0. Since µ = χ′ ∈Irr(G′ , b′ ) (Iµ (χ′ ) × χ′∗ ), Iµ (χ′ )(eχ ) is nonzero and Iµ (χ′ )
is equal to χ. On the other hand, by the proof of [3, Theorem 1.5] the isomorphism ρ′K also sends
eχ′ onto eχ . So ρ′K = ρK .
3. Morita equivalences and the ∗-structure
In the section, we will prove that if two blocks with abelian defect groups are basically Morita
equivalent, then there is a Morita equivalence between the two corresponding block algebras, which
induces a perfect isometry compatible with a local system (see Proposition 3.6). In order to do
that, we firstly recall the computation of generalized decomposition numbers in [27].
3.1. Throughout the section, all O-algebras and O-modules are O-free of finite O-rank. For any
ring R, denote by R∗ the multiplicative group of R. Let G be a finite group. An O-algebra A
is an interior G-algebra if there is a group homomorphism G → A∗ . The conjugation of G in
A induces a group homomorphism G → Aut(A) so that A becomes a G-algebra, where Aut(A)
denotes the automorphism group of the O-algebra A. For any subgroup H of G, we denote by AH
the subalgebra of all H-fixed elements in the G-algebra A. A pointed group on A is a pair (H, α),
where H is a subgroup of G and α is an (AH )∗ -conjugacy class of primitive idempotents in AH . We
often write (H, α) as Hα , and say that α is a point of H on A. The pointed group Hα is contained
in another pointed group Kβ , denoted by Hα ≤ Kβ , if H is a subgroup of K and there are i ∈ α
and j ∈ β such that ji = ij = i.
K
H
3.2. Let H and K be two subgroups of G such that K ≤ H. We denote by TrH
K : A → A the
H
relative trace map and by AH
K the image of TrK . We consider the quotient
.
X
J(O)AH +
AH
A(H) = AH
K
K
where K runs over all proper subgroups of H, and denote by BrA
H the canonical surjective algebra
homomorphism from AH to A(H). If A(H) is not zero, then H has to be a p-subgroup. Clearly
the inclusion G ⊂ OG endows OG an interior G-algebra structure. Let Q be a p-subgroup of G.
Then the inclusion OCG (Q) ⊂ (OG)Q induces an algebra isomorphism (OG)(Q) ∼
= kCG (Q). We
identify (OG)(Q) and kCG (Q) in the sequel.
3.3. A pointed group Hα on A is local if H is a p-subgroup of G and the image BrA
H (α) 6= {0}. In
is
a
surjective
algebra
homomorphism,
this case, α is called a local point of H on A, and since BrA
H
6
BrA
H (α) is a set of primitive idempotents in A(H) and by the lifting theorem of idempotents, the
correspondence α 7→ BrA
H (α) gives a bijection between the local points of H on A and the points of
H on A(H). Let x be an element of G. We call xα a local pointed element if hxiα is a local pointed
group on A. The local pointed element xα is contained in a pointed group Kβ if hxiα is contained
in Kβ .
3.4. Now we consider the group algebra OG. Let b a block of G. Then {b} is a point of G on OG.
Let xα be a local pointed element contained in the pointed group G{b} on OG, and take j ∈ α.
Let χ ∈ Irr(G, b), afforded by a KG-module M . The product jM of j and M is an Ohui-module,
whose character is denoted by χα . Denote by POG (x) the set of all α such that xα is a local pointed
element on OG. By [27, Corollary 4.4], for any p′ -element s of CG (x), we have
X
3.4.1
χ(xs) =
χα (x)ϕα (s).
α∈POG (x)
If xα is not contained in G{b} , then χα (x) is equal to 0.
3.5. By [27, Theorem 1.2], all maximal local pointed groups contained in G{b} are G-conjugate
to each other. Let Pγ be a maximal local pointed group contained in G{b} , choose i ∈ γ and set
Aγ = iAi. Obviously Aγ is an O-algebra and admits a group homomorphism ψ : P → A∗γ , u 7→ ui.
So Aγ is an interior P -algebra. Such an interior P -algebra Aγ is called a source algebra of the block
algebra OGb. By [27, Corollary 3.5], Aγ and OGb are Morita equivalent. In particular, given a
simple KG-module M in the block b, iM is a simple K ⊗O Aγ -module, and the correspondence
3.5.1
M 7→ iM
gives a bijection between the set of isomorphism classes of simple KG-modules in the block b and
the set of isomorphism classes of simple K ⊗O Aγ -modules.
We say that the blocks b and b′ are basically Morita equivalent if there is an O(G × G′ )-module
with endopermutation source, inducing a Morita equivalence between the block algebras OGb and
OG′ b′ . In this case, such a Morita equivalence is called a basic Morita equivalence between OGb
and OG′ b′ (see [23]).
Proposition 3.6. Let G and G′ be finite groups and b and b′ blocks of G and G′ . Assume that the
blocks b and b′ are basically Morita equivalent and that defect groups of b and b′ are abelian. Then
there is a Morita equivalence between the block algebras OGb and OG′ b′ , which induces a perfect
isometry Iν : ZIrr(G′ , b′ ) → ZIrr(G, b), so that IνK is compatible with a local system {IT }{T (cyclic)⊂P } .
Proof. Since b and b′ are basically Morita equivalent, by [23, Corollary 7.4] they have isomorphic
defect groups. For convenience, we assume without loss of generality that b and b′ have a common
defect group P . By [23, 6.9.3 and Corollary 7.4], there are maximal local pointed groups Pγ on
OGb and Pγ ′ on OG′ b′ and an endopermutation OP -module N with vertex P , such that setting
(OG)γ = i(OG)i and (OG′ )γ ′ = i′ (OG′ )i′ for some i ∈ γ and some i′ ∈ γ ′ , we have an interior
P -algebra embedding
f : (OG)γ → EndO (N ) ⊗O (OG′ )γ ′ .
Explicitly, f is an injective O-algebra homomorphism, the image of f is equal to f (i) EndO (N ) ⊗O
(OG′ )γ ′ f (i), and f preserves interior P -algebra structures (which is an interior P -algebra homomorphism in the sense of [27, Definiton 3.1]); the interior P -algebra structure on EndO (N ) ⊗O
7
(OG′ )γ ′ is determined diagonally by the interior P -algebra structures on EndO (N ) and (OG′ )γ ′ .
We identify (OG)γ as its image through f , so that
3.6.1
(OG)γ = i EndO (N ) ⊗O (OG′ )γ ′ i.
Since b and b′ are basically Morita equivalent, (OG)γ and (OG′ )γ ′ have the same numbers of
isomorphism classes of simple (OG)γ - and (OG′ )γ ′ -modules. Clearly EndO (N ) ⊗O (OG′ )γ ′ and
(OG′ )γ ′ have the same numbers of isomorphism classes of simple (EndO (N ) ⊗O (OG′ )γ ′ )- and
(OG′ )γ ′ -modules. So (OG)γ and EndO (N ) ⊗O (OG′ )γ ′ have the same numbers of isomorphism
classes of simple (OG)γ - and (EndO (N ) ⊗O (OG′ )γ ′ )-modules. By [30, Theorem 9.9], the correspondence
3.6.2
W → i(N ⊗O W )
gives a Morita equivalence between the categories of (OG′ )γ ′ - and (OG)γ -modules. By composing
the Morita equivalences between OG′ b′ and (OG′ )γ ′ (see 3.5), between (OG′ )γ ′ and (OG)γ and
between (OG)γ and OGb (see 3.5), we get a new Morita equivalence between OG′ b′ and OGb.
We remind that the new Morita equivalence between OG′ b′ and OGb may be different from
the Morita equivalence between OG′ b′ and OGb in the assumption. For a K-algebra A, we denote
by M(A) the set of isomorphism classes of simple A-modules. The purpose of constructing such a
new Morita equivalence is to get the following commutative diagram
M(KG′ b′ )
→
M(KGb)
↓
↓
′
M(K ⊗O (OG )γ ′ ) → M(K ⊗O (OG)γ ),
where the vertical arrows are obtained by the correspondence 3.5.1 applied to OGb and OG′ b′ , the
top arrow is induced by the new Morita equivalence between OGb and OG′ b′ and the bottom arrow
is determined by the correspondence 3.6.2. The diagram is a consequence of the new Morita equivalence between OGb and OG′ b′ in the last paragraph. The bijection M(KG′ b′ ) → M(KGb) induced
by the Morita equivalence in the assumption may not make the above diagram commutative.
We say that a local pointed group Qδ on OG is associated to a Brauer pair (Q, g), if
BrQ (f )BrQ (δ) = BrQ (δ)
and that a local pointed element uα is associated to a Brauer element (u, g) if the local pointed
group huiδ is associated to the Brauer pair (hui, g). Let (P, f ) and (P, f ′ ) be Brauer pairs, to which
Pγ and Pγ ′ are associated. Let M be a simple KGb-module and χ its ordinary character. Let
(u, h) be a Brauer element contained in (P, f ). Since P is abelian, any local pointed element uα
associated to the Brauer element (u, h) is contained in Pγ and conversely any local pointed element
uα contained in Pγ is associated to (u, h). By 3.4.1, we have
X
(u, h)
χα (u)ϕα ,
3.6.3
dG (χ) =
uα ∈P(u, h)
where P(u, h) denotes the set of all local pointed elements associated to the Brauer element (u, h).
Let M ′ be a simple KG′ b′ -module corresponding to M under the bijection M(KG′ b′ ) → M(KGb),
and χ′ its ordinary character. Let (u, h′ ) be the Brauer pair contained in (P, f ′ ). Similarly, we have
X
′
(u, h′ )
3.6.4
dG′ (χ′ ) =
(χ′ )α (u)ϕα′ .
uα′ ∈P(u, h′ )
8
Given uα ∈ P(u, h), by [30, Proposition 15.1], the intersection α ∩ (OG)γ is still a local point
of hui on (OG)γ and the correspondence α 7→ α ∩ (OG)γ gives a bijection between P(u, h) and the
set of local points of hui on (OG)γ . Through the bijection, we identify P(u, h) as a set of local
pointed elements on (OG)γ . By [23, 7.6.2] applied to the equality 3.6.1, we get a bijection
P(u, h) → P(u, h′ ), uα 7→ uα′
where uα and uα′ correspond to each other if and only if there are j ∈ α and j ′ ∈ α′ such that
j(ℓ ⊗ j ′ ) = j = (ℓ ⊗ j ′ )j,
where ℓ is a primitive idempotent in the unique local point of hui on EndO (N ). We are going to
prove that
′
χα (u) = ω(u)(χ′ )α (u)
3.6.5
under the bijection, where ω denotes the character of the Ohui-module ℓ(N ).
Since the KG-module M and the KG′ -module M ′ correspond to each other under the bijection
M(KG′ b′ ) → M(KGb) induced by the new Morita equivalence between OGb and OG′ b′ , by the
construction of the new Morita equivalence and 3.6.2, we have
iM ∼
= i(N ⊗O i′ M )
as K ⊗O (OG)γ -modules. On the other hand, since uα is contained in Pγ , there is j ′′ ∈ α ∩ (OG)γ
such that ij ′′ = j ′′ = j ′′ i. So
j ′′ M ∼
= j ′′ (N ⊗O i′ M )
as (K ⊗O (OG)α )-modules, where (OG)α = j ′′ (OG)j ′′ . We note that both j and j ′′ are inside
α ∩ (OG)γ and thus that the idempotents j ′′ and j are conjugate in (EndO (N ) ⊗ (OG′ )γ ′ )hui . In
particular, we have
j ′′ (N ⊗O i′ M ) ∼
= j(N ⊗O i′ M )
as Ohui-modules. Now, summarizing the above three isomorphisms, we conclude that the value of
χα at u is equal to the value at u of the character of the Ohui-module j(N ⊗O i′ M ).
Since j(ℓ ⊗ j ′ ) = j = (ℓ ⊗ j ′ )j, we have j(N ⊗O i′ M ) = j(ℓ(N ) ⊗O j ′ M ). By [25, 5.6.3],
EndO (N )⊗O (OG′ )
′
γ
Brhui
(ℓ ⊗ j ′ − j) = 0. So the Ohui-module (ℓ ⊗ j ′ − j)(N ⊗O i′ M ) is the direct sum
of indecomposable Ohui-modules with vertex properly contained in hui, its character has the value
0 at u, and the value at u of the character of the Ohui-module j(N ⊗O i′ M ) is equal to the value
′
at u of the character of the the Ohui-module ℓ(N ) ⊗O i′ M , which is equal to ω(u)(χ′ )α (u). Until
now, the equality 3.6.5 is proved.
Since the bijection Irr(G′ , b′ ) → Irr(G, b), χ′ 7→ χ is induced by the new Morita equivalence
between OGb and OG′ b′ , it can be extended to a perfect isometry Iν : ZIrr(G′ , b′ ) → ZIrr(G, b) by
the Z-linearity. We define a K-linear isomorphism
Iu : CF(CG′ (u), h′ , K) → CF(CG (u), h, K),
which sends ϕα′ onto ω(u)ϕα if local pointed elements uα ∈ P(u, h) and uα′ ∈ P(u, h′ ) correspond
(u, h)
under the bijection P(u, h) → P(u, h′ ). By the equalities 3.6.3, 3.6.4 and 3.6.5 we have dG ◦IνK =
′
(u, h )
Iu ◦ dG′ . In particular, IνK is compatible with a local system {Iu |u ∈ P }.
9
4. Reduction
In the section, we give several lemmas for the reduction of the proof of Theorem.
4.1. Let G be a finite group and b a block of G with maximal b-Brauer pair (P, f ). Obviously
NG (P, f ) acts on the simple algebra k ⊗O OCG (P )f by the conjugation. We denote by N̂G (P, f ) the
set of all pairs (x, ax ), where x ∈ NG (P, f ) and ax is an invertible element of k ⊗O OCG (P )f such
that dx = dax for all d ∈ k ⊗O OCG (P )f . Obviously N̂G (P, f ) is a subgroup of the direct product
N̂G (P, f ) × (k ⊗O OCG (P )f )∗ . The maps k∗ → N̂G (P, f ), λ 7→ (1, λ) and CG (P ) → N̂G (P, f ), x 7→
(x, x(1 ⊗ f )) are injective group homomorphisms. The image of k∗ is central in N̂G (P, f ), the
image of CG (P ) is normal in N̂G (P, f ) and they intersect trivially. We identify k∗ and CG (P ) with
their images through the two homomorphisms. The quotient of N̂G (P, f ) by k∗ is isomorphic to
NG (P, f ) and N̂G (P, f ) is a central extension of NG (P, f ) by k∗ . Set EG (P, f ) = NG (P, f )/CG (P )
and ÊG (P, f ) = N̂G (P, f )/CG (P ). The quotient group EG (P, f ) is the inertial quotient of the block
b if P is abelian. The homomorphism k∗ → N̂G (P, f ), λ 7→ (1, λ) induces a new injective group
homomorphism k∗ → ÊG (P, f ), and the quotient of ÊG (P, f ) by the image of k∗ is isomorphic to
EG (P, f ). So ÊG (P, f ) is a central extension of EG (P, f ) by k∗ .
4.2. In this section, we always assume that P is abelian and that the central extension ÊG (P, f )
splits. The NG (P, f )-conjugation induces actions of EG (P, f ) on P . Set K = P ⋊ EG (P, f ).
We say that a perfect isometry Iν : ZIrr(K) → ZIrr(G, b) is compatible with the ∗-structure if
Iν (λ ∗ χ) = λ ∗ Iν (χ) for any χ ∈ Irr(K) and any NG (P, f )-stable character λ of P , and that a
Morita equivalence between OGb and OK is a Morita equivalence compatible with the ∗-structure
if the Morita equivalence induces a perfect isometry compatible with the ∗-structure.
4.3. Let N be a normal subgroup of G and h a block of N covered by the block b. Let H be
the stabilizer
P of h under the G-conjugation action. Then there is a suitable block e of H such
that b = x xex−1 , where x runs over a set of representatives of left cosets of H in G, and such
that eex = 0 for any x ∈ G − H. The blocks b and e have common defect groups and the block
algebra OGb is isomorphic to the induction IndG
H (OHe) of the block algebra OHe (see [30, §16]).
In particular, there is a Morita equivalence between the block algebras OGb and OHe induced
by an indecomposable O(G × H)-module with trivial sources. We adjust the choice of P , so that
P is also a defect group of the block e. By [23, 7.6.5], ÊG (P, f ) is isomorphic to ÊH (P, f ′ ) as
central extensions for some maximal e-Brauer pair (P, f ′ ). In particular, EG (P, f ) is isomorphic to
EH (P, f ′ ). We identify EG (P, f ) and EH (P, f ′ ). Since the central extension ÊG (P, f ) is assumed
to split, ÊH (P, f ′ ) splits.
Lemma 4.4. Keep the notation and the assumptions in 4.2 and 4.3. Assume that there is a
Morita equivalence between the algebras OHe and OK compatible with the ∗-structure. There is a
Morita equivalence between the block algebras OGb and OK compatible with the ∗-structure.
Proof. Since there is a Morita equivalence between the block algebras OGb and OHe induced by an
indecomposable O(G×H)-module with trivial sources, by Propositions 2.7 and 3.6, there is a Morita
equivalence inducing a perfect isometry Iν : ZIrr(H, e) → ZIrr(G, b) such that Iν (λ ∗ χ) = λ ∗ Iν (χ)
for any χ ∈ Irr(H, e) and any NG (P, f )-stable character of P . The composition of Iν and the perfect
isometry induced by the Morita equivalence between OHe and OK is a perfect isometry compatible
with the ∗-structure, which is induced by the composition of the latter Morita equivalence between
OGb and OHe and the Morita equivalence between OHe and OK.
10
Lemma 4.5. Keep the notation and the assumption in 4.2 and 4.3. Set Q = P ∩ N . Then Q is
a defect group of the block h. Assume that H = G and that the block h is nilpotent. Then there is
a finite group L such that
4.5.1. Q is normal in L and P is a Sylow p-subgroup of L,
4.5.2. the quotient groups G/N and L/Q are isomorphic, and
4.5.3. there are a p′ -central extension L̃ of L and a block ℓ of L̃ with defect group P , such that
the blocks b and ℓ are basically Morita equivalent.
Proof. This lemma follows from [14, Theorem 1.12].
By [23, 7.6.5], ÊG (P, f ) is isomorphic to ÊL̃ (P, f ′ ) as central extensions for some maximal ℓBrauer pair (P, f ′ ). In particular, EG (P, f ) is isomorphic to EL̃ (P, f ′ ). We identify EG (P, f ) and
EL̃ (P, f ′ ). Since the central extension ÊG (P, f ) is assumed to split, ÊL̃ (P, f ′ ) splits.
Lemma 4.6. Keep the notation and the assumptions in Lemma 4.5. Assume that there is a
Morita equivalence between the algebras OL̃ℓ and OK compatible with the ∗-structure. Then there
is a Morita equivalence between the algebras OGb and OK compatible with the ∗-structure.
Proof. One can prove the lemma, using the same proof method for Lemma 4.4.
5. Extensions of Morita equivalences
In this section, we mainly give an extension of Morita equivalences compatible with the ∗structure (see Proposition 5.9 below).
5.1. Let G be a finite group, H a normal subgroup of G with p-power index, and b a G-stable
block of H. Then b is also a block of G. We always assume that b as a block of G has abelian
defect group P . Set Q = P ∩ H. By [14, Proposition 5.3], Q is a defect group of the block b of
H and G is equal to P H. Let (P, f ) be a maximal Brauer pair of the block b of G. Denote by T
the subgroup of all EG (P, f )-fixed elements in P and by R the commutator [P, NG (P, f )]. We have
P = T × R, R = [P, NH (P, f )] ⊂ H, and R is contained in Q. In particular, we have G = T H. Set
E = EG (P, f ), K = P ⋊ E and L = Q ⋊ E. Then K = LT .
Lemma 5.2. Keep the notation and the assumption in 5.1. Any simple module in the block b of
H is extendible to G and any irreducible ordinary character in the block b of H is extendible to G.
Proof.
See [12, Theorem].
We label characters in Irr(K) as ζj , j ∈ J, and choose a subset I ⊂ J, such that all these
restrictions (ζi )L , i ∈ I, are exactly all characters in Irr(L). For any i ∈ I, set φi = (ζi )L . Then
Irr(L) = {φi |i ∈ I}.
Lemma 5.3. Keep the notation and the assumption in 5.1. Assume that Iν : ZIrr(K) →
ZIrr(G, b) is a perfect isometry compatible with the ∗-structure. Then there is a perfect isometry Iν ′ : ZIrr(L) → ZIrr(H, b) compatible the ∗-structure such that the diagram
ZIrr(K)
5.3.1
−→
Iν
↓
ZIrr(L)
ZIrr(G, b)
↓
Iν ′
−−→ ZIrr(H, b)
11
is commutative, where the two down arrows are determined by the restriction of characters. More′
over, ν ′ has a suitable extension ν̂ ′ to (H × L)∆(P ) such that ν = IndG×K
(H×L)∆(P ) (ν̂ ).
Proof.
Since Iν : ZIrr(K) → ZIrr(G, b) is a perfect isometry compatible with the ∗-structure,
5.3.2
Iν (λ ∗ χ) = λ ∗ Iν (χ)
for any χ ∈ ZIrr(K) and any NG (P, f )-stable character λ of P .
For any i, j ∈ I, there exist ǫi , ǫj ∈ {±1} and χi , χj ∈ Irr(G, b), such that ǫi χi = Iν (ζi ) and
ǫj χj = Iν (ζj ). By Lemma 5.2, the restrictions ψi and ψj of χi and χj to H are irreducible. Suppose
ψi = ψj for different i and j. Set S = T ∩ H. The quotient group G/H is naturally isomorphic
to T /S. Through the isomorphism, any linear character λ of T with kernel containing S can be
regarded as a character of G with kernel containing H. Since ψi = ψj , there exists a linear character
λ of T with kernel containing S, such that χi = χj λ. Obviously λ is EG (P, f )-stable, and by [5,
Corollary] χj λ is equal to λ ∗ χj . By 5.3.2, λǫj ζj = ǫi ζi , ǫi = ǫj and ζi = ζj . Thus we have φi = φj .
That contradicts with the characters φi and φj being different for different i and j. On the other
hand, denoting by Irr(T /S) the set of irreducible ordinary characters of T with kernel containing
S, by 5.3.2, we have Irr(G, b) = {λχi | i ∈ I; λ ∈ Irr(T /S)}. Therefore Irr(H, b) = {ψi |i ∈ I}.
By 5.3.2,
the perfect isometry Iν sends λζi onto
P P
Pǫi λχi for any i ∈ I and any λ ∈ Irr(T /S).
′
So ν = i∈I λ∈Irr(T /S) ǫi (λχi ) × (λζi )∗ . Set ν ′ = i∈I ǫi (χi × ζi∗ ). Then ν = IndG×K
(H×L)∆(P ) (ν ).
Considering ν ′ as a generalized character on H ×L by restriction, the map Iν ′ : ZIrr(L) → ZIrr(H, b)
sends φi onto ǫi ψi for any i ∈ I and it is a bijective isometry. Moreover, it is trivial to see that Iν
and Iν ′ satisfies the diagram 5.3.1. By the following lemma, the proof is done.
Lemma 5.4. Keep the notation and the assumptions in Lemma 5.3 and its proof. Then the map
Iν ′ is a perfect isometry compatible with the ∗-structure.
Proof. Let (x, fhxi ) be a b-Brauer element contained in (P, f ). The pair (P, f ) is also a maximal bhxi -Brauer pair and the quotient group NCG (x) (P, f )/CG (P ) is isomorphic to CE (hxi). Set
Khxi = P ⋊ CE (hxi). By Proposition 2.7, there is a K-linear isomorphism Ihxi : CFp′ (Khxi , K) →
CFp′ (CG (x), bhxi , K), such that
(x, bhxi )
Ihxi ◦ dxK = dG
5.4.1
◦ IνK .
G
Denote by ResG
H the restriction map from CF(G, b, K) → CF(H, b, K) and by ResH, p′ the reG
striction map from CFp′ (G, b, K) → CFp′ (H, b, K). By Lemma 5.2 ResG
H is surjective and ResH, p′
is a K-linear isomorphism.
Assume that x lies in Q. Note that bhxi is a block of CH (x) with defect group Q. We set
5.4.2
(x, bhxi )
We claim that dH
5.4.3
K
C (x)
hxi
′
G
−1
Lhxi = Q ⋊ CE (hxi) and Ihxi
= ResCH
(x), p′ ◦ Ihxi ◦ (ResLhxi , p′ ) .
C (x)
′ ◦ dx . Indeed, it is easy to prove that the equalities
◦ IνK′ = Ihxi
L
(x, bhxi )
G
ResCH
(x), p′ ◦ dG
(x, bhxi )
= dH
K
hxi
x
x
K
◦ ResG
H and ResLhxi ◦ dK = dL ◦ ResL
12
hold. Since
(x, bhxi )
dH
(x, bhxi )
◦ IνK′ ◦ ResK
L
= dH
K
◦ ResG
H ◦ Iν
(x, bhxi )
C (x)
G
= ResCH
(x), p′ ◦ dG
◦ IνK
C (x)
G
x
= ResCH
(x), p′ ◦ Ihxi ◦ dK
K
′ ◦ Res hxi
x
= Ihxi
Lhxi , p′ ◦ dK
′ ◦ dx ◦ ResK
= Ihxi
L
L
where the first equality is obtained by 5.3.1, the second by 5.4.3, the third by 5.4.1, the fourth
by 5.4.2 and the fifth by 5.4.3, the claim is done. Therefore IνK′ is compatible with a local system
′ |x ∈ P } in the sense of [3, Definition 4.3]. By Proposition 2.7, I K is compatible with the
{Ihxi
ν′
∗-structure.
In order to prove that Iν ′ is a perfect isometry, by [3, Lemma 4.5], it remains to prove that
Khxi
′ = ResCG (x)
−1
′
Ihxi maps CFp′ (Lhxi , O) onto CFp′ (CH (x), bhxi , O). Since Ihxi
CH (x), p′ ◦ Ihxi ◦ (ResLhxi , p′ ) , it
suffices to show that Ihxi maps CFp′ (Khxi , O) onto CFp′ (CG (x), bhxi , O). This follows from Lemma
2.8.
Set L′ = R ⋊ E. Then L is equal to the direct product L′ × T ′ , where T ′ is the subgroup of
all E-fixed elements in Q. Let Iψ′ : ZIrr(L′ ) → ZIrr(L′ ) be a perfect isometry compatible with the
∗-structure. Since any irreducible ordinary character of L is of the form φ×λ, where φ ∈ Irr(L′ ) and
λ ∈ Irr(T ′ ), we extend Iψ′ to a Z-linear map Iψ : ZIrr(L) → ZIrr(L) sending φ×λ onto Iψ′ (φ)×λ for
any φ ∈ Irr(L′ ) and any λ ∈ Irr(T ′ ). Set ψ̃ ′ = ψ ′ × 1∆(T ′ ) , since (L′ × L′ )∆(T ′ ) = (L′ × L′ ) × ∆(T ′ ).
Lemma 5.5. Keep the notation and the assumption as above. Then Iψ is a perfect isometry
L×L
′
compatible with the ∗-structure, ψ = Ind(L
′ ×L′ )∆(T ′ ) (ψ̃ ), and the diagram
−→
ZIrr(L)
Iψ
ZIrr(L)
↓
↓
Iψ ′
ZIrr(L′ )
−−→ ZIrr(L′ )
is commutative, where the downarrows are induced by the restriction of characters.
Proof.
The last statement of the lemma is obvious. Since
X
X
ψ=
(Iψ′ (φ) × λ) × (φ∗ × λ∗ ) = ψ ′ ×
φ∈Irr(L′ ) λ∈Irr(T ′ )
X
λ∈Irr(T ′ )
λ × λ∗ ,
L×L
′
it is easy to see ψ = Ind(L
′ ×L′ )∆(T ′ ) (ψ̃ ). It remains to show that Iψ is a perfect isometry compatible
with the ∗-structure.
Since any irreducible ordinary character of L is of the form φ × λ, where φ ∈ Irr(L′ ) and
λ ∈ Irr(T ′ ), there is an obvious group isomorphism
O
ZIrr(L) ∼
ZIrr(T ′ ).
= ZIrr(L′ )
Z
We identify ZIrr(L) and ZIrr(L′ )
have a K-linear isomorphism
N
Z ZIrr(T
′)
through the isomorphism. Since Q = R × T ′ , we also
CFp′ (Q, K) ∼
= CFp′ (R, K)
O
K
13
CFp′ (T ′ , K), 1◦Q 7→ 1◦R ⊗ 1◦T ′
where 1◦Q denotes
the unique irreducible Brauer character on Q. We identify CFp′ (Q, K) and
N
CFp′ (R, K) K CFp′ (T ′ , K). With the identifications as above, we have
5.5.1
Iψ = Iψ′ ⊗ IdZIrr(T ′ )
where IdZIrr(T ′ ) is the identity map on ZIrr(T ′ ), and
x
u
dxu
L = dL′ ⊗ dT ′
5.5.2
for any x ∈ R and any u ∈ T ′ . On the other hand, since Iψ′ is a perfect isometry compatible with
the ∗-structure, by Proposition 2.7, for any x ∈ R, there is a K-linear map Ix : CFp′ (CL′ (x), K) →
CFp′ (CL′ (x), K) such that
5.5.3
dxL′ ◦ IψK′ = dxL′ ◦ Ix .
Now set Ixu = Ix ⊗K IdCFp′ (T ′ , K) , where IdCFp′ (T ′ , K) is the identity map on CFp′ (T ′ , K). By
xu
K
5.5.1, 5.5.2 and 5.5.3, we have Ixu ◦ dxu
L = dL ◦ Iψ . By Proposition 2.7, Iψ is compatible with
the ∗-structure. Since Iψ′ is a perfect isometry, by Lemma 2.8, Ix maps CFp′ (CL′ (x), O) onto
CFp′ (CL′ (x), O). Thus for any x ∈ R and u ∈ T ′ , Ixu maps CFp′ (CL (xu), O) onto CFp′ (CL (xu), O).
By [3, Lemma 4.5], Iψ is a perfect isometry.
Lemma 5.6. Keep the notation and the assumption in 5.1. Assume that E is cyclic and acts
freely on R − {1} and that the map Iψ : ZIrr(L) → ZIrr(L) is a perfect isometry compatible
with the ∗-structure. Then ψ has a suitable extension ψ̃ to (L × L)∆(K), such that setting ψ̂ =
IndK×K
(L×L)∆(K) (ψ̃), the map Iψ̂ : ZIrr(K) → ZIrr(K) is a perfect isometry compatible with the ∗structure and makes the following diagram commutative
ZIrr(K) −→ ZIrr(K)
Iψ̂
5.6.1
↓
ZIrr(L)
↓
Iψ
−→
ZIrr(L)
where the downarrows are induced by the restriction of characters.
Proof. By Lemma 5.3, there is a generalized character ψ ′ of (L′ × L′ )∆(T ′ ) such that ψ =
L×L
′
′
′
′
′
′
Ind(L
′ ×L′ )∆(T ′ ) (ψ ), and such that, restricting ψ to L × L , the map Iψ ′ : ZIrr(L ) → ZIrr(L )
is a perfect isometry compatible with the ∗-structure and makes the following diagram commutative
ZIrr(L) −→ ZIrr(L)
Iψ
↓
↓
Iψ ′
ZIrr(L′ ) −−→ ZIrr(L′ )
where the two down arrows are determined by the restriction of characters.
L×L
′
Set ψ̂ ′ = ψ ′ × 1∆(T ′ ) and ψ ′′ = Ind(L
′ ×L′ )∆(T ′ ) (ψ̂ ). By Lemma 5.5 applied to Iψ ′ and the group
L, the map Iψ′′ : ZIrr(L) → ZIrr(L) is a perfect isometry compatible with the ∗-structure and the
diagram
ZIrr(L) −−→ ZIrr(L)
Iψ ′′
↓
ZIrr(L′ )
↓
Iψ ′
−−→
14
ZIrr(L′ )
is commutative where the downarrows are induced by the restriction of characters. Note (L′ ×
′
L′ )∆(K) = (L′ × L′ ) × ∆(T ). Set ψ̃ ′ = ψ ′ × 1∆(T ) and ψ ′′′ = IndK×K
(L′ ×L′ )∆(T ) (ψ̃ ). Similarly, the map
Iψ′′′ : ZIrr(K) → ZIrr(K) is a perfect isometry compatible with the ∗-structure and the diagram
ZIrr(K) −−→ ZIrr(K)
Iψ ′′′
↓
ZIrr(L′ )
↓
Iψ ′
−−→
ZIrr(L′ )
is commutative, where the downarrows are induced by the restriction of characters.
(L×L)∆(T )
Notice that the restriction of ψ̃ ′ to (L′ × L′ )∆(T ′ ) is equal to ψ̂ ′ . Set ψ̂ ′′ = Ind(L′ ×L′ )∆(T ) (ψ̃ ′ ).
′′
′′′
′′
The restriction to L × L of ψ̂ ′′ is equal to ψ ′′ and IndK×K
(L×L)∆(T ) (ψ̂ ) is equal to ψ . Moreover, Iψ
and Iψ′′′ are perfect isometries compatible with the ∗-structure and making the following diagram
commutative
ZIrr(K) −−→ ZIrr(K)
Iψ ′′′
↓
ZIrr(L)
↓
Iψ ′′
−−→
ZIrr(L)
is commutative, where the downarrows are induced by the restriction of characters. Therefore, to
extend Iψ to a perfect isometry from ZIrr(K) to itself compatible with the ∗-structure satisfying
the diagram 5.6.1, is equivalent to extend the composition Iψ ◦ Iψ−1
′′ to a perfect isometry from
ZIrr(K) to itself compatible with the ∗-structure satisfying the diagram 5.6.1.
′
We replace Iψ by the composition Iψ ◦ Iψ−1
′′ . Then, for any φ ∈ Irr(L ), there is some suitable
λφ ∈ Irr(T ′ ) such that Iψ (φ × 1T ′ ) = φ × λφ . We claim that λφ = λφ′ for any φ, φ′ ∈ Irr(L′ ).
In order to prove the claim, we take φ′ to be 1L′ . Since the degree of φ is coprime to p, by [17,
Theorem 5.16], there is some u ∈ R − {1} such that φ(u) 6= 0. Note that CL (uv) = Q since E acts
freely on R − {1}. For any v ∈ T ′ , since Iψ is a perfect isometry compatible with the ∗-structure,
there is a K-linear isomorphism Ihuvi : CFp′ (Q, K) → CFp′ (Q, K) such that
uv
K
Ihuvi ◦ duv
L = dL ◦ Iψ .
Applying both sides of the equality to φ × 1T ′ and φ′ × 1T ′ , we get Ihuvi (φ(u)1◦Q ) = φ(u)λφ (v)1◦Q ,
Ihuvi (φ′ (u)1◦Q ) = φ′ (u)λφ′ (v)1◦Q , and thus λφ (v) = λφ′ (v) for any v ∈ T ′ . The claim is done.
We set π = λ1L′ and fix an extension ̟ of π to T . We extend Iψ to a Z-linear map Iψ̂ :
ZIrr(K) → ZIrr(K) sending φ × λ onto φ × ̟λ for any φ ∈ Irr(L′ ) and any λ ∈ Irr(T ). It is trivial
to verify that Iψ̂ is a perfect isometry compatible with the ∗-structure, that ψ̂ = IndK×K
(L×L)∆(T ) (ψ̃)
for some suitable extension ψ̃ of ψ to (L × L)∆(T ), and that Iψ and Iψ̂ satisfy the diagram 5.6.1.
Proposition 5.7. Keep the notation and the assumption in 5.1 and assume that E is cyclic and
acts freely on R − {1} and that Iν ′ : ZIrr(L) → ZIrr(H, b) is a perfect isometry compatible with
the ∗-structure. Then there is a suitable extension ν̃ ′ of ν ′ to (H × L)∆(P ), such that setting
′
ν = IndG×K
(H×L)∆(P ) (ν̃ ), the map Iν : ZIrr(K) → ZIrr(G, b) is a perfect isometry compatible with the
∗-structure and the diagram is commutative
ZIrr(K)
−→
Iν
↓
ZIrr(L)
ZIrr(G, b)
↓
Iν ′
−−→ ZIrr(H, b)
15
where the down arrows are induced by the restriction of characters.
Proof. Denote by ℓG (b) the number of isomorphism classes of simple modules in the block b of G.
Since ℓH (b) is equal to |EG (P, f )| and ℓG (b) is equal to ℓH (b) (see Lemma 5.2), ℓG (b) is equal to
EG (P, f ). By [32, Theorem], there is a perfect isometry ZIrr(K) → ZIrr(G, b) compatible with the
∗-structure. By Lemma 5.3, there is a perfect isometry Iµ′ : ZIrr(L) → ZIrr(H, b) compatible with
G×K
′
′
′
the ∗-structure, µ = Ind(H×L)∆(P
) (µ̃ ) for some extension µ̃ of µ to (H × L)∆(P ), and the diagram
ZIrr(K)
−→
Iµ
↓
ZIrr(L)
ZIrr(G, b)
↓
Iµ′
−−→ ZIrr(H, b)
is commutative, where the two down arrows are determined by the restriction of characters.
−1
Obviously Iµ−1
′ ◦Iν ′ is a perfect isometry compatible with the ∗-structure. Suppose Iη ′ = Iµ′ ◦Iν ′ .
′
By Lemma 5.6, η ′ has an extension η̃ ′ to (L × L)∆(K), such that setting η = IndK×K
(L×L)∆(K) (η ), the
map Iη : ZIrr(K) → ZIrr(K) is a perfect isometry compatible with the ∗-structure and makes the
following diagram commutative
ZIrr(K) −→ ZIrr(K)
Iη
↓
ZIrr(L)
↓
Iη ′
−−→
ZIrr(L)
where the downarrows are induced by the restriction of characters. Now we compose Iµ and Iη and
suppose Iν = Iµ ◦ Iη . Then Iν is a perfect isometry compatible with the ∗-structure, and it makes
commutative the diagram in the proposition. Notice that Iν uniquely determines Iν ′ . By Lemma
′
5.3, ν ′ has an extension ν̃ ′ to (H × L)∆(P ) and ν = IndG×K
(H×L)∆(P ) (ν̃ ).
5.8.
Assume that there are perfect isometries
Iν : ZIrr(K) → ZIrr(G, b) and Iν ′ : ZIrr(L) → ZIrr(H, b),
G×K
′
′
′
′
that ν = Ind(H×L)∆(P
) (ν̃ ) for a suitable extension ν̃ of ν to (H × L)∆(P ), and that Iν and Iν
make the following diagram commutative
5.8.1
ZIrr(K) → ZIrr(G, b)
↓
↓
ZIrr(L) → ZIrr(H, b)
where the downarrows are induced by the restrictions of characters. By [3, Theorem 1.5], the
perfect isometries Iν and Iν ′ induce algebra isomorphisms
π : Z(OL) ∼
= Z(OHb) and τ : Z(OK) ∼
= Z(OGb).
Obviously Z(OGb) and Z(OK) are P/Q-graded algebras with the ū-components Z(OGb) ∩ OHu
and Z(OK) ∩ OLu, where ū ∈ P/Q and u is a representative of ū in P .
Lemma 5.9. Keep the notation and the assumption in 5.8. Then τ is a P/Q-graded algebra
isomorphism and π is the restriction of τ to the 1-components of Z(OGb) and Z(OK).
Proof. For
on G such
P any r ∈ Z(KGb) and any s ∈ Z(KK), denote by r(g) the class function
P
−1 )t. We
that r = g∈G r(g−1 )g, and by s(t) the class function on K such that s =
s(t
t∈K
16
P
P
1
define a K-linear map Rν◦ : Z(KGb) → Z(KK), r 7→ t∈K |G|
ν(g,
t)r(g)
t−1 , and a Kg∈G
P
1 P
−1
linear map Iν◦ : Z(KK) → Z(KGb), s 7→ g∈G |K|
t∈K ν(g, t)s(t) g . By [3, Theorem 1.5],
′
the isomorphism τ maps r onto Iν◦ (rRν◦ (b)) for any r ∈ Z(OK). Since ν = IndG×K
(H×L)∆(P ) (ν ), it
◦
◦
is trivial to prove that Iν and Rν are P/Q-graded linear maps and thus that τ is a P/Q-graded
algebra isomorphism.
We extend the isomorphisms τ and π by the K-linearity, and get K-algebra isomorphisms
K
τ : Z(KK) ∼
= Z(KGb) and π K : Z(KL) ∼
= Z(KHb). Note that Z(KL) is the 1-component of
Z(KK). Since τ K (Z(KL)) = Z(KGb) ∩ Z(KHb) and the algebras Z(KL) and Z(KHb) have the
same K-dimension, we have τ K (Z(KL)) = Z(KHb).
Since P is abelian, by Lemma 5.2, any character χ ∈ Irr(H, b) is extendible to G. Consequently,
for any χ ∈ Irr(H, b) and any χ̂ ∈ Irr(G, b), χ̂ is an extension of χ if and only if eχ eχ̂ 6= 0.
Fix χ ∈ Irr(L) and χ̂ ∈ Irr(K) such that χ̂ is an extension of χ. Since eχ is primitive in Z(KL),
τ K (eχ ) = eϕ for a character ϕ ∈ Irr(H, b). Similarly, τ K (eχ̂ ) = eϕ̂ for some ϕ̂ ∈ Irr(G, b). By the
proof of [3, Theorem 1.5], Iν (χ̂) = ±ϕ̂. Then we use the diagram 5.8.1 and obtain Iν ′ (χ) = ±ϕ.
Again, by the proof of [3, Theorem 1.5], we have π K (eχ ) = eϕ . The proof is done.
Proposition 5.10. Keep the notation and the assumption in 5.1. Assume that E is cyclic and
acts freely on R − {1} and that there is a Morita equivalence between OHb and OL compatible
with the ∗-structure. Then there is a Morita equivalence between OGb and OK compatible with the
∗-structure.
Proof. We suppose that the O(H × L)-module M induces the Morita equivalence between OHb
and OL, and denote by µ′ the character of M . Then the map Iµ′ : ZIrr(L) → ZIrr(H, b) is a perfect
isometry compatible with the ∗-structure, which induces an algebra isomorphism
σ : Z(OL) ∼
= Z(OHb).
By Proposition 5.7, the character µ′ has a suitable extension µ̃′ to (H × L)∆(P ), such that setting
′
µ = IndG×K
(H×L)∆(P ) (µ̃ ), the map Iµ : ZIrr(K) → ZIrr(G, b) is a perfect isometry compatible with
the ∗-structure and the diagram
ZIrr(K)
−→
Iµ
↓
ZIrr(L)
ZIrr(G, b)
↓
Iµ′
−−→ ZIrr(H, b).
is commutative, where the two down arrows are induced by the restrictions of characters. Since Iµ′
maps ordinary characters of L onto ordinary characters of H, by the diagram, Iµ maps ordinary
characters of K onto ordinary characters of G. In particular, µ is an ordinary character of G × K.
The map Iµ induces an algebra isomorphism
ς : Z(OK) ∼
= Z(OGb).
By Lemma 5.9, the isomorphism ς is a P/Q-graded algebra isomorphism and the restriction of
ς to the 1-component of Z(OGb) coincides with σ. Since T is contained in Z(OK), for any
t ∈ T , ς(t) is contained in the t̄-component of Z(OGb) and tζ(t−1 ) is inside OHb. The map
ρ : T → (OHb)∗ , t 7→ tς(t−1 ) is a group homomorphism. Now we extend the O(H × L)-module
M to (H × L)∆(P ) by the equality (t, t) · m = ρ(t)m for any t ∈ T and any m ∈ M . We claim
17
that such an extension is well defined. Since O(H × L)-module M induces the Morita equivalence
between OHb and OL, by Lemma 2.10, we have σ(a)m = ma for any a ∈ Z(OL) and any m ∈ M .
So for any t ∈ T ∩ H and any m ∈ M , we have
(t, t)m = tmt−1 = tσ(t−1 )m = tς(t−1 )m = (t, t) · m.
The claim is done. Now by [15, Theorem 3.4 (a)], the O(G × K)-module IndG×K
(H×L)∆(P ) (M ) induces
a Morita equivalence between OGb and OK. Now it remains to show that the Morita equivalence
is compatible with the ∗-structure.
G×K
Set M̃ = Ind(H×L)∆(P
) (M ). The Morita equivalence induced by M̃ induces an algebra isomorphism (see Paragraph 2.9)
ρ : Z(OK) ∼
= Z(OGb)
such that for any a ∈ Z(OGb) and any a′ ∈ Z(OK), a and a′ correspond to each other if and only
if am = ma′ for any m ∈ M̃ . We claim that ρ and ς are the same. For any a ∈ Z(OL) and any
m ∈ M , we have
(σ(a) ⊗ 1) ⊗ m = (1 ⊗ 1) ⊗ σ(a)m = (1 ⊗ 1) ⊗ ma = (1 ⊗ a◦ ) ⊗ m,
where a◦ is the image of a through the opposite ring isomorphism OL → OL sending x onto x−1
for any x ∈ L. So ρ(a) = σ(a) = ς(a). On the other hand, given t ∈ T and m ∈ M , we have
(ς(t) ⊗ t) ⊗ m = (t ⊗ t)(t−1 ς(t) ⊗ 1) ⊗ m = (t ⊗ t) ⊗ t−1 ς(t)m = (1 ⊗ 1) ⊗ m
P
and thus (ς(t) ⊗ 1) ⊗ m = (1 ⊗ t−1 ) ⊗ m. So ρ maps t onto ς(t). Since Z(OK) = t∈T Z(OL)t,
the claim is done.
Finally, we claim that the character of IndG×K
(H×L)∆(P ) (M ) is equal to µ. In particular, the
Morita equivalence between OGb and OK induced by IndG×K
(H×L)∆(P ) (M ) induces a perfect isometry
compatible with the ∗-structure. Indeed, since the isomorphisms ρ and ς coincide with each other,
setting ρ(eχ ) = eχ′ for any χ
P∈ Irr(K), by the proof of [3, Theorem 1.5], both µ and the character
of M̃ are equal to the sum χ∈Irr(K) χ′ × χ∗ . The claim is done.
6. Proof of the Theorem
In this section, we prove Theorem in the introduction. We borrow the notation and the
assumption in Theorem. So p is 2 in the section. However, some results, such as Lemmas 6.2, 6.5
and 6.6, hold in general. Let (P, f ) be a maximal Brauer pair associated with the block b. The
inertial quotient EG (P, f ) may be of order 1, 3, 7 or 21. We break the proof into two propositions.
Lemma 6.1. Let n be a nilpotent block of a normal subgroup N of the group G covered by the
block b, and set N ′ = N Z(G). Then there is a nilpotent block n′ of N ′ covered by the block b.
Proof. We consider the group homomorphism N × Z(G) → N ′ , (x, y) 7→ xy and the induced
algebra homomorphism O(N × Z(G)) → ON ′ . Since the homomorphism O(N × Z(G)) → ON ′
maps n ⊗ 1 onto n, by the same proof method as that of Lemma 2.1, we prove that any block of
N ′ covering n is the image of the block n ⊗ w of N × Z(G) for some suitable block w of Z(G), that
its inertial quotient has order 1, and that it is nilpotent.
18
Lemma 6.2. Assume that there is a finite group G̃, such that G is normal in G̃ and G̃ has a
nilpotent block b̃ covering b. Then the block b and its Brauer correspondent in NG (P ) are basically
Morita equivalent.
Proof. By the main theorem of [18], the block b is inertial. That is to say, the blocks b and c are
basically Morita equivalent.
Lemma 6.3. Let G be a finite group and N a normal subgroup of G with index an odd prime.
Let b be a block of G with defect group P , and n a G-stable block of N covered by b. Assume that
P is Z2n × Z2n × Z2n , that the inertial quotient of the block b is of order 7 or 21 and that the block
n is not nilpotent. Then the inertial quotient of the block n is of order 7 or 21.
Proof. Since N is a normal subgroup of G with index an odd prime and b covers n, P is also a
defect group of n; moreover, we can choose a maximal n-Brauer pair (P, h) so that f covers h.
Suppose that N contains CG (P ). Then f = h, NN (P, h)/CN (P ) is a normal subgroup of
NG (P, f )/CG (P ), and NN (P, h)/CN (P ) has order 7 or 21 since the block n is not nilpotent.
Suppose that N does not contain CG (P ) and that CG (P ) = CG (P, h), where CG (P, h) is the
intersection of CG (P ) and the stabilizer NG (P, h) of (P, h) under the G-conjugation. Then h is a
central idempotent in OCG (P ) and we have hf = f . For any x ∈ NG (P, f ), we have xhx−1 f = f ,
xhx−1 = h and x ∈ NG (P, h). So NG (P, f ) ≤ NG (P, h). Since the index of N in G is an odd prime
and N does not contain CG (P ), G = N CG (P ) and NG (P, h) = NN (P, h)CG (P ). The inclusion
NG (P, f ) ⊂ NG (P, h) induces an injective group homomorphism
NG (P, f )/CG (P ) → NG (P, h)/CG (P ) ∼
= NN (P, h)/CN (P ).
This implies that the inertial quotient of the block n has order 7 or 21.
Suppose that N does not contain CG (P ) and that CG (P ) 6= CG (P, h). Since the index of
N in G is an odd prime q, the index of CN (P ) in CG (P ) is q and CN (P ) = CG (P, h). That
is toP
say, the stabilizer of h under the CG (P )-conjugation is CN (P ). Since f covers h, we have
f = x∈I xhx−1 , where I is a complete set of representatives of CN (P ) in CG (P ) and xhx−1 h = 0
for any x ∈ I outside CN (P ). Obviously we have NG (P, h) ⊂ NG (P, f ). For any y ∈ NG (P, f ), there
is z ∈ CG (P ) such that z −1 y centralizes h. So NG (P, f ) ⊂ NG (P, h)CG (P ) and then NG (P, f ) =
NG (P, h)CG (P ). Now it is easily seen that the inclusion NN (P, h) ⊂ NG (P, f ) induces an injective
group homomorphism
NN (P, h)/CN (P ) → NG (P, f )/CG (P ),
whose image is normal. So the inertial quotient of the block n has order 7 or 21.
Proposition 6.4. Assume that P is Z2n × Z2n × Z2n and that EG (P, f ) is of order 7 or 21. If
n ≥ 2, then the block b and its Brauer correspondent c in NG (P ) are basically Morita equivalent.
Proof. Suppose that b is a block of G with defect group P and the inertial quotient of order 7 or
21, such that (|G : Z(G)|, |G|) is minimal in the lexicographic ordering and that the blocks b and
c are not basically Morita equivalent. Then the block b has to be quasiprimitive.
Suppose that O2 (G) is not trivial. Set H = CG (O2 (G)). Let h be a block of H covered by the
block b. Then P is a defect group of the block h. Since CG (P ) ⊂ H, we may adjust the choice
of h so that (P, f ) is a maximal Brauer pair of the block h. By [19, Proposition 15.10], h is equal
to b and the quotient group G/H is an odd group. The inertial quotient EH (P, f ) is a normal
subgroup of EG (P, f ). Since EG (P, f ) is of order 7, EH (P, f ) is of order 1, or 7, or 21. Suppose
that EH (P, f ) is of order 1. Then the block b of H is nilpotent and by [33, Theorem], the block b
19
of G and the block c are basically Morita equivalent. That contradicts with the choice of the block
b. Suppose that EH (P, h) is of order 7 or 21. Then the commutator subgroup of P and EH (P, h) is
P . But since O2 (G) is in the center of H, that is impossible. So O2 (G) has to be trivial and Z(G)
is a subgroup of odd order.
Suppose that the block b covers a nilpotent block n of a normal subgroup N of G. Set N ′ =
N Z(G). By Lemma 6.1, there is a nilpotent block n′ covered by the block b. By Lemma 4.5 applied
to the normal subgroup N ′ and its block n′ and by the minimality of (|G : Z(G)|, |G|), N ′ has to
be equal to Z(G) and thus N is contained in Z(G).
Denote by E(G) the layer of the group G and by F (G) the Fitting subgroup of G. By the
choice of the block b, F (G) is equal to O2′ (G). Set F ∗ (G) = E(G)F (G), the generalized Fitting
subgroup of G. Suppose that E(G) is the central product L1 ∗ L2 ∗ · · · ∗ Lt , where L1 , L2 , · · · , Lt
are components of G. Since CG (F ∗ (G)) ≤ F ∗ (G), by the choice of the block b, t ≥ 1.
Let e be the unique block of E(G) covered by the block b and ei a block of Li covered by the
block e. Suppose t > 1. Since the block b has defect group Z2n × Z2n × Z2n , by Lemma 2.1 some
block ei has to be nilpotent and then the block b covers a nilpotent block of a normal subgroup
generated by all G-conjugates of Li . So Li ≤ Z(G) and this contradicts with Li being a component
of G. So t = 1 and then F ∗ (G) = L1 O2′ (G).
By the Schreier conjecture, G/L1 is sovable. We claim that G/L1 is trivial.
Suppose that G has a normal subgroup H of index 2. By [14, Proposition 5.3] G = P H. Since
P = [P, EG (P, f )], by [22, Proposition 4.2] P is contained in H and so G = H. That contradicts
with the choice of H. So G has no normal subgroup H of index 2.
Suppose that G has a normal subgroup H with an odd prime index containing L1 . Let h
be the unique block of H covered by b. Note that P is a defect group of the block h. By [33,
Theorem], the block h is not basically Morita equivalent to its Brauer correspondent in NH (P ). By
Lemma 6.3, the inertial quotient of the block h has order 7 or 21. This is against the minimality
of (|G : Z(G)|, |G|).
So G is equal to L1 . Since P is Z2n × Z2n × Z2n for n ≥ 2, only the case (iii) among the four
cases of [10, Theorem 6.1] happens for the block b of G. In this case, there is a finite group G̃, such
that G is normal in G̃ and G̃ has a nilpotent block b̃ covering b. By Lemma 6.2, the blocks b and
c are basically Morita equivalent. That contradicts with the choice of the block b.
Lemma 6.5. Let H and H ′ be finite groups and h and h′ blocks of H and H ′ . Assume that M
is an indecomposable O(H × H ′ )-module inducing a Morita stable equivalence between OHh and
OH ′ h′ . Let R̈ be a vertex of M and R and R′ the images of R̈ through the projections H × H ′ → H
and H × H ′ → H ′ . Then R and R′ are defect groups of h and h′ .
Proof.
This lemma follows from [23, Theorem 6.9].
Lemma 6.6. Let H, H ′ and H ′′ be finite groups, and h, h′ and h′′ blocks of H, H ′ and H ′′ .
Assume that R is a common p-subgroup of H, H ′ and H ′′ , that M is an indecomposable O(H ×H ′ )module with vertex ∆σ (R) = {(u, σ(u))|u ∈ R} for some group automorphism σ on R, and that
M ′ is an indecomposable O(H ′ × H ′′ )-module with vertex ∆σ′ (R) = {(u, σ ′ (u))|u ∈ R} for some
group automorphism σ ′ on R. Then the order of vertex of any indecomposable direct summand of
M ⊗OH ′ M ′ is less than |R|.
Proof. Let O∆σ (R)-module S be a source of M and O∆σ′ (R)-module S ′ a source of M ′ . Then
′
H ′ ×H ′′
′
′
M is a direct summand of IndH×H
∆σ (R) (x) and M is a direct summand of Ind∆ ′ (R) (S ). We have an
σ
20
O(H × H ′′ )-module isomorphism
′′
′
H ′ ×H ′′
′′
′
′
H×H ′
∼
IndH×H
R×R (OH ⊗O Resτ (S ) ⊗O Resτ ′ (S )) = Ind∆σ (R) (x) ⊗OH ′ Ind∆ ′ (R) (S ),
σ
mapping (x ⊗ y) ⊗ (z ⊗ s ⊗ s′ ) onto ((x ⊗ z) ⊗ s′ ) ⊗ ((1 ⊗ y) ⊗ s′ ) for any x ∈ OH, any y ∈ OH ′′ , any
z ∈ OH ′ , any s ∈ S and s′ ∈ S ′ , where τ is the homomorphism R × R → ∆σ (R), (u, v) 7→ (u, σ(u)),
τ ′ is the homomorphism R × R → ∆σ′ (R), (u, v) → (σ ′−1 (v), v) and OH ′ is an O(R × R)-module
defined by the equality
(u, v)w = σ(u)wσ ′ (v −1 )
R×R
(O), where R\H ′ /R denotes
for any u, v ∈ R and any w ∈ OH. Since OH ′ ∼
= ⊕x∈R\H ′ /R IndR
x
′
a set of representatives of the double cosets of R in H and Rx = {(σ −1 (z), σ ′−1 (x−1 zx))|z ∈
xRx−1 ∩ R} for any x ∈ R\H ′ /R, the order of vertex of any indecomposable direct summand of
′
H ′ ×H ′′
′
′
′
′
IndH×H
∆σ (R) (x) ⊗OH Ind∆ ′ (R) (S ) is less than |R|, and so is that of M ⊗OH M .
σ
Proposition 6.7. Let h be a block of H with defect group R = Z2n × Z2n , where n ≥ 2. Then
the block h and its Brauer correspondent d in NH (R) are basically Morita equivalent.
Proof. Let (R, e) be a maximal h-Brauer pair and set K = R ⋊ EH (R, e). By the structure
theorem of blocks with normal defect groups, there is a Morita equivalence between ONH (P )d and
OK induced by a p-permutation bimodule. Therefore, in order to prove the lemma, it suffices to
show that OHh and OK are basically Morita equivalent.
By [10], OHh and OK are Morita equivalent. Suppose that O(H × K)-module M induces
the Morita equivalence. By [24, Remarque 6.8] there is a stable equivalence between OHh and
OK induced by an indecomposable O(H × K)-module M̃ with endopermutation module sources
and vertex ∆(R). Denote by M̃ ∗ the dual of the O(H × K)-module M̃ . Then M̃ ∗ ⊗OG M is
an indecomposable O(K × K)-module inducing a Morita stable self-equivalence on OK. By [6,
Corollary 3.3], there is an integer r such that Ωr (M̃ ∗ ⊗OG M ) induces a Morita self-equivalence on
OK, where Ω denotes the Heller translate. By [35, Theorem 2] and [23, Corollary 7.4], Ωr (M̃ ∗ ⊗OG
M ) has vertex ∆σ (R) = {(u, σ(u))|u ∈ R} for some group automorphism σ on R, and so does
M̃ ∗ ⊗OG M . Since M is a direct summand of M̃ ⊗OK M̃ ∗ ⊗OG M and the order of vertex of any
indecomposable direct summand of M̃ ⊗OK M̃ ∗ ⊗OG M is less than |R| (see Lemma 6.6), by Lemma
6.5 the order of vertex of M has to be equal to |R|. Then by [23, Corollary 7.4], M induces a basic
Morita equivalence between OHh and OK.
Lemma 6.8. Let G be a finite group and N a normal subgroup of G with index an odd prime.
Let b be a block of G with defect group P , and n a G-stable block of N covered by b. Assume that
P is Z2n × Z2n × Z2m , that the inertial quotient of the block b is of order 3 and that the block n is
not nilpotent. Then the inertial quotient of the block n has order 3 or 21.
Proof.
One uses the proof of Lemma 6.3 to prove the lemma.
Lemma 6.9. Let G be a finite group and b a block of G with defect group P . Assume that P is
Z2n × Z2n × Z2m and that the inertial quotient of the block b is of order 3. Then ℓG (b) is 3.
Proof.
See [11, Theorem 1.1 (i)] or alternatively use [31, Theorem 1] to prove the lemma.
Proposition 6.10. Assume that P is Z2n × Z2n × Z2m and that EG (P, f ) is of order 3. If n ≥ 2,
then there is a Morita equivalence between the block algebras OGb and ONG (P )c compatible with
the ∗-structure, where c denotes the Brauer correspondent of b in NG (P ).
21
Proof. Set K = P ⋊ EG (P, f ). By the structure theorem of blocks with normal defect groups,
there is a Morita equivalence between ONG (P )c and OK induced by a p-permutation bimodule.
Therefore, in order to prove the lemma, by Propositions 2.7 and 3.6, it suffices to show that there
is a Morita equivalence between OGb and OK compatible with the ∗-structure.
When m = 0, by Propositions 6.7, 2.7 and 3.6 there is a Morita equivalence between OGb and
OK compatible with the ∗-structure. We go by induction on m. Suppose that b is a block of G
with defect group P and the inertial quotient of order 3 such that (|G : Z(G)|, |G|) is minimal in the
lexicographical order and that there is not a Morita equivalence between OGb and OK compatible
with the ∗-structure. By Lemma 3.4, the block b has to be quasiprimitive.
Set H = CG (O2 (G)). Let h be the block of H covered by the block b. Then P is a defectgroup
of the block h. Since CG (P ) ⊂ H, we may assume without loss that (P, f ) is a maximal Brauer
pair of the block h. By [19, Proposition 15.10], h is equal to b and the quotient group G/H is an
odd group. The inertial quotient EH (P, f ) is a normal subgroup of EG (P, f ). Since EG (P, f ) is of
order 3, EH (P, f ) is of order 1 or 3. If EH (P, f ) is of order 1, then the block b of H is nilpotent
and by [33], OGb and OK are basically Morita equivalent. By Proposition 3.6, there is a Morita
equivalence between OGb and OK compatible with the ∗-structure. That contradicts with the
choice of the block b. So EH (P, f ) is of order 3 and NG (P, f ) = NH (P, f )CG (P ). By a Frattini
argument, we have G = NG (P, f )H. So G = H and O2 (G) is central.
Suppose that the block b covers a nilpotent block of a normal subgroup N of G. Set N ′ =
N Z(G). By Lemma 6.1, there is a nilpotent block n′ covered by the block b. By Lemma 3.5
applied to the normal subgroup N ′ and its block n′ and the minimality of (|G : Z(G)|, |G|), N ′ has
to be equal to Z(G) and thus N is contained in Z(G).
Denote by E(G) the layer of the group G and by F (G) the Fitting subgroup of G. Then F (G)
is equal to Z(G). Set F ∗ (G) = E(G)F (G), the generalized Fitting subgroup of G. Suppose that
E(G) is the central product L1 ∗ L2 ∗ · · · ∗ Lt , where L1 , L2 , · · · , Lt are the components of G. Since
CG (F ∗ (G)) ≤ F ∗ (G), by the choice of the block b, t ≥ 1.
Let e be the unique block of E(G) covered by the block b and ei a block of Li covered by the
block e. By the choice of the block b, the order of the inertial quotient of the block e may be 3, 7,
or 21. Suppose t > 1. By Lemma 2.1, some block ei has to be nilpotent and then the block b covers
a nilpotent block of a normal subgroup generated by all G-conjugates of Li . So Li is contained in
Z(G). This contradicts with Li being a component of G. So t = 1 and F ∗ (G) = L1 Z(G).
By the Schreier conjecture, G/L1 is sovable. We claim that G/L1 is trivial.
Suppose that G has a normal subgroup H of index 2. Since EG (P, f ) is of order 3, we may
assume without loss of generality that the commutator [P, EG (P, f )] is equal to Z2n × Z2n and that
the subgroup of EG (P, f )-fixed elements of P is Z2m . Let h be the unique block of H covered by
the block b. Then the intersection Q = P ∩ H is a defect group of the block h. By [22, Proposition
4.2] Z2n × Z2n ≤ Q and so Q is Z2n × Z2n × Z2m−1 since G = P H (see [14, Proposition 5.3]). By [34,
Lemma 3.6] the block h has the inertial quotient of order 3. By induction on m, there is a Morita
equivalence between OHh and OL compatible with the ∗-structure, where L = Q ⋊ EG (P, f ).
By Proposition 5.10 there is a Morita equivalence between OGb and OK compatible with the ∗structure. This contradicts with the choice of the block b. So G has no normal subgroup H of
index 2.
Suppose that G has a normal subgroup H of an odd prime index containing L1 . Let h be the
unique block of H covered by the block b. By Lemma 6.8, the inertial quotient of the block h has
order 3 or 21. By Proposition 6.4 and [33, Theorem], we may exclude the case that the inertial
quotient of the block h has order 21. Now by Lemma 6.9, ℓG (b) = ℓH (h) = 3. Since H has odd
prime index in G, by the last paragraph of the proof of [7, Theorem 1.1], that is impossible.
22
So G = L1 . Then as in the last paragraph of the proof of Proposition 6.4, we prove that
the block algebra OGb and OK are basically Morita equivalent. Then by Propositions 2.7 and
3.6, there is a Morita equivalence between OGb and OK compatible with the ∗-structure. That
contradicts with the choice of the block b.
6.11.
Proof of Theorem
When EG (P, f ) is 1, the block b of G is nilpotent and Theorem is true (see [25]). The rest of
Theorem easily follows from Propositions 6.4 and 6.10.
References
[1] F.W. Anderson, K.R. Fuller, Rings and Categories of Modules, Grad. Texts in Math., vol. 13,
Springer-Verlag, New York, 1974.
[2] M. Broué, Equivalences of blocks of group algebras, in: Finite Dimensional Algebras and
Related Topics, Kluwer, 1994, pp. 1-26.
[3] M. Broué, Isométries parfaites, types de blocs, catégories dérivées, Astérisque 181-182 (1990),
61-92.
[4] M. Broué, Radical, Hauteurs, p-Sections et Blocs, Ann. Math., 107, (1978), 89-107.
[5] M. Broué & L. Puig, Characters and local structure in G-algebras, J. Algebra 63 (1980),
306-317.
[6] J. F. Carlson & R. Rouquier, Self-equivalences of stable module categories. Math. Z. 233
(2000), 165-178.
[7] D. Craven & C. Eaton & R. Kessar & M. Linckelmann, The structure of blocks with a Klein
four defect group, Math. Z. 268 (2011), 441-476.
[8] E. C. Dade, Block extensions, Illinois J. Math. 17 (1973), 198-272.
[9] C. Eaton, Morita equivalence classes of 2-blocks of defect three, Proc. Amer. Math. Soc. 144
(2016), 1961-1970.
[10] C. Eaton & R. Kessar & B. Külshammer & B. Sambale, 2-blocks with abelian defect groups,
Adv. Math. 254 (2014), 706-735.
[11] X. Hu & Y. Zhou, 2-blocks with hyperfocal subgroup C2n × C2n , arXiv:1709.05983.
[12] R. Knörr, A remark on covering blocks. J. Algebra 103 (1986), 208-210.
[13] B. Külshammer, Morita equivalent blocks in Clifford theory of finite groups, Astérisque No.
181-182 (1990), 209-215.
[14] B. Külshammer & L. Puig, Extensions of nilpotent blocks, Invent. Math. 102 (1990), 17-71.
[15] A. Marcus, On equivalences between blocks of group algebras: reduction to the simple components, J. Algebra 184 (1996) 372–396.
23
[16] H. Nagao & Y. Tsushima, Representations of Finite Groups, Academic Press, Inc., Boston,
MA, 1989.
[17] G. Navarro, Characters and blocks of finite groups, Cambridge University Press, 1998.
[18] L. Puig, Nilpotent extensions of blocks, Math. Z. 269 (2011), 115-136.
[19] L. Puig, Frobenius Categories Versus Brauer Blocks, Birkhäuser, 2009.
[20] L. Puig, Block source algebras in p-solvable groups, Michigan Math. J. 58 (2009), 323-338.
[21] L. Puig, Source algebras of p-central group extensions, J. Algebra 235 (2001), 359-398.
[22] L. Puig, The hyperfocal subalgebra of a block, Invent. Math. 141 (2000), 365-397.
[23] L. Puig, On the local structure of Morita and Rickard equivalences between Brauer blocks,
Progress in Math., 178 (1999), Birkhäuser, Basel.
[24] L. Puig, Une correspondance de modules pour les blocs à groupes de défaut abéliens, Geom.
Dedicata 37 (1991), 9-43.
[25] L. Puig, Nilpotent blocks and their source algebras, Invent. Math. 93 (1988), 77-116.
[26] L. Puig, Pointed groups and construction of modules, J. Algebra 116 (1988), 7-129.
[27] L. Puig, Pointed groups and construction of characters, Math. Z. 176 (1981), 265–292.
[28] L. Puig & Y. Usami, Perfect isometries for blocks with abelian defect groups and Klein four
inertial quotients, J. Algebra 160 (1993), 192-225.
[29] J. Rickard, Splendid equivalences: derived categories and permutation modules, Proc. London
Math. Soc. (3) 72 (1996), 331-358.
[30] J. Thévenaz, G-algebras and Modular Representation Theory, Oxford Math. Monogr., Oxford
Science Publications, Clarendon Press, Oxford, 1995.
[31] A. Watanabe, Notes on p -blocks of characters of finite groups, J. Algebra 136 (1991), 109-116.
[32] A. Watanabe, On perfect isometries for blocks with abelian defect groups and cyclic hyperfocal
subgroups, Kumamoto J. Math. 18 (2005) 85-92.
[33] Y. Zhou, On the p′ -extensions of inertial blocks, Proc. Amer. Math. Soc. 144 (2016), 41-54.
[34] Y. Zhou, On the p-extensions of inertial blocks, J. Algebra 423 (2015), 538-549.
[35] Y. Zhou, Morita equivalences between some blocks for p-solvable groups, Proc. Amer. Math.
Soc. 133 (2005), 3133-3142
24
| 4math.GR
|
1
Effect of Cell-Selection on the Effective Fading
Distribution in a Downlink K -tier HetNet
arXiv:1701.01164v2 [cs.IT] 31 May 2017
Mustafa A. Kishk and Harpreet S. Dhillon
Abstract—This paper characterizes the statistics of effective
fading gain in multi-tier cellular networks with strongest base
station (BS) cell association policy. First, we derive the probability
of association with the n-th nearest BS in the k-th tier. Next, we
use this result to derive the probability density function (PDF)
of the channel fading gain (effective fading) experienced by the
user when associating with the strongest BS. Interestingly, our
results show that the effective channel gain distribution solely
depends upon the original channel fading and the path-loss
exponent. Moreover, we show that in the case of Nakagami-m
fading channels (Gamma distribution), the distribution of the
effective fading is also Gamma but with a gain of α2 in the shape
parameter, where α is the path-loss exponent.
Index Terms—Stochastic geometry, order statistics, heterogeneous cellular networks, Poisson Point Process, fading.
I. I NTRODUCTION
With the dawn of fifth-generation (5G) cellular systems
almost upon us, we are currently witnessing unprecedented
changes in the cellular use-cases and consequently the architecture of cellular networks. In the context of this letter, two
parallel trends are of particular interest. First, cell sizes are
shrinking due to the organic deployment of low-power BSs
called small cells [1]. Second, 5G may need to support mobility at speeds upto 500km/hour [2]. These trends together mean
that the mobile users may experience orders of magnitude
higher handover rates compared to the current networks. This
issue has indeed attracted a lot of attention, with cloud radio
access networks (C-RANs) and more recently Fog networks
presented as possible solutions [3]. The occurrence of more
frequent handovers necessitates the need to understand the
system characteristics during the handover phase, which we
refer to as the cell selection phase. In this letter, we focus
on characterizing the distribution of fading gain experienced
by a mobile user during the cell selection phase, which we
term as effective fading gain. Due to the occurence of order
statistics in cell selection, the effective fading gain is shown
to deviate significantly from the fading gain observed by
the users at other times (when they are not performing cell
selection), which is referred to as original fading gain. It is
worth noting that this problem has never been studied before
perhaps because it was not quite as meaningful in conventional
cellular networks that had far lower handover rates.
Although this exact problem has not been studied before,
order statistics have been used to study several problems that
have a somewhat similar flavor [4]. For instance, in multiantenna receivers, it is important to exploit the existence of
multiple paths by optimally combining them in order to enhance the performance gain. This is usually achieved by giving
larger weights to the strongest paths [5]–[7]. Also in multi-user
The authors are with Wireless@VT, Department of ECE, Virginia Tech,
Blacksburg, VA. Email: {mkishk, hdhillon}@vt.edu. The support of the U.S.
NSF (Grants CCF-1464293 and CNS-1617896) is gratefully acknowledged.
systems, similar procedure appears during scheduling when the
transmitter picks the users with strongest channel state [8]–
[10]. Similar to diversity techniques, selecting stronger users
during scheduling necessitates using order statistics in the
analysis. Similar decision needs to be made during cell selection in cellular networks, which is the main focus of this
letter. The usual procedure is to measure the received signal
strengths from all nearby BSs at the mobile user and then
attach it to the BS that provides the strongest signal. While
this typically reduces to closest-BS association in conventional
single-tier cellular networks, the same is not true in multitier HetNets due to different transmit powers across tiers [1].
Assuming same fading gain statistics across all wireless links,
in this letter we study the effect of cell selection procedure on
the effective fading gain observed at the mobile user.
Contributions. Using tools from stochastic geometry [11],
we characterize the PDF of the effective channel fading gain
experienced by the typical user in a multi-tier cellular network
during cell-selection phase under strongest-BS association
policy. As an important intermediate result, we first derive
the probability of association of the typical user with the
n-th nearest BS in the k-th tier, which itself is a new
and much finer characterization of association probabilities
under strongest BS association policy in HetNets. Our results
concretely demonstrate that the PDF solely depends upon
the distribution of the original fading gain and the path-loss
exponent (does not depend upon infrastructure properties, such
as the number of BS tiers, their densities, or their transmission
powers). In addition, if we assume independent Nakagamim fading (Gamma distribution) over all links, the resulting
effective fading is also Gamma distributed with the same scale
parameter but a gain of α2 in the shape parameter.
II. S YSTEM M ODEL
We consider a K-tier cellular network were the locations
of BSs in each tier are modeled by an independent Poisson
point process (PPP) Φk = {xk,i } ∈ R2 with density λk . The
transmission power of BSs in the k-th tier is Pk . TheSlocations
of all BSs can then be modeled by the PPP Ψ = k∈K Φk ,
where K = {1, 2, . . . , K} is the set of indices of the tiers.
The locations of the users are modeled by an independent
PPP Φu = {ui } ∈ R2 with density λu . Without loss of
generality, the analysis is performed at an arbitrarily chosen
user (termed typical user), whose position can be translated to
the origin due to the stationarity of this setup. The received
signal power at the typical user from a k-th tier BS located
at x ∈ Φk is Pk hx kxk−α , where hx models fading gain and
kxk−α represents standard power-law pathloss with exponent
α > 2. To differentiate it from the effective fading gain studied
in the next Section, we refer to hx as the original fading gain.
For notational ease, the set of all original fading gains {hx },
2
x ∈ Φk , ∀k ∈ K over all links between the typical user and
the BSs in the network is denoted by H. It is assumed that
hx is independent and identically distributed across all links
with PDF fh (h) and cumulative distribution function (CDF)
Fh (h). While all the results will be derived for this general
distribution, we will also specialize them to Nakagami-m
fading case for providing insights and numerical comparisons.
m y m−1
m
In this case, we have fh (y) = m
Ω
Γ(m) exp(− Ω y), y ≥ 0,
Ω
which is Gamma distribution with scale parameter
m and
Ω
shape parameter m (i.e. h ∼ Gamma m , m ), where Ω is the
mean of h (usually assumed to bem 1 in Nakagami-m fading
Γ(m, Ω y)
channels). Also Fh (y) = 1 − Γ(m)
. For cell selection, we
assume that the typical user connects to the BS that provides
maximum received power. For this setup, we define effective
fading gain experienced at the typical user as follows.
Definition 1 (Effective Fading Gain). The effective fading
gain h∗ represents the fading gain experienced by the typical
user during cell-selection phase when associating with the BS
located at x∗ , where
x∗ = arg max Pk hx kxk−α .
(1)
x∈Φk ,k∈K
III. D ISTRIBUTION OF E FFECTIVE FADING G AIN
Since fading gain appears in the cell selection criterion
given by Eq. 1, the distribution of the effective fading gain h∗
experienced by the typical user will, in general, be different
from the distribution of the original fading gain h (due to the
effect of order statistics). Before starting the derivation of the
PDF of h∗ , we introduce an indicator function δk,n for which
δk,n = 1 if the n-th nearest BS in the k-th tier is the serving
BS and δk,n = 0 otherwise. This function can be expressed as
Y
δk,n =
1(kxk,n k−α hk,n ≥ kxk k−α hxk )
(2)
xk ∈Φk \xk,n
Y
Y
1 Pk kxk,n k−α hk,n ≥ Pj kxj k−α hxj ,
j∈K\k xj ∈Φj
where 1(Ξ) is an indicator function that equals to 1 when the
condition Ξ is satisfied and equals to zero otherwise, and hk,n
is the original fading gain for the channel between the typical
user and the n-th nearest BS in the k-th tier. Using the above
function, the CDF of h∗ is defined as follows:
P(h∗ ≤ y) =
=
(a)
=
∞
XX
k∈K n=1
∞
XX
k∈K n=1
∞
XX
k∈K n=1
=
P(hk,n ≤ y, δk,n = 1)
EΨ,H [1(hk,n ≤ y)δk,n ]
∞ Z
XX
k∈K n=1
fh (hk,n )P(k,n)|hk,n dhk,n ,
(3)
0
A. Conditional Association Probability
We first derive an expression for P(k,n)|hk,n in Lemma 1.
Lemma 1 (Association Probability). The conditional probability of the typical user associating with the n-th nearest BS
in the k-th tier is:
n
1
g1 (hk,n )n−1 ,
(4)
P(k,n)|hk,n =
g2 (hk,n ) + 1
R1
2
where g1 (hk,n ) = α2 0 Fh(yhk,n ) y α −1 dy, g2 (hk,n )
=
R
R
2
2
∞
2
2 ∞
−1
−1
α
dy + Bk α 0 F̄h (yhk,n ) y α dy ,
α 1 F̄h (yhk,n ) y
2
λk Pk α
P
j∈K\k
2
λj (Pj ) α
, and F̄h (h) = 1 − Fh (h).
Proof: See Appendix A.
Remark 1. Note that the effect of BS density and their transmit
powers appears in the term Bk in g2 (hk,n ) in Eq. 4. Consistent
with the intuition, the conditional association probability is an
increasing function of both λk and Pk .
For the case of Nakagami-m original fading, the integrals
in g1 (hk,n ) and g2 (hk,n ) can be reduced to closed forms.
Corollary 1. In the case of Nakagami-m original fading,
g1 (hk,n ) and g2 (hk,n ) can be simplified as follows:
α2
mhk,n
Ω
2 mhk,n
γ
m
+
,
+
Γ
m,
mhk,n
α
Ω
Ω
g1 (hk,n ) = 1 −
Γ(m)
α2
Γ(m+ 2 )
mh
Ω
Γ m + α2 , Ωk,n + Bk α
mhk,n
g2 (hk,n ) =
Γ(m)
mhk,n
Γ m, Ω
−
,
(5)
Γ(m)
where Γ(a, b), and γ(a, b) are the upper and lower incomplete
Gamma functions, respectively.
B. Effective Fading Distribution Analysis
In the following Theorem we provide the PDF of the
effective fading gain h∗ in terms of general original fading.
Theorem 1 (Effective Fading Distribution with General Fading). The PDF of the effective fading gain in a K-tier network
under general cell-selection policy given by Definition 1 is:
fh∗ (y) =
Ehk,n [1(hk,n ≤ y)P(δk,n = 1|hk,n )]
y
where (a) follows from the fact that 1(hk,n ≤ y) and
δk,n are conditionally independent, conditioned on hk,n , and
P(k,n)|hk,n in (b) is the probability of association with the nth nearest BS in the k-th tier conditioned on hk,n . Hence, we
first derive an expression for P(k,n)|hk,n in Lemma 1, using
which we will derive the PDF of h∗ in Theorem 1.
Bk =
Note that, in principle, hx can be interpreted as the cumulative fading gain incorporating both slow and fast fading effects.
This generality is mathematically provided by the general
distribution of h. Also note that the cell-selection policy can
be specialized to both instantaneous and average power-based
cell selection by choosing appropriate distribution fh (·).
×
(b)
2
α α
2 y fh (y)
.
R∞
2
F̄h (z)z α −1 dz
0
Proof: See Appendix B.
(6)
3
f h*(y)
1
Original fading m=1
Simulation m=1
Theoretical m=1
0.5
0
0
1
2
3
4
5
6
7
f h*(y)
1
Original fading m=2
Simulation m=2
Theoretical m=2
0.5
0
f h*(y)
0
1
2
3
4
5
6
7
Original fading m=10
Simulation m=10
Theoretical m=10
1
0.5
0
0
1
2
3
4
5
6
the simulation results are also provided for K = 2 tiers. As
discussed already in Remark 2, these results do not depend
upon the number of tiers (value of K), which was also verified
separately with simulations. By comparing these distributions
with the corresponding original fading distributions, we deduce that the PDF of h∗ almost coincides with the PDF of
the original fading h as the value of m is increased, whereas
they are significantly different for smaller values of m. This
is consistent with insights gained in Remark 3.
7
y
Fig. 1. PDF of both original channel fading gain and effective channel fading
gain for Nakagami-m original fading case with different values of m (P1 = 1,
P2 = 2, λ1 = 1, and λ2 = 2). Results are independent of K.
Remark 2. Note that in Eq. 6 the distribution of the effective
channel fading gain is agnostic of the number of tiers, the
transmission powers, and the densities of the tiers. Interestingly, the distribution only depends on the original fading
distribution and the path-loss exponent α.
In the case of Nakagami-m fading, the distribution of the
effective fading gain is presented in the following corollary.
Corollary 2 (Effective Fading Distribution with Nakagami-m
Original Fading). In the case of Nakagami-m original fading
Ω
with h ∼ Gamma m
, m , we have
m m+ α2 2
exp(− m
Ω y)
fh∗ (y) =
y α +m−1
.
(7)
Ω
Γ(m + α2 )
Remark 3. From Corollary 2, we note that in the case of
Ω
, m , the distriNakagami-m original fading h ∼ Gamma m
bution of the effective fading gain is also Nakagami-m but
Ω
with a different shape parameter: h∗ ∼ Gamma( m
, m + α2 ).
2
The α gain in the shape parameter represents the decrease in
the severity of the effective fading gain distribution as a result
of base station selection. It is also noteworthy that this gain
gets negligible as the value of m gets larger.
Remark 4. The reduction in the perceived severity of the
fading at the receiver (discussed in Remark 3) can be used
to lower the transmission power of the BS for about one
channel coherence time immediately after the cell selection
phase. This can be argued using multiple performance metrics.
For instance, due to reduction in the channel fluctuations, one
needs a relatively lower fade margin during this time, which
directly translates to lower transmit power at the BS. Similarly,
due to less severe fading, the bit error rate will decrease, which
means the BS can transmit at a lower power to achieve the
same target bit error rate as before.
C. Simulation Results
In Fig. 1, we present the PDF of the effective fading gain
h∗ for the case of Nakagami-m original fading with Ω = 1
for several values of m. Along with the analytical results,
IV. C ONCLUSION
In this letter, we derived the PDF of the effective fading
gain observed by the typical user during the handover phase
in a K-tier HetNet under strongest-BS association policy.
As an intermediate result (which is in fact important in its
own right), we provided a much finer characterization of
association probabilities by deriving the probability that the
typical user is served by the n-th nearest BS in the k-th tier.
Our results concretely demonstrate that the effective channel
fading distribution during cell selection phase is significantly
different from the original fading distribution. In fact, association with the strongest BS reduces perceived severity of fading
in general due to order statistics. One useful consequence of
this result is the possibility of lowering the transmit power
of the serving BS to the typical user for about one channel
coherence time immediately after the cell selection phase.
A PPENDIX
A. Proof of Lemma 1
The association probability conditioned on hk,n can be
derived as follows:
"
Y
!xk,n
P(k,n)|hk,n = Exk,n EΦk EΨ\Φk EH
1(kxk,n k−α hk,n
xk ∈Φk \xk,n
≥ kxk k
−α
Y
hxk ) ×
1 Pk kxk,n k hk,n ≥ Pj kxj k hxj
Y
−α
−α
j∈K\k xj ∈Φj
(a)
=
"
!x
Exk,n EΦkk,n EΨ\Φk
Y
P(kxk,n k−α hk,n ≥ kxk k−α hxk )
xk ∈Φk \xk,n
#
Y
×
Y
−α
P Pk kxk,n k
hk,n ≥ Pj kxj k
−α
hxj
j∈K\k xj ∈Φj
(b)
=
"
!x
Exk,n EΦkk,n EΨ\Φk
Y
Fh
xk ∈Φk \xk,n
kxk kα
hk,n
kxk,n kα
#
Pk kxj kα
×
Fh
hk,n
Pj kxk,n kα
j∈K\k xj ∈Φj
"
"
#
Y
kxk kα
(c)
!xk,n
= Exk,n EΦk
Fh
hk,n
kxk,n kα
xk ∈Φk \xk,n
{z
}
|
Y
Y
I1 (xk,n )
"
Y
×
j∈K\k
|
EΦj
Y
Fh
xj ∈Φj
Pk kxj kα
hk,n
Pj kxk,n kα
{z
I2 (xk,n )
= Exk,n [I1 (xk,n )I2 (xk,n )] ,
##
}
(8)
#
4
where (a) is due to the independence of fading gain across all
links, (b) follows by substituting for the CDF of the original
fading gain Fh (h), and step (c) follows from the assumption
that the K-tiers are modeled by independent PPPs.
The term I1 (xk,n ) can be derived by splitting Φk into Φk ∩
B(0, kxk,n k) and Φk ∩ B(0, kxk,n k)c , where B(0, kxk,n k)
is the ball of radius kxk,n k centered at the origin, and
B(0, kxk,n k)c is its compliment. Note that conditioned on the
location of the n-th nearest BS xk,n , the number of points
inside the ball Φk ∩ B(0, kxk,n k) is n − 1, and their locations
are uniformly distributed (follows from the definition of PPP).
Hence, I1 (xk,n ) can be derived as follows:
"
#
Y
kxkα
!xk,n
hk,n
I1 (xk,n ) = EΦk
Fh
kxk,n kα
x∈Φk ∩B(0,kxk,n k)
"
#
Y
kxkα
!xk,n
× EΦk
Fh
hk,n
kxk,n kα
x∈Φk ∩B(0,kxk,n k)c
!n−1
Z
1
kxkα
(d)
=
Fh
hk,n dx
πkxk,n k2 x∈R2 ∩B(0,kxk,n k)
kxk,n kα
!
Z
kxkα
hk,n dx
× exp −λk
1 − Fh
kxk,n kα
x∈R2 ∩B(0,kxk,n k)c
n−1
Z 1
2
2
(e)
−1
α
dy
Fh (yhk,n ) y
=
α 0
!
Z ∞
2
2πλk rk,n
2
−1
× exp −
F̄h (yhk,n ) y α dy ,
(9)
α
1
where F̄h (h) = 1 − Fh (h). The first term in (d) is due to the
uniform distribution of the points inside Φk ∩ B(0, kxk,n k),
the second term in (d) follows by applying PGFL of PPP [12].
Note that the point process Φk ∩ B(0, kxk,n k)c remains a PPP
due to Slyvniak’s theorem. Step (e) follows from converting
to polar coordinates where rk,n = kxk,n k and simple manipulations to the integrals. The term I2 (xk,n ) can also be derived
using PGFL of PPP and similar procedure to steps (d) and (e)
which results in:
!
!
Z ∞
Y
Pk r α
I2 (xk,n ) =
exp −2πλj
F̄h
α hk,n rdr
Pj rk,n
0
j∈K\k
!
α2 Z ∞
2
Y
2πλj rk,n
2
Pj
−1
α
F̄h (yhk,n ) y
dy
=
exp −
α
Pk
0
j∈K\k
!
Z ∞
2
2πλk rk,n
2
= exp −
F̄h (yhk,n ) y α −1 dy ,
(10)
αBk
0
2
where Bk =
λk Pk α
P
j∈K\k
2
λj (Pj ) α
. Now substitute Eq. 9, 10 in
Eq. 8 , which gives
2
P(k,n)|hk,n = Erk,n g1 (hk,n )n−1 exp −λk πrk,n
g2 (hk,n )
Z ∞
=
frk,n (r)g1 (hk,n )n−1 exp −λk πr2 g2 (hk,n ) dr
0
(11)
R1
2
α −1
dy, g2 (hk,n ) =
where g1 (hk,n ) = α2 0 Fh (yhk,n ) y
R
R∞
2
2
2 ∞
2
α −1 dy
F̄
(yh
)
y
+
F̄
(yhk,n ) y α −1 dy.
h
k,n
h
α 1
αBk 0
2
(λk π)k r(2k−1) exp(−λk πr2 ) [12]
Substituting frk,n (r) = Γ(k)
in Eq. 11 with some algebraic manipulations leads to the
final result in Lemma 1.
B. Proof of Theorem 1
Using Eq. 3 and the result in Lemma 1, the CDF of the
effective fading gain h∗ can be derived as follows:
∞
XZ y
X
P(h∗ ≤ y) =
fh (hk,n )
P(k,n)|hk,n dhk,n (12)
k∈K
(f )
=
0
XZ
k∈K
0
n=1
y
fh (hk,n )
1
dhk,n ,
1 + g2 (hk,n ) − g1 (hk,n )
where step (f ) comes from using the summation of
a geometric series. It can easily be shown with simple
manipulations that 1 + g2 (hk,n ) − g1 (hk,n ) =
algebraic
2 R
2
Bk
2 −α ∞
1
α −1 dz, where B̃
k = Bk +1 . In this
α hk,n 0 F̄h (z)z
B̃
k
expression, we note that B̃k is the only term that is function
of k. Moreover, note that B̃k is not function of hk,n . Hence,
the summation in Eq. 12 can be handled as follows:
2
Z y
α
X
αh
k,n
dh
B̃k
P(h∗ ≤ y) =
fh (hk,n ) R ∞
k,n
2
2 0 F̄ (z)z α −1 dz
0
k∈K
2
Z y
α
αh
(g)
k,n
fh (hk,n ) R ∞
=
dhk,n ,
(13)
2
α −1 dz
2
F̄
(z)z
0
0
P
where step (g) is due to k∈K B̃k = 1. Taking the derivative
of this CDF w.r.t y, the final result in Theorem 1 follows.
R EFERENCES
[1] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling
and analysis of K-tier downlink heterogeneous cellular networks,” IEEE
Journal on Sel. Areas in Commun., vol. 30, no. 3, pp. 550–560, Apr.
2012.
[2] ITU, “Framework and overall objectives of the future development of
IMT for 2020 and beyond”, Rec. ITU-R M.2083-0, Sep. 2015.
[3] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras, M. S.
Berger, and L. Dittmann, “Cloud RAN for mobile networks–A technology overview,” IEEE Commun. Surveys and Tutorials, vol. 17, no. 1,
pp. 405–426, 2015.
[4] H. C. Yang and M. S. Alouini, Order statistics in wireless communications: diversity, adaptation, and scheduling in MIMO and OFDM
systems. Cambridge University Press, 2011.
[5] T. Eng, N. Kong, and L. B. Milstein, “Comparison of diversity combining techniques for Rayleigh-fading channels,” IEEE Trans. on Commun.,
vol. 44, no. 9, pp. 1117–1129, Sep 1996.
[6] M. S. Alouini and M. K. Simon, “An MGF-based performance analysis
of generalized selection combining over Rayleigh fading channels,”
IEEE Trans. on Commun., vol. 48, no. 3, pp. 401–415, Mar 2000.
[7] H. C. Yang and M. S. Alouini, “Generalized switch-and-examine combining (GSEC): a low-complexity combining scheme for diversity-rich
environments,” IEEE Trans. on Commun., vol. 52, no. 10, pp. 1711–
1721, Oct 2004.
[8] G. Song and Y. Li, “Asymptotic throughput analysis for channel-aware
scheduling,” IEEE Trans. on Commun., vol. 54, no. 8, pp. 1514–1514,
Aug 2006.
[9] Y. Ma, J. Jin, and D. Zhang, “Throughput and channel access statistics
of generalized selection multiuser scheduling,” IEEE Trans. on Wireless
Commun., vol. 7, no. 8, pp. 2975–2987, August 2008.
[10] L. Yang and M. S. Alouini, “Performance analysis of multiuser selection
diversity,” IEEE Trans. on Veh. Technology, vol. 55, no. 6, pp. 1848–
1861, Nov 2006.
[11] J. G. Andrews, A. K. Gupta, and H. S. Dhillon, “A primer on
cellular network analysis using stochastic geometry,” available online:
arxiv.org/abs/1604.03183.
[12] M. Haenggi, Stochastic geometry for wireless networks. Cambridge
University Press, 2012.
| 7cs.IT
|
1
The Age of Information: Real-Time Status
Updating by Multiple Sources
arXiv:1608.08622v2 [cs.IT] 12 Dec 2017
Roy D. Yates and Sanjit K. Kaul
Abstract
We examine multiple independent sources providing status updates to a monitor through simple
queues. We formulate an Age of Information (AoI) timeliness metric and derive a general result for the
AoI that is applicable to a wide variety of multiple source service systems. For first-come first-served
and two types of last-come first-served systems with Poisson arrivals and exponential service times, we
find the region of feasible average status ages for multiple updating sources. We then use these results
to characterize how a service facility can be shared among multiple updating sources. A new simplified
technique for evaluating the AoI in finite-state continuous-time queueing systems is also derived. Based
on stochastic hybrid systems, this method makes AoI evaluation to be comparable in complexity to
finding the stationary distribution of a finite-state Markov chain.
Index Terms
Age of information, status updates, queueing systems, random processes, communication
networks.
I. I NTRODUCTION
Increasingly ubiquitous connectivity to communication networks and availability of portable
devices have engendered a host of applications in which sources – people and environmental
sensors – send updates of their status to interested recipients. These include news and weather
reports and updates by individuals on Twitter about what is keeping them busy, updates by
environmental sensors [1], and vehicular status (position, velocity, acceleration) updates that can
Roy Yates is with WINLAB and the ECE Department, Rutgers University, NJ, USA, e-mail: [email protected].
Sanjit Kaul is with Wireless Systems Lab, IIIT-Delhi, India, e-mail: [email protected].
This work was presented in part at the 2012 Conference on Information Sciences and Systems and the 2012 IEEE International
Symposium on Information Theory. This work was supported by NSF award CCF-1422988.
2
assist drivers of nearby vehicles in an intelligent transportation system [2]. These applications
need status updates at one or more monitors to be as timely as possible; however, this is typically
constrained by limited network resources.
For example, consider various sensors (location, acceleration, tire pressure, etc.) in a vehicle
generating status updates which the in-vehicle radio delivers to other vehicles in vicinity or other
networked monitoring systems. The update packets are queued while they wait to be serviced
by the car radio. The packet currently being serviced by the radio waits for medium access and
transmission before it is received by other cars. Note that each sensor in the car may be a source
or the car may aggregate a collection of sensor measurements into a status update message that
is transmitted as a single packet. The packet service time will depend on the wireless channel
and may or may not incorporate retransmissions due to channel errors and backoff due to the
activity of other wireless transmitters. While system models that incorporate these effects can
be arbitrarily complex, we observe that optimal updating policies are not well understood even
in the simple setting of M/M/1 queues.
Maintaining the timeliness of data and state information in a network is a problem that has
appeared in many forms, including, for example, data freshness in warehouses [3] and web
caches [4], periodic updating of real time databases [5], and route caches in ad hoc networks
[6]. However, no consistent analytic methodology has emerged. This paper focuses on an age
of information (AoI) timeliness metric as a basis for the evaluation and design of status update
systems.
When a monitor’s most recently received update at time t is timestamped u(t), the status
update age or simply the age, is the random process ∆(t) = t − u(t) and the AoI is the
average ∆(t). The monitor’s requirement of timely updating corresponds to small AoI. AoI is
an application-independent metric that permits evaluation of the network performance, separate
from application-specific metrics that may be too complex to employ in the design of the network.
However, AoI can also be useful in specific applications by designing the communication network
to meet statistical requirements, such as expected value and variance, of the age process. For
example, if a status updating system is reporting sample values of a Wiener process X(t) with
variance αt [7], then the monitor’s MMSE estimate of X(t) given the status age ∆(t) is X̂(t) =
X(t − ∆(t)). The variance of this estimate is α∆(t).
Traditionally, network performance has been characterized by tradeoffs in rate, delay, throughput and loss. The data rate can be increased, but this induces additional delay in lossless systems
3
or increased packet dropping in lossy systems. Furthermore, comparisons between lossless and
lossy networks are generally problematic. By contrast, we will see that AoI is fundamentally
different; the age metric enables direct comparison of lossless and lossy systems. Moreover, the
goal of timely updating is neither the same as maximizing the throughput or utilization of the
communication system, nor of ensuring that generated status updates are received with minimum
delay. Utilization is maximized when sensors send updates as fast as possible. However, this can
lead to the monitor receiving delayed updates because the status messages become backlogged
in the communication system. Instead, we will see that sources can minimize their AoI by
optimizing their updating rates in response to the available system resources.
We further observe that it may also be desirable to redesign systems to facilitate timely
updating. A basic property of the first-come first-served (FCFS) queue model is that new update
messages can be queued behind outdated messages that were generated earlier. This can be
viewed as an undesirable consequence of protocol layering or of the hardware design. However,
among all status update packets in the wireless interface, the transmission of the youngest packet
will minimize the status age at the monitor. Moreover, under the assumption that a status update
carries the Markov state of the source, the transmission of the youngest status update obviates the
need for transmission of the older outdated packets in the queue. Thus it is desirable to implement
a lossy last-come first-served (LCFS) queueing discipline in which a new status update packet
will preempt any previously queued update packets and this preempted packet will be discarded.
A. Prior Work and Related Applications
This paper expands on our analyses of status age in single-source single-server queues [8], the
M/M/1 LCFS queue with preemption in service [9], and the M/M/1 FCFS system with multiple
sources [10]. Other contributions to AoI analysis have also appeared recently. To evaluate AoI
for a single source sending updates through a network cloud [11] or through an M/M/2 server
[12], [13], out-of-order packet delivery was the key analytical challenge. A related (and generally
more tractable) metric, peak age of information (PAoI), was introduced in [14]. Properties of
PAoI have also been studied in [15] for an FCFS M/G/1 multiclass queue. In [14], [16], the
authors analyzed AoI and PAoI for M/M/1/1 and M/M/1/2 queues that discard arriving updates
if the system is full and also for a third queue, dubbed M/M/1/2*, in which an arriving update
would preempt a waiting update. In this work, the M/M/1/2* queue is called the M/M/1 LCFS-
4
W (Last Come First Served with preemption only in Waiting) queue and here we extend AoI
results to a LCFS-W system with multiple sources.
Most recently, optimality properties of a Last Generated First Served (LGFS) service discipline
when updates arrive out of order are identified in [17], packet deadlines are found to improve
AoI in [18], AoI in the presence of errors is evaluated in [19], and LCFS with non-memoryless
gamma-distributed service times is considered in [20]. There have also been recent studies
of energy-constrained updating [21]–[23] in which updates are submitted to the server with
knowledge of the server state.
In addition to these queue-theoretic AoI analyses, the theme of ensuring “freshness” has
also appeared in various application areas, including that of networks, real time databases and
warehousing.
In [24], we look at minimizing the age of status updates sent by vehicles over a carriersense multiple access (CSMA) network. A local minimum for the age can be approached using
gradient descent; however, it is not known if this is a global minimum and is only seen to exist
in simulations. In [25], we show that allowing nodes to piggyback other nodes’ status updates
can lead to a smaller age.
For safety-related intelligent transport system applications, an Awareness Quality metric [26]
captures how fine and up-to-date the application information is. The authors observe that default
metrics like throughput and delay are unable to capture awareness and propose Update Delay,
which is the elapsed time between application updates. In [27], the authors propose to use an
oldest packet drop mechanism instead of a tail drop policy to reduce the delay of the received
information, via beacons, in vehicular networks.
In [3], the authors want to maximize the freshness of data in warehouses to meet user demands.
They estimate the queue length and delay at the warehouse staging area where updates wait before
they are committed to the warehouse database. Experiments lead them to conclude that small
queues are desirable.
Web caching reduces the latency in returning a web page to a client. However, unless refreshed
often enough, a cache will return stale web pages. The refresh rate is limited by the finite time it
takes for a cache to be updated after the page has been updated at the server. In [4] the authors
propose an architecture that limits the “degree of staleness” of a cache. Our work, for fairly
simple descriptions of the time it takes to update a cache, answers how often the cache must be
refreshed such that its age is minimized.
5
Source
1
λ1
..
.
Source
N
µ
Monitor
λN
Fig. 1. Independent sources send status updates through a shared queue to a monitor.
In [5] the authors look at periodic transactions updating real time databases. Each transaction
updates the database with data that is associated with a deadline relative to when it is generated. In
their work, there is no assumed limit on available processing power (service rate). The objective
is to find the combination of update period and deadline such that all transactions complete
before their deadlines, thus ensuring the freshness of data while minimizing the CPU utilization.
Ad hoc networking protocols typically use cached routes to forward packets to their destinations. In [6] the authors propose a mechanism that avoids propagation of stale route information.
To avoid the overhead associated with periodic broadcasts of new route information, their method
uses an epoch numbering system that helps network nodes to reject older information. In [28]
the authors consider the issue of frequency of hello messages in ad-hoc networks. The frequency
must not be so large as to congest the network but also not too small that the nodes have stale
information.
Finally, data dissemination in sensor networks has been looked at under varied constraints.
For example, in works like [29] and [30] the authors consider energy efficient dissemination of
state in sensor networks. More frequent updates lead to greater energy consumption and smaller
sensor lifetime. Our work suggests strategies that a sensor, when awake, can use to minimize
the average age of its status updates.
B. Paper Overview
This work is based on the system depicted in Figure 1 in which a server delivers the updates
of N sources to a monitor. Each source i offers update packets as a rate λi Poisson process. The
service rate is µ for updates from any source. This is sufficient to model systems in which the
updating sources submit packets with identically distributed lengths but heterogeneous timeliness
requirements. Although the M/M/1 queue models that we examine are often too simple to
6
describe practical networks, status age is a new metric that is not well understood. We start
here with these simple models to develop an understanding of the age of information in shared
queues, in order to go on to characterize status age in more complex practical systems.
In this system, source i offers updating load ρi = λi /µ and the total offered load is
ρ=
N
X
ρi .
(1)
i=1
The updates of source i compete for the server against the aggregate other-source updating load
ρ−i = ρ − ρi =
X
ρj .
(2)
j6=i
In Section II, we derive Theorem 3, a general result that describes the AoI ∆i for each source
i in terms of the stationary properties of the interarrival times and system times of delivered
source i updates. We then apply Theorem 3 to M/M/1 systems in which update packets arrive as
Poisson processes and have memoryless service times. In particular, Section II-A uses Theorem 3
to derive the following result.
Theorem 1: N sources with offered loads ρ1 , . . . , ρN and total load ρ =
P
j
ρj at a rate µ
M/M/1 FCFS queue have average ages ∆1 , . . . , ∆N such that
ρ2i (1 − ρρ−i )
1
1
1
∆i =
+
+
.
µ (1 − ρ)(1 − ρ−i )3 1 − ρ−i ρi
We continue the study of Poisson updaters with AoI results for a pair of lossy N -source
M/M/1 LCFS systems. First, under LCFS with preemption-in-service (denoted LCFS-S), a new
update packet preempts any update packet currently in service. Second, under the LCFS with
preemption only in waiting (LCFS-W) queue discipline, a new packet replaces any older packet
waiting in the queue; however, the new packet has to wait for any update packet currently in
service to finish. In this work, preemption is assumed to be source agnostic; we will allow a
source’s packet to be preempted by that of another source. Prioritized preemption policies are
of considerable interest but beyond the scope of this work. For M/M/1 systems, the main result
is summarized here:
Theorem 2: N sources with offered loads ρ1 , . . . , ρN at a rate µ M/M/1 LCFS queue with
P
total load ρ = N
i=1 ρi have average ages ∆1 , . . . ∆N such that
7
(a) with preemption allowed in service (LCFS-S),
∆i =
1
1
(1 + ρ) ,
µ
ρi
(b) and with preemption allowed only in waiting (LCFS-W),
1
ρ2
1
∆i =
αW (ρ) + 1 +
µ
1 + ρ ρi
where
(1 + ρ + ρ2 )2 + 2ρ3
.
αW (ρ) =
(1 + ρ + ρ2 )(1 + ρ)2
We note that while αW (ρ) is a ratio of fourth order polynomials, direct calculation will verify
that
0.837 < αW (ρ) < 1.09,
ρ ≥ 0.
(3)
The proof of Theorem 2 appears in parts in various sections of this paper. In Section II-B, we
use Theorem 3 to derive Theorem 2(a) for AoI in the LCFS-S queue. The method is similar to
that used in [9], but with some algebraic simplifications that went previously unrecognized. As it
is based on Theorem 3, this analysis is conceptually similar to the FCFS analysis in Section II-A.
We note that for a single source with ρ1 = ρ, Theorem 2(b) can be shown to reduce to the AoI
of the M/M/1/2* queue, as given in [16, Equation (65)]. We also note Theorem 2(b) corrects
an error in [9, Equation (23)]. In the context of a single-source system, this error was identified
and explained in [16, Appendix]. That explanation serves to highlight how easily mistakes can
be made in using the approach of Theorem 3 for AoI analysis, even in simple memorylessservice systems. At the conclusion of [16, Appendix], the authors argue “In the LCFS system
with preemption we expect that, for very large arrival rates, the age would increase without
bound, as no packet finishes service.” We note that this speculation is contrary to the result of
Theorem 2(a), which in the special case of a single-source with ρ1 = ρ = λ/µ, shows that the
average age approaches 1/µ as λ → ∞.1
Nevertheless, the error in [9] and the skepticism expressed in [16] reflect on how difficult it can
1
For the single-source LCFS-S system, this asymptotic result is a consequence of memoryless service. With fixed service
rate µ and arrival rate λ → ∞, the server is always occupied and (because the service is memoryless) the queue departure
rate approaches µ. That is, the queue inter-departures approach a Poisson process of rate µ. While the fraction µ/(λ + µ) of
those updates that complete service goes to zero, those that do complete service have system time T that approaches zero. In
this limiting case, the interarrival time Y of a delivered update becomes an exponential (µ) random variable. In the context of
Theorem 3 and the sawtooth age process in Figure 2, E[T Y ] → 0, E[Y 2 ] → 2/µ and E[Y ] → 1/µ.
8
be to use Theorem 3 to prove and verify AoI results, even for relatively simple service facilities.
Thus we introduce in Section III an analysis technique, namely stochastic hybrid systems (SHS)
[31], that has not been previously applied to status updating systems. A stochastic hybrid system
has a state with discrete components described by a Markov chain and continuous components
that are subject to reset mappings in discrete state transitions. In AoI analysis, the queue state
describing the number and source type of each update in the system is discrete while the age
process at the monitor and the age of each update in the system varies continuously but is subject
to reset mappings as updates enter or complete service, or get preempted.
We will see that AoI tracking can be implemented as a simplified SHS with non-negative
linear reset maps in which the continuous state is a piecewise linear process [32], [33], a special
case of piecewise deterministic processes [34], [35]. In this case, the SHS approach leads to
a system of ordinary first order differential equations describing the temporal evolution of the
expected value of the age process. The SHS approach may not appear to be simple at first;
however, in Section III-C it yields Theorem 4, a simple, systematic (and largely mechanical)
procedure for the calculation of AoI in finite-state queues with memoryless service.
In Section IV, we demonstrate the power of Theorem 4 by using it to prove Theorem 2.
In particular, Section IV exercises Theorem 4 in a sequence of three SHS-based derivations
of Theorem 2(a). The first derivation is a straightforward application of Theorem 4 with three
discrete states to track whether the system is idle or busy, and whether an update in service
is from the source of interest. The second derivation demonstrates how the discrete state space
can be reduced to two states (idle or busy) by careful embedding of some elements of the
discrete state in the continuous state. The third derivation shows how fake updates can be used
to reduce the LCFS-S system to a single discrete state. The embedding approach is then used
in Section IV-D to provide an SHS derivation of Theorem 2(b) for AoI in the LCFS-W system.
We note that SHS analysis of the LCFS-S and LCFS-W systems is far simpler than analyses
based on Theorem 3. In Section V, we rewrite the equations of Theorem 4 in a non-negative
matrix form in order to prove Theorem 4.
Finally, in Section VI we return to examine the performance of the updating system shown in
Figure 1. We use Theorems 1 and 2 to examine achievable AoI regions for two-source FCFS and
LCFS systems. In addition, resource sharing issues for N sources are also explored. Our results
show that there are nontrivial gains in trunking efficiency when N sources share the system
capacity with coordinated load balancing of the sources. In particular, high offered load at an
9
FCFS system induces high AoI through queueing delays. A lossy LCFS discipline can mitigate
this problem but its packet discarding policy may encourage sources to operate at excessively
high offered loads. A short conclusion follows in Section VII.
C. Notation
For integers m ≤ n, m : n = {m, m + 1, . . . , n}; otherwise m : n is an empty set. The vectors
0n and 1n denote the row vectors [0 0 · · ·
0] and [1 1 · · · 1] in Rn . A vector x ∈ Rn is a
1 × n row vector with elements [x0 x1 . . .
xn−1 ]. Similarly, a matrix B ∈ Rn×n has elements
[B]i,j for i, j ∈ 0 : (n − 1). For a vector x and a matrix B, [x]j and [B]j denote the jth element
and jth column respectively for j = 0, . . . , n − 1. For a vector process x(t), we use ẋ and ẋ(t)
to denote the derivative dx(t)/dt.
II. T IME -AVERAGE AGE A NALYSIS
Figure 2 shows a sample variation of age ∆i (t), for source i as a function of time t, at the
monitor. Without loss of generality, assume that we begin observing at t = 0 when the queue is
empty and the age is ∆i (0). The first status update of source i is timestamped t1 and is followed
by updates timestamped t2 , t3 , . . . , tn . The status age of source i at the monitor increases linearly
in time in the absence of any updates and is reset to a smaller value when an update is received.
Update j of source i, generated at time tj , finishes service and is received by the monitor at
time t0j . At t0j , the age ∆i (t0j ) at the monitor is reset to the age Tj = t0j − tj of the received status
update. The age Tj is also the system time of update packet j. Thus the age function ∆i (t)
exhibits the sawtooth pattern shown in Figure 2. The time average age of the status updates is
the area under the age graph in Figure 2 normalized by the time interval of observation.
Over an interval (0, T ), the average age is
1
h∆i iT =
T
Z
T
∆i (t)dt.
(4)
0
For simplicity of exposition, the length of the observation interval is chosen to be T = t0n , as
depicted in Figure 2. We decompose the area defined by the integral (4) into the sum of the
polygon area Q̃1 , the trapezoidal areas Qj for j ≥ 2 (Q2 and Qn are highlighted in the figure),
and the triangular area of width Tn over the time interval (tn , t0n ). From Figure 2, we see that
Qj can be calculated as the difference between the area of the isosceles triangle whose base
10
∆i (t)
Q̃1
Q2
Qn
∆i (0)
N
t1
N
t2
Y2
H
t01
N N H
t3 t4 t02
Y3
H
t03
H
t04
Y4
N H
tn−1
Tn
H
t
t0n
N
tn
Yn
T2
Tn
T3
Fig. 2. Example change in status update age at a monitor for a system with a FCFS queue. Updates from source
1 arrive at times marked N and are received at the monitor at times marked H.
connects the points tj−1 and t0j and the area of the isosceles triangle with base connecting the
points tj and t0j . Defining
Yj = tj − tj−1
(5)
to be the interarrival time of update j, it follows that
1
1
Qj = (Tj + Yj )2 − Tj2 = Yj Tj + Yj2 /2.
2
2
(6)
With Ni (T ) = max{n|tn ≤ T } denoting the number of source i updates by time T , this
decomposition, along with some rearrangement, yields the time-average age
PNi (T )
Q̃ (Ni (T ) − 1) j=2 Qj
h∆i iT = +
T
T
Ni (T ) − 1
(7)
where Q̃ = Q̃1 + Tn2 /2. We observe that the age contribution Q̃ represents a boundary effect
that is finite with probability 1, so the first term in (7) will vanish as T grows.
Definition 1: A status updating system for a source i is stationary and ergodic if (Yj , Tj ) is a
stationary sequence with marginal distribution identical to (Y, T ), and as T → ∞,
PNi (T )
Ni (T )
1
j=2 Qj
→
, and
→ E[Q]
T
E[Y ]
Ni (T ) − 1
with probability 1.
11
For such systems, the AoI of source i is ∆i = limT →∞ h∆i iT and (7) implies the next claim.
Theorem 3: For a stationary ergodic status updating system in which Y is the interarrival time
between delivered source i updates and T is the system time of such a delivered packet, the AoI
for source i is
∆i =
E[Q]
E[Y T ] + E[Y 2 ]/2
=
.
E[Y ]
E[Y ]
We note that Theorem 3 is more akin to a bookkeeping identity such as Little’s Law in that
sufficient conditions for the ergodicity of the age process are not explicitly provided but can
be verified for most reasonably-designed service systems. As a consequence, Theorem 3 can be
applied to a broad class of service systems, including both lossless FCFS systems as well as
lossy LCFS systems in which updates are preempted and discarded. Furthermore, it makes no
specific assumptions regarding other traffic that might share the queue with the update packets
of source i.
With respect to Theorem 3, we emphasize that Y is the interarrival time between delivered
updates of source i, and T is the system time of a delivered update. These somewhat cumbersome
definitions are a consequence of the generality of the approach. For example, Section II-A
employs Theorem 3 to evaluate a work-conserving M/M/1 FCFS system in which the Yj are
independent identically distributed (iid) exponential interarrival times and the Tj are the packet
system times. On the other hand, Section II-B uses Theorem 3 to analyze the LCFS-S system
that supports preemption of the packet in service. In this system, packet j refers to the jth packet
that completes service and is delivered to the monitor. There may be an arbitrarily large number
of update packets that arrive between packets j − 1 and j that are preempted and discarded
before completing service.
A. M/M/1 First-Come First-Served
In prior work [8], we analyzed M/M/1 FCFS queues serving the status updates of a single
source. In that work, it was shown that the average status age for an M/M/1 queue with arrival
rate λ, service rate µ and offered load ρ = λ/µ is given by
1
1 ρ2
+1+ .
∆=
µ 1−ρ
ρ
(8)
The average age ∆ in (8) is minimized at ρ∗ ≈ 0.53. In this section, we generalize this result
to an N source system.
12
Following Theorem 3, the system time of a source i update packet is T = W + S, where W
and S are the respective waiting and service times. Since S is independent of Y , it follows that
E[Y T ] = E[Y W ] + E[Y ] E[S]. We note that E[S] = 1/µ and that the rate λi Poisson arrival
process implies E[Y ] = 1/λi . and E[Y 2 ] = 2/λ2i . It follows from Theorem 3 that
∆i = λi E[Y W ] +
1
1
+ .
µ λi
(9)
The expectation E[Y W ] is nontrivial because Y and W are negatively correlated; a large
interarrival time Y can allow the queue to empty, yielding a small waiting time W . Evaluation
of E[Y W ] is provided in Appendix A in the proof of the following lemma.
Lemma 1:
1
ρi (1 − ρρ−i )
ρ−i
E[Y W ] = 2
+
.
µ (1 − ρ)(1 − ρ−i )3 ρi (1 − ρ−i )
Applying Lemma 1 to (9) yields Theorem 1. We note that Theorem 1 reduces to the single
source result (8) when ρi = ρ and ρ−i = 0.
B. LCFS With Preemption In Service: Analysis
In this system, a packet arrival preempts the packet currently in service, if any. The number
of packets in such a system is at most 1. To analyze this system, we start with Theorem 3. As
shown in Figure 2, update packets generated by source i at time instants tj are those updates
that complete service and Yj = tj − tj−1 is the time between such arrivals. The service time (and
also system time) of this jth packet is Tj .
In order to calculate the average age ∆i , let Dj (see Figure 3) be the time interval between
the departures j − 1 and j. This interval starts with an idle period and may see zero or more
arrivals of other sources, some of which may complete service, while others are preempted. Any
arrivals of the given source during Dj , other than arrival j, are preempted. Thus the interval
Dj consists of one or more blocks of the server being idle followed by it being busy. Note that
if the system consists of just one source, then Dj consists of just one block, which starts with
the idle period that follows the departure of j − 1. This idle period is followed by the server
busy period that ends in departure j. Figure 3 shows Dj , which contains a random L number
of blocks. The figure shows blocks 1 and L. A block k, say of length Bk , consists of an idle
13
∆1 (t)
Dj
NH N
tj−2 tj−1
···
X10 S1
H
0
XL0
tj−1
Yj−1
SL
N H0
tj tj
t
Yj
Tj−1
Tj
Fig. 3. Example change in update age of source i under LCFS with preemption in service. On the time axis, updates
from source i arrive at times marked N and are received at the monitor at times marked H.
period of length Xk0 followed by a busy period of length Sk . We have
Dj =
L
X
k=1
Bk =
L
X
(Xk0 + Sk ).
(10)
k=1
Note that packet j arrives during SL and then spends time Tj in service.
We will now calculate the terms E[Y ], E[Y 2 ] and E[Y T ] in Theorem 3 in terms of Dj and
Tj . Consider the interval Yj , for any j. We observe from Figure 3 that
Yj + Tj = Tj−1 + Dj .
(11)
Because T =st Tj−1 =st Tj , Y =st Yj , and D =st Dj ,
E[Y ] = E[Yj ] = E[Dj ] = E[D].
(12)
Note that Yj and Tj are independent. Thus (12) implies
E[Yj Tj ] = E[Yj ] E[Tj ] = E[Y ] E[T ] = E[D] E[T ].
(13)
14
Furthermore, since Dj and Tj−1 are also independent, it also follows from (11) that
Var[Yj ] + Var[Tj ] = Var[Tj−1 ] + Var[Dj ].
(14)
It then follows from (12) that E[Y 2 ] = E[D2 ]. This fact, combined with (12) and (13), simplify
Theorem 3 to
∆i = E[T ] +
E[D2 ]
.
2 E[D]
(15)
The remainder of the proof of Theorem 2(a), specifically the calculation of the moments in (15),
appears in Appendix B.
III. S TOCHASTIC H YBRID S YSTEMS FOR AO I A NALYSIS
We start in Section III-A with an introduction to the key elements in the general SHS method.
In Section III-B, we consider a special case of SHS in which the continuous state x(t) is a
piecewise linear process that is subject to linear reset maps during a discrete state transition.
For the piecewise linear SHS, we derive a set of first order differential equations for the first
order moments of the continuous state. Section III-C employs the piecewise linear SHS to derive
the AoI of a general finite state queueing system described by a continuous-time Markov chain.
The resulting methodology makes AoI computation practically as simple as the calculation of
stationary probabilities of the Markov chain of the queue.
A. A Brief Introduction to SHS
There are many SHS variations [36], but in this work we follow the model and notation in
[31]. In an SHS, the state is partitioned into a discrete component q(t) ∈ Q = {0, 1, . . . , m}
that evolves as a point process and a continuous component x(t) = [x0 (t) · · ·
xn (t)] ∈ Rn+1 .
Given the discrete set Q and the k-vector z(t) of independent Brownian motion processes, an
SHS is defined by a stochastic differential equation
ẋ = f (q, x, t) + g(q, x, t)ż
(16)
for mappings f : Q × Rn+1 × [0, ∞) → Rn+1 and g : Q × Rn+1 × [0, ∞) → R(n+1)×k , and a set
of transitions L = {1, . . . , `0 − 1}, such that each l ∈ L defines a discrete transition/reset map
(q 0 , x0 ) = φl (q, x, t),
φl : Q × Rn+1 × [0, ∞) → Q × Rn+1 ,
(17a)
15
with transition intensity
λ(l) (q, x, t),
λ(l) : Q × Rn+1 × [0, ∞) → [0, ∞).
(17b)
When the system is in discrete state q, x(t) evolves according to (16); but in a discrete transition
from q to q 0 , the continuous state can have a discontinuous jump from x to x0 , as described by
(17a). The resulting x(t) process has piecewise continuous sample paths. Associated with each
transition l is a counting process Nl (t) that counts the number of occurrences of transition l in
the interval [0, t]. The probability that Nl jumps in the interval (t, t + dt] is λ(l) (q(t), x(t), t) dt.
Because of the generality and power of the SHS model, characterization of the q(t) and
x(t) processes can be complicated and often intractable. The approach in [31] is to define test
functions ψ(q, x, t) whose expected values E[ψ(q(t), x(t), t)] can be evaluated as functions of
time. We refer the reader to [31] and the survey [36] for additional background.
B. Piecewise linear SHS with linear reset maps
In the setting of status updates passing through queues with memoryless service processes,
we restrict our attention to systems in which q(t) is a continuous-time finite-state Markov chain
that describes the occupancy of a service facility and x(t) ∈ Rn+1 describes the continuous-time
evolution of a collection of age-related processes.
In particular, we consider a restricted class of SHS in which the components of x(t) are
deterministic constant-slope ramp processes that can have discontinuous jumps during discrete
state transitions. We will see that this will be sufficient to capture the sawtooth age processes.
In terms of the general SHS model given by (16) and (17), we have
f (q, x, t) = bq ,
(18a)
g(q, x, t) = 0,
(18b)
λ(l) (q, x, t) = λ(l) δql ,q ,
φl (q, x, t) = (ql0 , xAl ).
(18c)
(18d)
In the graphical representation of the Markov chain q(t), each state q ∈ Q is a node and each
transition l is a directed edge (ql , ql0 ) with transition rate λ(l) while q(t) = ql . In (18c), the
Kronecker delta function δql ,q ensures that transition l occurs only in state ql . For each transition
l, the transition reset map will be a linear mapping of the continuous state x of the form x0 = xAl .
16
That is, transition l causes the system to jump to discrete state ql0 and resets the continuous state
from x to x0 = xAl . In addition, we note that (18a) and (18b) imply that the continuous state
evolution (16) in each discrete state q(t) = q is
ẋ(t) = bq .
(19)
Thus, the evolution of x(t) in each state is specified by B = {bq : q ∈ Q}. Furthermore, the
transition links l are described by the tuples al = (ql , ql0 , λ(l) , Al ) and the set of transitions is
A = {al : l ∈ L}. Thus we refer to a piecewise linear SHS with linear reset maps by the tuple
(Q, B, A).
The transition rates λ(l) correspond to the transition rates associated with the continuoustime Markov chain for the discrete state q(t); but there are some differences. Unlike an ordinary
continuous-time Markov chain, the SHS may include self-transitions in which the discrete state
is unchanged because a reset occurs in the continuous state. Furthermore, for a given pair of
states i, j ∈ Q, there may be multiple transitions l and l0 in which the discrete state jumps from
i to j but the transition maps Al and Al0 are different.2
For each q̂ ∈ Q, it will be sufficient for average age analysis to employ test functions of the
form ψ(q, x) = ψq̂ (q, x) and ψ(q, x) = ψq̂j (q, x) such that
ψq̂ (q, x) = δq̂,q
(20a)
and
ψq̂j (q, x) = xj δq̂,q ,
j ∈ 0 : n.
(20b)
Based on these test functions, we define for all q̂ ∈ Q,
πq̂ (t) = E[ψq̂ (q(t), x(t))] = E[δq̂,q(t) ],
vq̂j (t) = E[ψq̂j (q(t), x(t))] = E[xj (t)δq̂,q(t) ],
(21a)
j ∈ 0 : n,
(21b)
and the vector functions
vq̂ (t) = [vq̂0 (t), . . . , vq̂n (t)] = E[x(t)δq̂,q(t) ].
2
(21c)
For example, consider a queueing system in which an update in service can either complete service or be discarded in
the middle of service. Under either transition, the next discrete state reflects the departure or discard of the update in service.
However, a service completion yields a reduction in age while discarding an update in service results in no reduction in age.
17
We note that πq̂ (t) denotes the discrete Markov state probabilities, i.e.,
πq̂ (t) = E[δq̂,q(t) ] = P[q(t) = q̂].
(22)
Similarly, vq̂ (t) measures correlation between the age process x(t) and the discrete state q(t).
Associated with an SHS is a mapping ψ → Lψ known as the extended generator. From [31,
Theorem 1], it follows from the conditions (18) and the time invariance of ψ(q, x) in (20) that
the extended generator of a piecewise linear SHS is given by
(Lψ)(q, x) =
X
∂ψ(q, x)
· bq +
(ψ(φl (q, x)) − ψ(q, x))λ(l) (q),
∂x
l∈L
(23)
where ∂ψ(q, x)/∂x denotes the gradient. Each test function ψ(q(t), x(t)) must satisfy Dynkin’s
formula
d E[ψ(q(t), x(t))]
= E[(Lψ)(q(t), x(t))].
dt
(24)
L0q̄ = {l ∈ L : ql0 = q̄},
(25a)
Lq̄ = {l ∈ L : ql = q̄}
(25b)
Defining
as the respective sets of incoming and outgoing transitions for each state q̄, we can now prove
that πq̄ (t) and vq̄ (t) obey the system of first order ordinary differential equations given in the
following lemma.
Lemma 2: For a piecewise linear SHS with linear reset maps (Q, B, A),
π̇q̄ (t) =
X
λ(l) πql (t) − πq̄ (t)
l∈L0q̄
v̇q̄ (t) = bq̄ πq̄ (t) +
X
λ(l) ,
q̄ ∈ Q,
(26a)
l∈Lq̄
X
λ(l) vql (t)Al − vq̄ (t)
l∈L0q̄
X
λ(l) ,
q̄ ∈ Q.
(26b)
l∈Lq̄
Proof of Lemma 2 appears in Appendix C. From a given initial condition at time t = 0, we can
use Lemma 2 to compute the temporal evolution of the discrete state probabilities πq̄ (t) and the
expected values vq̄ (t) = E[x(t)δq̄,q(t) ]. Moreover, since
x(t) =
X
q̄∈Q
x(t)δq̄,q(t) ,
(27)
18
Lemma 2 enables us to compute the expected value
E[x(t)] =
X
E[x(t)δq̄,q(t) ] =
q̄∈Q
X
vq̄ (t).
(28)
q̄∈Q
C. An SHS for AoI
We now employ a piecewise SHS with linear reset maps (Q, B, A) for age tracking in a
system described by a continuous-time Markov chain. Our approach is to label the source of
interest as source 1 and to employ the continuous state x(t) as a vector of age-related processes
that enables tracking of the age of source 1 updates at the monitor.
In the LCFS-S and LCFS-W systems with x(t) = [x0 (t) . . . xn (t)] that we examine in
Section IV, x0 (t) is the age at the monitor, n is the maximum number of updates in the system,
and updates in the system are indexed 1, 2, . . . , n such that if update i is from source 1, then
xi (t), 1 ≤ i ≤ n, specifies the age of the update. In such systems, we will calculate the AoI
∆ = limt→∞ E[x0 (t)] of the source 1 update process. The sequence of SHS analyses of the
LCFS-S system will reveal there is considerable flexibility in choosing the continuous state
x(t). Moreover, careful definition of x(t) can reduce the size of the discrete state space.
For example, in SHS analyses of the LCFS-S queue in Section IV, x(t) = [x0 (t), x1 (t)], where
x0 (t) tracks the age (of source 1 updates) and x1 (t) is a state variable that specifies the age of a
source 1 update currently in service, if any. When a source 1 update is delivered at time t, the
age x0 (t) will be reset to x1 (t). In the following, when we refer to age, we specifically mean
the age of source 1 updates at the monitor.
In using a piecewise linear SHS for AoI, the elements of bq will be binary. We will see
that the ones in bq correspond to certain relevant components of x(t) that grow at unit rate in
state q while the zeros mark components of x(t) that are irrelevant in state q to the age process
and need not be tracked. For tracking of the age process, the transition reset maps are binary:
Al ∈ {0, 1}(n+1)×(n+1) . The set of linear mappings {Al } will depend on the specific queue
discipline, and the indexing scheme for updates in the system.
Definition 2: An age-of-information SHS (Q, B, A) is an SHS in which the discrete state
q(t) ∈ Q is a continuous-time Markov chain with transitions l ∈ L from state ql to ql0 at rate
λ(l) and the continuous state evolves according to ẋ(t) = bq ∈ {0, 1}n+1 in each discrete state
q ∈ Q and is subject to the linear transition reset map x0 = xAl in transition l.
19
Note that column j, 0 ≤ j ≤ n, of Al determines how x0j is set when transition l takes place.
Typically, we will construct transition mappings Al that have no more than a single 1 in each
column. In particular if [Al ]i,j = 1, then transition l firing causes x0j = xi . This corresponds to
transition l relabeling update i as update j. This occurs, for example, in a FCFS queue when the
service completion of update 1 causes update i occupying queue position i to be relabeled as
update i − 1, because its queue position changes to i − 1. Another important case occurs when a
transition l inserts a fresh source 1 update in queue position j at time t. Immediately following
the transition, i.e. after that update is inserted, x0j = 0 because that update is fresh. In this case,
[Al ]j is an all-zero column.
On the other hand, if transition l corresponds to the service completion of a source 1 update
indexed j, then Al must encode the resulting age reduction. This would require [Al ]j,0 = 1
and, for k 6= j, [Al ]k,0 = 0, so that the mapping x0 = xAl yields x00 = xj . That is, the age is
reset to the age of this most recently delivered update and the corresponding reduction in age
is x0 − x00 = x0 − xj .
We note that the SHS method specifies the continuous state x(t) for all discrete states q ∈ Q.
However, not all components of x(t) are relevant in all states. Since the dimensionality of x(t)
is fixed to be n + 1, we choose n to be the maximum number of updates in the system over
all states q. However, when a state q has kq < n updates in the system, then xkq +1 (t), . . . , xn (t)
are irrelevant variables in state q as there are no corresponding updates in the system that
could complete service. We also note that not all updates in the system are relevant to the future
trajectory of the age process. Since we are tracking the age of source 1 updates, xj (t) is irrelevant
when there is an update from source s > 1 in queue position j.
We define Iq as the index set of irrelevant variables in discrete state q. That is, j ∈ Iq if
xj (t) is irrelevant in state q. We will see that the irrelevant variables in discrete state q have no
impact on the subsequent state in a transition out of state q. Hence, when the system enters state
q, we can arbitrarily set an irrelevant variable to any value. For algebraic clarity, we adopt the
convention that each irrelevant xj (t) is zero in state q. For each state q, we set
[Al ]j = 0>
n,
l ∈ L0q , j ∈ Iq ,
(29)
and
[bq ]j = 0,
j ∈ Iq .
(30)
20
When transition l occurs and state q is entered with j ∈ Iq , (29) implies xj (t) is reset to x0j = 0
and (30) ensures xj (t) remains zero while in state q. On the other hand, if in state q the update
j is of source 1, then xj (t) is relevant in state q. As this relevant update ages at unit rate in state
q, we set [bq ]j = 1 for j 6∈ Iq .
A foundational assumption for age analysis is that the Markov chain q(t) is ergodic; otherwise,
time-average age analysis makes little sense. Under this assumption, the state probability vector
π(t) = [π0 (t) · · ·
πm (t)] always converges to the unique stationary vector π̄ = [π̄0 · · ·
π̄m ]
satisfying
X
π̄q̄
λ(l) =
λ(l) π̄ql ,
(31a)
l∈L0q̄
l∈Lq̄
X
X
π̄q̄ = 1.
(31b)
q̄∈Q
Moreover, we see in Lemma 2 that convergence to π̄ is disconnected from the evolution of the
age process. This is as expected since the age process x(t) is a measurements process that does
not influence the evolution of the queue state.
When π(t) has converged to the stationary probability vector π̄, we see from Lemma 2 that
(26b) is reduced to a system of first order differential equations
v̇q̄ (t) = bq̄ π̄q̄ +
X
λ(l) vql (t)Al − vq̄ (t)
l∈L0q̄
in v(t) = [v0 (t) · · ·
X
λ(l) ,
q̄ ∈ Q,
(32)
l∈Lq̄
vm (t)]. While Lemma 2 holds for any set of reset maps {Al }, the
differential equation (32) may or may not be stable. Stability depends on the collection of reset
maps3 . When (32) is stable, each vq̄ (t) = E[x(t)δq̄,q(t) ] converges to a limit v̄q̄ as t → ∞. In
this case, it then follows from (27) that
E[x] = lim E[x(t)] = lim
t→∞
t→∞
X
q̄∈Q
E[x(t)δq̄,q(t) ] =
X
v̄q̄
(33)
q̄∈Q
and that the average age of the process of interest is then ∆ = E[x0 ] =
P
q̄∈Q
v̄q̄0 . We can
calculate these limiting values by setting the derivatives π̇q̄ (t) and v̇q (t) in Lemma 2 to zero and
solve for the limiting values π̄q̄ and v̄q̄ . A consequence is the following theorem.
3
For example, it would be easy to construct reset maps such that x0 (t) = t, i.e. x0 (t) simply tracks the passage of time and
vq0 (t) grows without bound for all states q ∈ Q.
21
Theorem 4: If the discrete-state Markov chain q(t) is ergodic with stationary distribution π̄
and we can find a non-negative solution v̄ = [v̄0 · · · v̄m ] such that
v̄q̄
X
λ(l) = bq̄ π̄q̄ +
X
λ(l) v̄ql Al ,
q̄ ∈ Q,
(34a)
l∈L0q̄
l∈Lq̄
then the differential equation (32) is stable and the average age of the AoI SHS is given by
∆=
X
v̄q̄0 .
(34b)
q̄∈Q
Proof of Theorem 4 is deferred to Section V as some elements of the proof will be more
clear after we employ the LCFS-S and LCFS-W systems as examples. In particular, Section IV
uses Theorem 4 to prove Theorem 2. We will see that the construction of a simple table that
enumerates the transitions al = (ql , ql0 , λ(l) , Al ) will be sufficient to immediately write down and
solve the equations of Theorem 4.
We note that Theorem 4 is in a form convenient for deriving closed form AoI expressions for
simple queues. However, this form is not concise in that multiple instances of vql on the right
side of (34a) may refer to the same vq . In Section V, we rewrite these equations in matrix form
that is convenient for numerical evaluation and also facilitates a proof of Theorem 4. For the
non-negative linear reset maps Al that we employ for age analysis, we will show that stability of
the differential equation (32) is equivalent to an eigenvalue constraint on a non-negative matrix.
IV. LCFS AGE : SHS A NALYSIS
A. LCFS With Preemption In Service: SHS Analysis
Without loss of generality, we assume a two-source LCFS-S system and we solve for the
average age ∆1 of source 1. In terms of the N source system, source 2 represents the composition
of all other sources. We can represent the LCFS-S system discrete state q(t) = q ∈ Q = {0, 1, 2}
such that q = 0 indicates that the server is idle and q ∈ {1, 2} denotes the source of the update
packet in service. The continuous state is x(t) = [x0 (t) x1 (t)] where x0 (t) is the current age
∆1 (t) of the source 1 process, and x1 (t) encodes what ∆1 (t) will become if the packet-in-service
is delivered. We note that x1 (t) is irrelevant in state 0. In state 1, x1 (t) is the age of the source
1 update in service. In state 2, a source 2 update is in service. Because a service completion by
this update has no effect on the source 1 age, x1 (t) is also irrelevant in state 2.
22
0
1
2
3
4
1
5
7
6
2
Fig. 4. The SHS Markov chain for updates of source 1 in the two-source LCFS-S system. In state 0 the system is idle while in
state i ∈ {1, 2} a source i update is in service. The transition rates and transition/reset maps for links l = 1, . . . , 7 are shown
in Table I.
l ql → ql0 λ(l) xAl
Al
vql Al
1
0
1 0 → 1 λ1 [x0 0] [ 0 0 ] [v00 0]
2
0→2
λ2 [x0 0] [ 10 00 ] [v00 0]
3
1→0
µ [x1 0] [ 01 00 ] [v11 0]
4
1→1
λ1 [x0 0] [ 10 00 ] [v10 0]
5
1→2
λ2 [x0 0] [ 10 00 ] [v10 0]
6
2→0
µ [x0 0] [ 10 00 ] [v20 0]
7
2→1
λ1 [x0 0] [ 10 00 ] [v20 0]
TABLE I
TABLE OF TRANSITIONS FOR THE M ARKOV CHAIN IN F IGURE 4.
A Markov chain for the discrete state q(t) is shown in Figure 4. The corresponding SHS
transitions al are shown in Table I. In the figure, a link l from node ql to ql0 indicates that
transitions in state ql to state ql0 occur at exponential rate λ(l) as given in the table. In constructing
the table, we first identify the x → x0 mapping for each transition l, from which we infer the
matrix Al such that x0 = xAl . Given Al , it is convenient to also include vql Al in the table to
facilitate the use of Theorem 4. We now explain each transition l:
l = 1 A source 1 update arrives at an empty queue. With this arrival, x00 = x0 is unchanged
because the arrival does not yield an age reduction until it departs. However x01 = 0
because the arriving source 1 update is fresh and its age is zero at that instant.
l = 2 A source 2 update arrives at an empty queue. The age x00 = x0 is unchanged because
the arrival does not change the age. However x01 = 0 because x1 is irrelevant in state 2.
l = 3 A source 1 update completes service and is delivered to the monitor. In this transition,
23
x00 = x1 , corresponding to the age being reset to the age of the source 1 that just
completed service. Also note that x01 = 0 since x1 becomes irrelevant when the system
enters state 0.
l = 4 The source 1 update in service is preempted by a fresh source 1 update. The age x0
remains unchanged while x1 is reset to zero because the new update is fresh.
l = 5 The source 1 update in service is preempted by a source 2 update. The age x00 = x0 is
unchanged and x01 = 0 since x1 becomes irrelevant in state 2.
l = 6 A source 2 update completes service. The source 1 age x0 is unchanged. In the transition
to state 0, x1 remains irrelevant and is set to zero.
l = 7 The source 2 update in service is preempted by a fresh source 1 update. The age x0 is
unchanged while x01 = 0 because the new update is fresh.
We note that this SHS includes a self-transition in which the discrete state is unchanged but
a reset occurs in the continuous state. Specifically, in state 1, the self-transition of link 4 marks
the arrival of a source 1 update packet that preempts the source 1 packet in service. This leaves
the discrete state q(t) and the current age x0 (t) unchanged, but the more recent timestamp of
the new source 1 update resets x1 (t).
The evolution of x(t) depends on the discrete state q(t). Specifically, when q(t) = q,
[1 0] q = 0, 2,
ẋ(t) = bq =
[1 1] q = 1.
(35)
The interpretation of (35) is that the age ∆1 (t) = x0 (t) increases at unit rate with time t in all
discrete states but x1 (t) increases at unit rate only in state q = 1 in which there is a relevant
update in service.
To employ Theorem 4, we first use (31a) to show that the stationary probability vector π̄
satisfies π̄D = π̄Q with
D = diag[λ, µ + λ, µ + λ1 ],
Applying
P2
i=0
0 λ1 λ2
Q = µ λ1 λ2 .
µ λ1 0
(36)
π̄i = 1, the stationary probabilities are
h
i
h
i
−1
=
(1
+
ρ)
π̄0 π̄1 π̄2
1 ρ1 ρ2 .
(37)
24
Since b0 = [1 0], evaluation of (34a) in Theorem 4 at q̄ = 0 yields
λ[v̄00 v̄01 ] = [π̄0 0] + µ[v̄11 0] + µ[v̄20 0].
(38)
We see from (38) that v̄01 = 0. This is a consequence of x1 being irrelevant in state q = 0. In
particular, x1 (t)δ0,q(t) = 0 for all t because x1 (t) is held at 0 when q(t) = 0 (by our convention
for irrelevant variables) and δ0,q(t) = 0 when q(t) 6= 0. Thus v01 (t) = E[x1 (t)δ0,q(t) ] = 0 and we
refer to v01 as irrelevant. In general, when xj is irrelevant in state q, vqj (t) = 0 for all t and we
refer to vqj (t) as irrelevant.
Evaluating (34a) at q̄ = 1 and q̄ = 2 produces
(µ + λ)[v̄10 v̄11 ] = [π̄1 π̄1 ] + λ1 [v̄00 0] + λ1 [v̄10 0] + λ1 [v̄20 0],
(µ + λ1 )[v̄20 v̄21 ] = [π̄2 0] + λ2 [v̄00 0] + λ2 [v̄10 0].
(39a)
(39b)
In terms of the vectors
π̄ = [π̄0 π̄1 π̄2 ],
(40)
v̄ = [v̄0 v̄1 v̄2 ] = [v̄00 v̄01 v̄10 v̄11 v̄20 v̄21 ],
(41)
v̄D = π̄B + v̄R
(42)
we have
where
D = diag[λ, λ, µ + λ, µ + λ, µ + λ1 , µ + λ1 ],
(43)
1 0 0 0 0 0
B = 0 0 1 1 0 0 ,
0 0 0 0 1 0
0 0 λ1 0 λ2 0
0 0 0 0 0 0
0 0 λ 0 λ 0
1
2
R=
.
µ 0 0 0 0 0
µ 0 λ1 0 0 0
0 0 0 0 0 0
(44)
We observe that the columns and rows of R corresponding to the irrelevant variables v̄01 and
25
2
4
1
0
1
5
3
Fig. 5. The simplified SHS Markov chain for updates of source 1 in the two-source LCFS-S system. In state 0 the system
is idle while in state 1 an update of either source 1 or 2 is in service. The transition rates and transition/reset maps for links
l = 1, . . . , 5 are shown in Table II.
v̄21 are zero. Gathering the relevant variables, we obtain
ρ
−ρ1
0
−ρ2
0 1+ρ
0
−ρ2
1
2
[π̄0 π̄1 π̄1 π̄2 ] = [v̄00 v̄10 v̄11 v̄20 ]
.
−1
µ
0
1
+
ρ
0
−1 −ρ1
0
1 + ρ1
(45)
It follows from (37) and (45) that
v̄00
v̄10
v̄20
1 + ρ2
1
1
+
,
=
µ(1 + ρ)
ρ1
1+ρ
1
ρ1
=
1+ρ+
,
µ(1 + ρ)
1+ρ
ρ2 (1 + ρ)
ρ2
1
+
.
=
µ(1 + ρ)
ρ1
1+ρ
(46a)
(46b)
(46c)
From (39), it can be seen that v̄11 is also non-negative. Thus, Theorem 4 implies that the average
P
age for source 1 is ∆ = 2q=0 v̄q0 . Applying (46) yields Theorem 2(a) for source i = 1.
B. LCFS-S: A simpler SHS analysis
We note that the preceding analysis of the LCFS-S system used the discrete state to track the
source of the update in service. However, since all updates are served at rate µ, specifying the
source of an update in service is not essential for a Markov model for the server occupancy. We
will now show that we can track the source 1 age without specifying the source of the update
in service. By employing transition reset maps that depend on the source of the arriving new
update, we can forgo using the discrete state to track the source of an update in service. We now
demonstrate this technique with a simpler SHS derivation of the LCFS-S age. In this analysis,
the discrete state tracks only whether the server is busy.
26
l ql → ql0 λ(l)
xAl
Al
vql Al
[x0 0]
[ 10 00 ]
[v00 0]
1
0→1
λ1
2
0→1
λ2 [x0 x0 ] [ 10 10 ] [v00 v00 ]
3
1→0
µ
[x1 0]
[ 01 00 ]
[v11 0]
4
1→1
λ1
[x0 0]
[ 10 00 ]
[v10 0]
5
1→1
λ2 [x0 x0 ] [ 10 10 ] [v10 v10 ]
TABLE II
TABLE OF TRANSITIONS FOR THE M ARKOV CHAIN IN F IGURE 5.
Just as in the original LCFS-S SHS, the continuous state is x(t) = [x0 (t) x1 (t)] where x0
is the current source 1 age. However, x1 (t) is now what the age would become if the update in
service is delivered. In state 1, both x0 (t) and x1 (t) increase at unit rate; i.e. b1 = [1 1]. On
the other hand, in state 0, x1 is meaningless and b0 = [1 0]. The transitions are:
l = 1 A fresh source 1 update goes into service; x01 = 0 because the update is fresh.
l = 2 A fresh source 2 update goes into service and x01 = x0 . If the source 2 update does
complete service, it doesn’t reduce the age of the process of interest.
l = 3 The update in service is delivered. The age x0 is reset to x00 = x1 . If this delivered
update is from source 1, then x00 < x0 . However, if this update is from source 2, then
x00 = x0 and no age reduction occurs. Note that this age reduction was encoded in the
prior transition that put this update in service.
l = 4 The update in service is replaced by a fresh source 1 update. This reset map is essentially
the same as for transition l = 1.
l = 5 The update in service is replaced by a fresh source 2 update. This reset map is essentially
the same as for transition l = 2.
The Markov chain for the discrete state has stationary probabilities
π̄ = [π̄0 π̄1 ] =
h
1
1+ρ
ρ
1+ρ
i
.
(47)
In this system, v0 = [v00 v01 ] and v1 = [v10 v11 ]. Evaluating (34a) at q̄ = 0, 1 produces
λ[v̄00 v̄01 ] = [π̄0 0] + µ[v̄11 0],
(48a)
27
2
1
0
3
Fig. 6. The simplified SHS Markov chain for updates of source 1 in the two-source LCFS-S system. The system is always busy
serving either a real or fake update. The transition rates and transition/reset maps for links l = 1, 2, 3 are shown in Table III.
(µ + λ)[v̄10 v̄11 ] = [π̄1 π̄1 ] + λ1 [v̄00 0] + λ2 [v̄00 v̄00 ] + λ1 [v̄10 0] + λ2 [v̄10 v̄10 ].
(48b)
As expected, we see from (48a) that v̄01 = 0 because x1 is irrelevant in state 0. Normalizing
by the service rate µ, we obtain
ρv̄00 = π̄0 /µ + v̄11 ,
(49a)
(1 + ρ)v̄10 = π̄1 /µ + ρv̄00 + ρv̄10 ,
(49b)
(1 + ρ)v̄11 = π̄1 /µ + ρ2 v̄00 + ρ2 v̄10 .
(49c)
Solving (49), it can be shown that v̄00 , v̄10 and v̄11 are all non-negative. Moreover, calculation
of ∆ = v̄00 + v̄10 yet again yields Theorem 2(a) for source i = 1.
C. LCFS-S: An even simpler SHS analysis with fake updates
We note that the preceding analysis of the LCFS-S system used the discrete state to track
whether an update is in service. However, it turns out that this is not essential and we now
analyze AoI in the multi-source LCFS-S queue using a one-state SHS in which there is always
an update in service. The key idea is that when an update completes service and the server would
become idle, we create a “fake” update to keep the server busy. This fake update is a duplicate
of the previous update and has its same timestamp. The reason this trick works is two-fold. First,
if the fake update were to complete service, the age at the monitor is unchanged because the
update timestamp is the same as that of the previously delivered update. Second, when a new
(true) update is submitted, it immediately preempts any fake update that may have been keeping
the server busy.
The one-state SHS is shown in Figure 6 and the corresponding table of transitions is given
in Table III. Just as in the previous LCFS-S SHS, the continuous state is x(t) = [x0 (t) x1 (t)]
where x0 is the current source 1 age and x1 (t) is what the age would become if the update in
28
l ql → ql0 λ(l)
xAl
Al
vql Al
[x0 0]
[ 10 00 ]
[v00 0]
1
0→0
λ1
2
0→0
λ2 [x0 x0 ] [ 10 10 ] [v00 v00 ]
3
0→0
µ [x1 x1 ] [ 01 01 ] [v01 v01 ]
TABLE III
TABLE OF TRANSITIONS FOR THE M ARKOV CHAIN IN F IGURE 5.
service is delivered. Both x0 (t) and x1 (t) increase at unit rate and b0 = [1 1]. The transitions
are:
l = 1 A fresh source 1 update goes into service; x01 = 0 because the update is fresh.
l = 2 A fresh source 2 update goes into service and x01 = x1 . If the source 2 update does
complete service, it doesn’t reduce the age of the process of interest.
l = 3 The update in service is delivered. The age x0 is reset to x00 = x1 but x1 is unchanged:
x01 = x1 . This corresponds to creating a fake update with the same timestamp as the
update that was just delivered.
The Markov chain for the discrete state has the trivial stationary probability π0 = 1. In this
system, v0 = [v00 v01 ]. Evaluating (34a) at q̄ = 0 produces
(µ + λ)[v̄00 v̄01 ] = [1 1] + λ1 [v00 0] + λ2 [v00 v00 ] + µ[v01 v01 ].
(50)
Solving these two equations for v00 and v01 , the average age ∆ = v00 yet again yields Theorem 2(a) for source i = 1.
D. LCFS With Preemption Only In Waiting: SHS Analysis
Following the simplified SHS method introduced in Section IV-B, we now model the LCFS-W
system as a stochastic hybrid system. Once again we assume a two-source system and we solve
for the average age ∆1 of source 1. In terms of the N source system, source 2 represents the
composition of all other sources.
The LCFS-W system with discrete states q ∈ Q = {0, 1, 2} is shown in Figure 7 with the
corresponding transition rates λ(l) and reset maps Al given in Table IV. Much like the two-state
analysis of the SHS for the LCFS-S system, the discrete state tracks the number of updates in
29
2
5
1
4
0
1
3
7
2
8
6
Fig. 7. The simplified SHS Markov chain for updates of source 1 in the two-source LCFS-W system. The state i indicates the
number of updates in the system. The transition rates and transition/reset maps for links l = 1, . . . , 8 are shown in Table IV.
the system but not the source of each update. Whether a delivered update reduces the age of
source 1 is embedded in the continuous state.4
The continuous state is x(t) = [x0 (t) x1 (t) x2 (t)] where x0 (t) is the current age ∆1 (t) of
the source 1 process, x1 (t) is what the age would be if the update in service were delivered
at time t, and x2 (t) is what the age would be if the update-in-waiting were delivered at time
t. In state q = 0, x1 and x2 are irrelevant. In state q = 1, x2 is irrelevant. Following our prior
convention, relevant components in each state grow at unit rate while irrelevant components are
fixed at zero. Consequently, in discrete state q(t) = q, the continuous state evolves according to
[1 0 0], q = 0,
ẋ(t) = bq = [1 1 0], q = 1,
(51)
[1 1 1], q = 2.
For x0 (t) and x1 (t), the transition maps are similar to those in the two-state SHS for the
LCFS-S system. The additional complications involve how x2 (t) modifies x1 (t) when an update
completes service (and a waiting update goes into service) and how x2 (t) is modified when a
waiting update is preempted. In particular, the reset maps A5 and A8 for transitions l = 5 and
l = 8 are less straightforward than the others. Under transition l = 5, the update in service
is joined by a new source 2 update that waits in the queue. Assuming it is not preempted in
waiting, this new source 2 update enters service and is eventually delivered only after the update
in service is delivered. This delivery of the update in service reduces the age to x1 and puts
the new source 2 update into service. However, as the SHS is tracking the source 1 age, the
eventual delivery of the new source 2 update will not reduce the age. Thus transition l = 5 sets
4
The fake updates method fails for the LCFS-W system because the discrete state must track whether the update in service
is real or fake because an update in service would be preempted if it were fake but not if it were real.
30
l ql → ql0 λ(l)
xAl
1 0 → 1 λ1
[x0 0
2 0 → 1 λ2 [x0 x0
3 1→0 µ
[x1 0
4 1 → 2 λ1 [x0 x1
5 1 → 2 λ2 [x0 x1
6 2 → 1 µ [x1 x2
7 2 → 2 λ1 [x0 x1
8 2 → 2 λ2 [x0 x1
vql Al
0]
[v00 0
0] [v00 v00
0]
[v11 0
0] [v10 v11
x1 ] [v10 v11
0] [v21 v22
0] [v20 v21
x1 ] [v20 v21
0]
0]
0]
0]
v11 ]
0]
0]
v21 ]
TABLE IV
TABLE OF TRANSITIONS FOR THE M ARKOV CHAIN IN F IGURE 7.
x02 = x1 so that the age upon delivery of the new source 2 update will leave the age unchanged
from the age that will be established by the prior service completion. The same effect occurs in
transition l = 8 in which an arriving source 2 update preempts the update-in-waiting.
From the Markov chain in Figure 7, it is easy to see that the discrete state has stationary
distribution
[π̄0 π̄1 π̄2 ] = Cπ [1 ρ ρ2 ]
(52)
where Cπ = (1 + ρ + ρ2 )−1 is the normalizing constant. From (34a) with q̄ ∈ Q, we obtain
λv̄0 = [π̄0 0 0] + µ[v̄11 0 0],
(λ + µ)v̄1 = [π̄1 π̄1 0] + λ1 [v̄00 0 0] + λ2 [v̄00 v̄00 0] + µ[v̄21 v̄22 0],
(53a)
(53b)
(λ + µ)v̄2 = [π̄2 π̄2 π̄2 ] + λ1 [v̄10 v̄11 0] + λ2 [v̄10 v̄11 v̄11 ]
+ λ1 [v̄20 v̄21 0] + λ2 [v̄20 v̄21 v̄21 ].
(53c)
We see from (53a) that v̄01 and v̄02 are zero because x1 (t) and x2 (t) are irrelevant in state 0.
Similarly, (53b) implies v̄12 = 0 because x2 (t) is irrelevant in state 1. Gathering the relevant
variables and normalizing by the service rate µ, we obtain
ρv̄00 = π̄0 /µ + v̄11 ,
(54a)
(1 + ρ)v̄10 = π̄1 /µ + ρv̄00 + v̄21 ,
(54b)
(1 + ρ)v̄11 = π̄1 /µ + ρ2 v̄00 + v̄22 ,
(54c)
v̄20 = π̄2 /µ + ρv̄10 ,
(54d)
31
v̄21 = π̄2 /µ + ρv̄11 ,
(1 + ρ)v̄22 = π̄2 /µ + ρ2 v̄11 + ρ2 v̄21 .
(54e)
(54f)
We employ (54e) and (54f) to write
v̄22 =
1 + ρ2
π̄2 + ρ2 v̄11 .
1+ρ
(55)
We now apply (54e) and (55) to the other equations in (54), yielding
ρv̄00 = π̄0 + v̄11 ,
v̄10 =
(56a)
1
+ v̄11 ,
µ(1 + ρ)
(1 + ρ)v̄11 = π̄1 /µ + ρ2 v̄00 +
(56b)
1 + ρ2
π̄2 + ρ2 v̄11 ,
µ(1 + ρ)
v̄20 = π̄2 /µ + ρv̄10 .
(56c)
(56d)
From (56), some algebra will show
v̄11 =
Cπ (1 + ρ + ρ3 )
ρ
1
−
.
µ(1 + ρ) ρ1
µ(1 + ρ)2
(57)
To verify that v̄11 is non-negative, we note that ρ1 ≤ ρ and that for fixed ρ, v̄11 is minimized
over all ρ1 at ρ1 = ρ. Some algebra will verify that v̄11 ≥ 0 when ρ1 = ρ. It then follows from
(54) and (56) that all components of v̄ are non-negative. Moreover, it also follows from (34b)
and (56) that the average age is
∆ = v̄00 + v̄10 + v̄20
=
1 π 0 + π2 1 + ρ + ρ 2
+
+
v̄11 .
µ
ρ
ρ
(58)
The claim of Theorem 2(b) then follows from substitution of (52) and (57) in (58).
V. SHS M ATRIX R EFORMULATION
In this section, we derive a matrix representation of (32) as well as (34a) in Theorem 4. This
reformulation will facilitate a proof of Theorem 4. Starting with the differential equations (32),
we define the departure rate from state q̄ as
dq̄ =
X
l∈Lq̄
λ(l) .
(59)
32
We also define
Lij = {l ∈ L : ql = i, ql0 = j},
i, j ∈ Q,
(60)
as the set of SHS transitions from state i to state j. With the observation that L0q̄ = ∪i Liq̄ , we
now can rewrite (32) as
v̇q̄ (t) = bq̄ π̄q̄ +
XX
i
λ(l) vql (t)Al − dq̄ vq̄ (t),
q̄ ∈ Q.
(61)
l∈Liq̄
With the substitution q̄ = j and the observation that ql = i for all l ∈ Lij , we obtain
v̇j (t) = bj π̄j +
X
vi (t)
i
X
λ(l) Al − dj vj (t),
j ∈ Q.
(62)
l∈Lij
We define the block matrix R such that block i, j of R is given by
Rij =
X
λ(l) Al ,
i, j ∈ Q.
(63)
l∈Lij
We also define the block diagonal matrices5
B = diag[b0 , b1 , . . . , bm ],
(64)
D = diag[d0 In , d1 In , . . . , dm In ].
(65)
With the definition of the long row vector v(t) = [v0 (t) · · ·
vm (t)], we can write (62) in
vector form as
v̇(t) = π̄B + v(t)(R − D).
(66)
We note that setting v̇(t) = 0 and solving for v(t) = v̄ yields
v̄D = π̄B + v̄R,
(67)
just as we observed in (42) for the LCFS-S system.
We note that (66) and (67) are vectorized forms of (32) and (34a). In vector form, the claim
of Theorem 4 is that a non-negative solution v̄ for (67) implies the differential equation (66) is
stable and thus E[x0 (t)] converges to the average age.
5
Note that B is an (m + 1) × (m + 1)(n + 1) matrix with ith row [0(n+1)i
bi
0(n+1)(m−i) ], i = 0, 1, . . . , m.
33
As we saw in (44) for the LCFS-S example, there may be irrelevant variables that yield zero
columns in B and corresponding zero rows and zero columns in R. These irrelevant variables
will have zero derivatives and will be perpetually zero. Stability of the differential equations
depends only on the stability of the relevant variables. Thus we omit the irrelevant variables
and form v̂, a long row vector of the relevant variables in v̄. Similarly, we form B̂ and D̂
by deleting the rows/columns of B and D corresponding to irrelevant variables. It follows that
relevant variables satisfy
dv̂(t)
= π̄ B̂ + v̂(t)(R̂ − D̂).
dt
(68)
If there is a non-negative solution v̄ for (67), then the relevant components have a fixed point
v̂(t) = ṽ that satisfies
ṽD̂ = π̄ B̂ + ṽR̂.
(69)
Let s = maxi di , then sI − D̂ is a non-negative diagonal matrix. Adding ṽ(sI − D̂) to both sides
of (69) yields
ṽs = π̄ B̂ + ṽ(sI + R̂ − D̂).
(70)
Because the reset maps Al are binary, the matrices R and R̂ are non-negative and thus the matrix
sI + R̂ − D̂ is also non-negative. It follows that sI + R̂ − D̂ has a dominant real eigenvalue
r(s) ≥ 0 with an associated non-negative non-zero right eigenvector u> such that || ≤ r(s) for
any other eigenvalue [37, Exercise 1.12]6 . Right multiplying (70) by u> , we obtain
sṽu> = π̄ B̂u> + r(s)ṽu> ,
(71)
[s − r(s)]ṽu> = π̄ B̂u>
(72)
which simplifies to
Because the irrelevant variables have been omitted, π̄ B̂ > 0. Since u ≥ 0 and is not trivially
zero, it follows that π̄ B̂u> > 0. This implies r(s) < s. Moreover, if is an eigenvalue of
6
This is a weak form of the Perron-Frobenius theorem that does not require irreducibility of the non-negative matrix.
34
sI + R̂ − D̂ then − s is an eigenvalue of R̂ − D̂ and has real part
Re( − s) = Re() − s
≤ || − s
≤ r(s) − s < 0.
(73)
Thus the differential equation (68) for the relevant variables is stable and it follows that the
differential equation for v(t) is also stable. This completes the proof of Theorem 4.
VI. P ERFORMANCE E VALUATION
Here we use Theorems 1 and 2 to examine achievable AoI regions for two-source FCFS and
LCFS systems. In addition, resource sharing issues for N sources are explored in Section VI-C.
These issues are pertinent, for example, to embedded systems that are a part of the Internet
of Things (IoT) [38]. As in our models, such systems have multiple sensors (sources), each
generating updates independently. These updates are queued for transmission (service) by the
system’s wireless interface. Their transmission time is modeled by the exponential distribution
in this work. Alternatively, the queueing disciplines may be implemented at an access point,
which queues updates from a network of a large number of distributed sensors and forwards
them to a server for storage and analysis.
A. M/M/1 FCFS: Two Sources
For a two-user system with normalized service rate µ = 1, Theorem 1 yields the contours of
achievable age pairs (∆1 , ∆2 ) for fixed load ρ that are shown in Figure 8. The set of feasible
age pairs (∆1 , ∆2 ) is given by the union of all such contours. The lower left “corner” point
(marked ∗) where the sum ∆1 + ∆2 is minimized is obtained at ρ1 = ρ2 = 0.306, yielding
∆1 = ∆2 = 5.30. It then follows from Theorem 1 that if those two sources were to share a rate
µ = 2 server, then each source would obtain average age ∆1 = ∆2 = 2.65. By comparison,
if server resources were partitioned and each source employed an individual rate µ = 1 server
with optimal load ρ1 = 0.531, Equation (8) will yield ∆1 = 3.48. Thus we observe a trunking
efficiency in having two status-updating sources share a combined service facility.
We note that the ρ = 0.612 age contour appears in Figure 8 to be superior to all other age
contours. In fact, this contour marks the Pareto frontier of achievable ages at the operating point
35
8
ρ=0.40
ρ=0.43
ρ=0.75
ρ=0.53
ρ=0.684
ρ=0.612
7.5
7
∆2
6.5
6
5.5
5
4.5
4
4
5
6
7
8
∆1
Fig. 8. Age contours (∆1 , ∆2 ) for fixed total load ρ1 + ρ2 = ρ for two sources sharing a rate µ = 1 FCFS M/M/1 queue. The
minimum sum age point marked ∗ in the lower left is achieved by ρ1 = ρ2 = 0.306. The point 4 marks the Nash equilibrium
age pair achieved by unilateral optimization.
20
ρ=0.612
0.45 < ρ < 0.6
∆2
15
10
5
3.5
4
4.5
5
5.5
∆1
Fig. 9. Age contours (∆1 , ∆2 ) for fixed total load ρ1 + ρ2 = ρ for two sources sharing a rate µ = 1 FCFS M/M/1 queue.
In the lower right corner, the ρ = 0.612 contour is seen to be Pareto optimal for ρ1 ≈ ρ2 . However, in the upper left corner,
ρ ≈ 0.53 can reduce ∆1 for constant ∆2 as ρ2 → 0 and ρ1 → ρ.
∗, corresponding to offered loads ρ1 = ρ2 . However, the optimal load ρ will vary along the Pareto
frontier. For example, as ρ1 → ρ and ρ2 → 0, the ρ = 0.53 (the optimal load for a single source)
contour offers reduced ∆1 for fixed ∆2 . This is shown in Figure 9 where the solid line marks
the ρ = 0.612 contour while dotted lines mark contours for ρ in the interval 0.45 ≤ ρ ≤ 0.6.
The general insight for FCFS systems is that multiuser age optimization depends on both the
total load ρ and the allocation of load among individual sources.
36
12
ρ=0.612 FCFS
ρ=0.612 LCFS-S
ρ=0.612 LCFS-W
ρ=1 LCFS-S
ρ=1 LCFS-W
ρ=3 LCFS-W
ρ=3 LCFS-S
10
∆2
8
6
4
2
2
4
6
8
10
12
∆1
Fig. 10. Update ages of two users under the queuing policies of FCFS, LCFS with preemption in service (LCFS-S), and LCFS
with preemption only in waiting (LCFS-W). The service rate is µ = 1.
B. M/M/1 LCFS: Two Sources
With N = 1 source, some algebra applied to Theorem 2 will verify that LCFS-S (with
preemption in service) yields smaller age ∆1 than LCFS-W (with preemption only in waiting).
For N = 2 sources, Figure 10 compares the LCFS policies for different choices of the arrival
rate ρ. The achievable age pair contour for a given ρ is obtained by varying ρ1 and ρ2 such
that ρ1 + ρ2 = ρ. The service rate is µ = 1. We observe that LCFS-W is better than LCFS-S
when arrival rates ρ are low but somewhat worse when arrival rates are high. Because this same
behavior does not hold for the single source system, we speculate that LCFS-W benefits at low
arrival rates from not preempting an update in service with an update from some other source
whose update is not currently in service.
For comparison with FCFS, we also plot the FCFS age contour for total load ρ = 0.612,
which was shown in Figure 8 to be optimal in the neighborhood of ρ1 ≈ ρ2 and near-optimal
elsewhere. For the same total offered load ρ = 0.612, Figure 10 shows that the age contours
obtained under both LCFS policies are better than those obtained under FCFS. To summarize
Figure 10, if a system can choose all of ρ, ρ1 , and ρ2 , LCFS-S is the policy that minimizes sum
age.
In general, the choice of policy is not as straightforward. Figure 11 shows, for N = 2 sources
37
and µ = 1, that the policy that minimizes the sum age over points in the cartesian product of
{0 < ρ1 /ρ ≤ 0.5} and {0 < ρ < 2}. FCFS is the policy of choice for ρ < 0.4. It is also
the policy of choice for larger ρ < 1, however, only when source 1 updates constitute a small
enough fraction of ρ. This fraction gets smaller as ρ → 1. Similarly, LCFS-W is a policy of
choice over LCFS-S even for large ρ when the load due to source 1 updates is a small enough
fraction of the total load ρ.
C. Multiple Source Resource Allocation
P
We can compare the FCFS and LCFS systems in terms of the sum of ages ∆Σ = N
i=1 ∆i
PN
when users share the system capacity in fixed proportions such that ρi = αi ρ with i=1 αi = 1.
We note that Theorem 2 implies that the following observations hold for both types of LCFS
systems:
•
Each user i has age ∆i that decreases monotonically with total load ρ.
•
The sum age ∆Σ is minimized by equal offered loads ρi = ρ/N .
In addition, it follows from Theorem 2 that
•
∆i in a LCFS-W system is strictly less than that under LCFS-S iff
N
1 X −1
α > (1 + ρ)αW (ρ).
N i=1 i
(74)
The right side of Equation (74) is a nondecreasing O(ρ) function. Thus given N updaters and
the proportions αi in which they share the load, there is a maximum ρ such that LCFS-W is
better than LCFS-S. We had observed this for N = 2 sources in Figures 10 and 11.
Optimization of the offered load ρ is more complicated in the FCFS system. Theorem 1 implies
that for each user i, the value of total load ρ that minimizes the average age ∆i depends on
the proportional load αi . Load optimization of the FCFS system will depend on a performance
metric derived from ∆1 , . . . , ∆N . However, it is easy to see that for any given 0 < ρ < 1, the
sum of ages ∆Σ is minimized when αi = 1/N for all i. This follows from the fact that for a
given ρ,
∂ 2 ∆i
∂ρ2i
> 0 for ρi ∈ (0, ρ). Also, as a result, the sum of ages is a convex function of ρi ,
i = 1, . . . , N . Further, all the ∆i are the same function over ρi ∈ (0, ρ).
To compare the FCFS, LCFS-S and LCFS-W systems, we focus on symmetric systems with
ρi = ρ/N . In this case, each user has identical average age ∆i given by Theorem 1 or Theorem 2
38
LCFS−S
ρ
1.5
LCFS−W
1
0.5
FCFS
0
0
0.1
0.2
0.3
0.4
0.5
ρ1/ρ
Fig. 11. For the policies of FCFS, LCFS-W, and LCFS-S, we show the region where they are sum age minimizing. FCFS
minimizes the sum age for all points (ρ1 /ρ, ρ) that lie below the squares. LCFS-W minimizes the sum age for the points above
the squares and to the left of the circles. LCFS-S is sum age minimizing for all points above the circles. The service rate is
µ = 1.
with ρi = ρ/N and ρ−i = (N − 1)ρ/N . For symmetric LCFS systems, Equation (74) simplifies
to
N > (1 + ρ)αW (ρ).
(75)
Therefore for fixed ρ, LCFS-W outperforms LCFS-S when the number of sources N is large. To
further examine this, we now assume a large system with N 1 sources. As N becomes large,
ρ−i → ρ and it follows from (3) that αW (ρ)/N → 0. With these limits, it is straightforward to
N
show that Theorems 1 and 2 imply that ∆i approaches ∆N
F (ρ) (FCFS) or ∆S (ρ) (LCFS-S) or
∆N
W (ρ) (LCFS-W) where
N (1 + ρ)ρ2
1
1
=
+
+ ,
µ N 3 (1 − ρ)3 N (1 − ρ) ρ
1
N
N
∆S (ρ) =
1+ ,
µ
ρ
1
N
N
1+
.
∆W (ρ) =
µ
ρ(1 + ρ)
∆N
F (ρ)
(76a)
(76b)
(76c)
From (76b) and (76c), we see that both LCFS systems have AoI that decreases with total load
ρ. By contrast, the FCFS system is subject to the stability constraint ρ < 1 and it benefits from
matching the load ρ to the number of sources N . Let ρ∗N minimize ∆N
F (ρ) over 0 ≤ ρ ≤ 1.
Because ρ < 1 and N is large, the first term on the right side of (76a) becomes negligible for ρ
near ρ∗N as N becomes large. Thus ρ∗N converges to the minimizer of [N (1 − ρ)]−1 + ρ−1 , i.e.,
√
N
∗
ρN = √
.
(77)
N +1
39
∗
It follows that ∆N
F (ρN ) → N/µ as N becomes large. Similarly, for the LCFS systems, if ρ
N
grows large as N grows large, then ∆N
S (ρ) and ∆W (ρ) also approach N/µ. In this sense, as
the number of users becomes large, all three systems become equivalent. On the other hand, in
many settings the offered load ρ may be physically constrained. For example, if each update
requires energy for a wireless transmission, then it would be appropriate to compare the FCFS
and LCFS systems on an equal power basis. In this spirit, we note that if all three systems
∗
N ∗
N ∗
N ∗
operate at offered load ρ = ρ∗N , then ∆N
W (ρN )/∆F (ρN ) → 1.5 and ∆S (ρN )/∆F (ρN ) → 2 as
N becomes large. Thus, one can argue that FCFS is more efficient than either LCFS discipline
for large symmetric systems.
D. Non-cooperative Rate Adaptation
We now examine how sources may individually adapt their updating rates. When source i is a
status updater in the presence of “interfering” traffic with aggregate load ρ−i from other sources,
it may be in the interest of source i to unilaterally optimize its updating load ρi in response to
the aggregate other load ρ−i .
For the N -user FCFS system, it was observed [10] that the adaptation
ρ∗i (ρ−i ) =
1 − ρ−i [1 − ρ−i ]2
+
2
32
(78)
is essentially the same as the best-response normalized load that exactly minimizes ∆i . It was
also noted that since the minima over ρi is broad and nearly flat, ρ̂i = 0.5(1 − ρ−i ), a rule of
thumb that a source should use half the the residual capacity, was a good linear approximation.
If each other source is a status updater that selects an update load to minimize its respective age
given the other sources’ update loads, we obtain the synchronous iterative algorithm
1 − ρ−i (n) [1 − ρ−i (n)]2
ρi (n + 1) =
+
.
2
32
(79)
This iterative algorithm was shown to work reasonably well for N = 2 users as it converges to
a fixed point with ρ1 = ρ2 = 0.342, and corresponding ages ∆1 = ∆2 = 5.4390 [10]. However,
it is easily verified that this iteration is unstable for N > 2 users. Similar distributed algorithms
ρ̂i (n + 1) = ωN [1 − ρ−i (n)], in which each node uses a fraction ωN (depending on the number of
sources N ) of the residual capacity can be shown to be stable. However, a network mechanism
that sends ωN to a source i might just as well send the appropriate load ρi to that source.
40
For both LCFS systems, each source i has incentive to generate updates as fast as possible. To
be precise, when a source i is subject to a maximum load constraint ρi ≤ ρ̄i and other sources
offer combined load ρ−i , then source i can decrease its average age ∆i by unilaterally increasing
ρi to ρ̄i . The Nash equilibrium of the system is for each node i to operate at maximum updating
load ρ̄i . The average age each node will obtain will depend on ρ̄i and the total updating load
PN
k=1 ρ̄k . This may be a desirable operating point for some nodes but not for others. Moreover,
if each node bears some cost for its offered load ρi , this operating point may be undesirable
even if the resulting age is small.
An alternate approach would be for node i to view its age as a function ∆i (ρi , ρ−i ) of its offered
load and the interfering load, and to adjust ρi to meet a target age constraint ∆i (ρi , ρ−i ) = δi .
Such an approach shares many commonalities with game-theoretic power control in wireless
CDMA systems [39] although discussion of these issues is beyond the scope of this work.
VII. C ONCLUSION
We have looked at the problem of multiple sources generating timely status updates at
interested recipients. We have employed a simple approach in which the communication network
is modeled by an M/M/1 queue with either FCFS service or one of two variants of LCFS service.
For these systems, we have derived and characterized the region of feasible status ages. Our
results show that there are nontrivial gains in trunking efficiency when N users share the system
capacity. However, achieving these gains appears to require coordinated load balancing of the
sources. In particular, high offered load at an FCFS system induces high AoI through queueing
delays. A lossy LCFS discipline can mitigate this problem but its packet discarding policy may
encourage sources to operate at excessively high offered loads.
The preliminary insights from these simple models lead us to believe that the age of information represents a new and useful metric for the analysis of status updating systems. Moreover,
it should be apparent that many challenges remain in modeling, characterizing and optimizing
practical status updating systems.
As a step in this direction, we have introduced stochastic hybrid systems as a new way to
evaluate AoI in queues with memoryless service. The SHS method, in the form of Theorem 4,
provides a way to derive closed-form AoI results for simple queues described by finite-state
Markov chains. In addition, Theorem 4 will permit evaluation of AoI in substantially more
complex queueing systems that capture more realistic service facilities. These include facilities
41
that prioritize service of updates of certain sources over others, heterogeneous service facilities,
state-dependent updating policies, and facilities in which arrival and service rates change with
time. More generally, continuous-time Markov chains are a widely used tool for the modeling
and performance evaluation of complex service systems [40]. When these systems are delivering
status updates, Theorem 4 enables straightforward numerical evaluation of AoI.
Theorem 4 is based on non-negative linear transition reset maps. We note that this choice
is not unique and that there are other ways to embed AoI tracking in the continuous state
of an SHS with linear reset maps. However, stability results for the corresponding differential
equations will need to be derived. We note that there is substantial literature on SHS stability
and ergodicity [36], but this is not yet fully understood in the specific context of age systems.
We further observe that a general SHS with time-varying and state-dependent transition rates
also may prove to be useful in further characterizing age processes.
A PPENDIX A
P ROOF OF L EMMA 1
The proof of Lemma 1 relies on the following basic properties of Poisson processes and
exponential random variables.
Lemma 3: Let X1 and X2 be independent exponential random variables with E[Xi ] = 1/αi .
Let V = X2 − X1 .
(a) P[X1 < X2 ] = α1 /(α1 + α2 ).
(b) Given X1 < X2 , X1 and V are conditionally independent and have conditional exponential probability density functions (PDFs)
fX1 |X1 <X2 (x) = (α1 + α2 )e−(α1 +α2 )x ,
x ≥ 0,
fV |X1 <X2 (v) = α2 e−α2 v ,
v ≥ 0.
Lemma 4: Given a rate λ Poisson process N (t) and an independent exponential (α) random
variable X, the number of arrivals N (X) in the interval [0, X] has the geometric PMF
PN (X) (n) = (1 − γ)γ n ,
with γ = λ/(α + λ).
n ≥ 0,
42
To prove Lemma 1, let Yj , Wj and Tj to refer to the interarrival time, waiting time and system
time of the jth packet of source i. We now examine Wj via the partition
Bj = {Yj < Tj−1 },
Lj = {Tj−1 < Yj }.
(80)
That is, Bj denotes the event that the jth interarrival time for source i is brief, specifically, less
than the system time of the preceding packet from source i, and Lj is the complementary event
that Yj is long. With the partition {Bj , Lj }, we write
E[Yj Wj ] = E[Yj Wj |Lj ] P[Lj ] + E[Yj Wj |Bj ] P[Bj ].
(81)
Since source i packets and packets from other sources have identical exponential (µ) service
times, the combined queue is just M/M/1 with offered load ρ = ρi + ρ−i . In steady state, the
system time Tj−1 has the exponential (µ − λ) PDF
fT (t) = (µ − λ)e−(µ−λ)t ,
t ≥ 0.
(82)
Furthermore, Tj−1 , which depends on packets (and their service times) that arrived prior to
packet j − 1, is independent of Yj . Given Bj , packet j − 1 is still in the system when packet
j is generated. The waiting time Wj depends on both the residual system time Tj−1 − Yj and
also on the workloads of packets from other sources that arrive during the source i interarrival
period of length Yj . Specifically, let M = N−i (Yj ) denote the number of other-source (i.e. not
source i) packets that arrive during the source i interarrival period and let S1 , S2 , . . . SM denote
their service requirements. As these packets are all queued between packets j − 1 and j,
Wj = (Tj−1 − Yj ) +
M
X
Sk .
(83)
E1 = E[Yj (Tj−1 − Yj )|Bj ],
X
M
E 2 = E Yj
Sk |Bj .
(84a)
k=1
This implies E[Yj Wj |Bj ] = E1 + E2 where
k=1
By Lemma 3(b),
E1 = E[(Tj−1 − Yj )|Bj ] E[Yj |Bj ]
(84b)
43
1
1
1
=
= 2
.
µ − λ λi + (µ − λ)
µ (1 − ρ)(1 − ρ−i )
(85)
For the second term, iterated expectation yields
∞
Z
E2 =
X
M
E Yj
Sk |Bj , Yj = y fYj |Bj (y) dy
0
k=1
M
∞ X
E y
Sk |Yj = y fYj |Bj (y) dy.
Z
=
0
(86)
k=1
Given that Yj = y, M = N−i (Yj ) = N−i (y) is the number of other-source arrivals in a period of
length y and is Poisson with conditional expectation E[M |Yj = y] = λ−i y. In addition, each Sk
is an exponential (µ) random variable, independent of Yj , implying E[Sk |Yj = y] = 1/µ. This
implies
X
M
E y
Sk |Yj = y = y E[M |Yj = y] E[Sk |Yj = y] = y(λ−i y)(1/µ) = ρ−i y 2 .
(87)
k=1
By Lemma 3, Yj given Bj is an exponential (α) random variable with α = λi + (µ − λi − λ−i ) =
µ − λ−i . This implies
Z
∞
2ρ−i
2ρ−i
= 2
.
2
α
µ (1 − ρ−i )2
(88)
1
1
2ρ−i
E[Yj Wj |Bj ] = 2
+
.
µ (1 − ρ−i )2 (1 − ρ−i )(1 − ρ)
(89)
E2 = ρ−i
0
y 2 αe−αy dy =
It follows from (85) and (88) that
Given event Lj , packet j − 1 has departed the system prior to the arrival of packet j. In this
case, the waiting time for packet j depends on the number of other-source packets in the system
when packet j arrives. To characterize this, we now let M denote the number of other-source
packets in the system at the departure instant of packet j − 1. Since the queue is FCFS, M is
the number of other-source packets that arrived and were queued during the system time Tj−1
of packet j − 1. Given Tj−1 is exponential and independent of Yj , Lemma 3(b) tells us that Tj−1
given Lj is conditionally an exponential (α) random variable with α = (µ − λ) + λi = µ − λ−i .
As Tj−1 is independent of the subsequent Poisson arrivals of the other sources, Lemma 4 implies
that M has the geometric PMF
PM (m) = (1 − γ)γ m ,
m ≥ 0,
(90)
44
where γ = λ−i /(α + λ−i ) = ρ−i .
From (90), we see at the departure instant of packet j −1 that M is described by the stationary
distribution of an M/M/1 queue serving only other-source packets at rate λ−i . Going forward
from this instant, we wait an additional time Yj − Tj−1 for packet j from source i. In this time
period, there may be either arrivals or departures of other-source packets. Nevertheless, as the
queue holds zero source i packets, the operation of the queue is identical to an M/M/1 queue for
just other-source packets. At all times up to the arrival of packet j, the number of other-source
packets in the queue remains stochastically identical to M . If the kth queued other-source packet
P
has service requirement Sk , then Wj = M
k=1 Sk and E[Wj |Lj ] = E[M ]/µ. It follows that when
packet j does arrive, the number of queued packets M and the service times Sk are independent
of both the additional delay Yj − Tj−1 until the arrival of packet j and the prior system time
Tj−1 . This implies
E[Yj Wj |Lj ] = E[(Tj−1 + (Yj − Tj−1 ))Wj |Lj ]
= E[Tj−1 + (Yj − Tj−1 )|Lj ] E[Wj |Lj ]
1
1
ρ−i
+
.
=
µ − λ−i λi
µ(1 − ρ−i )
(91)
Next we recall from Lemma 3 that independence of Tj−1 and Yj implies P[Bj ] = ρi /(1 − ρ−i ).
Combining this fact with (81), (89), and (91), some algebra yields the claim.
A PPENDIX B
T HEOREM 2( A ): P ROOF C OMPLETION
Note that T = Tj is the time that packet j from source i spends in service. This packet
completes service (is not preempted) and hence no new arrivals occur during the interval SL .
The distribution of T is that of the time to service completion, say Xµ , after packet j arrives,
conditioned on Xµ being smaller than the time to the next packet arrival, say Xλ , from any
source. Thus P [T > z] = P [Xµ > z|Xµ < Xλ ]. By Lemma 3(b), Tj is exponential (λ + µ) and
E[T ] = 1/(λ + µ).
Now we calculate the moments of D = Dj . From (10) we know that D is a random sum of
random variables Bk , 1 ≤ k ≤ L. Also, D ends with the departure of a source i update packet.
Since the arrival rate λ is the sum of rates of independent Poisson sources, the probability that
any block Bk ends in the departure of the update packet of source i is λi /λ. Note that D consists
45
of L = l blocks if l − 1 consecutive blocks end in departures of other-source packets, followed
by block l ending in a source i departure. Thus,
P[L = l] = (1 − q)l−1 q,
l ≥ 1,
(92)
where q = λi /λ. Note that Bk = Xk0 + Sk , where Xk0 is an idle period and Sk is the server busy
period. During the busy period a random number of packets in service may be preempted, but
the service rate (i.e., the instantaneous departure rate) is µ throughout the busy period. Thus the
busy period Sk is memoryless and is independent of the number of arrivals during it that get
preempted and the source whose packet departs at its end. From these observations and given
Poisson arrivals of rate λ, we can write
E[Xk0 ] =
1
1 1
1
, E[Sk ] = , and E[Bk ] = + .
λ
µ
λ µ
(93)
The memoryless nature of the arrival and service processes also implies that each Bk is independent of L. Using this fact and equations (10), (92) and (93), we can write
E[D] = E[L] E[Bk ] =
µ+λ
.
λi µ
(94)
To calculate E[D2 ], let the random variable B be stochastically identical to block lengths Bk ,
k = 1, . . . , L. Using arguments we used to calculate E[D], and noting that Bk and Bk0 are
independent for k 6= k 0 , we can write
E[D2 ] = E[L] E[B 2 ] + E[L(L − 1)](E[B])2 .
(95)
Also note that the idle period Xk0 and busy period Sk that constitute Bk are mutually independent.
This implies
E[B 2 ] =
2
2
2
+ 2+
.
2
λ
µ
λµ
(96)
Using equations (92), (93) and (96), we can write (95) as
λ
E[D2 ] = 2
λi
!
2
λ 1 1
1
+
−
.
λi λ µ
λµ
Applying the moments E[T ], E[D] and E[D2 ] to equation (15) yields Theorem 2(a).
(97)
46
A PPENDIX C
P ROOF OF L EMMA 2
From (23), we calculate Lψq̄ and Lψq̄j for each j and q̄ ∈ Q:
Lψq̄ (q, x) = Λq̄ (q, x),
(98a)
Lψq̄j (q, x) = [bq ]j δq̄,q + Λq̄j (q, x), j ∈ 0 : n,
(98b)
where
Λq̄ (q, x) =
X
Λq̄j (q, x) =
X
[ψq̄ (φl (q, x)) − ψq̄ (q, x)]λ(l) (q),
(99a)
[ψq̄j (φl (q, x)) − ψq̄j (q, x)]λ(l) (q).
(99b)
l∈L
l∈L
When φl (q, x) = (ql0 , xAl ),
ψq̄ (φl (q, x)) = ψq̄ (ql0 , xAl ) = δq̄,ql0 ,
ψq̄j (φl (q, x)) = ψq̄j (ql0 , xAl ) = [xAl ]j δq̄,ql0 , j ∈ 0 : n.
(100a)
(100b)
Since λ(l) (q) = λ(l) δql ,q , it follows from (99) and (100) that
Λq̄ (q, x) =
X
Λq̄j (q, x) =
X
λ(l) δq̄,ql0 − δq̄,q δql ,q ,
(101a)
λ(l) [xAl ]j δq̄,ql0 − xj δq̄,q δql ,q , j ∈ 0 : n.
(101b)
l∈L
l∈L
We observe that
δq̄,ql0 δql ,q =
δq̄,q δql ,q =
δq ,q
l
l ∈ L0q̄ ,
0
δq̄,q
l ∈ Lq̄ ,
0
otherwise.
(102a)
otherwise,
(102b)
It follows from (25), (101) and (102) that
Λq̄ (q, x) =
X
l∈L0q̄
λ(l) δql ,q − δq̄,q
X
l∈Lq̄
λ(l) ,
(103a)
47
Λq̄j (q, x) =
X
λ(l) [xAl ]j δql ,q − xj δq̄,q
l∈L0q̄
X
λ(l) , j ∈ 0 : n.
(103b)
l∈Lq̄
With ψ(q, x) = ψq̄ (q, x), (21a), (24) and (103a) imply
π̇q̄ (t) = E[Λq̄ (q(t), x(t))]
X
X
= E[
λ(l) δql ,q(t) − δq̄,q(t)
λ(l) ]
l∈L0q̄
=
X
l∈Lq̄
λ(l) πql (t) − πq̄ (t)
l∈L0q̄
X
λ(l) ,
q̄ ∈ Q.
(104)
l∈Lq̄
For j ∈ 0 : n with ψ(q, x) = ψq̄j (q, x), (21b), (24) and (103b) imply
v̇q̄j (t) = E[Lψq̄j (q(t), x(t))]
= E[ bq(t) j δq̄,q(t) ] + E[Λq̄j (q(t), x(t))].
(105)
From (103b), we observe that
E[Λq̄j (q(t), x(t))] =
X
X
λ(l) E[x(t)δql ,q(t) ]Al j − E[xj (t)δq̄,q(t) ]
λ(l)
l∈L0q̄
=
X
l∈Lq̄
λ(l) [vql (t)Al ]j − vq̄j (t)
l∈L0q̄
X
λ(l) .
(106)
l∈Lq̄
It then follows from (105) and (106) that for j > 0,
v̇q̄j (t) = [bq̄ ]j πq̄ (t) +
X
λ(l) [vql (t)Al ]j − vq̄j (t)
l∈L0q̄
X
λ(l) .
(107)
l∈Lq̄
The claim for v̇q̄ (t) follows by gathering the equations (107) for j ∈ 0 : n and rewriting as row
vectors.
R EFERENCES
[1] A. Mainwaring, D. Culler, and et al., “Wireless sensor networks for habitat monitoring,” in Proceedings of the 1st ACM
international workshop on Wireless sensor networks and applications, ser. WSNA ’02. Atlanta, Georgia, USA: ACM,
2002, pp. 88–97, ACM ID: 570751.
[2] P. Papadimitratos, A. La Fortelle, K. Evenssen, R. Brignolo, and S. Cosenza, “Vehicular communication systems: Enabling
technologies, applications, and future outlook on intelligent transportation,” IEEE Communications Magazine, vol. 47,
no. 11, pp. 84–95, Nov. 2009.
[3] A. Karakasidis, P. Vassiliadis, and E. Pitoura, “ETL queues for active data warehousing,” in Proceedings of the 2nd
international workshop on Information quality in information systems, ser. IQIS ’05. Baltimore, Maryland: ACM, 2005,
pp. 28–39, ACM ID: 1077509.
[4] H. Yu, L. Breslau, and S. Shenker, “A scalable web cache consistency architecture,” SIGCOMM Comput. Commun. Rev.,
vol. 29, no. 4, pp. 163–174, Aug. 1999, ACM ID: 316219.
[5] M. Xiong and K. Ramamritham, “Deriving deadlines and periods for real-time update transactions,” in The 20th IEEE
Real-Time Systems Symposium, 1999. Proceedings. IEEE, 1999, pp. 32–43.
48
[6] Y. C. Hu and D. B. Johnson, “Ensuring cache freshness in on-demand ad hoc network routing protocols,” in Proceedings
of the second ACM international workshop on Principles of mobile computing, 2002, pp. 25–30.
[7] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed. New York: McGraw-Hill, 1991.
[8] S. Kaul, R. Yates, and M. Gruteser, “Real-time status: How often should one update?” in Proc. IEEE INFOCOM Mini
Conference, 2012.
[9] ——, “Status updates through queues,” in Conf. on Information Sciences and Systems (CISS), Mar. 2012.
[10] R. Yates and S. Kaul, “Real-time status updating: Multiple sources,” in Proc. IEEE Int’l. Symp. Info. Theory, Jul. 2012.
[11] C. Kam, S. Kompella, and A. Ephremides, “Age of information under random updates,” in Proc. IEEE Int’l. Symp. Info. Theory, 2013, pp. 66–70.
[12] ——, “Effect of message transmission diversity on status age,” in Proc. IEEE Int’l. Symp. Info. Theory, June 2014, pp.
2411–2415.
[13] C. Kam, S. Kompella, G. D. Nguyen, and A. Ephremides, “Effect of message transmission path diversity on status age,”
IEEE Trans. Info Theory, vol. 62, no. 3, pp. 1360–1374, March 2016.
[14] M. Costa, M. Codreanu, and A. Ephremides, “Age of information with packet management,” in Proc. IEEE
Int’l. Symp. Info. Theory, June 2014, pp. 1583–1587.
[15] L. Huang and E. Modiano, “Optimizing age-of-information in a multi-class queueing system,” in Proc. IEEE
Int’l. Symp. Info. Theory, Jun. 2015.
[16] M. Costa, M. Codreanu, and A. Ephremides, “On the age of information in status update systems with packet management,”
IEEE Trans. Info Theory, vol. 62, no. 4, pp. 1897–1910, April 2016.
[17] A. M. Bedewy, Y. Sun, and N. B. Shroff, “Optimizing data freshness, throughput, and delay in multi-server informationupdate systems,” in Proc. IEEE Int’l. Symp. Info. Theory, 2016, pp. 2569–2574.
[18] C. Kam, S. Kompella, G. D. Nguyen, J. Wieselthier, and A. Ephremides, “Age of information with a packet deadline,” in
Proc. IEEE Int’l. Symp. Info. Theory, 2016, pp. 2564–2568.
[19] K. Chen and L. Huang, “Age-of-information in the presence of error,” in Proc. IEEE Int’l. Symp. Info. Theory, 2016, pp.
2579–2584.
[20] E. Najm and R. Nasser, “The age of information: The gamma awakening,” in Proc. IEEE Int’l. Symp. Info. Theory, 2016,
pp. 2574–2578.
[21] B. T. Bacinoglu, E. T. Ceran, and E. Uysal-Biyikoglu, “Age of information under energy replenishment constraints,” in
Proc. Info. Theory and Appl. (ITA) Workshop, Feb. 2015, la Jolla, CA.
[22] R. Yates, “Lazy is timely: Status updates by an energy harvesting source,” in Proc. IEEE Int’l. Symp. Info. Theory, 2015.
[23] Y. Sun, E. Uysal-Biyikoglu, R. Yates, C. E. Koksal, and N. B. Shroff, “Update or wait: How to keep your data fresh,” in
IEEE INFOCOM 2016 - The 35th Annual IEEE International Conference on Computer Communications, April 2016, pp.
1–9.
[24] S. Kaul, M. Gruteser, V. Rai, and J. Kenney, “Minimizing age of information in vehicular networks,” in IEEE Conference
on Sensor, Mesh and Ad Hoc Communications and Networks (SECON), Salt Lake City, Utah, USA, 2011.
[25] S. Kaul, R. Yates, and M. Gruteser, “On piggybacking in vehicular networks,” in IEEE Global Telecommunications
Conference, GLOBECOM 2011, Dec. 2011.
[26] B. Kloiber, C. Garcia, J. Härri, and T. Strang, “Update delay: A new information-centric metric for a combined
communication and application level reliability evaluation of cam based safety applications,” in ITS World Congress,
2012.
[27] M. van Eenennaam, L. Hendriks, G. Karagiannis, and G. Heijenk, “Oldest packet drop (opd): A buffering mechanism for
beaconing in ieee 802.11p vanets (poster),” in Vehicular Networking Conference (VNC), 2011 IEEE, nov. 2011, pp. 252
–259.
[28] V. C. Giruka and M. Singhal, “Hello protocols for ad-hoc networks: overhead and accuracy tradeoffs,” in World of Wireless
Mobile and Multimedia Networks, WoWMoM, Sixth IEEE International Symposium on. IEEE, Jun. 2005, pp. 354–361.
[29] T. He, Krishnamurthy, and et al., “Energy-efficient surveillance system using wireless sensor networks,” in Proceedings of
the 2nd international conference on Mobile systems, applications, and services, ser. MobiSys ’04. Boston, MA, USA:
ACM, 2004, pp. 270–283, ACM ID: 990096.
[30] C. Schurgers, V. Tsiatsis, S. Ganeriwal, and M. Srivastava, “Optimizing sensor networks in the energy-latency-density
design space,” IEEE Transactions on Mobile Computing, vol. 1, no. 1, pp. 70–80, Mar. 2002.
[31] J. Hespanha, “Modelling and analysis of stochastic hybrid systems,” IEE Proceedings-Control Theory and Applications,
vol. 153, no. 5, pp. 520–535, 2006.
[32] D. Vermes, “Optimal dynamic control of a useful class of randomly jumping processes,” International Institute for Applied
Systems Analysis, Tech. Rep. PP-80-015, 1980.
[33] B. Gnedenko and I. Kovalenko, “Introduction to the theory of mass service,” Nau (trad, ingles: 1968, Jerusalem), 1966.
[34] M. H. A. Davis, “Piecewise-deterministic Markov processes: a general class of nondiffusion stochastic models,” J. Roy.
Statist. Soc., vol. 46, pp. 353–388, 1984.
[35] L. DeVille, S. Dhople, A. D. Domı́nguez-Garcı́a, and J. Zhang, “Moment closure and finite-time blowup for piecewise
deterministic markov processes,” SIAM Journal on Applied Dynamical Systems, vol. 15, no. 1, pp. 526–556, 2016.
[36] A. R. Teel, A. Subbaraman, and A. Sferlazza, “Stability analysis for stochastic hybrid systems: A survey,” Automatica,
vol. 50, no. 10, pp. 2435–2456, 2014.
[37] E. Seneta, Non-negative Matrices and Markov Chains, 2nd ed. Springer Verlag, 1981.
49
[38] A. Munir, A. Gordon-Ross, and S. Ranka, Modeling and optimization of parallel and distributed embedded systems. John
Wiley & Sons, 2015.
[39] C. Saraydar, N. B. Mandayam, and D. Goodman, “Efficient power control via pricing in wireless data networks,”
Communications, IEEE Transactions on, vol. 50, no. 2, pp. 291–303, 2002.
[40] G. Bolch, S. Greiner, H. de Meer, and K. S. Trivedi, Queueing networks and Markov chains: modeling and performance
evaluation with computer science applications. John Wiley & Sons, 2006.
| 7cs.IT
|
Randomized algorithms for finding a majority element
Pawel Gawrychowski1 , Jukka Suomela2 , and Przemyslaw Uznański3
arXiv:1603.01583v3 [cs.DS] 28 Apr 2016
1
Institute of Informatics, University of Warsaw, Poland
2
Helsinki Institute for Information Technology HIIT,
Department of Computer Science, Aalto University, Finland
3
Department of Computer Science, ETH Zürich, Switzerland
Abstract
Given n colored balls, we want to detect if more than bn/2c of them have the same color, and if so find
one ball with such majority color. We are only allowed to choose two balls and compare their colors, and
the goal is to minimize the total number of such operations. A well-known exercise is to show how to find
such a ball with only 2n comparisons while using only a logarithmic number of bits for bookkeeping. The
resulting algorithm is called the Boyer–Moore majority vote algorithm. It is known that any deterministic
method needs d3n/2e − 2 comparisons in the worst case, and this is tight. However, it is not clear what is
the required number of comparisons if we allow randomization. We construct a randomized algorithm
which always correctly finds a ball of the majority color (or detects that there is none) using, with high
probability, only 7n/6 + o(n) comparisons. We also prove that the expected number of comparisons used
by any such randomized method is at least 1.019n.
1
Introduction
A classic exercise in undergraduate algorithms courses is to construct a linear-time constant-space algorithm
for finding the majority in a sequence of n numbers a1 , a2 , . . . , an , that is, a number x such that more than
bn/2c numbers ai are equal to x, or detect that there is no such x. The solution is to sweep the sequence
from left to right while maintaining a candidate and a counter. Whenever the next number is the same as
the candidate, we increase the counter; otherwise we decrease the counter and, if it drops down to zero, set
the candidate to be the next number. It is not difficult to see that if the majority exists, then it is equal to
the candidate after the whole sweep, therefore we only need to count how many times the candidate occurs
in the sequence. This simple yet beautiful solution was first discovered by Boyer and Moore in 1980; see [4]
for the history of the problem.
The only operation on the input numbers used by the Boyer–Moore algorithm is testing two numbers for
equality, and furthermore at most 2n such checks are ever being made. This suggests that the natural way
to think about the algorithm is that the input consists of n colored balls and the only possible operation is
comparing the colors of any two balls. Now the obvious question is how many such comparisons are necessary
and sufficient in the worst possible case. Fischer and Salzberg [12] proved that the answer is d3n/2e − 2.
Their algorithm is a clever modification of the original Boyer–Moore algorithm that reuses the results of some
previously made comparisons during the verification phase. They also show that no better solution exists by
an adversary-based argument. However, this argument assumes that the strategy is deterministic, so the next
step is to allow randomization.
Surprisingly, not much seems to be known about randomized algorithms for computing the majority in
the general case. For the special case of only two colors, Christofides [5] gives a randomized algorithm that
uses 23 (1 − 3 )n comparisons in expectation and returns the correct answer with probability 1 − , and he
also proves that this is essentially tight; this improves on a previous lower bound of Ω(n) by De Marco and
Pelc [16]. Note that in the two-color case any deterministic algorithm needs precisely n − B(n) comparisons,
1
where B(n) is the number of 1s in the binary expansion of n, and this is tight [2, 18, 20]. For a random input,
with each ball declared to be red or blue uniformly at random, roughly 2n/3 comparisons are sufficient and
necessary in expectation to find the majority color [3]. To the best of our knowledge upper and lower bounds
on the expected number of comparisons with unbounded number of colors have not been studied before.
Related work include oblivious algorithms studied by Chung et al. [6], that is, algorithms in which
subsequent comparisons do not depend on the previous answers, and finding majority with larger queries [10,
15, 19]. Another generalization is finding a ball of plurality color, that is, the color that occurs more often
than any other [1, 13, 14].
We consider minimizing the number of comparisons mostly as an academic exercise, and believe that a
problem with such a simple formulation deserves to be thoroughly studied. However, it is possible that a
single comparison is so expensive that their number is the bottleneck. Such a line of thought motivated a
large body of work studying the related questions of the smallest number of comparisons required to find the
median element; see [8, 9, 17] and the references therein. Of course, the simplest Boyer–Moore algorithm has
the advantage of using only two sequential scans over the input and a logarithmic number of bits, while our
algorithm needs more space and random access to the input.
Given that the original motivation of Boyer and Moore was fault-tolerant computing, we find it natural to
consider Las Vegas algorithms, that is, the number of comparisons depends on the random choices of the
algorithm but the answer is always correct. This way the result is correct even if the source of random bits is
compromised; an adversary controlling the random number generator can only influence the running time.
Model. We identify balls with numbers 1, 2, . . . , n. We write cmp(i, j) for the result of comparing the
colors of balls i and j (true for equality, false for inequality). We consider randomized algorithm that, after
performing a number of such comparisons, either finds a ball of the majority color or detects that there is
no such color. A majority color is a color with the property that more than bn/2c balls are of such color.
The algorithm should always be correct, irrespectively of the random choices made during the execution.
However, the colors of the balls are assumed to be fixed in advance, and therefore the number of comparisons
is a random variable. We are interested in minimizing its expectation.
Contributions. We construct a randomized algorithm, which always correctly determines a ball of the
majority color or detects that there is none, using 7n/6+o(n) comparisons with high probability (in particular,
in expectation). We also show that the expected number of comparisons used by any such algorithm must be
at least 1.019n.
2
Preliminaries
We denote the set of balls (items) by M = {1, 2, . . . , n}. We write color(x) for the color of ball x, and
cmp(x, y) returns true if the colors of balls x and y are identical.
An event occurs with very high probability (w.v.h.p.) if it happens with probability at least 1 −
exp(−Ω(log2 n)). Observe that the intersection of polynomially many very high probability events also
happens with very high probability.
Lemma 2.1 (symmetric
Chernoff bound). The number of successes for n independent coin flips is w.v.h.p.
√
at most n2 + O( n log n).
Lemma 2.2 (sampling). Let X ⊆ M such that |X| = m. Let m0 denote the number of hits on elements
from X if we sample uniformly at random k ≤ n elements from M without replacement. Then w.v.h.p.
|m0 /k − m/n| = O(k −1/2 log n).
Proof. Let δ = Θ(k −1/2 log n). From the tail bound for hypergeometric distribution (see Chvátal [7]), we
have
m0
m
Pr
−
≥ δ ≤ 2 exp(−2δ 2 k) = exp(−Θ(log2 n)).
k
n
2
Worth noting is that the error bound here would be essentially the same if we replace “without replacement”
with “with replacement”. In that case, we would be invoking a tail bound for binomial distribution.
Now we consider a process of pairing the items without replacement (choosing a random perfect matching
on M ; if n is odd then one item remains unpaired). For any X ⊆ M , let uXX be a random variable counting
the pairs with both elements belonging to X when choosing uniformly at random n2 pairs of elements from
M without replacement. Of course E[uXX ] = |X|(|X|−1)
2(n−1) .
Lemma 2.3 (concentration for pairs). For any X ⊆ M w.v.h.p.
p
|X|2
= O |X| log n .
2n
p
Proof. For simplicity, we can assume that n is even. Set ∆ = Θ( |X| log n). Instead of choosing a random
perfect matching on M , we consider an equivalent two-step random process:
uXX −
1. We choose uniformly at random a partition of M into M1 , M2 such that |M1 | = |M2 | =
partition induces a partition of X into X1 , X2 , such that X1 = M1 ∩ X and X2 = M2 ∩ X.
Observe that, by Lemma 2.2, |X1 |, |X2 | ∈ 21 |X| − ∆, 12 |X| + ∆ w.v.h.p.
n
2.
This
2. Now, instead of pairing uniformly at random the elements from M1 with the elements from M2 , for our
purposes it is enough to choose uniformly at random a set M20 ⊆ M2 of elements that are paired with
X1 , and count how many elements of X2 we have chosen. Thus uXX = |M20 ∩ X2 |.
By Lemma 2.2 we have
|M20 ∩ X2 | |M20 |
−
= O(|X2 |−1/2 log |M2 |) w.v.h.p.,
|X2 |
|M2 |
so the following holds (since |M20 | = |X1 |) w.v.h.p.:
uXX −
Hence
uXX −
p
|X1 ||X2 |
= O( |X2 | log |M2 |).
n/2
2
p
|X|2
∆
= O(∆).
= O( |X2 | log |M2 |) + O
2n
n/2
Lemma 2.4 (pairs in partition). Let F = {X1 , . . . , Xm } be a partition of M . Then w.v.h.p.
X
uXX −
X∈F
X |X|2
= O(n2/3 log n).
2n
X∈F
Proof. Let F 0 be a partition {X10 , . . . , X 0 m0 } obtained from F by merging all sets Xi such that |Xi | < n2/3
into larger sets of size between n2/3 and 2n2/3 (this can be done in a greedy fashion). Let A be the family of
original large sets, B the family of original small sets, and C the family of new larger sets.
Since B is a finer partition than C, we have that
X
uXX ≤
X∈B
X
X |X|2
X |X|2
≤
.
2n
2n
uXX ,
X∈C
X∈B
X∈C
Since all sets in C are smaller than 2n2/3 , we obtain
X
X |X|2
≤
|X| n−1/3 ≤ n2/3 .
2n
X∈C
X∈C
3
Then, because all sets in C are large, by a a direct application of Lemma 2.3, w.v.h.p.
X
X |X|2
Xp
p
|X| ≤ O(n2/3 log n).
uXX ≤
+ O( |X| log n) ≤ n2/3 + O(log n)
2n
X∈C
X∈C
X∈C
Similarly, because all sets in F 0 are large, by an application of Lemma 2.3, w.v.h.p.
X
uXX −
X∈F 0
X
X |X|2
p
≤
O |X| log n = O(n2/3 log n).
2n
0
0
X∈F
X∈F
Now our goal is to bound
X
uXX −
X∈F
X |X|2
2n
X∈F
which, because |a + b| ≤ |a| + |b|, is at most
X
uXX −
X∈B
X
X∈C
uXX +
X
X∈F 0
uXX −
X |X|2
X |X|2
X |X|2
+
−
.
2n
2n
2n
0
X∈F
X∈B
X∈C
The first and the third addend can be bounded by O(n2/3 log n) because if 0 ≤ a ≤ b then |b − a| ≤ b, hence
by plugging in the previously derived bounds we obtain
X
uXX −
X∈F
X |X|2
≤ O(n2/3 log n) + O(n2/3 log n) + O(n2/3 ) = O(n2/3 log n).
2n
X∈F
Lemma 2.5. Let X ⊆ M such that |X| = m. Let k(m0 , m, n) denote the number of draws without replacement
until we hit m0 elements from X. Then w.v.h.p.
n
n
k(m0 , m, n) ≤ m0 + O √ · log n .
m
m
Proof. Denoting by S the number hits on X when drawing without replacement k items, then by Lemma 2.2
w.v.h.p.
m
S ≥ k − C · m1/2 log n
n
n
for some constant C > 0. Substituting k = m m0 + C · √nm log n we obtain
S ≥ m0 + C · m1/2 log n − C · m1/2 log n = m0
meaning that indeed w.v.h.p. after k draws we will have at least m0 elements from X.
3
Algorithm
In this section we describe a randomized algorithm for finding majority. Recall that the algorithm is required
to always either correctly determine a ball of the majority color or decide that there is no such color, and the
majority color is a color of more than bn/2c balls. For simplicity we will assume for the time being that n is
even, as the algorithm can be adjusted for odd n in a straightforward manner without any change to the
asymptotic cost. Hence to prove that there is a majority color, it is sufficient to find n/2 + 1 balls with the
same color. In such case our algorithm will actually calculate the multiplicity of the majority color. To prove
that there is no majority color, it is sufficient to partition the input into n/2 pairs of balls with different
colors.
4
The algorithm consists of three parts. Intuitively, by choosing a small random sample we can approximate
the color frequencies and choose the right strategy:
(i) There is one color with a large frequency. We use algorithm heavy. In essence, we have only one
candidate for the majority, and we compute the frequency of the candidate in a naive manner. If the
frequency is too small, we need to form sufficiently many pairs of balls with different colors among the
balls that are not of the candidate color. This can be done by virtually pairing the non-candidate color
elements, and testing these pairs until we find enough of them that have distinct colors. Additionally,
we show that one sweep through the pairs is enough.
(ii) There are two colors with frequencies close to 0.5. Now we use algorithm balanced. In essence, we
can now reduce the size of the input by a pairing process, and then find the majority recursively. If the
recursion finds the majority, the necessary verification step is speeded up by reusing the results of the
comparisons used to form the pairs.
(iii) All frequencies are small. We use light which, as balanced, applies pairing and recursion. However,
if the recursive call reports the majority, we construct enough pairs with different colors: whenever we
find a pair of elements with both colors different than the majority color found by the recursive call, we
pair them with elements of the majority color. Here we speeded up the process by reusing the results of
the comparisons used to form the pairs as well.
We start with presenting the main procedure of the algorithm; see Algorithm 1. The parameters are
chosen by setting α = 13 , ε = n−1/10 and β = 0.45. In fact we could chose any β ∈ (β1 , β2 ), where
β1 = 1 − √13 ≈ 0.4226 and β2 ≈ 0.47580 is a root to p3 − 19p2 − 8p + 8 = 0.
Algorithm 1: majority(M )
1
2
3
4
5
6
7
8
9
if |M | = 1 then return M [1] is the majority with multiplicity 1 in M sample M 0 ⊆ M such that
|M 0 | = nα
let v1 , v2 , . . . , vk be the representatives of the colors in M 0
let qi |M 0 | be the frequency of color(vi ) in M 0 , where q1 ≥ q2 ≥ . . . ≥ qk
if q1 , q2 ∈ [ 21 − 4ε, 12 + 4ε] then
return balanced(M )
else if q1 ≥ β and q12 ≥ q22 + . . . + qk2 + 2ε then
return heavy(M, v1 )
else
return light(M )
Before we proceed to describe the subprocedures, we elaborate on the sampling performed in line 3.
Intuitively, we would like to compute the frequencies of all colors in M . This would be too expensive, so
we select a small sample M 0 and claim that the frequencies of all colors in M 0 are not too far from the
frequencies of all colors in M . Formally, let p1 , p2 , p3 , . . . , p` be the frequencies of all colors in M , that is
0
there are pi · n balls of color i in M and let qi be the frequency
Lemma 2.2,
P 2 of color i in the sample M
P . By
−α/2
w.v.h.p. |pi − qi | = O(n
log n) = o(ε). We argue that i qi is a good estimation of i p2i .
Lemma 3.1. Let pi be the frequency of color i in M and qi be its frequency in M 0 , where M 0 ⊆ M a random
sample without replacement of size nα . Then w.v.h.p.
X
i
p2i −
X
qi2 = O(n−α/3 log n) = o(ε).
i
Proof. Let m = nα . We analyze the following two sampling methods.
5
1. Partition the elements of M into n2 disjoint pairs uniformly at random. Select m
2 of these pairs uniformly
at random. Denote by A1 and A2 the pairs with both elements of
the
same
colors
in the first and the
P
second pairing, respectively. By Lemma 2.4, w.v.h.p. |A1 | − n2 i p2i = O(n2/3 log n). Observe that
1/2
by Lemma 2.2 w.v.h.p. |A2 | − m
log n). Thus, by the triangle inequality, w.v.h.p.
n |A1 | = O(m
|A2 | X 2
−
pi = O(n−1/3 log n) + O(m−1/2 log n).
m/2
i
2. Partition the elements of M 0 into m
and denote by B all pairs
2 disjoint pairs uniformly at random, P
2
2/3
with both elements of the same color. By Lemma 2.4, w.v.h.p. |B| − m
log n), or
i qi = O(m
2
P 2
|B|
−1/3
log n).
equivalently m/2 − i qi = O(m
Now, because A2 and B have identical distributions, by the triangle inequality we have
X
i
p2i −
X
qi2 = O(n−1/3 log n) + O(m−1/2 log n) + O(m−1/3 log n) = O(m−1/3 log n).
i
Now we present the subprocedures; see Algorithms 2–4.
Lemma 3.2. Algorithm 1 always returns the correct answer.
Proof. We analyze separately every subprocedure.
balanced(M ). If the majority exists then removing two elements with different colors preserves it.
Hence if the recursive call returns that there is no majority in X then indeed there is no majority in M ,
and otherwise color(v) is the only possible candidate for the majority in M . The remaining part of the
subprocedure simply verifies it.
heavy(M, v). The subprocedure first checks if color(v) is the majority. Hence it is enough to analyze
what happens if color(v) is not the majority. Then X contains all elements with other colors. We partition
the elements in X into pairs and check which of these pairs consists of elements with different colors. If the
number of elements in all the remaining pairs is smaller than the number of elements of color color(v), then
clearly we can partition all elements in M into disjoint pairs of elements with different colors, hence indeed
there is no majority. Otherwise, we revert to the simple 2n algorithm, which is always correct.
light(M ). Again, if the majority exists then removing two elements with different color preserves it.
Hence we can assume that color(v) is the only possible candidate for the majority. Then, Y consists of pairs
of two elements with different colors. From the recursive call we also know what is the frequency of color(v)
in M \ Y . We iterate through the elements of Y and check if their color is color(v). However, if the color of
the first element in a pair is color(v), then the second element has a different color. So the subprocedure
either correctly determines the frequency of the majority color(v), or find sufficiently many elements with
different colors to conclude that color(v) is not the majority.
Theorem 3.3. Algorithm 1 w.v.h.p. uses at most 76 n + o(n) comparisons on an input of size n. The expected
number of comparisons is also at most 76 n + o(n).
Proof. Let T (n) be a random variable counting the comparisons on the given input of size n. We will
inductively prove that T (n) ≤ 76 n + C · n9/10 for a fixed constant C that is sufficiently large. In the analysis
we will repeatedly invoke Lemmas 2.2, 2.3, 2.4, 2.5, 3.1 and Chernoff bound to bound different quantities. We
will assume that each such the application succeeds. Since there will be a polynomial number of applications,
each on a polynomial number of elements, this happens w.v.h.p. with respect to the size of the input. We
also assume that n is large enough. Algorithm 1 uses at most O(n2α ) = O(n2/3 ) comparisons in the sampling
stage. We bound the number of subsequent comparisons used by each subprocedure as follows.
6
Algorithm 2: balanced(M )
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
randomly shuffle M
X ← [], Y ← []
for i = 1 to |M |/2 do
if cmp(M [2i − 1], M [2i]) then
append M [2i] to X
else
append M [2i − 1] and M [2i] to Y
run majority(X)
if there is no majority in X then return no majority in M
let color(v) be the majority with multiplicity k in X
cnt ← 2k
for i = 1 to |Y |/2 do
if cmp(v, Y [2i − 1]) then
cnt ← cnt + 1
else if cmp(v, Y [2i]) then
cnt ← cnt + 1
if cnt ≤ |M |/2 then
return no majority in M
else
return color(v) is the majority with multiplicity k in X
Algorithm 3: heavy(M, v)
1
2
3
4
5
6
7
8
9
10
11
12
cnt ← 0, X ← []
for i = 1 to |M | do
if cmp(v, M [i]) then
cnt ← cnt + 1
else
append M [i] to X
if cnt > |M |/2 then return color(v) is the majority with multiplicity k in M
k ← |M |/2 − cnt
randomly shuffle X
for i = 1 to |X|/2 do
if ¬cmp(X[2i − 1], X[2i]) then k ← k − 1 if k = 0 then return no majority in M
return Boyer–Moore(M)
. fallback, 2n comparisons
7
Algorithm 4: light(M )
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
randomly shuffle M
X ← [], Y ← []
for i = 1 to |M |/2 do
if cmp(M [2i − 1], M [2i]) then
append M [2i] to X
else
append M [2i − 1] and M [2i] to Y
run majority(X)
if there is no majority in X then return no majority in M
let color(v) be the majority with multiplicity k in X
cnt ← 2k − |X|
for i = 1 to |Y | do
if ¬cmp(v, Y [2i − 1]) then
if ¬cmp(v, Y [2i]) then
cnt ← cnt − 1
if cnt = 0 then return no majority in M
return color(v) is the majority with multiplicity (|M |/2 + cnt) in M
P 2
1
balanced(M ). We have that p1 , p2 = 12 ± O(ε). Thus also
i pi = 2 ± O(ε). By Lemma 2.4,
P
1
n
2
2/3
|X| = ( 2 i pi ) ± O(n log n), thus |X| = ( 4 ± O(ε))n. Also |Y | = n − 2|X| = ( 21 ± O(ε))n.
List Y consists of pairs of elements with different colors. Because at most O(εn) of all elements are not
of color 1 or 2, there are at most O(εn) pairs not of the form {1, 2}. Since the relative order of elements
Y [2i − 1] and Y [2i] is random, for each pair {1, 2} we pay 1 with probability 1/2 and pay 2 with probability
1/2, and for any other pair we pay always 2. Thus the total cost incurred by the loop in line 12 is (by Chernoff
bound) at most
p
3
3
O(εn) · 2 + |Y |/2 + O( |Y | log n) ≤ n ± O(εn).
2
8
Thus the total cost is
9/10
T (n) ≤ T ( 41 + ε)n + 12 n + 38 n + O(εn) ≤ 76 n + O(n9/10 ) + C · 31 n
and 76 n + O(n9/10 ) + C · ( 13 )9/10 · n9/10 ≤ 76 n + C · n9/10 for a large enough C.
heavy(M, v). If p1 > 12 , then we terminate in line 7 afterPn comparisons. Thus we can assume that
p1 ∈ [0.45 − ε, 12 ]. Because P
by Lemmas 2.2 and 3.1 both p21 and i p2i are estimated within an absolute error
2
of o(ε), we have that p1 − i≥2 p2i ≥ 2ε − 2o(ε) ≥ ε.
We argue that the loop in line 10 will eventually find sufficiently many pairs of elements with different
colors, and thus return without falling back to the 2n algorithm. By definition, |X| = (1 − p1 )n and initially
k = (1/2 − p1 )n. By Lemma 2.4, after the random shuffle the number D of pairs of elements (X[2i − 1], X[2i])
with different colors can be bounded by
P
2
|X|
j≥2 (pj n)
D≥
−
− O |X|2/3 log |X| ≥
2
2|X|
1 − p1
p21 − ε
≥
n−
n − o(εn) ≥
2
2(1 − p1 )
1 − 2p1
ε
≥
n + n − o(εn) ≥
2(1 − p1 )
2
≥ 21 − p1 n,
8
thus indeed there are sufficiently many pairs. Hence, because the pairs are being considered in a random
order, the total cost can be bounded using Lemma 2.5 by
|X| 1
|X|
T (n) ≤ n +
− p1 n + O √ log |X| ≤
D 2
D
r
2
(1 − p1 )
ε
n + O n/
n log n ≤
≤n+
2
3
r
n
≤ (1 + 0.552 /2)n + O(εn) + O
log n =
ε
= 1.15125n + O(n9/10 ),
where we used D ≥
ε
2
− o(n) ≥ 3ε n for a large enough n.
P
light(M ). We start with bounding |X| and |Y |. By Lemma 2.4, |X| = n2 i p2i ± O(n2/3 log n), and by
Lemma 2.3 there are n2 p21 ± O(n1/2 log n) elements from A1 in X, thus there are n(p1 − p21 ) ± O(n1/2 log n)
of elements from A1 in Y (each paired with a non-A
P 1 element).
We know that either p1 ≤ 0.45 + ε or p21 − i≥2 (p2i ) ≤ ε. If there is no majority in X, then p1 ≤ 12 and
the total cost is bounded by
T (n) ≤
n
n 7
+ T (|X|) ≤ + |X| + C · |X|9/10 ,
2
2
6
which, because |X| ≤ n4 + O(n2/3 log n), is less than 19
24 n + o(n). Hence we can assume that there is a majority
in X. In such case, cnt is set to
n 2 X 2
c=
p1 −
pi ± O(n2/3 log n).
2
i≥2
We denote by I the total number of iterations of the loop in line 12. By Lemma 2.5
I≤
1
2 |Y
1
2 |Y
|
· c + O(E),
| − |A1 ∩ Y |
q
P
where E = |Y |/ 12 |Y | − |A1 ∩ Y |. Substituting S = i≥2 p2i , by Lemma 2.4 we have
|Y | = 1 − p21 − S ± O(n−1/3 log n) n,
|Y | − 2|A1 ∩ Y | = (1 − p1 )2 − S ± O(n−1/3 log n) n,
1
c = p21 − S ± O(n−1/3 log n) n.
2
Since p1 ≤ 12 and p2 ≤ 12 − 3ε (as for a larger p2 the sampled q2 would be sufficiently large for other
subprocedure to be used), we have
2
2
(1 − p1 )2 − S − o(ε) ≥ 21 − 21 − 3ε − (3ε)2 − o(ε) = 3ε − 18ε2 − o(ε) ≥ 2ε.
pn
Thus E ≤ 2ε
. Now, since |Y | = Θ(n) we can bound I from above by
|Y |
1
· (p21 − S)n + O(1/ε) · O(n2/3 log n) + O(E) ≤
|Y | − 2|A1 ∩ Y | 2
2/3
r
1 1 − p21 − S + O(n−1/3 log n)
n log n
n
2
≤
(p
−
S)n
+
O
+
O
≤
2 (1 − p1 )2 − S − O(n−1/3 log n) 1
ε
4ε
I≤
≤
1
1 − p21 − S
O(n2/3 log n)
2
(p
−
S)n
+
+ O(n23/30 log n),
1
2 (1 − p1 )2 − S − O(n−1/3 log n)
2ε
9
which, because (1 − p1 )2 − S is sufficiently large, can be bounded by
1 1 − p21 − S
O(n−1/3 log n)
2
(p
−
S)n
·
1
+
+ O(n23/30 log n) ≤
2 (1 − p1 )2 − S 1
(1 − p1 )2 − S
1 1 − p21 − S
O(n−1/3 log n)
2
≤
(p − S)n · 1 +
+ O(n23/30 log n) ≤
2 (1 − p1 )2 − S 1
2ε
I≤
O(n2/3 log n)
1 1 − p21 − S
2
+ O(n23/30 log n) ≤
(p
−
S)n
+
1
2 (1 − p1 )2 − S
8ε2
1
p21 − S
≤ (1 − p21 − S)
n + O(n26/30 log n).
2
(1 − p1 )2 − S
≤
For each of c iterations we pay 2, and for each of the remaining I −c iterations we pay only
(for each iteration independently). Thus, by Chernoff bound the total cost is
T (n) ≤
3
2
in expectation
√
3
1
n + T (|X|) + (I − c) + O( I − c log(I − c)) + 2c ≤
2
2
1
7
3
p21 − S
1
n + n(p21 + S) + (1 − p21 − S)
n + (p21 − S)n + O(n26/30 log n) + C · |X|9/10 =
2
12
4
(1 − p1 )2 − S
4
n
3
1 − p21 − S − ((1 − p1 )2 − S) 3 2
1 2
7
=
+
(p
−
S)
+
(p
−
S)
+ O(n9/10 ) =
1 + (p21 + S) + (p21 − S)
2
6
2
(1 − p1 )2 − S
2 1
2 1
19 2 5
p1 − p21
n
2
1 + p1 − S + 3(p1 − S)
+ O(n9/10 ).
=
2
6
6
(1 − p1 )2 − S
≤
We reason that, for a fixed p1 , the quantity
1+
19 2 5
p1 (1 − p1 )
p − S + 3(p21 − S)
6 1 6
(1 − p1 )2 − S
is a decreasing function of S, since p21 ≤ (1 − p1 )2 . Now we consider two cases. If p21 − S ≤ ε then simplifying
p2 −S
above estimation either with p21 − S ≤ 0, or, since (1 − p1 )2 − S ≥ 2ε, with 0 ≤ (1−p11 )2 −S ≤ 12 , we get either
n
T (n) ≤
2
7 2
1 + p1 + O(n9/10 ),
3
which, since p1 ≤ 12 , is bounded from above by 19
24 n + o(n),
7
3
n
1 + p21 + p1 −
T (n) ≤
2
3
2
or
3 2
p
2 1
+ O(n9/10 ),
which is bounded from above by 47
48 n + o(n).
Otherwise, p1 ≤ 0.45 + ε and substituting S = 0 (since the cost is decreasing in S) we obtain
n
19
p3
T (n) ≤
1 + p21 + 3 1
+ O(n9/10 ) = 1.06915n + O(n9/10 ).
2
6
1 − p1
Wrapping up. We see that in each subprocedure, the number of comparisons is bounded by 76 n+C·n9/10 .
Each subprocedure makes at most one recursive call, where the size of the input is reduced by at least a
factor of 2. Thus the worst-case number of comparison is always bounded by O(n). Recall that the bound on
the number of comparisons used by every recursive call holds w.v.h.p. with respected to the size of the input
10
to the call. Eventually, the size of the input might become very small, and then w.v.h.p. with respect to the
size of the input is no longer w.v.h.p. with respect to the original n. However, as soon as this size decreases
to, say, n0.1 , the number of comparisons is O(n) irrespectively of the random choices made by the algorithm.
Thus w.v.h.p. the number of comparisons is at most 76 n + O(n9/10 ), and the expected number of comparisons
is also bounded by 76 n + O(n9/10 ).
4
Lower bound
We consider Las Vegas algorithms. That is, the algorithm must always correctly determine whether a majority
element exists. We will prove that the expected number of comparisons used by such an algorithm is at least
c · n − o(n), for some constant c > 1. By Yao’s principle, it is sufficient to construct a distribution on the
inputs, such that the expected number of comparisons used by any deterministic algorithms run on an input
chosen from the distribution is at least c · n − o(n). The distribution is that with probability n1 every ball
has a color chosen uniformly at random from a set of n colors. With probability 1 − n1 every ball is black
or white, with both possibilities equally probable. We fix a correct deterministic algorithm A and analyze
its behavior on an input chosen from the distribution. As a warm-up, we first prove that A needs n − o(n)
comparisons in expectation on such input.
4.1
A lower bound of n − o(n)
In every step A compares two balls, thus we can describe its current knowledge by defining an appropriate
graph as follows. Every node corresponds to a ball. Two nodes are connected with a negative edge if the
corresponding balls have been compared and found out to have different colors. Two nodes are connected
with a positive edge if the corresponding balls are known to have the same colors under the assumption that
every ball is either black or white (either because they have been directly compared and found to have the
same color, or because such knowledge has been indirectly inferred from the assumption). After every step of
the algorithm the graph consists of a number of components C1 , C2 , . . .. Every components is partitioned into
two parts Ci = Ai ∪· Bi , such that both Ai and Bi are connected components in the graph containing only
positive edges and there is at least one (possibly more than one) negative edge between Ai and Bi . There are
no other edges in the graph. Now we describe how the graph changes after A compares two balls x ∈ Ci and
y ∈ Cj assuming that every ball is either black or white. If i = j then the result of the comparison is already
determined by the previous comparisons and the graph does not change. Otherwise, i 6= j and assume by
symmetry that x ∈ Ai , y ∈ Aj . The following two possibilities are equally probable:
1. color(x) = color(y), then we merge both components into a new component C = A ∪· B, where
A = Ai ∪· Aj and B = Bi ∪· Bj by creating new positive edges (x, y) and (x0 , y 0 ) for some x0 ∈ Bi , y 0 ∈ Bj
(if Bi , Bj 6= ∅).
2. color(x) 6= color(y), then we merge both components into a new component C = A ∪· B, where
A = Ai ∪· Bj and B = Bi ∪· Aj by creating new positive edges (x, y 0 ) for some y 0 ∈ Bj (if Bj 6= ∅) and
(x0 , y) for some x0 ∈ Bi (if Bi =
6 ∅). We also create a new negative edge (x, y). Here we crucially use
the assumption that every ball is either black or white.
The graph exactly captures the knowledge of A about a binary input.
Any binary input contains a majority and A must report so. However, because with very small probability
the input is arbitrary, this requires some work due to the following lemma.
Lemma 4.1. If A reports that a binary input contains a majority element, then the graph contains a
component C = A ∪· B such that |A| > n2 or |B| > n2 .
Proof. Assume otherwise, that is, A reports that a binary input contains a majority element even though
both parts of every component are of size less than n2 . Construct another input by choosing, for every
component C = A ∪· B, two fresh colors cA and cB and setting color(x) = cA for every x ∈ A, color(y) = cB
11
for every y ∈ B. Every comparison performed by A is an edge of the graph, so its behavior on the new input
is exactly the same as on the original binary input. Hence A reports that there is a majority element, while
the frequency of every color in the new input is less than n2 , which is a contradiction.
From now on we consider only binary inputs. If we can prove that the expected number of comparisons
used by A on such input is n − o(n), then the expected number of comparisons on an input chosen from
our distribution is also n − o(n). Because every comparison decreases the number of components by one, it
is sufficient to argue that the expected size of some component when A reports that there is a majority is
n − o(n). We already know that there must exist a component C = A ∪· B such that (by symmetry) |A| > n/2.
·
We will argue that |B| must also be large. To this end,
P define balance of a component Ci = Ai ∪ Bi as
balance(Ci ) = (|Ai | − |Bi |)2 , and the total balance as i balance(Ci ). By considering the situation before
and after a single comparison, we obtain the following.
Lemma 4.2. The expected total balance at termination of algorithm A is n.
Proof. In the very beginning the total balance is n because every component is a singleton. Recall that when A
compares two balls x ∈ Ci and y ∈ Cj then Ci and Cj are merged into a new component C. It is easy to verify
that if balance(Ci ) = b2i and balance(Cj ) = b2j then either balance(C) = (bi + bj )2 or balance(C) = (bi − bj )2 ,
with both possibilities equally probable. Hence the expected total balance after the comparison is equal to
the previous total balance increased by −b2i − b2j + 12 ((bi − bj )2 + (bi + bj )2 ) = 0. Consequently, total balance
is a martingale and is preserved in expectation with respect to arbitrarily branching computation.
Total balance when A reports a majority is a random variable with expected value n. By Markov’s
inequality, with probability 1 − 1/n1/3 its value is at most n4/3 , which implies that for any component
Ci = Ai ∪· Bi , we have balance(Ci ) ≤ n4/3 . If we apply this inequality to the component C = A ∪· B with
|A| > n/2, we obtain |B| ≥ n/2 − n2/3 . Hence with probability 1 − 1/n1/3 there is a component with at least
n − n2/3 nodes, which means that the algorithm must have performed at least n − n2/3 − 1 comparisons.
Therefore the expected number of comparisons is at least (1 − 1/n1/3 )(n − n2/3 − 1) = n − o(n).
4.2
A stronger lower bound
To obtain a stronger lower bound, we extend the definition of the graph that captures the current knowledge
of A. Now a positive edge can be verified or non-verified. A verified positive edge (x, y) is created only after
comparing two balls x and y such that color(x) = color(y). All other positive edges are non-verified. The
algorithm can also turn a non-verified positive edge (x, y) into a verified positive edge by comparing x and y.
By the same reasoning as in Lemma 4.1 we obtain the following.
Lemma 4.3. If A reports that a binary input contains a majority element, then the graph consisting of all
verified positive edges contains a connected component with at least n2 nodes.
Now the goal is to construct a large component in the graph that consists of all verified positive edges, so
it makes sense for A to compare two balls from the same component. However, without loss of generality,
such comparisons are executed after having identified a large component in the graph consisting of all positive
edges. Then, A asks sufficiently many queries of the form (x, y), where (x, y) is a non-verified edge from the
identified component. In other words, A first isolates a candidate for a majority, and then makes sure that
all inferred equalities really hold, which is necessary because with very small probability the input is not
binary. This allows us to bound the total number of comparisons from below as follows. We define that a
majority edge is an edge between two nodes of the majority color.
Lemma 4.4. The expected number of comparisons used by A on a binary input is at least n − o(n) plus the
expected number of non-verified majority edges.
Proof. Recall that if there exists a component C = A ∪· B with |A| > n/2 then with probability 1 − 1/n1/3
we also have |B| ≥ n/2 − n2/3 . Set A consists of nodes of the majority color, although possibly not all nodes
of the majority color are there. However, because B is large, there are at most n2/3 nodes of the majority
12
color outside of √
A. Also, because we consider binary inputs chosen uniformly at random, by Chernoff bound
|A| ≤ n/2 + O( n log n) with probability 1 − 1/n.
The expected number of comparisons used by A to construct a component C = A ∪· B such that |A| > n/2
is at least n − n2/3 − 1. Then, A needs to verify sufficiently many non-verified edges inside A to obtain a
connected component of size n/2 in the graph that consists of verified positive edges. By construction, there
1/3
are no cycles in the graph that
√ consists of positive edges. Hence with probability 1 − 1/n − 1/n there will
2/3
be no more than n + O( n log n) non-verified positive edges between nodes outside of B when A reports
a majority. Consequently,
√ the additional verification cost is the expected number of non-verified majority
edges minus n2/3 + O( n log n) = o(n).
In the remaining part of this section we analyze the expected number of non-verified majority edges
constructed during the execution of the algorithm. We show that this is at least (c − 1)n − o(n) for some
c > 1. Then, Lemma 4.4 implies the claimed lower bound.
A component C = A ∪· B is called monochromatic when A = ∅ or B = ∅ (by symmetry, we will assume
the latter) and dichromatic otherwise. With probability 1 − 1/n1/3 , when A reports a majority there is one
large dichromatic component with at least n − n2/3 nodes, and hence the total number of components is
at most n2/3 + 1. It is convenient to interpret the execution of A as a process of eliminating components
by merging two components into one. Each such merge might create a new non-verified edge. We define
that the cost of such a non-verified edge is the probability that it is a majority edge. We want to argue that
because all but n2/3 components will be eventually eliminated, the total cost of all non-verified edges that we
create is (c − 1)n − o(n).
We analyze in more detail the merging process in terms of mono- and dichromatic components. Let
predictk be the random variable denoting the probability that, after k steps of A, a node from the larger part
of a component is of the majority color. It is rather difficult to calculate predictk exactly, so we will use a
crude upper bound instead. An important property of the upper bound will be that it is nondecreasing in k.
When A compares two balls x ∈ Ci and y ∈ Cj with i 6= j to obtain a new component C = A ∪· B there are
three possible cases:
1. Ci and Cj are monochromatic. Then with probability
and with probability 21 it is dichromatic.
1
2
the new component C is also monochromatic,
2. Ci is dichromatic and Cj is monochromatic. The new component is dichromatic. With probability 12 we
have a new non-verified edge, and with probability at least 12 (1 − predictk ) we have a new non-verified
majority edge.
3. Ci and Cj are dichromatic. Then with probability
and B, and one of them is a majority edge.
1
2
we create a new non-verified edge inside both A
We analyze the expected total cost of all non-verified edges when only one component remains. When A
reports a majority up to n2/3 components might remain, but this changes only the lower order terms of the
bound.
Lemma 4.5. The expected total cost of all non-verified edges when only one component remains is at least
P2n/3
1 1
k=1 E min 6 , 2 (1 − predictk ) .
Proof. We start with n components and need to eliminate all but at most one of them. To each component
we associate credit, 12 to each dichromatic and 16 to each monochromatic one. The algorithm can collect the
credit from both of the components it merges, but it has to pay for credit of newly created one. Additionally
algorithm has to pay for any non-verified majority edge created by merging.
In every step we have three possibilities:
1. Merge two monochromatic components into one. With probability 12 the new component is dichromatic,
and with probability 12 the new component is monochromatic. Thus the expected amortized cost for
this step is 0.
13
2. Merge a monochromatic components with a dichromatic component. Then the total number of
monochromatic components decreases by 1 and we add with probability at least 12 (1 − predictk ) a
non-verified majority edge. The expected amortized cost for this step is 12 (1 − predictk ) − 16 .
1
2
3. Merge two dichromatic components while adding with probability
expected amortized cost for this step is 0.
a non-verified majority edge. The
In total algorithm has to pay for initial credits and for each step, making the total expected cost at least
2/3n
n−1
n X
1 X
+
≥
E min
E min 0, 12 (1 − predictk ) −
6
6
1 1
6 , 2 (1
− predictk ) .
k=1
k=1
We note that by truncating the sum at 23 n we do not lose any cost estimation, as for k ≥ 23 n our estimation
for predictk gives 1.
Now we focus deriving an upper bound for the expression obtained in Lemma 4.5. To bound predictk
we use an approach due to Christofides [5]. At any given step k we will look at all components with a
nonzero balance. Specifically, we introduce two new random variables: Mk being the largest balance of a
component, and Nk being the number of components with a nonzero balance. Since at each step, Nk is
decreased in expectation at most by 32 , we have E[Nk ] ≥ n − 32 (k − 1), and w.v.h.p., by Chernoff bound
√
Nk ≥ n − 32 k − O( k log n).
Since by Lemma 4.2 the expected sum of balances is n, and each nonzero component contributes at least
1 to the sum, we have E[Mk ] ≤ n − E[Nk − 1] = 32 k − 12 .
Now to proceed, for a component Ci = Ai ∪· Bi we define δi = ||Ai | − |Bi ||, a positive value such that
δi2 = balance(Ci ). Thus, at any given step k, the algorithm observes the nonzero values δ1 , δ2 , . . . , δNk .
Without loss of generality we can narrow our focus on a component C1 . We are interested in bounding the
probability
Pr(A1 in majority) = Pr(δ1 + ε2 δ2 . . . + εNk δNk ≥ 0) =
1
2
+
1
2
Pr(ε2 δ2 . . . + εNk δNk ∈ [−δ1 , δ1 ]),
where 2 , 3 , . . . , Nk ∈ {−1, 1} are drawn independently and uniformly at random. By a result of Erdős [11],
if δ2 , . . . , δNk ≥ 1 then the above is maximized for δ2 = . . . = δNk = 1.
We now approximate binomial distribution using the symmetric case of de Moivre–Laplace Theorem.
Recall that
Z x
2
1
Φ(x) = √
e−t /2 dt
2π −∞
is the cumulative distribution function of the normal distribution.
Theorem 4.6 (De Moivre–Laplace). Let Sn be the number of successes in n independent coin flips. Then
n
√
√
n
Pr
+ x1 n ≤ Sn ≤ + x2 n ∼ Φ(2x2 ) − Φ(2x1 ).
2
2
In our case we are interested in Nk − 1 coin flips and the number of successes in the range [(Nk − 1)/2 −
δ1 /2, (Nk − 1)/2 + δ1 /2]. Thus probability that 1 is the majority can be bounded from above by
1
δ1
δ1
1
δ1
√
√
√
Pr(A1 is the majority) ≤
Φ
−Φ −
+ =Φ
.
2
2
Nk − 1
Nk − 1
Nk − 1
√
Because Mk is the largest balance of√a component, δ1 , δ2 , . . . , δNk are bounded from above by Mk . Additionally, w.v.h.p. Nk ≥ n − 32 k − O( n log n), thus
s
predictk ≤ Φ
Mk
√
3
n − 2 k − O( n log n)
14
!
.
√
Since Φ( x/const) is a concave function, we can apply expected value, and get
s
s
!
!
3
E[Mk ]
2k
E[predictk ] ≤ Φ
∼Φ
.
√
n − 23 k − O( n log n)
n − 32 k
Now we are ready to bound the sum from Lemma 4.5. Using the linearity of expectation and inequality
min( 16 , 12 x) ≥ 16 x for x ∈ [0, 1] we obtain:
2n/3
2n/3
X
X
1
1
1 1
E
min , (1 − predictk )
=
E min , (1 − predictk )
≥
6 2
6 2
k=1
k=1
s
2n/3
Z 2/3
X 1
1
≥
(1 − E[predictk ]) ≥ n ·
6
6
0
1−Φ
k=1
1
3
2x
− 32 x
!!
dx − o(n).
Finally, we calculate
Z
1+
0
2/3
1
6
s
1−Φ
1
3
2x
− 23 x
!!
dx ≈ 1.0191289.
Theorem 4.7. Any algorithm that reports majority exactly requires in expectation at least 1.019n comparisons.
5
Conclusions
We have presented a Las Vegas algorithm for finding a majority color ball using, with high probability,
7
6 n + o(n) comparisons. We have also shown that the expected number of comparisons needs to be at least
1.019n. We believe that a more careful application of our methods might slightly increase the lower bound,
but achieving 76 n, which we believe to be the answer, requires a new approach. Another interesting question
is to consider Monte Carlo algorithms.
Acknowledgments
Most of the work has been done while PU was affiliated to Aalto University, Finland.
References
[1] M. Aigner, G. D. Marco, and M. Montangero. The plurality problem with three colors and more. Theor.
Comput. Sci., 337(1-3):319–330, 2005.
[2] L. Alonso, E. M. Reingold, and R. Schott. Determining the majority. Inf. Process. Lett., 47(5):253–255,
1993.
[3] L. Alonso, E. M. Reingold, and R. Schott. The average-case complexity of determining the majority.
SIAM J. Comput., 26(1):1–14, 1997.
[4] R. S. Boyer and J. S. Moore. MJRTY: A fast majority vote algorithm. In Automated Reasoning: Essays
in Honor of Woody Bledsoe, pages 105–118, 1991.
[5] D. Christofides. On randomized algorithms for the majority problem. Discrete Applied Mathematics,
157(7):1481–1485, 2009.
[6] F. R. K. Chung, R. L. Graham, J. Mao, and A. C. Yao. Oblivious and adaptive strategies for the
majority and plurality problems. Algorithmica, 48(2):147–157, 2007.
15
[7] V. Chvátal. The tail of the hypergeometric distribution. Discrete Mathematics, 25(3):285–287, 1979.
[8] D. Dor and U. Zwick. Selecting the median. SIAM J. Comput., 28(5):1722–1758, 1999.
[9] D. Dor and U. Zwick. Median selection requires (2 + )n comparisons. SIAM J. Discrete Math.,
14(3):312–325, 2001.
[10] D. Eppstein and D. S. Hirschberg. From discrepancy to majority. In LATIN, volume 9644 of Lecture
Notes in Computer Science, pages 390–402. Springer, 2016.
[11] P. Erdős. On a lemma of Littlewood and Offord. Bull. Amer. Math. Soc., 51(12):898–902, 12 1945.
[12] M. Fischer and S. Salzberg. Finding a majority among n votes: solution to problem 81-5. Journal of
Algorithms, 1982.
[13] D. Gerbner, G. O. H. Katona, D. Pálvölgyi, and B. Patkós. Majority and plurality problems. Discrete
Applied Mathematics, 161(6):813–818, 2013.
[14] D. Král, J. Sgall, and T. Tichý. Randomized strategies for the plurality problem. Discrete Applied
Mathematics, 156(17):3305–3311, 2008.
[15] G. D. Marco and E. Kranakis. Searching for majority with k-tuple queries. Discrete Math., Alg. and
Appl., 7(2), 2015.
[16] G. D. Marco and A. Pelc. Randomized algorithms for determining the majority on graphs. Combinatorics,
Probability & Computing, 15(6):823–834, 2006.
[17] M. Paterson. Progress in selection. In Algorithm Theory–SWAT’96, pages 368–379. Springer, 1996.
[18] M. E. Saks and M. Werman. On computing majority by comparisons. Combinatorica, 11(4):383–387,
1991.
[19] M. Vizer, D. Gerbner, B. Keszegh, D. Pálvölgyi, B. Patkós, and G. Wiener. Finding a majority ball
with majority answers. Electronic Notes in Discrete Mathematics, 49:345–351, 2015.
[20] G. Wiener. Search for a majority element. Journal of Statistical Planning and Inference, 100(2):313–318,
2002.
16
| 8cs.DS
|
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
1
A One-step Approach to Computing a Polytopic Robust Positively Invariant
Set
Paul Trodden, Member, IEEE
arXiv:1504.06488v3 [cs.SY] 21 Jul 2016
Abstract
A procedure and theoretical results are presented for the problem of determining a minimal robust
positively invariant (RPI) set for a linear discrete-time system subject to unknown, bounded disturbances.
The procedure computes, via the solving of a single LP, a polytopic RPI set that is minimal with respect
to the family of RPI sets generated from a finite number of inequalities with pre-defined normal vectors.
Index Terms
Linear systems; Uncertain systems; Computational methods; Optimization; Invariant sets
I. I NTRODUCTION
We consider the problem of finding, for the discrete-time, linear time-invariant system,
x+ = Ax + w,
(1)
a robust positively invariant (RPI) set. That is, a set R ⊂ Rn with the property
Ax + w ∈ R, ∀x ∈ R, w ∈ W.
(2)
In this problem, x ∈ Rn is the current state and x+ its successor. The disturbance w ∈ Rn
is unknown but lies in a polytopic (compact and convex) set W that contains the origin in its
interior.
Robust or disturbance invariant sets are important in control, and their theory and computation
have attracted significant attention; see, for example, [1]–[4] and references therein. One set that
is of particular interest is the minimal RPI (mRPI) set—that is, the RPI set that is smallest in
volume among all the RPI sets for a system—which is also the set of states reachable from the
origin in the presence of a bounded disturbance. This set is an essential ingredient in many robust
P. A. Trodden is with the Department of Automatic Control & Systems Engineering, University of Sheffield, Mappin Street,
Sheffield S1 3JD, UK (e-mail: [email protected]).
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
2
control algorithms. For example, in tube-based robust model predictive control (MPC) [5], an RPI
set is used to guarantee robust stability and feasibility in the presence of bounded uncertainty;
moreover, since the constraints in the MPC optimization problem are tightened according to the
size of the RPI set, then the smallest RPI set (i.e., the mRPI set) is desirable. However, computing
an exact representation of the mRPI is generally impossible (except for special instances of A, as
identified later), and instead an approximation is usually sought. A seminal contribution in this
regard is [3], which proposes a method for computing an abitrarily close outer-approximation
to the mRPI set, which is itself RPI.
The essence of the problem of computing exactly the mRPI set is that this set is, in general,
not finitely determined. Methods for computing approximations to the mRPI set, including [3],
rely on finding finite representations of the set. Recently, in the context of tube-based MPC, [6]
introduced and studied the notion of a polytopic RPI set defined by a finite number, r, of apriori selected linear inequalities. For a non-autonomous system x+ = Ax + Bu + w controlled
by a continuous positively homogeneous control law, u = κ(x), the authors showed that the
RPI set dynamic condition (2) has an equivalent representation as r functional inequalities. It
was established that a fixed-point solution to the functional equation corresponds to an RPI set
that is minimal, in volume, with respect to the entire family of RPI sets defined by the preselected inequalities, and is an invariant outer-approximation to the mRPI set. To compute this
set, the authors of [6] give an iterative procedure, based on solving a sequence of LPs, for which
convergence is guaranteed.
In this note, we adopt the notions of [6] and specialize their results to the case of the
linear autonomous system (1) (alternatively, the linear non-autonomous system with linear state
feedback control law) in order to develop a one-step approach, based on solving a single LP, to
the computation of the smallest RPI set defined by a pre-selected system of inequalities. Though
simple, to the author’s knowledge this has not appeared in the literature, although there are
related results; for example, it is known that checking the invariance of an existing polytope
is an LP [2]. On the other hand, the ability to synthesize a near-minimal RPI set by solving a
single LP potentially paves the way for robust control methods that re-compute the disturbance
invariant sets on-line, as done in, for example, the recently developed “plug-and-play” approach
to distributed MPC [7].
The proposed approach differs to the one of [3] in one important assumption: the number
and normal vectors of the inequalities that represent the RPI set are, as in [6], defined a priori,
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
3
while in [3] both are unknown until termination of the algorithm. This a-priori definition, first
proposed and studied by [6], has two consequences: firstly, the RPI set obtained is not necessarily
the mRPI set, or even an abitrarily close outer-approximation (as it is in [3]); however, it is the
smallest RPI set that can be represented by the finite number, r, of chosen inequalities with
normal vectors {Pi⊤ : i = 1 . . . r} [6]. To make a clear distinction, in this note we term this the
(P, r)-mRPI set when the number of chosen inequalities is r and the matrix of normal vectors
(the left-hand side of the defining system of inequalities) is P . Secondly, the method of [3]
involves solving a sequence of LPs and then computing a Minkowski summation, but here only
the solving of a single LP is required. The development of the procedure here comprises two
steps, the enumeration of which also serves to clarify the contribution of this note with respect
to [6]: first, we show that, for the studied linear autonomous system (1), the fixed-point solution
to the functional equation, which [6] showed is guaranteed to exist, is in fact unique. Secondly,
we show that the corresponding RPI set—which [6] proved to be minimal with respect to the
family of RPI sets represented by (P, r)—can be computed via a single linear program (LP), as
an alternative the iterative sequence of LPs proposed by [6].
Another method that uses a single LP to compute a disturbance invariant set is the optimized
robust control invariance approach of [4], applicable to the linear non-autonomous system x+ =
Ax+Bu+w. Because a robust control invariant (RCI) set—and the associated control policy—is
obtained, then this subsumes the robust positive invariance (where a fixed control law is assumed)
considered here. However, that approach optimizes over only those control policies that guarantee
a finitely determined set, achieved by employing a relaxed variation of the assumption, for (1),
that Ak W ⊆ αW for some α ∈ [0, 1) and finite integer k. In this note, the assumption that
A has eigenvalues inside the unit circle is required, which is different to the assumption used
for finite determination of RCI sets in [4], but weaker than the assumption required for finite
determination of the mRPI set for (1).
The organization of this note is as follows. First, in Section II, it is shown that for the
system (1), the fixed-point solution is, under suitable assumptions, unique. Subsequently, in
Section III, it is shown that the (P, r)-mRPI set for (1) may be computed via a single LP.
Finally, examples are given in Section IV to illustrate the practicality of the proposed approach,
before conclusions are made in Section V.
Notation: The sets of non-negative and positive reals are, respectively, R0+ and R+ . For
a, b ∈ Rn , a ≤ b applies element by element. A matrix M is non-negative, denoted M ≥ 0, if
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
4
Mij ≥ 0 for all i and j. λX is the scaling of a set X by λ ∈ R, defined as {λx : x ∈ X }.
AX denotes the image of a set X ⊂ Rn under the linear map A : Rn 7→ Rp , and is given by
{Ax : x ∈ X }. The support function of a set X is h(X , v) , sup{v ⊤ x : x ∈ X }. A polyhedron is
the convex intersection of a finite number of halfspaces, and a polytope is a closed and bounded
(hence compact) polyhedron.
II. E XISTENCE
AND UNIQUENESS OF A
(P, r)- M RPI
SET
For the system (1), we consider the case of a polytopic disturbance set
W , w ∈ Rn : F w ≤ g ,
(3)
where F ∈ Rp×n , g ∈ Rp0+ , and make the following two standing assumptions.
Assumption 1: The set W contains the origin in its interior.
Assumption 2: The eigenvalues of A are strictly within the unit circle.
The former assumption requires that g ∈ Rp+ . The latter assumption implies, as shown in [1],
that for a given compact disturbance set W there exists a compact RPI set, R, for the system (1),
satisfying (2).
Assumption 3: The RPI set R is a polytope that contains the origin in its interior.
Note that Assumptions 1 and 3 imply that the support functions to W and R, respectively,
are positive—a key technical property that will be used in this note to establish the existence
and uniqueness of the RPI set that we aim to compute.
In this note, following [6], we consider the RPI set constructed from a finite number, r, of
inequalities with pre-defined normal vectors. That is, R , R(q), defined as
R(q) , z ∈ Rn : P z ≤ q ,
(4)
where P ∈ Rr×n , Pi⊤ : i ∈ {1, . . . , r} spans Rn , Pi is the ith row of matrix P , and q ∈ Rr0+ .
The left-hand side of the inequalities—the matrix P —is to be chosen a priori by the designer.
The following result, which is an application of Farkas’ Lemma, establishes basic conditions on
the matrices A, P and F for the existence of an RPI set for the system (1) given the disturbance
polytope (3).
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
5
Theorem 1 (Adapted from Hennet and Castellan [8]): Suppose Assumptions 1–3 hold. Then
the set R(q) with some q = q̄ is robust positive invariant for the system (1) if and only if there
exist non-negative matrices H ∈ Rr×r and M ∈ Rr×p such that
HP = P A
(5a)
MF = P
(5b)
H q̄ + Mg ≤ q̄
(5c)
We will assume that P is chosen so that an RPI set exists:
Assumption 4: For the chosen P , and the system (A, W), there exists a q̄ ∈ Rr+ such that (2)
holds for all x ∈ R(q̄).
Remark 1: While Assumption 4 may appear strong, it is needed to narrow the class of matrices
that we consider to those that admit an RPI set. However, the procedure presented in the next
section includes a easy certification of existence of an RPI set for a chosen P : if an RPI set exists,
the (P, r)-mRPI set is returned. If no RPI set exists, the optimization problem is unbounded.
The authors of [6] show—in the more general setting of a linear non-autonomous system
controlled by a positively homogeneous state-feedback control law—that RPI condition (2) is
equivalent to the functional inequality
c(q) + d ≤ b(q),
(6)
where, for i = 1 . . . r, bi (q) , h(R(q), Pi⊤ ), ci (q) , h(AR(q), Pi⊤ ), di , h(W, Pi⊤ ). That is,
the set inclusion requirement is replaced by support function inequalities, which is a standard
technique [9]. Note that b(q) may be different to q; for example, in the case of redundant inequalities defining R(q). The topological properties of these functions described in the following two
lemmas are essential to establishing existence and uniqueness of the fixed-point solution to (6).
Lemma 1 (Adapted from Proposition 1 of [6]): Suppose that Assumptions 1–3 hold. Then the
functions b : Rr0+ 7→ Rr0+ , c : Rr0+ 7→ Rr0+ are continuous and monotonically non-decreasing; that
is, b(a1 ) ≤ b(a2 ) for a1 ≤ a2 . Also, d ∈ Rr+ .
Lemma 2 (Positive homegeneity of b, c): Suppose Assumptions 2 and 3 hold. Then the functions
b(·) and c(·) are positively homogeneous; that is b(λa) = λb(a) for λ ≥ 0, with a similar
expression for c(·).
Proof: Consider bi (λa) = h (R(λa), Pi ) for some a ∈ Rr0+ , λ ≥ 0 and i ∈ {1, . . . , r}.
By definition of R(·), R(λa) = λR(a). Thus, h (R(λa), Pi ) = h (λR(a), Pi ) = λh (R(a), Pi ),
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
6
for λ ≥ 0, where the latter equality follows directly from the definition of the support function [9]. Hence, bi (λa) = λbi (a), therefore b(λa) = λb(a). Positive homogeneity of c(·) may be
established by the same arguments.
The next result, which concerns the existence of a fixed-point solution to (6), was established
by [6] in the setting of a linear non-autonomous system controlled by positively homogeneous
state-feedback control law, and hence immediately applies to the more specialized case considered
in this note.
Theorem 2 (Theorem 1 of [6]): Suppose Assumptions 1–3 hold. Let Q , q ∈ Rr0+ : 0 ≤
q ≤ q̄ . Then, (i) for all q ∈ Q, c(q) + d ∈ Q and (ii) there exists at least one q ∗ ∈ Q satisfying
c(q ∗ ) + d = b(q ∗ ) = q ∗ if and only if Assumption 4 holds.
Remark 2: The necessity and sufficiency of Assumption 4 follows by definition. In particular,
if Assumption 4 does not hold, then there does not exist an RPI set for the system (A, W) with
the chosen P .
Remark 3: Note that, in view of the assumptions on g and the properties of b(·), c(·), and d,
a fixed-point solution q ∗ must be strictly positive.
With respect to computing a fixed-point solution, the sequence generated by the iterative
procedure q [p+1] = c(q [p]) + d, with q [0] = 0, converges to the fixed-point solution q ∗ with the
smallest 1-norm value, kq ∗ k1 [6, Theorem 2]. As the following result states, the corresponding
set R(q ∗ ) is RPI, and, in fact, is the minimal (smallest volume) RPI set over the family of RPI
sets defined by the r inequalities with left-hand side P .
T
Lemma 3 (Corollary 1 of [6]): R(q ∗ ) = X∈S X where
S , {R(q) : q ∈ H} , and H , q ∈ Rr0+ : c(q) + d ≤ b(q)
For convenience, we define this set R(q ∗ ) as the (P, r)-mRPI set.
Definition 1 ((P, r)-mRPI set): The (P, r)-mRPI set for system (1) is R(q ∗ ) where q ∗ =
b(q ∗ ) = c(q ∗ ) + d.
In this note, we propose an alernative to the iterative procedure of [6]. To this end, the next
result shows that the fixed-point solution to (6) is, in fact, unique. This result is then exploited
in Section III, wherein the problem of finding the fixed-point solution is cast as an LP.
Theorem 3 (Uniqueness of fixed-point solution): Suppose Assumptions 1–4 hold. Then there
exists a unique q ∗ ∈ Rr+ satisfying c(q ∗ ) + d = b(q ∗ ) = q ∗ .
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
7
Proof: Existence is established by Theorem 2, so it remains to show that q ∗ is unique. Let
l(q) = c(q)+d−b(q) and f (q) = b(q)−q. Finding the fixed-point solution c(q ∗ )+d = b(q ∗ ) = q ∗
is equivalent to finding q ∗ such that l(q ∗ ) = f (q ∗) = 0. Suppose there exist q 1 ∈ Rr+ and q 2 ∈ Rr+
such that l(q 1 ) = f (q 1 ) = 0, l(q 2 ) = f (q 2 ) = 0, and q 2 6= q 1 , i.e., q 2 − q 1 6= 0. There are two
possibilities:
(i) qi2 > qi1 for at least one i ∈ {1, . . . , r}, with qj2 ≤ qj1 otherwise;
(ii) q 2 ≤ q 1 , with qi2 < qi1 for at least one i ∈ {1, . . . , r}.
Consider case (i). Let
α = min
i=1...r
qi1
qi2
=
qp1
>0
qp2
Strict positivity follows from the discussion in Remark 3. Since qi2 > qi1 for at least one i,
then α < 1. Let s = αq 2 < q 2 . It follows, from positive homogeneity of b(·) and the fact that
b(q 2 ) − q 2 = 0, that f (s) = b(s) − s = b(αq 2 ) − αq 2 = α b(q 2 ) − q 2 ) = 0. Similarly,
l(s) = c(s) + d − b(s)
= c(αq 2 ) + d − b(αq 2 )
= αc(q 2 ) + d − αb(q 2 )
= α c(q 2 ) − b(q 2 ) + d
>0
where the second line follows from the positive homegeneity of c(·) and b(·), and the strict
inquality with zero follows from c(q 2 ) − b(q 2 ) = −d, α < 1 and d > 0. Now, by definition of α,
and since α < 1, then s ≤ q 1 with sp = qp1 . For the same p, we have fp (q 1 ) = bp (q 1 ) − qp1 = 0,
fp (s) = bp (s) − sp = 0, and, since s ≤ q 1 , then bp (s) ≤ bp (q 1 ). In fact, bp (s) = bp (q 1 ), as we
have already shown that bp (s) = sp = qp1 . We also have lp (q 1 ) = cp (q 1 ) + dp − bp (q 1 ) and lp (s) =
cp (s) + dp − bp (s). Because bp (q 1 ) = bp (s) and cp (s) ≤ cp (q 1 ), it follows that lp (s) ≤ lp (q 1 ). But
then 0 = lp (q 1 ) ≥ lp (s) > 0, and we have a contradiction: therefore, we conclude that case (i)
cannot hold, and either case (ii) holds or q 2 = q 1 . Now consider case (ii), and its equivalent
statement: qi1 > qi2 for at least one i ∈ {1, . . . , r}, with qj1 ≥ qj2 otherwise. Following the same
set of arguments, starting with the opposite definitions of α = mini=1...r {qi2 /qi1 } and s = αq 1 ,
we find that that case (ii) cannot hold either. Therefore, q 1 = q 2 = q ∗ , and the solution is unique.
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
III. C OMPUTING
8
THE
(P, r)- M RPI
SET VIA A SINGLE
LP
The problem of computing the (P, r)-mRPI set is that of finding the q that satisfies the
functional inequality (RPI condition) (6) while attaining the smallest value of kqk1 . The results
in the previous section show that this q in fact satisfies (6) with equality; it is the fixed-point
solution q ∗ . Therefore, the problem of finding q ∗ is
q ∗ = arg min {kqk1 : c(q) + d ≤ b(q)}
(7)
q
This is not tractable, as, by the definitions of b(·) and c(·), the constraints are maximization
problems involving the optimization variable:
max {Pi Ax : P x ≤ q} + max {Pi w : F w ≤ g}
≤ max {Pi x : P x ≤ q}
for i = 1 . . . r. However, by noting that the fixed-point solution is unique, we may replace the
problem of (7) with the maximization problem
q ∗ = arg max {kqk1 : c(q) + d = b(q)}
q
This problem then easily converts to a linear program, as shown by the following. Introduce
auxiliary variables ξ i ∈ Rn and ω i ∈ Rn for each RPI inequality i ∈ {1, . . . , r}. Then, noting
that q = b(q) = c(q) + d at the desired fixed-point solution, eliminate q and b(q), leading to the
problem
P : q ∗ = c∗ + d∗ , where (c∗ , d∗ ) = arg
max
i
r
X
{ci ,di ,ξ ,ω i }
∀i∈{1,...,r} i=1
ci + d i
(8)
subject to, for all i ∈ {1, . . . , r},
ci ≤ Pi Aξ i ,
(9a)
P ξ i ≤ c + d,
(9b)
d i ≤ Pi ω i ,
(9c)
F ω i ≤ g.
(9d)
In this problem, maximizing each ci subject to constraints (9a) and (9b) represents finding the
vector of support functions to AR. Constraint (9b) represents P x ≤ b(q), with the condition
c(q) + d = b(q) enforced. Constraints (9c) and (9d) represent finding d, the vector of support
functions to W.
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
9
Remark 4: Note that, by definition, di = h(W, Pi⊤ ) is constant and does not depend on
q. Therefore, d could be computed prior to solving P, by solving a sequence of LPs, before
entering the optimization as a parameter. However, our aim is to formulate a single LP (a onestep procedure) that computes, simultaneously, d, c and hence q.
Note that each di and ωi is bounded, via (9c) and (9d) and the assumptions on W. Further note
that this problem always has a feasible solution, since one can choose, for example, ci = di = 0
and ξ i = ω i = 0. The question, then, is whether an optimal solution exists, or the problem is
unbounded. To this end, we require the following result, which specializes Theorem 1 to the
fixed-point solution.
Proposition 1: Suppose Assumptions 1–4 hold. A vector q ∗ satisfies the fixed-point relation
c(q ∗ ) + d = b(q ∗ ) = q ∗ if and only if there exist non-negative matrices H ∈ Rr×r and M ∈ Rr×p
such that
HP = P A
(10a)
MF = P
(10b)
Hq ∗ + Mg = q ∗
(10c)
Proof: Consider the ith element of each of c(q ∗ ), d and b(q ∗ ), defined by the (primal) LPs
ci (q ∗ ) = max Pi Ax : P x ≤ q ∗
di = max Pi w : F w ≤ g
bi (q ∗ ) = max Pi x : P x ≤ q ∗
(11a)
(11b)
(11c)
If Assumptions 1–4 hold, then by the previous results there exists a q ∗ satisfying the fixed-point
equation. Moreover, each of the terms in (11) is well defined, which is the case if and only if
each LP is feasible and attains an finite optimum. Therefore, by weak duality, the dual of each
LP
∗
⊤
ci (q ∗ ) : min h⊤
i q : hi P = Pi A, hi ≥ 0 ,
⊤
di : min m⊤
i g : mi F = Pi , mi ≥ 0 ,
bi (q ∗ ) : min yi⊤ q ∗ : yi⊤ P = Pi , yi ≥ 0 ,
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
10
is feasible. Examining these dual problems, dual feasible solutions exist if and only if there exist
non-negative hi ∈ Rr , mi ∈ Rp , yi ∈ Rr such that
h⊤
i P = Pi A,
m⊤
i F = Pi ,
yi⊤ P = Pi .
Applying strong duality, which holds in view of the previous arguments, to each of the three
LPs
∗
ci (q ∗ ) = h⊤
i q ,
di = m⊤
i g,
bi (q ∗ ) = yi⊤ q ∗ .
Collecting all rows i = 1 . . . r,
c(q ∗ ) = Hq ∗
d = Mg,
b(q ∗ ) = Y q ∗ ,
where HP = P A, MF = P , Y P = P . Therefore, it follows that if the fixed-point equation
c(q ∗ ) + d = b(q ∗ ) = q ∗
is satisfied, then so are the conditions (10); conversely, if (10) are satisfied, then so is the
fixed-point equation.
Then the main result of this section follows.
Theorem 4: Suppose Assumptions 1–3 hold. If P satisfies Assumption 4, then problem
P admits an optimal solution corresponding to the fixed-point solution q ∗ . Otherwise, P is
unbounded above.
Proof: We use duality to prove the theorem. Our goal is to prove that the optimal solution
to P satisfies the conditions (10), for some non-negative H and M, if and only if Assumption 4
holds, and that P is otherwise unbounded. Since the primal LP problem P is known to be feasible,
it suffices to show that the dual problem is feasible—and the solution is as claimed—if and only
if P satisfies Assumption 4; on the other hand, if the dual is infeasible, then by weak duality
the primal problem P is unbounded.
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
11
The dual problem is
D:
r
X
min
{λi ,νi ,µi ,ηi }
∀i∈{1,...,r} k=1
(η k )⊤ g
(12)
subject to, for all i ∈ {1, . . . , r},
r
X
µk = 1,
(13a)
µk = 1,
(13b)
P ⊤ µi − A⊤ Pi⊤ λi = 0,
(13c)
F ⊤ η i − Pi⊤ νi = 0,
(13d)
λi , νi ≥ 0
(13e)
µi , η i ≥ 0
(13f)
λ−
k=1
ν−
r
X
k=1
where λi ∈ R, µi ∈ Rr , νi ∈ R, η i ∈ Rp are the dual variables associated with constraints (9a)–
(9d) respectively.
We first suppose the dual problem D is feasible. From (13a) and (13b), λi = νi = 1+
Pr
k
k=1 µi ,
for all i = 1, . . . , r, where µki is the ith element of µk ∈ Rr . From this and (13c), (13d), it follows
that
Pi A =
Pi =
Pr
1+
1+
(µi )⊤
Pr
k
k=1 µi
i ⊤
(η )
Pr
k
k=1 µi
P,
F,
k
k=1 µi
≥ 0. Collecting all rows i = 1 . . . r of P , it
P
follows that a feasible solution to D satisfies (10a) and (10b) with Hij = µij /(1 + rk=1 µki ) ≥ 0,
P
j = 1 . . . r, Mij = ηji /(1 + rk=1 µki ) ≥ 0, j = 1 . . . p; therefore, H and M are non-negative
where the division is permitted since
matrices.
Now we study the optimal solution to D. Since the primal problem P is known to be feasible,
and we assumed D to be feasible, then by strong duality (which holds regardless of the feasibility
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
12
of D) the optimal solutions to P and D are attained and equal in objective value. So, applying
complementary slackness to (9a) and (9c),
r
X
λ∗i (c∗i − Pi Aξ i∗ ) = 0
i=1
r
X
νi∗ (d∗i − Pi ω i∗ ) = 0.
i=1
∗
where denotes a variable in the optimal solution. Since each term in these sums is non-positive,
λ∗i (c∗i − Pi Aξ i∗ ) = 0,
νi∗ (d∗i − Pi ω i∗ ) = 0,
for i = 1, . . . , r. Morever, because (by (13a) and (13b)) λ∗i > 0 and νi∗ > 0, then, at the optimum,
c∗i = Pi Aξ i∗ ,
d∗i = Pi ω i∗.
Hence,
c∗i + d∗i = Pi Aξ i∗ + Pi ω i∗
=
=
1+
1+
(µi∗ )⊤
Pr
P ξ i∗ +
k∗
k=1 µi
1
Pr
k∗
k=1 µi
r
X
1+
(η i∗ )⊤
Pr
i∗
µi∗
j Pj ξ
j=1
k∗
k=1 µi
+
p
X
j=1
F ω i∗
ηji∗ Fj ω i∗
(14)
!
.
Now consider the inequality (9d). Suppose F ω i∗ < g for some i ∈ {1, . . . , r} (i.e., Fj ω i∗ < gj for
all j = 1 . . . p). Complementary slackness implies that η i∗ = 0 which in turn implies (from (13d),
assuming that Pi is not trivially all zeros) that νi∗ = 0; but νi∗ ≥ 1 by (13b), which is a
contradiction. Hence, there must exist a subset K ⊂ {1, . . . , p} of active constraints for which
Fk ω i∗ = gk for k ∈ K. But for any j ∈
/ K, ηji∗ = 0.
Similarly, consider the inequality (9b). By complementary slackness, if P ξ i∗ < c∗ + d∗ then
µi∗ = 0. By (13c), this implies that A⊤ Pi⊤ λ∗i = 0. There are two cases to consider: (i) if any
elements of Pi A are non-zero then λ∗i = 0; (ii) if Pi A = 0 then λ∗i > 0 is permitted. We leave
case (ii) for now and consider (i) first. λ∗i = 0 contradicts (13a), which requires λ∗i ≥ 1. Hence,
there must exist a subset J ⊂ {1, . . . , r} of active constraints for which Pj ξ i∗ = c∗j + d∗j for
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
13
j ∈ J . But for any k ∈
/ J , µi∗
k = 0. As a consequence of the preceding arguments, (14) may
be re-written as
c∗i + d∗i =
1+
1
Pr
k∗
k=1 µi
X
j∈J
i∗
µi∗
j Pj ξ +
X
k∈K
ηki∗ Fk ω i∗
!
= Hi (c∗ + d∗ ) + Mi g
where Hi is the ith row of H and Mi is the ith row of M. The second line follows because
Hij = 0 for j ∈
/ J and Mik = 0 for k ∈
/ K, while Pj ξ i∗ = c∗j + d∗j for j ∈ J and Fk ω i∗ = gk
for k ∈ K.
Now case (ii). If Pi A = 0 then c∗i = 0. Moreover, λ∗i ≥ 1 is permitted, so the same
contradiction is not constructed. Then, however, either P ξ i∗ < d∗ , hence µi∗ = 0, or P ξj∗i = d∗j ,
i∗
with µi∗
/ J . Either way,
j ≥ 0, for j ∈ J ⊂ {1, . . . , r}, and µk = 0 for all k ∈
c∗i + d∗i = Hi (c∗ + d∗ ) + Mi g
as before.
Finally, collecting all rows i = 1 . . . r,
H(c∗ + d∗ ) + Mg = c∗ + d∗
which is the third condition in (10). This establishes that the solution to P, if it is attainable,
satisfies the conditions (10) for it to be the fixed-point solution. It is attainable if and only if
the dual problem D is feasible. Therefore, it remains to show that the D is feasible if and only
if Assumption 4 holds.
First, necessity of Assumption 4. Suppose Assumption 4 is not satisfied, but the dual D
is feasible. By definition, if Assumption 4 is not satisfied then for the chosen P and system
(A, W) there does not exist a q satisfying the functional inequality (6). Therefore, there exists
no q ∗ satisfying the functional equation and, by Proposition 1, the conditions (10). However, the
attainable optimal solution to P and D satisfies (10) with non-negative H and M, as has been
shown. Therefore, we have a contradiction, and conclude the optimal solution is attainable, and
D is feasible, only if Assumption 4 holds.
Second, sufficiency of Assumption 4. Writing the primal constraints (9) in the form Ax ≤ b,
where x is the vector of primal decision variables, it follows that the dual constraints (13) may
be written in the form A⊤ y = c, y ≥ 0, where y is the vector of dual variables and c is the
coefficients vector in the vectorized form, c⊤ x, of the objective function (8). By Farkas’ Lemma,
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
14
a feasible solution to A⊤ y = c, y ≥ 0 exists if and only if Ax ≥ 0 =⇒ c⊤ x ≥ 0. Hence,
we aim to show that, if Assumption 4 holds, then for all x satisfying Ax ≥ 0 we also have
c⊤ x ≥ 0. The system Ax ≥ 0 may be written in terms of the primal variables as
ci ≥ Pi Aξ i
P ξi ≥ c + d
d i ≥ Pi ω i
F ωi ≥ 0
for i = 1 . . . r. If Assumption 4 holds, then Hi P = Pi A and Mi F = Pi for some non-negative
Hi and Mi . Substituting into the system Ax ≥ 0,
ci ≥ H i P ξ i
P ξi ≥ c + d
di ≥ Mi F ω i
F ω i ≥ 0,
from which it follows that di ≥ 0 and ci ≥ Hi (c + d), hence c ≥ Hc. But we also have that,
if Assumption (4) holds, then there exists some q ∈ Rr+ for which 0 ≤ Hq ≤ Hq + Mg ≤ q.
Applying recursively, 0 ≤ H n q ≤ Hq ≤ q, H n ≥ 0 because H ≥ 0, and therefore limn→∞ H n ≥
0, if the limit exists. In fact, because HP = P A, the nullspace of P is A-invariant and P has
rank n, then the eigenvalues of H are are subset of the eigenvalues of A; hence, limn→∞ H n = 0
P
because ρ(A) < 1. Then c ≥ Hc ≥ limn→∞ H n c = 0. Consequently, c⊤ x = ri=1 ci + di ≥ 0.
Therefore, D is feasible if Assumption 4 holds.
IV. E XAMPLES
We consider the non-autonomous system
1 1
0.5
x + u + w,
x+ =
0 1
1
(15)
with w ∈ W = {w ∈ R2 : kwk∞ ≤ 0.1}. This is converted to the linear autonomous system (1)
by use of a state feedback control law u = Kx.
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
15
A. Computation of (P, r)-mRPI from selected inequalities
First, we use the feedback matrix K = [−0.4345, −1.0285], corresponding to the infinitehorizon LQR solution with cost matrices Q = I and R = 1. Note that in this example the mRPI
set is not finitely determined, and therefore an approximation is required.
Figure 1(a) shows the (P, r)-mRPI sets generated from r = 6, 20 and 48 inequalities, wherein
the ith row of P is designed as
h
i
2π(i−1)
,
Pi = sin 2π(i−1)
cos
r
r
(16)
i.e., so that P x ≤ 1 is the r-sided regular polygon. Also shown is the outer approximation to
the mRPI, which is itself RPI, computed using the algorithm of [3] and a tolerance ǫ = 10−4 .
This set, termed the ǫ-mRPI set, is defined by 48 non-redundant inequalities.
Figure 1(b) shows a similar comparison using K = [−0.0796, −0.4068], obtained as the LQR
solution with Q = I and R = 100. Now the ǫ-mRPI (ǫ = 10−4 ) comprises 172 non-redundant
inequalities, while the (P, r)-mRPI sets computed using the proposed method are shown for
r = 20, 60 and 172, again using (16) for P .
Table I compares the computation times and number of operations for computing the (P, r)mRPI with those for obtaining the ǫ-outer approximation using the algorithm of [3]. For the latter,
the Multi-Parametric Toolbox v3.0 [10] was used for set operations, with CPLEX 12.6 as the
LP solver for support function calculations. For the (P, r)-mRPI set computations (i.e., solving
the LP), CPLEX 12.6 was used as the LP solver. The platform was a 64-bit Intel Core i7-2600
at 3.40 GHz with 8 GB RAM. Times are reported as the mean elapsed time over 100 runs.
Comparison was also made with the iterative procedure of [6] for computing the (P, r)-mRPI.
The iterative procedure is
qk+1 = c(qk ) + d with q0 = 0
for which qk → q ∗ as k → ∞. This was implemented in M ATLAB using the MPT v3.0 [10] for
support function calculations (with CPLEX 12.6 as underlying LP solver). The function c(·) was
evaluated element by element at each iteration; that is, as r separate support function calculations.
For the simplest case considered of K = [−0.4345, −1.0285] and r = 6 (the first row of Table I),
the number of iterations to convergence (of |qk+1 − qk | to within a chosen tolerance of 10−6 ) was
34, which included the solving of 238 LPs and took a mean total time of 1.7 seconds. At the
other end of the scale, for the most difficult problem considered (K = [−0.0796, −0.4068] and
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
16
r=6
r = 20
r = 48
ǫ-mRPI
x2
0.2
0
−0.2
−0.5
−0.25
0
x1
0.25
0.5
(a) K = [−0.4345, −1.0285]
r = 20
r = 60
r = 172
ǫ-mRPI
0.4
x2
0.2
0
−0.2
−0.4
−2
−1
0
x1
1
2
(b) K = [−0.0796, −0.4068]
Fig. 1. Comparison of (P, r)-mRPI and ǫ-mRPI sets for the system (15) with different feedback matrices K.
r = 172), the iterative procedure required 70 iterations, the solving of over 12000 LPs, and took,
on average, 90 seconds. While these times can, of course, be shortened by using optimized code,
the intention here is merely to report the times obtained using standard computational tools.
B. Re-computing the (P, r)-mRPI set given P
An interesting use of the method is when an RPI set for the system is available, but is desired
to be re-computed or modified; for example, if the disturbance set changes. Potential applications
of this include “plug-and-play” tube-based approaches to distributed MPC, wherein a dynamic
subsystems’ disturbance set evolves over time as other subsystems are added to and removed
from the system of coupled subsystems [7]; in such situations, one needs a new RPI set that
takes into account the latest disturbance set. One could re-compute from scratch a new RPI set,
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
17
but it may be advantageous, in the interests of computation time, to modify an existing RPI set
instead. In the context of the approach proposed here, the P matrix of the known RPI set may
be used as a basis for computing the new RPI set.
For the system (15) with K = [−0.4345, −1.0285] and W = {w ∈ R2 : |w|∞ ≤ 0.1}, the P
matrix is obtained as that of the ǫ-mRPI set. For ǫ = 10−4, this comprises 48 inequalities. Now
suppose the disturbance set enlarges to
−0.3
0.1
2
W= w∈R :
≤w≤
−0.4
0.2
Figure 2 shows the (P, r)-mRPI and ǫ-mRPI sets based on the new disturbance set, using
for the former the P matrix from the old ǫ-mRPI set. The (P, r)-mRPI set, computed in 0.03 s
using the proposed method, is visually indistinguishable from the new ǫ-mRPI set.
V. C ONCLUSIONS
A procedure for computing a polytopic robust positively invariant set for a linear uncertain
system has been presented. The method, which requires the solution of a single LP, obtains the
an RPI set that is the smallest among those represented by a finite number inequalities with
pre-defined normal vectors, and offers an alternative method of computation to the iterative
procedure of [6]. Existence and uniqueness of a solution has been established. The practicality
of the approach has been demonstrated via examples.
TABLE I
C OMPARISON OF
COMPUTATION TIMES AND OPERATIONS FOR
LPs solved
(P, r)- M RPI AND ǫ- M RPI SETS .
Minkowski sums
Mean time (s)
K = [−0.4345, −1.0285]
r=6
1
0
0.005
r = 20
1
0
0.007
r = 48
1
0
0.019
369
11
2.9
ǫ-mRPI [3] (r = 48)
K = [−0.0796, −0.4068]
r = 20
1
0
0.008
r = 60
1
0
0.036
r = 172
1
0
0.30
3250
42
25
ǫ-mRPI [3] (r = 172)
July 22, 2016
DRAFT
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
18
New r-mRPI
New ǫ-mRPI
Original ǫ-mRPI
x2
0.5
0
−0.5
−1
0
−0.5
0.5
x1
Fig. 2. Comparison of (P, r)-mRPI and ǫ-mRPI sets for the system (15) with K = [−0.4345, −1.0285] and different disturbance
sets.
ACKNOWLEDGMENTS
The author would like to thank the anonymous reviewers for their constructive and useful
comments.
R EFERENCES
[1] I. Kolmanovsky and E. G. Gilbert, “Theory and computation of disturbance invariant sets for discrete-time linear systems,”
Mathematical Problems in Engineering, vol. 4, pp. 317–367, 1998.
[2] F. Blanchini, “Set invariance in control,” Automatica, vol. 35, no. 11, pp. 1747–1767, 1999.
[3] S. V. Raković, E. C. Kerrigan, K. I. Kouramas, and D. Q. Mayne, “Invariant approximations of the minimal robust positively
invariant set,” IEEE Transactions on Automatic Control, vol. 50, no. 4, pp. 406–410, April 2005.
[4] S. V. Raković, E. C. Kerrigan, D. Q. Mayne, and K. I. Kouramas, “Optimized robust control invariance for linear discretetime systems: Theoretical foundations,” Automatica, vol. 43, no. 5, pp. 831–841, May 2007.
[5] D. Q. Mayne, M. M. Seron, and S. V. Raković, “Robust model predictive control of constrained linear systems with
bounded disturbances,” Automatica, vol. 41, no. 2, pp. 219–224, 2005.
[6] S. V. Raković, B. Kouvaritakis, and M. Cannon, “Equi-normalization and exact scaling dynamics in homethetic tube model
predictive control,” Systems & Control Letters, vol. 62, no. 2, pp. 209–217, February 2013.
[7] S. Riverso, M. Farina, and G. Ferrari-Trecate, “Plug-and-play model predictive control based on robust control invariant
sets,” Automatica, vol. 50, pp. 2179–2186, 2014.
[8] J.-C. Hennet and E. B. Castelan, “Invariant polyhedra and control of large-scale systems,” in Proceedings of the 9th IFAC
Symposium on Large Scale Systems: Theory and Applications, 2001.
[9] R. Schneider, Convex bodies: the Brunn-Minkowski theory, ser. Encyclopedia of Mathematics and its Applications, G.-C.
Rota, Ed.
Cambridge University Press, 1993, vol. 44.
[10] M. Herceg, M. Kvasnica, C. N. Jones, and M. Morari, “Multi-Parametric Toolbox 3.0,” in Proc. of the European Control
Conference, Zürich, Switzerland, July 17–19 2013, pp. 502–510, http://control.ee.ethz.ch/~mpt.
July 22, 2016
DRAFT
| 3cs.SY
|
arXiv:cs/0403037v3 [cs.DS] 2 Nov 2004
Under consideration for publication in Theory and Practice of Logic Programming
1
Schedulers and Redundancy for a Class of
Constraint Propagation Rules
Sebastian Brand and Krzysztof R. Apt
Abstract
We study here schedulers for a class of rules that naturally arise in the context of rule-based
constraint programming. We systematically derive a scheduler for them from a generic
iteration algorithm of (Apt 2000). We apply this study to so-called membership rules
of (Apt and Monfroy 2001). This leads to an implementation that yields a considerably
better performance for these rules than their execution as standard CHR rules. Finally, we
show how redundant rules can be identified and how appropriately reduced sets of rules
can be computed.
KEYWORDS: constraint programming, rule-based programming, constraint propagation
1 Introduction
In this paper we identify a class of rules that naturally arise in the context of constraint programming represented by means of rule-based programming and study
efficient schedulers for these rules. We call these rules propagation rules, in short
prop rules. An important class of prop rules are the membership rules, introduced
in (Apt and Monfroy 2001). An example of a membership rule is
x ∈ {3, 4, 8}, y ∈ {1, 2} → z 6= 2.
Informally, it should be read as follows: if the domain of x is included in {3,4,8}
and the domain of y is included in {1,2}, then 2 is removed from the domain of z.
In the computations of constraint programs the variable domains gradually
shrink. So if the domain of x is included in {3,4,8}, then it will remain so during the computation. In turn, if 2 is removed from the domain of z, then this
removal operation does not need to be repeated. The prop rules generalize these
observations to specific conditions on the rule condition and body.
In the resulting approach to constraint programming the computation process
is limited to a repeated application of the prop rules intertwined with splitting
(labeling). So the viability of this approach crucially depends on the availability
of efficient schedulers for such rules. This motivates the work here reported. We
provide an abstract framework for such schedulers and use it as a basis for an
implementation.
More precisely, to obtain appropriate schedulers for the prop rules we use the
generic approach to constraint propagation algorithms introduced in (Apt 1999)
and (Apt 2000). In this framework one proceeds in two steps. First, a generic iteration algorithm on partial orderings is introduced and proved correct in an abstract
setting. Then it is instantiated with specific partial orderings and functions to obtain specific constraint propagation algorithms. In this paper, as in (Apt 2000), we
take into account information about the scheduled functions, which are here the
prop rules. This yields a specific scheduler in the form of an algorithm called R.
We then show by means of an implementation how this abstract framework can
be used to obtain a scheduler for the membership rules. The relevance of the membership rules for constraint satisfaction problems (CSPs) with finite domains stems
from the following observations made in (Apt and Monfroy 2001):
• constraint propagation can be naturally achieved by repeated application of
the membership rules;
• in particular the notions of arc consistency and hyper-arc consistency can be
characterized in terms of the membership rules;
• for constraints explicitly defined on small finite domains all valid membership rules can be automatically generated; (For a more referent
work on the subject of an automatic generation of such rules see
(Abdennadher and Rigotti 2001).)
• many rules of the CHR language (Constraint Handling Rules) of
(Frühwirth 1998) that are used in specific constraint solvers are in fact membership rules. In the logic programming approach to constraint programming
CHR is the language of choice to write constraint solvers.
The implementation is provided as an ECLi PSe program that accepts a set of
membership rules as input and constructs an ECLi PSe program that is the instantiation of the R algorithm for this set of rules. Since membership rules can be
naturally represented as CHR propagation rules, one can assess this implementation
by comparing it with the performance of the standard implementation of membership rules in the CHR language. By means of various benchmarks we found that
our implementation is considerably faster than CHR. It is important to stress that
this implementation was obtained by starting from “first principles” in the form
of a generic iteration algorithm on an arbitrary partial ordering. This shows the
practical benefits of studying the constraint propagation process on an abstract
level.
Additionally, we clarify how to identify prop rules that are redundant for the
considered computations and how to compute appropriately reduced sets of rules.
The concept of redundancy is formalized here in a “semantic” sense that takes
into account the type of computations performed by means of the considered rules.
We provide a simple test for redundancy that leads to a natural way of computing
minimal sets of rules in an appropriate sense. The computation of a specific minimal
set for the membership rules is then implemented in ECLi PSe .
CHR is available in a number of languages including the ECLi PSe and the Sicstus
Prolog systems. In both cases CHR programs are compiled into the host language,
so either ECLi PSe or the Sicstus Prolog. There is also a recent implementation in
Java, see (Abdennadher et al. 2001). To make CHR usable it is important that its
2
implementation does not incur too much overhead. And indeed a great deal of effort
was spent on implementing CHR efficiently. For an account of the most recent implementation see (Holzbaur et al. 2001). Since, as already mentioned above, many
CHR rules are membership rules, our approach provides a better implementation
of a subset of CHR. This, hopefully, may lead to new insights into a design and
implementation of languages appropriate for writing constraint solvers.
The paper is organized as follows. In the next section we briefly recall the original
generic iteration algorithm of (Apt 2000) and modify it successively to obtain the
iteration algorithm R for prop rules. An important novelty is the preprocessing phase
during which we analyze the mutual dependencies between the rules. This allows
us to remove permanently some rules during the iteration process. This permanent
removal of the scheduled rules is particularly beneficial in the context of constraint
programming where it leads to accumulated savings when constraint propagation
is intertwined with splitting.
In Section 3 we recall the membership rules of (Apt and Monfroy 2001) and show
that they are prop rules. Then in Section 4 we recall the relevant aspects of the
CHR language, discuss the implementation of the R algorithm and present several
benchmarks. Finally, in Section 5 we deal with the subject of redundancy of prop
rules.
2 Revisions of the Generic Iteration Algorithm
2.1 The Original Algorithm
Let us begin our presentation by recalling the generic algorithm of (Apt 2000). We
slightly adjust the presentation to our purposes by assuming that the considered
partial ordering also has the greatest element ⊤.
So we consider a partial ordering (D, ⊑) with the least element ⊥ and the greatest
element ⊤, and a set of functions F := {f1 , . . . , fk } on D. We are interested in
functions that satisfy the following two properties.
Definition 2.1
• f is called inflationary if x ⊑ f (x) for all x.
• f is called monotonic if x ⊑ y implies f (x) ⊑ f (y) for all x, y.
The following algorithm is used to compute the least common fixpoint of the
functions from F .
Generic Iteration Algorithm (GI)
d := ⊥;
G := F ;
while G 6= ∅ and d 6= ⊤ do
choose g ∈ G;
G := G − {g};
G := G ∪ update(G, g, d);
d := g(d)
end
3
where for all G, g, d the set of functions update(G, g, d) from F is such that
A {f ∈ F − G | f (d) = d ∧ f (g(d)) 6= g(d)} ⊆ update(G, g, d),
B g(d) = d implies that update(G, g, d) = ∅,
C g(g(d)) 6= g(d) implies that g ∈ update(G, g, d).
Intuitively, assumption A states that update(G, g, d) contains at least all the
functions from F − G for which the “old value”, d, is a fixpoint but the “new
value”, g(d), is not. So at each loop iteration such functions are added to the set G.
In turn, assumption B states that no functions are added to G in case the value of
d did not change. Assumption C provides information when g is to be added back
to G as this information is not provided by A. On the whole, the idea is to keep in
G at least all functions f for which the current value of d is not a fixpoint.
The use of the condition d 6= ⊤, absent in the original presentation, allows us
to leave the while loop earlier. Our interest in the GI algorithm is clarified by the
following result.
Theorem 2.2 (Correctness)
Suppose that all functions in F are inflationary and monotonic and that (D, ⊑) is
finite and has the least element ⊥ and the greatest element ⊤. Then every execution
of the GI algorithm terminates and computes in d the least common fixpoint of the
functions from F .
Proof. Consider the predicate I defined by:
I := (∀f ∈ F − G f (d) = d) ∧ (∀f ∈ F f (⊤) = ⊤).
Note that I is established by the assignment G := F . Moreover, it is easy to check
that by virtue of the assumptions A, B and C the predicate I is preserved by each
while loop iteration. Thus I is an invariant of the while loop of the algorithm.
Hence upon its termination
(G = ∅ ∨ d = ⊤) ∧ I
holds, which implies
∀f ∈ F f (d) = d.
This implies that the algorithm computes in d a common fixpoint of the functions
from F .
The rest of the proof is the same as in (Apt 2000). So the fact that this is the least
common fixpoint follows from the assumption that all functions are monotonic.
In turn, termination is established by considering the lexicographic ordering of
the strict partial orderings (D, ⊐) and (N , <), defined on the elements of D × N
by
(d1 , n1 ) <lex (d2 , n2 )
iff
d1 ⊐ d2 or (d1 = d2 and n1 < n2 ).
Then with each while loop iteration of the algorithm the pair (d, card G), where
card G is the cardinality of the set G, strictly decreases in the ordering <lex .
4
2.2 Removing Functions
We now revise the GI algorithm by modifying dynamically the set of functions that
are being scheduled. The idea is that, whenever possible, we remove functions from
the set F . This will allow us to exit the loop earlier which speeds up the execution
of the algorithm.
To this end we assume that for each function g ∈ F and each element d ∈
D, two lists of functions from F are given, friends(g, d) and obviated (g, d), to be
instantiated below. We then modify the GI algorithm in such a way that each
application of g to d will be immediately followed by the applications of all functions
from friends(g, d) and by a removal of the functions from friends(g, d) and from
obviated (g, d) from F and G. Below we identify a condition, (1), on friends(g, d)
and obviated (g, d) that ensures correctness of this scheduling strategy. Informally,
this condition states that after an application of all the functions from friends(g, d)
the functions from friends(g, d) and from obviated (g, d) will not change anymore
the subsequent values of d.
This modified algorithm has the following form.1
Revised Generic Iteration Algorithm (RGI)
d := ⊥;
F0 := F ;
G := F ;
while G 6= ∅ and d 6= ⊤ do
choose g ∈ G;
G := G − {g};
F := F − (friends(g, d) ∪ obviated (g, d));
G := G − (friends (g, d) ∪ obviated (g, d));
G := G ∪ update(G, h, d),
where friends(g, d) = [g1 , . . . , gk ] and h = g ◦ g1 ◦ . . . ◦ gk ;
d := h(d)
end
We now formalize the condition under which the Correctness Theorem 2.2 holds
with the GI algorithm replaced by the RGI algorithm. To this end we consider the
following property.
Definition 2.3
Suppose d ∈ D and f ∈ F . We say that f is stable above d if d ⊑ e implies f (e) = e.
We then say that f is stable if it is stable above f (d), for all d.
That is, f is stable if for all d and e, f (d) ⊑ e implies f (e) = e. So stability
implies idempotence, which means that f (f (d)) = f (d), for all d. Moreover, if d
and f (d) are comparable for all d, then stability implies inflationarity. Indeed, if d
⊑f (d), then the claim holds vacuously. And if f (d) ⊑ d, then by stability f (d) = d.
1
We need in it lists instead of sets since the considered functions will be applied in a specific
order. But in some places, for simplicity, we identify these lists with the sets.
5
Consider now the following condition
∀d ∀e ⊒ g ◦ g1 ◦ . . . ◦ gk (d) ∀f ∈ friends(g, d) ∪ obviated (g, d) (f (e) = e),
(1)
where friends(g, d) = [g1 , . . . , gk ]. That is, for all elements d, each function f in
friends(g, d) ∪ obviated (g, d) is stable above g ◦ g1 ◦ . . . ◦ gk (d), where friends(g, d)
is the list [g1 , . . . , gk ]. The following result holds.
Theorem 2.4
Suppose that all functions in F are inflationary and monotonic and that (D, ⊑) is
finite and has the least element ⊥ and the greatest element ⊤. Additionally, suppose
that for each function g ∈ F and d ∈ D two lists of functions from F are given,
friends(g, d) and obviated (g, d) such that condition (1) holds.
Then the Correctness Theorem 2.2 holds with the GI algorithm replaced by the
RGI algorithm.
Proof. In view of condition (1) the following statement is an invariant of the while
loop:
∀f ∈ F −G(f (d) = d) ∧ ∀f ∈ F (f (⊤) = ⊤) ∧ ∀f ∈ F0 −F ∀e ⊒ d (f (e) = e). (2)
So upon termination of the algorithm the conjunction of this invariant with the
negation of the loop condition, i.e.,
G=∅ ∨ d=⊤
holds, which implies that ∀f ∈ F0 (f (d) = d).
The rest of the proof is the same.
2.3 Functions in the Form of Rules
In what follows we consider the situation when the scheduled functions are of a
specific form b → g, where b is a condition and g a function, that we call a body.
We call such functions rules.
First, we explain how rules are applied. Given an element d of D a condition b is
evaluated in d. The outcome is either true, that we denote by holds (b, d), or false.
Given a rule b → g we define then its application to d as follows:
g(d) if holds(b, d)
(b → g)(d) :=
d
otherwise
We are interested in a specific type of conditions and rules.
Definition 2.5
Consider a partial ordering (D, ⊑).
• We say that a condition b is monotonic if for all d, e we have that holds(b, d)
and d ⊑ e implies holds(b, e).
• We say that a condition b is precise if the least d exists such that holds(b, d).
We call then d the witness for b.
• We call a rule b → g a prop rule if b is monotonic and precise and g is stable.
To see how natural this class of rules is consider the following example.
6
Example 2.6
Take a set A and consider the partial ordering
(P(A), ⊆).
In this ordering the empty set ∅ is the least element and A is the greatest element.
We consider rules of the form
B → G,
where B, G ⊆ A.
To clarify how they are applied to subsets of A we first stipulate for E ⊆ A
holds(B, E) iff B ⊆ E.
Then we view a set G as a function on P(A) by putting
G(E) := G ∪ E.
This determines the rule application of B → G.
It is straightforward to see that such rules are prop rules. In particular, the
element B of P(A) is the witness for the condition B. For the stability of G take
E ⊆ A and suppose G(E) ⊆ F . Then G(E) = G ∪ E, so G ∪ E ⊆ F , which implies
G ∪ F = F , i.e., G(F ) = F .
Such rules can be instantiated to many situations. For example, we can view the
elements of the set A as primitive constraints. Then each rule B → G is a natural
operation on the constraint store: if all constraints in B are present in the store,
then add to it all constraints in G.
Alternatively, we can view A as a set of some atomic formulas and each rule
B → G as a proof rule, usually written as
B
G
The minor difference with the usual proof-theoretic framework is that rules have
then a single conclusion. An axiom is then a rule with the empty set ∅ as the
condition. A closure under such a set of rules is then the set of all (atomic) theorems
that can be proved using them.
The algorithm presented below can in particular be used to compute efficiently
the closure under such proof rules given a finite set of atomic formulas A.
We now modify the RGI algorithm for the case of prop rules. In the algorithm
below we take into account that an application of a rule is a two step process:
testing of the condition followed by a conditional application of the body. This will
allow us to drop the parameter d from the lists friends(g, d) and obviated (g, d) and
consequently to construct such lists before the execution of the algorithm. The list
friends(g) will be constructed in such a way that we shall not need to evaluate
the conditions of its rules: they will all hold. The specific construction of the lists
friends(g) and obviated (g) that we use here will be provided in the second algorithm,
called Friends and Obviated Algorithm.
7
Rules Algorithm (R)
d := ⊥;
F0 := F ;
G := F ;
while G 6= ∅ and d 6= ⊤ do
choose f ∈ G; suppose f is b → g;
G := G − {b → g};
if holds(b, d) then
F := F − (friends(b → g) ∪ obviated (b → g));
G := G − (friends (b → g) ∪ obviated (b → g));
G := G ∪ update(G, h, d),
where friends(b → g) = [b1 → g1 , . . . , bk → gk ] and h = g ◦ g1 ◦ . . . ◦ gk ;
d := h(d)
else if ∀e ⊒ d ¬holds (b, e) then
F := F − {b → g}
end
end
Again, we are interested in identifying conditions under which the Correctness
Theorem 2.2 holds with the GI algorithm replaced by the R algorithm. To this end,
given a rule b → g in F and d ∈ D define as follows:
friends(b → g) if holds(b, d),
friends(b → g, d) :=
[]
otherwise
and
obviated (b → g) if holds (b, d),
obviated (b → g, d) :=
[b → g]
if ∀e ⊒ d ¬holds (b, e),
[]
otherwise
We obtain the following counterpart of the Correctness Theorem 2.2.
Theorem 2.7 (Correctness)
Suppose that all functions in F are prop rules of the form b → g, where g is
inflationary and monotonic, and that (D, ⊑) is finite and has the least element ⊥
and the greatest element ⊤. Further, assume that for each rule b → g the lists
friends(b → g, d) and obviated (b → g, d) defined as above satisfy condition (1) and
the following condition:
∀d( b′ → g ′ ∈ friends(b → g) ∧ holds (b, d) → ∀e ⊒ g(d) holds(b′ , e) ).
(3)
Then the Correctness Theorem 2.2 holds with the GI algorithm replaced by the
R algorithm.
Proof. It suffices to show that the R algorithm is an instance of the RGI algorithm.
On the account of condition (3) and the fact that the rule bodies are inflationary
functions, holds(b, d) implies that
((b → g) ◦ (b1 → g1 ) ◦ . . . ◦ (bk → gk ))(d) = (g ◦ g1 ◦ . . . ◦ gk )(d),
8
where friends(b → g) = [b1 → g1 , . . . , bk → gk ]. This takes care of the situation in
which holds(b, d) is true.
In turn, the definition of friends(b → g, d) and obviated (b → g, d) and assumption
B take care of the situation when ¬holds(b, d). When the condition b fails for all
e ⊒ d we can conclude that for all such e we have (b → g)(e) = e. This allows us to
remove at that point of the execution the rule b → g from the set F . This amounts
to adding b → g to the set obviated (b → g, d) at runtime. Note that condition (1)
is then satisfied.
We now provide an explicit construction of the lists friends and obviated for a
rule b → g in the form of the following algorithm. GI(d) stands here for the GI
algorithm activated with ⊥ replaced by d and the considered set of rules as the set
of functions F . Further, given an execution of GI(e), we call here a rule g relevant
if at some point g(d) 6= d holds after the “choose g ∈ G” action.
Friends and Obviated Algorithm (F & O)
w := witness of b;
d := GI(g(w));
friends(b → g) := list of the relevant rules h ∈ F in the execution of GI(g(w));
obviated (b → g) := [ ];
for each (b′ → g ′ ) ∈ F − friends(b → g) do
if g ′ (d) = d or ∀e ⊒ d ¬holds (b′ , e) then
obviated (b → g) := [ b′ → g ′ | obviated (b → g) ]
end
end
Note that b → g itself is not contained in friends(b → g) as it is a prop rule, however
it is in obviated (b → g), since by the stability of g g(d) = d holds.
The following observation now shows the adequacy of the F & O algorithm for
our purposes.
Lemma 2.8
Upon termination of the F & O algorithm conditions (1) and (3) hold, where the
lists friends(b → g, d) and obviated (b → g, d) are defined as before Theorem 2.7.
Let us summarize now the findings of this section that culminated in the R algorithm. Assume that all functions are of the form of the rules satisfying the conditions
of the Correctness Theorem 2.7. Then in the R algorithm, each time the evaluation
of the condition b of the selected rule b → g succeeds,
• the rules in the list friends(b → g) are applied directly without testing the
value of their conditions,
• the rules in friends(b → g) ∪ obviated (b → g) are permanently removed from
the current set of functions G and from F .
9
2.4 Recomputing of the Least Fixpoints
Another important optimization takes place when the R algorithm is repeatedly
applied to compute the least fixpoint. More specifically, consider the following sequence of actions:
• we compute the least common fixpoint d of the functions from F ,
• we move from d to an element e such that d ⊑ e,
• we compute the least common fixpoint above e of the functions from F .
Such a sequence of actions typically arises in the framework of CSPs, further studied
in Section 3. The computation of the least common fixpoint d of the functions
from F corresponds there to the constraint propagation process for a constraint
C. The moving from d to e such that d ⊑ e corresponds to splitting or constraint
propagation involving another constraint, and the computation of the least common
fixpoint above e of the functions from F corresponds to another round of constraint
propagation for C.
Suppose now that we computed the least common fixpoint d of the functions from
F using the RGI algorithm or its modification R for the rules. During its execution
we permanently removed some functions from the set F . These functions are then
not needed for computing the least common fixpoint above e of the functions from
F . The precise statement is provided in the following simple, yet crucial, theorem.
Theorem 2.9
Suppose that all functions in F are inflationary and monotonic and that (D, ⊑)
is finite. Suppose that the least common fixpoint d0 of the functions from F is
computed by means of the RGI algorithm or the R algorithm. Let Ffin be the final
value of the variable F upon termination of the RGI algorithm or of the R algorithm.
Suppose now that d0 ⊑ e. Then the least common fixpoint e0 above e of the
functions from F coincides with the least common fixpoint above e of the functions
from Ffin .
Proof. Take a common fixpoint e1 of the functions from Ffin such that e ⊑ e1 .
It suffices to prove that e1 is common fixpoint of the functions from F . So take
f ∈ F − Ffin . Since condition (2) is an invariant of the main while loop of the RGI
algorithm and of the R algorithm, it holds upon termination and consequently f is
stable above d0 . But d0 ⊑ e and e ⊑ e1 , so we conclude that f (e1 ) = e1 .
Intuitively, this result means that if after splitting we relaunch the same constraint propagation process we can disregard the removed functions.
In the next section we instantiate the R algorithm by a set of rules that naturally arise in the context of constraint satisfaction problems with finite domains. In
Section 4 we assess the practical impact of the discussed optimizations.
3 Concrete Framework
We now proceed with the main topic of this paper, the schedulers for the rules that
naturally arise in the context of constraint satisfaction problems. First we recall
briefly the necessary background on the constraint satisfaction problems.
10
3.1 Constraint Satisfaction Problems
Consider a sequence of variables X := x1 , . . . , xn where n > 0, with respective
domains D1 , . . . , Dn associated with them. So each variable xi ranges over the
domain Di . By a constraint C on X we mean a subset of D1 × . . . × Dn . Given an
element d := d1 , . . . , dn of D1 × . . . × Dn and a subsequence Y := xi1 , . . . , xiℓ of X
we denote by d[Y ] the sequence di1 , . . . , diℓ . In particular, for a variable xi from X,
d[xi ] denotes di .
Recall that a constraint satisfaction problem, in short CSP, consists of a finite
sequence of variables X with respective domains D, together with a finite set C of
constraints, each on a subsequence of X. We write it as hC ; x1 ∈ D1 , . . . , xn ∈ Dn i,
where X := x1 , . . . , xn and D := D1 , . . . , Dn .
By a solution to hC ; x1 ∈ D1 , . . . , xn ∈ Dn i we mean an element d ∈ D1 ×. . .×Dn
such that for each constraint C ∈ C on a sequence of variables X we have d[X] ∈ C.
We call a CSP consistent if it has a solution. Two CSPs with the same sequence of
variables are called equivalent if they have the same set of solutions.
3.2 Partial Orderings
With each CSP P := hC ; x1 ∈ D1 , . . . , xn ∈ Dn i we associate now a specific
partial ordering. Initially we take the Cartesian product of the partial orderings
(P(D1 ), ⊇), . . . , (P(Dn ), ⊇). So this ordering is of the form
(P(D1 ) × . . . × P(Dn ), ⊇)
where we interpret ⊇ as the Cartesian product of the reversed subset ordering. The
elements of this partial ordering are sequences (E1 , . . . , En ) of respective subsets of
(D1 , . . . , Dn ) ordered by the component-wise reversed subset ordering. Note that
in this ordering (D1 , . . . , Dn ) is the least element while
(∅, . . . , ∅)
| {z }
n times
is the greatest element. However, we would like to identify with the greatest element
all sequences that contain as an element the empty set. So we divide the above
partial ordering by the equivalence relation R∅ according to which
(E1 , . . . , En ) R∅ (F1 , . . . , Fn ) iff (E1 , . . . , En ) = (F1 , . . . , Fn )
or (∃i Ei = ∅ and ∃j Fj = ∅).
It is straightforward to see that R∅ is indeed an equivalence relation.
In the resulting quotient ordering there are two types of elements: the sequences
(E1 , . . . , En ) that do not contain the empty set as an element, that we continue to
present in the usual way with the understanding that now each of the listed sets
is non-empty, and one “special” element equal to the equivalence class consisting
of all sequences that contain the empty set as an element. This equivalence class is
the greatest element in the resulting ordering, so we denote it by ⊤. In what follows
we denote this partial ordering by (DP , ⊑).
11
3.3 Membership Rules
Fix now a specific CSP P := hC ; x1 ∈ D1 , . . . , xn ∈ Dn i with finite domains. We
recall the rules introduced in (Apt and Monfroy 2001). They are called membership
rules and are of the form
y1 ∈ S1 , . . . , yk ∈ Sk → z1 6= a1 , . . . , zm 6= am ,
where
• y1 , . . . , yk are pairwise different variables from the set {x1 , . . . , xn } and
S1 , . . . , Sk are subsets of the respective variable domains,
• z1 , . . . , zm are variables from the set {x1 , . . . , xn } and a1 , . . . , am are elements
of the respective variable domains.2
Note that we do not assume that the variables z1 , . . . , zm are pairwise different.
The computational interpretation of such a rule is:
if for i ∈ [1..k] the current domain of the variable yi is included in the set Si , then for
j ∈ [1..m] remove the element ai from the domain of zi .
When each set Si is a singleton, we call a membership rule an equality rule.
Let us mention here that in (Apt and Monfroy 2001) the interpretation of the
conditions of an equality rule is slightly different, as it is stipulated that the current domain of the variable yi is to be equal to the singleton set Si . However, in
the discussed algorithms the membership rules are applied only when all variable
domains are non-empty and then both interpretations coincide.
Let us reformulate this interpretation so that it fits the framework considered in
the previous section. To this end we need to clarify how to
• evaluate the condition of a membership rule in an element of the considered
partial ordering,
• interpret the conclusion of a membership rule as a function on the considered
partial ordering.
Let us start with the first item.
Definition 3.1
Given a variable y with the domain Dy and E ∈ P(Dy ) we define
holds(y ∈ S, E)
iff E ⊆ S,
and extend the definition to the elements of the considered ordering (DP , ⊑) by
putting
holds(y ∈ S, (E1 , . . . , En ))
holds(y ∈ S, ⊤).
2
iff Ek ⊆ S,
where we assumed that y is xk , and
In (Apt and Monfroy 2001) it is also assumed that the lists y1 , . . . , yk and z1 , . . . , zm have no
variable in common. We drop this condition so that we can combine the membership rules.
12
Furthermore we interpret a sequence of conditions as a conjunction, by putting
holds ((y1 ∈ S1 , . . . , yk ∈ Sk ), (E1 , . . . , En ))
iff holds(yi ∈ Si , (E1 , . . . , En )) for i ∈ [1..k].
Concerning the second item we proceed as follows.
Definition 3.2
Given a variable z with the domain Dz we interpret the atomic formula z 6= a as a
function on P(Dz ), defined by:
(z 6= a)(E) := E − {a}.
Then we extend this function to the elements of the considered ordering (DP , ⊑)
as follows:
• on the elements of the form (E1 , . . . , En ) we put
(z 6= a)(E1 , . . . , En ) := (E1′ , . . . , En′ ),
where
— if z ≡ xi , then Ei′ = Ei − {a},
— if z 6≡ xi , then Ei′ = Ei .
If the resulting sequence (E1′ , . . . , En′ ) contains the empty set, we replace it
by ⊤,
• on the element ⊤ we put (z =
6 a)(⊤) := ⊤
Finally, we interpret a sequence z1 6= a1 , . . . , zm 6= am of atomic formulas by interpreting each of them in turn.
As an example take the CSP
P := hC ; x1 ∈ {a, b, c}, x2 ∈ {a, b, c}, x3 ∈ {a, b, c}, x4 ∈ {a, b, c}i
and consider the membership rule
r := x1 ∈ {a, b}, x2 ∈ {b} → x3 6= a, x3 6= b, x4 6= a.
Then we have
r({a}, {b}, {a, b, c}, {a, b}) =
r({a, b, c}, {b}, {a, b, c}, {a, b}) =
r({a, b}, {b}, {a, b}, {a, b}) =
({a}, {b}, {c}, {b}),
({a, b, c}, {b}, {a, b, c}, {a, b}),
⊤.
In view of the Correctness Theorem 2.7 the following observation allows us to
apply the R algorithm when each function is a membership rule and when for each
rule b → g the lists friends(b → g) and obviated (b → g) are constructed by the
F & O algorithm.
13
Note 3.3
Consider the partial ordering (DP , ⊑).
(i) Each membership rule is a prop rule.
(ii) Each function z1 6= a1 , . . . , zm 6= am on DP is
• inflationary,
• monotonic.
To be able to instantiate the algorithm R with the membership rules we still
need to define the set update(G, g, d). In our implementation we chose the following
simple definition:
F − G if holds (b, d) and g(d) 6= d,
update(G, b → g, d) :=
∅
otherwise.
To illustrate the intuition behind the use of the lists friends(b → g) and
obviated (b → g) take the above CSP
P := hC ; x1 ∈ {a, b, c}, x2 ∈ {a, b, c}, x3 ∈ {a, b, c}, x4 ∈ {a, b, c}i
and consider the membership rules
r1 :=
r2 :=
r3 :=
x1 ∈ {a, b}
→
x1 ∈ {a, b}, x2 ∈ {b, c} →
x2 ∈ {b}
→
x2 =
6 a, x4 =
6 b,
x3 =
6 a,
x3 =
6 a, x4 =
6 b.
Then upon application of rule r1 rule r2 can be applied without evaluating its
condition and subsequently rule r3 can be deleted without applying it. So we can
put rule r2 into friends(r1 ) and rule r3 into obviated (r1 ) and this is in fact what
the F & O algorithm does.
4 Implementation
In this section we discuss the implementation of the R algorithm for the membership
rules and compare it by means of various benchmarks with the CHR implementation
in the ECLi PSe system.
4.1 Modelling of the Membership Rules in CHR
Following (Apt and Monfroy 2001) the membership rules are represented as CHR
propagation rules with one head. Recall that the latter ones are of the form
H ==> G1 , . . . , Gl | B1 , . . . , Bm .
where
• l > 0, m > 0,
• the atom H of the head refers to the defined constraints,
• the atoms of the guard G1 , . . . , Gl refer to Prolog relations or built-in constraints,
14
• the atoms of the body B1 , . . . , Bm are arbitrary atoms.
Further, recall that the CHR propagation rules with one head are executed as
follows. First, given a query (that represents a CSP) the variables of the rule are
renamed to avoid variable clashes. Then an attempt is made to match the head of
the rule against the first atom of the query. If it is successful and all guards of the
instantiated version of the rule succeed, the instantiated version of the body of the
rule is executed. Otherwise the next rule is tried.
Finally, let us recall the representation of a membership rule as CHR a propagation
rule used in (Apt and Monfroy 2001). Consider the membership rule
y1 ∈ S1 , . . . , yk ∈ Sk → z1 6= a1 , . . . , zm 6= am .
related to the constraint c on the variables X1 , . . . , Xn . We represent its condition
by starting initially with the atom c(X1 , . . . , Xn ) as the head. Each atomic condition
of the form yi ∈ {a} is processed by replacing in the atom c(X1 , . . . , Xn ) the
variable yi by a. In turn, each atomic condition of the form yi ∈ Si , where Si is
not a singleton, is processed by adding the atom in(yi , LSi ) to the guard of the
propagation rule. The in/2 predicate is defined by
in(X,L) :- dom(X,D), subset(D,L).
So in(X,L) holds if the current domain of the variable X (yielded by the built-in
dom of ECLi PSe ) is included in the list L. In turn, LSi is a list representation of the
set Si .
Finally, each atomic conclusion zi 6= ai translates to the atom zi ## ai of the
body of the propagation rule.
As an example consider the membership rule
X ∈ {0}, Y ∈ {1, 2} → Z 6= 2
in presence of a constraint c on the variables X, Y, Z. It is represented by the
following CHR propagation rule:
c(0,Y,Z) ==> in(Y,[1,2]) | Z##2.
In ECLi PSe the variables with singleton domains are automatically instantiated.
So, assuming that the variable domains are non-empty, the condition of this membership rule holds iff the head of the renamed version of the above propagation rule
matches the atom c(0,Y,Z) and the current domain of the variable Y is included
in [1,2]. Further, in both cases the execution of the body leads to the removal of
the value 2 from the domain of Z. So the execution of both rules has the same effect
when the variable domains are non-empty.
Execution of CHR. In general, the application of a membership rule as defined
in Section 3 and the execution of its representation as a CHR propagation rules
coincide. Moreover, by the semantics of CHR, the CHR rules are repeatedly applied
until a fixpoint is reached. So a repeated application of a finite set of membership
rules coincides with the execution of the CHR program formed by the representations
15
of these membership rules as propagation rules. An important point concerning the
standard execution of a CHR program is that, in contrast to the R algorithm, every
change in the variable domains of a constraint causes the computation to restart.
4.2 Benchmarks
In our approach the repeated application of a finite set of membership rules is
realized by means of the R algorithm of Section 2 implemented in ECLi PSe . The
compiler consists of about 1500 lines of code. It accepts as input a set of membership
rules, each represented as a CHR propagation rule, and constructs an ECLi PSe
program that is the instantiation of the R algorithm for this set of rules. As in CHR,
for each constraint the set of rules that refer to it is scheduled separately.
In the benchmarks below for each considered CSP we used the sets of all minimal
valid membership and equality rules for the “base” constraints which were automatically generated using a program discussed in (Apt and Monfroy 2001). In the first
phase the compiler constructs for each rule g the lists friends(g) and obviated (g).
Time spent on this construction is comparable with the time needed for the generation of the minimal valid equality and membership rules for a given constraint. For
example, the medium-sized membership rule set for the rcc8 constraint, containing
912 rules, was generated in 166 seconds while the construction of all friends and
obviated lists took 142 seconds.
To see the impact of the accumulated savings obtained by permanent removal of
the rules during the iteration process, we chose benchmarks that embody several
successive propagation steps, i.e., propagation interleaved with domain splitting or
labelling.
In Table 1 we list the results for selected single constraints. For each such
constraint, say C on a sequence of variables x1 , . . . , xn with respective domains
D1 , . . . , Dn , we consider the CSP hC ; x1 ∈ D1 , . . . , xn ∈ Dn i together with randomized labelling. That is, the choices of variable, value, and action (assigning or
removing the value), are all random. The computation of simply one or all solutions
yields insignificant times, so the benchmark program computes and records also all
intermediate non-solution fixpoints. Backtracking occurs if a recorded fixpoint is encountered again. In essence, this benchmark computes implicitly all possible search
trees. As this takes too much time for some constraints, we also impose a limit on
the number of recorded fixpoints.
In turn, in Table 2 we list the results for selected CSPs. We chose here CSPs that
formalize sequential automatic test pattern generation for digital circuits (ATPG),
see (Brand 2001). These are rather large CSPs that employ the and constraints of
Table 1 and a number of other constraints, most of which are implemented by rules.
We measured the execution times for three rule schedulers: the standard CHR
representation of the rules, the generic chaotic iteration algorithm GI, and its improved derivative R. The codes of the latter two algorithms are both produced by
our compiler and are structurally equal, hence allow a direct assessment of the
improvements embodied in R.
An important point in the implementations is the question of when to remove
16
Constraint
rcc8
fork
and3
and9
and11
37% / 22%
58% / 46%
66% / 49%
26% / 15%
57% / 25%
membership
relative
absolute 147/396/686 0.36/0.62/0.78 0.27/0.41/0.55 449/1727/2940 1874/3321/7615
equality
relative
97% / 100%
98% / 94%
92% / 59%
absolute 359/368/359 21.6/21.9/22.9 0.36/0.39/0.61
95% / 100%
96% / 101%
386/407/385
342/355/338
Table 1. Randomized search trees for single constraints
Logic
3-valued
9-valued
11-valued
membership
relative
61% / 44%
65% / 29%
73% / 29%
absolute
1.37/2.23/3.09
111/172/385
713/982/2495
relative
63% / 29%
40% / 57%
36% / 51%
absolute
0.77/1.22/2.70
2.56/6.39/4.50
13.8/38.7/26.7
equality
Table 2. CSPs formalizing sequential ATPG
solved constraints from the constraint store. The standard CHR representation of
membership rules as generated by the algorithm of (Apt and Monfroy 2001) does
so by containing, beside the propagation rules, one CHR simplification rule for each
tuple in the constraint definition. Once its variables are assigned values that correspond to a tuple, the constraint is solved, and removed from the store by the
corresponding simplification rule. This ‘solved’ test takes place interleaved with
propagation. The implementations of GI and R, on the other hand, check after
closure under the propagation rules. The constraint is considered solved if all its
variables are fixed, or, in the case of R, if the set F of remaining rules is empty (see
the following subsection). Interestingly, comparing CHR and GI, the extra simplification rules sometimes constitute a substantial overhead while at other times their
presence allows earlier termination.
We mention briefly that our specific implementation deviates slightly from the
description of R inside the else branch. The test ∀e ⊒ d ¬holds(b, e) in the case
of a membership condition y ∈ S corresponds to testing whether the intersection
17
Dy ∩ S is empty. Performing this always turned out to be more costly than doing
so only when Dy is a singleton set.
The platform for all benchmarks was a Sun Enterprise 450 with 4 UltraSPARCII 400 MHz processors and 2 GB memory under Solaris, and ECLi PSe 5.5 (nonparallel). In the tables we provide for each constraint or CSP the ratio of the
execution times in seconds between, first, R and GI, and second, R and CHR. This is
followed by the absolute times in the order R / GI / CHR.
Recently, we have been experimenting with various ways of optimizing our implementation of the R algorithm. In particular, we considered a better embedding
into the constraint-handling mechanism of ECLi PSe , for instance by finer control
of the waking conditions and a joint removal of the elements from the same domain.
At this stage we succeeded in achieving an average speed-up by a factor of 4. This
work is in progress but already shows that further improvements are possible.
4.3 Recomputing of the Least Fixpoints
Finally, let us illustrate the impact of the permanent removal of the rules
during the least fixpoint computation, achieved here by the use of the lists
friends(g) and obviated (g). Given a set F of rules call a rule g ∈ F solving if
friends(g) ∪ obviated (g) = F .
Take now as an example the equivalence relation ≡ from three valued logic of
(Kleene 1952, page 334) that consists of three values, t (true), f (false) and u (unknown). It is defined by the truth table
≡
t
f
u
t
f
u
t
f
u
f
t
u
u
u
u
The program of (Apt and Monfroy 2001) generates for it 26 minimal valid membership rules. Out of them 12 are solving rules. For the remaining rules the sizes of
the set friends ∪ obviated are: 17 (for 8 rules), 14 (for 4 rules), and 6 (for 2 rules).
In the R algorithm a selection of a solving rule leads directly to the termination
(G = ∅) and to a reduction of the set F to ∅. For other rules, also a considerable
simplification in the computation takes place. For example, one of the 8 rules with
17 rules in the set friends ∪ obviated is
r := x ∈ {f}, z ∈ {f, u} → y 6= f.
Consider now the CSP h≡ ; x ∈ {f}, y ∈ {f, t, u}, z ∈ {f, u}i. In the R algorithm the
selection of r is followed by the application of the rules from friends and the removal
of the rules from friends ∪ obviated . This brings the number of the considered rules
down to 26−17 = 9. The R algorithm subsequently discovers that none of these nine
rules is applicable at this point, so this set F remains upon termination. Then in a
18
subsequent constraint propagation phase, launched after splitting or after constraint
propagation involving another constraint, the fixpoint computation by means of the
R algorithm involves only these nine rules instead of the initial set of 26 rules. For
solving rules this fixpoint computation immediately terminates.
Interestingly, as Table 3 shows, the solving rules occur quite frequently. We list
there for each constraint and each type of rules the number of solving rules divided
by the total number of rules, followed in a new line by the average number of rules
in the set friends(g) ∪ obviated (g).
Constraints and2
and3
and9
and11
fork
rcc8
allen
equality
6/6
6
13/16
14
113/134
130
129/153
148
9/12
11
183/183
183
498/498
498
membership
6/6
6
4/13
7
72/1294
810
196/4656
3156
0/24
9
0/912
556
n.a./26446
n.a.
Table 3. Solving rules
The fork constraint is taken from the Waltz language for the analysis of polyhedral scenes. The rcc8 is the composition table for the Region Connection Calculus
with 8 relations from (Egenhofer 1991). It is remarkable that all its 183 minimal
valid equality rules are solving. While none of its 912 minimal valid membership rule
for rcc8 is solving, on the average the set friends(g)∪obviated (g) contains 556 membership rules. Also all 498 minimal valid equality rules for the allen constraint,
that represents the composition table for Allen’s qualitative temporal reasoning,
are solving. The number of minimal valid membership rules exceeds 26,000 and
consequently they are too costly to analyze.
Simplification rules. The CHR language supports besides propagation rules also
so-called simplification rules. Using them one can remove constraints from the
constraint store, so one can affect its form. In (Abdennadher and Rigotti 2001)
a method is discussed that allows one to automatically transform CHR propagation
rules into simplification rules that respects their semantics. It is based on identifying
or constructing propagation rules that are solving.
In contrast, our method captures the degree to which a rule is solving, by the
ratio of the sizes of U (r) = friends(r) ∪ obviated (r) and the full rule set. If the
sets are equal, then the ratio is 1 and r is a solving rule. Consider now two nonsolving rules r1 , r2 , that means with U (r1 ) ⊂ R and U (r2 ) ⊂ R, but let also
U (r1 ) ∪ U (r2 ) = R. Suppose that during a fixpoint computation the conditions of
both rules have succeeded, and their bodies have been applied. The R algorithm
would now immediately detect that the constraint is solved, and consequently terminate. CHR, for which r1 and r2 are ordinary (propagation) rules, cannot detect
this possibility for immediate termination.
19
5 Redundancy of prop Rules
The cost of a fixpoint computation by the GI algorithm or one of its derivatives
depends on the number of functions involved, in particular in absence of a good
strategy for selecting the functions, represented in the algorithms by the “choose”
predicate. It is therefore important to identify functions or rules that are not needed
for computing fixpoints. In the following we shall examine the issue of rule redundancy. We shall again start with arbitrary functions before moving on to (prop)
rules. The redundancy concept we employ is based on fixpoints. In the following,
for brevity, we drop the word “common” when referring to common fixpoints of a
set of functions.
Definition 5.1
• Consider a set F ∪ {f } of functions on a partial ordering. A function f is
called redundant with respect to F if the sets of fixpoints of F and F ∪ {f }
are equal.
• A set of functions F is called minimal with respect to redundancy (or simply
minimal ), if no function f ∈ F is redundant with respect to F − {f }.
Equivalently, we can say that a function f is redundant w.r.t. F if every fixpoint
of F is also a fixpoint of f .
5.1 Redundant Rules
We now focus on the subject of redundancy for prop rules. The following simple
test is then useful.
Theorem 5.2
Consider a set F of prop rules and a prop rule r := b → g with the witness w for
b. Let e be the least fixpoint of F greater than or equal to w. If g(e) = e, then the
rule r is redundant with respect to F .
Proof. We show that g(e) = e implies that an arbitrary fixpoint d of F is a fixpoint
of r by a case condition.
b holds for d: We have w ⊑ d since w is the witness for b. Also, w ⊑ e ⊑ d since
e is the least fixpoint of F greater than or equal to w. From e ⊑ d, g(e) = e, and
the stability of g we conclude g(d) = d. Hence r(d) = (b → g)(d) = g(d) = d.
b does not hold for d: Then r(d) = (b → g)(d) = d.
This test is of interest to us since it allows us to compute only one fixpoint of F
instead of all fixpoints. It is effective if
• the witness can be computed,
• the equality g(e) = e can be determined, and
• the fixpoint computations are effective.
For the sake of fixpoint computations a rule r = b → g with a body g = g1 , . . . , gn
(describing a function composition) such that any two different gi , gj commute can
20
be identified with the collection b → g1 , . . . , b → gn of the rules, and vice versa.
Indeed, the respective fixpoints and the rule properties are the same. We consider
here these two representations as equivalent. If a rule with such a “compound” body
is not redundant it might be so in part. That is, some part of its body might be
redundant or, in other words, some sub-rules of its decomposition might be. This
is what we mean below by saying that a rule is partially redundant.
Let us consider now the task of computing minimal sets of prop rules. Such
sets can of course be generated by a simple bounded loop: select an untested rule,
test whether it is redundant and, if so, remove it from the current set. In general,
however, the obtained minimal sets depend on the selection order for testing; see
an example below. In our experiments we used a strategy that selects first the rules
the execution of which is computationally expensive, for instance due to conditions
on many variables. In this way we hope to obtain a set of computationally cheap
rules.
5.2 An Example: Redundant Membership Rules
Let us illustrate now a number of issues by means of an example. Consider the
constraint c(x, y, z, u) defined by
x y
z
u
0
1
1
0
0
1
1
1
0
1
0
1
The underlying domain for all its variables is {0, 1}. Hence the induced corresponding partial order is
({(A, B, C, D) | A, B, C, D ⊆ {0, 1}}, ⊇).
The algorithm of (Apt and Monfroy 2001) generates eleven membership rules listed
in Figure 1. Since the rule conditions are only equality tests, we use an alternative
notation that should be self-explanatory.
For example, rule (11) states that if c(x, y, z, u), then it is correct to conclude
from x = 1 and u = 1 that y 6= 1 (validity), and furthermore that neither x = 1
nor u = 1 suffices for this conclusion (minimality).
Suppose we are interested in computing the smallest fixpoint greater than or
equal to E1 = {1} × {0, 1} × {0, 1} × {1}. Suppose rule (11) is considered. Its
application yields E2 = {1} × {0} × {0, 1} × {1} from where rule (4) leads to
E3 = {1} × {0} × {0} × {1}. This is indeed a fixpoint since for each rule either its
condition does not apply or the application of its body results again in E3 .
A second possible iteration from E1 that stabilises in E3 is by rule (5) followed by
rule (10). Rule (11) can be applied at this point but its body does not change E3 .
21
c(x, y, z, 0)
→
x 6= 0, y 6= 0, z 6= 0
(1)
c(x, y, 1, u)
→
u 6= 1, x 6= 0, y 6= 0
(2)
c(0, y, z, u)
→
u 6= 0, y 6= 0, z 6= 1
(3)
c(x, 0, z, u)
→
u 6= 0, x 6= 0, z 6= 1
(4)
c(x, y, z, 1)
→
z 6= 1
(5)
c(x, y, 0, u)
→
u 6= 0
(6)
c(1, 1, z, u)
→
u 6= 1, z 6= 0
(7)
c(x, 1, 0, u)
→
x 6= 1
(8)
c(x, 1, z, 1)
→
x 6= 1
(9)
c(1, y, 0, u)
→
y 6= 1
(10)
c(1, y, z, 1)
→
y 6= 1
(11)
Fig. 1. Membership rules for the constraint c
Indeed, E3 is a fixpoint of all rules including rule (11). We conclude that rule (11)
is redundant – we just performed the test of Theorem 5.2.
The process of identifying redundant rules can then be continued for the rule
set {(1), . . . , (10)}. One possible outcome is depicted in Figure 1, where redundant
parts of rules are underlined. From the 20 initial atomic conclusions 13 remain, thus
we find here a redundancy ratio of 35%.
Consider now the justification for the redundancy of rule (11), and observe that
rule (11) has no effect since rule (10), which has the same body, was applied before.
Suppose now that the process of redundancy identification is started with rule
(10) instead of rule (11). This results in identifying rule (10) as redundant, with a
relevant application of rule (11).
Note moreover that one of the rules (10), (11) must be present in any minimal
set since their common body y 6= 1 occurs in no other rule. It would seem difficult
to find a criterion that prefers one rule over the other as their structure is the same.
5.3 Experiments
We implemented in ECLi PSe an algorithm that computes minimal sets of membership rules. The results for some benchmark rule sets are listed in Table 4.
For each constraint the set of minimal membership or equality rules (indicated
respectively by the subscript “M ” or “E ”) was computed by the rule generation
algorithm of (Apt and Monfroy 2001). The constraints are taken from the experiments discussed in Table 1. Additionally a 5-ary constraint fula (standing for the
well-known fulladder constraint) is analyzed.
The table shows the size of the rule set, the number of fully and, in parentheses,
partially redundant rules. The redundancy ratio for the entire rule set shows the
percentage of the atomic disequalities that are removed from the rule conclusions
on the account of redundancy.
22
total
redundant
(partially)
redundancy
ratio
and11M
and11E
and3M
equ3M
fula2E
forkE
forkM
4656
153
18
26
52
12
24
4263 (2)
0 (6)
5 (0)
8 (0)
24 (0)
0 (9)
6 (6)
81%
4%
30%
26%
35%
35%
40%
Table 4. Minimizing rule sets
Computation times are negligible in so far as they are considerably smaller than
the corresponding rule generation times.
5.4 Schedulers and Minimal Rule Sets
There is no simple connection between redundancy and the rule sets friends and
obviated of the R scheduler. For instance, it is not the case that a rule is redundant
if it is contained in friends(r) ∪ obviated (r) of every rule r. Nor is a redundant rule
necessarily contained in friends(r) ∪ obviated (r) of every rule r. To examine this in
an example, recall the rules in Figure 1. All except (5) and (6) are solving rules, i.e.,
each respective set friends ∪ obviated is the complete set {(1), . . . , (11)} of rules,
while for rules (5) and (6) this set is {(1), (3), (5), (6)}. Further, neither (5) nor (6)
is redundant with respect to all other rules, whereas (10) and (11) are.
Benchmarks. We reran the benchmarks from Tables 1 and 2 with all involved rule
sets subjected to a removal of redundant rules and subsequent recomputation of
the sets friends and obviated . The results are reported in Tables 5 and 6 below.
The rule sets of rcc8 were already minimal; therefore this constraint is omitted.
Constraint
fork
and3
and9
and11
membership
relative
60% / 46%
69% / 48%
28% / 18%
50% / 29%
absolute
0.32/0.53/0.70
0.27/0.39/0.56
167/589/924
157/316/543
equality
relative
97% / 93%
97% / 64%
96% / 101%
96% / 101%
absolute
21.6/22.2/23.2
0.37/0.38/0.58
386/404/384
341/353/339
Table 5. Randomized search trees for single constraints (without redundant rules)
23
3-valued
9-valued
11-valued
relative
66% / 46%
62% / 33%
68% / 35%
absolute
1.32/2.00/3.05
37/59/114
70/103/199
relative
61% / 26%
40% / 58%
33% / 48%
absolute
0.72/1.18/2.73
2.57/6.41/4.46
13.8/41.0/28.6
Logic
membership
equality
Table 6. CSPs formalizing sequential ATPG (without redundant rules)
When comparing the redundancy and non-redundancy benchmarks versions we
observe that the absolute execution times are enormously reduced in the case of
the constraints on higher-valued logics. This is in line with the much smaller sizes
of the reduced rule sets. The ratios of the execution times, however, are barely
affected. The type of a scheduler and minimality w.r.t. redundancy appear to be
rather orthogonal issues.
It is interesting to examine in one case the distribution of the solving degrees,
i.e., the ratios of the sizes of friends ∪ obviated and the full rule set. Recall that a
ratio of 1 means that the constraint is solved when the rule body has been executed.
Such a rule could be represented as a simplification rule in CHR (see Section 4.3).
In Figure 2 two membership rule sets for the constraint and9 are compared.
One set contains redundant rules, the other set is minimal w.r.t. redundancy. The
rules in the minimal set are solving to a lesser degree. In particular, none is a
proper solving rule. The good performance of the R algorithm in the benchmarks
of Tables 5,6 may thus be attributed not to distinguishing solving (simplification)
rules and non-solving propagation rules, but to the accumulated effect of removing
rules from the fixpoint computation.
385 nonredundant rules
1
0.8
0.8
Solving degree
Solving degree
1294 rules, some redundant
1
0.6
0.4
0.2
0
0.6
0.4
0.2
0
0
200
400
600 800 1000 1200
Rule
50
Fig. 2. and9M : Solving degree and redundancy
24
100 150 200 250 300 350
Rule
Acknowledgments
We thank Christian Holzbaur and Eric Monfroy for helpful discussions on the implementation and on an early version of this paper. The reviewers made useful
comments which helped to improve the paper.
References
Abdennadher, S., Krämer, E., Saft, M., and Schmaus, M. 2001. JACK: A Java
Constraint Kit. In International Workshop on Functional and (Constraint) Logic Programming (WFLP 2001), Technical Report No. 2017. University of Kiel, Kiel, Germany.
Abdennadher, S. and Rigotti, C. 2001. Using confluence to generate rule-based constraint solvers. In Proceedings of the 3rd Int. Conf. on Principles and Practice of Declarative Programming (PPDP 2001). ACM, Firenze, Italy, 127–135.
Apt, K. R. 1999. The essence of constraint propagation. Theoretical Computer Science 221, 1–2, 179–210. Available via http://arXiv.org/archive/cs/.
Apt, K. R. 2000. The role of commutativity in constraint propagation algorithms. ACM
Transactions on Programming Languages and Systems 22, 6, 1002–1036. Available via
http://arXiv.org/archive/cs/.
Apt, K. R. and Monfroy, E. 2001. Constraint programming viewed as rule-based
programming. Theory and Practice of Logic Programming 1, 6, 713–750. Available via
http://arXiv.org/archive/cs/.
Brand, S. 2001. Sequential automatic test pattern generation by constraint programming.
In CP2001 Post Conference Workshop Modelling and Problem Formulation. Available
via http://homepages.cwi.nl/~sbrand/.
Egenhofer, M. 1991. Reasoning about binary topological relations. In Proceedings of the
2nd International Symposium on Large Spatial Databases (SSD), O. Günther and H.-J.
Schek, Eds. Lecture Notes in Computer Science, vol. 525. Springer-Verlag, 143–160.
Frühwirth, T. 1998. Theory and practice of Constraint Handling Rules. Journal of
Logic Programming 37, 1–3 (October), 95–138. Special Issue on Constraint Logic Programming (P. J. Stuckey and K. Marriot, Eds.).
Holzbaur, C., de la Banda, M. G., Jeffery, D., and Stuckey, P. J. 2001. Optimizing compilation of constraint handling rules. In Proceedings of the 2001 International Conference on Logic Programming. Lecture Notes in Computer Science, vol. 2237.
Springer-Verlag, 74–89.
Kleene, S. C. 1952. Introduction to Metamathematics. van Nostrand, New York.
25
| 2cs.AI
|
CROSS-MODAL EMBEDDINGS FOR VIDEO AND AUDIO RETRIEVAL
Didac Surı́s1 , Amanda Duarte2 , Amaia Salvador1 , Jordi Torres2 and Xavier Giró-i-Nieto1
1
2
Universitat Politècnica de Catalunya (UPC)
Barcelona Supercomputing Center (BSC-CNS)
arXiv:1801.02200v1 [cs.IR] 7 Jan 2018
ABSTRACT
The increasing amount of online videos brings several opportunities for training self-supervised neural networks. The
creation of large scale datasets of videos such as the YouTube8M allows us to deal with this large amount of data in manageable way. In this work, we find new ways of exploiting
this dataset by taking advantage of the multi-modal information it provides. By means of a neural network, we are able
to create links between audio and visual documents, by projecting them into a common region of the feature space, obtaining joint audio-visual embeddings. These links are used
to retrieve audio samples that fit well to a given silent video,
and also to retrieve images that match a given a query audio.
The results in terms of Recall@K obtained over a subset of
YouTube-8M videos show the potential of this unsupervised
approach for cross-modal feature learning. We train embeddings for both scales and assess their quality in a retrieval
problem, formulated as using the feature extracted from one
modality to retrieve the most similar videos based on the features computed in the other modality.
Index Terms— Sonorization, embedding, retrieval,
cross-modal, YouTube-8M
1. INTRODUCTION
Videos have become the next frontier in artificial intelligence.
The rich semantics contained in them make them a challenging data type posing several challenges in both perceptual,
reasoning or even computational level. Mimicking the learning process and knowledge extraction that humans develop
from our visual and audio perception remains an open research question, and video contain all this information in a
format manageable for science and research.
Videos are used in this work for two main reasons. Firstly,
they naturally integrate both visual and audio data, providing
a weak labeling of one modality with respect to the other. Secondly, the high volume of both visual and audio data allows
training machine learning algorithms whose models are governed by a high amount of parameters. The huge scale video
archives available online and the increasing number of video
cameras that constantly monitor our world, offer more data
than computation power available to process them.
The popularization of deep neural networks among the
computer vision and audio communities has defined a common framework boosting multimodal research. Tasks like
video sonorization, speaker impersonation or self-supervised
feature learning have exploited the opportunities offered by
artificial neurons to project images, text and audio in a feature space where bridges across modalities can be built.
This work exploits the relation between the visual and
audio contents in a video clip to learn a joint embedding
space with deep neural networks. Two multilayer perceptrons
(MLPs), one for visual features and a second one for audio
features, are trained to be mapped into the same cross-modal
representation. We adopt a self-supervised approach, as we
exploit the unsupervised correspondence between the audio
and visual tracks in any video clip.
We propose a joint audiovisual space to address a retrieval
task formulating a query from any of the two modalities. As
depicted in Figure 1, whether a video or an audio clip can be
used as a query to search its matching pair in a large collection
of videos. For example, an animated GIF could be sonorized
by finding an adequate audio track, or an audio recording illustrated with a related video.
In this paper, we present a simple yet very effective model
for retrieving documents with a fast and light search. We do
not address an exact alignment between the two modalities
that would require a much higher computation effort.
The paper is structured as follows. Section 2 introduces
the related work on learned audiovisual embeddings with neural networks. Section 3 presents the architecture of our model
and Section 4 how it was trained. Experiments are reported
in Section 5 and final conclusions drawn in Section 6. The
source code and trained model used in this paper is publicly available from https://github.com/surisdi/
youtube-8m.
2. RELATED WORK
In the past years, the relationship between the audio and the
visual content in videos has been researched in several contexts. Overall, conventional approaches can be divided into
four categories according to the task: generation, classification, matching and retrieval.
As online music streaming and video sharing websites
Fig. 1. A learned cross-modal embedding allows retrieving
video from audio, and vice versa.
have become increasingly popular, some research has been
done on the relationship between music and album covers
[1, 2, 3, 4] and also on music and videos (instead of just images) as the visual modality [5, 6, 7, 8] to explore the multimodal information present in both types of data.
A recent study [9] also explored the cross-modal relations
between the two modalities but using images with people talking and speech. It is done through Canonical Correlation
Analysis (CCA) and cross-modal factor analysis. Also applying CCA, [10] uses visual and sound features and common subspace features for aiding clustering in image-audio
datasets. In a work presented by [11], the key idea was to use
greedy layer-wise training with Restricted Boltzmann Machines (RBMs) between vision and sound.
The present work is focused on using the information
present in each modality to create a joint embedding space to
perform cross-modal retrieval. This idea has been exploited
especially using text and image joint embeddings [12, 13, 14],
but also between other kinds of data, for example creating a
visual-semantic embedding [15] or using synchronous data
to learn discriminative representations shared across vision,
sound and text [16].
However, joint representations between the images
(frames) of a video and its audio have yet to be fully exploited,
being [17] the work that most has explored this option up to
the knowledge of the authors. In their paper, they seek for a
joint embedding space but only using music videos to obtain
the closest and farthest video given a query video, only based
on either image or audio.
The main idea of the current work is borrowed from [14],
which is the baseline to understand our approach. There, the
authors create a joint embedding space for recipes and their
images. They can then use it to retrieve recipes from any food
image, looking to the recipe that has the closest embedding.
Apart from the retrieval results, they also perform other experiments, such as studying the localized unit activations, or
doing arithmetics with the images.
images of the video and the vector of features representing
the audio. These features are already precomputed and provided in the YouTube-8M dataset [18]. In particular, we use
the video-level features, which represent the whole video clip
with two vectors: one for the audio and another one for the
video. These feature representations are the result of an average pooling of the local audio features computed over windows of one second, and local visual features computed over
frames sampled at 1 Hz.
The main objective of the system is to transform the two
different features (image and audio, separately) to other features laying in a joint space. This means that for the same
video, ideally the image features and the audio features will
be transformed to the same joint features, in the same space.
We will call these new features embeddings, and will represent them with Φi , for the image embeddings, and Φa , for the
audio embeddings.
The idea of the joint space is to represent the concept of
the video, not just the image or the audio, but a generalization
of it. As a consequence, videos with similar concepts will
have closer embeddings and videos with different concepts
will have embeddings further apart in the joint space. For
example, the representation of a tennis match video will be
close to the one of a football match, but not to the one of a
maths lesson.
Thus, we use a set of fully connected layers of different
sizes, stacked one after the other, going from the original features to the embeddings. The audio and the image network
are completely separated. These fully connected layers perform a non-linear transformation on the input features, mapping them to the embeddings, being the parameters of this
non-linear mapping learned in the optimization process.
After that, a classification from the two embeddings is
done, also using a fully connected layer from them to the different classes, using a sigmoid as activation function. We will
get more insight on this step in section 4.
The number of hidden layers is not necessarily fixed, as
well as the number of neurons per layer, since we experimented with different configurations. Each hidden layer uses
ReLu as activation function, and all the weights in each layer
are regularized using L2 norm.
4. TRAINING
In this section we present the used losses as well as their
meaning and intuition.
4.1. Similarity Loss
3. ARCHITECTURE
In this section we present the architecture for our joint embedding model, which is depicted in the Figure 2.
As inputs, we have the vector of features representing the
The objective of this work is to get the two embeddings of the
same video to be as close as possible (ideally, the same), while
keeping embeddings from different videos as far as possible.
Formally, we are given a video vk , represented by the audio and visual features vk = {ik , ak } (ik represents the image
Fig. 2. Schematic of the used architecture.
features and ak the audio features of vk ). The objective is to
maximize the similarity between Φik , the embedding obtained
by transformations on ik , and Φak , the embedding obtained by
transformations on ak .
At the same time, however, we have to prevent embeddings from different videos to be “close” in the joint space. In
other words, we want them to have low similarity. However,
the objective is not to force them to be opposite to each other.
Instead of forcing them to have similarity equal to zero, we
allow a margin of similarity small enough to force the embeddings to be clearly not in the same place in in the joint
space. We call this margin α.
During the training, both positive and negative pairs are
used, being the positive pairs the ones for which ik and ak
correspond to the same video vk , and the negative pairs the
ones for which ik1 and ak2 do not correspond to the same
video, this is, k1 6= k2. The proportion of negative samples
is pnegative .
For the negative pairs, we selected random pairs that did
not have any common label, in order to help the network to
learn how to distinguish different videos in the embedding
space. The notion of “similarity” or “closeness” is mathematically translated into a cosine similarity between the embeddings, being the cosine similarity definedNas:
P
xk zk
k=1
s
similarity = cos(x, z) = s
(1)
N
N
P
P
2
2
xk
zk
k
i
for any pair of real-valued vectors x and z.
From this reasoning we get to the first and most important
loss:
Lcos ((Φa ,Φi ), y) =
1 − cos(Φa , Φi ),
if
=
max(0, cos(Φa , Φi ) − α), if
y=1
y = −1
(2)
where y = 1 denotes positive sampling, and y = −1
denotes negative sampling.
4.2. Classification Regularization
Inspired by the work presented in [14], we provide additional
information to our system by incorporating the video labels
(classes) provided by the YouTube-8M dataset. This information is added as a regularization term that seeks to solve
the high-level classification problem, both from the audio and
from the video embeddings, sharing the weights between the
two branches. The key idea here is to have the classification
weights from the embeddings to the labels shared by the two
modalities.
This loss is optimized together with the previously explained similarity loss, serving as a regularization term. Basically, the system learns to classify the audio and the images of
a video (separately) into different classes or labels provided
by the dataset. We limit its effect by using a regularization
parameter λ.
To incorporate the previously explained regularization to
the joint embedding, we use a single fully connected layer, as
shown in Figure 2. Formally, we can obtain the label probabilities as pi = softmax(W Φi ) and pa = softmax(W Φa ),
where W represents the learned weights, which are shared
between the two branches. The softmax activation is used in
order to obtain probabilities at the output. The objective is to
make pi as similar as possible to ci , and pa as similar as possible to ca , where ci and ca are the category labels for the video
represented by the image features and the audio features, respectively. For positive pairs, ci and ca are the same.
The loss function used for the classification is the well
known cross entropy loss:
L(x, z) = −
X
k
xk log(zk )
(3)
Thus, the classification loss is:
X
Lclass (pi , pa , ci , ca ) = −
(pik log(cik )+(pak log(cak )) (4)
k
Finally, the loss function to be optimized is:
L = Lcos + λLclass
(5)
4.3. Parameters and Implementation Details
For our experiments we used the following parameters:
• Batch size of 1024.
• We saw that starting with λ different than zero led to
a bad embedding similarity because the classification
accuracy was preferred. Thus, we began the training
with λ = 0 and set it to 0.02 at step number 10,000.
• Margin α = 0.2.
• Percentage of negative samples pnegative = 0.6.
• 4 hidden layers in each network branch, the number of
neurons per layer being, from features to embedding,
2000, 2000, 700, 700 in the image branch, and 450,
450, 200, 200 in the audio branch.
• Dimensionality of the feature vector = 250.
• We trained a single epoch.
The simulation was programmed using Tensorflow [19],
having as a baseline the code provided by the YouTube-8M
challenge authors1 .
5. RESULTS
5.1. Dataset
The experiments presented in this section were developed
over a subset of 6,000 video clips from the YouTube-8M
dataset [18]. This dataset does not contain the raw video files,
but their representations as precomputed features, both from
audio and video. Audio features were computed using the
method explained in [20] over audio windows of 1 second,
while visual features were computed over frames sampled at
1 Hz with the Inception model provided in TensorFlow [19].
The dataset provides video-level features, which represent
all the video using a single vector (one for audio and another
for visual information), and thus does not maintain temporal information; and also provides frame-level features, which
consist on a single vector representing each second of audio,
and a single vector representing each frame of the video, sampled at 1 frame per second.
The main goal of this dataset is to provide enough data to
reach state of the art results in video classification. Nevertheless, such a huge dataset also permits approaching other tasks
related to videos and cross-modal tasks, such as the one we
1 https://www.kaggle.com/c/youtube8m
Table 1. Evaluation of Recall from audio to video
Number of elements Recall@1 Recall@5 Recall@10
256
21.5%
52.0%
63.1%
512
15.2%
39.5%
52.0%
1024
9.8%
30.4%
39.6%
Table 2. Evaluation of Recall from video to audio
Number of elements Recall@1 Recall@5 Recall@10
256
22.3%
51.7%
64.4%
512
14.7%
38.0%
51.5%
1024
10.2%
29.1%
40.3%
approach in this paper. For this work, and as a baseline, we
only use the video-level features.
5.2. Quantitative Performance Evaluation
We divide our results in two different categories: quantitative
(numeric) results and qualitative results.
To obtain the quantitative results we use the Recall@k
metric. We define Recall@k as the recall rate at top K for
all the retrieval experiments, this is, the percentage of all the
queries where the corresponding video is retrieved in the top
K, hence higher is better.
The experiments are performed with different dimension
of the feature vector. The Table 1 shows the results of recall from audio to video. In other words, from the audio embedding of a video, how many times we retrieve the embedding corresponding to the images of that same video. Table 2
shows the recall from video to audio.
To have a reference, the random guess result would be
k/Number of elements. The obtained results show a very
clear correspondence between the embeddings coming from
the audio features and the ones coming from the video features. It is also interesting to notice that the results from audio
to video and from video to audio are very similar, because the
system has been trained bidirectionally.
5.3. Qualitative Performance Evaluation
In addition to the objective results, we performed some insightful qualitative experiments. They consisted on generating the embeddings of both the audio and the video for a list
of 6,000 different videos. Then, we randomly chose a video,
and from its image embedding, we retrieved the video with
the closest audio embedding, and the other way around (from
one video’s audio we retrieved the video with the closest image embedding). If the closest embedding corresponded to
the same video, we took the second one in the ordered list.
The Figure 3 shows some experiments. On the left, we
can see the results given a video query and getting the closest
audio; and on the right the input query is an audio. Examples depicting the real videos and audio are available online
Fig. 3. Qualitative results. On the left we show the results obtained when we gave a video as a query. On the right, the results
are based on an audio as a query.
2
. It shows both the results when going from image to audio,
and when going from audio to image. Four different random
examples are shown in each case. For each result and each
query, we also show their YouTube-8M labels, for completeness.
The results show that when starting from the image features of a video, the retrieved audio represents a very accurate
fit for those images. Subjectively, there are non negligible
cases where the retrieved audio actually fits better the video
than the original one, for example when the original video has
some artificially introduced music, or in cases where there
is some background commentator explaining the video in a
foreign (unknown) language. This analysis can also be done
similarly the other way around, this is, with the audio colorization approach, providing images for a given audio.
6. CONCLUSIONS AND FUTURE WORK
We presented an effective method to retrieve audio samples
that fit correctly to a given (muted) video. The qualitative
results show that the already existing online videos, due to its
variety, represent a very good source of audio for new videos,
even in the case of only retrieving from a small subset of this
large amount of videos. Due to the existing difficulty to create
new audio from scratch, we believe that a retrieval approach
is the path to follow in order to give audio to videos.
The range of possibilities to extend the presented work is
excitingly broad. The first idea would be to make use of the
YouTube-8M dataset variety and information. The temporal
2 https://goo.gl/NAcJah
information provided by the individual image and audio features is not used in the current work. The most promising
future work implies using this temporal information to match
audio and images, making use of the implicit synchronization
the audio and the images of a video have, without needing
any supervised control. Thus, the next step in our research is
introducing a recurrent neural network, which will allow us to
create more accurate representations of the video, and also retrieve different audio samples for each image, creating a fully
synchronized system.
Also, it would be very interesting to study the behavior of
the system depending on the class of the input. Observing the
dataset, it is clear that not all the classes have the same degree
of correspondence between audio and image, as for example
some videos have artificially (posterior) added music, which
is not related at all to the images.
In short, we believe the YouTube-8M dataset allows for
promising research in the future in the field of video sonorization and audio retrieval, for it having a huge amount of samples, and for it capturing multi-modal information in a highly
compact way.
7. ACKNOWLEDGEMENTS
This work was partially supported by the Spanish Ministry
of Economy and Competitivity and the European Regional
Development Fund (ERDF) under contract TEC2016-75976R. Amanda Duarte was funded by the mobility grant of the
Severo Ochoa Program at Barcelona Supercomputing Center
(BSC-CNS).
8. REFERENCES
deep learning,” in Proceedings of the 28th international
conference on machine learning, 2011, pp. 689–696. 2
[1] Eric Brochu, Nando De Freitas, and Kejie Bao, “The
sound of an album cover: Probabilistic multimedia and
information retrieval,” in Artificial Intelligence and
Statistics (AISTATS), 2003. 2
[12] Liwei Wang, Yin Li, and Svetlana Lazebnik, “Learning deep structure-preserving image-text embeddings,”
CoRR, vol. abs/1511.06078, 2015. 2
[2] Rudolf Mayer, “Analysing the similarity of album art
with self-organising maps,” in International Workshop
on Self-Organizing Maps. Springer, 2011, pp. 357–366.
2
[13] Ryan Kiros, Ruslan Salakhutdinov, and Richard S.
Zemel, “Unifying visual-semantic embeddings with
multimodal neural language models,”
CoRR, vol.
abs/1411.2539, 2014. 2
[3] Janis Libeks and Douglas Turnbull, “You can judge an
artist by an album cover: Using images for music annotation,” IEEE MultiMedia, vol. 18, no. 4, pp. 30–37,
2011. 2
[14] Amaia Salvador, Nicholas Hynes, Yusuf Aytar, Javier
Marin, Ferda Ofli, Ingmar Weber, and Antonio Torralba,
“Learning cross-modal embeddings for cooking recipes
and food images,” in CVPR, 2017. 2, 3
[4] Jiansong Chao, Haofen Wang, Wenlei Zhou, Weinan
Zhang, and Yong Yu, “Tunesensor: A semantic-driven
music recommendation service for digital photo albums,” in 10th International Semantic Web Conference,
2011. 2
[15] Andrea Frome, Greg Corrado, Jonathon Shlens, Samy
Bengio, Jeffrey Dean, Marc’Aurelio Ranzato, and
Tomas Mikolov, “Devise: A deep visual-semantic embedding model,” in Neural Information Processing Systems, 2013. 2
[5] Alexander Schindler and Andreas Rauber, “An audiovisual approach to music genre classification through
affective color features,” in European Conference on
Information Retrieval. Springer, 2015, pp. 61–67. 2
[16] Yusuf Aytar, Carl Vondrick, and Antonio Torralba, “See,
hear, and read: Deep aligned representations,” arXiv
preprint arXiv:1706.00932, 2017. 2
[6] Xixuan Wu, Yu Qiao, Xiaogang Wang, and Xiaoou
Tang, “Bridging music and image via cross-modal ranking analysis,” IEEE Transactions on Multimedia, vol.
18, no. 7, pp. 1305–1318, 2016. 2
[7] Esra Acar, Frank Hopfgartner, and Sahin Albayrak,
“Understanding affective content of music videos
through learned representations,” in International Conference on Multimedia Modeling. Springer, 2014, pp.
303–314. 2
[8] Olivier Gillet, Slim Essid, and Gal Richard, “On the
correlation of automatic audio and visual segmentations
of music videos,” IEEE Transactions on Circuits and
Systems for Video Technology, vol. 17, no. 3, pp. 347–
355, 2007. 2
[9] Dongge Li, Nevenka Dimitrova, Mingkun Li, and Ishwar K Sethi, “Multimedia content processing through
cross-modal association,” in Proceedings of the eleventh
ACM international conference on Multimedia. ACM,
2003, pp. 604–611. 2
[10] Hong Zhang, Yueting Zhuang, and Fei Wu, “Crossmodal correlation learning for clustering on imageaudio dataset,” in 15th ACM international conference
on Multimedia. ACM, 2007, pp. 273–276. 2
[11] Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan
Nam, Honglak Lee, and Andrew Y Ng, “Multimodal
[17] Sungeun Hong, Woobin Im, and Hyun S. Yang, “Deep
learning for content-based, cross-modal retrieval of
videos and music,” CoRR, vol. abs/1704.06761, 2017. 2
[18] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul
Natsev, George Toderici, Balakrishnan Varadarajan, and
Sudheendra Vijayanarasimhan, “Youtube-8m: A largescale video classification benchmark,” CoRR, vol.
abs/1609.08675, 2016. 2, 4
[19] Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S
Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin,
et al., “Tensorflow: Large-scale machine learning
on heterogeneous distributed systems,” arXiv preprint
arXiv:1603.04467, 2016. 4
[20] S. Hershey, S. Chaudhuri, D. P. W. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A.
Saurous, B. Seybold, M. Slaney, R. J. Weiss, and K. Wilson, “Cnn architectures for large-scale audio classification,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), March
2017, pp. 131–135. 4
| 1cs.CV
|
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC
GROUPS
arXiv:1801.04750v1 [math.GT] 15 Jan 2018
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
Abstract. Let φ be an atoroidal outer automorphism of the free group Fn . We study
the Gromov boundary of the hyperbolic group Gφ = Fn oφ Z. We explicitly describe a
family of embeddings of the complete bipartite graph K3,3 into ∂Gφ . To do so, we define
the directional Whitehead graph and prove that an indecomposable Fn -tree is Levitt type
if and only if one of its directional Whitehead graphs contains more than one edge. As an
application, we obtain a direct proof of Kapovich-Kleiner’s theorem [KK00] that ∂Gφ is
homeomorphic to the Menger curve if the automorphism is atoroidal and fully irreducible.
1. Introduction
The family of free-by-cyclic groups is an intriguing example of a collection of groups that
is simply defined, yet has a rich geometric structure theory. Denote by Fn = hx1 , . . . , xn i the
free group on n ≥ 2 generators. Let Aut(Fn ) be the group of automorphisms of Fn , and let
Out(Fn ) be the quotient of Aut(Fn ) by the normal subgroup of inner automorphisms. Given
an outer automorphism φ ∈ Out(Fn ), the group Gφ is defined by the HNN presentation
(1.1)
Gφ := Fn oφ Z = h x1 , . . . , xn , t | t−1 xi t = Φ(xi ), 1 ≤ i ≤ n i
where Φ ∈ Aut(Fn ) is any automorphism of Fn in the outer class φ. (Different choices of Φ
yield Tietze-equivalent presentations of Gφ .) There is a satisfying correspondence between
geometric properties of Gφ and algebraic properties of φ. Brinkmann [Bri00] proved that
Gφ is hyperbolic precisely when φ is atoroidal, which means that no conjugacy class of Fn
is invariant under a power of φ. In the case that Gφ is Gromov hyperbolic, a theorem of
Bowditch [Bow98, Theorem 6.2] implies that the visual boundary of Gφ contains a local cut
point if and only if Gφ splits over a virtually infinite cyclic subgroup. Kapovich–Kleiner
[KK00, proof of Corollary 15] proved that in this case some power of φ preserves a free
splitting of Fn . Thus, there is a trichotomy: ∂Gφ contains a local cut point if and only if
Gφ splits over a virtually infinite cyclic subgroup if and only if a power of φ preserves a free
splitting of Fn .
Date: January 16, 2018.
The first and third authors were supported by ISF grant 1941/14. The second author was supported by
the grant ANR-16-CE40-0006 (ANR Dagger). The third author was supported by the Azrieli Foundation
and Zuckerman Foundation.
1
2
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
If Gφ is a hyperbolic free-by-cyclic group, then the cohomological dimension of Gφ is two,
hence the boundary of Gφ is 1-dimensional by work of Bestvina–Mess [BM91]. Kapovich–
Kleiner [KK00, Theorem 4] proved that if G is a hyperbolic group with one-dimensional
boundary, and G does not split over a finite or virtually cyclic subgroup, then the boundary
of G is homeomorphic to either the unit circle S 1 , the Sierpinski carpet, or the Menger
curve. A deep theorem of Tukia–Gabai–Casson–Jungreis [Tuk88, Gab92, CJ94] classifies
the hyperbolic groups with boundary homeomorphic to S 1 as precisely the groups that
act discretely and cocompactly by isometries on the hyperbolic plane. Kapovich–Kleiner
[KK00, Theorem 5] characterize the structure of hyperbolic groups that have boundary
homeomorphic to the Sierpinski carpet, and it follows from their characterization that
any such group has negative Euler characteristic (see Section 7 of the present article).
Consequently, a hyperbolic free-by-cyclic group cannot have boundary homeomorphic to
S 1 or the Sierpinski carpet, since a free-by-cyclic group has Euler characteristic equal to
zero. Thus, the boundary of a hyperbolic free-by-cyclic group which does not split over a
virtually cyclic group must be homeomorphic to the Menger curve.
The starting point of this project was to explicitly find a non-planar set inside ∂Gφ when
Gφ is hyperbolic. Kuratowski [Kur30] proved that a graph is non-planar if and only if it
admits an embedding of the complete bipartite graph K3,3 or the complete graph K5 .
Main Theorem. If φ is an atoroidal automorphism in Out(Fn ), then ∂Gφ contains a copy
of the complete bipartite graph K3,3 .
In order to find this non-planar set, we use the following facts and constructions. If
φ is fully irreducible, then φ acts on the set of projective classes of very small minimal
Fn -trees by north-south dynamics [LL03]. Denote the attractor by T+ and the repeller
by T− . Levitt and Lustig [LL03] defined for any free Fn tree T with dense orbits a map
Q : ∂Fn → Tb where Tb is the compactification of the metric completion of T . Of particular
interest to us are the maps Q− and Q+ associated to the trees T− and T+ . Mitra (Mj)
[Mit98] proved there exists a continuous surjection ι̂ : ∂Fn → ∂Gφ , that extends the injection
ι : Fn → Gφ in the presentation (1.1). The map ι̂ is called the Cannon-Thurston map for
the subgroup Fn ≤ Gφ . In [KL15], Kapovich and Lustig proved that preimages of ι̂ are finite
and uniformly bounded. Moreover, they showed that ι̂ factors through both Q+ and Q− .
That is, there exist continuous surjections R± : Tb± → ∂Gφ so that ι̂ = R+ ◦ Q+ = R− ◦ Q− .
The maps R± are both onto, and restricted to T± they are injective and their images are
disjoint.
To prove the Main Theorem, we realize the embedded K3,3 in ∂Gφ as a union of a part
that lies in the tree T+ and a part that lies in the tree T− . If φ admits the existence of these
subsets in a particular way, then we say that φ satisfies the T± -pattern; see Definition 6.2.
Given a tree T with dense orbits, we introduce the directional Whitehead graph (in Definition 5.1) to capture certain asymptotic relations between leaves of the lamination dual
to T . We prove in Lemma 6.6 that for any indecomposable tree T , there is the following
correspondence.
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
A directional Whitehead graph of T
contains more than one edge.
⇐⇒
3
T has Levitt type; in particular, there exists a system of partial isometries associated to T for which the Rips Machine
never halts. See Definition 3.5.
In Lemma 6.5 we prove
A directional Whitehead graph of
T− contains more than one edge
⇐⇒
φ satisfies the T± -pattern.
The T± -pattern maps into ∂Gφ to form an embedded K3,3 , as shown in Proposition 6.4. In
Section 7, we explain how the Main Theorem gives a direct proof of a result of Kapovich–
Kleiner [KK00, Section 6] which states that, if φ is fully irreducible, then ∂Gφ is homeomorphic to the Menger curve.
There is a striking and well-developed analogy between fully irreducible automorphisms
of free groups and pseudo-Anosov homeomorphisms of surfaces. Let Σ be a surface with
boundary. A pseudo-Anosov homeomorphism of Σ that fixes the boundary components of
Σ gives rise to a fully irreducible automorphism ψ of π1 (Σ), a non-abelian free group. In
this case, the automorphism ψ is not atoroidal. While the associated free-by-cyclic group
Gψ is not a hyperbolic group, the group is CAT(0) with isolated flats. Hence, the group
has a unique CAT(0) boundary by work of Hruska–Kleiner [HK09]. Ruane [Rua05] proved
that the CAT(0) boundary of Gψ in this case is homeomorphic to the Sierpinski carpet;
in particular, the CAT(0) boundary of Gψ is planar. A key distinction between these
settings is the difference in the systems of partial isometries representing the attracting and
repelling trees associated to the automorphisms. The process of iterating the Rips Machine
on the system associated to the pseudo-Anosov will halt, producing an interval exchange
transformation; while on a system associated to a fully irreducible atoroidal automorphism,
the Rips Machine will not halt, and will produce a band system with three overlapping
bands (Proposition 3.15). Our methods illustrate how this difference in dynamics leads to
the difference in the topology of the boundary: one is planar, while the other is not.
The quasi-conformal homeomorphism type of the visual boundary of a hyperbolic group is
a complete quasi-isometry invariant [Pau96]. The quasi-isometry classification of hyperbolic
free-by-cyclic groups is open. Our hope is that the new-found understanding of these
structures in the boundary will allow us to understand more about the quasi-conformal
structure on the boundary of these groups as well, but, this is beyond the scope of the
current paper.
In the process of proving our main theorem, we establish new results which may be of
independent interest to specialists in the field. For instance, we further investigate systems
of partial isometries on compact trees in continuation of the work of Coulbois, Hilion and
Lustig [CHL09, CH14, CH12]. Precise definitions of the objects considered here can be a
bit technical and will be given in Sections 2 and 3. In particular, combining our results with
previous results of [CH12], we obtain the following characterization of the type (Levitt or
surface) of a free mixing Fn -tree.
4
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
Theorem 1.1 (Corollary 3.16). Let T be a free mixing Fn -tree, and let A be a basis of
Fn . Let S0 = (K0 , A0 ) be the associated system of partial isometries, and let Si = (Ki , Ai )
denote the output after the ith iteration of the Rips Machine. Then
• T is pseudo-surface if and only if vol Ki≥3 = 0 for some i ∈ N;
• T is Levitt type if and only if vol Ki≥3 > 0 for all i ∈ N,
where Ki≥3 is the subset of Ki where at least 3 distinct partial isometries in Ai are defined,
and vol Ki≥3 denotes the volume of Ki≥3 .
Another noteworthy result relates two possible definitions of the ideal Whitehead graph
of a fully irreducible outer automorphism φ ∈ Out(Fn ). Handel–Mosher [HM11, Chapter 3]
introduced the ideal Whitehead graph to study the asymptotic relations between singular
leaves of the attracting lamination of φ; see Section 2.4. They define the graph Whφ as the
Fn -quotient of the graph whose vertex set is the union of nonrepelling fixed points in ∂Fn of
principal automorphisms representing φ; see Definition 2.33. An edge of Whφ corresponds
to a leaf of the attracting lamination Λφ+ of φ. For the second definition, let the graph WhΛφ
+
be the Fn -quotient of the graph whose vertex set is the union of endpoints of singular leaves
of Λφ+ , where a singular leaf has an asymptotic class containing more than one element. An
edge of WhΛφ corresponds to a singular leaf of Λφ+ . We prove the following in Section 2.7.
+
Theorem 1.2. The ideal Whitehead graphs Whφ and WhΛφ of a fully irreducible outer
+
automorphism φ ∈ Out(Fn ) are isomorphic.
Acknowledgements. We wish to thank the Mathematical Science and Research Institute
for their hospitality in the fall of 2016.
2. Preliminaries
2.1. Trees, directions, and the observers’ topology. A metric space (T, d) is an R-tree
if for any two points x, y ∈ T , there is a unique topological arc px,y : [0, 1] → T connecting
x to y and so that the image of px,y is isometric to the segment [0, d(x, y)]. We denote the
image of px,y by [x, y], and we refer to this arc as the segment in T from x to y. An arc is
non trivial if it contains at least 2 distinct points.
If x ∈ T is a point, a direction at x in T is a component of T r {x}. A point x ∈ T is
a branch point if there are at least 3 directions at x; it is an extremal point if there is only
one direction at x. Let Int(T ) denote the interior of a tree T , i.e. the set of points of T
which are not extremal. Two arcs [x, y], [x, y 0 ] are germ-equivalent if [x, y] ∩ [x, y 0 ] 6= {x};
an equivalence class is a germ at x. The map that associates to the arc [x, y] the direction
containing y induces a bijection from the set of germs at x to the set of directions at x.
A ray in T is the image of an immersion R+ → T . Two rays ρ and ρ0 are asymptotic if
ρ ∩ ρ0 has infinite diameter. Let ∂T denote the Gromov boundary of T ; as a set, ∂T consists
of asymptotic classes of rays in T . Let T denote the metric completion of T . The points in
T r T are extremal in T . Let Tb = T ∪ ∂T .
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
5
As explained by Coulbois–Hilion–Lustig [CHL07], the set Tb may be equipped with the
observers’ topology, which makes the space homeomorphic to a dendrite. The observers’
topology on Tb is the topology generated (in the sense of a subbasis) by the set of directions
in Tb, where the notion of direction in T extends to Tb, and points of ∂T are extremal in
Tb. This topology is weaker than the topology induced by the metric on an R-tree. On any
interval of T , the metric topology and the observers’ topology agree.
2.2. Systems of partial isometries and the band complex.
Definition 2.1. A compact forest is a finite union of compact R-trees. A partial isometry
of a compact forest K is an isometry a : J → J 0 , where J and J 0 are compact subtrees
of K. The domain of a, denoted dom(a), is J, and the range of a is J 0 . A partial isometry is
non-empty if its domain is non-empty. A system of partial isometries is a pair S = (K, A),
where K is a compact forest and A is a finite collection of non-empty partial isometries
of K.
Definition 2.2. Let S = (K, A) be a system of partial isometries, and let I = [0, 1] denote
the unit interval. For each ai ∈ A, let bi ⊂ K be the domain of ai . Let Bi := bi × I be
called a band. The band complex B is the quotient of K ti Bi , where bi × {0} is identified
to the domain of ai , and bi × {1} is identified to the range of ai by isometries.
Definition 2.3. Each band B = b × I ⊂ B is foliated by leaves of the form x × I for x ∈ b.
The foliation of the bands yields a foliation of B. A finite, infinite, or bi-infinite path γ in
B is an admissible leaf path if γ is a locally isometric, immersed path contained in a leaf.
A half-leaf based at x ∈ K is an admissible leaf path ρ : [0, ∞) → B with ρ(0) = x. A
finite admissible path γ travels through a finite sequence of bands B1 , . . . , Bk and, in turn,
corresponds to the sequence of partial isometries a1 , . . . , ak that give rise to the bands. The
domain of γ is then dom(ak ◦ · · · ◦ a1 ). Note that the domain contains γ(0). This definition
extends to admissible rays and bi-infinite lines. The limit set Ω of S = (K, A) is the set
of elements of K which are in the domain of a bi-infinite admissible reduced path. The
lamination L(B) is the set of bi-infinite admissible leaf paths in B. For an R-tree J, let
µJ denote the Lebesgue measure on J, which consists of the Lebesgue measures on the
segments of J. The foliated space B has a transverse measure associated to the Lebesgue
measure on the components of K.
Notation 2.4. Let x ∈ K, and let ` = . . . z−2 z−1 z0 z1 z2 . . . ⊂ L(B) be a bi-infinite admissible leaf path through x; so, zi ∈ A ∪ A−1 for i ∈ Z. Suppose `1 = z0 z1 z2 . . . and
−1 −1
`2 = z−1
z−2 . . . denote the half-leaves based at x. Then, ` = `1 ∪ `2 , and to record the
point x in the leaf `, we use the notation ` = . . . z−2 z−1 .z0 z1 z2 . . ..
Definition 2.5. Let S = (K, A) be a system of partial isometries, and let B be the associated band complex. Let Be denote the universal cover of B. The foliation and transverse
e Collapsing each leaf in Be to
measure on B lift to a foliation and transverse measure on B.
a point yields an R-tree TB , called the dual tree to B.
6
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
2.3. The Q-map and dual lamination.
Definition 2.6. An Fn -tree is an R-tree T together with a homomorphism ρ : Fn →
Isom(T ); the homomorphism is often repressed. The action of Fn on T is minimal if Fn
does not leave invariant any non-trivial subtree of T . The action is very small if (1) all
edge stabilizers are cyclic (i.e. {1} or Z); (2) for every non-trivial g ∈ Fn , the fixed subtree
F ix(g) is isometric to a subset of R; and (3) F ix(g) is equal to F ix(g p ) for all p ≥ 2.
Theorem 2.7. [LL03] [CHL07, Proposition 2.3] Let T be a minimal, very small, Fn -tree
with dense orbits. There exists an Fn -equivariant surjective map
Q : ∂Fn → Tb
which is continuous with respect to the observers’ topology. In addition, points in ∂T have
exactly one pre-image by Q.
The Q-map given in the previous theorem may be used to define a lamination of Fn as
follows.
Definition 2.8. The double boundary of Fn is ∂ 2 Fn := (∂Fn × ∂Fn ) r ∆, where ∆ is the
diagonal. Let i : ∂ 2 Fn → ∂ 2 Fn denote the involution that exchanges the factors. The double
boundary ∂ 2 Fn is endowed with the topology induced by the product topology, and Fn acts
diagonally on ∂ 2 Fn . A lamination of Fn is a non-empty, closed, Fn -invariant, i-invariant
subset of ∂ 2 Fn .
Definition 2.9. Let Fn act by isometries on an R-tree T so that the action is minimal,
very small, and has dense orbits. Let Q : ∂Fn → Tb be the Q-map given in Theorem 2.7.
The dual lamination of T is
L(T ) = {(X, Y ) ∈ ∂ 2 Fn | Q(X) = Q(Y )}.
A leaf of the dual lamination L(T ) is a pair (X, Y ) ∈ L(T ).
2.4. Attracting and repelling trees and laminations of an outer automorphism.
Definition 2.10. (Outer space.) For n ≥ 2, Culler–Vogtmann outer space, denoted CVn ,
is the projectivized space of minimal, free, discrete actions by isometries of the free group
Fn on R-trees. Its topology is induced by embedding CVn into the space of length functions
[CM87]. Let CV n denote the compactification of CVn , which is the set of projective classes
of minimal, very small, Fn -trees [CM87, CL95, BF94]. Let ∂CVn = CVn r CVn . These
spaces admit an action of Out(Fn ); an element φ ∈ Out(Fn ) sends an Fn -tree (T, ρ) to
(T, ρ ◦ Φ), where Φ ∈ Aut(Fn ) is in the class φ. (For background, see [CV86], [Vog02].)
Definition 2.11. An outer automorphism φ ∈ Out(Fn ) is fully irreducible (or, iwip) if no
conjugacy class of a proper free factor of Fn is fixed by a positive power of φ.
Theorem 2.12. [LL03] If φ ∈ Out(Fn ) is fully irreducible, then φ acts on CV n with NorthSouth dynamics and projectively fixes two trees T+ , T− ∈ ∂CVn .
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
7
Definition 2.13. We refer to the tree T+ = T+φ as the attracting tree of φ and to the tree
T− = T−φ as the repelling tree of φ. We omit the φ-notation when the outer automorphism
−1
is clear from context. Notice that T+φ = T−φ
−1
and T−φ = T+φ .
If φ is a fully irreducible outer automorphism, then the group Fn acts on the trees T+ and
T− with dense orbits. In fact, these actions exhibit a much stronger dynamical property.
Definition 2.14. The action of Fn on an R-tree T is mixing if for any non-degenerate
segments I and J in T the segment I is covered by finitely many translates of J: there
exist finitely many elements u1 , . . . , uk ∈ Fn so that I ⊂ u1 J ∪ u2 J ∪ . . . ∪ uk J. The
action is indecomposable if, in addition, the elements u1 , . . . , uk ∈ Fn may be chosen so that
ui J ∩ ui+1 J is a non-degenerate segment for any i = 1, . . . , k − 1.
Lemma 2.15. [CH12, Theorem 2.1] The actions of Fn on the attracting and repelling trees
of a fully irreducible outer automorphism of Fn are indecomposable.
Definition 2.16. Let T+ and T− be the attracting and repelling trees of a fully irreducible
free group automorphism. By Theorem 2.7, there are Fn -equivariant, surjective maps
Q+ : ∂Fn → Tb+ and Q− : ∂Fn → Tb−
which are continuous with respect to the observers’ topology. In addition, points in ∂T+ and
∂T− have exactly one pre-image by Q+ and Q− , respectively. Define the dual laminations
L(T+ ) and L(T− ) as in Definition 2.9.
Proposition 2.17. [KL15, Proposition 3.22] L(T+ ) ∩ L(T− ) = ∅.
Remark 2.18. Let T be an R-tree with a very small, minimal action of Fn by isometries
and dense orbits, let Q : ∂Fn → Tb be the Q-map, and let L(T ) be the dual lamination.
The map Q induces a map
Q : L(T ) → Tb
by
Q(X, Y ) = Q(X) = Q(Y ).
Thus, there are maps Q+ : L(T+ ) → Tb+ and Q− : L(T− ) → Tb− .
Bestvina–Feighn–Handel [BFH97] define attracting and repelling laminations Λ+ = Λφ+
and Λ− = Λφ− , respectively, associated to a fully irreducible φ ∈ Out(Fn ). As we do not use
the definition of these laminations explicitly, we refer the reader to [BFH97] for details and
briefly recall the facts relevant to this paper. Similar to the structure of the attracting and
φ−1
repelling trees T+ and T− , the attracting and repelling laminations satisfy Λφ+ = Λ−
and
−1
Λφ− = Λφ+ . Moreover, a strong relationship between Λ+ and L(T− ) can be seen using the
following construction:
diag(Λ± ) = {(ξ1 , ξm ) ∈ ∂ 2 Fn | ∃ξ2 , . . . , ξm−1 ∈ ∂Fn with (ξi , ξi+1 ) ∈ Λ± for 1 ≤ i ≤ m − 1}.
Lemma 2.19. [CHR15, KL14] L(T− ) = diag(Λ+ ) and L(T+ ) = diag(Λ− ).
We shall use another relationship between Λ+ and L(T− ), given as follows.
8
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
Definition 2.20. [CHR15] Let L be a lamination in ∂ 2 Fn . The derived lamination of L,
denoted L0 , is the set of limit points in L. That is,
` ∈ L0 ⇐⇒ ∃{`i } ⊂ L with lim `i = `.
i→∞
Lemma 2.21. [CHR15] Let φ be a fully irreducible outer automorphism, let Λ+ denote its
attracting lamination, and let L(T− ) denote the dual lamination of its repelling tree. Then
Λ+ = L(T− )0 .
Proof. Let φ be a fully irreducible outer automorphism. By Lemma 2.15, the action of Fn
on T− is indecomposable. By [CHR15, Theorem A], the lamination L(T− ) contains a unique
minimal sublamination which equals its derived lamination L(T− )0 . On the other hand, the
attracting lamination Λ+ is minimal and contained in L(T− ). Therefore, L(T− )0 = Λ+ .
2.5. Realization of leaves.
Definition 2.22. Let L be a lamination of Fn . The ends of the lamination L is the set
EL = {X ∈ ∂Fn | ∃Y ∈ ∂Fn such that (X, Y ) ∈ L}.
Lemma 2.23. Let φ ∈ Out(Fn ) be a fully irreducible outer automorphism, let T± be the
attracting and repelling trees of φ, and let Q+ : ∂Fn → Tb+ be the Q-map. For every
X ∈ EL(T− ), Q+ (X) ∈ ∂T+ .
Proof. This argument is given in [HM11]. We include an outline for clarity (keeping the
notation used by the reference), and omitting the definitions; consult [HM11] for more
e denote the
details. Let g : Γ → Γ be an affine train track representative of φ, and let Γ
e
universal cover of Γ. Via the marking on Γ, the boundary ∂ Γ may be identified with ∂Fn ,
e There
and the attracting lamination Λ+ may be identified with a set of geodesic lines in Γ.
e
is an Fn -equivariant edge-isometry fg : Γ → T+ which is an Λ+ -isometry; that is, for each
e the map fg ◦ ` : R → T+ is an
leaf ` ∈ Λ+ viewed as an isometric embedding ` : R → Γ,
isometric embedding. Therefore, fg (`(±∞)) := limt→±∞ fg ◦ `(t) exists and lies in ∂T+ .
Since the limit exists, Q+ (`(±∞)) = fg (`(±∞)) [LL03]. Hence, Q+ maps the ends of Λ+
to ∂T+ . By Lemma 2.19, EL(T− ) = EΛ+ , so Q+ (X) ∈ ∂T+ .
Definition 2.24. Let ` = (X, Y ) ∈ L(T− ). By Lemma 2.23, Q+ (X), Q+ (Y ) ∈ ∂T+ , and
by Proposition 2.17, Q+ (X) 6= Q+ (Y ). The realization of ` in T+ is the bi-infinite geodesic
`+ in T+ connecting Q+ (X) and Q+ (Y ).
Proposition 2.28 below compares the convergence of leaves of the lamination L(T− ) in
the topology on ∂ 2 Fn to the convergence of the realization of leaves in the attracting tree
T+ in the Hausdorff topology.
+
Definition 2.25. If {`+
i | i ∈ N} is a sequence of bi-infinite geodesics in T+ and ` is a
+
+
bi-infinite geodesic in T+ , then limi→∞ `i = ` in the Hausdorff topology on T+ if for any
subarc I ⊂ `+ , there exists N ∈ N so that I ⊂ `+
i for all i > N .
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
9
Definition 2.26. Let A be a basis for Fn . The set ∂ 2 Fn may be identified with the space
of pairs (X, Y ) of infinite reduced words in A ∪ A−1 with X 6= Y . For an infinite reduced
word X, let X1 denote the first letter of X. Let
CA := {(X, Y ) ∈ ∂ 2 Fn | X1 6= Y1 }
be the unit cylinder associated to A.
Lemma 2.27. Let A be a basis of Fn . The unit cylinder associated to A is open and
compact in ∂ 2 Fn .
Proposition 2.28. Let A a basis of Fn . Let CA ⊂ ∂ 2 Fn be the unit cylinder associated
to A and let {`i | i ∈ N} ∪ {`} ⊂ L(T− ) ∩ CA . Then, limi→∞ `i = ` ∈ ∂ 2 Fn if and only if
+
limi→∞ `+
i = ` ⊂ T+ in the Hausdorff topology.
Proof. Let ` = (ξ, η) ∈ L(T− ) ∩ CA and `i = (ξi , ηi ) ∈ L(T− ) ∩ CA so that limi→∞ `i = `.
Let I = [a, b] ⊂ `+ , and assume that a is between Q+ (ξ) and b. Let d1 be the direction
at a containing Q+ (ξ). Since Q+ is continuous with respect to the observers’ topology,
there exists N1 ∈ N so that for all i > N1 , Q+ (ξi ) ∈ d1 . Similarly, if d2 is the direction
at b containing Q+ (η), then there exists N2 ∈ N so that for all i > N2 , Q+ (ηi ) ∈ d2 . Let
i > max{N1 , N2 }. Then `+
i contains [a, b] as desired.
+ in the Hausdorff topology. Since L(T ) ∩ C
Suppose now that limi→∞ `+
−
A is
i = `
compact by Lemma 2.27, the sequence {`i } has a convergent subsequence. Let τ be a
partial limit; that is, limj→∞ `ij = τ ∈ ∂ 2 Fn . By the arguments in the previous paragraph,
+ in the Hausdorff topology on T . Thus, `+ = τ + . By Theorem 2.7, the
limj→∞ `+
+
ij = τ
map Q+ is injective on EL(T− ), so τ = `. Therefore, the sequence {`i } has a unique partial
limit; thus, {`i } converges to `.
We conclude this section by describing a property of the realization of leaves of L(T− ) in
T+ that we will use in Section 6.
Definition 2.29. A star is a wedge of intervals or a wedge of rays. The wedge point is
called the middle of the star.
Proposition 2.30. [HM11] Let `1 , `2 , . . . , `k be leaves of the lamination L(T− ) such that `i
S
is asymptotic to `i+1 for each i = 1, . . . , k − 1. Then ki=1 `+
i is a star in T+ .
To prove this property we will need some facts from the next subsection.
2.6. Automorphisms, branch points, and homotheties. We will use the following
facts about the correspondence between a fully irreducible outer automorphism φ and an
automorphism in Aut(Fn ) representing φ.
Definition 2.31. Let Rn denote the graph with one vertex v and n edges. Choosing a
basis A of Fn and identifying each oriented edge with a distinct element of A identifies
π1 (Rn , v) with Fn . Moreover, each automorphism Φ ∈ Aut(Fn ) is represented by a map
f : Rn → Rn sending the vertex v to itself and an edge of Rn to an immersed edge path
so that f is a homotopy equivalence. The correspondence between such self-maps of Rn
10
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
and elements of Aut(Fn ) is a bijection. Let G be a graph and µ : Rn → G a homotopy
equivalence. A homotopy equivalence f : G → G gives rise to an outer class [f∗ ] ∈ Out(Fn ).
If f (e) is an immersed edge path for every edge e ∈ E(G), then we say that f is a topological
representative of [f∗ ]. Let φ ∈ Out(Fn ), let f : G → G be a topological representative of
e be the universal cover of G. The identification, via the homotopy equivalence
φ, and let G
e by deck
µ, of π1 (G, ∗) with Fn (up to conjugation) gives rise to an action of Fn on G
e
transformations ρ : Fn → Aut(G). The map ρ is well defined up to precomposing it with
ig , where ig ∈ Aut(Fn ) denotes conjugation by g.
e
Lemma 2.32. Let φ ∈ Out(Fn ) and f : G → G be a topological representative of φ. Let G
e
denote the universal cover of G and ρ : Fn → Aut(G) the action of Fn by deck transformae→G
e of g given by
tions. There is a bijection between automorphisms Φ ∈ φ and lifts ge : G
the equation:
(2.1)
ge ◦ ρ(γ) = ρ(Φ(γ)) ◦ ge.
Definition 2.33. Let φ ∈ Out(Fn ) be fully irreducible, and let Φ ∈ Aut(Fn ) so that Φ ∈ φ.
We say that Φ is a principal automorphism if the extension of Φ to ∂Fn , which we denote
by ∂Φ, fixes at least three nonrepelling points.
Lemma 2.34. [GJLL98, HM11] Let φ ∈ Out(Fn ) be fully irreducible, and let T+ be its
attracting tree.
(1) If Φ is a principal automorphism representing φ, then there exists a unique homothety h+ : T+ → T+ so that h+ (γx) = Φ(γ)h+ (x) for all γ ∈ Fn and x ∈ T+ . This
homothety stretches the distances in T+ by the dilatation of φ.
(2) The homothety h+ from the previous item fixes a branch point b ∈ T+ .
(3) The correspondence Φ 7→ b defines a bijection between the set of principal automorphisms representing φ and the set of branch points of T+ .
(4) For each automorphism Φ representing φ there exists a homothety h− : T− → T− so
that h− (γx) = Φ(γ)h− (x) for all γ ∈ Fn and for all x ∈ T− .
Proof. Item (4) is the only one that does not appear in [HM11]. Since φ acts on CV n by
North-South dynamics, T− is equivalent to T− ·φ. This means that there exists a homothety
h− : T− → T− which is Fn -equivariant. That is, if ρ : Fn → Aut(T− ) is the Fn -action on T−
then,
h− (ρ(γ)x) = ρ(Φ(γ))h− (x).
Suppressing ρ yields the statement in the lemma.
It will be easier to work with rotationless automorphisms and topological representatives.
One can find the definitions of rotationless automorphisms and rotationless train track maps
in [FH09] or [HM11, pages 24,27]. Here we will only need the following facts due to FeighnHandel [FH09].
(1) Each fully irreducible φ ∈ Out(Fn ) has a rotationless power φp , and p is bounded
p
p
by a constant depending only on n. Moreover, T±φ = T±φ and Λφ± = Λφ± .
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
11
(2) Let φ be a fully irreducible automorphism and f : G → G a train track representative
of φ. Then, φ is rotationless if and only if f is rotationless.
(3) If f : G → G is rotationless, then for each vertex v, f 2 (v) = f (v). For each directed
edge e, the germ of f 2 (e) equals that of f (e).
Lemma 2.35. Let φ ∈ Out(Fn ) be fully irreducible. There exists φp , a rotationless power of
φ, so that for any asymptotic leaves `1 , `2 ∈ Λφ+ , there exists an automorphism Φ ∈ Aut(Fn )
representing φp so that the endpoints of `1 and `2 are non-repelling fixed points of ∂Φ.
Proof. We call a triplet (X, Y, Z) of distinct points in ∂Fn special if (X, Y ), (Y, Z) ∈ Λφ+ . The
set of special triplets is Fn -invariant. Moreover, if (X, Y, Z) is special, then by Lemma 2.19,
Q− (X) = Q− (Y ) = Q− (Z). Since the Q− -index is bounded (by 2n − 2) [CH14] (see [CH14]
for the definition), there are finitely many Fn -orbits of special triplets. Therefore, there
exist q, m ∈ N so that for every automorphism Ψ representing φ, and for each special triplet
W = (X, Y, Z), the special triplets Ψm+q (W ) and Ψq (W ) are in the same Fn -orbit.
Let φp be a rotationless power of φ. Let f : G → G be a train-track representative of
e with ∂Fn , let f0 : G
e→G
e be any lift of f , and let Φ0 be the corresponding
Identify ∂ G
automorphism given by Lemma 2.32 representing φp . By the previous paragraph, there
exists g ∈ Fn so that each special triplet W = (X, Y, Z) satisfies ∂f0m+q (W ) = g∂f0q (W ).
The map f is rotationless; hence, the image of each germ is fixed by f . So, m = 1. Moreover,
we may assume q = 0 or q = 1. If q = 0, then ∂f0 (W ) = gW . Let Φ = ig−1 ◦ Φ0 , where
ig denotes conjugation by g. Then, ∂Φ(W ) = W = (X, Y, Z). The points X, Y, Z are
attracting fixed points in this case, since they belong to EΛφ+ . If q = 1, then ∂f02 (W ) =
−1
−1 = i ◦ Φ−1 for w = Φ−2 (g). Then, W is
g∂f0 (W ). Thus, W = Φ−2
w
0
0
0 (g)∂Φ0 (W ). Let Φ
fixed by Φ−1 , which represents φ−p . Therefore, Φ also fixes W = (X, Y, Z), and the points
X, Y, Z must be attracting fixed points, since they belong to EΛφ+ .
φp .
Lemma 2.36. [CH12, Equation 4.2] Let Φ be an automorphism and h± : T± → T± the
corresponding homotheties guaranteed by Lemma 2.34. Then, for each ξ ∈ ∂Fn ,
Q± (∂Φ(ξ)) = h± (Q± (ξ)).
Proof of Proposition 2.30. By Lemma 2.35 there exists a principal automorphism Φ representing φp for some p ∈ N so that the endpoints of `1 and `2 are attracting fixed points of
∂Φ. Similarly, since `2 and `3 are asymptotic, there exists an automorphism Φ0 representing
φp so that the endpoints of `2 and `3 are attracting fixed points of ∂Φ0 . Attracting fixed
points of two automorphisms in the same outer class are disjoint [Hil07, Theorem 1.1] (see
alternatively [HM11, Corollary 2.9] for two principal automorphisms), so Φ = Φ0 . Continuing in this fashion proves that the end-points of `1 , . . . , `k are attracting fixed points of ∂Φ.
By Lemma 2.34(1), there exists a homothety h : T+ → T+ representing Φ. Let S ⊂ T+ be
the union of the realizations of {`i }ki=1 . Then, S is invariant by h by Lemma 2.36. Moreover,
S has finitely many vertices, and, h permutes the finitely many vertices of S. Since h is a
homothety, S contains only one vertex.
12
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
2.7. Proof of Theorem 1.2. Let Whφ be the Fn -quotient of the graph whose vertex set
is the union of nonrepelling fixed points in ∂Fn of principal automorphisms representing φ;
see Definition 2.33. An edge of Whφ corresponds to a leaf of the attracting lamination Λφ+
of φ. Let WhΛφ be the Fn -quotient of the graph whose vertex set is the union of endpoints
+
of singular leaves of Λφ+ , where a singular leaf has an asymptotic class containing more than
one element. An edge of WhΛφ corresponds to a singular leaf of Λφ+ . By Lemma 2.35,
+
W hφ ∼
= WhΛφ , proving Theorem 1.2.
+
2.8. System of partial isometries associated to a free group automorphism.
Coulbois–Hilion–Lustig in [CHL09] use the Q-map to construct a system of partial isometries for an R-tree T with a very small, minimal action of Fn by isometries and dense
orbits.
Construction 2.37. [CHL09]. Let T be a very small, minimal Fn -tree with dense orbits,
let A be a basis of Fn , and let CA ⊂ ∂ 2 Fn be its unit cylinder. Let
ΩA := Q(L(T ) ∩ CA ) ⊂ T .
The compact heart KA of T relative to A is the convex hull of ΩA in T . By [CHL09], KA
is indeed a compact subtree of T . Let S = (KA , A) be the system of partial isometries so
that for each a ∈ A, the partial isometry associated to a is the maximal restriction of a−1
to KA . As defined in Section 2.2, let B be the associated band complex of S, and let TB
denote the Fn -tree dual to B.
Remark 2.38. If KA is the compact heart of T relative to the basis A of Fn , then gKA
is the compact heart of T relative to the basis gAg −1 of Fn . Moreover, the unit cylinder
satisfies CgAg−1 = gCA and ΩgAg−1 = gΩA . We leave the verification of these facts to the
reader.
Theorem 2.39. [CHL09] Let T be a minimal, very small, Fn -tree with dense orbits, and
let A, CA , B, and TB be defined as in Construction 2.37.
(1) The tree T is equal to the minimal subtree of the tree TB .
(2) L(T ) ∩ CA = L(B), where L(T ) ∩ CA and L(B) are identified with bi-infinite reduced
words in A ∪ A−1 as in the above construction.
Remark 2.40. When we apply Construction 2.37 and Theorem 2.39 to the attracting or
repelling trees T+ or T− of a fully irreducible outer automorphism, we denote each object
with the appropriate subscript; we write S+ , Ω+ , . . ., and so on.
3. Rips Induction and overlapping bands
3.1. Rips Induction.
Definition 3.1. Let S = (K, A) be a system of partial isometries. The output of the Rips
Machine applied to S is a new system of partial isometries S 0 = (K 0 , A0 ) defined as follows.
K 0 := {x ∈ K | x ∈ dom(a) ∩ dom(a0 ) for some a 6= a0 ∈ A ∪ A−1 }.
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
13
Since A is finite and the intersection of two domains is a compact R-tree, F 0 is a compact forest. Let A0 be the set of all maximal restrictions of the elements of A to pairs of
components of F 0 . Then S 0 = (F 0 , A0 ) is a system of partial isometries.
Definition 3.2. [CH14, Definition 3.11] Let S0 = (K0 , A0 ) be a system of partial isometries,
and let S1 = (K1 , A1 ) denote the output of the Rips Machine. The system of partial
isometries S0 is reduced if for any partial isometry a ∈ A±
0 , the set of extremal points of the
domain of a is contained in K1 .
Proposition 3.3. [CH14, Propositions 5.6, 3.14] Let φ ∈ Out(Fn ) be a fully irreducible
+
−
outer automorphism, and let S+ = (KA
, A) and S− = (KA
, A) be the systems of partial
isometries defined in Construction 2.37. Then S+ and S− are reduced, as is any output of
S+ or S− under the Rips Machine.
Definition 3.4. Let S0 = (K0 , A0 ) be a system of partial isometries. Let Si = (Ki , Ai )
denote the output of the ith iteration of the Rips Machine. If for some i, Ki = Ki+1 , then
the Rips Machine halts on Si , and the Rips Machine eventually halts on S0 .
Definition 3.5. Let S0 be a system of partial isometries. If the Rips Machine eventually
halts on S0 , then S0 is called surface type. If the Rips Machine does not eventually halt on
S0 and
lim max diam a = 0,
i→∞ a∈Ai
then S0 is called of Levitt type.
The definition of ‘Levitt type’ given here is equivalent to the original one in [CH12,
Section 5], which states that the limit set Ω, see Definition 2.3, is totally disconnected.
Notation 3.6. Let φ ∈ Out(Fn ) be a fully irreducible outer automorphism, and let S+ and
S− be the systems of partial isometries defined in Construction 2.37. If S+ , respectively
S− , is of Levitt type, we say T+ , respectively T− , is of Levitt type.
The following proposition is a direct consequence of a result of Coulbois–Hilion [CH12].
Proposition 3.7. [CH12, Theorem 5.2] Let φ ∈ Out(Fn ) be a fully irreducible atoroidal
outer automorphism. Then either both T+ and T− are of Levitt type, or one of the trees T+
and T− is of Levitt type and the other one is of surface type.
The proof of our Main Theorem uses the fact that at least one of T+ and T− is of Levitt
type.
3.2. Volume of an R-tree and overlapping domains. The main aim of this section is
to prove Proposition 3.15, which is a converse to [CH14, Proposition 4.3], and states that
if T is an Fn -tree with dense orbits and of Levitt type, then the associated Rips machine
(beginning from the compact heart) has the property that at each step there are three
overlapping bands.
14
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
Definition 3.8. A compact R-tree K is finite if K has a finite number of extremal points.
In this case, K has also a finite number of branch points. Removing these branch points
from K yields to a finite set of arcs; the volume vol(K) of K is the sum of the lengths of
these arcs. The volume vol(K) of a compact R-tree K is the supremum of the volume of
the finite subtrees contained in K. A compact forest K is a finite disjoint union of compact
P
trees K1 , . . . , Kp ; its volume is vol(K) = 1≤i≤p vol(Ki ).
This volume of an R-tree may be finite or infinite. In either case, the following proposition
applies.
Proposition 3.9. Let T be a compact R-tree. Let > 0 and let E be a set of disjoint arcs
in T of length . Then E is finite.
For the proof of Proposition 3.9, we will use the following combinatorial lemma (that we
state in the context of R-trees even if it is valid for more general notions of trees).
Lemma 3.10. Let T be an R-tree. Let E be an infinite set of pairwise disjoint subtrees
of T . There exists a sequence (Tn )n∈N of pairwise distinct elements of E with the property
S
that, for all n, Tn+1 does not intersect the convex hull Kn of ni=1 Ti .
Proof. First, suppose there exists an arc γ in T intersecting an infinite number of elements
of E. Choose an orientation on γ, which induces a linear ordering ≺ on the set of elements
of E that intersect γ. Any countable subset of E gives rise to a sequence (Tn )n∈N of pairwise
disjoint subtrees of T such that for all n ∈ N, Tn ≺ Tn+1 . In particular Tn+1 does not
S
intersect the convex hull of ni=1 Ti .
In the remaining case, every arc of T intersects a finite number of elements of E. We build
the sequence (Tn )n∈N by induction. First, pick an element T1 of E. Suppose T1 , . . . , Tn have
S
S
been chosen so that no element of E − ni=1 Ti intersects the convex hull Kn of ni=1 Ti . Let
S
S be an element of E − ni=1 Ti . By assumption, S does not intersect Kn . Let γ be the arc
in T joining S and Kn . Set Tn+1 to be the element of E closest to Kn that intersects γ.
Proof of Proposition 3.9. Assume towards a contradiction there is a sequence (γn )n∈N of
disjoint arcs in T of length > 0. By Lemma 3.10, we can assume that up to reordering
S
indices, γn+1 does not lie in the convex hull Kn of ni=1 γi . Since γn+1 has length , at least
one of its extremal points, denoted by xn+1 , is at distance more than /2 from Kn . Thus,
for all i 6= j ∈ N, d(xi , xj ) > /2. Hence, no subsequence of (xk )k∈N is convergent (since
such a subsequence is not a Cauchy sequence), which contradicts the compactness of T .
Remark 3.11. Let K be a (non-empty) compact tree. Then vol(K) = 0 if and only if K
is point. Indeed, if K contains two distinct points, then K contains the arc joining these
two points, and hence the volume of K is as least as big as the length of this interval.
Notation 3.12. Let S = (K, A) be a system of partial isometries. Let A−1 = {a−1 | a ∈ A},
and let A± = A ∪ A−1 . We suppose from now on that A ∩ A−1 = ∅. Let v(x) be the band
valence of x; that is, v(x) = #{a ∈ A± | x ∈ dom(a)}. Let K =i = {x ∈ K | v(x) = i}.
Define K ≥i and K ≤i similarly.
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
Lemma 3.13. Let S = (K, A) be a system of partial isometries, such that vol K
Then:
15
≥3
= 0.
(i) K ≥3 is a finite set of points in K.
(ii) If, moreover, the valence of every point in K is finite, then the set {δ direction at x | x ∈
K ≥3 } of directions in K at a point in K ≥3 is finite.
Proof. Property
(ii) follows immediately from Property (i). According to Remark 3.11,
since vol K ≥3 = 0, the components of K ≥3 are points. It remains to show there are only
finitely many components. Since the domains of the partial isometries in A±1 are subtrees
of the tree K, an intersection of domains is also a tree (and possibly the empty set). Thus,
the set of components of K ≥3 injects in the set of subsets of A±1 of cardinality at least 3,
which is a finite set since A±1 is itself a finite set.
Let T be a free R-tree with dense orbits and of Levitt type, and let A be a basis of Fn .
Let K0 be the compact heart of T relative to A, and let S0 = (K0 , A0 ) be the associated
system of partial isometries as in Construction 2.37. Since T is of Levitt type, the Rips
machine does not halt. Let Si = (Ki , Ai ) denote the output after the ith iteration of the
Rips Machine. Then, vol(Ki=1 ) > 0 for all i ∈ N.
Remark 3.14. (i) All orbits of S0 = (K0 , A0 ) are infinite; that is, every leaf in the
band complex B is infinite. Indeed, [CH14, Proposition 5.6] ensures the system S0 is
reduced, which, by definition, implies every orbit is infinite.
(ii) The valence of the points in K0 is bounded. Indeed, K0 is a subtree of T , and the
valence of the points of T is bounded (by 2n [GJLL98]).
Proposition 3.15. Let T be an Fn -tree with dense orbits and of Levitt type, and let A
be a basis of Fn . Let S0 = (K0 , A0 ) be the associated system of partial isometries, and
let Si = (Ki , Ai ) denote the output after the ith iteration of the Rips Machine. Then
vol Ki≥3 > 0 for all i ∈ N.
Proof. By contradiction, suppose vol Ki≥3
= 0 for some i0 ∈ N. By definition of the Rips
0
≥3
machine, Ki+1
⊆ Ki≥3 . In particular, the sequence vol Ki≥3
is decreasing, and thus
i
vol Ki≥3 = 0 for all i ≥ i0 . We can suppose that i0 = 0.
We define by induction a sequence (γn ) of subarcs of K0 . Let γ0 be an arc contained in
K0=1 . There is only one partial isometry a defined on γ0 ; let γ1 = γ0 · a, the image of γ0
by a. Since S0 has no finite orbit, γ1 ⊆ K0≥2 . In particular, γ0 and γ1 are disjoint.
Case 1: γ1 ⊆ K0=2 . There are exactly two partial isometries defined on γ1 : one is a−1 ;
let b denote the other. Set γ2 = γ1 · b. Again, γ2 ⊆ K0≥2 since S0 has no finite orbits.
Case 2: γ1 ∩ K0≥3 6= ∅. Replace γ0 by a subarc so γ1 ⊆ K0=2 ; this procedure is possible
since K0≥3 is a finite subset of K0 by Lemma 3.13. Define γ2 as in Case 1.
16
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
Iterate this process. Since the valence of the points of K is finite (see Remark 3.14),
statement (ii) of Lemma 3.13 ensures that after a finite number of iterations, Case 2 stops
to occur.
Finally, the arcs γn are pairwise disjoint. Indeed, γ0 ⊆ K0=1 and γi ⊆ K0=2 for all i ∈ N.
(In other words, one can perform a sequence of “Rips moves” successively based on γ0 , γ1 ,
γ2 , and so on.) This collection contradicts Proposition 3.9.
By [CH14, Proposition
4.3], if S0 = (K0 , A0 ) is a pseudo-surface system of partial isome
≥3
tries, then vol K0
= 0. As shown in [CH14, Proposition 5.14], a free mixing Fn -tree
T is either pseudo-surface – in which case vol K0≥3 = 0 by [CH14, Proposition 4.3] – or
Levitt type – in which case vol Ki≥3 > 0 for all i ∈ N by Proposition 3.15. Combining
these two propositions yields the following corollary.
Corollary 3.16. Let T be a free mixing Fn -tree and let A be a basis of Fn . Let S0 = (K0 , A0 )
be the associated system of partial isometries, and let Si = (Ki , Ai ) denote the output after
the ith iteration of the Rips Machine. Then
• T is pseudo-surface if and only if vol Ki≥3 = 0 for some i ∈ N,
• T is Levitt type if and only if vol Ki≥3 > 0 for all i ∈ N.
4. Cannon–Thurston maps
The existence of a Cannon–Thurston map in the setting of hyperbolic free-by-cyclic
groups was shown by Mitra (Mj) [Mit98]. The structure of Cannon–Thurston maps for
hyperbolic free-by-cyclic groups which do not virtually split over Z was investigated by
Kapovich–Lustig [KL15]. We will use the following results.
Theorem 4.1. [Mit98] Let Gφ = Fn oφ Z be a hyperbolic group. There exists a continuous
surjection ι̂ : ∂Fn → ∂Gφ . The map ι̂ is the Cannon–Thurston map.
Theorem 4.2. [Mit97, KL15] Suppose (X, Y ) ∈ ∂ 2 Fn , and let ι̂ : ∂Fn → ∂Gφ be the
Cannon–Thurston map. Then, ι̂(X) = ι̂(Y ) if and only if (X, Y ) ∈ L(T+ ) ∪ L(T− ).
Proposition 4.3. [KL15, Proposition 4.8, Lemma 4.9] The Cannon–Thurston map
ι̂ : ∂Fn → ∂Gφ
factors through the maps
Q+ : ∂Fn → Tb+ and Q− : ∂Fn → Tb− ,
and thus induces well-defined maps
R+ : Tb+ → ∂Gφ and R− : Tb− → ∂Gφ
which are surjective and Fn -equivariant. Furthermore,
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
17
(a) R+ (T + ) ∩ R+ (Tb+ \T + ) = ∅.
Likewise, R− (T − ) ∩ R− (Tb− \T − ) = ∅.
(b) R+ (T + ) ∩ R− (T − ) = ∅.
(c) The restriction R+ |T + of R+ to the metric completion of T+ is injective.
Likewise, the restriction R− |T − of R− to the metric completion of T− is injective.
5. The directional Whitehead graph
In this section, we introduce a tool called the directional Whitehead graph to study certain
finer asymptotic relations between singular leaves. We define the graph in two ways: first,
using the Q-map and its dual lamination, and second, using the band complex associated
to a system of partial isometries and its dual lamination. For fully irreducible outer automorphisms, there is a well-studied correspondence between these objects, and we prove in
Lemma 5.5 that these two definitions agree in this setting. Our main theorem hinges on a
certain property of the directional Whitehead graphs of T−φ .
5.1. Directional Whitehead graph.
Definition 5.1. Let T be an R-tree with a very small, minimal action of Fn by isometries
and dense orbits, and let Q : ∂Fn → T be the Q-map given in Theorem 2.7. Let L(T ) be
the lamination dual to T . Let x ∈ T be a branch point, and let d be a component of T \ {x}.
The directional Whitehead graph of T at d, denoted WhT (x, d), is defined as follows. There
is an edge (Y, Y 0 ) in the graph if there exists a leaf ` = (Y, Y 0 ) ∈ L(T ) such that Q(`) = x
and for which there exists a sequence {`i }∞
i=1 ⊂ L(T ) limiting to ` and Q(`i ) ∈ d for all i.
Edges share a vertex if the corresponding leaves are asymptotic.
Remark 5.2. In this paper, the only property of the directional Whitehead graph that we
are interested in is whether the graph contains more than one edge. An interesting problem
is to investigate other properties of the graph such as connectivity.
The edges in the directional Whitehead graph, by definition, correspond to leaves which
are limits of other leaves; that is, every edge of a directional Whitehead graph corresponds
to a leaf in the derived lamination; see Definition 2.20. Recall Lemma 2.21: if φ ∈ Out(Fn )
is fully irreducible, then L(T− )0 = Λφ+ .
The second definition of the directional Whitehead graph is given in terms of systems of
partial isometries. The second definition is used in Lemma 6.6.
Definition 5.3. Let (K, A) be a system of partial isometries, and let B be the corresponding
band complex. Let x ∈ K, and let d be a direction of K at x. The directional Whitehead
graph of d with respect to (K, A), denoted Wh(K,A) (x, d), is defined as follows. There is an
edge (ξ, η) in the graph Wh(K,A) (x, d) if there exists a leaf (ξ, η) ∈ L(B) based at x and
sequences of leaves {(ξi , ηi )}i∈N based at xi ∈ d for all i and so that ξi → ξ, ηi → η. See
Figure 5.1 for an illustration.
18
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
α1
α3
B
Wh(x)
d1
x
d2
α1
Wh(x, d1 )
α3
α1
α3
α2
α2
α2
α1
Wh(x, d2 )
α2
α3
B0
Wh(x)
d1
x
α2
d2
α1
α3
α2
Wh(x, d1 )
α1
Wh(x, d2 )
α3
α3
α2
α2
Figure 5.1. On the left are local pictures of band complexes B and B0 ; the colored lines represent leaves of the lamination. To the right of each complex is drawn
Wh(x), the ideal Whitehead graph at x, and Wh(x, di ), the directional Whitehead
graph at x in direction di . Note that if the bands overlap, then a directional Whitehead graph may contain more than one edge.
Remark 5.4. In the above definition, limi→∞ xi = x since Q is continuous with respect to
the observers’ topology by Theorem 2.7.
Lemma 5.5. Let T be a very small, indecomposable Fn -tree. For every x ∈ T and direction
d at x there exists a compact subtree K ⊂ T and a reduced system of partial isometries
S = (K, A) so that x ∈ K, the subtree K contains a germ in the direction d, and,
WhT (x, d) = Wh(K,A) (x, d).
Proof. Let e be an edge of WhT (x, d). By Definition 5.1, there exists a leaf ` ∈ L(T )
corresponding to e so that Q(`) = x, and there exists a sequence {`i }∞
i=1 ⊂ L(T ) so that
Q(`i ) = xi ∈ d and limi→∞ `i = `. Let A be a basis of Fn so that ` ∈ CA (a translation of
the original basis will do the trick). Since CA is an open neighborhood of ` ∈ ∂ 2 Fn there
exists an N ∈ N so that for all i > N , `i ∈ CA . We truncate the first N elements of the
sequence {`i }. Thus, x = Q− (`) ∈ ΩA and xi = Q− (`i ) ∈ ΩA . The space KA is the convex
hull of ΩA in T , so {xi }∞
i=1 ∪ {x} ⊂ KA . Hence, KA contains x and the germ corresponding
to the direction d. By Theorem 2.39, L(T ) ∩ CA = L(B). Therefore, `i ∈ L(B) for all i, and
converges to ` ∈ L(B). Hence, there is a corresponding edge in Wh(K,A) (x, d).
As for the other direction, an edge e in Wh(K,A) (x, d) corresponds to a leaf ` ∈ L(B) =
CA ∩ L(T ) based at x such that there exist leaves `i ∈ L(B) based at xi ∈ d so that
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
19
limi→∞ `i = `. Thus, Q(`i ) = xi , Q(`) = x, and hence ` corresponds to an edge of
WhT (x, d).
6. The T± -pattern and K3,3 subcomplexes.
Notation 6.1. Throughout this section, suppose that φ ∈ Out(Fn ) is fully irreducible, and
let T+ and T− denote the attracting and repelling trees for φ, respectively.
An illustration of the following definition appears in Figure 6.1.
Definition 6.2. The outer automorphism φ satisfies the T± -pattern if the following holds.
There exists a point a ∈ T− such that |Q−1
− (a)| ≥ 3, and there exists a direction d at a
containing two points b, c ∈ d with
−1
|Q−1
− (b)| = |Q− (c)| = 2,
which have the following properties. There exist leaves `1 , `2 , `b , `c ∈ L(T− ) such that
Q− (`1 ) = Q− (`2 ) = a
Q− (`b ) = b
Q− (`c ) = c,
+
+
+
+
+
+
+
and (`+
1 ∩ `b ) ∪ (`2 ∩ `c ) ⊂ T+ is a star with midpoint y ∈ Int(`1 ∩ `b ) ∩ Int(`2 ∩ `c ).
Remark 6.3. Suppose φ ∈ Out(Fn ) can be represented by an automorphism of Fn which is
induced by a pseudo-Anosov homeomorphism of a surface with negative Euler characteristic
and non-empty boundary. Then, φ does not satisfy the T± -pattern. Indeed, the leaves of
the attracting and repelling laminations can be realized as (non-crossing) embedded lines
in H2 with dual trees Tb+ and Tb− , respectively. In this case, a direction at a point a ∈ T−
corresponds to a half-space in H2 . We leave the details to the reader.
Proposition 6.4. If φ satisfies the T± -pattern (or the T∓ -pattern), then there exists an
embedding of K3,3 in ∂Gφ , where K3,3 denotes the complete bipartite graph with two vertex
sets of size three.
Tb+
`+
1
`+
b
`+
b
`+
c
`+
1
Tb−
`+
2
`+
c
T−
b
a
c
`+
2
Figure 6.1. The T± -pattern. The right side of the figure is contained in T − . The
configuration in Tb+ is either like the left side or the middle of the figure depending
+
+
+
on whether (`+
1 ∩ `b ) ∪ (`2 ∩ `c ) ⊂ T+ in Definition 6.2 has 3 or 4 prongs.
20
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
B1
A1
Tb+
C1
p1
p1
B1
y
C1
b
a
y
p3
p3
p2
B2
Tb−
A4
A1
C2
A2
A3
p2
B2
c
C2
A3
A2
R+
Υ
R−
α
β
γ
∂Gφ
π1
π2
π3
Figure 6.2. An embedding of K3,3 in ∂Gφ built using the T± -pattern. The points
πi and Υ are the images of pi and y under the map R+ , respectively. In addition,
α = R+ (Ai ) = R− (a), β = R+ (Bi ) = R− (b), and γ = R+ (Ci ) = R− (c).
Proof. Suppose φ satisfies the T± -pattern. An embedding of K3,3 in ∂Gφ is illustrated in
Figure 6.2 and is described as follows. Recall from Section 4 that the Cannon–Thurston
map ι̂ : ∂Fn → ∂Gφ factors through the maps Q+ : ∂Fn → T+ and Q− : ∂Fn → ∂T− . The
induced maps R+ : T+ → ∂Gφ and R− : T− → ∂Gφ are embeddings restricted to Int(T+ )
and Int(T− ), and the images of Int(T+ ) and Int(T− ) in ∂Gφ are disjoint. Thus, the images
of the interiors of the colored paths in Tb+ drawn in Figure 6.1 do not intersect in ∂Gφ . The
endpoints of these paths on ∂T+ are identified, via R+ and R− , to points in the interior of
T− to form a K3,3 .
Lemma 6.5. An outer automorphism φ satisfies the T± -pattern if and only if there exists
a point a ∈ T− and a direction d at a such that WhT− (a, d) contains at least three vertices.
Proof. Suppose there exist a point a ∈ T− and a direction d at a so that WhT− (a, d) contains
three vertices, corresponding to points α1 , α2 , α3 ∈ ∂Fn . By definition, Q− (αi ) = a and each
αi is an end of a leaf of the lamination L(T− ). Thus, there exist two leaves `1 6= `2 ∈ L(T− )
so that {α1 , α2 , α3 } ⊂ (`1 ∪ `2 ). By the definition of the directional Whitehead graph, there
∞
exist sequences {σk }∞
k=1 , {ρk }k=1 ⊂ L(T− ) such that limk→∞ σk = `1 , limk→∞ ρk = `2 , and
Q− (σk ), Q− (ρk ) ∈ d for all k. By Proposition 2.30, the realization in T+ of the leaves
+
corresponding to edges of a Whitehead graph is a star. Therefore, `+
1 ∩ `2 is either a point y
or a ray initiating at a point y. Since limk→∞ σk = `1 , by Proposition 2.28 σk+ ∩ `+
1 is a
non-trivial segment containing y in its interior for a large enough k. Likewise, for large
+
enough k, ρ+
y in
k ∩ `2 is a non-trivial segment containing
its interior. Set `a = σk and
+
+
+ is a star with y its midpoint.
`b = ρk for large enough k so that `+
∩
`
∪
`
∩
`
c
1
2
b
Therefore, φ satisfies the T± -pattern, concluding one direction of the proof.
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
21
Conversely, suppose φ satisfies the T± -pattern. We will show `1 and `2 yield distinct
edges of W hT− (a, d). By Lemma 2.34, there exists a homothety h+ : T+ → T+ that fixes
+
+
+
the middle y of the star (`+
1 ∩ `b ) ∪ (`2 ∩ `c ) in T+ . The homothety h+ corresponds to
a principal automorphism Φ representing φ. We replace φ by a rotationless power so that
h+ fixes all directions at y and ∂Φ(`j ) = `j for j = 1, 2. Let h− be the corresponding
homothety of T− guaranteed by Lemma 2.34. By Lemma 2.36,
h− (a) = h− (Q− (`1 )) = Q− (Φ(`1 )) = Q− (`1 ) = a,
where a ∈ T− is the point given in the definition of the T± -pattern. Thus, h− permutes the
directions at a. We replace Φ, h+ , and h− by powers so that h− fixes the directions at a.
+
Since h+ is a homothety fixing the directions at y, limk→∞ hk+ (`+
b ) = `1 . By Lemma 2.28,
k
k
limk→∞ Φ (`b ) = `1 . Likewise, limk→∞ Φ (`c ) = `2 . Recall that b = Q− (`b ) and c = Q− (`c )
belong to the same component d of T− \ {a} by assumption. Since h− fixes the directions
at a,
Q− (Φk (`b )) = hk− (Q− (`b )) ∈ d
and likewise, Q− (Φk (`c )) ∈ d. Therefore, `1 and `2 yield distinct edges in WhT− (a, d),
concluding the proof.
Lemma 6.6. There exists a directional Whitehead graph of a point in T− with at least three
vertices if and only if the Rips Machine for T− never halts.
Proof. Suppose the Rips machine never halts, and let (Km , Am ) denote the output after
the mth iteration of the Rips machine. This means that the tree T− is not of surface type,
and thus, Proposition 3.7 ensures that T− is of Levitt type. In particular, if am ∈ Am (for
each m ∈ N), then limm→∞ diam(dom(am )) = 0.
By Proposition 3.15, for each m there exists a compact non-trivial interval Im ⊂ Km and
three distinct elements am , bm , cm ∈ Am such that Im ⊂ dom(am ) ∩ dom(bm ) ∩ dom(cm ).
We may choose {Im } to be nested because for each m, if an interval is contained in the
domain of three partial isometries in Am , then it is contained in the domain of three partial
isometries in Am−1 . We may also choose the sequences of partial isometries so that for each
∞
\
f = a, b, c, fm |Km+1 = fm+1 . Thus, there exists a point x with x ∈
Im . Moreover,
m=1
because there are only finitely many directions at x, there exists at least one germ d based
at x, so that for each m, the interval Im contains a subsegment in the germ d.
To find an infinite sequence of leaves, let xm 6= x be an extremal point of dom(am+1 ) so
that the segment [x, xm ] is contained in the germ d. By Proposition 3.3, xm ∈ Ω, where
Ω is the limit set of the system (see Definition 2.3). So, there exists a leaf σm ∈ L(B− )
with Q− (σm ) = xm . In the alphabet A0 , σm = · · · .a0 · · · for all m ∈ N; see Notation 2.4.
Since L(B− ) is compact, there is a subsequence of {σm }m∈N that converges in ∂ 2 Fn to a
leaf `1 . Moreover, diam(dom(am )) → 0, so limm→∞ xm = x. The map Q− is continuous
with respect to the observers’ topology, so Q− (`1 ) = limm→∞ Q− (σm ) = limm→∞ xm = x.
Thus, there exists an edge e in WhK (x, d) corresponding to `1 . Since xm ∈ dom(am ), in the
alphabet Am , σm = · · · .am · · · . Thus, one half-leaf of `1 is the limit of half-leaves which
22
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
begin with {am }m∈N . Without loss of generality, assume that the other end of `1 is the limit
of half-leaves which begin with {bm }m∈N . Perform the same construction with {cm }m∈N
instead of {am }m∈N to obtain a sequence {ym }m∈N ⊂ d and a sequence {ρm }m∈N ⊂ L(B− )
so that ρm = · · · .cm · · · in the alphabet Am and limm→∞ ρm = `2 . Then Q− (`2 ) = x. The
leaf `2 is the limit of half-leaves which begin with {cm }m∈N , so `1 6= `2 . Therefore, there
exists another edge e0 6= e in Wh(K,A) (x, d) where e0 corresponds to `2 , and e corresponds
to `1 . By Lemma 5.5, WhT− (x, d) contains two edges as well.
Conversely, suppose WhT− (x, d) contains three vertices. By Lemma 5.5, there exists
a system of partial isometries (K, A) so that Wh(K,A) (x, d) contains three vertices. So,
Wh(K,A) (x, d) contains two edges corresponding to leaves `1 , `2 in L(B− ). By definition,
there exist sequences {σk }k∈N , {ρk }k∈N ⊂ L(B− ) so that limk→∞ σk = `1 , limk→∞ ρk = `2 ,
and Q− (σk ), Q− (ρk ) ∈ d. For each m ∈ N, there are two distinct pairs (a, b), (c, f ) of
partial isometries a, b, c, f ∈ Am so that in the alphabet of Am , `1 = · · · b−1 .a · · · and `2 =
· · · f −1 .c · · · . (Note that the pairs may share one element.) Therefore, x ∈ dom(a)∩dom(b)∩
dom(c) ∩ dom(f ). Let k be large enough so that in the alphabet Am , σk = · · · b−1 .a · · · and
ρk = · · · f −1 .c · · · . Thus, Q− (σk ) ∈ dom(a) ∩ dom(b) and Q(ρk ) ∈ dom(c) ∩ dom(f ). Since
the domain of a partial isometry is convex, the arc [Q− (σk ), x] is contained in dom(a) ∩
dom(b), and the arc [Q− (ρk ), x] is contained in dom(c) ∩ dom(f ). Hence, the domains of
a, b, c, f intersect in a non-trivial arc (in the direction d). By [CH14], there exists a free
band, that is, an interval J ∈ K and a unique partial isometry g ∈ Am so that J ⊂ dom(g).
Thus, for all m ∈ N the Rips Machine does not stop at the m-th step.
7. The boundary ∂Gφ
Main Theorem. If Gφ is a hyperbolic group, then ∂Gφ contains a copy of the complete
bipartite graph K3,3 .
Proof. If φ is fully irreducible, let T+ and T− be the attracting and repelling trees of φ in
the boundary of CVn . By Proposition 3.7, either T− or T+ has Levitt type. Without loss of
generality, suppose T− has Levitt type. By Lemma 6.6, some directional Whitehead graph
of T− contains more than one edge. By Lemma 6.5, φ satisfies the T± pattern; hence, by
Lemma 6.4, the boundary ∂Gφ contains an embedded copy of K3,3 .
If φ is not fully irreducible, then there exists a free factor A0 < Fn and a power φk
so that [A0 ], the conjugacy class of A0 in Fn , is φk -invariant. Let A be a minimal such
factor. Since Gφ is hyperbolic, φ is atoroidal, hence rank(A) ≥ 2. Moreover, φ induces
a well-defined outer automorphism φ0 on A [HM13, Fact 1.4]. Since φ is atoroidal, so is
φ0 , and hence rank(A) ≥ 3. By the minimality of A, the outer automorphism φ0 is fully
irreducible. Therefore, the boundary of G0 = A oφ0 Z contains an embedded copy of K3,3 by
the previous paragraph. The subgroup G0 quasi-isometrically embeds in Gφk [Mit98]. Thus,
∂G0 embeds into ∂Gφk . The group Gφk is a finite-index subgroup of Gφ , so the boundaries
of Gφ and Gφk are homeomorphic.
In what follows, we describe the work of Kapovich-Kleiner [KK00] relevant to this paper.
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
23
Theorem 7.1. [KK00, Theorem 4] Let G be a hyperbolic group which does not split over
a finite or virtually cyclic subgroup, and suppose ∂G is 1-dimensional. Then one of the
following holds:
(1) ∂G is a Menger curve,
(2) ∂G is a Sierpinski carpet,
(3) ∂G is homeomorphic to S1 , and G maps onto a Schwartz triangle group with finite
kernel.
Theorem 7.1 quickly follows from a compilation of results: the characterization of the
Menger curve [And58a, And58b], the characterization of the Sierpinski carpet [Why58],
and the structure of local and global cut points in the boundary of a hyperbolic group
[BM91, Bow99, Swa96].
Corollary 7.2. [KK00, Corollary 15] Let F be a finitely generated free group, and φ : F → F
an atoroidal automorphism. If no power of φ preserves a free splitting of F , then the Gromov
boundary of Gφ := F oφ Z is the Menger curve. In particular, if φ is fully irreducible, then
∂Gφ is the Menger curve.
Remark 7.3. The hypothesis of the above corollary in [KK00] states that φ is irreducible,
not fully irreducible. However, in the proof they obtain a contradiction by producing a φj invariant free decomposition of F for some positive j > 0. This property only contradicts
that φ is fully irreducible.
To prove the above corollary, Kapovich-Kleiner [KK00] first show that if φ ∈ Out(Fn ) is
fully irreducible (or no power of φ preserves a free splitting of F ), then Gφ does not split over
a trivial or cyclic subgroup. (See also Brinkmann [Bri02].) Then, Theorem 7.1 implies ∂Gφ
is either the circle, the Sierpinski carpet, or the Menger curve. At this point, using our Main
Theorem, one can directly rule out the circle and Sierpinski carpet, since these spaces are
planar. Alternatively, Kapovich–Kleiner rule out the circle by applying the work of Tukia–
Gabai–Casson–Jungreis [Tuk88, Gab92, CJ94], which classifies the hyperbolic groups with
boundary homeomorphic to S 1 as precisely the groups that act discretely and cocompactly
by isometries on the hyperbolic plane. Finally, Kapovich–Kleiner prove a rather deep result:
if G is a hyperbolic group with Sierpinski carpet boundary, then G together with the
stabilizers of the peripheral circles of the Sierpinski carpet forms a Poincaré duality pair
(see [KK00, Corollary 12]). In particular, they conclude that in this case χ(G) < 0. Since
χ(Gφ ) = 0, the boundary ∂Gφ is the Menger curve.
References
[And58a] R. D. Anderson. A characterization of the universal curve and a proof of its homogeneity. Ann.
of Math. (2), 67:313–324, 1958.
[And58b] R. D. Anderson. One-dimensional continuous curves and a homogeneity theorem. Ann. of Math.
(2), 68:1–16, 1958.
[BF94]
Mladen Bestvina and Mark Feighn. Outer limts, 1994. Preprint available at
http://andromeda.rutgers.edu/ feighn/research.html.
24
YAEL ALGOM-KFIR, ARNAUD HILION, AND EMILY STARK
[BFH97] M. Bestvina, M. Feighn, and M. Handel. Laminations, trees, and irreducible automorphisms of
free groups. Geom. Funct. Anal., 7(2):215–244, 1997.
[BM91] Mladen Bestvina and Geoffrey Mess. The boundary of negatively curved groups. J. Amer. Math.
Soc., 4(3):469–481, 1991.
[Bow98] Brian H. Bowditch. Cut points and canonical splittings of hyperbolic groups. Acta Math.,
180(2):145–186, 1998.
[Bow99] B. H. Bowditch. Connectedness properties of limit sets. Trans. Amer. Math. Soc., 351(9):3673–
3686, 1999.
[Bri00]
P. Brinkmann. Hyperbolic automorphisms of free groups. Geom. Funct. Anal., 10(5):1071–1089,
2000.
[Bri02]
Peter Brinkmann. Splittings of mapping tori of free group automorphisms. Geom. Dedicata,
93:191–203, 2002.
[CH12]
Thierry Coulbois and Arnaud Hilion. Botany of irreducible automorphisms of free groups. Pacific
J. Math., 256(2):291–307, 2012.
[CH14]
Thierry Coulbois and Arnaud Hilion. Rips induction: index of the dual lamination of an R-tree.
Groups Geom. Dyn., 8(1):97–134, 2014.
[CHL07] Thierry Coulbois, Arnaud Hilion, and Martin Lustig. Non-unique ergodicity, observers’ topology
and the dual algebraic lamination for R-trees. Illinois J. Math., 51(3):897–911, 2007.
[CHL09] Thierry Coulbois, Arnaud Hilion, and Martin Lustig. R-trees, dual laminations and compact
systems of partial isometries. Math. Proc. Cambridge Philos. Soc., 147(2):345–368, 2009.
[CHR15] Thierry Coulbois, Arnaud Hilion, and Patrick Reynolds. Indecomposable FN -trees and minimal
laminations. Groups Geom. Dyn., 9(2):567–597, 2015.
[CJ94]
Andrew Casson and Douglas Jungreis. Convergence groups and Seifert fibered 3-manifolds. Invent.
Math., 118(3):441–456, 1994.
[CL95]
Marshall M. Cohen and Martin Lustig. Very small group actions on R-trees and Dehn twist
automorphisms. Topology, 34(3):575–617, 1995.
[CM87] Marc Culler and John W. Morgan. Group actions on R-trees. Proc. London Math. Soc. (3),
55(3):571–604, 1987.
[CV86]
Marc Culler and Karen Vogtmann. Moduli of graphs and automorphisms of free groups. Invent.
Math., 84(1):91–119, 1986.
[FH09]
M. Feighn and M. Handel. Abelian subgroups of Out(Fn ). Geom. Topol., 13(3):1657–1727, 2009.
[Gab92] David Gabai. Convergence groups are Fuchsian groups. Ann. of Math. (2), 136(3):447–510, 1992.
[GJLL98] Damien Gaboriau, Andre Jaeger, Gilbert Levitt, and Martin Lustig. An index for counting fixed
points of automorphisms of free groups. Duke Math. J., 93(3):425–452, 1998.
[Hil07]
Arnaud Hilion. On the maximal subgroup of automorphisms of a free group FN which fix a point
of the boundary ∂FN . Int. Math. Res. Not. IMRN, (18):Art. ID rnm066, 21, 2007.
[HK09]
G. Christopher Hruska and Bruce Kleiner. Erratum to: “Hadamard spaces with isolated flats”
[Geom. Topol. 9 (2005), 1501–1538; mr2175151]. Geom. Topol., 13(2):699–707, 2009.
[HM11] M. Handel and L. Mosher. Axes in Outer Space. Number 1004. Amer Mathematical Society, 2011.
[HM13] Michael Handel and Lee Mosher. Subgroup decomposition in Out(Fn ), part I: Geometric Models.
arXiv:1302.2378, 2013.
[KK00]
Michael Kapovich and Bruce Kleiner. Hyperbolic groups with low-dimensional boundary. Ann.
Sci. École Norm. Sup. (4), 33(5):647–669, 2000.
[KL14]
Ilya Kapovich and Martin Lustig. Invariant laminations for irreducible automorphisms of free
groups. Q. J. Math., 65(4):1241–1275, 2014.
[KL15]
Ilya Kapovich and Martin Lustig. Cannon-Thurston fibers for iwip automorphisms of FN . J. Lond.
Math. Soc. (2), 91(1):203–224, 2015.
[Kur30] Kazimierz Kuratowski. Sur le problme des courbes gauches en topologie. Fund. Math., 15(3):271–
283, 1930.
THE VISUAL BOUNDARY OF HYPERBOLIC FREE-BY-CYCLIC GROUPS
[LL03]
25
Gilbert Levitt and Martin Lustig. Irreducible automorphisms of Fn have north-south dynamics
on compactified outer space. J. Inst. Math. Jussieu, 2(1):59–72, 2003.
[Mit97]
Mahan Mitra. Ending laminations for hyperbolic group extensions. GAFA, pages 379–402, 1997.
[Mit98]
Mahan Mitra. Cannon-Thurston maps for hyperbolic group extensions. Topology, 37(3):527–538,
1998.
[Pau96] Frédéric Paulin. Un groupe hyperbolique est déterminé par son bord. J. London Math. Soc. (2),
54(1):50–74, 1996.
[Rua05] Kim Ruane. CAT(0) boundaries of truncated hyperbolic space. Topology Proc., 29(1):317–331,
2005. Spring Topology and Dynamical Systems Conference.
[Swa96] G. A. Swarup. On the cut point conjecture. Electron. Res. Announc. Amer. Math. Soc., 2(2):98–
100, 1996.
[Tuk88] Pekka Tukia. Homeomorphic conjugates of Fuchsian groups. J. Reine Angew. Math., 391:1–54,
1988.
[Vog02] Karen Vogtmann. Automorphisms of free groups and outer space. In Proceedings of the Conference
on Geometric and Combinatorial Group Theory, Part I (Haifa, 2000), volume 94, pages 1–31,
2002.
[Why58] G. T. Whyburn. Topological characterization of the Sierpiński curve. Fund. Math., 45:320–324,
1958.
| 4math.GR
|
arXiv:1602.01942v3 [math.ST] 15 Jan 2018
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES BY
AUTO-REGRESSION
FRANÇOIS ROUEFF AND ANDRÉS SÁNCHEZ-PÉREZ
Abstract. In this contribution we introduce weakly locally stationary time series through
the local approximation of the non-stationary covariance structure by a stationary one.
This allows us to define autoregression coefficients in a non-stationary context, which, in
the particular case of a locally stationary Time Varying Autoregressive (TVAR) process,
coincide with the generating coefficients. We provide and study an estimator of the time
varying autoregression coefficients in a general setting. The proposed estimator of these
coefficients enjoys an optimal minimax convergence rate under limited smoothness conditions. In a second step, using a bias reduction technique, we derive a minimax-rate estimator for arbitrarily smooth time-evolving coefficients, which outperforms the previous one
for large data sets. In turn, for TVAR processes, the predictor derived from the estimator
exhibits an optimal minimax prediction rate.
1. Introduction
In many applications, one is interested in predicting the next values of an observed time
series. It is the case in various areas like finance (stock market, volatility on prices), social
sciences (population studies), epidemiology, meteorology and network systems (Internet
traffic). Autoregressive processes have been used successfully in a stationary context for
several decades. On the other hand, in a context where the number of observations can be
very large, the usual stationarity assumption has to be weakened to take into account some
smooth evolution of the environment.
Several prediction methods developed in signal processing are well known to adapt to
a changing environment. This is the case of the wide spread recursive least square algorithms. The initial goal of these methods is to provide an online procedure for estimating a
regression vector with low numerical cost. Such methods usually rely on a forgetting factor
or a gradient step size γ and they can be shown to be consistent in a stationary environment
when γ decreases adequately to zero (see e.g. Duflo (1997)). Even when the environment is
changing, that is, when the regression parameter evolves along the time, a “small enough”
γ often yields a good tracking of the evolving regression parameter. In order to have a
sound and comprehensive understanding of this phenomenon, an interesting approach is to
consider a local stationarity assumption, as successfully initiated in Dahlhaus (1996b) by
relying on a non-stationary spectral representation introduced in Priestley (1965); see also
Dahlhaus (2012) and the references therein for a recent overview. The basic idea is to provide an asymptotic analysis for the statistical inference of non-stationary time series such as
2000 Mathematics Subject Classification. 62M20, 62G99, 62M10, 68W27.
Key words and phrases. locally stationary time series, auto-regression coefficients, time varying autoregressive processes, minimax-rate prediction.
This work has been partially supported by the Conseil régional d’Île-de-France under a doctoral allowance of
its program Réseau de Recherche Doctoral en Mathématiques de l’Île de France (RDM-IdF) for the period 2012
- 2015 and by the Labex LMH (ANR-11-IDEX-003-02).
1
2
François Roueff and Andrés Sánchez-Pérez
time varying autoregressive (TVAR) processes using local stationary approximations. The
analysis of the Normalized Least Mean Squares (NLMS) algorithm for tracking a changing
autoregression parameter in this framework is tackled in Moulines et al. (2005). Such an
analysis is based on the usual tools of non-parametric statistics. The TVAR parameter θ is
seen as the regular samples of a smooth Rd -valued function. An in-fill asymptotic allows
one to derive the rates of convergence of the NLMS estimator for estimating this function
within particular smoothness classes of functions. As shown in Moulines et al. (2005),
it turns out that the NLMS algorithm provides an optimal minimax rate estimator of the
TVAR parameter with Hölder smoothness index β ∈ (0, 1]. However it is no longer optimal
for β > 1, that is, when the TVAR parameter is smoother than a continuously differentiable
function. An improvement of the NLMS is proposed in Moulines et al. (2005) to cope
with the case β ∈ (0, 2] but, to the best of our knowledge, there is no available method
neither for the θ minimax-rate estimation nor for the minimax-rate prediction when β > 2,
that is when the TVAR parameter is smoother than a two-times continuously differentiable
function.
In the present work, our main contribution is twofold. First we extend the concept
of time-varying linear prediction coefficients to a general class of weakly locally stationary processes, which includes the class of locally stationary processes as introduced in
Dahlhaus (1996b). In the specific case of a TVAR process, these coefficients correspond to
the time-varying autoregression parameters. Second, we show that the tapered Yule-Walker
estimator introduced in Dahlhaus and Giraitis (1998) for TVAR processes also applies to
this general class and is minimax-rate for Hölder indices up to β = 1 for asymmetric tapers
and up to β = 2 for symmetric ones. Moreover, by applying a bias reduction technique,
we derive a new estimator which is minimax-rate for any arbitrarily large Hölder index
β. By achieving this goal, we provide a theoretically justified construction of predictors
that can be chosen optimally, depending on how smoothly the time varying spectral density evolves along the time. On the other hand, in practical situations, one may not have
a clear view on the value of the smoothness index β and one should rely on data driven
methods that are therefore called adaptive. This problem was recently tackled in Giraud
et al. (2015) . More precisely, using aggregation techniques introduced in the context of
individual sequences prediction (see e.g. Cesa-Bianchi and Lugosi (2006)) and statistical
learning (see e.g. Barron (1987)), one can aggregate sufficiently many predictors in order
to build a minimax predictor which adapts to the unknown smoothness β of the time varying parameter of a TVAR process. However, a crucial requirement in Giraud et al. (2015)
is to rely on β-minimax-rate sequences of predictors for any β > 0. Our main contribution
here is to fill this gap, hence achieving to solve the problem of the adaptive minimax-rate
linear forecasting of locally stationary TVAR processes with coefficients of any (unknown,
arbitrarily large) Hölder smoothness index.
The paper is organized as follows. In Section 2, we introduce a definition of weakly
locally stationary time series, the regression problem investigated in this work in relation
with the practical prediction problem, and the tapered Yule-Walker estimator under study.
General results on this estimator are presented in Section 3 and a minimax rate estimator
is derived. The particular case of TVAR processes is treated in Section 4. Numerical
experiments illustrating these results can be found in Section 5. Postponed proofs and
useful lemmas are provided in the appendices.
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
3
2. General setting
In the following, non-random vectors and sequences are denoted using boldface symP
bols, kxk denotes the Euclidean norm of x, kxk = ( i |xi |2 )1/2 , and kxk1 its `1 norm,
P
kxk1 = i |xi |. If f is a function, k f k∞ = sup x | f (x)| corresponds to its sup norm. If A
is a matrix, kAk denotes its spectral norm, kAk = sup{kAxk , kxk ≤ 1}. We moreover denote
`1 (N) = {x ∈ RN s.t. kxk1 < ∞} and
`+1 (N) = {x ∈ RN
+ s.t. kxk1 < ∞} .
2.1. Main definitions. We consider a doubly indexed time series (Xt,T )t∈Z,T ∈N∗ . Here t
refers to a discrete time index and T is an additional index indicating the sharpness of the
local approximation of the time series (Xt,T )t∈Z by a stationary one. Coarsely speaking,
(Xt,T )t∈Z,T ∈N∗ is considered to be weakly locally stationary if, for T large, given a set S T of
sample indices such that t/T ≈ u over t ∈ S T , the sample (Xt,T )t∈S T can be approximately
viewed as the sample of a weakly stationary time series depending on the rescaled location
u. Note that u is a continuous time parameter, sometimes referred to as the rescaled time
index. Following Dahlhaus (1996b), T is usually interpreted as the number of available
observations, in which case all the definitions are restricted to 1 ≤ t ≤ T and u ∈ [0, 1].
However this is not essential in our mathematical derivations and it is more convenient to
set t ∈ Z and u ∈ R for presenting our setting.
We use the following class of functions. For α ∈ (0, 1] the α−Hölder semi-norm of a
function f : R → Cd is defined by
|f|α,0 =
kf(s) − f(s0 )k
.
|s − s0 |α
0<|s−s0 |<1
sup
This semi-norm is used to build a norm for any β > 0 as it follows. Let k ∈ N and α ∈ (0, 1]
be such that β = k + α. If f is k times differentiable on R, we define
|f|β = f (k) α,0 + max f (s)
0≤s≤k
∞
,
and |f|β = ∞ otherwise. For R > 0 and β > 0, the (β, R)− Hölder ball of dimension d is
denoted by
n
o
Λd (β, R) = f : R → Cd , such that |f|β ≤ R .
We first introduce definitions for the time varying covariance and the local covariance
functions.
Definition 2.1 (Time varying covariance function). Let (Xt,T )t∈Z,T ∈N∗ be an array of random
variables with finite variances. The local time varying covariance function γ∗ is defined for
all t ∈ Z, T ∈ N∗ and ` ∈ Z as
γ∗ (t, T, `) = cov Xt,T , Xt−`,T .
(2.1)
Definition 2.2 (Local covariance function and local spectral density). A local spectral
density f is a R2 → R+ function, (2π)-periodic and locally integrable with respect to the
second variable. The local covariance function γ associated with the local spectral density
f is defined on R × Z by
Z π
γ (u, `) =
ei`λ f (u, λ) dλ .
(2.2)
−π
In (2.2), the variable u should be seen as rescaled time index (in R), ` as a (non-rescaled)
time index and λ as a frequency (in [−π, π]). Recall that, by the Herglotz theorem (see
Brockwell and Davis (2002, Theorem 4.3.1)), Equation (2.2) guaranties that for any u ∈ R,
4
François Roueff and Andrés Sánchez-Pérez
(γ (u, `))`∈Z is indeed the autocovariance function of a stationary time series. Now, we can
state the definition of weakly locally stationary processes that we use here.
Definition 2.3 (Weakly locally stationary processes). Let (Xt,T )t∈Z,T ≥T0 be an array of
random variables with finite variances and T 0 , β, R > 0. We say that (Xt,T )t∈Z,T ≥T0 is
(β, R)-weakly locally stationary with local spectral density f if, for all λ ∈ R, we have
f (·, λ) ∈ Λ1 (β, R), and the time varying covariance function γ∗ of (Xt,T )t∈Z,T ≥T0 and the
local covariance function γ associated with f satisfy, for all t ∈ Z and T ≥ T 0 ,
t
(2.3)
γ∗ (t, T, `) − γ , ` ≤ R T − min(1,β) .
T
Let us give some examples fulfilling this definition.
Example 1. Locally stationary processes were introduced in a general fashion by Dahlhaus
(1996b) using the spectral representation
Z π
Xt,T =
eitλ A0t,T (λ) ξ (dλ) ,
(2.4)
−π
where ξ(dλ) is the spectral representation of a white noise and (A0t,T )t∈Z,T ∈N∗ is a collection
of transfer functions such that there exist a constant K and a (unique) 2π− periodic function
A : R × R → C with A(u, −λ) = A(u, λ) such that for all T ≥ 1,
t
K
sup A0t,T (λ) − A , λ ≤
.
(2.5)
T
T
t∈Z,λ∈R
Provided adequate smoothness assumptions on the time varying transfer function A, this
class of locally stationary processes satisfies Definition 2.3 (see Dahlhaus (1996a, Section 1)) for some β ≥ 1 and f (u, λ) = |A(u, λ)|2 . The case β ∈ (0, 1] can be obtained by
raising T to the power β in (2.5).
Example 2 (Time-varying Causal Bernoulli Shift (TVCBS)). Let ϕ : R × RN → R. Consider, for all T ≥ 1 and t ∈ Z, a mapping ϕ0t,T : RN → R defining the random variables
Xt,T = ϕ0t,T ((ξt−k )k≥0 ) ,
(2.6)
where (ξt )t∈Z are i.i.d. We assume that E[|ξ0 |2r ] < ∞ for some r ≥ 1 and that there exist
β, K > 0 and (ψk )k≥0 ∈ `+1 (N), such that, for all T ≥ 1, t ∈ Z, u, u0 ∈ R and x ∈ RN ,
r
∞
X
0
ϕt,T (x) ≤ K 1 +
ψk |xk | ,
(2.7)
k=0
t
ϕ0t,T (x) − ϕ , x
T
≤
KT
− min(1,β)
r
∞
X
1 +
ψk |xk | ,
(2.8)
k=0
It is easy to see that (Xt,T )t∈Z,T ≥T0 satisfies (2.3) with a constant R only depending on
K, β, (ψk )k∈N and E[|ξ0 |2r ], and with local covariance function γ(u, ·) defined as the covariance function of the stationary causal Bernoulli shift process (Xt (u))t∈Z defined by
Xt (u) = ϕ(u, ((ξt−k )k≥0 ). To obtain that (Xt,T )t∈Z,T ≥T0 is (β, R)-weakly locally stationary,
it thus only remains to check that (Xt (u))t∈Z admits a spectral density f (u, ·) and that the
resulting local spectral density satisfies f (·, λ) ∈ Λ1 (β, R) for all λ ∈ R.
Example 3 (TVAR(p) model). Under suitable assumptions, the TVAR process is a particular case both of Example 1 (see Dahlhaus (1996b, Theorem 2.3)) and Example 2 (see
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
Section 4). It is defined as the (unique) stable solution of the recursive equation
p
t
t
X
Xt,T =
θj
Xt− j,T + σ
ξt ,
T
T
j=1
5
(2.9)
where θ = [θ1 . . . θ p ]0 : R → R p are the time varying autoregressive coefficients and (ξt )t∈Z
are i.i.d. centred and with variance 1. This example is detailed in Section 4.
2.2. Statement of the problem. Consider a weakly locally stationary (Xt,T )t∈Z,T ≥T0 , which
we assume to have mean zero for convenience. Let d ∈ N∗ . For each t = 1, . . . , T , define
the prediction vector of order d by
2
d
X
∗
θk Xt−k,T
θt,T =
arg min E Xt,T −
θ=[θ1 ... θd ]0 ∈Rd
k=1
h
i
0
= arg min E Xt,T − θ Xt−1,T 2 ,
(2.10)
θ∈Rd
where A denotes the transpose of matrix A and X s,T = X s,T . . . X s−(d−1),T 0 .
Let Γ∗t,T be the time varying covariances matrix Γ∗t,T = (γ∗ (t − i, T, j − i); i, j = 1, . . . , d)
where γ∗ is the time varying covariance function as defined in (2.1). Provided that Γ∗t,T is
non-singular, the solution of (2.10) is given by
−1
∗
θt,T
= Γ∗t,T
γ∗t,T ,
(2.11)
0
where γ∗t,T = [γ∗ (t, T, 1) . . . γ∗ (t, T, d)]0 . Analogously to (2.11), and with the aim of
approximating the local solution of the stationary Yule-Walker equations, we set
θu
=
Γ−1
u γu ,
(2.12)
where γu = [γ(u, 1) . . . γ(u, d)] , Γu is the covariances matrix Γu = (γ(u, i − j); i, j =
1, . . . , d) and γ is the local covariance function as defined in (2.2). To ensure the above
matrices to be non-singular, we assume a lower bound on the local spectral density, yielding
the main assumption used on the model which depends on some positive constants β, R and
f− and on a local spectral density f .
(M-1) The sequence (Xt,T )t∈Z,T ≥T 0 is a (β, R)–weakly locally stationary process with
local spectral density f in the sense of Definition 2.3. Assume moreover that
the spectral density f satisfies f (u, λ) ≥ f− for all u, λ ∈ R.
The following lemma allows us to control the error of the approximation of the optimal lin∗
ear prediction coefficients θt,T
by the local ones θt/T . Its proof is postponed to appendix B.1
for convenience.
0
Lemma 2.4. Let d ∈ N∗ , β > 0, R > 0 and f− > 0. Suppose that Assumption (M-1) holds.
Then, there exist two constants C1 , T 0 > 0 depending only on d, β, R and f− such that, for
all t ∈ Z and T ≥ T 0 ,
∗
− θt/T ≤ C1 T − min(1,β) .
θt,T
(2.13)
An estimator b
θ of θ is studied in Dahlhaus and Giraitis (1998) for the model of Example 1. In the following we improve these results by deriving minimax rate properties of the
estimator of Dahlhaus and Giraitis (1998) and extensions of it in a more general setting.
In the following, the problem that we are interested is to derive a minimax rate estimator
e
θ at a given smoothness index β > 0, which means that, for such a β, the estimation
∗ 2
risk, say the quadratic risk E[ke
θt,T − θt,T
k ], can be bounded uniformly over all processes
satisfying (M-1) (among with additional assumptions), and that the corresponding rate of
6
François Roueff and Andrés Sánchez-Pérez
convergence as T → ∞ cannot be improved by any other estimator. The case β ≤ 2 is
solved in Moulines et al. (2005) for the subclass of TVAR models.
b∗ denote the best linear pre2.3. Minimax estimation for adaptive prediction. Let X
d,t,T
dictor of order d of Xt,T , which as a consequence of (2.10), reads
0
∗
∗
bd,t,T
X
= θt,T
Xt−1,T ,
b∗ the best predictor of Xt,T given its past, that is, the conditional expectaWe denote by X
t,T
tion
h
i
∗
bt,T
X
= E Xt,T X s,T , s ≤ t − 1 .
(2.14)
As explained before, the goal of this paper is to derive estimators, say e
θt,T ∈ Rd , of θt/T ,
∗
e
which is a local approximation of θt,T . In this section, we assume that θt,T is a function of
0
Xt−1,T is a legitimate predictor of Xt,T and we have the
the past X s,T , s ≤ t − 1. Then e
θt,T
following decomposition of the corresponding prediction quadratic risk
2
0
∗ 2
0
∗ 2
bt,T
bt,T
.
+E e
θt,T
Xt−1,T − X
E Xt,T − e
θt,T
Xt−1,T
= E Xt,T − X
The first term is the minimal prediction error that one would achieve with the conditional
expectation (which
the true distribution of the whole process). Furthermore, in requires
0
b∗ = θ∗ Xt−1,T and using the Minkowskii and Cauchy-Schwartz inequalities,
serting X
t,T
d,t,T
the square root of the second term can be bounded as
1/2
1/2
0
∗ 2
∗
∗ 2
bt,T
bd,t,T
bt,T
θt,T
Xt−1,T − X
≤ E X
−X
E e
4 1/4
4 1/4
∗
+ E Xt−1,T
E e
θt,T − θt,T
.
The first term in the upper bound is due to the approximation of the best predictor by the
best linear predictor of order d and can only be improved by increasing d. Note that, in the
case of the TVAR(p) model with p ≤ d, this error term vanishes. The quantity E[kXt−1,T k4 ]
is typically bounded by a constant independent of (t, T ) over the class of processes under
consideration. Hence, for a given d, the control of the prediction risk boils down to the
∗ 4
control of the estimation risk E[ke
θt,T − θt,T
k ].
To do so, we can further decompose the loss as
∗
∗
e
θt,T − θt,T
≤ e
θt,T − θt/T + θt/T − θt,T
.
(2.15)
Note that the second term is a deterministic error basically accounting for the approximation precision of the non-stationary model by a stationary one, a bound of which is
provided in Lemma 2.4 stated above.
As a result of the successive error developments above, our efforts in the following focus
on controlling the estimation risk E[ke
θt,T − θt/T k4 ] uniformly over a class of weakly locally
stationary processes with given smoothness index β > 0.
2.4. Tapered Yule-Walker estimate. Following Dahlhaus and Giraitis (1998), a local
empirical covariance function is defined as follows. It relies on a real data taper function h
and a bandwidth M which may depend on T .
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
7
Definition 2.5 (Local empirical covariance function). Consider a function h : [0, 1] → R
and M ∈ 2N∗ . The empirical local covariance function b
γT,M with taper h is defined in R × Z
as
M
1 X t1 t2
b
γT,M (u, `) =
h
h
XbuT c+t1 −M/2,T XbuT c+t2 −M/2,T ,
H M t ,t =1 M
M
1 2
t1 −t2 =`
R1
PM 2
where H M = k=1
h (k/M) ∼ M 0 h2 (x)dx is the normalizing factor. If H M = 0, we set
b
γT,M (u, `) = 0, by convention.
For h ≡ 1 in Definition 2.5 we obtain the classical covariance estimate for a centred
sample {X s , buT c − M/2 ≤ s ≤ buT c + ` + M/2}. For any d ≥ 1, based on the local empirical
covariance function b
γT,M , the d-order local empirical Yule-Walker prediction coefficients
are then defined as
b
θt,T (M) = b
Γ−1
γt,T,M ,
(2.16)
t,T,M b
where b
γt,T,M = [b
γT,M (t/T, 1) . . . b
γT,M (t/T, d)]0 , b
Γt,T,M is the matrix of empirical covarib
ances Γt,T,M = (b
γT,M (t/T, i − j); i, j = 1, . . . , d). The only way b
Γt,T,M can be singular
is when b
γT,M (t/T, `) = 0 for all ` ∈ Z (see Lemma A.5), in which case we just set
b
θt,T (M) := 0. Hence b
θt,T (M) is always well defined and always satisfies the following
(see again Lemma A.5 for the bound)
b
Γt,T,Mb
θt,T (M) = b
γt,T,M ,
and kb
θt,T (M) k ≤ 2d − 1 .
(2.17)
Using this trick, we do not find it necessary to add additional assumptions on the model
to guarantee that b
Γt,T,M is non-singular a.s., as done for instance in Dahlhaus and Giraitis
(1998), where P(Xt,T = 0) = 0 for all t ∈ Z is assumed.
3. Main results in the general framework
3.1. Additional notation and assumptions. For convenience, we introduce the following
notation. Let p > 0, q, r, s ∈ N∗ , µ be a probability distribution on R, u : R → R,
a, b : Rr → R, c ∈ Rq and a collection of random matrices {U M ∈ Rr×s , M ∈ N∗ }. We write
(i) U M = OL p ,c (u(M)) if there exists C p,c > 0, depending continuously and at most on
(p, c0 ), such that for all M ∈ N∗
h
p i1/p
max
E U Mi,, j
≤ C p,c |u (M)| ,
(3.1)
1≤i≤r,1≤ j≤s
where U M,i, j is the (i, j)-th entry of the matrix U M .
(ii) U M = OL• (µ),c (u(M)) if U M = OL p ,m p ,c (u(M)) for all p ∈ [1, ∞), where m
R p is a constant only depending on the absolute moments of the distribution µ, |x|q µ(dx),
q ≥ 1.
(iii) a(x) = O c (b(x)) if and only if there exists a constant C c depending continuously
and at most on the index c, such that for all x ∈ Rr
|a (x)| ≤ C c |b (x)| .
Concerning the function h we have the following assumption.
(H) The function h : [0, 1] → R is piecewise continuously differentiable, that is, for
0 = u0R < u1 < . . . < uN = 1, h is C1 on (ui−1 , ui ], i = 1, . . . , N . Moreover we as1
sume 0 h2 = 1, khk∞ = supu∈[0,1] |h(u)| < ∞ and kh0 k∞ = max1≤i≤N supu∈(ui−1 ,ui ] |h0 (u)| <
∞.
8
François Roueff and Andrés Sánchez-Pérez
Provided a piecewise continuously differentiable funtion h (as in (H)) and a local spectral density function f continuously differentiable on its first argument, we also consider
the following assumption, which depends on a constant C > 0 and on a probability distribution µ on R.
(C) For all ` ∈ Z and h satisfying (H), we have, for all t ∈ Z and T ≥ T 0 ,
b
γT,M (u, `) − E b
γT,M (u, `) = OL• (µ),`,khk∞ ,kh0 k∞ ,C M −1/2 .
Assumption (C) amounts to say that the tapered empirical covariance estimator b
γT,M from
a sample of length M satisfies a standard deviation rate M −1/2 in all Lq -norms. Locally
stationary processes of Example 1 satisfy it under suitable assumptions (see Dahlhaus and
Giraitis (1998, Eq. (4.4) in Theorem 4.1)). We conclude this section with a result that can
be used for processes of Example 2.
Theorem 3.1. Let (Xt,T )t∈Z,T ≥T0 be an array of random variables defined as in (2.6) where
(ξt )t∈Z are i.i.d. satisfying E[|ξ0 |q ] < ∞ for all q ≥ 1 and ϕ0t,T satisfies (2.7) for some
(ψk )k∈N ∈ `+1 (N), K > 0 and r ≥ 1. Assume moreover that there exist k0 ∈ N, (ζk )k∈N ∈
`+1 (N) such that for all t ∈ Z, T ≥ T 0 and all x, x0 ∈ RN satisfying xk = xk0 for 1 ≤ k ≤ k0 ,
r−1
X
X
ϕ0t,T (x) − ϕ0t,T (x0 ) ≤ K ζk |xk0 +k − xk0 0 +k | 1 +
ψk (|xk | + |xk0 |)
.
(3.2)
k≥0
k≥0
Suppose moreover that
X
kζk < ∞ .
(3.3)
k≥0
Then there exists a constant C only depending on r, K, k0 , (ψk )k∈N , (ζk )k∈N and the distribution of ξ0 such that (C) holds.
The proof is postponed to Appendix C.1.
3.2. Bound of the estimation risk. Our first result on the estimation risk is a uniform
approximation for the estimation error of b
θt,T (M).
Theorem 3.2. Suppose that Assumption (M-1) holds with some β > 0, f− > 0 and R > 0,
and let h : [0, 1] → R satisfying (H). Let k ∈ N and α ∈ (0, 1] be uniquely defined by
the relation β = k + α. Suppose that Assumption (C) holds for some constant C > 0 and
distribution µ. Then, for any d ≥ 1, the estimator b
θt,T (M) defined by (2.16) satisfies
!
k
X
1 M β
M `
b
+ Od, f− ,khk∞ ,kh0 k∞ ,β,R
+
+ vM ,
(3.4)
θt,T (M) − θt/T =
ah, f,`
T
M
T
`=1
where ah, f,` ∈ Rd depends only on h, f and ` and v M = OL• (µ),d, f− ,khk∞ ,kh0 k∞ ,β,R,C (M −1/2 ).
Moreover, if h(x) = h(1 − x) for x ∈ [0, 1], then ah, f,1 = 0.
The proof is postponed to Appendix B.3.
Remark 3.3. In (3.4), the choice of the taper may influence the rate of convergence through
the constant ah, f,1 , which vanishes if the taper is symmetric, that is, if h(x) = h(1 − x)
for x ∈ [0, 1]. Other constants depend on the choice of the taper but one cannot choose
tapers that ensure a further systematic improvement of the rate. The reason is given by
the definition of the constant ch,2 appearing in the proof of Theorem B.2, which implies
ch,2 > 0 for all tapers h. Consequently, for any taper h, one have that ah, f,2 , 0, except
perhaps for some particular local density functions f . Hence, as far as rates of convergence
are concerned, the only important property of the taper is that of being symmetric.
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
9
Theorem 3.2 suggests to combine several b
θt,T (M) to obtain a more accurate estimation by canceling out the first k bias terms in (3.4). This technique was already used
for eliminate one term of bias in Moulines et al. (2005, Theorem 8). It is inspired by
the Romberg’s method in numerical analysis (see Baranger and Brezinski (1991)). Let
ω = [ω0 . . . ωk ]0 ∈ Rk+1 , be the solution of the equation
Aω = e1 ,
(3.5)
where e1 = [1 0 . . . 0] is the R - vector having a 1 in the first position and zero
everywhere else and A is a (k + 1) × (k + 1) matrix with entries Ai, j = 2i j for 0 ≤ i, j ≤ k.
0
k+1
Theorem 3.4. Under the same assumptions as Theorem 3.2, the estimator
e
θt,T (M) =
k
X
ω jb
θt,T (2 j M) ,
(3.6)
j=0
with ω defined by (3.5), satisfies
!
1 M β
e
+
θt,T (M) − θt/T = Od, f− ,khk∞ ,kh0 k∞ ,β,R
+ OL• (µ),d, f− ,khk∞ ,kh0 k∞ ,β,R,C (M −1/2 ) . (3.7)
M
T
The proof is postponed to Appendix B.4.
Remark 3.5. If h(x) = h(1 − x) for x ∈ [0, 1] then the first order term of (3.4) is zero; in this
case we can remove the term j = k in (3.6) and define ω = [ω0 . . . ωk−1 ]0 ∈ Rk by (3.5)
with the second row and last column of A removed and e1 = [1 0 . . . 0]0 ∈ Rk .
It is straightforward to check that the optimal bandwidth for minimizing the order of the
right term of Equation (3.7) is M ∝ T 2β/(2β+1) , yielding the next result.
Corollary 3.6. Let β, R, f− > 0 and h : [0, 1] → R. Let k ∈ N and ω ∈ (0, 1] be uniquely
defined such that β = k + ω and set M := 2bT 2β/(2β+1) c in the following. Suppose that
Assumptions (M-1), (H) and (C) hold. Let e
θt,T (M) be obtained as in Theorem 3.4. Then,
for any q > 0 there exist a constant C0 only depending on h, q, d, f− , R, µ and β, and a
T 0 > 0 only depending on d, R, β, f− and C such that, for all T ≥ T 0 and all t ∈ Z,
q 1/q
E e
θt,T (M) − θt/T
≤ C0 T −β/(2β+1) .
(3.8)
It is interesting to note that in the decomposition (2.15), the bound of the error term
ke
θt,T (M) − θt/T k in (3.8) always has a slower decaying rate that that of the bound of the
∗
(deterministic) error term kθt/T − θt,T
k in Lemma 2.4.
4. Application to TVAR processes
TVAR processes (see Example 3) are a handful model to illustrate our results. Under
suitable assumptions, they have the specific property that, when d = p, the linear predictor
∗
coefficients in θt,T
∈ Rd as defined by Equation (2.10) coincide with the time-varying
autoregressive coefficients given by θ(t/T ) of the TVAR(p) equation (2.9) and also with
the local solution θt/T of the Yule-Walker equations defined by (2.12), see (4.4) below.
In the following, we introduce some smoothness assumptions on the time-varying parameters, similar to (and actually yielding) the one required on the local spectral density
in 2.3. Additional stability conditions are also required, based on the following definitions.
10
François Roueff and Andrés Sánchez-Pérez
h
For θ : R → R p , u 7→ θ1 (u)
polynomial by
...
i0
θ p (u) we define the time varying autoregressive
θ(z; u) := 1 −
p
X
θ j (u)z j .
j=1
Let us define, for any p ∈ N and δ > 0,
p
X
h
i0
p
j
−1
θ
.
.
.
θ
,
θ
=
∈
R
s.t.
1
−
θ
z
,
0,
∀|z|
<
δ
s(p) (δ) :=
1
p
j
j=1
n
o
s p (δ) := θ : R → s(p) (δ)
∗
(4.1)
(4.2)
= {θ : R → R s.t. θ(z; u) , 0, ∀|z| < δ , u ∈ R} .
p
−1
Define, for β > 0, R > 0, δ ∈ (0, 1), ρ ∈ [0, 1] and σ+ > 0, the class of parameters
n
o
C (β, R, δ, ρ, σ+ ) = (θ, σ) : R → R p × [ρσ+ , σ+ ] s.t. θ ∈ Λ p (β, R) ∩ s p (δ), σ ∈ Λ1 (β, R) .
The first result to provide sufficient conditions on the TVAR coefficients for the existence
of a stable solution of the TVAR equations goes back to Künsch (1995). Here we use
Giraud et al. (2015, Proposition 1), which guarantees the following: given a centered i.i.d.
sequence (ξt )t∈Z with unit variance and given the constants δ ∈ (0, 1), ρ ∈ [0, 1], σ+ > 0,
β > 0 and R > 0, there exists a large enough T 0 only depending on δ, β and R such that,
for all (θ, σ) ∈ C (β, R, δ, ρ, σ+ ), there exists a unique process (Xt,T )t∈Z,T ≥T0 satisfying (2.9)
for all t ∈ Z and T ≥ T 0 and such that, for all T ≥ T 0 , Xt,T0 is bounded in probability as
t → −∞. We use this result as our definition of the TVAR process with time varying AR
coefficients θ1 , . . . , θ p , time varying standard deviation σ, and innovations (ξt )t∈Z . For later
reference, we summarize this in the following assumption.
(M-2) Let (ξt )t∈Z be an i.id. sequence with zero mean and unit variance. Assume
that (θ, σ) ∈ C (β, R0 , δ, ρ, σ+ ) with δ ∈ (0, 1), β > 0, R0 > 0 and ρ ∈ [0, 1]. The
array (Xt,T )t∈Z,T ≥T 0 is a TVAR process as previously defined with time varying
AR coefficients θ1 , . . . , θ p , time varying standard deviation σ, and innovations
(ξt )t∈Z .
In this assumption the constant T 0 is set to have the existence and uniqueness of the stable
solution of the TVAR equation for all T ≥ T 0 . It may change hereafter from line to line
to guarantee additional properties of the solution but always in a way where it depends at
most on the constants β, R0 , δ, ρ and σ+ .
The following assumption can be used to control the moments of any order of the TVAR
process.
(I) For all q > 0 the innovations (ξt )t∈Z satisfy E |ξ0 |q < ∞.
Time varying autoregressive processes are well known to be locally stationary under certain conditions on their parameters and moments, see Dahlhaus (1996b, Theorem 2.3).
Adapting these results to our context, we have the following.
Theorem 4.1. Assumption (M-2) implies the two following assertions.
(i) There exist constants R and T 0 only depending on p, δ, σ+ , β and R0 such that
(Xt,T )t∈Z,T ≥T0 is (β, R)-weakly locally stationary in the sense of Definition 2.3 with
local spectral density defined by
σ2 (u) −iλ −2
f (u, λ) =
θ e ;u
.
(4.3)
2π
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
11
Moreover, we have, for all T ≥ T 0 and t ∈ Z,
∗
θ(t/T ) = θt,T
= θt/T ,
(4.4)
∗
θt,T
where
and θt/T are the optimal and local prediction coefficients respectively
defined by (2.10) and (2.12) in the case d = p.
(ii) If ρ ∈ (0, 1], then Assumption (M-1) holds with the same β and some constants R,
T 0 and f− > 0 only depending on p, δ, ρ, σ+ , β and R0 .
(iii) If P(ξ0 = x) = 0 for all x ∈ R, then P(Xt,T = 0) = 0 for all t ∈ Z and T ≥ T 0 .
(iv) If (I) holds, then (Xt,T )t∈Z,T ≥T0 satisfies Assumption (C) with C only depending on
R0 , β, δ, σ+ and with µ defined as the distribution of ξ0 .
The proof is postponed to Appendix C.2. Theorem 4.1 basically shows that the results
of Section 3 apply to TVAR processes, as defined by (M-2) provided that ρ > 0 and (I) is
assumed on the innovations. We specifically state the following result which provides a
useful complement to Moulines et al. (2005, Corollary 9) where the same minimax rate is
exhibited for a different estimator but only for smoothness index β ≤ 2.
Corollary 4.2. Let δ ∈ (0, 1), β > 0, R > 0 and ρ ∈ (0, 1]. Suppose that Assumptions (M-2)
and (I) hold. Let M = 2bT 2β/(2β+1) c and e
θt,T (M) be the estimator defined by (2.16) and (3.6)
with p, the order of the TVAR process equal to d, the order of the prediction vector. Then,
for any q ∈ N there exists a constant C only depending on q, h, p, δ, ρ, σ+ , β, R0 and the
moments of the distribution of ξ0 such that, for all T ≥ T 0 and t ∈ Z, we have
"
t q #!1/q
e
≤ C T −β/(2β+1) .
(4.5)
E θt,T (M) − θ
T
Proof. By Theorem 4.1, we can apply Corollary 3.6. Recall that in the case of the TVAR
∗
with order p equal to the prediction length, we have θ (t/T ) = θt/T = θt,T
.
The estimator e
θ proposed in Corollary 4.2 achieves the β-minimax-rate for TVAR processes according to the lower Moulines et al. (2005, Theorem 4). Hence, it is also βminimax-rate in the class of weakly locally stationary processes satisfying Assumption (M1). Giraud et al. (2015, Section A.1) explains how to construct minimax-rate predictors
from minimax-rate estimators of θ. Applying their approach, Corollary 4.2 also provides a
crucial ingredient in building β-minimax-rate predictors for any β > 0.
5. Numerical work
We test both methods on data simulated according to a TVAR process with p = 3. The
parameter function u 7→ θ(u) within s p (δ) for some δ ∈ (0, 1) is chosen randomly as follows. First we pick randomly some smoothly time varying partial autocorrelation functions
P
2
up to the order p that are bounded between −1 and 1, θ̌k,k (u) ∝ F−1
j=1 a j,k j cos ( ju), where
a j,k are random numbers in [−1, 1]. Here θ̌k,k (u) is defined up to a multiplicative constant;
dividing, for example, by F(F −1)(2F −1)/6 guarantees its values to remain within (−1, 1).
Then, for any required t, T , we use Algorithm 1 with input θ̌k,k (t/T ) and assign the output
to θ(t/T ). Based on the classical Levinson-Durbin recurrence (see for example Brockwell
and Davis (2006, Proposition 5.2.1)), the θ̌ in Algorithm 1 is in s(p) (1) as defined in (4.1),
and it follows that the output θ ∈ s(p) (δ). The randomly obtained three components of our
θ(t) are displayed in Figure 5.1, for t ∈ [0, 1].
For each T ∈ {22 j , j = 5, . . . , 15} we generate 100 realizations of a TVAR process from
innovation sequences (ξt )t∈Z of i.i.d. centred Gaussian random variables with unit variance
by sampling the previous θ at a rate T −1 , and taking σ ≡ 1.
12
François Roueff and Andrés Sánchez-Pérez
Then we compare b
θ and e
θ for estimating θ(1/2) using h ≡ 1 and different values of M.
∗
Recall that θ(1/2) = θT/2,T
. Figure 5.2 shows the boxplots corresponding to this evaluation
for two different T s. In Figure 5.2 we observe that for T = 220 the estimation error of b
θ is
minimized in M = 215 while that of e
θ is minimized in M = 217 . The estimator e
θ beats b
θ
for the two biggest values of M. In the case T = 230 , the error of b
θ reaches its minimum
in M = 223 and that of e
θ in M = 226 . The estimator e
θ beats b
θ for the four biggest values
of M. These experiences illustrate the theoretical result established in Theorem 3.2 and
Corollary 4.2 that after optimizing in M, e
θ outperforms b
θ for T large enough.
To corroborate these conclusions over a wider range of T ’s, we refer to Figure 5.3.
The plot on the left-hand side shows the oracle errors min M kb
θT/2,T (M) − θ(1/2)k and
min M ke
θT/2,T (M) − θ(1/2)k for all T ∈ {22 j , 5 ≤ j ≤ 15}. The slope corresponding to e
θ (in
blue) is steeper than the one corresponding to b
θ (in red), meaning that, in average, e
θ outperforms b
θ by an increasing order of magnitude as T increases. The boxplots on the right-hand
side of Figure 5.3 represent the ratios min M ke
θT/2,T (M) − θ(1/2)k/ min M kb
θT/2,T (M) − θ(1/2)k
computed for each T and realization of the TVAR process. Observe that for 214 ≤ T ≤ 218
the estimator e
θ beats b
θ in at least half of the cases. For T ≥ 220 , it happens in at least
75% of the cases. We conclude that the estimator with reduced bias is of interest when the
length of the data set becomes very large.
Algorithm 1: Adapted Levinson-Durbin algorithm.
parameters the stability parameter δ > 0 and the time varying partial autocorrelation
functions θ̌k,k , k = 1, . . . , p;
for k = 2 to p do
for j = 1 to k − 1 do
θ̌ j,k = θ̌ j,k−1 − θ̌k,k θ̌k− j,k−1 ;
for j = 1 to p do
θ j,p = δ j θ̌ j,p ;
return θ = [θ1,p . . . θ p,p ]0 .
Acknowledgments
We thank Tobias Kley for his thoughtful reading of the paper and appropriate remarks
(concerning specially Lemma 2.4) and also for bringing forward to our attention the related
work Kley et al. (2016).
Appendix A. Useful lemmas
We gather here some useful lemmas that are (simple extensions of) standard results
for Yule-Walker estimation of the linear prediction coefficients. Most of them are similar
to those used in Dahlhaus and Giraitis (1998). Short proofs are provided for the sake of
completeness. Different bounds can be found in Kley et al. (2016), in order to better control
the case d → ∞.
Lemma A.1. Let d be a positive integer. Consider the d × d matrices Γ and b
Γ and vectors
γ, b
γ, θ, b
θ ∈ Rd satisfying the relations
Γθ = γ ,
(A.1)
b
Γb
θ =b
γ.
(A.2)
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
0.10
0.05
0.00
0.05
0.10
0.0
0.15
0.10
0.05
0.00
0.05
0.10
0.15
0.20
0.0
0.15
0.10
0.05
0.00
0.05
0.10
0.15
0.0
13
0.2
0.4
0.6
0.8
1.0
0.2
0.4
0.6
0.8
1.0
0.2
0.4
0.6
0.8
1.0
Figure 5.1. Plots of θ1 (t) (top), θ2 (t) (middle) and θ3 (t) (bottom) on the
interval t ∈ [0, 1].
10-1
10-2
θb loss
e
θ loss
10-2
θb loss
e
θ loss
10-3
10-4
10-3
loss
loss
10-5
10-6
10-4
10-7
10-5
10-6
10-8
211
212
213
214
M
215
216
217
10-9
218
221
222
223
224
M
225
226
227
228
Figure 5.2. Box plots of the quadratic losses for estimating θ(1/2) using
b
θT/2,T (M) (red boxes) and e
θT/2,T (M) (blue boxes) for various bandwidths
M, when T = 220 (left) and T = 230 (right).
Then, for any k ∈ N we have, if Γ is non-singular,
k
k+1
X
`
`
X −1
−1
−1
b
b
θ − θ = Γ +
Γ Γ − Γ b
γ−γ +
Γ Γ −b
Γ θ
`=1
`=1
k+1
b
+ Γ−1 Γ − b
Γ
θ − θ . (A.3)
Proof. From Equations (A.1) and (A.2) we get
h
i
b
θ − θ = Γ−1 Γ − b
Γ b
θ +b
γ−γ .
The result follows by applying a recursion on k = 0, 1, . . . .
14
François Roueff and Andrés Sánchez-Pérez
10-1
102
θb loss
θb loss slope
e
θ loss
θe loss slope
10-2
10-3
oracle losses ratio
oracle loss
10-4
101
10-5
10-6
10-7
100
10-1
10-8
10-9
210
212
214
216
218
220
T
222
224
226
228
10-2
230
210
212
214
216
218
220
T
222
224
226
228
230
Figure 5.3. Oracle losses (using the best choice for the bandwidth M)
for estimating θ(1/2) using b
θT/2,T (M) (red points) and e
θT/2,T (M) (blue
points) for various values of T . The left-hand side plot displays the
losses over all the Monte Carlo simulations and the two resulting loglog regression lines. The right-hand side plot displays boxplots of the
corresponding losses ratio.
Lemma A.2. Let γ be the autocovariance function associated with a spectral density funcR 2π
tion f , γ (s) = 0 eisλ f (λ) dλ , for all s ∈ Z, and denote by Γd the corresponding covariance matrix of size d × d,
γ (0)
..
Γd =
.
γ (d − 1)
...
..
.
...
γ (d − 1)
..
.
.
γ (0)
(A.4)
Then the following assertions hold for any d ∈ N∗ .
R 2π
(i) If 0 f > 0 then Γd is positive definite.
(ii) If f is valued in [ f− , f+ ] with f− ≤ f+ , then all the eigenvalues of Γd belong to
[2π f− , 2π f+ ].
Proof. These well known facts (see e.g. Brockwell and Davis (2006, Proposition 4.5.3))
2
Z π X
d
0
i jλ
f (λ) dλ , for all a = [a1 . . . ad ]0 ∈
follow from the identity a Γd a =
aj e
−π j=1
Rd .
The next lemmas allow us to control the norms of b
θt,T and θt/T .
Lemma A.3. Let p be a positive integer and δ > 0. The set s(p) (δ) defined in (4.1) is a
closed subset of the ball {θ ∈ R p s.t. kθk ≤ (1 + δ) p − 1}.
Proof. Hurwitz’s theorem (see Conway (1973, Theorem 2.5) or Gamelin (2001, Section 3,
Chapter VIII)) implies that s(p) (δ) is a closed subset of R p . It is also bounded (see Moulines
et al. (2005, Lemma 1)). Hereafter we provide a slightly different bound using Euclidean
norm instead of the supnorm.
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
15
Take now θ ∈ s(p) (δ). Let z1 , . . . , z p denote the complex roots of the polynomial θ(z) :=
P
1 − pj=1 θ j z j . They satisfy |z j | ≥ δ−1 for any j. The following holds
1
1 + kθk2 =
2π
Z
π
1−
−π
p
X
2
θ je
j=1
i jλ
1
dλ =
2π
Z
π
2
θ eiλ dλ .
(A.5)
−π
Q
−1
≤ δ, we
On the other hand we have θ (z) = pj=1 1 − zz−1
j , so that for |z| = 1, since |z j |
p
get |θ (z) | ≤ (1 + δ) . Putting this into (A.5) the proof is completed.
The next lemma is similar in flavor to the statistical result of Whittle (1963, Section 3).
It is also a classical property of orthogonal polynomials (see Grenander and Szegő (1984,
Section 2.4)). We provide an elementary proof.
Lemma A.4. Let γ be an autocovariance function. Let d ≥ 1 such that the covariance
matrix Γd defined by Equation (A.4) is positive-definite. Let θ denote the solution of the
0
d-order Yule-Walker equation, θ = [θ1 . . . θd ]0 = Γ−1
d γd with γd = [γ(1) . . . γ(d)] . Then
d
we have θ ∈ s(d) (1) and kθk ≤ 2 − 1.
Proof. We only need to prove θ ∈ s(d) (1) since kθk ≤ 2d − 1 is then implied by Lemma A.3
with p = d and δ = 1.
For j = 1, . . . , d, let e j = [0 . . . 1 . . . 0]0 be the Rd - vector having a 1 in the j-th posih
i0
tion and zero everywhere else. Consider also the companion matrix A = θ e1 . . . ed−1 .
Since the roots of θ(z) are the inverses of the eigenvalues of A, or A0 , we only need to prove
that the eigenvalues of A0 are inside the closed unit disk. Observe that
0
θ0 Γd e1 . . . . . .
θ0 Γd ed−1
θ Γd θ
e01 Γd θ
e01 Γd e1 . . . . . .
e01 Γd ed−1
0
e02 Γd e1 . . . . . .
e02 Γd ed−1 .
Γd − AΓd A0 = Γd − e2 Γd θ
.
..
..
..
..
.
.
.
.
..
0
0
0
ed−1 Γd θ ed−1 Γd e1 . . . . . . ed−1 Γd ed−1
Because Γd is a Toeplitz matrix, its (i, j)-th entries, and those of AΓd A0 are equal for i, j ≥ 2.
The definition of θ implies also the equality of the (i, j)-th entries of both matrices when
i = 1, j ≥ 2 and i ≥ 2, j = 1. Hence Γd − AΓd A0 is a d × d symmetric matrix with zero
entries except at the top-left where it takes value γ(0) − θ0 γd . This value is non-negative
since it is the variance of the prediction error or order d. Hence we conclude that for
v ∈ Cd , v0 (Γd − AΓd A0 )v ≥ 0. Consider now λ, an eigenvalue of A0 and the corresponding
eigenvector v ∈ Cd \ {0}. We get
0 ≤ v̄0 (Γd − AΓd A0 )v = v̄0 (Γd − AΓd A0 )v = v̄0 Γd v(1 − |λ|2 ) .
We conclude that |λ| ≤ 1 since v̄0 Γd v > 0.
Lemma A.5. Let d ≥ 1, (Xt,T )t∈Z,T ≥T0 be an array process and h : [0, 1] → R. For any
M ∈ N∗ , define the local tapered empirical covariance function b
γT,M as in Definition 2.5
and let, for any t ∈ Z and T ≥ T 0 , b
Γt,T,M = (b
γT,M (t/T, i − j); i, j = 1, . . . , d) be the
corresponding d × d empirical covariance matrix. Then, either b
Γt,T,M is non-singular, or
b
γT,M (t/T, `) = 0 for all ` ∈ Z. Moreover, in the case where it is non-singular, the YuleWalker estimate b
θt,T (M) defined by (2.16) satisfies kb
θt,T (M)k ≤ 2d − 1.
16
François Roueff and Andrés Sánchez-Pérez
Proof. First note that for all u ∈ R, the sequence b
γT,M (u, `) `∈Z is the covariance function
associated with the spectral density
2
M
1 X t
−iλ t
b
h
XbuT c+t−M/2,T e
f M (u, λ) =
H M t=1
M
We conclude by applying Lemmas A.2 (i) and A.4.
Appendix B. Bounds of the estimation risk in the general setting
B.1. Proof of Lemma 2.4. Let us first bound the approximation error Γt/T − Γ∗t,T .
t
t − i
t
γ , i − j − γ∗ (t − i, T, j − i) ≤ γ
, j − i − γ ,i − j
T
T
T
t − i
+ γ
, j − i − γ∗ (t − i, T, j − i) . (B.1)
T
The second line term of Inequality (B.1) is upper bounded by RT − min(1,β) because of Inequality (2.3). Using the local covariance expression (2.2), Cauchy-Schwartz inequality
and f (, λ) ∈ Λ1 (β, R), the following holds for T ≥ d ≥ |i|,
Z π
t − i
t
t
t − i
γ
, j − i − γ ,i − j
=
,λ − f
, λ dλ
ei( j−i)λ f
T
T
T
T
−π
≤ 2πdR T − min(1,β) .
(B.2)
Inequality (2.3) implies that kγ∗t,T − γt/T k ≤ d1/2 R T − min(1,β) and inequalities (B.1), (B.2)
and again (2.3) imply that for T ≥ d,
kΓt/T − Γ∗t,T k ≤ d(2πdR + R)T − min(1,β) .
The smallest eigenvalue of the matrix Γt/T is greater or equal to 2π f− (see Lemma A.2 (ii)).
Observe that
n
o
inf inf a0 Γ∗t,T a = inf inf a0 Γ∗t,T − Γt/T a + a0 Γt/T a
t kak=1
t kak=1
≥ inf inf a0 Γ∗t,T − Γt/T a + inf inf a0 Γt/T a ≥ 2π f− − d3/2CT − min(1,β) .
t kak=1
t kak=1
Then, for T ≥ T 0 = (dR(2πd + 1)/(2π f− ))
, we have that Γ∗t,T is invertible and
∗ −1
−1
k(Γt,T ) k ≤ (π f− ) . Now, from equations (2.11) and (2.12) we obtain that
−1 h
i
∗
θt,T
− θt/T = Γ∗t,T
Γt/T − Γ∗t,T θt/T + γ∗t,T − γt/T .
1/ min(1,β)
Applying matrix inequalities (specifically with the spectral norm) we get
−1
∗
− θt/T k ≤ Γ∗t,T
Γt/T − Γ∗t,T θt/T + γ∗t,T − γt/T .
kθt,T
Lemma A.4 ensures that kθt/T k ≤ 2d and the result follows with C1 = (π f− )−1 (d(2πdR +
R)2d + R).
B.2. Bias Approximation. The following elementary lemma will be useful.
Lemma B.1. Let h : [0, 1] → R satisfying (H). Then, for all ` = 0, 1, 2, . . . and M ≥ j ≥ 0,
we have
Z 1
M
1 X s s − j s `
h
h
=
h2 (u)u` du + O j,`,khk∞ ,kh0 k∞ M −1 .
M s= j+1 M
M
M
0
In particular, in the case j = ` = 0, H M = M + Okhk∞ ,kh0 k∞ (1).
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
17
Proof. The proof is straightforward using, for M large enough, Riemann approximations
on the blocks defined by s/M ∈ (ui−1 + j/M, ui ], for i = 1, . . . , N, and neglecting the terms
from the indices s such that s/M ∈ (ui−1 , ui−1 + j/M], the number of which is bounded.
We can now derive the following approximation of the bias.
Theorem B.2. Suppose that Assumption (M-1) holds with some β > 0 and R > 0, and
let h : [0, 1] → R satisfying (H). Let k ∈ N and α ∈ (0, 1] be uniquely defined such that
β = k + α. Then, for all j ∈ Z and M ∈ 2N∗ , we have
!
k
M `
t X
t
1 M β
ch, f, j,`
,j =γ ,j +
+ O j,khk∞ ,kh0 k∞ ,β,R
+
,
E b
γT,M
T
T
T
M
T
`=1
where ch, f, j,` ∈ C only depends on h, the spectral density f , j and `. If h(x) = h(1 − x) for
all x ∈ [0, 1], then ch, f, j,1 = 0.
Proof. Without loss of generality we let moreover assume M ≥ j ≥ 0, in which case
t
E b
γT,M
,j
T
=
M
1 X s s − j ∗
M
h
h
γ t + s − , T, j .
H M s= j+1 M
M
2
Since |h| is also piecewise continuously differentiable, Lemma B.1 gives that
M
1 X s s − j
h
h
= O j,khk∞ ,kh0 k∞ (1) .
|H M | s= j+1
M
M
With Inequality (2.3), we obtain that
t
E b
γT,M
, j = γ M, j + O j,khk∞ ,kh0 k∞ ,R T − min(1,β) ,
T
!
M
1 X s s − j t + s − M/2
h
where γ M, j :=
h
γ
,j .
H M s= j+1 M
M
T
(B.3)
Since f (·, λ) ∈ Λ1 (β, R), a Taylor expansion yields
! X
!
!
k
∂`1 f (t/T, λ) −M/2 + s `
t − M/2 + s
t −M/2 + s
,λ =
+ fk
,
,λ ,
f
T
`!
T
T
T
`=0
with
| fk (t/T, (−M/2 + s)/T, λ)| = Oβ,R ((M/T )β ) .
sup
(B.4)
t∈Z,1≤s≤M
Then
γ M, j =
1
HM
Z
π
ei jλ
−π
k Z π
X
`=0
−π
M
X
s= j+1
h
s s − j t − M/2 + s !
h
f
, λ dλ =
M
M
T
M
s s − j −M/2 + s !`
(t/T, λ) i jλ 1 X
e
h
h
dλ
`!
H M s= j+1 M
M
T
!
Z π
M
1 X s s − j
t −M/2 + s
+
ei jλ
h
h
fk
,
, λ dλ . (B.5)
H M s= j+1 M
M
T
T
−π
∂`1 f
18
François Roueff and Andrés Sánchez-Pérez
Lemma B.1 yields that for all ` = 1, . . . , k,
!`
M
1 X s s − j −M/2 + s
h
h
=
H M s= j+1 M
M
T
!`
!
M
M ` M 1 X
s s − j 1
M `
s
1 M `
h
h
= ch,`
+ O j,khk∞ ,kh0 k∞ ,`
,
− +
T H M M s= j+1 M
M
2 M
T
M T
!`
Z 1
1
2
with ch,` =
h (u) u −
du. Observe that ch,0 = 1 by assumption in (H), and, if
2
0
h(x) = h(1 − x) for all x ∈ [0, 1], then moreover ch,1 = 0. From this it follows that
!!
Z π `
M
M `
s s − j −M/2 + s !`
∂1 f (t/T, λ) ei jλ X
1
h
dλ =
ch, f, j,` + O j,khk∞ ,kh0 k∞ ,β,R
,
h
`!
H M s= j+1 M
M
T
T
M
−π
where, in particular, ch, f, j,0 = γ(t/T, j) and ch, f, j,1 = 0 if h(x) = h(1 − x). Finally, by (B.4),
and since |h| is also piecewise continuously differentiable, the reminder term in (B.5) satisfies
!
Z π
M
s s − j
M β !
X
t − M/2 + s
i jλ 1
e
h
h
fk
, λ dλ = O j,khk∞ ,kh0 k∞ ,β,R
.
H M s= j+1 M
M
T
T
−π
We thus finally obtain that
!
k
M `
t X
1 M β
ch, f, j,`
+ Od,khk∞ ,kh0 k∞ ,β,R
+
γ M, j = γ , j +
.
T
T
M
T
`=1
This approximation and the bound (B.3) allow us to conclude the proof.
B.3. Proof of Theorem 3.2. Gathering together Assumption (C) and Theorem B.2 yields,
for any j = −d, . . . , d,
!
k
t
t X
M `
1 M β
b
γT,M
,j =γ ,j +
ch, f, j,`
+ Od,khk∞ ,kh0 k∞ ,β,R
+
T
T
T
M
T
`=1
t
+ uM
, j , (B.6)
T
where u M (t/T, j) = OL• (µ),d,khk∞ ,kh0 k∞ ,C (M −1/2 ) and ch, f, j,1 = 0 if h(x) = h(1 − x) for all
x ∈ [0, 1].
For the sake of simplicity, we drop t, T in the notation and set γ = γt/T , b
γM = b
γt,T,M ,
Γ := Γt/T and b
Γ M := b
Γt,T,M . Using the expression (B.6), we obtain
!
k
M `
X
1 M β
Γ −b
ΓM =
Ch, f,`
+ Od,khk∞ ,kh0 k∞ ,β,R
+
+ UM ,
(B.7)
T
M
T
`=1
!
k
M `
X
1 M β
b
γM − γ =
ch, f,`
+ Od,khk∞ ,kh0 k∞ ,β,R
+
+ uM ,
(B.8)
T
M
T
`=1
where the matrices Ch, f,` ∈ Rd×d and the vectors ch, f,` ∈ Rd only depend on d, h, f and `.
Furthermore U M and u M both are OL• (µ),d,khk∞ ,kh0 k∞ ,C (M −1/2 ). Again Ch, f,1 = 0 and ch, f,1 = 0
if h(x) = h(1 − x) for all x ∈ [0, 1].
Note now that Lemma A.2 with the assumption f− > 0 further says that Γt/T is also
non-singular with Γ−1
t/T = O f− (1) and, with (2.17), we can thus apply Lemma A.1, showing
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
19
that (A.3) holds with θ = θt/T , b
θ =b
θt,T (M), γ = γt/T , b
γ =b
γM = b
γt,T,M , Γ := Γt/T and
b
Γ =b
Γ M := b
Γt,T,M . Next the bounds provided by Lemmas A.4 and (2.17) further imply kθk
and kb
θk to be less that 2d . This, with the approximations (B.7) and (B.8), yields (3.4).
B.4. Proof of Theorem 3.4. For each j = 0, . . . , k replace M by M/2 j in (3.4), multiply
the resulting expression by ω j and sum over j. Matrix A (definition below Equation (3.5))
is a non singular Vandermonde matrix and ω is well defined.
Appendix C. Postponed proofs for TVCBS and TVAR processes
C.1. Proof of Theorem 3.1. Let us denote, for any L p random variable Z, kZk p = E |Z| p 1/p
its L p -norm. The proof relies on the Burkhölder inequality for non-stationary dependent
sequences. Namely, an immediate consequence of Dedecker et al. (2007, Proposition 5.4)
and the Hölder and Minkowskii inequalities is that if p ≥ 2 and (Zt )t∈Z is a L p process
adapted to the filtration (Ft )t∈Z , then, for all s ∈ Z and n ≥ 1
s+n
X
2
(Zt − E[Zt ])
t=s+1
!
∞
X
≤ 2pn sup kZt − E[Zt ]k p sup
kE [Zt+k |Ft ] − E[Zt+k ]k p .
t∈Z
p
t∈Z k=0
Applying this inequality with s = |`| + 1, n = M − |`| to
t t − |`| !
Zt = h
h
XbuT c+t−M/2,T XbuT c+t−|`|−M/2,T ,
M
M
ξ
Ft = FbuT
,
c+t−M/2
where Ftξ = σ(ξ s , s ≤ t) denotes the natural filtration of (ξt )t∈Z , we obtain that, for any
M > |`|,
b
γT,M (u, `) − E b
γT,M (u, `)
p
≤
2(p(M − |`|))1/2
khk2∞ sup kX s,T k2p
HM
s∈Z
∞
X
i
h
× sup
E Xt+k,T Xt−|`|+k,T | Ftξ − E[Xt+k,T Xt−|`|+k,T ]
t∈Z k=0
p
1/2
.
Under Assumption (H), H M /M ∈ (1/2, 3/2) for M large enough (see Lemma B.1) and it is
thus now sufficient to show that, for any p ≥ 2,
sup kX s,T k2p = OK,r,(ψk ) (1)
(C.1)
s∈Z
sup
∞
X
t∈Z k=0
h
i
E Xt+k,T Xt−|`|+k,T | Ftξ − E[Xt+k,T Xt−|`|+k,T ]
p
= OK,r,k0 ,`,(ψk ),(ζk ) (1)
(C.2)
The bound (C.1) is a direct consequence of (2.6) and (2.7) and the assumption on (ξt )t∈Z .
To prove (C.2), let us define, for all t ∈ Z, T ≥ T 0 and x ∈ RN ,
Φ0t,T (x) = ϕ0t,T (x) ϕ0t−|`|,T ((x j+|`| ) j≥0 ) .
It then follows that Xt,T Xt−|`|,T = Φ0t,T ((ξt− j ) j≥0 ) and is then straightforward to show that
the assumptions on ϕ0 yields that there exists some constant K 0 only depending on K and
(ψ j ) j∈N and (ζ j ) j∈N such that, for all t ∈ Z, T ≥ T 0 and all x, x0 ∈ RN satisfying x j = x0j for
20
François Roueff and Andrés Sánchez-Pérez
1 ≤ j ≤ k0 + |`|,
Φ0t,T
2r
∞
X
(x) ≤ K 1 +
ψ̃ j x j ,
0
j=0
Φ0t,T (x)
−
Φ0t,T (x0 )
X
0
≤ K ζ̃ j |xk0 + j − xk0 + j |
0
j≥0
2r−1
X
0
1 +
,
ψ̃ j (|x j | + |x j |)
j≥0
where ψ̃ j = ψ j and ζ̃ j = ζ j for 0 ≤ j < |`|, and ψ̃ j = max(ψ j−|`| , ψ j ) and ζ̃ j = ζ j−|`| + ζ j for
j ≥ |`|. By Jensen’s inequality, we have that, for all k ≥ 0,
h
i
0
E Xt+k,T Xt−|`|+k,T | Ftξ − E[Xt+k,T Xt−|`|+k,T ] ≤ Φ0t+k,T ((ξt+k− j ) j≥0 ) − Φ0t+k,T ((ξt+k−
j ) j≥0 ) p ,
p
0
where
is i.i.d. with same distribution as (ξ s ) s∈Z and such that ξt+k− j = ξt+k−
j for all
0
)
and
(ξ
)
are
independent.
With
the
above
bounds
on
j = 0, 1, . . . , k − 1 and (ξt−
t− j j≥0
j j≥0
0
Φ and using the Minkowskii and Hölder inequalities, we thus get that, for all t ∈ Z and
k ≥ 0,
h
i
E Xt+k,T Xt−|`|+k,T | Ftξ − E[Xt+k,T Xt−|`|+k,T ] ≤ 2K 0 (1 + kψ̃k1 kξ0 k2rp )2r ,
(ξ0s ) s∈Z
p
and, if k > k0 + |`|, the same quantity is bounded from above by
∞
2r−1 X
0
2K kξ0 k2p 1 + 2kψ̃k1 kξ0 k2p(2r−1)
ζ̃ j .
j=k−k0
Summing these bounds over all k ≥ 0, we obtain
∞
X
k=0
h
i
E Xt+k,T Xt−|`|+k,T | Ftξ − E[Xt+k,T Xt−|`|+k,T ]
p
∞
X X
2r
0
≤ 2K 1 + 2kψ̃k1 kξ0 k4rp k0 + |`| +
ζ̃ j
k>k0 +|`| j=k−k0
X
2r
= 2K 0 1 + 2kψ̃k1 kξ0 k4rp k0 + |`| +
( j − |`|)ζ̃ j .
j>|`|
By (3.3), the term between parentheses is a finite constant only depending on k0 , |`| and ζ.
We thus have shown (C.2), which concludes the proof.
C.2. Proof of Theorem 4.1. We first need to recall some basic facts on the representation
of a TVAR process as a TVCBS. Let us set e1 = [1 0 . . . 0]0 ∈ R p and introduce the
companion p × p matrices defined for all u ∈ R by
θ1 (u) θ2 (u) . . . . . . θ p (u)
1
0
... ...
0
1
0 ...
0 .
A(u) = 0
.
..
..
..
..
.
.
.
.
..
0
...
0
1
0
By Giraud et al. (2015, Proposition 1 and its proof), the TVAR process (Xt,T )t∈Z,T ≥T0 is a
special case of TVCBS introduced in Example 2. Namely, it satisfies a representation (2.6)
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
with a linear form
∞
X
ϕ0t,T (x) =
at,T ( j)σ((t − j)/T ) x j ,
with at,T ( j) =
e01
21
j
Y t − i
e1 ,
A
T
i=1
j=0
and moreover, there exist some constants K̄ > 0 and δ1 ∈ (0, 1) only depending on p, δ, β
and R0 such that, for all T ≥ T 0 , t ∈ Z and j ∈ N∗ ,
j
t − i
Y
A
≤ K̄ δ1j .
(C.3)
T
i=1
Hence in particular
|at,T ( j)| ≤ K̄δ1j ,
implying that ϕ0 above satisfies (2.7) and (3.2) with r = 1, K = 1 and ψ j = ζ j = K̄σ+ δ1j .
We can now proceed with the proof of Theorem 4.1, starting with (i). To show that it
is weakly locally stationary with the local spectral density of the AR(p) with local standard deviation σ(u) and autoregresive coefficients θ1 (u), . . . , θ p (u), it only remains to show
that (2.8) holds with r = 1 and
∞
X
e01 A j (u) e1 x j .
ϕ(u, x) = σ(u)
j=0
This is a simple consequence of (C.3) and the fact that σ ∈ Λ1 (min(1, β), R0 ) and θ ∈
Λ p (min(1, β), R0 ) by assumption. Then (4.4) follows from the representation (2.6) : using
∗
that it is causal and (2.9), we get in the case d = p that θ(t/T ) = θt,T
. On the other hand,
the mere definition of the local spectral density in (4.3) (being that of an AR(p) process at
fixed u), the case d = p yields that θu = θ(u). Finally, an additional consequence of the
definition of f is that the spectral density f (·, λ) belongs to Λ1 (β, R) for any
λ ∈ R with
R only depending on p, δ, β, σ+ and R0 , provided that we can show that θ e−iλ ; u can be
bounded from below by a positive constant only depending on δ and p. By Lemma A.3
P
and continuity of (λ, θ) 7→ 1 − pj=1 θ j e−i jλ , and since δ−1 > 1 we have that
0<
inf
θ∈s(p) (δ),λ∈R
1−
p
X
j=1
θ j e−i jλ ≤
sup
θ∈s(p) (δ),λ∈R
1−
p
X
θ j e−i jλ < ∞ .
(C.4)
j=1
Of course these
two constants only depend on δ and p and the inf one can serve as a lower
bound of θ e−iλ ; u , concluding the proof of (i).
Next, we prove (ii) and (iii), which respectively require the two add-on properties
(a) f (u, λ) ≥ f− for all u, λ ∈ R,
(b) if ξ0 has a diffuse distribution, then P(Xt,T = 0) = 0 for all t ∈ Z and T ≥ T 0 .
The fact (a) follows from (4.3), σ(u) ≥ ρσ+ and the upper bounds in (C.4), which shows
that we can find such an f− > 0 only depending on p, δ, ρ and σ+ . Fact (b) is a consequence
of the TVCBS representation above and the assumptions on (ξt ) which implies that for all
t, T and j ∈ N, we can write Xt,T = at,T ( j)ξt− j + Zt,T ( j) with Zt,T ( j) independent of ξt− j .
Hence, if ξ0 has a diffuse distribution, it only remains to prove that for all T ≥ T 0 and
all t ∈ Z, there exists j ∈ N such that at,T ( j) , 0. Using again the TVCBS representation
above, this is equivalent to show that for all T ≥ T 0 and all t ∈ Z, γ∗ (t, T, 0) = var(Xt,T ) > 0.
Now observe that by (a) and since (Xt,T )t∈Z,T ≥T0 is (β, R)-weakly locally stationary, we have
γ∗ (t, T, 0) ≥ γ(t/T, 0) − R T − min(1,β) ≥ 2π f− − R T − min(1,β) > 0 ,
22
François Roueff and Andrés Sánchez-Pérez
where the last inequality holds by taking T 0 large enough (only depending on f− and R and
thus on p, δ, ρ, σ+ and R0 .).
We conclude with the proof of (iv). We use the above TVCBS representation which
were mentioned to satisfy (3.2) with ζ j = K̄σ+ δ1j for some δ1 < 1. Hence (3.3) holds as
well and we can apply Theorem 3.1, which shows that the TVAR process satisfies (C).
References
Jacques Baranger and Claude Brezinski. Analyse numérique, volume 38. Hermann, Paris
(1991). Collection Méthodes.
Andrew R. Barron. Are bayes rules consistent in information?
In Thomas M.
Cover and B. Gopinath, editors, Open Problems in Communication and Computation, pages 85–91. Springer New York (1987).
ISBN 978-1-4612-9162URL http://dx.doi.org/10.1007/
6.
doi:10.1007/978-1-4612-4808-8 22.
978-1-4612-4808-8_22.
Peter J. Brockwell and Richard A. Davis. Introduction to time series and forecasting.
Springer Texts in Statistics. Springer-Verlag, New York, second edition (2002). ISBN
0-387-95351-5. doi:10.1007/b97391. URL http://dx.doi.org/10.1007/b97391.
With 1 CD-ROM (Windows).
Peter J. Brockwell and Richard A. Davis. Time series: theory and methods. Springer Series
in Statistics. Springer, New York (2006). ISBN 978-1-4419-0319-8; 1-4419-0319-8.
Reprint of the second (1991) edition.
Nicolò Cesa-Bianchi and Gábor Lugosi. Prediction, learning, and games. Cambridge University Press, Cambridge (2006).
ISBN 978-0-521-84108-5; 0-52184108-9. doi:10.1017/CBO9780511546921. URL http://dx.doi.org/10.1017/
CBO9780511546921.
John B. Conway. Functions of one complex variable. Springer-Verlag, New YorkHeidelberg (1973). Graduate Texts in Mathematics, 11.
R. Dahlhaus. Maximum likelihood estimation and model selection for locally stationary processes. J. Nonparametr. Statist. 6 (2-3), 171–191 (1996a). ISSN 10485252. doi:10.1080/10485259608832670. URL http://dx.doi.org/10.1080/
10485259608832670.
R. Dahlhaus. On the Kullback-Leibler information divergence of locally stationary processes. Stochastic Process. Appl. 62 (1), 139–168 (1996b). ISSN 0304-4149. doi:10.
1016/0304-4149(95)00090-9. URL http://dx.doi.org/10.1016/0304-4149(95)
00090-9.
Rainer Dahlhaus. Locally stationary processes. In T.S. Rao, S.S. Rao and C.R. Rao,
editors, Time Series Analysis: Methods and Applications, volume 30 of Handbook
of Statistics, pages 351–413. North Holland (2012). ISBN 9780444538581. doi:
10.1016/B978-0-444-53858-1.00013-2. URL https://books.google.fr/books?
id=9wXOytMWHQoC.
Rainer Dahlhaus and Liudas Giraitis. On the optimal segment length for parameter estimates for locally stationary time series. J. Time Ser. Anal. 19 (6), 629–655 (1998).
ISSN 0143-9782. doi:10.1111/1467-9892.00114. URL http://dx.doi.org/10.
1111/1467-9892.00114.
Jérôme Dedecker, Paul Doukhan, Gabriel Lang, José Rafael León R., Sana Louhichi and
Clémentine Prieur. Weak dependence: with examples and applications, volume 190 of
Lecture Notes in Statistics. Springer, New York (2007). ISBN 978-0-387-69951-6.
PREDICTION OF WEAKLY LOCALLY STATIONARY PROCESSES
23
Marie Duflo. Random iterative models, volume 34 of Applications of Mathematics
(New York). Springer-Verlag, Berlin (1997). ISBN 3-540-57100-0. doi:10.1007/
978-3-662-12880-0. URL http://dx.doi.org/10.1007/978-3-662-12880-0.
Translated from the 1990 French original by Stephen S. Wilson and revised by the author.
Theodore W. Gamelin. Complex analysis. Undergraduate Texts in Mathematics. SpringerVerlag, New York (2001). ISBN 0-387-95093-1; 0-387-95069-9. doi:10.1007/
978-0-387-21607-2. URL http://dx.doi.org/10.1007/978-0-387-21607-2.
Christophe Giraud, François Roueff and Andres Sanchez-Perez. Aggregation of predictors
for nonstationary sub-linear processes and online adaptive forecasting of time varying
autoregressive processes. Ann. Statist. 43 (6), 2412–2450 (2015). ISSN 0090-5364.
doi:10.1214/15-AOS1345. URL http://dx.doi.org/10.1214/15-AOS1345.
Ulf Grenander and Gábor Szegő. Toeplitz forms and their applications. Chelsea Publishing
Co., New York, second edition (1984). ISBN 0-8284-0321-X.
Tobias Kley, Philip Preuß and Piotr Fryzlewicz. Predictive, finite-sample model choice for
time series under stationarity and non-stationarity. Technical report, ArXiv (2016). URL
https://arxiv.org/abs/1611.04460.
Hans Rudolf Künsch. A note on causal solutions for locally stationary ar-processes (1995).
Preprint ETH Zürich.
Eric Moulines, Pierre Priouret and François Roueff. On recursive estimation for time varying autoregressive processes. Ann. Statist. 33 (6), 2610–2654 (2005). ISSN 00905364. doi:10.1214/009053605000000624. URL http://dx.doi.org/10.1214/
009053605000000624.
M. B. Priestley. Evolutionary spectra and non-stationary processes.(With discussion). J.
Roy. Statist. Soc. Ser. B 27, 204–237 (1965). ISSN 0035-9246.
P. Whittle. On the fitting of multivariate autoregressions, and the approximate canonical
factorization of a spectral density matrix. Biometrika 50, 129–134 (1963). ISSN 00063444.
LTCI, Télécom ParisTech, Université Paris-Saclay
46, rue Barrault,
75634 Paris Cedex 13, France.
E-mail address: [email protected]
URL: https://perso.telecom-paristech.fr/˜roueff/
| 10math.ST
|
arXiv:1601.05622v1 [math.AC] 21 Jan 2016
POSTULATION AND REDUCTION VECTORS OF MULTIGRADED
FILTRATIONS OF IDEALS
PARANGAMA SARKAR AND J. K. VERMA
Abstract. We study relationship between postulation and reduction vectors of admissible multigraded
filtrations F = {F(n)}n∈Zs of ideals in Cohen-Macaulay local rings of dimension at most two. This is
enabled by a suitable generalisation of the Kirby-Mehran Complex. An analysis of its homology leads
to an analogue of Huneke’s Fundamental Lemma which plays a crucial role in our investigations. We
also clarify the relationship between the Cohen-Macaulay property of the multigraded Rees algebra of
F and reduction vectors with respect to complete reductions of F.
1. Introduction
The objective of this paper is to study properties of Hilbert functions and Hilbert polynomials of
multigraded filtrations of ideals under certain cohomological conditions. Among the themes presented
are: (1) an analogue of Huneke’s Fundamental Lemma in terms of the homology of generalised KirbyMehran complex for multigraded filtrations of ideals using complete reductions (2) relationship between
postulation vectors and reduction vectors for multigraded filtrations of ideals in Cohen-Macaulay local
rings of dimension at most two (3) providing necessary and sufficient conditions for the equality of
multigraded Hilbert functions and polynomials in terms of reduction numbers with respect to complete reductions and finally (4) relationship between the Cohen-Macaulay property of the Rees algebra
of multigraded filtrations of ideals and reduction numbers in two-dimensional Cohen-Macaulay local
rings.
Hilbert functions of mutigraded filtrations of ideals have been found useful in works of B. Teissier
[17] who used them in his investigations of Milnor numbers of singularities of complex analytic hypersurfaces. To wit, let (R, m) be a d-dimensional local ring of an isolated singularity of a complex
analytic hypersurface and let f be the defining polynomial. Then the Jacobian ideal J := J(f ) is an
m-primary ideal and the function B(r, s) = λ(R/mr J s ) is a polynomial P (r, s) of degree d in r and s.
Here λ denotes length. Teissier proved that the normalised coefficients of monomials of degree d in
P (r, s) are the Milnor numbers of linear sections of the isolated singularity. Joint reductions and the
2000 AMS Subject Classification: Primary 13H15
Key words and phrases: Hilbert function, joint reductions, complete reductions, Rees algebra, Kirby-Mehran complex,
postulation and reduction vectors.
The first author is supported by CSIR Fellowship of Government of India.
2
Bhattacharya function B(r, s) = λ(R/I r J s ) for m-primary ideals I and J have been used by D. Rees in
several contexts. For example, he characterised pseudo-rational local rings of dimension two in terms
of the constant term of the normal Hilbert polynomial for the normal Hilbert function λ(R/I r J s ).
We now describe the contents of the paper. We recall a few definitions and set up notation to explain
the results of this paper.
Throughout this paper let (R, m) be a Noetherian local ring of dimension d and I1 , . . . , Is be m-primary
ideals of R. For s ≥ 1, we put e = (1, . . . , 1), 0 = (0, . . . , 0) ∈ Zs and ei = (0, . . . , 1, . . . , 0) ∈ Zs where 1
+
occurs at ith position. Let n = (n1 , . . . , ns ) ∈ Zs , then we write I n = I1n1 · · · Isns and n+ = (n+
1 , . . . , ns )
where
(
n+
i =
ni
if ni > 0
0
if ni ≤ 0.
For α = (α1 , . . . , αs ) ∈ Ns , we put |α| = α1 + · · · + αs . Define m = (m1 , . . . , ms ) ≥ n = (n1 , . . . , ns )
if mi ≥ ni for all i = 1, . . . , s. By the phrase “for all large n” we mean n ∈ Ns and ni ≫ 0 for all
i = 1, . . . , s.
Definition 1.1. A set of ideals F = {F(n)}n∈Zs is called a Zs -graded I = (I1 , . . . , Is )-filtration if
for all m, n ∈ Zs , (i) I n ⊆ F(n), (ii) F(n)F(m) ⊆ F(n + m) and (iii) if m ≥ n, F(m) ⊆ F(n).
Let t1 , . . . , ts be indeterminates. For n ∈ Zs , put tn = tn1 1 · · · tns s and denote the Ns -graded Rees
L
ring of F by R(F) =
F(n)tn and the Zs -graded extended Rees ring of F by R′ (F) =
s
n∈N
L
L
L
F(n)tn . For an Ns -graded ring S =
Sn , we denote the ideal
Sn by S++ . For F = {I n }n∈Zs ,
n∈Zs
n≥0
n≥e
we set R(F) = R(I), R′ (F) = R′ (I) and R(I)++ = R++ .
Definition 1.2. A Zs -graded I = (I1 , . . . , Is )-filtration F = {F(n)}n∈Zs of ideals in R is called an
I = (I1 , . . . , Is )-admissible filtration if F(n) = F(n+ ) for all n ∈ Zs and R′ (F) is a finite R′ (I)-
module.
For an m-primary ideal I, the Hilbert function HI (n) is defined as HI (n) = λ (R/I n ) for all
n ∈ Z. Here we adopt the convention that I n = R if n ≤ 0. P. Samuel [16] showed that for sufficiently
large n, HI (n) coincides with a polynomial PI (n) of degree d, called the Hilbert polynomial of I.
For all n ∈ Z, PI (n) is often written in the form
PI (n) =
d
X
i=0
n+d−1−i
(−1) ei (I)
.
d−i
i
The coefficients ei (I) are integers for all i = 0, 1, . . . , d, called the Hilbert coefficients of I. The
leading coefficient e0 (I) is sometimes denoted by e(I) and called the multiplicity of I.
Let s ≥ 2 and F = {F(n)}n∈Zs be an I-admissible filtration of ideals in a Noetherian local ring (R, m)
3
of dimension d. For the Hilbert function HF (n) = λ (R/F(n)) of F, Rees [15] proved that there
exists a polynomial of degree d, called the Hilbert polynomial of F,
X
n1 + α1 − 1
ns + αs − 1
d−|α|
PF (n) =
(−1)
eα (F)
···
α1
αs
s
α=(α1 ,...,αs )∈N
|α|≤d
such that PF (n) = HF (n) for all large n. Here eα (F) are integers called the Hilbert coefficients of
F. This was proved by P. B. Bhattacharya for the filtration F = {I r J s }r,s∈Z in [1] where I and J are
m-primary ideals. Teissier [17] showed the existence of PF (n) for the filtration F = {I n }n∈Zs .
Let I be an m-primary ideal in a local ring (R, m) of dimension d ≥ 1. An integer n(I) is called the
postulation number of I if PI (n) = HI (n) for all n > n(I) and PI (n(I)) 6= HI (n(I)). An ideal
J ⊆ I is called a reduction of I if JI n = I n+1 for some n. We say J is a minimal reduction of I
if whenever K ⊆ J and K is a reduction of I, then K = J. Let
rJ (I) = min{m : JI n = I n+1 for n ≥ m} and r(I) = min{rJ (I) : J is a minimal reduction of I}.
The analogues of reduction and postulation number for I-admissible multigraded filtration F =
{F(n)}n∈Zs are described below.
Definition 1.3. A vector n ∈ Zs is called a postulation vector of F if HF (m) = PF (m) for all
m ≥ n.
Rees [15] introduced the concept of complete reduction for {I n }n∈Zs . In a similar way we define
complete reduction of an I-admissible filtration F = {F(n)}n∈Zs .
Definition 1.4. A set of elements A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s} is called a complete
reduction of F if J = (y1 , . . . , yd ) satisfies JF(n) = F(n+e) for all large n ∈ Ns where yj = x1j · · · xsj
for all j = 1, . . . , d.
Definition 1.5. A complete reduction A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s} of F is called a good
complete reduction if for all large m ∈ Ns , F(m) ∩ (y1 ) = y1 F(m − e) where y1 = x11 · · · xs1 .
Let A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s} be a complete reduction F, yj = x1j · · · xsj for all
j = 1, . . . , d and J = (y1 , . . . , yd ).
Definition 1.6. A vector r ∈ Ns is called a reduction vector of F with respect to A if for all
n ≥ r, JF(n) = F(n + e).
Definition 1.7. An integer k ∈ N is called the complete reduction number of F with respect
to A if JF(n) = F(n + e) for all n ≥ ke and whenever k 6= 0, there does not exist any 0 ≤ t < k such
that JF(n) = F(n + e) for all n ≥ te.
4
We use the following notation
(1) P(F) = {n ∈ Zs | n is a postulation vector of F}.
(2) RA (F) = {n ∈ Zs | n is a reduction vector of F with respect to A}.
(3) rA (F) = The complete reduction number of F with respect to A.
We now describe the main results proved in this paper. In section two, we prove some preliminary
results about the coefficients of Hilbert polynomial of a multigraded filtration of ideals which we use to
prove our main results. Let f (n) : Zs → Z be an integer valued function. Define the first difference
function of f (n) by ∆1 (f (n)) = f (n+e)−f (n). For all k ≥ 2, we define ∆k (f (n)) = ∆k−1 (∆1 (f (n))).
In [7], C. Huneke proved the following fundamental lemma:
Lemma 1.8. [7, Lemma 2.4] Let (R, m) be a 2-dimensional local Cohen-Macaulay ring and let x, y ∈ m
be any system of parameters of R. Let I be any ideal integral over (x, y). Then for all n ≥ 1,
n+1
n
I
(I : (x, y))
λ
−λ
= ∆2 (PI (n − 1) − HI (n − 1)) .
(x, y)I n
I n−1
S. Huckaba [5] extended this result for dimension d ≥ 1. In section three, for l ≥ 1, 1 ≤ k ≤ d
and y [l] = (y1 l , . . . , yk l ), we introduce an analogue of the Kirby-Mehran complex [9] for multigraded
filtrations: C.(y [l] , F(n)) :
R
dk
0 −→
−→
F(n)
(k)
( k )
1 d
k−1
R
R
R
d2
d1
d0
k−1
−→ · · · −→
−→
−→
0
F(n + le)
F(n + (k − 1)le)
F(n + kle)
and prove an analogue of Huneke’s Fundamental Lemma for multigraded filtration of ideals.
Theorem 1.9. Let (R, m) be a Cohen-Macaulay local ring of dimension d ≥ 1 with infinite residue
field, I1 , . . . , Is be m-primary ideals of R and F = {F(n)}n∈Zs be an I-admissible filtration of ideals
in R. Let A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s} be any complete reduction of F, yj = x1j · · · xsj for
all j = 1, . . . , d. Let y = y1 , . . . , yd and J = (y). Then for all n ∈ Zs ,
∆d (PF (n) − HF (n)) = λ
F(n + de)
JF(n + (d − 1)e)
−
d
X
i=2
(−1)i λ(Hi (C.(y, F(n)))).
1
In section four, for d ≥ 2, we compute n degree component of local cohomolgy module HR
(R(F))n
++
1
for all n ∈ Ns and give an equivalent criterion for vanishing of HR
(R(F))n for all n ∈ Ns . We discuss
++
vanishing of Hilbert coefficients and generalize some results due to T. Marley [11] in Cohen-Macaulay
local ring of dimension 1 ≤ d ≤ 2.
Theorem 1.10. Let (R, m) be a Cohen-Macaulay local ring of dimension 1 ≤ d ≤ 2 with infinite
residue field, I1 , . . . , Is be m-primary ideals of R and F = {F(n)}n∈Zs be an I-admissible filtration of
ideals in R. Let e(d−1)ei (F) = 0 for i = 1, . . . , s. Then
5
(1) For d = 1, PF (n) = HF (n) for all n ∈ Ns .
1
(2) For d = 2, if HR
(R(F))n = 0 for all n ∈ Ns then PF (n) = HF (n) for all n ∈ Ns and
++
e0 (F) = 0.
Theorem 1.11. Let (R, m) be a Cohen-Macaulay local ring of dimension 1 ≤ d ≤ 2 with infinite
residue field and I, J be m-primary ideals of R. Let F = {F(r, s)}r,s∈Z be a Z2 -graded (I, J)-admissible
filtration of ideals in R. Then the following statements are equivalent.
(1) e(d−1)ei (F) = 0 for i = 1, 2.
(2) I and J are generated by system of parameters, PF (r, s) = HF (r, s) for all r, s ∈ N and
F(r, s) = I r J s for all r, s ∈ Z.
(3) eα (F) = 0 for |α| ≤ d − 1.
In [10], Marley proved that for Cohen-Macaulay local ring of dimension d ≥ 1, r(I) = n(I) + d
under some depth condition of the associated graded ring of I. In his thesis [11], Marley extended
this result for Z-graded I-admissible filtrations. We generalize this result for multigraded filtration of
ideals when d = 1, 2. In section five, we prove the following theorem.
Theorem 1.12. Let (R, m) be a Cohen-Macaulay local ring of dimension one with infinite residue
field and I1 , . . . , Is be m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of
ideals in R and A = {ai ∈ Ii : i = 1, . . . , s} be a complete reduction of F. Then
P(F) ⊆ Ns and P(F) = RA (F).
Moreover, the set RA (F) is independent of any complete reduction A of F.
We also show that for one-dimensional Cohen-Macaulay local ring (R, m), rA (F) is independent of
any complete reduction A of F.
In section six, we provide a relation between reduction vectors of good complete reductions and
postulation vectors of multigraded filtration of ideals in two-dimensional Cohen-Macaulay local rings.
For bigraded filtration we prove a result which relates the Cohen-Macaulayness of the bigraded Rees
algebra, the complete reduction number, reduction numbers and the joint reduction number.
Theorem 1.13. Let (R, m) be a Cohen-Macaulay local ring of dimension two with infinite residue
field and I1 , . . . , Is be m-primary ideals of R and s ≥ 2. Let F = {F(n)}n∈Zs be an I-admissible
filtration of ideals in R and A = {xij ∈ Ii : j = 1, 2; i = 1, . . . , s} be a good complete reduction of F.
1
Let HR
(R(F))n = 0 for all n ≥ 0. Then P(F) ⊆ Ns and there exists a one-to-one correspondence
++
f : P(F) ←→ {r ∈ RA (F) | r ≥ e}
defined by f (n) = n + e where f −1 (r) = r − e.
6
Theorem 1.14. Let (R, m) be a Cohen-Macaulay local ring of dimension two with infinite residue field
and I1 , . . . , Is be m-primary ideals of R and s ≥ 2. Let F = {F(n)}n∈Zs be an I-admissible filtration
1
of ideals in R and HR
(R(F))n = 0 for all n ≥ 0. Then the following statements are equivalent.
++
(1) P(F) = Ns , i.e. PF (n) = HF (n) for all n ≥ 0.
(2) rA (F) ≤ 1 for any good complete reduction A of F.
(2′ ) There exists a good complete reduction A of F such that rA (F) ≤ 1.
In order to state the final result, we recall the definition of joint reduction of multigraded filtrations
[12]. The joint reduction of F of type q = (q1 , . . . , qs ) ∈ Ns is a collection of qi elements xi1 , . . . , xiqi ∈
Ii for all i = 1, . . . , s such that q1 + · · · + qs = d and
qi
s X
X
i=1 j=1
xij F(n − ei ) = F(n) for all large n.
We say the joint reduction number of F with respect to a joint reduction {xij ∈ Ii : j =
1, . . . , qi ; i = 1, . . . , s} of type q is zero if
qi
s X
X
i=1 j=1
xij F(n − ei ) = F(n) for all n ≥
X
i∈A
ei , where A = {i|qi 6= 0}.
We say the joint reduction number of F of type q is zero if the joint reduction number of F
with respect to any joint reduction of type q is zero.
Theorem 1.15. Let (R, m) be a Cohen-Macaulay local ring of dimension two with infinite residue
field and I, J be m-primary ideals of R. Let F = {F(n)}n∈Z2 be a Z2 -graded (I, J)-admissible filtration
of ideals in R. Then the following are equivalent.
(1) The Rees algebra R(F) is Cohen-Macaulay.
(2) P(F) = N2 , i.e. PF (n) = HF (n) for all n ≥ 0.
1
(3) For any good complete reduction A of F, rA (F) ≤ 1 and HR
(R(F))n = 0 for all n ≥ 0.
++
1
(3′ ) There exists a good complete reduction A of F such that rA (F) ≤ 1 and HR
(R(F))n = 0
++
for all n ≥ 0.
(4) For the filtrations F (i) = {F(nei )}n∈Z , r(F (i) ) ≤ 1 where i = 1, 2 and the joint reduction
number of F of type e is zero.
2. Preliminary results
In this section we discuss the existence of good complete reduction of an I-admissible filtration
F = {F(n)}n∈Zs and prove some results about Hilbert coefficients which we need in the following
sections. For an admissible multigraded filtration F = {F(n)}n∈Zs , by Rees’ Lemma [15, Lemma 1.2],
[12, Lemma 2.2], we get elements xi ∈ Ii for all i = 1, . . . , s, called superficial elements for F, such that
7
for each i, there exists an integer ri and (xi ) ∩ F(n) = xi F(n − ei ) for all n ≥ ri ei . In [15, Theorem
1.3], Rees proved existence of complete reduction of the filtration {I n }n∈Zs . Using the same lines of
proof of this theorem and existence of superficial elements we get the following theorem.
Theorem 2.1. Let (R, m) be a Cohen-Macaulay local ring of dimension d with infinite residue field
and I1 , . . . , Is be m-primary ideals of R. Let F be an I-admissible filtration of ideals in R. Then there
exists a good complete reduction of F.
Lemma 2.2. Let (R, m) be a Noetherian local ring of dimension d ≥ 1 and I1 , . . . , Is be m-primary
ideals of R. Put I = I1 · · · Is . Then
X
α=(α1 ,...,αs )∈Ns
|α|=d
d!eα (I)
= e0 (I) and e0 (I) = ed (I).
α1 ! · · · αs !
Proof. Since for large n,
PI (ne) = HI (ne) = λ
R
I ne
=λ
R
In
= PI (n),
comparing the coefficients of nd and constant terms we get the required result.
Proposition 2.3. Let s ≥ 1 be a fixed integer and i1 , . . . , is ∈ N such that g = i1 + · · · + is ≥ 1. Then
∆g (n1 i1 · · · ns is ) = g! where nk ∈ Z for all k = 1, . . . , s.
Proof. We use induction on g. Let g = 1. Then without loss of generality assume i1 = 1 and ik = 0
for all k 6= 1. Therefore ∆1 (n1 ) = (n1 + 1) − n1 = 1. Let g ≥ 2 and assume the result is true up to
g − 1. Now
h
i
∆g (n1 i1 · · · ns is ) = ∆g−1 ∆1 (n1 i1 · · · ns is )
h
i
= ∆g−1 (n1 + 1)i1 · · · (ns + 1)is − n1 i1 · · · ns is
= ∆g−1
s
hX
k=1
=
=
s
X
k=1
s
X
k=1
i
ik (n1 i1 · · · nk ik −1 · · · ns is )
ik ∆g−1 (n1 i1 · · · nk ik −1 · · · ns is )
ik (g − 1)!
= g!.
8
Proposition 2.4. Let (R, m) be a Noetherian local ring of dimension d ≥ 1 and I1 , . . . , Is be m-primary
ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of ideals in R. Then
(1) eα (F) = eα (I) for all α ∈ Ns where |α| = d.
(2) ∆d (PF (n)) = ∆d (PI (n)) = e0 (I1 · · · Is ).
(3) For an ideal J ⊆ F(e) such that JF(n) = F(n + e) for all large n ∈ Ns , we have ∆d (PF (n)) =
e0 (J) = e0 (I1 · · · Is ).
Proof. (1) This follows from [15, Theorem 2.4].
(2) Using Proposition 2.3, we get
n
+
α
−
1
n
+
α
−
1
1
1
s
s
(−1)d−|α| eα (F)
···
∆d (PF (n)) = ∆d
α1
αs
|α|≤d
X eα (F)
= ∆d
n1 α1 · · · ns αs
α1 ! · · · αs !
X
|α|=d
=
X
i
eα (F) h d α1
∆ (n1 · · · ns αs )
α1 ! · · · αs !
X
eα (F)
(α1 + · · · + αs )!
α1 ! · · · αs !
|α|=d
=
|α|=d
=
X
|α|=d
d!
eα (F)
.
α1 ! · · · αs !
In a similar way we get ∆d (PI (n)) =
X
d!
|α|=d
∆d (PI (n)) = e0 (I1 · · · Is ).
eα (I)
. By part (1) and Lemma 2.2, ∆d (PF (n)) =
α1 ! · · · αs !
(3) Since JF(n) = F(n + e) for all large n, there exists an integer k ∈ N such that JF(n) = F(n + e)
for all n ≥ ke. Now for all n ≥ k, J n−k F(ke) = F(ne) and hence J n ⊆ F(ne) ⊆ J n−k . This implies
R
R
R
λ
≤λ
≤λ
n−k
J
F(ne)
Jn
for all n ≥ k. Therefore for all n ≥ k, we have
PJ (n − k)
PF (ne)
PJ (n)
≤ lim
≤ lim d
d
d
n→∞
n→∞ n /d!
n→∞ n /d!
n /d!
lim
which implies ∆d (PF (n)) = e0 (J). Hence using part (2), we get the required result.
Proposition 2.5. Let (R, m) be a Noetherian local ring of dimension d ≥ 1, depth R ≥ 1 and
I1 , . . . , Is be m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of ideals in
9
R and A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s} be a complete reduction of F. Then Ji = (xi1 , . . . , xid )
is a reduction of Ii for all i = 1, . . . , s.
Proof. For all large n, JF(n) = F(n+e) where yj = x1j · · · xsj for all j = 1, . . . , d and J = (y1 , . . . , yd ).
Since F is an I-admissible filtration, for each i ∈ {1, . . . , s}, there exists ri ∈ N such that for all n ≥ ri ei ,
F(n + ei ) = Ii F(n). Hence for all large n,
Ii F(n + e − ei ) ⊇ Ji F(n + e − ei ) ⊇ JF(n) = F(n + e) ⊇ Ii F(n + e − ei ).
Now by [15, Lemma 1.5], Ji is reduction of Ii for all i = 1, . . . , s.
Proposition 2.6. Let f (n) : Zs → Z be an integer valued function such that for all large n ∈ Ns ,
f (n) = 0. Let B = {n ∈ Zs : f (m) = 0 for all m ≥ n} and Cj = {n ∈ Zs : ∆j (f (m)) = 0 for all m ≥ n}
for all j ≥ 0. Then for j ≥ 0, B = Cj .
Proof. For j = 0, the result is true due to the definition of B. It is enough to prove the statement
for j = 1. Let n ∈ B. Then ∆1 (f (m)) = f (m + e) − f (m) = 0 for all m ≥ n. This implies n ∈ C1 .
Conversely, let n ∈ C1 . Then for all m ≥ n, 0 = ∆1 (f (m)) = f (m + e) − f (m). Let k ∈ Ns such that
f (r) = 0 for all r ≥ k. Let m ≥ n. For all i = 1, . . . , s, define
(
ki − mi
if ki > mi
u(m)i =
0
if ki ≤ mi
Let u(m) = max{u(m)1 , . . . , u(m)s } + 1. Then for all m ≥ n,
0 = f (m + u(m)e) = · · · = f (m + e) = f (m).
Hence n ∈ B.
Proposition 2.7. Let R =
L
n∈Ns
.b
Rn be an Ns -graded ring, S be an Ns -graded R-algebra and b ∈ Re .
Let Sn −→ Sn+e be an injective map for all large n and grade(S++ ) ≥ 1. Then b is a nonzerodivisor
of S.
.b
Proof. Let m ∈ Ns such that for all n ≥ m, Sn −→ Sn+e is an injective map. Let x ∈ (0 :S b) ∩ Sk for
some k ∈ Ns . We show that x(S++ )m+1 = 0 where m = max{m1 , . . . , ms }. Let 0 6= z ∈ (S++ )m+1 ∩ Sp .
Now xz ∈ Sk+p and bxz = 0. Since k + p ≥ (m + 1)e, xz = 0. Thus x ∈ (0 :S (S++ )m+1 ) = 0.
3. An analogue of Huneke’s fundamental lemma
Throughout this section (R, m) is a Cohen-Macaulay local ring of dimension d ≥ 1 with infinite
residue field and I1 , . . . , Is are m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration
of ideals in R. In [8], A. V. Jayanthan and J. K. Verma generalized the Kirby-Mehran complex [9],
10
[6] for the bigraded filtration {I r J s }r,s∈Z where I, J are m-primary ideals and studied the relation
between cohomology modules of this complex and local cohomology modules of R(I, J). We construct
a multigraded analogue of the Kirby-Mehran complex and compute its homology modules. As a
consequence of this we prove an analogue of Huneke’s Fundamental Lemma [7],[5].
Let y1 , . . . , yk be elements in I1 · · · Is where 1 ≤ k ≤ d. For l ≥ 1 and (yt)[l] = y1 l tle , . . . , yk l tle , consider
the Koszul complex K.((yt)[l] , R′ (F)) :
k
k
0 −→ R′ (F) −→ R′ (F)(le)( 1) −→ · · · −→ R′ (F)((k − 1)le)(k−1) −→ R′ (F)(kle) −→ 0.
This complex has a Zs -graded structure inherited from R′ (F). The graded component of degree n of
the above Koszul complex is K.n ((yt)[l] , R′ (F)) :
k
k
0 −→ F(n) −→ (F(n + le))( 1) −→ · · · −→ (F(n + (k − 1)le))(k−1) −→ F(n + kle) −→ 0.
Let y [l] = y1 l , . . . , yk l . We can consider the above complex as a subcomplex of the Koszul complex
K.(y [l] , R) :
k
k
0 −→ R −→ R(1) −→ · · · −→ R(k−1) −→ R −→ 0.
Hence we have a chain map of complexes 0 −→ K.n ((yt)[l] , R′ (F)) −→ K.(y [l] , R) which produces a
quotient complex C.(y [l] , F(n)) :
R
dk
−→
0 −→
F(n)
(k)
( k )
1 d
k−1
R
R
R
d2
d0
d1
k−1
−→ · · · −→
−→
0.
−→
F(n + le)
F(n + (k − 1)le)
F(n + kle)
In the following proposition we compute homology modules of the above complex.
Proposition 3.1. For all l ≥ 1, n ∈ Zs and 1 ≤ k ≤ d,
R
,
(1) H0 (C.(y [l] , F(n))) =
(F(n + kle), y [l] )
(F(n + le) : (y [l] ))
(2) Hk (C.(y [l] , F(n))) =
,
F(n)
(3) if y1 , . . . , yk is a regular sequence then H1 (C.(y [l] , F(n))) =
((y [l] ) ∩ F(n + kle))
(y [l] )F(n + (k − 1)le)
.
Proof. (1) Since ker d0 = R/F(n + kle) and im d1 = (y [l] ) + F(n + kle) /F(n + kle), we get
H0 (C.(y [l] , F(n))) =
R
ker d0
=
.
im d1
(F(n + kle), y [l] )
11
(2) Since im dk+1 = 0,
Hk (C.(y [l] , F(n))) = ker dk
= {x + F(n) ∈ R/F(n)|xyi l ∈ F(n + le) for all i = 1, . . . , k}
= {x + F(n) ∈ R/F(n)|x ∈ ∩ki=1 (F(n + le) : (yi l ))}
= {x + F(n) ∈ R/F(n)|x ∈ (F(n + le) : (y [l] ))}.
(3) Since y1 , . . . , yk is a regular sequence, the following sequence is exact
k
k
R(k−2) −→ R(k−1) −→ (y [l] ) −→ 0.
Tensoring by
R
F (n+(k−1)le) ,
φ2
φ1
φ0
we get an exact sequence
( k )
( k )
k−2
k−1
(y [l] )
R
R
φ2
φ1
φ0
−→
−→
−→
0.
[l]
F(n + (k − 1)le)
F(n + (k − 1)le)
(y )F(n + (k − 1)le)
Hence im φ2 = im d2 and we get the following commutative diagram of exact rows:
0 −−−−→ im d2 −−−−→
iy
0 −−−−→ ker d1 −−−−→
R
F (n+(k−1)le)
idy
R
F (n+(k−1)le)
k
k−1
(
)
k
k−1
(
)
−−−−→
(y [l] )
(y [l] )F (n+(k−1)le)
−−−−→
θy
−−−−→ 0
R
F (n+kle)
where i is the inclusion map and id is the identity map. Then by the Snake Lemma,
H1 (C.(y [l] , F(n))) =
((y [l] ) ∩ F(n + kle))
ker d1 ∼
.
= ker θ = [l]
im d2
(y )F(n + (k − 1)le)
Theorem 3.2. (Analogue of Huneke’s Fundamental Lemma) Let (R, m) be a Cohen-Macaulay local
ring of dimension d ≥ 1 with infinite residue field, I1 , . . . , Is be m-primary ideals of R and F =
{F(n)}n∈Zs be an I-admissible filtration of ideals in R. Let A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s}
be any complete reduction of F, yj = x1j · · · xsj for all j = 1, . . . , d. Let y = y1 , . . . , yd and J = (y).
Then for all n ∈ Zs ,
d
∆ (PF (n) − HF (n)) = λ
F(n + de)
JF(n + (d − 1)e)
−
d
X
i=2
(−1)i λ(Hi (C.(y, F(n)))).
12
Proof. By Propositions 2.4 and 3.1, we get
∆d (PF (n) − HF (n)) = e0 (J) − ∆d (HF (n))
= e0 (J) −
R
= λ
(y)
−
d
X
i=2
d
X
i=0
(−1)i λ(Hi (C.(y, F(n))))
R
−λ
(F(n + de), (y))
+λ
(y) ∩ F(n + de)
(y)F(n + (d − 1)e)
(−1)i λ(Hi (C.(y, F(n))))
= λ
F(n + de)
(y)F(n + (d − 1)e)
= λ
F(n + de)
JF(n + (d − 1)e)
−
−
d
X
i=2
d
X
i=2
(−1)i λ(Hi (C.(y, F(n))))
(−1)i λ(Hi (C.(y, F(n)))).
4. Vanishing of Hilbert coefficients
1
In this section we compute the local cohomology module HR
(R(F))n for all n ≥ 0 [2],[8]. We
++
discuss the vanishing of Hilbert coefficients of an I-admissible filtration F and generalize some results
due to Marley [11] in Cohen-Macaulay local ring of dimension 1 ≤ d ≤ 2.
For a Zs -graded I-filtration F = {F(n)}n∈Zs , the filtration F̆ = {F̆ (n)}n∈Zs of ideals is called
S
(F(n + ke) : F(e)k ) for all n ∈ Ns and F̆(n) =
Ratliff-Rush closure filtration of F where F̆ (n) =
k≥1
1
(R(F))n we follow the lines of the proof [2, Theorem
F̆(n+ ) for all n ∈ Zs [12]. To compute HR
++
3.3].
Proposition 4.1. Let (R, m) be a Cohen-Macaulay local ring of dimension d ≥ 2 with infinite residue
field, I1 , . . . , Is be m-primary ideals of R and F = {F(n)}n∈Zs be an I-admissible filtration of ideals
in R. Then for all n ∈ Ns ,
F̆(n)
1
HR
(R(F))n ∼
.
=
++
F(n)
Proof. Let A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s} be any complete reduction of F, yj = x1j · · · xsj
for all j = 1, . . . , d. For each n, l ≥ 1, (yt)[l] = y1 l tle , . . . , yd l tle and y [l] = y1 l , . . . , yd l , we have the
following exact sequence
0 −→ K.n ((yt)[l] , R′ (F)) −→ K.(y [l] , R) −→ C.(y [l] , F(n)) −→ 0.
13
For each i ∈ {1, . . . , d} the following commutative diagram of complexes :
yi l
K.(yi l , R) :
0 −−−−→ R −−−−→ R −−−−→ 0
yi y
idy
yi l+1
K.(yi l+1 , R) : 0 −−−−→ R −−−−→ R −−−−→ 0
gives a map K.(y1 l , . . . , yd l , R) = ⊗di=1 K.(yi l , R) → K.(y1 l+1 , . . . , yd l+1 , R) = ⊗di=1 K.(yi l+1 , R). The
maps can be restricted to K.n (y1 l tle , . . . , yd l tle , R′ (F)). Hence for all l ≥ 1, we get morphisms of exact
sequences
0 −−−−→ K.n ((yt)[l] , R′ (F))
y
−−−−→ K.(y [l] , R) −−−−→ C.(y [l] , F(n)) −−−−→ 0
y
y
0 −−−−→ K.n ((yt)[l+1] , R′ (F)) −−−−→ K.(y [l+1] , R) −−−−→ C.(y [l+1] , F(n)) −−−−→ 0
which produce an inductive system of exact sequences of complexes. Applying lim
to the long exact
−→
l
sequence of cohomology modules, we get
0
0
1
0 −→ H(yt)
(R′ (F))n −→ H(y)
(R) −→ lim
(R′ (F))n −→ · · · .
H 0 (C.(y [l] , F(n))) −→ H(yt)
−→
l
i (R) = 0 for 0 ≤ i ≤ d − 1. Hence
Since (R, m) is Cohen-Macaulay, H(y)
1
(R′ (F))n ∼
H 0 (C.(y [l] , F(n))) = lim
Hd (C.(y [l] , F(n))) = lim
H(yt)
= lim
−→
−→
−→
l
Since
√
R++ =
q
l
l
(F(n + le) : (y [l] ))
.
F(n)
(yt), by [12, Proposition 3.1, Proposition 4.2], for all n ∈ Ns ,
F̆ (n)
1
1
.
HR
(R(F))n ∼
(R′ (F))n ∼
= HR
=
++
++
F(n)
For all i = 1, . . . , s, we denote the associated multigraded ring of F with respect to F(ei ) by
M F(n)
Gi (F) =
. For F = {I n }n∈Zs , we set Gi (F) = Gi (I). In the next proposition we give
F(n
+
e
)
i
s
n∈N
1
an equivalent criterion for the vanishing of HR
(R(F))n for all n ≥ 0 in terms of grade(Gi (F)++ )
++
for all i = 1, . . . , s.
Proposition 4.2. Let (R, m) be a Cohen-Macaulay local ring of dimension d ≥ 2 with infinite residue
field, I1 , . . . , Is be m-primary ideals of R and F = {F(n)}n∈Zs be an I-admissible filtration of ideals
in R. Then the following statements are equivalent.
1
(1) For all n ∈ Ns , HR
(R(F))n = 0.
++
(2) For all i = 1, . . . , s, grade(Gi (F)++ ) ≥ 1.
14
Proof. Fix i. Denote
R′ (F)
by G′i (F). Consider the short exact sequence of R(I)-modules,
R′ (F)(ei )
0 −→ R′ (F)(ei ) −→ R′ (F) −→ G′i (F) −→ 0.
This induces a long exact sequence of local cohomology modules,
0
1
0 −→ HR
(G′i (F))n −→ HR
(R′ (F))n+ei −→ · · · .
++
++
(4.2.1)
1
1
(R′ (F))n = 0 for all
Since HR
(R(F))n = 0 for all n ∈ Ns , by [12, Proposition 4.2], we have HR
++
++
n ∈ Ns . Hence from the exact sequence (4.2.1) and [12, Proposition 4.2], for all n ∈ Ns , we get
0
HG
i (F )
++
0
0
(G′i (F))n ∼
(G′i (F))n = 0.
(Gi (F))n ∼
= HG
= HR
++
i (F )++
Conversely, suppose grade(Gi (F)++ ) ≥ 1 for all i = 1, . . . , s. By [12, Theorem 3.3], for all n ∈ Ns ,
F̆(n + ei ) ∩ F(n) = F(n + ei ).
We show that if F̆(n) = F(n) for some n ≥ 0 then F̆(m) = F(m) for all m ≥ n. Let ti = mi − ni for
all i = 1, . . . , s. For each i,
F(n + ei ) = F̆(n + ei ) ∩ F(n) = F̆(n + ei ) ∩ F̆(n) = F̆(n + ei ).
Continuing this process ti times for each i we get F̆ (m) = F(m). Since F̆ (0) = F(0), by Proposition
4.1, we get the required result.
Remark 4.3. Note that if (R, m) is an analytically unramified local ring of dimension d ≥ 1, I1 , . . . , Is
1
(R(F))n =
are m-primary ideals of R then by [12, Corollary 3.4], for the filtration F = {I n }n∈Zs , HR
++
0 for all n ∈ Ns .
Proposition 4.4. Let (R, m) be a Cohen-Macaulay local ring of dimension d ≥ 2 with infinite residue
field and I1 , . . . , Is be m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of
ideals in R and A = {xij ∈ Ii : j = 1, . . . , d; i = 1, . . . , s} be a good complete reduction of F. Put
1
y1 = x11 . . . xs1 . Let HR
(R(F))n = 0 for all n ∈ Ns . Then
++
(1) yi1 = y1 + I e+ei ∈ Gi (I)e is Gi (F)-regular for all i = 1, . . . , s.
(2) (y1 ) ∩ F(n) = y1 F(n − e) for all n ≥ e.
Proof. (1) Fix i. Let m ≥ e such that (y1 ) ∩ F(n) = y1 F(n − e) for all n ≥ m. We show that
.yi1
(Gi (F))n −→ (Gi (F))n+e is injective for all n ≥ m. Let (z + F(n + ei ))yi1 = F(n + e + ei ). Then y1 z ∈
F(n + e + ei ). Since n ≥ m, z ∈ F(n + ei ). Hence by Propositions 2.7 and 4.2, yi1 is a nonzerodivisor
15
of Gi (F).
(2) For all i = 1, . . . , s, consider the Koszul complex Ki . = Ki .(yi1 , Gi (F)) :
0 −→ Gi (F) −→ Gi (F)(e) −→ 0.
The nth component of this complex is Ki .(yi1 , Gi (F), n) :
0 −→ Gi (F)(n) −→ Gi (F)n+e −→ 0.
Hence for all i = 1, . . . , s and n ≥ 0, we have exact sequence of complexes
0 −→ Ki .(yi1 , Gi (F), n) −→ C.(y1 , F(n + ei )) −→ C.(y1 , F(n)) −→ 0,
which gives a long exact sequence of homology modules
· · · −→ Hj (Ki .(yi1 , Gi (F), n)) −→ Hj (C.(y1 , F(n + ei ))) −→ Hj (C.(y1 , F(n))) −→ · · · .
(4.4.1)
Since yi1 is a nonzerodivisor of Gi (F) for all i = 1, . . . , s, we have H1 (Ki .(yi1 , Gi (F), n)) = 0 for all n
and i = 1, . . . , s. Since H1 (C.(y1 , F(0))) = 0, applying (4.4.1) several times for all i = 1, . . . , s, we get
H1 (C.(y1 , n)) = 0 for all n ≥ 0. Hence F(n) ∩ (y1 ) = y1 F(n − e) for all n ≥ e.
Proposition 4.5. Let (R, m) be a Cohen-Macaulay local ring of dimension two and I1 , . . . , Is be
m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of ideals in R and A =
{xij ∈ Ii : j = 1, 2; i = 1, . . . , s} be a good complete reduction of F. Let yj = x1j · · · xsj for all j = 1, 2
1
and HR
(R(F))n = 0 for all n ∈ Ns . Then H2 (C.(y1 , y2 , F(n))) = 0 for all n ≥ 0.
++
Proof. Fix i. Let yij = yj + I e+ei ∈ Gi (I)e for all j = 1, 2 and i = 1, . . . , s. Consider the Koszul
complex K.(yi1 , yi2 , Gi (F)):
0 −→ Gi (F) −→ (Gi (F)(e))2 −→ Gi (F)(2e) −→ 0
whose nth component is
0 −→ Gi (F)n −→ (Gi (F)n+e )2 −→ Gi (F)n+2e −→ 0.
For all n ≥ 0, we have the exact sequence
0 −→ K.(yi1 , yi2 , Gi (F), n) −→ C.(y1 , y2 , F(n + ei )) −→ C.(y1 , y2 , F(n)) −→ 0.
This gives a long exact sequence of homology modules
· · · −→ Hj (K.(yi1 , yi2 , Gi (F), n)) −→ Hj (C.(y1 , y2 , F(n + ei ))) −→ Hj (C.(y1 , y2 , F(n))) −→ · · ·
for all n ≥ 0. Since by Proposition 4.4, yi1 is a regular element in Gi (F) for all i = 1, . . . , s, we have
H2 (K.(yi1 , yi2 , Gi (F))) = 0 for all n and i = 1, . . . , s. Since H2 (C.(y1 , y2 , F(0))) = 0, using the above
exact sequence several times for all i = 1, . . . , s, we get H2 (C.(y1 , y2 , F(n))) = 0 for all n ≥ 0.
16
Theorem 4.6. Let (R, m) be a Cohen-Macaulay local ring of dimension 1 ≤ d ≤ 2 with infinite residue
field, I1 , . . . , Is be m-primary ideals of R and F = {F(n)}n∈Zs be an I-admissible filtration of ideals
in R. Let e(d−1)ei (F) = 0 for i = 1, . . . , s. Then
(1) For d = 1, PF (n) = HF (n) for all n ∈ Ns .
1
(2) For d = 2, if HR
(R(F))n = 0 for all n ∈ Ns then PF (n) = HF (n) for all n ∈ Ns and
++
e0 (F) = 0.
Proof. (1) For d = 1, since e0 (F) = 0, we get PF (0) = HF (0). Hence by the difference formula [12,
1
(R(F))
Theorem 4.3], we get λR HR
= PF (0) − HF (0) = 0. Since dim R = 1, by [12, Lemma
0
++
1
2.11], for all n ∈ Ns , HR
(R(F))n = 0. Therefore using the difference formula [12, Theorem 4.3]
++
again, we get PF (n) − HF (n) = 0 for all n ∈ Ns .
(2) Let d = 2, A = {xij ∈ Ii : j = 1, 2; i = 1, . . . , s} be any good complete reduction of F, yj =
x1j · · · xsj for j = 1, 2 and J = (y1 , y2 ). Fix i. Let R′ = R/(xi1 ) and ′ denote the image of an ideal in
R′ . For all large n, consider the following exact sequence
0 −→
R
R
R
(F(n) : (xi1 ))
.xi1
−→
−→
−→ 0.
−→
F(n − ei )
F(n − ei )
F(n)
(xi1 , F(n))
Since A is a good complete reduction, for all large n, (F(n) : (xi1 )) = F(n − ei ) and
λ
R
(xi1 , F(n))
=λ
R
F(n)
−λ
R
F(n − ei )
.
Therefore for the filtration F ′ = {F(n)R′ }n∈Zs , we have PF ′ (n) = PF (n) − PF (n − ei ). This implies
that the constant term of PF ′ (n) is eei (F) = 0.
Since A′ = {x′i2 ∈ Ii : i = 1, . . . , s} is a complete reduction of F ′ , J ′ = (y2′ ) and dim R′ = 1, by
Theorem 3.2, Proposition 2.6 and part (1), we have
λ
F(n + e)R′
J ′ F(n)R′
= ∆1 (PF ′ (n) − HF ′ (n)) = 0 for all n ∈ Ns .
Thus we get F(n + e) = y2 F(n) + ((xi1 ) ∩ F(n + e)) for all n ∈ Ns . We show that (xi1 ) ∩ F(n + e) =
xi1 F(n + e − ei ) for all n ∈ Ns . It is clear that xi1 F(n + e − ei ) ⊆ (xi1 ) ∩ F(n + e). Let axi1 ∈ F(n + e).
1
Then ay1 ∈ F(n +2e−ei ). Since HR
(R(F))n = 0 for all n ∈ Ns and A is a good complete reduction,
++
by Proposition 4.4, a ∈ F(n + e − ei ). Hence we get
F(n + e) = y2 F(n) + xi1 F(n + e − ei ) for all n ∈ Ns and i = 1, . . . , s.
17
We show that F(n + 2e) = JF(n + e) for all n ∈ Ns . Let n ∈ Ns . Then
F(n + 2e) = y2 F(n + e) + x11 F(n + 2e − e1 )
= y2 F(n + e) + x11 (y2 F(n + e − e1 ) + x21 F(n + 2e − e1 − e2 ))
⊆ y2 F(n + e) + x11 x21 F(n + 2e − e1 − e2 )
..
.
⊆ y2 F(n + e) + x11 · · · xs1 F(n + e) = (y1 , y2 )F(n + e) = JF(n + e).
1
Since HR
(R(F))n = 0 for all n ∈ Ns , by Theorem 3.2 and Proposition 4.5, for all n ∈ Ns , we get
++
F(n + 2e)
2
∆ (PF (n) − HF (n)) = λ
JF(n + e)
and hence by Proposition 2.6, we have PF (n) = HF (n) for all n ∈ Ns . Putting n = 0 in the above
equality we get e0 (F) = 0.
Theorem 4.7. Let (R, m) be a Cohen-Macaulay local ring of dimension 1 ≤ d ≤ 2 with infinite
residue field and I, J be m-primary ideals of R. Let F = {F(r, s)}r,s∈Z be a Z2 -graded (I, J)-admissible
filtration of ideals in R. Then the following statements are equivalent.
(1) e(d−1)ei (F) = 0 for i = 1, 2.
(2) I and J are generated by system of parameters, PF (r, s) = HF (r, s) for all r, s ∈ N and
F(r, s) = I r J s for all r, s ∈ Z.
(3) eα (F) = 0 for |α| ≤ d − 1.
Proof. (1) ⇒ (2) : Let F (1) = {F(r, 0)}r∈Z and F (2) = {F(0, s)}s∈Z . Since F is an (I, J)-admissible
filtration, F (1) and F (2) are I-admissible and J-admissible filtrations respectively. By [13], [11, Lemma
3.19] and [12, Theorem 5.5], we have
0 ≤ e1 (I) ≤ e1 (F (1) ) ≤ e(d−1)e1 (F) = 0 and 0 ≤ e1 (J) ≤ e1 (F (2) ) ≤ e(d−1)e2 (F) = 0.
Then by [11, Theorem 3.21], we get I and J are generated by system of parameters, F(r, 0) = I r and
F(0, s) = J s for all r, s ∈ Z.
Let d = 1. Then by Theorem 4.6, PF (r, s) = HF (r, s) for all r, s ∈ N. It is enough to prove that
F(r, s) = I r J s for all r, s ≥ 1. Since I, J are generated by system of parameters, for r, s ≥ 1, we have
r
R
R
R
I
R
R
λ
+
λ
=
λ
+
λ
=
λ
.
= PF (r, s) = re(I) + se(J) = λ
F(r, s)
Ir
Js
Ir
IrJ s
IrJ s
This implies F(r, s) = I r J s for all r, s ≥ 1.
Let d = 2. Since I, J are parameter ideals, we have
R
R
R
R
e(I) − ee1 (F) = e(I) = λ
=λ
and e(J) − ee2 (F) = e(J) = λ
=λ
.
I
F(1, 0)
J
F(0, 1)
18
Therefore by [12, Theorem 7.3], we get PF (r, s) = HF (r, s) for all r, s ∈ N. It is enough to prove that
F(r, s) = I r J s for all r, s ≥ 1. By [12, Theorem 7.3], the joint reduction number of F of type e is zero.
Let (a, b) be a joint reduction of F of type e. Then
F(r, s) = aF(r − 1, s) + bF(r, s − 1) for all r, s ≥ 1.
We use induction on r + s. Let r, s ≥ 1. If r + s = 2 then r = s = 1 and
F(1, 1) = aF(0, 1) + bF(1, 0) = aJ + bI ⊆ IJ ⊆ F(1, 1).
Let r + s > 2. Then r ≥ 2 or s ≥ 2. Without loss of generality assume r ≥ 2. If s = 1 then using
induction we get
F(r, 1) = aF(r − 1, 1) + bF(r, 0) = aI r−1 J + bI r ⊆ I r J ⊆ F(r, 1).
Hence we may assume s ≥ 2. Therefore
F(r, s) = aF(r − 1, s) + bF(r, s − 1) = aI r−1 J s + bI r J s−1 ⊆ I r J s ⊆ F(r, s).
(2) ⇒ (3) : For d = 1, putting r = s = 0 in the equation PF (r, s) = HF (r, s), we get the required
result. Let d = 2. Since PF (r, s) = HF (r, s) for all r, s ∈ N, we have
R
R
e0 (F) = 0, e(I) − ee1 (F) = λ
and e(J) − ee2 (F) = λ
.
F(1, 0)
F(0, 1)
Since I, J are parameter ideals and F(r, s) = I r J s for all r, s ∈ Z, by [12, Theorem 5.5], we have
ee1 (F) = e1 (F (1) ) = e1 (I) = 0 and ee2 (F) = e1 (F (2) ) = e1 (J) = 0.
(3) ⇒ (1) : It follows directly.
Example 4.8. Let R = k[|X, Y |]. Then R is a regular local ring of dimension two. Let I = (X, Y 2 ) and
J = (X 2 , Y ). Then I, J are complete parameter ideals in R. Consider the filtration F = {I r J s }r,s∈Z .
Since I, J are complete ideals, by [18, Theorem 2′ , Appendix 5], I r , J s , I r J s are complete ideals. By
[14, Theorem 1.2],
ee1 (F) = e1 (I) = e1 (I) = 0 and ee2 (F) = e1 (J) = e1 (J) = 0.
Since
R
R
R
R
and e(J) − ee2 (F) = e(J) = λ
,
=λ
=λ
e(I) − ee1 (F) = e(I) = λ
I
J
I
J
by [12, Theorem 7.3], we get PF (r, s) = HF (r, s) for all r, s ∈ N.
19
5. Postulation and reduction vectors in dimension one
Let (R, m) be a Cohen-Macaulay local ring of dimension one with infinite residue field and I1 , . . . , Is
be m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of ideals in R. In this
section we prove that the set of reduction vectors of F with respect to any complete reduction is the
same as the set of postulation vectors of F. Thus the set of reduction vectors of F with respect to
any complete reduction is independent of the choice of complete reduction. Then we show that the
complete reduction number of F with respect to any complete reduction is independent of choice of
complete reduction.
Theorem 5.1. Let (R, m) be a Cohen-Macaulay local ring of dimension one with infinite residue field
and I1 , . . . , Is be m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of ideals in
R and A = {ai ∈ Ii : i = 1, . . . , s} be a complete reduction of F. Then
P(F) ⊆ Ns and P(F) = RA (F).
Moreover, the set RA (F) is independent of any complete reduction A of F.
Proof. First we prove that P(F) ⊆ Ns . Suppose there exists n ∈ Zs \ Ns such that n ∈ P(F). Then
there exists at least one i ∈ {1, . . . , s} such that ni < 0. Therefore
R
R
=λ
= PF (n)
PF (n + ei ) = λ
F(n + ei )
F(n)
implies e0 (Ii ) = PF (n + ei )− PF (n) = 0. This contradicts to the fact that e0 (Ii ) > 0. Thus P(F) ⊆ Ns .
Let J = (a1 · · · as ). By Theorem 3.2, for all n ≥ 0,
1
∆ (PF (n) − HF (n)) = λ
F(n + e)
JF(n)
.
Hence by Proposition 2.6, we get the required result.
Example 5.2. Let R = k[|t3 , t4 , t5 |]. Then R is a one-dimensional Cohen-Macaulay local ring with
unique maximal ideal m = (t3 , t4 , t5 ). Consider I = (t3 , t4 ) and J = (t3 ). Then JI 2 = I 3 . Since
!
3
t
(t6 )(IJ)2 = (IJ)3 , A =
is a complete reduction for the filtration I = {I r J s }r,s∈Z . We have
3
t
R
λ J n = 3n for all n ∈ N. Now λ RI = 2, λ IR2 = 4 and for n ≥ 3,
I n = J n−2 I 2 = (t3n−6 )(t6 , t7 , t8 ) = (t3n , t3n+1 , t3n+2 ).
Hence for all n ≥ 2, λ
R
In
= 3n − 2. Let PI (n) = ne0 (I) − e1 (I), PJ (n) = ne0 (J) − e1 (J), PIJ (n) =
ne0 (IJ)− e1 (IJ) and PI (n) = n1 e0 (I)+ n2 e0 (J)− e0 (I) denote the Hilbert polynomials of I, J, IJ and
20
I respectively where n ∈ Z and n = (n1 , n2 ) ∈ Z2 . Then e0 (I) = e0 (J) = 3, e1 (I) = 2 and e1 (J) = 0.
Now by Lemma 2.2, e0 (I) = e1 (IJ). For large n,
PIJ (n) = ne0 (IJ) − e1 (IJ) = λ
R
(IJ)n
=λ
R
I 2n
= 6n − 2.
Hence e0 (I) = e1 (IJ) = 2. This implies PI (n) = 3n1 + 3n2 − 2.
Since (t6 )(IJ) = IJ 3 6= I 2 J 2 and (t6 )(IJ)2 = I 2 J 4 = I 3 J 3 , we have rA (I) = 2.
Note that (t6 )I 2 = J 2 I 2 = JI 3 and (t6 )I = J 2 I 6= I 2 J. Let (1, n) ∈ Z2 such that n ≥ 1. Then
(t6 )IJ n = IJ n+2 = (t3 , t4 )(t3n+6 ) = (t3n+9 , t3n+10 ) 6= (t3n+9 , t3n+10 , t3n+11 ) = I 2 J n+1 . Hence RA (I) =
{m ∈ N2 | m ≥ (2, 0)}.
For all n1 ≥ 2 and n2 ≥ 0,
PI (n1 , n2 ) = 3n1 + 3n2 − 2 = λ
R
I n1 +n2
=λ
R
I n1 J n2
= HI (n1 , n2 )
and PI (1, 0) = 1 6= 2 = λ RI . Let (1, n) ∈ Z2 such that n ≥ 1. Then PI (1, n) = 3n + 1 and HI (1, n) =
λ IJRn = λ (t3n+3R,t3n+4 ) = 3n + 2. Hence for all n = (1, n) where n ≥ 0, PI (1, n) 6= HI (1, n). Thus
RA (I) = P(I).
In the following example we show that we cannot drop the condition of Cohen-Macaulayness in
Theorem 5.1.
k[|X, Y |]
. Then R is a one-dimensional local ring which is not Cohen(X 2 , XY )
Macaulay. Consider the ideals I = (x, y) and J = (y) of R. Then JI = I 2 . Since (y 2 )(IJ) = I 2 J 2 ,
!
y
is a complete reduction for the filtration I = {I r J s }r,s∈Z . We have λ JRn = n + 1 for
A =
y
all n ≥ 1. Now λ RI = 1 and for n ≥ 2, I n = J n−1 I = (y n−1 )(x, y) = (y n ). Hence for all n ≥ 2,
λ IRn = n + 1. Let PI (n) = ne0 (I) − e1 (I), PJ (n) = ne0 (J) − e1 (J), PIJ (n) = ne0 (IJ) − e1 (IJ)
Example 5.3. Let R =
and PI (n) = n1 e0 (I) + n2 e0 (J) − e0 (I) denote the Hilbert polynomials with respect to I, J, IJ and
I respectively where n ∈ Z and n = (n1 , n2 ) ∈ Z2 . Then e0 (I) = e0 (J) = 1. Now by Lemma 2.2,
R
= λ IR2n = 2n + 1. Hence
e0 (I) = e1 (IJ). For large n, PIJ (n) = ne0 (IJ) − e1 (IJ) = λ (IJ)
n
e0 (I) = e1 (IJ) = −1. This implies PI (n) = n1 + n2 + 1. Now IJ = (x, y)(y) = (y 2 ). Hence rA (I) = 0.
This implies RA (I) = N2 . But PI (0, 0) = 1 6= 0 = HI (0, 0). Hence RA (I) 6= P(I).
Theorem 5.4. Let (R, m) be a one-dimensional Cohen-Macaulay local ring and I1 , . . . , Is be mprimary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of ideals in R. Then the
complete reduction number of F with respect to any complete reduction is independent of choice of
complete reduction of F.
21
Proof. Let A = {ai ∈ Ii : i = 1, . . . , s} be a complete reduction of F, J = (a1 · · · as ) and rA (F) = k.
First we show that
k = min {max {t1 , . . . , ts : t = (t1 , . . . , ts ) ∈ RA (F)}}.
If k = 0 then it is true. Suppose k ≥ 1. Let n ∈ Ns such that ni < k for all i = 1, . . . , s and n ∈ RA (F).
Let u = max {n1 , . . . , ns }. Then u < k and n ≤ ue ≤ (k − 1)e. Hence JF(m) = F(m + e) for all m ≥
(k − 1)e. This contradicts to the fact that k is the complete reduction number of F with respect to A.
Thus t ∈ RA (F) implies ti ≥ k for at least one i. Since JF(n) = F(n+e) for all n ≥ ke, there exists r ∈
RA (F) such that max {r1 , . . . , rs } = k. Thus k = min {max {t1 , . . . , ts : t = (t1 , . . . , ts ) ∈ RA (F)}}.
Hence by Theorem 5.1, we get the required result.
6. Postulation and reduction vectors in dimension two
Let (R, m) be a Cohen-Macaulay local ring of dimension two with infinite residue field and I1 , . . . , Is
be m-primary ideals of R. Let F = {F(n)}n∈Zs be an I-admissible filtration of ideals in R. In this
section we provide a relation between the reduction vectors of F with respect to any good complete
reduction and the postulation vectors of F. For a bigraded filtration F, we prove a result which relates
the Cohen-Macaulayness of the bigraded Rees algebra, the complete reduction number, reduction
numbers and the joint reduction number.
Theorem 6.1. Let (R, m) be a Cohen-Macaulay local ring of dimension two with infinite residue
field and I1 , . . . , Is be m-primary ideals of R and s ≥ 2. Let F = {F(n)}n∈Zs be an I-admissible
filtration of ideals in R and A = {xij ∈ Ii : j = 1, 2; i = 1, . . . , s} be a good complete reduction of F.
1
Let HR
(R(F))n = 0 for all n ≥ 0. Then P(F) ⊆ Ns and there exists a one-to-one correspondence
++
f : P(F) ←→ {r ∈ RA (F) | r ≥ e}
defined by f (n) = n + e where f −1 (r) = r − e.
Proof. First we prove that P(F) ⊆ Ns . Suppose there exists n ∈ Zs \ Ns such that n ∈ P(F). Then
there exists at least one i ∈ {1, . . . , s} such that ni < 0. Therefore for any j ∈ {1, . . . , s} with j 6= i
and l ≥ 0,
PF (n + lej + ei ) = λ
R
F(n + lej + ei )
=λ
R
F(n + lej )
= PF (n + lej ).
Thus for l = 1, we get
0 = PF (n + ej + ei ) − PF (n + ej )
X
= (ni + 1)e0 (Ii ) +
nk eek +ei (F) + (nj + 1)eej +ei (F) − eei (F).
k6=i,j
(6.1.1)
22
Then for l = 0, by equation (6.1.1), we get
0 = PF (n + ei ) − PF (n)
X
= (ni + 1)e0 (Ii ) +
nk eek +ei (F) − eei (F)
k6=i
= −eej +ei (F).
This contradicts to the fact that eej +ei (F) > 0 [15, Theorem 2.4]. Hence P(F) ⊆ Ns .
Let yj = x1j · · · xsj for j = 1, 2 and J = (y1 , y2 ). Then by Theorem 3.2 and Proposition 4.5, for all
n ≥ 0,
2
∆ (PF (n) − HF (n)) = λ
F(n + 2e)
JF(n + e)
.
Hence by Proposition 2.6, we get the required result.
Theorem 6.2. Let (R, m) be a Cohen-Macaulay local ring of dimension two with infinite residue field
and I1 , . . . , Is be m-primary ideals of R and s ≥ 2. Let F = {F(n)}n∈Zs be an I-admissible filtration
1
of ideals in R and HR
(R(F))n = 0 for all n ≥ 0. Then the following statements are equivalent.
++
(1) P(F) = Ns , i.e. PF (n) = HF (n) for all n ≥ 0.
(2) rA (F) ≤ 1 for any good complete reduction A of F.
(2′ ) There exists a good complete reduction A of F such that rA (F) ≤ 1.
Proof. (1) ⇒ (2) : Let PF (n) = HF (n) for all n ≥ 0 and A = {xij ∈ Ii : j = 1, 2; i = 1, . . . , s} be any
good complete reduction of F. Let yj = x1j · · · xsj for j = 1, 2 and J = (y1 , y2 ). Then by Theorem 3.2,
Propositions 4.5 and 2.6, for all n ≥ 0,
F(n + 2e)
λ
= 0 =⇒ JF(n + e) = F(n + 2e).
JF(n + e)
Hence rA (F) ≤ 1.
The implication (2) ⇒ (2′ ) is trivial.
(2′ ) ⇒ (1) : Suppose there exists a good complete reduction A = {xij ∈ Ii : j = 1, 2; i = 1, . . . , s} of
F such that rA (F) ≤ 1. Let yj = x1j · · · xsj for j = 1, 2 and J = (y1 , y2 ). Then again by Theorem
3.2 and Proposition 4.5, for all n ≥ 0, ∆2 (PF (n) − HF (n)) = 0. Now using Proposition 2.6, we get
PF (n) − HF (n) = 0 for all n ≥ 0. Since P(F) ⊆ Ns , we get the required result.
1
In the following example we show that we cannot drop the condition on HR
(R(F))n in Theorem
++
6.2.
Example 6.3. Let R = k[|X, Y |]. Then R is a two-dimensional Cohen-Macaulay local ring with
unique maximal ideal m = (X, Y ). Let I = m2 and J = (X 2 , Y 2 ). Since (X 4 , Y 4 )IJ = (X 4 , Y 4 )m4 =
23
m8 = I 2 J 2 , we have A =
X2 Y 2
!
is a complete reduction for the filtration I = {I r J s }r,s∈Z . By [4,
X2 Y 2
Proposition 1.2.2], for all large n = (n1 , n2 ) ∈ N2 , we get
(X 4 ) ∩ I n1 J n2 = X 4 (I n1 J n2 : (X 4 )) = X 4 I n1 −1 J n2 −1 .
| }
Hence A is a good complete reduction for the filtration I. Note that, since I e2 = J˘ = (I k J 1+k : I k J k )
for some large k, JI = I 2 and m is parameter ideal, we have
1
J˘ = (I 2k+1 : I 2k ) = (m4k+2 : m4k ) = m2 6= J and hence HR
(R(I))(0,1) 6= 0.
++
Since (X 4 , Y 4 )IJ = I 2 J 2 , we get rA (I) ≤ 1.
For n ≥ 1,
λ
R
In
=λ
R
m2n
=
2n + 1
2
n+1
=4
− n = PI (n),
2
Since J is parameter ideal and JI = I 2 ,
n+1
R
=4
= PJ (n),
λ
Jn
2
R
4n + 1
n+1
R
=λ
=
= 16
− 6n = PIJ (n)
λ
(IJ)n
m4n
2
2
and
R
R
6n + 1
n+1
λ
=
λ
=
=
36
− 15n = PI 2 J (n).
I 2n J n
m6n
2
2
R
= PI (ne) and PI 2 J (n) = λ
Hence e0 (I) = e0 (J) = 4. Now for large n, PIJ (n) = λ (IJ)
n
PI (2n, n). Comparing the coefficients on both sides we get
r+1
s+1
PI (r, s) = 4
+4
+ 4rs − r − s.
2
2
Then PI (0, 1) = 3 6= 4 = λ R
J .
R
I 2n J n
=
Theorem 6.4. Let (R, m) be a Cohen-Macaulay local ring of dimension two with infinite residue field
and I, J be m-primary ideals of R. Let F = {F(n)}n∈Z2 be a Z2 -graded (I, J)-admissible filtration of
ideals in R. Then the following statements are equivalent.
(1) The Rees algebra R(F) is Cohen-Macaulay.
(2) P(F) = N2 , i.e. PF (n) = HF (n) for all n ≥ 0.
1
(3) For any good complete reduction A of F, rA (F) ≤ 1 and HR
(R(F))n = 0 for all n ≥ 0.
++
1
(3′ ) There exists a good complete reduction A of F such that rA (F) ≤ 1 and HR
(R(F))n = 0
++
for all n ≥ 0.
(4) For the filtrations F (i) = {F(nei )}n∈Z , r(F (i) ) ≤ 1 where i = 1, 2 and the joint reduction
number of F of type e is zero.
24
Proof. The implications (1) ⇒ (2) ⇒ (4) ⇒ (1) and (3) ⇒ (3′ ) ⇒ (2) follow from [12, Theorem 7.3]
and Theorem 6.2 respectively. Enough to show (1) ⇒ (3). Suppose R(F) is Cohen-Macaulay. Then
1
by [12, Proposition 7.2,Theorem 2.15], we get HR
(R(F))n = 0 for all n ≥ 0. Since (1) implies (2),
++
by Theorem 6.2, we get the required result.
The following example illustrates Theorems 6.1, 6.2 and 6.4.
Example 6.5. Let R = K[|X, Y |]. Then R is a 2-dimensional Cohen-Macaulay local ring with unique
maximal ideal m = (X, Y ). Let I = m2 and J = m3 . Since (X 5 , Y 5 )IJ = (X 5 , Y 5 )m5 = m10 = I 2 J 2 ,
!
X2 Y 2
we have A =
is a complete reduction for the filtration I = {I r J s }r,s∈Z and rA (I) ≤ 1. By
3
3
X Y
[4, Proposition 1.2.2], for all large n = (n1 , n2 ) ∈ N2 , we get
(X 5 ) ∩ I n1 J n2 = X 5 (I n1 J n2 : (X 5 )) = X 5 I n1 −1 J n2 −1 .
Hence A is a good complete reduction for the filtration I.
1
We show that HR
(R(I))n = 0 for all n ≥ 0. For n = (n1 , n2 ) ≥ 0, we have
++
|}
I n = (I n+ke : I ke ) = (m2n1 +3n2 +5k : m5k )
|}
for some large k. Since m is parameter ideal, I n = m2n1 +3n2 = I n for all n ≥ 0.
For n ≥ 1,
and
R
R
2n + 1
n+1
λ
=λ
=
=4
− n = PI (n),
In
m2n
2
2
R
R
3n + 1
n+1
λ
=λ
=
=9
− 3n = PJ (n),
Jn
m3n
2
2
R
5n + 1
n+1
R
=λ
=
= 25
λ
− 10n = PIJ (n)
(IJ)n
m5n
2
2
8n + 1
n+1
=λ
=
= 64
− 28n = PIJ 2 (n).
λ
2
2
R
Hence e0 (I) = 4 and e0 (J) = 9. Now for large n, PIJ (n) = λ (IJ)
= PI (ne) and PIJ 2 (n) =
n
R
λ I n J 2n = PI (n, 2n). Comparing the coefficients on both sides we get
n1 + 1
n2 + 1
PI (n1 , n2 ) = 4
+9
+ 6n1 n2 − n1 − 3n2 .
(6.5.1)
2
2
R
n
I J 2n
R
m8n
Hence by [12, Theorem 6.2], the joint reduction number of I of type e is zero. Since (X 2 , Y 2 )I = I 2
and (X 3 , Y 3 )J = J 2 , we have r(I (i) ) ≤ 1 for i = 1, 2, where I (1) = {I r }r∈Z and I (2) = {J s }s∈Z . By [3,
Corollary 2.3, Corollary 2.4], R(I (1) ) and R(I (2) ) are Cohen-Macaulay. Hence by [12, Theorem 7.1],
R(I) is Cohen-Macaulay.
Using (6.5.1), we get PI (n1 , n2 ) = HI (n1 , n2 ) for all n = (n1 , n2 ) ∈ N2 . Thus P(I) = N2 . Since
25
(X 5 , Y 5 )IJ = I 2 J 2 , (X 5 , Y 5 )I 6= I 2 J, (X 5 , Y 5 )J 6= IJ 2 , (X 5 , Y 5 )I 2 = I 3 J and (X 5 , Y 5 )J 2 = IJ 3 , we
have RA (I) = {m ∈ N2 | m ≥ e} ∪ {m ∈ N2 | m ≥ 2e1 } ∪ {m ∈ N2 | m ≥ 2e2 }.
References
[1] P. B. Bhattacharya, The Hilbert function of two ideals, Math. Proc. Cambridge Philos. Soc. 53 (1957), 568-575. 3
[2] C. Blancafort, On the Hilbert functions and cohomology, J. Algebra 192 (1997), no. 1, 439-459. 12
[3] M. Herrmann, E. Hyry and J. Ribbe, On the Cohen-Macaulay and Gorenstein properties of multigraded Rees algebras,
Manuscripta Math. 79 (1993), no. 3-4, 343-377. 24
[4] J. Herzog and T. Hibi, Monomial ideals, Graduate texts in Mathematics, vol. 260, Springer-Verlag, 2011. 23, 24
[5] S. Huckaba, A d-dimensional extension of a lemma of Huneke’s and formulas for the Hilbert coefficients, Proc.
Amer. Math. Soc. 124 (1996), no. 5, 1393-1401. 4, 10
[6] S. Huckaba and T. Marley, Hilbert coefficients and the depths of associated graded rings, J. London Math. Soc. (2)
56 (1997), 64-76. 10
[7] C. Huneke, Hilbert functions and symbolic powers, Michigan Math. J. 34 (1987), 293-318. 4, 10
[8] A. V. Jayanthan and J. K. Verma, Grothendieck-Serre formula and bigraded Cohen-Macaulay Rees algebras, J.
Algebra 254 (2002), 1-20. 9, 12
[9] D. Kirby and H. A. Mehran, Hilbert functions and Koszul complex, J. London Math. Soc. (2), 24 (1981), 459-466.
4, 9
[10] T. Marley, The coefficients of the Hilbert polynomial and the reduction number of an ideal, J. London Math. Soc.
(2) 40 (1989), no. 1, 1-8. 5
[11] T. Marley, Hilbert Functions of Ideals in Cohen-Macaulay Rings, Ph.D. Thesis, Purdue University, (1989). 4, 5, 12,
17
[12] S. K. Masuti, P. Sarkar and J. K. Verma, Hilbert polynomials of multigraded filtrations of ideals, J. Algebra 444
(2015), 527-566. 6, 12, 13, 14, 16, 17, 18, 24
[13] D. G. Northcott, A note on the coefficients of the abstract Hilbert function, London Math. Soc. 35 (1960), 209-214.
17
[14] D. Rees, Hilbert functions and pseudo-rational local rings of dimension two, J. London Math. Soc. (2) 24 (1981),
467-479. 18
[15] D. Rees, Generalizations of reductions and mixed multiplicities, J. London Math. Soc. (2) 29 (1984), no. 3, 397-414.
3, 6, 7, 8, 9, 22
[16] P. Samuel, La notion de multiplicité en algèbre et en géométrie algébrique, J. Math. Pures Appl. 30 (1951), 159-274.
2
[17] B. Teissier, Cycles évanescents, sections planes et conditions de Whitney, Singularités à cargèse, 1972. Astèrisque
7-8, Soc. Math. France, Paris, (1973), 285-362. 1, 3
[18] O. Zariski and P. Samuel, Commutative algebra, Vol 2, van Nostrand, 1960. 18
Department of Mathematics, Indian Institute of Technology Bombay, Mumbai, 400076, India
E-mail address: [email protected]
J. K. Verma, Department of Mathematics, Indian Institute of Technology Bombay, Mumbai, India
E-mail address: [email protected]
| 0math.AC
|
arXiv:0807.0471v2 [math.AC] 6 Sep 2017
GORENSTEIN APPROXIMATION, DUAL FILTRATIONS
AND APPLICATIONS
TONY J. PUTHENPURAKAL
Abstract. We give a two step method to study certain questions regarding
associated graded module of a Cohen-Macaulay (CM) module M with respect
to an m-primary ideal a in a complete Noetherian local ring (A, m). The first
step, we call it Gorenstein
enables us to reduce to the case
L approximation,
n n+1 are Gorenstein and M is a maximal
when both A, Ga (A) =
n≥0 a /a
CM A-module. The second step consists of analyzing the classical filtration
{HomA (M, an )}n∈Z of the dual HomA (M, A). We give many applications of
this point of view. For instance let (A, m) be equicharacteristic & CM. Let
a(Ga (A)) be the a-invariant of Ga (A). We prove:
(1) a(Ga (A)) = − dim A if and only if a is generated by a regular sequence.
(2) If a is integrally closed and a(Ga (A)) = − dim A + 1 then a has minimal
multiplicity.
We extend to modules a result of Ooishi relating symmetry of h-vectors. Finally we prove a conjecture of Itoh, if a is a normal ideal with ea3 (A) = 0 then
Ga (A) is Cohen-Macaulay.
Contents
1. Introduction
2. Notation and Preliminaries
3. CI-Approximation: Some special cases
4. A classical filtration of the dual
5. Some preliminary applications of Gorenstein approximation
6. The initial degree of the dual filtration and the a-invariant
7. a-invariant: Borderline cases
8. Dual Filtrations of MCM modules over hypersurface rings
9. A generalization of a result due to Ooishi
10. Application of our generalization of Ooishi’s result
11. CI approximation: The general case
12. Dual Filtrations mod a superficial sequence
13. Some Preliminaries to prove Itoh’s conjecture
14. Proof of Itoh’s-conjecture
References
2
6
8
10
13
15
17
19
23
26
29
32
34
35
41
Date: September 7, 2017.
1991 Mathematics Subject Classification. Primary 13A30, 13D45 ; Secondary 13H10, 13H15.
Key words and phrases. multiplicity, reduction, Hilbert polynomial, a-invariant .
1
2
TONY J. PUTHENPURAKAL
1. Introduction
Let (A, m) be a Noetherian local ring of dimension d, a an m-primary ideal in
A
and
let M be a Cohen-Macaulay (CM) module of dimension r. Set Ga (A) =
L
n
a
/an+1 ; the associated graded ring of A with respect to a and let Ga (M ) =
Ln≥0 n
n+1
M be the associated graded module of M with respect to a.
n≥0 a M/a
In this paper we give a two step method to study certain questions regarding Ga (M ) when A is complete. Surprisingly this is also useful in the case of
M = A. We call the first step Gorenstein approximation. This step enables us
to reduce to the case when A, Ga (A) are Gorenstein and M is a maximal CohenMacaulay (MCM) A-module. This reduction surprised a few and attracted some
skepticism. Nevertheless it is true. The second step consists in studying the filtration {Homa (M, an )}n∈Z of the dual HomA (M, A). This filtration is classical cf.
[28, p. 12]. However it has not been used before in the study of blow-up algebra’s
(modules) of Cohen-Macaulay rings (modules).
We begin by a general definition of our notion of approximation. Let P be a
property of Noetherian rings. For instance P = regular, complete intersection (CI),
Gorenstein, Cohen-Macaulay etc.
Definition 1. Let (A, m) be a Noetherian local ring and let a be a proper ideal in
A. We say A admits a P-approximation with respect to a if there exists a local ring
(B, n), an ideal b of B and ψ : B → A, a local ring homomorphism with ψ(b)A = a
such that the following three properties hold:
(1) A is a finitely generated B module ( via ψ).
(2) dim A = dim B.
(3) B and Gb (B) have property P.
If the above conditions hold we say [B, n, b, ψ] is a P-approximation of [A, m, a].
Remark 2. • Regular approximations seems to be rare while Cohen-Macaulay
approximations don’t seem to have many applications.
• For applications we will often insist that ψ, b satisfy some additional properties.
When A is complete we prove the following general result(see 11.3). Let s(a)
denote the analytic spread of a, see 2.8.
Theorem 3. Let (A, m) be a complete Noetherian local ring and let a be a proper
in A such that s(a) + dim A/a = dim A. Then A admits a CI-approximation with
respect to a.
Notice the hypothesis of Theorem 3 is satisfied if a is m-primary. Another significant case when Theorem 3 holds is when A is equidimensional and a is equimultiple;
see [6, 2.6]. For definition of an equimultiple ideal see 2.8.
Remark 4. • All the examples of rings and ideals in this paper have CI (and
so Gorenstein) approximation. However I have only used the Gorenstein
property of B, Gb (B) for applications. I have been unable to use the fact that
B, Gb (B) are CI and not just Gorenstein.
• However the author believes that in the case of ideals having non-zero analytic
deviation (see 2.8), it may not be possible to get CI-approximation. We might have
to settle for only Gorenstein approximation.
The proof of Theorem 3 is fairly involved. The entire section 11 is devoted for a
proof of it. The following two special cases are easy to prove.
GORENSTEIN APPROXIMATION
3
(1) Let (A, m) be a quotient of a regular local ring. Then [A, m, m] admits a CI
approximation [B, n, n, φ] with φ-onto. (see 3.5).
(2) Let (A, m) be a complete equicharacteristic local ring and let a be an mprimary ideal. Then [A, m, a] admits a CI approximation [B, n, b, φ] with
b = n and B/n ∼
= A/m and µ(n) = µ(a). (see 3.6).
Before proceeding
further we need some notation. Set an = A for n ≤ 0. Let
L
n n
respect to a. We
Ra (A) =
n∈Z a t be the extended Rees-algebra of A with
L
n
n
consider Ra (A) as a subring of A[t, t−1 ]. Set R(a, M ) =
n∈Z a M t Clearly
R(a, M ) is a finitely generated Ra (A)-module respectively.
Definition 5. By the dual filtration of M ∗ = HomA (M, A) with respect to a we
mean the filtration {HomA (M, an )}n∈Z .
We study the dual filtration in detail when (A, m) is Gorenstein, a is m-primary,
M is MCM and Ga (A) is Gorenstein. Let F = {HomA (M, an )}n∈Z be the dual
filtration of M ∗ with respect to a. We prove that if Ga (M ) is CM then
G(F, M ∗ ) ∼
=
*
HomGa (A) (Ga (M ), Ga (A));
see 4.8.
Even when Ga (M ) is not Cohen-Macaulay we find the following relation: Set
reda (A) = reduction number of A with respect to a (see 2.10), a(Ga (M )) = the
a-invariant of Ga (M ) (see 2.9), and let
αa (M ) = min{n | G(F, M ∗ )n 6= 0}.
We prove
a(Ga (M )) ≥ reda (A) − αa (M ) − dim A
(see 6.6).
Applications
I. a-invariant of associated graded rings of m-primary ideals
Let a be an m-primary ideal in a Noetherian local ring (A, m). It is easy to see
that a(Ga (A)) ≥ − dim A. We analyze the case when equality holds when A is
Cohen-Macaulay. If dim A = 1, it follows immediately from a result of T.Marley
** that a is prinicipal. When d = 2, it follows from Lemma ** and a result of
Northcott that a is also generated by regular sequence of length two.
We cannot use induction to analyze the case when a(Ga (A)) = −d sinice if
x is A-superficial with respect to A then it does not immediately follow that
a(Ga/(x) (A/(x)) = −d + 1
We prove that if A is Cohen-Macaulay then
a(Gm (A)) = − dim A if and only if
A is regular (see 7.2).
See 7.3 for an example of a non-Cohen-Macaulay ring A with a(Gm (A)) = − dim A.
For arbitrary m-primary ideals in a Cohen-Macaulay local ring we unfortunately
have to assume that A is equicharacteristic. In this case we prove that
a(Ga (A)) = − dim A ⇐⇒ a is generated by a regular sequence (see 7.4).
Next we try to classify ideals such that a(Ga (A)) = − dim A + 1. When A is an
equicharacteristic Cohen-Macaulay local ring and a is integrally closed we prove
a(Ga (A)) = − dim A + 1 ⇐⇒ a has minimal multiplicity (see 7.5).
The result above need not hold if a is not integrally closed (see 7.7). When dim A =
2 we have been able to characterize m-primary ideals a (not necessarily integrally
closed) with a(Ga (A)) ≤ −1.
4
TONY J. PUTHENPURAKAL
II. Symmetric
L h-vectors.
Let R = n≥0 Rn be a standard graded algebra over a field. For definition of
hR (z); the h-polynomial of R see 2.11. If R is Gorenstein then hR (z) is symmetric
cf. [2, 4.4.6]. The converse need not hold even if R is Cohen-Macaulay cf. [2,
4.4.7(b)]. A celebrated theorem due to Stanley asserts that if R is Cohen-Macaulay
domain and hR (z) is symmetric then R is Gorenstein ([30, 4.4]).
If R = Ga (A) the associated graded ring of an m-primary ideal then Ooishi,
[20, 1.6], proves that if A is Gorenstein, Ga (A) is Cohen-Macaulay and ha (z) is
symmetric then Ga (A) is Gorenstein. We generalize Ooishi’s result as follows:
Let (A, m) be CM with a canonical module ωA . Let a be an m-primary ideal
and let M be a CM A-module. Set M † = HomA (M, ωA ), r = red(a, M ) and
c(M ) = dim A − dim M . Assume Ga (A), Ga (M ) and Ga (M † ) are CM. Let Ωa be
the canonical module of Ga (A). We prove (see Theorem 9.6)
ha (M † , z) = z r h(M, z −1 ) ⇐⇒ Ga (M † ) ∼
=
*
c
ExtGa (A) (Ga (M ), Ωa ) (up to a shift).
The implication (⇐) follows from [2, 4.4.5]. The assertion ⇒ is new and generalizes
Ooishi’s result.
In section 10 we use of a MCM module M to show that the dual filtration of
a MCM module M with respect to m, over a Gorenstein local ring A with Gm (A)
Gorenstein, is a shift of the usual m-adic filtration in the following cases.
(1) M is Ulrich (i.e., deg hM (z) = 0; equivalently e(M ) = µ(M )).
(2) type(M ) = e(M ) − µ(M ) and M has minimal multiplicity.
III. m-primary ideals a with µ(a) = dim A + 1
Using Gorenstein approximation of [A, m, a] when A is complete and the structure
of dual filtration of MCM modules over a hypersurface ring we prove (see 8.17):
Suppose (A, m) is an equicharacteristic Gorenstein local ring of dimension d ≥ 0
and a is an m-primary ideal of A with µ(a) = d + 1. Then
Ga (A) is Gorenstein if and only if ha (z) = ℓ(A/a)(1 + z + . . . + z s ) for some s ≥ 1.
Here ℓ(−) denotes length. The surprising thing is: ha (z) determines that Ga (A) is
Gorenstein without a priori assuming Ga (A) is CM.
IV. Associated graded modules of the canonical module
Let (A, m) be CM local ring with a canonical module ωA . Let a be an m-primary
ideal such that Ga (A) is CM. Let Ωa be the canonical module of Ga (A). A natural
question is when is Ga (ωA ) isomorphic to Ωa up to a shift? Set r = red(a, A).
Ooishi proves that if Gm (A) is CM then Gm (ωA ) ∼
= Ωm (up to a shift) if and only if
Gm (ωA ) is CM and h(ωA , z) = z r h(A, z −1 ); see [19, 3.5].
The hypothesis Ga (ωA ) is CM is difficult to verify. So there is a need to bypass
this assumption. We first consider the case when µ(a) = d + 1. We prove
Ga (ωA ) ∼
= ΩA (up to a shift) ⇐⇒ ha (A, z) = ℓ(A/a) 1 + z + . . . + z r ; see 8.16 .
a
The assumption Ga (ωA ) is CM holds automatically if dim A = 0. Surprisingly
the following general result holds: Let x, . . . , xd be a maximal A-superficial sequence
with respect to a. Set B = A/(x) and b = a/(x). We prove (see 10.6).
If h(ωB , z) = z r h(B, z −1 ) then Ga (ωA ) ∼
= Ωa up to a shift.
This result is quite practical. Using COCOA [3] & MACAULAY [5], we can determine whether Ga (ωA ) ∼
= Ωa up to a shift. See Example 10.7. Another application
of 10.6 (see 10.8) is: If red(A) = 2 then
GORENSTEIN APPROXIMATION
5
Gm (ωA ) ∼
= Ωm (up to a shift) if and only if type(A) = e0 (A) − h1 (A) − 1.
Ooishi derives the same result essentially assuming Gm (ωA ) is CM, see [19, 4.4].
V. Lengths of duals of associated graded modules over Gorenstein Artin local rings.
Let (A, m) be a Gorenstein Artin local ring, a an ideal in A and let M be a
finitely generated A-module. Set G = Ga (A) and G(M ) = Ga (M ). We prove
ℓG (G(M )) ≤ ℓG * HomG (G(M ), G)
This is surprising since Ga (A) need not be Gorenstein.
VI. Generalization of some results from CM rings to CM modules.
Let (A, m) be CM local ring of dimension d with infinite residue
P field and let a be
an m-primary ideal. Let c be a minimal reduction of a. Set δ = n≥0 ℓ(an+1 /can ).
Then for 0 ≤ δ ≤ 2 one has depth Ga (A) ≥ d − δ. This is due to Valabrega and
Valla [32, 2.3] (for δ = 0), Guerrieri [7, 3.2] (for δ = 1) and Wang [33, 2.6,3.1] (for
δ = 2).
We extend these results to CM modules. Using Gorenstein approximation we
prove it by reducing it to the case of rings, see 5.1.
VII. Relation between ea1 (M ), ea0 (M ) and a(Ga (M )).
Let M be a CM A module of dimension = 2. Let a be an m-primary ideal. For
i ≥ 0 let eai (M ) be the Hilbert-coefficients of M with respect to a. It can be easily
proved that
(1.0.1)
ea1 (M ) ≤ ea0 (A)(a(Ga (A)) + 2)
It is natural to investigate when equality holds occurs in 1.0.1. We prove that
equality holds in 1.0.1 if and only if Ga (M ) is generalized Cohen-Macaulay and
H 2 (Ga (A))∨ is generalized Ulrich Ga (A)-module.
VIII. It is well-known that the multiplicity of a Noetherian local ring A is positive.
When A is Cohen-Macaulay, Northcott [18] proved that ea1 (A) ≥ 0 with equality if
and only if a is generated by a regular sequence. Narita showed that ea2 (A) ≥ 0, see
[17]. Furthermore he gave an example which showed that ea3 (A) can be negative.
Recall an ideal a is said to be normal if an is integrally closed for all n ≥ 1.
Itoh conjectured that if a is normal then ea3 (A) ≥ 0. Furthermore he conjectured
that if eA
3 (A) = 0 (and a is normal) then Ga (A) is Cohen-Macaulay. A major
consequence of our theory of Gorenstein approximation, Dual filtration is a proof
of Itoh’s conjecture.
Here is an overview of the contents of this paper. In section two we introduce
notation and discuss a few preliminary facts that we need. In section 3 we prove the
result on Gorenstein approximations when a = m and A is a quotient of a regular
local ring. We also prove it when A is complete, a is m-primary and contains a
field. In section 4 we introduce dual filtrations. Furthermore application V is
proved in this section. For applications VI and VII see section 5. In section six
we discuss on the initial degree of the dual filtration. We also relate it to the ainvariant of Ga (M ). For application I see section 7. Application III is discussed in
section 8. In section 9 we deal with symmetric h-vectors and prove generalization
of Ooishi’s result (i.e., application II). In section 10 we use application II to get
application IV. In section 11 we prove our result on Gorenstein approximation of
equimultiple ideals. In section twelve we discuss behaviour of dual filtrations with
respect to superficial elements. In section thirteen we discuss a few preliminaries
6
TONY J. PUTHENPURAKAL
that are needed to prove Itoh’s theorem. Finally in section fourteen we prove Itoh’s
theorem.
2. Notation and Preliminaries
In this paper all rings are commutative Noetherian and all modules are assumed
to be finitely generated. Recall an a-filtration G = {Mn }n∈Z on M is a collection
of submodules of M with the properties
(1) Mn ⊃ Mn+1 for all n ∈ Z.
(2) Mn = M for all n ≪ 0.
(3) aMn ⊆ Mn+1 for all n ∈ Z.
If aMn = Mn+1 for all n ≫ 0 then we say G is a-stable. We usually set Gn = Mn
for all n ∈ Z.
2.1. Let G = {Mn }n∈Z be an a-filtration on M and let N be a submodule of M .
By the quotient filtration on M/N we mean the filtration G = {(Mn + N )/N }n∈Z .
If G is an a-stable filtration on M then G is an a-stable filtration on M/N . Usually
for us N = xM for some x = x1 , . . . , xs ∈ a.
L
2.2. If G = {Mn }n∈Z is an a stable filtration on M then set R(G, M ) = n∈Z Mn tn
the extended Rees-module of M with respect to G. Notice that R(F, M ) is a
finitely generated graded Ra (A)-module. If G is the usual a-adic filtration then
set R(a, M ) = R(G,L
M ).
Set G(G, M ) =
n∈Z Mn /Mn+1 , the associated graded module of M with respect to G. Notice G(G, M ) is a finitely generated graded module over Ga (A).
Furthermore R(G, M )/t−1 R(G, M ) = G(G, M ).
2.3. Let G = {Mn }n∈Z be an a-filtration on M and let s ∈ Z. By the s-th shift of
G, denoted by G(s) we mean the filtration {G(s)n }n∈Z where G(s)n = Gn+s . Clearly
G(G(s), M ) = G(G, M )(s)
and R(G(s), M ) = R(G, M )(s).
T
2.4. All filtration’s in this paper G = {Mn }n∈Z will be separated i.e., n∈Z Mn =
{0}. This is automatic if A is local, a 6= A and G is a-stable. If m is a non-zero
element of M and if j is the largest integer such that m ∈ Mj , then we let m∗G
denote the image of m in Mj \ Mj+1 and we call it the initial form of m with
respect to G. If G is clear from the context then we drop the subscript G.
The following result is well-known and easy to prove.
Proposition 2.5. Let (A, m) be a local ring, a an m-primary ideal and let M be
an A-module. Let F be an a-stable filtration of M with respect to a. The following
conditions are equivalent:
(i) G(F, M ) ∼
= G(M ) up to a shift.
(ii) F is the a-adic filtration up to a shift.
2.6. For definition and basic properties of superficial sequences see [21, p. 86-87].
2.7. Flat Base Change: In our paper we do many flat changes of rings. The
general set up we consider is as follows:
Let φ : (A, m) → (A′ , m′ ) be a flat local ring homomorphism with mA′ = m′ . Set
a′ = aA′ and if N is an A-module set N ′ = N ⊗ A′ . Set k = A/m and k ′ = A′ /m′ .
Properties preserved during our flat base-changes:
(1) ℓA (N ) = ℓA′ (N ′ ). So Ha (M, n) = Ha′ (M ′ , n) for all n ≥ 0.
GORENSTEIN APPROXIMATION
7
(2) dim M = dim M ′ and grade(K, M ) = grade(KA′ , M ′ ) for any ideal K of A.
(3) depth Ga (M ) = depth Ga′ (M ′ ).
A′ ′ ′
′ ∼
(4) For each i ≥ 0 we have TorA
i (k, k) ⊗A A = Tori (k , k ). In particular A is
′
regular local if and only if A is.
Specific flat Base-changes:
(a) A′ = A[X]S where S = A[X] \ mA[X]. The maximal ideal of A′ is n = mA′ .
The residue field of A′ is k ′ = k(X). Notice that k ′ is infinite.
b the completion of A with respect to the maximal ideal.
(b) A′ = A
(c) Applying (a) and (b) successively ensures that A′ is complete with k ′ infinite.
(d) A′ = A[X1 , . . . , Xn ]S where S = A[X1 , . . . , Xn ]\mA[X1 , . . . , Xn ]. The maximal
ideal of A′ is n = mA′ . The residue field of A′ is l = k(X1 , . . . , Xn ). Notice
that if I is integrally closed then I ′ is also integrally closed. When dim A ≥ 2
and I is normal, Itoh [11, Theorem 1] shows that in A′ (for n = µ(I) in (ii))
there exists a superficial element y ∈ I ′ such that the A′ /(y) ideal J = I ′ /(y)
is asymptotically normal i.e., J n is integrally closed for all n ≫ 0. We call A′
general extension of A.
L
2.8. The fiber cone of a is the k-algebra F (a) = n≥0 an /man. Set s(a) = dim F (a),
the analytic spread of a. Recall dim A ≥ s(a) ≥ ht a. We say a is equimultiple if
s(a) = ht a. Clearly m-primary ideals are equimultiple. The number s(a) − ht a is
called the analytic deviation of a.
i
(Ga (M )) = 0 for i > r.
2.9. Assume a is m-primary and dim M = r. Then HG
a (A)+
Recall that the a-invariant of Ga (M ) is
r
(Ga (M ))n 6= 0}.
a (Ga (M )) = max{n | HG
a (A)+
i
(Ga (M ))n 6= 0}.
For i ≥ 0 set aai (M ) = max{n | HG
a (A)+
2.10. We assume k = A/m is infinite. Let c = (x1 , . . . , xl ) be a minimal reduction
of a with respect to M . We denote by redc (a, M ) := min{n | can M = an+1 M } the
reduction number of a with respect to c and M . Let
red(a, M ) = min{redc (a, M ) | c is a minimal reduction of a}
be the reduction number of M with respect to a. Set reda (A) = red(a, A).
L
2.11. Let R = n≥0 Rn be a standard algebra over an Artin local ring (R0 , m0 )
L
i.e., there exists x1 , . . . xs ∈ R1 such that R = R0 [x1 , . . . , xs ]. Let M = n∈Z Mn
be an R-module. By Hilbert-Serre theorem
X
hM (z)
here hM (z) ∈ Z[z, z −1 ] & hR (1) 6= 0.
H(M, z) =
ℓR0 (Mn )z n =
(1 − z)dim M
n∈Z
We call HM (z) the Hilbert series of M and
hM (z) = h−p z −p +. . .+h−1 z −1 +h0 +h1 z +. . .+hs z s
(with hs 6= 0 and h−p 6= 0)
the h-polynomial of M . Notice hM (z) ∈ Z[z] if and only if Mn = 0 for all n < 0.
2.12. We say hM (z) is symmetric if hs−i = h−p+i for i = 0, 1, . . . , s + p.
2.13. Let R, M , hM (z) be as in 2.11. We also assume that Mn = 0 for n < 0. If f
is a polynomial we use f (i) to denote the i-th formal derivative of f . The integers
(i)
ei (M ) = hM /i! for all i ≥ 0 are called the Hilbert coefficients of M . The number
e0 (M ) is also called the multiplicity of M .
8
TONY J. PUTHENPURAKAL
2.14. Let (A, m) be a local ring, a an m-primary ideal and let M be an A-module.
Set Ha (M, z) = H(Ga (M )(z), z) and ha (M, z) = hGa (M) (z). For i ≥ 0 set eai (M ) =
ei (Ga (M )). If a = m we usually drop the ”label” m and write h(M, z) = hm (A, z),
ei (M ) = em
i (M ) and red(A) = red(m, A).
3. CI-Approximation: Some special cases
In this section we prove that if (A, m) is a quotient of a regular local ring then
[A, m, m] has a CI-approximation [B, n, n, φ] with φ onto. In Theorem 3.3 we state
our main result regarding CI-approximations. This will be proved in section 10.
Finally in Theorem 3.6 we prove a result regarding CI-approximation of an mprimary ideal a in a complete equicharacteristic local ring (A, m). The following
Lemma regarding annihilators is crucial.
Lemma 3.1. Let M be an A-module and let a be an ideal. Set G = Ga (A). Suppose
there exists ξ1 , . . . , ξs ∈ a such that ξi∗ ∈ annG Ga (M ) for each i. Also assume that
ξ1∗ , . . . , ξs∗ is a G-regular sequence. Then there exists u1 , . . . , us ∈ a such that
(1) u1 , . . . , us is an A-regular sequence.
(2) ui ∈ ann M for each i = 1, . . . , s.
(3) For each i ∈ {1, . . . , s} there exists ni ≥ 1 such that u∗i = (ξi∗ )ni .
(4) u∗1 , . . . , u∗s ∈ annG Ga (M ).
(5) u∗1 , . . . , u∗s is a G-regular sequence.
Proof. Suppose we have constructed u1 , . . . , us satisfying (2) and (3) then u1 , . . . , us
satisfy all the remaining properties. This can be seen as follows:
(4) follows from (3); since ξi∗ ∈ annG Ga (M ) for each i.
(1) and (5). As ξ1∗ , . . . , ξs∗ is a G-regular sequence we also get (ξ1∗ )n1 , . . . , (ξs∗ )ns is
a G-regular sequence [14, 16.1]. Thus u∗1 , . . . , u∗s is a G-regular sequence. It follows
from [32, 2.4] that u1 , . . . , us is an A-regular sequence. Thus it suffices to show
there exists u1 , . . . , us satisfying (2) and (3).
Fix i ∈ {1, . . . , s}. Set ξ = ξi . Say ξ ∈ ar \ ar+1 for some r ≥ 1. Since
∗
ξ ∈ annG Ga (M ) we have ξM ⊆ ar+1 M . Set q = ar+1 . By the determinant trick,
[14, 2.1], there exists a monic polynomial f (X) ∈ A[X] such that
f (X) = X n + a1 X n−1 + . . . + an−1 X + an
with
ai ∈ qi , for i = 1, . . . n
and u = f (ξ) = ξ n + a1 ξ n−1 + . . . + an−1 ξ + an ∈ ann M.
As ξ ∗ is G-regular we have ξ n ∈ anr \ anr+1 . However for each i ≥ 1 we have
ai ξ n−i ∈ qi a(n−i)r = ai(r+1) anr−ir = anr+i ⊆ anr+1 .
Thus u∗ = (ξ n )∗ = (ξ ∗ )n (since ξ ∗ is G-regular). Set ui = u and ni = n.
The following Corollary is useful.
Corollary 3.2. Let (A, m) be Cohen-Macaulay local ring, a an m-primary ideal
and let M be an A-module. Set c = dim A − dim M . If Ga (A) is Cohen-Macaulay
then there exists u1 , . . . , uc ∈ a such that
(1) u1 , . . . , uc ∈ annA M .
(2) u∗1 , . . . , u∗c ∈ annGa (A) Ga (M ).
(3) u1 , . . . , uc is an A-regular sequence.
(4) u∗1 , . . . , u∗c is a Ga (A)-regular sequence.
(5) u1 t, . . . , uc t ∈ annRa (A) R(M ).
GORENSTEIN APPROXIMATION
9
(6) u1 t, . . . , uc t is a Ra (A)-regular sequence.
Proof. Since Ga (A) is Cohen-Macaulay, we have that
grade annGa (A) Ga (M ) = ht annGa (A) Ga (M )
= dim Ga (A) − dim Ga (M ),
since Ga (A) is *-local
= dim A − dim M.
Therefore (1), (2), (3) and (4) follow from Lemma 3.1. The assertion (5) is clear.
(6). Since t−1 is Ra (A)-regular and u∗1 , . . . , u∗c is Ga (A) = Ra (A)/t−1 Ra (A)regular sequence it follows that t−1 , u1 t, . . . , uc t is a Ra (A)-regular sequence. Notice
Ra (A) is a *-local. It follows that u1 t, . . . , uc t is a Ra (A)-regular sequence.
We state our general result regarding CI-Approximations
Theorem 3.3. Let (A, m) be a complete local ring and let a be a proper ideal in A
with dim A/a + s(a) = dim A. Then A has a CI-approximation with respect to a.
Furthermore if a = (x1 , . . . , xm ) then we may choose a CI-approximation [B, n, b, ϕ]
of [A, m, a] such that there exists u1 , . . . , um ∈ b with ϕ(ui ) = xi for i = 1, . . . , m.
Theorem 3.3 is proved in section 10.
Remark 3.4. The hypothesis of the theorem holds when a is m-primary. It is also
satisfied when a is equimultiple and A is equidimensional [16, 34.5]. Our hypothesis
ensures that Ga (A) has a homogeneous system of parameters, [6, 2.6].
For the case when a = m and A is a quotient of a regular local ring T we have:
Theorem 3.5. Let (A, m) be a local ring such that A is a quotient of a regular local
ring (T, t). Then [A, m, m] has a CI-approximation [B, n, n, φ] with φ-onto.
Proof. Set k = A/m = T /t. Let ψ : T → A be the quotient map. We consider A as
a T -module. So Ga (A) = Gt (A). Set c = dim T − dim A = dim Gt (T ) − dim Ga (A).
Its well-known Gt (T ) ∼
= k[X1 , . . . , Xn ]; where n = dim T . In particular Gt (T ) is
Cohen-Macaulay. Let u1 , . . . uc be as in Corollary 3.2.
Set (B, n) = (T /(u), t/(u)) and let φ : B → A be the map induced by ψ. So φ is
onto. Clearly [B, n, n, φ] is a CI-approximation of [A, m, m].
For m-primary ideals in a complete equicharacteristic complete local ring we prove:
Theorem 3.6. Let (A, m) be a complete equicharacteristic local ring and let a be
an m-primary ideal. Then [A, m, a] has a CI-approximation [B, n, b, φ] with b = n
and B/n ∼
= A/m and µ(n) = µ(a).
Proof. A contains its residue field k = A/m.; see [14, 28.3]. Let a = (x1 , . . . , xn ).
Set S = k[|X1 , . . . Xn |] and let ψ : S → A be the natural map which sends Xi to xi
for each i. Since A/(X)A = A/a has finite length (as an S-module) it follows that
A is a finitely generated as an S-module, see [14, 8.4]. Set η = (X1 , . . . Xn ). Notice
ψ(η) = a and Gη (S) = k[X1∗ , . . . Xn∗ ] the polynomial ring in n-variables.
If dim A = dim S then set [B, n, b, φ] = [S, η, η, ψ]. Otherwise set c = dim S −
dim A = dim Gn (S) − dim Ga (A). Let u1 , . . . uc be as in Corollary 3.2. Set B =
S/(u), n = η/(u) and b = n. Clearly Gn (B) = Gη (S)/(u∗ ) is CI. The map ψ
induces φ : B → A. It can be easily checked that [B, n, n, φ] is a CI-approximation
of [A, m, a]. Finally by our construction its clear that µ(n) = µ(a).
10
TONY J. PUTHENPURAKAL
4. A classical filtration of the dual
Let M be an A-module. The following filtration of the dual of M , i.e, M ∗ =
HomA (M, A) is classical cf. [28, p. 12]. Set
Mn∗ = {f ∈ M ∗ | f (M ) ⊆ an } ∼
= HomA (M, an ).
It is easily verified
that FM = {Mn∗ }n∈Z is an L
a-stable filtration on M ∗ . Set
L
n n
R = R(a) = n∈Z a t and R(M ) = R(a, M ) = n∈Z an M tn . Let f ∈ Mn∗ . We
show that f induces a homogeneous R-linear map fˆ of degree n from R(M ) to R;
(see 4.2). Theorem 4.4 shows that ΨM is a R-linear isomorphism.
ΨM : R(F, M ∗ ) −→ * HomR (R(M ), R)
f 7→ fˆ
It is easily seen, see 4.5, that ΨM induces a natural map
ΦM : G(F, M ∗ ) −→
*
HomGa (A) (Ga (M ), Ga (A)).
We prove ΦM is injective and as a consequence deduce application V as stated in
the introduction. In Corollary 4.8 we give a sufficient condition for ΦM to be an
isomorphism.
The following result is well-known.
Proposition 4.1. Let (A, m) be local, a an ideal in A and let M an A-module.
Then a HomA (M, an ) = HomA (M, an+1 ) for all n ≫ 0.
4.2. Let f ∈ Mn∗ .
Claim: f induces a homogeneous R-linear map fˆ of degree n from R(M ) to R.
Since f (M ) ⊆ an , for each j ≥ 0 we have f (aj M ) ⊆ an+j . Furthermore for j < 0
as aj M = M and notice that f (M ) ⊂ an ⊂ an+j . Thus
f (aj M ) ⊆ an+j
(∗)
for each j ∈ Z.
This enables us to define
fˆ: R(M ) −→ R
X
X
mj tj 7→
f (mj )tn+j .
j∈Z
j∈Z
Next we prove that fˆ is R-linear. Clearly fˆ is A-linear. Notice
fˆ (xj tj ) • (mi ti ) = fˆ(xj mi ti+j ) = f (xj mi )ti+j+n = xj tj f (mi )ti+n
(xj tj ) • fˆ(mi ti ) = xj tj f (mi )ti+n .
Thus fˆ is R-linear.
4.3. We define a map
ΨM : R(F, M ∗ ) −→
*
HomR (R(M ), R)
f 7→ fˆ
Theorem 4.4. ΨM : R(F, M ∗ ) −→∗ HomR (R(M ), R) is a R-linear isomorphism.
GORENSTEIN APPROXIMATION
11
∗
ˆ . Thus
Proof. If a ∈ aj and f ∈ Mi∗ then af ∈ Mi+j
. Check that (ati ) • fˆ = af
ΨM is R-linear. Clearly if fˆ = 0 then f = 0. Thus ΨM is injective.
We show ΨM is surjective. Let g ∈ * HomR (R(M ), R)n . We write g as
X
X
g
gj (mj )tn+j .
m j tj =
j∈Z
j∈Z
For each j ∈ Z the map gj : aj M → an+j is A-linear. Define f = i ◦ g0 where
i : an → A is the inclusion map. Clearly f ∈ Mn∗ .
Claim: fˆ = g.
Let mj tj ∈ aj M tj .
Case 1. j = 0.
Set m = m0 . Notice
g(mt0 ) = g0 (m)tn = fˆ(mt0 ).
The last equality above holds since fˆ is R-linear.
In the next two cases we use Case 1 and the fact that fˆ is R-linear.
Case 2. j < 0.
Notice mj tj = tj • mj t0 . So
g(mj tj ) = tj • g(mj t0 ) = tj • fˆ(mj t0 ) = fˆ(mj tj ).
Case 3. j >
Ps0.
Set mj = l=1 ujl nl where ujl ∈ aj and nl ∈ M . Fix l. Notice
ujl nl tj = ujl tj • nl t0 .
Therefore for l = 1, . . . , s,
g(ujl nl tj ) = ujl tj • g(nl t0 ) = ujl tj • fˆ(nl t0 ) = fˆ(ujl nl tj ).
Note the last equality above is since fˆ is R-linear. So we have
!
s
s
s
X
X
X
g(mj tj ) = g
g(ujl nl ) =
fˆ(ujl nl ) = fˆ(mj tj ).
ujl nl =
l=1
l=1
l=1
Again the last equality above holds since fˆ is R-linear.
Observation 4.5. f ∈ Mn∗ induces fˆ: R(M ) → R which is homogeneous of degree
n. So fˆ induces a map f˜: Ga (M ) → Ga (A) which is also homogeneous of degree
∗
n. Clearly f˜ = 0 if and only if f ∈ Mn+1
. So we have
ΦM : G(F, M ∗ ) −→ * HomGa (A) (Ga (M ), Ga (A))
f + M ∗ 7→ f˜.
n+1
Clearly ΦM = ΨM ⊗
R
.
t−1 R
Corollary 4.6. Set G = Ga (A). There is an exact sequence
Φ
0 −→ G(F, M ∗ ) −−M→
*
HomG (Ga (M ), G) −→
*
1
ExtR (R(M ), R)(+1)
12
TONY J. PUTHENPURAKAL
t−1
Proof. The map 0 → R(+1) −−→ R → G → 0 induces the exact sequence
0 −→
−→
*
t−1
HomR (R(M ), R)(+1) −−→
*
HomR (R(M ), R) −→
*
HomR (R(M ), G)
1
*
ExtR (R(M ), R)(+1)
Note that * HomR (R(M ), G) ∼
= * HomG (Ga (M ), G). The result follows by using
Theorem 4.4 and Observation 4.5.
A consequence of Corollary 4.6 is Application V.
Corollary 4.7. Let (A, m) be a Artin Gorenstein local ring, a a proper ideal in A
and let M be a finitely generated A-module. Then
ℓ(Ga (M )) ≤ ℓ * HomGa (A) (Ga (M ), Ga (A)) .
Proof. Set G = Ga (A) and G(M ) = Ga (M ). Notice that
ℓ(Ga (M )) = e0 (Ga (M )) = ℓ(M )
ℓ(GF (M ∗ )) = e0 (Ga (M ∗ )) = ℓ(M ∗ ).
Since A is self-injective, ℓ(M ) = ℓ(M ∗ ). The result follows from 4.6.
Another important consequence of Corollary 4.6 is the following:
Corollary 4.8. Let (A, m) be a Gorenstein local ring and let a be an ideal such
that Ga (A) is Gorenstein. Let M be a maximal Cohen-Macaulay A-module with
Ga (M )-Cohen-Macaulay. Then
G(F, M ∗ ) ∼
=
*
HomGa (A) (Ga (M ), Ga (A)).
Proof. Notice R is also a Gorenstein ring and R(M ) is maximal Cohen-Macaulay.
1
It follows that * ExtR (R(M ), R) = 0. The result follows from Corollary 4.6.
4.9. In the theory of Hilbert functions over local rings the notion of superficial
elements plays an important role. In our case, first we assume for convenience that
(A,L
m) is also Gorenstein and M is a M CM A-module. Let x = x1 , . . . , xr be a
M
A superficial sequence. Set B = A/(x), b = a/(X), N = M/xM and
FM = {HomA (M, an )}n∈Z
and FN = {HomB (M, nn )}n∈Z .
We may ask when does
(4.9.1)
G(FM , M ∗ ) ∼
= G(FN , N ∗ ).
x∗ G(FM , M ∗ )
We prove
Theorem 4.10. [with hypothesis as in 4.9] If Ga (A) is Gorenstein and Ga (M ) is
Cohen-Macaulay then 4.9.1 holds.
See Theorem 12.5(2) for a proof of Theorem 4.10.
GORENSTEIN APPROXIMATION
13
5. Some preliminary applications of Gorenstein approximation
In this section we give give Applications VI and VII as stated in the introduction. First we generalize to modules some results of Valabrega and Valla[32, 2.3],
Guerrieri [7, 3.2] and Wang [33, 2.6,3.1]. If dim M = 2 and a an ideal of definition
for M , then it is easy to prove ea1 (M ) ≤ ea0 (M )(a2 (Ga (M ))+2). We give a complete
classification of when equality holds.
Theorem 5.1. [Appl. VI ] Let (A, m) be a local ring with infinite residue field,
M a Cohen-Macaulay A-module of dimension r and a an ideal of definition for M.
Let c = (x1 , . . . xr ) be a minimal reduction of a with respect to M . Set
n+1
∞
X
a
M ∩ cM
δ=
ℓ
.
can M
n=0
If δ = 0, 1, 2 then depth Ga (M ) ≥ d − δ.
Proof. We may go mod ann M. Thus we may assume dim M = dim A and a is
m-primary. Since M is a faithful A-module it can be easily checked that c is a
minimal reduction of a with respect to A. Our hypothesis nor conclusion change
under completion so we may assume A is complete.
Let [B, n, b, ϕ] be a Gorenstein approximation of [A, m, a]. We do base change
and consider M as a B-module. Notice Gb (M ) = Ga (M ). Let yi ∈ b be such
that ϕ(yi ) = xi . Set R = B ∝ M , the idealization of M and consider the ideal
q = b ∝ M . Then Gq (R) = Gb (B) ∝ Ga (M )(−1).
As Gb (B) is Cohen-Macaulay, depth Gq (R) = depth Ga (M ). Set zi = (yi , 0) for
i = 1, . . . , r. Notice qi = bi ∝ ai−1 M for i ≥ 0.
Therefore qi ∩ (z) = (bi ∩ (y), ai−1 M ∩ cM )
ai−1 M ∩ cM
qi ∩ (z)
.
=
0,
qi−1 (z)
aci−2 M
The result now follow from the case of rings.
and
The previous result did not use the fact the Gorenstein property of Gb (B). All
we used was that Gb (B) is Cohen-Macaulay. The next result uses the fact that
Gb (B) is Gorenstein. However we need the following elementary result .
L
Proposition 5.2. Let R =
Rn be a standard graded Noetherian ring with R0
Artin local. Let E be a graded R-module of dimension two. Then
e1 (E) ≤ e0 (E)(a2 (E) + 2).
0
Sketch. We may assume that HR
(E) = 0. Set t = a2 + 1. By the Grothendieck+
Serre formula
HE (t) − {e0 (t + 1) − e1 } = −ℓ(H 1 (M )t )
Hence e1 (E) ≤ e0 (E)(a2 + 2)
It is important to understand when equality holds above. In general we cannot
say much. However if M is Cohen-Macaulay of dimension two over a local ring
(A, m) then we can give a characterization when equality holds above. Before
proceeding we need the following notation: If D is a graded Artin Ga (A)-module
then D∨ denotes the dual of D with respect to to the injective hull of k as a
Ga (A)-module (note that if a is m-primary then Ga (A) is *-complete (see BH**).
14
TONY J. PUTHENPURAKAL
Furthermore a finitely generated C over Ga (A) is said to generalized Ulrich if the
h-polynomial of C is of the form hC (z) = az l for some l ∈ Z.
We give application VII of our notion of Gorenstein approximation.
Theorem 5.3. Let (A, m) be a local ring, a an m-primary ideal in A and M an
A-module with dim M = 2. Then the following are equivalent:
(i) ea1 (M ) = ea0 (M ) (aa2 (M ) + 2) .
(ii) Ga (M ) is generalized Cohen-Macaulay and H 2 (Ga (A))∨ is Generalized Ulrich
Ga (A)-module.
Proof. We may assume A is complete, a is m-primary and dim M = dim A. Let
[B, n, b, ϕ] be a Gorenstein approximation of [A, m, a]. Notice
abi (M ) = aai (M ) and ebi (M ) = eai (M ) for all i.
Thus we may assume (A, m) is Gorenstein local and Ga (A) is Gorenstein. We may
further assume A has an infinite residue field. This we do. Set
M
2
(G
(M
))
=
G = Ga (A), D = HG
Dn where α = aa2 (M ).
a
+
n≤α
2
Set (−)∨ = * HomG (−, HG
(G)) the Matlis dual functor over G. Notice D∨ =
+
L
∨
∨
n≥−α Dn where Di = HomA/a (D−i , EA/a (k)), and EA/a (k) is the injective hull
of k as an A/a-module. By Matlis duality D∨ is a finitely generated G-module.
By Local duality we have
D∨ = H 2 (Ga (M ))∨ ≃
*
HomG (Ga (M ), G(a))
(Here a = reda (A) − 2, the a-invariant of G).
Claim 1: D∨ is a Cohen-Macaulay G-module of dimension 2.
Proof of Claim 1: By [29, 17.1.10] it follows that the function n 7→ ℓ(Dn∨ ) is
polynomial of degree 1. So dim D∨ = 2. Let F1 → F0 → Ga (M ) → 0 be a free
presentation of Ga (M ) as a G-module. Notice
0→
*
HomG (Ga (M ), G) →
*
HomG (F0 , G) →
*
HomG (F1 , G)
As G is Cohen-Macaulay of dim 2 we get Claim 1.
Claim 2: ℓ(Dn ) = C0 (α −
Ln + 1) − C∨1 for all n <<∨0 where C0 > 0 and C1 ≥ 0.
. Notice ℓ(Dm ) = ℓ(D−m ) for all m. Set
Proof of Claim 2: D∨ = m≥−α Dm
Notice D∨ (−α)j = 0 for j < 0. So D∨ (−α) has Hilbert series
h0 + h1 z + . . . αhs z z
.
(1 − z)2
As D∨ (−α) is Cohen-Macaulay we have hi ≥ 0 for all i. Thus there exists C0 > 0
and C1 ≥ 0 such that ℓ(D∨ (−α)n ) = C0 (n + 1) − C1 for all n ≫ 0. Therefore
∨
ℓ(Dn ) = ℓ(D−n
) = C0 (−n + α + 1) − C1
for all n ≪ 0.
This proves Claim 2.
By a result due to Serre cf. [2, 4.4.3]
H(Ga (M ), n) − P (Ga (M ), n) =
2
X
i=0
(−1)i ℓ(H i (Ga (M ))n ).
GORENSTEIN APPROXIMATION
15
Clearly H 0 (Ga (M )))n = 0 for n < 0. Notice ℓ(H 1 (Ga (M ))n ) = ξ, a constant for
all n ≪ 0, see [1, 17.1.9]. So we get
− [ea0 (M )(n + 1) − ea1 (M )] = C0 (α + 1 − n) − C1 − ξ
Therefore C0 = ea0 (M ). Comparing coefficients we get
ea1 (M ) = ea0 (M )(α + 2) − ξ − C1 .
The result follows.
6. The initial degree of the dual filtration and the a-invariant
Let (A, m) be a Gorenstein local ring, a an equimultiple ideal in A and let M be
an MCM A-module. Let αa (M ) be the order of M with respect to a (see 6.1). We
prove that αa (M ) ≤ reda (A). If a is m-primary then we prove that
a(Ga (M )) ≥ reda (A) − αa (M ) − dim A.
We also prove that αm (M ) = red(A) if and only if M is an Ulrich module.
6.1. Let (A, m) be a local ring, a an ideal in A and M and A-module. Set M ∗ =
HomA (M, A). Let F = {Mn∗ = HomA (M, an )}n∈Z be the dual filtration of M ∗ with
respect to a. If f 6= 0 ∈ M ∗ set αa (f ) = max{n | f (M ) ⊆ an }. Set
αa (M ) = min{αa (f ) | f ∈ M ∗ , f 6= 0}.
Notice αa (M ) = max{n | Mn∗ 6= M ∗ } = min{n | G(F, M ∗ )n 6= 0}.
6.2. Let x1 , . . . , xr ∈ a be a sequence. Set (B, n) = (A/(x), m/(x)) and N =
M/xM . If f ∈ HomA (M, A) define ψ(f ) : N → B, by
ψ(f )(m + xM ) = f (m) + xA.
One can check readily ψ(f ) ∈ HomB (N, B). Define
ΥX : HomA (M, A) → HomB (N, B)
ΥX (f ) = ψ(f ).
Proposition 6.3. Let (A, m) be a Gorenstein local ring and let M be a MCM
A-module. If x1 , . . . , xr ∈ a is an A-regular sequence and then ΥX is onto.
x
π
Proof. It suffices to prove for r = 1. The exact sequence 0 → A −
→A−
→B →0
x
→ M ∗ → HomB (N, B) → Ext1A (M, A). Notice Ext1A (M, A) = 0.
yields 0 → M ∗ −
The map:
Hom(M,π)
Hom(M, A) −−−−−−−→ HomB (N, B)
f 7→ f
is nothing but Υx1 . Thus Υx1 is onto.
For equimultiple ideals we have
Corollary 6.4. [with hypothesis as in 6.3] If a is an equimultiple ideal then
αa (M ) ≤ reda (A).
16
TONY J. PUTHENPURAKAL
Proof. Let J = (x1 , . . . , xs ) be a minimal reduction of a such that
r = reda (A) = min{n | an+1 = Jan }.
By hypothesis x1 , . . . , xs is a regular sequence. Consider ΥX : HomA (M, A) →
HomA/J (M/JM, A/JM ) as defined earlier.
If α(M ) > r then note that ψ(f ) = 0 for each f ∈ M ∗ . By 6.3, Υ is onto. So
HomA/J (M/JM, A/J) = 0, a contradiction.
6.5. Set r = red(a) and d = dim A. Recall that if Ga (A) is Gorenstein then
Ωa = G(r − d) is the canonical module of Ga (A). Using graded local duality we get
(6.5.1)
a(Ga (M )) = − min{n |
*
HomGa (A) (Ga (M ), Ga (A)(r − d))n 6= 0}.
The following result gives a lower bound on a(Ga (M ) in terms of αa (M ).
Proposition 6.6. Let (A, m) be a d-dimensional Gorenstein local ring, a an mprimary ideal with Ga (A) Gorenstein and let M be a MCM A-module.
a(Ga (M )) ≥ reda (A) − αa (M ) − d.
Proof. Set G(M ) = Ga (M ), G = Ga (A), r = reda (A) and M ∗ = HomA (M, A). By
4.6 we have an inclusion 0 → G(F, M ∗ ) → HomG (G(M ), G). So
(*)
0 → G(F, M ∗ )(r − d) → HomG (G(M ), G(r − d)).
Notice G(F, M ∗ )((r − d))αa (M)−(r−d) = G(F, M ∗ )αa (M) 6= 0.
Looking at initial degrees of sequence (∗) and using 6.5.1 we get
−a(Ga (M )) ≤ αa (M ) − (r − d).
The result follows.
Remark 6.7. If a is m-primary then by Proposition 6.6 and 6.4 it follows that
a(Ga (M )) ≥ − dim A. But this also follows easily from other arguments.
The case when a = m.
Notation: Set G(−) = Gm (−) and α(−) = αm (−). If we have to specify the ring
then we write αA (−). Clearly α(M ) = 0 if and only if M has a free summand.
Lemma 6.8. LetL(A, m) be a Gorenstein local ring with G(A) Gorenstein. Let
x ∈ m \ m2 be M
A-superficial with respect to m. Then
αA (M ) ≤ αA/(x) (M/xM ).
Proof. Set N = M/xM and (B, n) = (A/(x), m/(x)). By Proposition 6.3
Υ : HomA (M, A) → HomB (N, B)
is onto. Set α = αA (M ). Then f (M ) ⊆ mα for all f ∈ M ∗ . Therefore Υ(f )(N ) ⊆
nα for all f ∈ M ∗ . Since Υ is onto, we get g(N ) ⊆ nα for all g ∈ N ∗ . Thus
αB (N ) ≥ α = αA (M ).
6.9. Reduction to dimension zero :
Let x = x1 . . . xd be a maximal M ⊕ A-superficial sequence. Set (B, n) =
(A/(x), m/(x)) and N = M/xM .
Remark 6.10. Recall an MCM a-module M is called Ulrich if e(M ) = µ(M ).
(a) M is Ulrich ⇐⇒ N is Ulrich. Notice N is Ulrich ⇐⇒ N ≃ k µ(N ) .
(b) Since G(A) is Cohen-Macaulay we have red(A) = red(B)
GORENSTEIN APPROXIMATION
17
(c) If αA (M ) = red(A) then αB (N ) = red(B). This is so, since
red(A) = αA (M ) ≤ αB (N ) ≤ red(B)
(inequalities due to 6.8 & 6.4).
s
(d) If αB (N ) = red(B) then N = k for some s ≥ 1. The converse also holds since
B is Gorenstein.
Proposition 6.11. Let (A, m) be a Gorenstein local ring with Gm (A) Gorenstein.
Also assume that A has infinite residue field. Let M be a MCM A-module. The
following conditions are equivalent:
(i) α(M ) = red(A).
(ii) M is an Ulrich A-module.
(iii) a(G(M )) = − dim A.
Proof. Set r = red(A), α = αA (M ) and d = dim A. Let x = x1 , . . . , xd be a
maximal M ⊕A-superficial sequence. Set (B, n) = (A/(x), m/(x)) and N = M/xM.
(i) ⇒ (ii): If α(M ) = red(A) then by 6.10(c), α(N ) = red(B) = r. We have
α(N ) = r. So N ∗ = Nr∗ = HomB (N, nr ). Since B is Gorenstein with nr 6= 0 and
nj = 0 for j > r we get nr ∼
= k. So N ∗ = HomA (N, k) = k µ(N ) . Therefore
∗∗
µ(N
)
N∼
. So N is Ulrich. By 6.10(a) M is an Ulrich A-module.
=N =k
(ii) ⇒ (iii): Trivial.
(iii) ⇒ (i): By Proposition 6.6 we get
a(G(M )) ≥ r − α − d.
So we obtain α ≥ r. But α ≤ r always (see 6.4). So α = r.
7. a-invariant: Borderline cases
Let (A, m) be CM. It is well-known that a(Gm (A)) ≥ − dim A. We prove that
a(Gm (A)) = − dim A if and only if A is regular local (see 7.2). If A is equicharacteristic we prove
• a(Ga (A)) = − dim A if and only if a is generated by a regular sequence.
• (assume a is also integrally closed). a(Ga (A)) = − dim A + 1 if and only if
a has minimal multiplicity.
7.1. Discussion: We can easily show that if dim A = 1, 2, then a(Ga (A)) = − dim A
if and only if a is generated by a regular sequence. When d = 1 it easily follows from
a result of Marley **. When d = 2 it follows from 5.2 that ea1 (A) = 0. It follows from
a result of Northcott ** that a is generated by a regular sequence. We cannot use
induction to analyze the case when a(Ga (A)) = −d sinice if x is A-superficial with
respect to A then it does not immediately follow that a(Ga/(x) (A/(x)) = −d + 1
We begin our investigations when a = m the maximal ideal of A.
Theorem 7.2. Let (A, m) be a Cohen-Macaulay local ring. The following conditions are equivalent:
(i) A is regular local.
(ii) a(G(A)) = − dim A.
The example below shows that the assumption, A is CM in 7.2, is crucial.
Example 7.3. Let (A, m) be a regular local ring of dimension d > 0. Let M be
an A-module with dim M < d. Let R = A ∝ M , the idealization of M . Notice R
is local with maximal ideal n = m ∝ M . Also dim R = dim A = d. Furthermore
depth R = min{depth M, depth A} < d. So R is not CM.
18
TONY J. PUTHENPURAKAL
Claim : a(Gn (R)) = −d.
Set T = Gm (A), N = Gm (M )(−1) and S = Gn (R). Notice S = T ∝ N ,
the idealization of N . Since N 2 = 0 in S we have that HSi + (−) = HTi + S (−).
Furthermore as S is a finite T -module we get that HTi + (S) = HTi + S (S) as T modules. Since S = T ⊕ N as T -modules we get that HTi + (S) = HTi + (T ) ⊕ HTi + (N )
for each i ≥ 0. Since dim N = dim M < d we get that HTd+ (S) = HTd+ (T ). Thus
HSd+ (S) = HTd+ (T ) as T -modules. The result follows.
Proof of Theorem 7.2. The assertion (i) =⇒ (ii) is clear. To prove the converse we
may assume that A is complete with infinite residue field k; see 2.7(c) and 2.7(5).
Since A is complete, A is the quotient of a regular local ring. So by 3.5 [A, m, m]
has a Gorenstein approximation [B, n, n, φ] with φ onto. We consider A as a Bmodule. Notice Gn (A) = Gm (A) as a Gn (B)-module. It is also be easily seen that
i
i
(Gm (A)) as a Gn (B)-module.
(Gn (A)) ∼
for each i ≥ 0 we have HG
= HG
m (A)+
n (B)+
Since dim B = dim A = d(say), we get a(Gn (A)) = − dim B. By Theorem 6.11
we get that A is Ulrich as a B-module. Let J = (u1 , . . . , ud ) be a minimal reduction
of A with respect to n. Set xi = φ(ui ) for i = 1, . . . , d. As A is an Ulrich B-module
we get JA = nA. It follows that m = (x1 , . . . , xd ). So A is regular local.
We prove an analogue of 7.2 for m-primary ideals. Unfortunately we have to
assume that A is equicharacteristic (i.e., it contains a field).
Theorem 7.4. Let (A, m) be an equicharacteristic CM local ring of dim d ≥ 1. Let
a be an m-primary ideal. Then a(Ga (A)) = − dim A ⇐⇒ a is a parameter ideal.
Proof. We may assume, without loss of any generality that A is complete and
has an infinite residue field. As A is equicharacteristic we choose a Gorenstein
approximation [B, n, b, ϕ] of [A, m, a] with b = n. Notice Gn (A) = Ga (A) and
i
i
(Ga (A)). If a(Ga (A)) = −d then A is Ulrich as
(Ga (A)) = HG
notice HG
a (A)+
n (B)+
an B-module. So nA = JA where J = (y1 , . . . , yd ) is a minimal reduction of n with
respect to A. But nA = aA. Set xi = ϕ(yi ). We get a = (x1 , . . . , xd ). It follows
that a is a parameter ideal.
Conversely if a is a parameter ideal then its clear that a(Ga (A)) = −d.
We might also ask what happens when a(Ga (A)) = −d + 1. We show
Theorem 7.5. Let (A, m) be an equicharacteristic CM local ring of dim d ≥ 1. Let
a be an m-primary ideal. If a has minimal multiplicity then a(Ga (A)) = −d + 1.
The converse holds if
(1) a is integrally closed.
(2) dim A = 1.
Proof. If a has minimal multiplicity then Ga (A) is Cohen-Macaulay. The description of Hilbert series of Ga (A) gives a(Ga (A)) = −d + 1. To prove the converse we
may assume that A is complete with an infinite residue field.
Case (1): a is integrally closed.
As A is equicharacteristic we choose a Gorenstein approximation [B, n, b, ϕ] of
[A, m, a] with b = n. From Lemma 6.8 we have
αn (A) ≤ αn/(x) (A/xA)
L
where x = x1 , . . . , xd is a maximal A B superficial sequence with respect to n.
GORENSTEIN APPROXIMATION
19
By 6.6 we have αa (A) ≥ r − 1 where r = red B = red B/(x). Therefore
αn/(x) (A/xA) ≥ r − 1. It follows that n2 A = 0. Thus n2 A ⊆ xA. Let yi = ϕ(xi )
for all i. Set q = (y1 , . . . , yd ). Then we have a2 ⊆ J. Since a is integrally closed we
have a2 = qa. So a has minimal multiplicity.
Case (2): dim A = 1.
By hypothesis a1 (Ga (A)) = a(Ga (A)) = 0. If Ga (A) is CM then we have nothing to
show. We assert that Ga (A) is CM. If not then by a result of Marley [13, 2.1(a)] we
0
(Ga (A)) is concentrated
get a0 (Ga (A)) < a1 (Ga (A)) a contradiction, since HG
a (A)+
in non-negative degrees.
When dim A = 2 there exists m-primary ideals a which do not have minimal
multiplicity but have a(Ga (A)) = −1 (see 7.7). In fact when dim A = 2 we give the
following characterization of m-primary ideals a with a(Ga (A)) ≤ −1.
Proposition 7.6. Let (A, m) be a Cohen-Macaulay local ring with dim A = 2. Let
a be an m-primary ideal. The following conditions are equivalent:
(i) a(Ga (A)) ≤ −1.
(ii) redan (A) = 1 for all n ≫ 0.
(iii) ea2 (A) = 0.
Proof. By [10, 2.6] we get a(Ga (A)) < 0 if and only if red(an ) = dim A − 1 for all
n ≫ 0. Thus the assertions (i) and (ii) are equivalent. By [17], we get that (ii) and
(iii) are equivalent.
The example below was constructed by Marley [12, 4.1] for a different purpose.
Example 7.7. Let A = k[X, Y ](X,Y ) and let a = (X 7 , X 6 Y, XY 6 , Y 7 ). Using
COCOA [3] one verifies that
ha (A, z) = 38 + 3z + 3z 2 + 3z 3 + 3z 4 + 3z 5 − 4z 6 .
Notice ea2 (A) = 0. So by 7.6 we get a(Ga (A)) < 0. Since a(Ga (A)) ≥ −2 and a is
not generated by a regular sequence it follows from 7.4 that a(Ga (A)) = −1.
Question 7.8. What are all the m-primary ideals in a Cohen-Macaulay local ring
(A, m) having a(Ga (A)) = − dim A + 1?
8. Dual Filtrations of MCM modules over hypersurface rings
Let (A, m) be a complete equicharacteristic hypersurface ring and let M be a
MCM A-module. When Gm (M ) is CM we give a necessary and sufficient condition
for the dual filtration, FM = {HomA (M, mn )}n∈Z , on M ∗ to be a shift of the m-adic
filtration on M ∗ : see 8.11. This result along with Gorenstein approximation is then
used to deduce results about associated graded rings of q-primary a in a CohenMacaulay local ring (R, q) with µ(a) = dim R + 1. We give proof of application III
and half of application IV, stated in the introduction: see 8.17 and 8.16.
8.1. Setup: Let Q = k[[X0 , X1 , X2 . . . , Xd ]] where k is an infinite field and let n
be the maximal ideal of Q. Let (A, m) = (Q/(f ), n/(f )) where f ∈ ne \ ne+1 and
e ≥ 2. Notice e0 (A) = e and Gm (A) = Gn (Q)/(f ∗ ). So Gm (A) is Gorenstein.
8.2. Some invariants of a MCM module over a hypersurface ring.
20
TONY J. PUTHENPURAKAL
Case 1 : dim A = 0. So Q is a DVR. Let n = (Π). As a Q-module
µ(M)
(8.2.1)
M≃
M
Q/(Πai (M) ) where
1 ≤ a1 (M ) ≤ . . . ≤ aµ(M) (M ) ≤ e.
i=1
Thus the decomposition above is as A-modules.
8.3. The Hilbert function of M is
H(M, z) =
µ(M)
X
i=1
1 + z + . . . + z ai (M)−1 .
Remark 8.4. Notice H(M, z) completely determines a1 (M ), . . . , aµ(M) (M ) in the
case when dim M = 0.
Case 2: d = dim A ≥ 1.
By an argument similar to [22, 7.6] we get that for sufficiently general x1 , . . . xd ∈
m\m2 , the Hilbert function of M/(x1 , . . . xd )M is constant. Using 8.4 we can define
ai (M ) = ai (M/(x)M ) for sufficiently general x.
Definition 8.5. We call the numbers a1 (M ), . . . , aµ(M) (M ) the generic superficial
invariants of M .
Corollary 8.6. (with assumptions as above)
µ(M)
e0 (M ) =
X
ai (M )
i=1
Proof. Let x = x1 , . . . , xd be sufficiently general. Set N = M/xM. Then
µ(M)
µ(M)
e0 (M ) = e0 (N ) =
X
ai (N ) =
i=1
X
ai (M ).
i=1
φ
8.7. Assume 0 → Qn −
→ Qn → M → 0 is a minimal presentation of M . Set
iφ = max{i | all entries of φ are in ni }.
It is well known that iφ is an invariant of M . We set i(M ) = iφ .
Proposition 8.8. [with hypothesis as above] i(M ) = a1 (M ).
Proof. When dim M = 0 the its clear that i(M ) = a1 (M ). We can choose x =
x1 , . . . , xd sufficiently general such that i(M ) = i(M/xM ) (see [22, 4.4]). The
result follows from the zero-dimensional case.
Remark 8.9. If E = M ⊕ N then
Ei = HomA (M ⊕ N, mi ) = Mi∗ ⊕ Ni∗ .
So G(FE , E ∗ ) = G(FM , M ∗ ) ⊕ G(FN , N ∗ ).
The Dual Filtration in dimension zero.
Proposition 8.10. Let (Q, Π) be a DVR. Set A = Q/(Πe ) for some e ≥ 2 and let
M be an A-module.
(1) If M = Q/(Πi ) then G(FM , M ∗ ) ≃ G(M ∗ )(e − i).
GORENSTEIN APPROXIMATION
(2) If M =
Lµ(M)
i=1
21
Q/(Πai (M) ) then
µ(M)
G(FM , M ∗ ) =
M
G(Q/(Πai (M) ))(e − ai (M )).
i=1
∗
(3) G(FM , M ) ≃ G(M ) (up to a shift) if and only if a1 (M ) = · · · = aµ(M) (M ).
Proof. (1). Set
Mj∗ = HomA (M, mj ) = HomQ/(Πe ) Q/(Πi ), (Πj )/(Πe ) .
Clearly Mj∗ = M ∗ for j = 0, 1, . . . , e − i.
For j > e − i let f ∈ Mj∗ . Let f (1) = αΠj = Πj−(e−i) · (αΠe−i ).
Define g ∈ M ∗ by g(1) = αΠe−i . Clearly g is A-linear. Also g ∈ Me−i = M ∗ .
Thus f (1̄) = Πj−(e−i) g(1) ∈ Πj−(e−i) Me−i . Therefore Mj∗ ⊆ mj−(e−i) M ∗ . But
mj−(e−i) M ∗ ⊆ Mj∗ always. So Mj∗ = mj−(e−i) M ∗ . Thus
G(FM , M ∗ ) ≃ G(M ∗ )(e − i).
(2). This follows from 8.9 and (1).
(3). This follows from (2).
Theorem 8.11. (with hypothesis as in 8.1) Assume G(M ) is Cohen-Macaulay.
Then the following conditions are equivalent:
(i) G(M ∗ ) ≃ Hom(G(M ), G(A)) up to a shift.
(ii) a1 (M ) = . . . = aµ(M) (M ).
(iii) hM (z) = µ(M )(1 + z + . . . + z i(M)−1 ).
(iv) hM (z) = µ(M )(1 + z + . . . + z s−1 ) for some s ≥ 1.
We need the following
Lemma 8.12. (with hypothesis as in 8.1) Set α = i(M ). Then
X
h(M, z) = µ(M )(1 + z + . . . + z α−1 ) +
hi (M )z i .
i≥α
Furthermore hα (M ) < µ(M ).
Proof. We prove the result by induction on dimension of A. When dim A = 0 then
the result follows by 8.3.
If dim A > 0 then let x ∈ m \ m2 be sufficiently general. Set N = M/xM . We
may choose x such that i(M ) = i(N ) = α, see [22, 4.4]. By induction hypothesis
X
h(N, z) = µ(N )(1 + z + . . . + z α−1 ) +
hi (N )z i and hα (N ) < µ(N ).
i≥α
By Singh’s Lemma [29, Theorem 1]
H(N, z) X
−
ℓ
H(M, z) =
(1 − z)
i≥0
mi+1 M : x
mi M
zi.
By [22, 4.6] we get (mi+1 M : x) = mi M for all i = 0, 1, . . . , α − 1.
X mi+1 M : x
zi.
So hM (z) = h(N, z) − (1 − z)d •
ℓ
mi M
i≥α
22
TONY J. PUTHENPURAKAL
It follows that hi (M ) = hi (N ) for i = 0, . . . , α − 1 and
α+1
m
M: x
≤ hα (N ) < µ(N ) = µ(M ).
hα (M ) = hα (N ) − ℓ
mα M
The result follows.
Corollary 8.13. If h(M, z) = µ(M )(1 + z + . . . + z s−1 ) than i(M ) = s and G(M )
is Cohen-Macaulay.
Proof. The assertion i(M ) = s follows from Lemma 8.12. The fact that G(M ) is
Cohen-Macaulay follows from [22, Theorem 2].
We now give a
Proof of Theorem 8.11. (iii) ⇔ (ii) follows from [22, Theorem 2].
(iii) =⇒ (i). Let FM be the dual filtration on M ∗ with respect to m. As G(M ) is
Cohen-Macaulay it follows from Corollary 4.8 that G(F, M ∗ ) ∼
= Hom(G(M ), G(A)).
Let x = x1 , . . . , xd ∈ m \ m2 be sufficiently general. Set N = M/xM . Then by 4.10
we have
G(FM , M ∗ )/x∗ G(FM , M ∗ ) ∼
= G(FN , N ∗ ).
Notice a1 (M ) = . . . = aµ(M) (M ) = i(M ). By 8.10(1) it follows that the Hilbert
series of G(FN , N ∗ ) is µ(M )z e−i(M) (1 + z + . . . + z i(M)−1 ). It follows that the
Hilbert series of G(FM , M ∗ ) is
µ(M )z e−i(M) (1 + z + . . . + z i(M)−1 )
(1 − z)d
∗
Note α = α(M ) = e − i(M ). For all i ≥ 0 we always have mi+1 M ∗ ⊆ Mi+α+1
. So
we have an exact sequence
M∗
M∗
→ ∗
→0
i+1
∗
m M
Mi+α+1
∗
for all
Since the lengths are equal it is an isomorphism. Thus mi+1 M ∗ = Mi+α+1
∗
i ≥ 0. Therefore G(M ) ≃ Hom(G(M ), G(A)) up to a shift.
(i) =⇒ (iii) By hypothesis G(M ∗ ) ≃ G(F, M ∗ ) (up to a shift). Thus G(M ∗ )
is Cohen-Macaulay. By 4.10 we go mod a maximal regular sequence to reduce to
dimension zero case. Here the assertion is true by 8.10(3).
(iii) =⇒ (iv) Nothing to show.
(iv) =⇒ (iii) Follows from Corollary 8.13.
8.14. The case when (A, m) is Cohen-Macaulay (need not be a hypersurface ring) but a is an m-primary ideal with µ(a) = d + 1.
Remark 8.15. Assume A is also complete and equicharacteristic. By 3.6 we get
that [A, m, a] has a Gorenstein approximation [B, n, n, ψ] with µ(n) = d + 1. It
follows that B is a hypersurface ring. Also A is a MCM as a B-module. Clearly
Ga (A) = Gn (A). It can be easily checked that ωA ∼
= HomB (A, B) = A∗ and
*
∼ B
ΩA
HomG(B) (Gn (A), ΩB
a =
n ). Also note that Gn (B) = Ωn up to a shift.
Theorem 8.16. Let (A, m) be a Cohen-Macaulay equicharacteristic local ring and
a canonical module ωA . Let a be an m-primary ideal with µ(a) = d + 1 and Ga (A)
Cohen-Macaulay with canonical module ΩA
a . The following conditions are equivalent:
GORENSTEIN APPROXIMATION
23
(i) Ga (ωA ) ≃ ΩA
a (up to a shift).
(ii) ha (A, z) = ℓ(A/a)(1 + z + . . . + z s−1 ) for some s ≥ 2.
Proof. This follows from 8.15 and 8.11.
Theorem 8.17. Let (A, m) be a equicharacteristic Gorenstein local ring and a an
m-primary ideal with µ(a) = d + 1. The following conditions are equivalent:
(i) Ga (A) is Gorenstein.
(ii) ha (A, z) = ℓ(A/a)(1 + z + . . . + z s−1 ) for some s ≥ 2.
Proof. The assertion (i) =⇒ (ii) follows from 8.16.
(ii) =⇒ (i). We use notation as in 8.15. Notice A is a MCM B-module and
ha (A, z) = hn (A, z). So by 8.13 we get that G(A) = Ga (A) is Cohen-Macaulay.
The result follows from 8.16.
Question 8.18. Are the results of Theorem 8.17 and Theorem 8.16 true when A
does not contain a field?
9. A generalization of a result due to Ooishi
9.1. Introduction & Setup: Let (A, m) be a CM local ring with a canonical module ωA . Let a an m-primary ideal and M a CM A-module. Set M † =
dim A−dim M
ExtA
(M, ωA ). Assume Ga (A), Ga (M ) are CM.
Let Ωa , Ka be the canonical module of Ga (A), Ra (A) respectively. Notice that
Ωa ∼
= Ka /t−1 Ka (−1)(see [2, 3.6.14]).
In Theorem 9.5 we show that if A is complete then there is a a-stable filtration
F on M † such that
R(F, M † ) ∼
=
(9.1.1)
equivalently
(9.1.2)
G(F(−1), M † ) ∼
=
*
dim A−dim M
ExtRa (A)
dim A−dim M
*
ExtGa (A)
(Ra (M ), Ka )
(Ga (M ), Ωa ) .
As an application we give a generalization of a result due to Ooishi.
Definition 9.2. [With hypotheses as in 9.1] If there exists a a-stable filtration F on
M † such that 9.1.1 (equivalently 9.1.2) holds then we say F is a canonical filtration
on M † with respect to a.
Two canonical filtrations (say F, G) on M † are equivalent, i.e., there exists an
A-linear isomorphism σ : M † → M † such that σ(Fn ) = Gn for all n ∈ Z. This is
due to the following well-known result.
Theorem 9.3. Let M be an A-module and let a be an ideal in A. Suppose
F = {Fn }n∈Z and G = {Gn }n∈Z are two a-stable filtration’s such that R(F, M ) ∼
=
R(G, M ) as Ra (A)-modules. Then F, G are equivalent filtration’s.
We need the following graded version of a result of Rees [25].
L
Remark 9.4. Let R =
n∈Z Rn be a graded ring and let M, N be graded Rmodules. Let x be a homogeneous element in R. If x is both M and R-regular
element and x · N = 0 then
*
i
ExtR/xR (N, M/xM ) ∼
=
*
i+1
ExtR (N, M )(− deg x)
for all i ≥ 0.
The proof is along the same lines as given (for example in) [2, 3.1.16]
24
TONY J. PUTHENPURAKAL
Theorem 9.5. [With hypotheses as in 9.1] Further assume A is complete. Then
M † has a canonical filtration. Furthermore if F, G are two such filtration’s then
they are equivalent as filtration’s on M † .
Proof. Uniqueness of the filtration up to equivalence follows by 9.3.
Case 1: M is a MCM A-module.
Let [B, n, b, φ] be a Gorenstein approximation of (A, m, a). Consider the dual filtration FM = {HomB (M, bn )}n∈Z on M † . Since Gb (B) is Gorenstein & Gb (M ) =
Ga (M ) is CM, we get
G(FM , M † ) ∼
=
(9.5.1)
*
HomGb (B) (Ga (M ), Gb (B))
(see 4.8).
The canonical module of Gb (B) is Gb (B)(s) for some s ∈ Z. So
Ωa ∼
=
*
HomGb (B) (Ga (A), Gb (B)(s))
(see [2, 3.6.12]).
Notice that we have the following isomorphisms of G = Ga (A)-modules
*
HomG (Ga (M ), Ωa ) ∼
= * HomG Ga (M ), * HomGb (B) (G, Gb (B)(s))
∼
= * HomGb (B) (Ga (M ), Gb (B)(s))
∼
= G(FM (s), M † ), using 9.5.1 and 2.3 .
Case 2: M is a Cohen-Macaulay A-module but not necessarily MCM
By case 2 we may assume that c = dim A − dim M > 0. Let u1 , . . . uc be as in
Corollary 3.2. Recall
(1) u1 , . . . , uc ∈ annA M and u∗1 , . . . , u∗c ∈ annGa (A) Ga (M ).
(2) u∗1 , . . . , u∗c is a Ga (A)-regular sequence.
Set A′ = A/(u), a′ = a/(u). Then M is a A′ -module. Also notice that Ga′ (M ) =
Ga (M ). Since Ωa is a MCM Ga (A) module, it follows that u∗1 , . . . , u∗c is Ωa -regular.
Notice also that u1 , . . . , uc ∈ annA M † and u∗1 , . . . , u∗c ∈ annGa (A) Ga (M † ).
Pc
Set w = i=1 deg u∗i . Notice Ωa /(u∗ Ωa ) (w) is the canonical module of Ga′ (A′ );
cf. [2, 3.6.14]. Using 9.4 repeatedly we get
(9.5.2)
*
HomGa′ (A) (Ga′ (M ), Ωa /(u∗ Ωa )) ∼
=
*
c
ExtGa (A) (Ga (M ), Ωa ) (−w) .
By case 1 we have a a′ -stable filtration F on M † such that
G(F, M † ) ∼
=
*
HomGA′ (A′ ) (Ga′ (M ), Ωa /(u∗ Ωa ) (w)) .
Notice F is an a-stable filtration on M † . Also Ga′ (M ) = Ga (M ). By 9.5.2 we get
G(F, M † ) ∼
=
*
dim A−dim M
ExtGa (A)
(Ga (M ), Ωa ) .
We now state our generalization of Ooishi’s result.
Theorem 9.6. [with hypotheses as in 9.1] Set r = red(a, M ) and assume further
that G(M † ) is Cohen-Macaulay. Then the following assertions are equivalent:
(i) h(M † , z) = z r · h(M, z −1 ).
∼ * Extdim A−dim M (Ga (M ), Ωa ) (up to a shift).
(ii) G(M † ) =
Ga (A)
For proving Theorem 9.6 we need the following Lemma which gives a sufficient
condition for an a-stable filtration to be a-adic
GORENSTEIN APPROXIMATION
25
Lemma 9.7. Let (A, m) be local, a an m-primary ideal and let M be a CohenMacaulay A-module of dimension r ≥ 1. Let F = {Mn }n∈Z be an a-stable filtration
on M such that
(1) Mn = M for n ≤ 0 and M1 6= M .
(2) G(F, M ) is Cohen-Macaulay.
Let x = x1 , . . . , xs be an M -superficial sequence with respect to M such that x∗ =
x∗1 , . . . , x∗s is G(F, M )-regular. Set B = A/(x), b = a/(x) and N = M/xM . Let
F = {(Mn + xM )/xM }n∈Z be the quotient filtration on N . If F is the b-adic
filtration on N then F is the a-adic filtration on M .
Proof. Since M0 = M we get ai M ⊆ Mi for all i ≥ 1. We prove by induction on n
that Mn = an M for n ≥ 1.
The case n = 1.
F is b-adic. So
M1 + xM
aM + xM
=
xM
xM
Thus M1 + xM = aM + xM . But xM ⊆ aM ⊆ M1 . Therefore M1 = aM .
Assume the result for n = p and prove for n = p + 1. By hypothesis on F,
Mp+1 + xM
ap+1 M + xM
=
.
xM
xM
(*)
Notice
(**)
Mp+1
Mp+1 + xM ∼
=
xM
xM ∩ Mp+1
and
ap+1 M + xM ∼
ap+1 M
.
=
xM
xM ∩ ap+1 M
Observe that
xap M ⊆ ap+1 M ∩ xM ⊆ Mp+1 ∩ xM
= xMp
p
= xa M
p
So xM ∩ Mp+1 = xa M = a
p+1
[4, 2.3]
by induction hypothesis
M ∩ xM . The result follows from (*) and (**).
Proof of Theorem 9.6. The assertion (ii) =⇒ (i) follows from [2, 4.4.5].
To prove the converse, we note that using the argument Case 2 in Theorem 9.5
we may assume that M is a MCM A-module. Furthermore we may assume that A
is complete with infinite residue field.
We first consider the case when dim A = 0. Consider the usual a-adic filtration:
M † ⊇ aM † ⊇ a2 M † ⊇ . . . ⊇ ar M † ⊇ ar+1 M † = 0
Let the canonical filtration of M † (up to a shift) be :
M † ⊇ q1 ⊇ q2 ⊇ . . . ⊇ qs ⊇ qs+1 = 0 with q1 6= M † .
Let the Hilbert series of M † with respect to a be
Ha (M, z) = h0 + . . . + hr z r .
Then the Hilbert series of the canonical module of Ga (M )† is
Ha (M, z −1 ) = hr z −r + hr−1 z −r+1 + . . . + h1 z −1 + h0
cf. [2, 4.4.5]. It follow that s = r. Also ℓ(qr ) = h0 = hr = ℓ(ar M † ). Since qr ⊇ ar M †
we have ar M † = qr .
By downward induction we show qi = ai M † for i ≤ r. Notice qi ⊇ ai M † .
26
TONY J. PUTHENPURAKAL
For i = r we just showed ar M † = qr . Assume for i = j + 1 and prove for i = j.
Notice that
j †
a M
qj
.
= hr−j = hj = ℓ
ℓ
qj+1
aj+1 M †
But qj+1 = aj+1 M † and qj ⊇ aj M † . So qj = aj M † . The result follows.
Next we consider the case when d = dim A > 0. By hypothesis Ga (A), Ga (M )
are Cohen-Macaulay. Let x1 , . . . , xd ∈ a/a2 be such that x∗1 , . . . , x∗d is Ga (A) and
Ga (M ) regular sequence. Set B = A/(x), N = M/xM and b = aB. Furthermore
Gb (B) = Ga (A)/(x∗ ) and Gb (N ) = Ga (M )/x∗ Ga (M ). So the h-vector of Gb (N )
is symmetric. By the previous case we have that the b-adic filtration on N † is the
canonical filtration on N † up to a shift. By 9.7 it follows that the a-adic filtration
on M † is the canonical filtration on M † up to a shift. Therefore
G(M † ) ∼
=
*
dim A−dim M
ExtGa (A)
(Ga (M ), Ωa ) (up to a shift).
10. Application of our generalization of Ooishi’s result
We use Theorem 9.6 to to show that the dual filtration with respect to m of a
MCM module M , over a Gorenstein local ring A with Gm (A) Gorenstein, is a shift
of the usual m-adic filtration in the following cases:
(1) M is Ulrich (i.e., deg hM (z) = 0; equivalently e(M ) = µ(M )).
(2) type(M ) = e(M ) − µ(M ) and M has minimal multiplicity.
Assume (R, n) is CM local with a canonical module ωR and Gq (R) is CM for some
n-primary ideal q. We use Theorem 9.6 to give an easily verifiable condition on
whether Gq (ωR ) is the canonical module of Gq (R) (up to a shift).
The following criterion is useful to show dual filtrations are a-adic up to a shift.
Proposition 10.1. Let (A, m) be Gorenstein, a an m-primary ideal with Ga (A)
Gorenstein. Let M be an MCM A-module with Ga (M ) Cohen-Macaulay. Let F be
the dual filtration of M with respect to a. Set c = red(a, M ). Let x = x1 , . . . , xd be
an A superficial sequence with respect to a. Set (B, n) = (A/(x), m/(x)), b = A/(x)
and N = M/xM . Set N ∗ = HomB (N, B). The following conditions are equivalent:
(i) h(N ∗ , z) = z c · h(M, z −1 ).
(ii) Ga (M ∗ ) ∼
= G(F, M ∗ ) (up to a shift).
(iii) F is the a-adic filtration of M ∗ (up to a shift).
Proof. (iii) =⇒ (ii). Is trivial
(ii) =⇒ (iii). Follows from 2.5.
(i) =⇒ (iii) Let G be the dual filtration of N ∗ with respect to b. Notice
h(N, z) = h(M, z), as Ga (M ) is CM. Using 9.6 we get that G is b-adic up to a shift.
Notice x∗ = x∗1 , . . . , x∗d is a Ga (A)-regular sequence, [21, 8]. So x∗ = x∗1 , . . . , x∗d
is a Ga (M )-regular sequence. Using 4.10 we get that G is the quotient filtration of
F. By 9.7 it follows that F is a-adic up to a shift.
(ii) =⇒ (i). By 4.8, Ga (M ∗ ) is a MCM Ga (A)-module. So h(N ∗ , z) = h(M ∗ , z).
We also have h(N, z) = h(M, z). The result follows from 9.6.
Applications:
Dual filtrations of Ulrich modules:
GORENSTEIN APPROXIMATION
27
Theorem 10.2. Let (A, m) be a Gorenstein local ring with Gm (A) Gorenstein.
Assume A has infinite residue field. Let r = red(A). Let M be a MCM A-module.
The following conditions are equivalent:
(i) Mr+i = mi M ∗ for all i ∈ Z.
(ii) G(F, M ∗ ) ≃ G(M ∗ )(−r).
(iii) M is an Ulrich A-module.
Proof. By 2.5 (i) and (ii) are equivalent.
(ii) =⇒ (iii) Notice α(M ) = r = red(A). So M is Ulrich; see 6.11.
(iii) =⇒ (ii) Let N = M/xM where x = x1 , . . . , xd is a maximal superficial
A-sequence. Notice N = k e0 (M) . Set (B, n) = (A/(x), m/(x)). Notice N ∗ =
HomB (N, B) ∼
= N since B is Gorenstein. Clearly h(N ∗ , z) = e0 (M ) = z 0 h(M, z −1 ).
So by 10.1 we get that G(F, M ∗ ) ∼
= G(M ∗ ) up to a shift; say t.
Determining t: Since G(M ) is Cohen-Macaulay, we may reduce to the zerodimensional case, see 4.10. So M = k µ(M) . In this case it is easy to show that
G(F, M ∗ ) ∼
= G(M ∗ )(−r). So t = r.
Dual filtration of non-Ulrich Modules having minimal multiplicity
Remark 10.3. Let type(M ) = Cohen-Macaulay-type of M = dimk ExtdA (k, M ).
If M has minimal multiplicity then τ (M ) ≥ e0 (M ) − µ(M ).
Proof : Let x1 , . . . , xd be a maximal M -superficial sequence. Set N = M/xM.
Note that e0 (M ) = e0 (N ) = ℓ(N ). Set (B, n) = (A/(x), m/(x)). Since M has
minimal multiplicity we get n2 N = 0. So nN ⊆ soc(N ). As Gn (N ) = N/nN ⊕
nN/(0). So ℓ(nN ) = e0 (N ) − µ(N ) = e0 (M ) − µ(M ). Since n2 N = 0 we get
nN ⊆ soc(N ). So e0 (M ) − µ(M ) ≤ ℓ(soc(N )) = type(N ) = type(M ).
Theorem 10.4. Let (A, m) be a Gorenstein local ring with an infinite residue
field. Assume Gm (A) is Gorenstein. Let M be a MCM A-module with minimal
multiplicity with respect to m and M is NOT Ulrich. The following conditions are
equivalent:
(i) type(M ) = e0 (M ) − µ(M ).
(ii) M ∗ has minimal multiplicity and G(FM , M ∗ ) ≃ Gm (M ∗ ) up to a shift.
(iii) G(FM , M ∗ ) ≃ Gm (M ∗ ) up to a shift.
Proof. Set G = Gm (A), G(M ) = Gm (A).
(i) =⇒ (ii). Since M has minimal multiplicity we get that
h(M, z) = µ(M ) + (e0 (M ) − µ(M ))z
Notice M is MCM. Let x = x1 , . . . , xd ∈ m \ m2 be sufficiently general. Set
(B, n) = (A/(x), m/(x)), N = M/xM . Also note that N ∗ ∼
= M ∗ /xM ∗ . Since
2
2 ∗
∗
∗
n N = 0, we get n N = 0. So H(N , z) = µ(N ) + ℓ(nN ∗ )z. By hypothesis
it follows that µ(N ∗ ) = type(N ) = type(M ) = e0 (M ) − µ(M ). Since ℓ(N ∗ ) =
e0 (N ∗ ) = e0 (N ) = e0 (M ) we get that
∗
h(N ∗ , z) = (e0 (M ) − µ(M )) + µ(M )z = zh(M, z −1 ).
So by 10.1 we get G(FM , M ∗ ) ∼
= G(M ∗ ) up to a shift. So G(M ∗ ) is Cohen-Macaulay
∗
∗ ∼
(by 4.8). Since N = M /xM ∗ and N ∗ has minimal multiplicity it follows that
M ∗ has minimal multiplicity.
(ii) =⇒ (iii) is clear
(iii) =⇒ (i) Using 10.1 and 4.8 it follows that h(M ∗ , z) = zh(M, z −1 ). It follows
that type(M ) = e0 (M ) − µ(M ).
28
TONY J. PUTHENPURAKAL
Associated graded module of the Canonical module
10.5. Let (A, m) be Cohen-Macaulay local ring with a canonical module ωA . Let
a be an m-primary ideal such that Ga (A) is Cohen-Macaulay. Let r = red(a, A).
Let Ωa be the canonical module of Ga (A). Let type(A) = ℓ(Ext d A (A/m, A)) be
the Cohen-Macaulay type of A.
Theorem 10.6. [hypothesis as in 10.5] Let x, . . . , xd be a maximal A-superficial sequence with respect to a. Set B = A/(x) and b = a/(x). If h(ωB , z) = z r h(B, z −1 )
then Ga (ωA ) is Cohen-Macaulay and isomorphic to Ωa up to a shift.
Proof. Notice that we may assume that A is complete. Then by 11.4 there exists
[R, n, q, φ], a Gorenstein approximation of [A, m, a] with the property that there
exists y1 , . . . , yd ∈ q such that φ(yi ) = xi for i = 1, . . . , d. Furthermore y1∗ , . . . , yd∗ is
Gq (R)-regular. Notice that ωA ∼
= HomR (A, R), i.e., the dual of the MCM R-module
A. The result follows from 10.1.
Theorem 10.6 is quite practical as shown by the following:
Example 10.7. Let A = Q[t13 , t18 , t23 , t28 , t33 ] and m = (t13 , . . . , t33 ). Set B =
A/(t13 ). By [15, 1.1]; Gm (Am ) is CM. Set R = Q[x, y, z, w]. Consider the natural
map φ : R → A with φ(x) = t13 , . . . , φ(w) = t33 . Using MACAULAY we can
compute q = ker φ. Set S = R/q and n = (x, y, z, w). Also set T = S/(x). Using
COCOA we get
h(A, z) = h(S, z) = 1 + 4z + 4z 2 + 4z 3
and h(B, z) = h(T, z) = 1 + 4z + 4z 2 + 4z 3
This proves that Gn (Sn ) is CM and x∗ is Gn (Sn )-regular. (Notice this also proves
∗
Gm (Am ) is CM and t13 is Gm (Am )-regular).
We consider R graded with deg x = 13, deg y = 18, deg z = 28 and deg w = 33.
Set D = Ext5R (R/(q + (x)), R). Note that
D∼
= ωA /t13 ωA .
= ωS /xωS ∼
= ωT ∼
n
n
n
m
m
Using MACAULAY we get that
h(D, z) = 4 + 4z + 4z 2 + z 3 = z 3 h(B, z −1 )
By 10.6 we get that G(ωAm ) is the canonical module of Gm (Am ) up to a shift.
An easy consequence of 10.6 is the following:
Corollary 10.8. Let (A, m) be Cohen-Macaulay local ring with red(A) = 2. Then
Gm (ωA ) ∼
= Ωm (up to a shift) if and only if type(A) = e0 (A) − h1 (A) − 1.
Proof. Notice Gm (A) is CM; [27, 2.1]. Set h(A, z) = 1+h1 z +h2 z 2 and e = e0 (A) =
e0 (ωA ).
(⇒) This implies h(ωA , z) = z 2 h(A, z −1 ). It follows that
type(A) = µ(ω) = h2 = e0 (A) − h1 − 1.
(⇐) We may assume A has infinite residue field. Notice h2 = type(A). Let
x = x1 , . . . , xb be a maximal superficial A-sequence. Set (B, n) = (A/(x), m/(x).
Notice ωB ∼
= ωA /(x)ωA . Check that
h(ωB , z) = type(B) + (e − type(B) − 1)z + z 2 = z 2 h(A, z −1 ).
The result follows from 10.6.
GORENSTEIN APPROXIMATION
29
11. CI approximation: The general case
In this section we prove our general result regarding complete intersection (CI)
approximation; see 11.3. In 3.6 we showed that every m-primary ideal a in an
equicharacteristic local ring A has a CI-approximation. Even if we are interested
only in the case of m-primary ideals; we have to take some care for dealing with
the case of local rings with mixed characteristics. The essential point is to show
existence of homogeneous regular sequences in certain graded ideals. We also prove a
Lemma regarding lifting of superficial sequence along a Gorenstein approximation;
see 11.4.
L
This section is divided into two subsections. In the first subsection R = n≥0 Rn
be a standard
L graded algebra over a local ring (R0 , m0 ). When R is Cohen-Macaulay
and M = n≥0 Mn is a finitely generated graded R-module generated by elements
in M0 , we give conditions to ensure a regular sequence ξ1 , . . . , ξc ∈ R+ ∩ annR M
where c = dim R − dim M (see Theorem 11.2). In the second subsection we prove
Theorem 11.3.
Homogeneous
regular sequence
L
Let R = n≥0 Rn be a standard algebra over a local ring (R0 , m0 ) and let M =
L
n≥0 Mn be a finitely generated graded R-module. In general a graded module M
need not have homogeneous regular sequence ξ1 , . . . , ξc ∈ R+L
∩ annR M where c =
grade M (see 11.1). When R is Cohen-Macaulay and M = n≥0 Mn is a finitely
generated graded R-module generated by elements in M0 , we give conditions to
ensure a regular sequence ξ1 , . . . , ξc ∈ R+ ∩ annR M where c = dim R − dim M .
We adapt an example from [2, p. 34] to show that a homogeneous regular sequence of length grade M in R+ ∩ annR M need not exist.
Example 11.1. Let A = k[[X]] and T = A[Y ]. Set R = T /(XY ). Notice that
R0 = A. Set M = (X, Y )R the unique graded maximal ideal of R and let M =
R/M = k. Clearly grade M = grade(M, R) = 1. It can be easily checked that
every homogeneous element in R+ is a zero-divisor.
The following result regarding existence of homogeneous regular sequence in
(ann M ) ∩ R+ (under certain conditions) is crucial in our proof of Theorem 11.3.
L
Theorem 11.2. Let (R0 , m0 ) be a local ring and let R = n≥0 Rn be a standard
L
graded R0 -algebra and let M = n≥0 Mn be a finitely generated R-module. Assume
R has a homogeneous s.o.p. Set c = dim R − dim M . Assume
(1) R0 is Cohen-Macaulay and R is Cohen-Macaulay.
(2) M is generated as an R-module by some elements in M0 .
(3) There exists y1 , . . . , ys ∈ R0 a system of parameters for both M0 and R0 and
a part of a homogeneous system of parameters for both M and R.
Then there exists a homogeneous regular sequence ξ1 , . . . , ξc ∈ R+ such that ξi ∈
annR M for each i.
Proof. We prove this by induction on s = dim R0 . When s = 0 R0 is Artinian. So
grade(ann M, R) = grade(ann M ∩R+ , R). As R is CM we have grade(ann M, R) =
c. The result follows from [2, 1.5.11].
We prove the
L assertion for s = r + 1 assuming it to be valid for s = r. Set
S = R/y1 R = n≥0 Sn and N = M/y1 M . Note that N is also generated in degree
0. Let yi = yi mod (y1 ) with 2 ≤ i ≤ s. Then y2 . . . ys is a system of parameters
for both S0 and N0 = M0 /y1 M . Notice c = dim R − dim M = dim S − dim N . Also
30
TONY J. PUTHENPURAKAL
note that S has h.s.o.p. By induction hypothesis there exists a S-regular sequence
ξ1 , . . . , ξc ∈ S+ such that ξi ∈ annS N for each i.
R is CM. So y1 is R-regular. Therefore y1 , ξ1 , ξ2 . . . , ξc is an R-regular sequence.
Set q = y1 R+ . Let M be generated as an R-module by u1 , . . . , um ∈ M0 . Fix
i ∈ {1, . . . , c} and set ξ = ξi . By construction ξuj ∈ y1 M for each j. However
ξ ∈ R+ and y1 ∈ R0 . So ξuj ∈ qM for each j. By the determinant trick (observing
that each uj has degree zero) there exists a homogeneous polynomial
∆(ξ) = ξ n + a1 ξ n−1 + . . . + an−1 ξ + an ∈ annR M
and ai ∈ qi for i = 1, . . . , n.
Clearly ∆(ξ) ∈ R+ . Notice that ∆(ξ) = ξ n mod (y1 ). So y1 , ∆(ξ1 ), . . . , ∆(ξc )
is a regular sequence, [14, 16.1]. Therefore ∆(ξ1 ), ∆(ξ2 ) . . . , ∆(ξc ) is an R-regular
sequence. This proves the assertion when s = r + 1.
CI approximation
In this subsection we prove our general result regarding CI-approximation.
Theorem 11.3. Let (A, m) be a complete local ring and let a be a proper ideal in
A with dim A/a + s(a) = dim A. Then A has a CI-approximation with respect to a.
Furthermore if a = (x1 , . . . , xm ) then we may choose a CI-approximation [B, n, b, ϕ]
of [A, m, a] such that there exists u1 , . . . , um ∈ b with ϕ(ui ) = xi for i = 1, . . . , m.
Proof. Notice Ga (A) has a homogeneous s.o.p (see proof of Proposition 2.6 in [6]).
Let y1 , . . . , ys ∈ A such that y1 , . . . , ys (their images in A/a ) is an s.o.p of A/a.
Case 1. A contains a field: As A is a complete A contains k = A/m. Set
S = k[[Y1 , . . . , Ys , X1 , . . . , Xn ]]. Define φ : S → A which maps Yi to yi and Xj to
xj for i = 1, . . . , s and j = 1, . . . , n. Set q = (Y1 , . . . , Ys ). Notice
A
A
=
(q, X1 , . . . , Xn )A
(a, y1 , . . . , ys )
has finite length.
As S is complete we get that A is a finitely generated S-module, cf. [14, 8.4]. Since
ψ((X))A = a, we get Ga (A) = G(X) (A) is a finitely generated G(X) (S)-module.
Notice G(X) (S) ∼
= S/(X)[X1∗ , . . . , Xn∗ ] and S/(X) ∼
= k[|Y1 , . . . , Ys |]. Furthermore
G(X) (A) = Ga (A) is generated in degree zero as a G(X) (S)-module. Set c =
dim S − dim A = dim G(X) (S) − dim Ga (A). Notice that Y1 , . . . , Ys is a s.o.p of
G(X) (S)0 and Ga (A)0 . We apply Theorem 11.2 to get ξ1 , . . . , ξc ∈ (X) such that
ξ1∗ , . . . , ξc∗ ∈ G(X) (S)+ is a regular sequence in G(X) (S) and ξi∗ ∈ annG(X) (S) Ga (A)
for all i. We now apply Lemma 3.1 to get u1 , . . . , uc ∈ (X) an S-regular sequence
in annS (A) such that u∗1 , . . . , u∗c is a G(X) (S)-regular sequence in annG(X) (S) Ga (A).
Set u = (u1 , . . . , uc ), B = S/u, b = (X + u)/u. Let n be the maximal ideal in
B and let φ : B → A be the map induced by φ : S → A. Then it is clear that
(B, n, b, φ) is a CI approximation of A with respect to a.
Case 2. A does not contain a field: There exists a DVR (D, π) and a local ring
homomorphism η : D → A such that η induces an isomorphism D/(π) → A/m.
Set S = D[|Y1 , . . . , Ys , X1 , . . . , Xn |]. Define φ : S → A, with φ(d) = η(d) for each
d ∈ D and maps Yi to yi and Xj to xj for i = 1, . . . , s and j = 1, . . . , n. Set
q = (Y1 , . . . , Ys ). Notice
A
A
=
(q, X1 , . . . , Xn )A
(a, y1 , . . . , ys )
has finite length.
As S is complete we get that A is a finitely generated S-module.
GORENSTEIN APPROXIMATION
31
Set T = S/(X) = GX (S)0 . Notice dim T is one more than that of A/a = Ga (A)0 .
To deal with this situation we use an argument from [24], which we repeat for the
convenience of the reader. Let t be the maximal ideal of T . Set Yi = image of Yi
in T . Set V = A/a. Note that
q
q
t = annT (V /YV ) = annT (V ) + (Y).
So there exists f ∈ annT (V ) \ (Y) such that f , Y1 , . . . , Ys is an s.o.p. of T . Since
T is CM; f, Y1 , . . . , Ys is a T -regular sequence. So X1 , . . . , Xn , f, Y1 , . . . , Ys is a
S-regular sequence. Since f ∈ annT (A/a) we get f A ⊆ aA = (X)A. Using the
determinant trick there exists
∆ = f m + α1 f m−1 + . . . + αm−1 f + αm ∈ annS A
&αi ∈ (X)i ; 1 ≤ i ≤ m.
Notice that ∆ = f m mod(X). Thus X1 , . . . , Xn , ∆, Y1 , . . . , Ys is a S-regular sequence. So ∆ is S-regular. Set U = S/(∆) and c = ((X) + (∆))/(∆). Notice that A is a finitely generated U -module and Gc (A) = Ga (A). Furthermore
Gc (U ) ∼
= T /(f m )[X1∗ , . . . , Xn∗ ] is CI.
Note that as a Gc (U ) module Ga (A) is generated in degree zero. Furthermore
dim Gc (U )0 = dim Ga (A)0 . Set c = dim U − dim A = dim Gc (U ) − dim Ga (A). The
subsequent argument is similar to that of Case 1.
The following result is needed for Theorem 10.6.
Lemma 11.4. Let (A, m) be a complete Cohen-Macaulay local ring of dimension d
and let a = (x1 , . . . , xd , w1 , . . . , ws ) be a m-primary ideal. Assume that x1 , . . . , xd is
an A-superficial sequence. Then there exist [B, n, b, φ], a Gorenstein approximation
of [A, m, a] such that
(1) there exists v1 , . . . , vd ∈ b with φ(vi ) = xi for i = 1, . . . , d.
(2) v1∗ , . . . , vd∗ is Gb (B)-regular.
Proof. The proof of Theorem 11.3 shows that there exists a Cohen-Macaulay local
ring (R, η) of dimension d+s, an η-primary ideal and a ring homomorphism ψ : R →
A such that
(1) A is a finitely generated as a R-module (via ψ).
(2) (a) In the equicharacteristic case R = k[[V1 , . . . , Vd , W1 , . . . , Ws ]].
(b) In the mixed characteristic case R = D/(f )[[V1 , . . . , Vd , W1 , . . . , Ws ]] with
(D, Π) a DVR and f ∈ (Π).
(3) In both cases q = (V1 , . . . , Vd , W1 , . . . , Ws ), ψ(Vi ) = xi for i = 1, . . . , d and
Ψ(Wi ) = wi for i = 1, . . . , s. Notice Gq (R) = k[V1∗ , . . . , Vd∗ , W1∗ , . . . , Ws∗ ] in
the equicharacteristic case and Gq (R) = D/(f )[V1∗ , . . . , Vd∗ , W1∗ , . . . , Ws∗ ] in the
mixed characteristic case.
Set c = dim R − dim A. Set S = R/(V1 , · · · , Vd ) and q = q/(V1 , . . . , Vd ). Note
Gq (S) = Gq (R)/(V1∗ , . . . , Vd∗ ). Set C = A/(x1 , . . . , xd )
We then choose ξ1 , . . . , ξc ∈ q such that ξ1∗ , . . . , ξc∗ ∈ annGq (R) Ga (C) and that it
is a regular sequence in Gq (R). After raising powers we may assume that deg ξi∗ = r
for all i. Set M = A (note A need not be cyclic as a R-module). Fix i. Set J = ar+1 .
We have ξi M ⊆ ar+1 M + (V1 , . . . , Vd )M . By the determinant trick there exists a
monic polynomial f (X) ∈ A[X] such that
f (X) = X n + a1 X n−1 + · · · + an−1 X + an
with ai ∈ J i + (V1 , . . . , Vd ).
32
TONY J. PUTHENPURAKAL
and ui = f (ξi ) ∈ ann M It can be eaily checked (as in proof of Lemma 3.1) that in
Gq (S) we have ui = (ξin )∗ = (ξi∗ )n . Thus V1 ,∗ · · · , Vd∗ , u∗1 , . . . , u∗c is a Gq (R)-regular
sequence. Then note that u∗1 , . . . , u∗c , V1 ,∗ · · · , Vd∗ is also a Gq (R)-regular sequence.
Set B = R/(u), b = q/(u) and φ = ψ ⊗ R/(u). Notice Gb (B) = Gq (R)/(u∗ ).
Set vi = image of Vi in B. Then note that v1 ,∗ , . . . , vd∗ is a Gq (B)-regular sequence
and [B, n, b, φ], a Gorenstein approximation of [A, m, a].
12. Dual Filtrations mod a superficial sequence
The behavior of Dual filtrations mod a superficial sequence is a crucial result in
our paper. Unfortunately the proof is technical and unappealing.
12.1. Setup: Let (A, m) be Gorenstein, a-equimultiple with Ga (A) Gorenstein and
let M beL
a M CM A-module. We assume A/m is an infinite field. Let x = x1 , . . . , xr
be a M
A superficial sequence with respect to a. Set B = A/(x), b = a/(X) and
N = M/xM . The natural map Ga (A)/(x∗ ) → Gb (B) is an isomorphism. Set
FM = {HomA (M, an )}n∈Z
and FN = {HomB (N, nn )}n∈Z .
A natural question is when is
G(FM , M ∗ ) ∼
= G(FN , N ∗ ).
x∗ G(FM , M ∗ )
Theorem 12.5(2) asserts that this is so when Ga (M ) is Cohen-Macaulay.
Some Natural Maps:
12.2. The natural maps HomA (M, an ) → HomB (N, bn ) induces a map of Ra (A)modules
∗
∗
ΓM
x : R(FM , M ) → R(FN , N ).
(12.2.1)
Set q = (x1 t, . . . xr t). Clearly qR(FM , M ∗ ) ⊆ ker ΓM
x . Since Ra (A) is CohenMacaulay, we have Ra (A)/qRa (A) = Rb (B). So we have a map Rb (B)-modules
ΓM
x :
(12.2.2)
R(FM , M ∗ )
→ R(FN , N ∗ ).
qR(FM , M ∗ )
12.3. The map 12.2.1 induces a map of Ga (A)-modules
∗
∗
ΘM
x : G(FM , M ) −→ G(FN , N ).
(12.3.1)
Clearly x∗ G(FM , M ∗ ) ⊆ ker ΘM
x . So we have a map Gb (B)-modules
ΘM
x :
(12.3.2)
G(FM , M ∗ )
−→ G(FN , N ∗ ).
x∗ G(FM , M ∗ )
12.4. The natural map Ra (A) → Rb (B) induces a natural map of Ra (A)-modules
ΠM
x :
*
HomRa (A) (R(a, M ), Ra (A)) −→ * HomRb (B) (R(b, N ), Rb (B)).
Clearly q * HomRa (A) (R(a, M ), Ra (A)) ⊆ ker ΠM
x .
Theorem 12.5. [with assumptions as in 12.1] If Ga (M ) is Cohen-Macaulay then
(1) ΓM
x is an isomorphism of Rb (B)-modules.
(2) ΘM
x is an isomorphism of Gb (B)-modules.
M
It suffices to prove for r = 1 case. Clearly if ΓM
x is an isomorphism then Θx is
an isomorphism. The following observation enables us to analyze the case r = 1.
GORENSTEIN APPROXIMATION
33
Observation 12.6. Let x ∈ a. We have a commutative diagram of Ra (A)-modules
ΓM
x
R(FM , M ∗ )
/ R(FN , N ∗ )
ΨM
ΨN
*
Hom(R(a, M ), Ra (A))
/
*
ΠM
x
Hom(R(b, N ), Rb (B))
Here ΨM is as in 4.3. Notice that by 4.4, ΨM and ΨN are isomorphisms.
Remark 12.7. It is convenient to look at 12.6 in the following way:
1. By 12.2.1 and 12.2.2 we have a complex of Ra (A)-modules:
C:
ΓM
xt
0 → R(FM , M ∗ )(−1) −→ R(FM , M ∗ ) −−x→ R(FN , N ∗ ) → 0.
2. Set R(M ) = R(a, M ), R = Ra (A) and R = Rb (B). By 12.4 we have a complex
D of Ra (A)-modules:
0→
*
xt
HomR (R(M ), R)(−1) −→
*
ΠM
HomR (R(M ), R) −−x→
*
HomR (R(N ), R) → 0.
3. We consider C, D as co-chain complexes starting at 0; i.e.,
C:
0 → C0 → C1 → C2 → 0
0 → D0 → D1 → D2 → 0.
& D:
4. By 12.6, it follows that we have chain map of complexes of Ra (A)-modules
Ξ : C −→ D
i
with Ξ an isomorphism for i = 0, 1, 2.
5. It can be easily verified that C is exact if and only if D is exact.
Proof of Theorem 12.5. It suffices
L to prove for r = 1. Notice that (1) =⇒ (2).
(1) By assumption x is M
A-superficial. Set u = xt ∈ Ra (A)1 . Since Ga (A)
is Cohen-Macaulay we have x∗ is Ga (A)-regular. So u is Ra (A)-regular. It is
clear that Ra (A)/uRa (A) = Rb (B). Set R(M ) = R(a, M ), R = Ra (A). Also set
R = Rb (B) and R(N ) = R(b, M ). Since x is M -superficial we have an isomorphism
ϑM : R(M )/uR(M ) → R(N ). The exact sequence
u
ρ
0 −→ R(−1) −
→R−
→ R −→ 0
yields an exact sequence of R-modules
0→
→
*
u
HomR (R(M ), R)(−1) −
→
1
*
ExtR (R(M ), R)(−1)
*
ρM
HomR (R(M ), R) −−→
*
Hom(R(M ), R)
→ ···
1
Since R is Gorenstein and R(M ) is a MCM R-module we get * ExtR (R(M ), R) =
0. Also notice that * HomR (R(M ), R) = * HomR (R(M )/uR(M ), R). Furthermore the isomorphism ϑM induces an isomorphism
νM :
*
HomR (R(M )/uR(M ), R) →
ΠM
x .
*
HomR (R(N ), R).
It can be easily checked that νM ◦ ρM =
Therefore the complex D in 12.7.2 is exact. By 12.7.5 we get that the complex
M
M
C in 12.7.1 is exact. Thus ΓM
x is onto and ker Γx = uR(FM , M ). Therefore Γx is
an isomorphism.
34
TONY J. PUTHENPURAKAL
13. Some Preliminaries to prove Itoh’s conjecture
In [23] we introduced a new technique to investigate problems relating to associated graded modules. In this section we collect all the relevant results which
we proved in [23]. Throughout thus section (A, m) is a Noetherian local ring with
infinite residue field, M is a Cohen-Macaulay module of dimension r ≥ 1 and I is
an m-primary ideal.
13.1. Set L
S = A[It]; the Rees Algebra of I. In [23, 4.2] we proved that
LI (M ) = n≥0 M/I n+1 M is a S-module. Note that LI (M ) is not finitely generated S-module.
i
13.2. Set M = m ⊕ S+ . Let H i (−) = HM
denote the ith -local cohomology functor
with respect to M. Recall a graded S-module L is said to be *-Artinian if every
descending chain of graded submodules of L terminates. For example if E is a
finitely generated S-module then H i (E) is *-Artinian for all i ≥ 0.
13.3. In [23, 4.7] we proved that
H 0 (LI (M )) =
n+1 M
M I^
n≥0
I n+1 M
.
13.4. For LI (M ) we proved that for 0 ≤ i ≤ r − 1
(a) H i (LI (M )) are *-Artinian; see [23, 4.4].
(b) H i (LI (M ))n = 0 for all n ≫ 0; see [23, 1.10 ].
(c) H i (LI (M ))n has finite length for all n ∈ Z; see [23, 6.4].
(d) λ(H i (LI (M ))n ) coincides with a polynomial for all n ≪ 0; see [23, 6.4].
13.5. The natural maps 0 → I n M/I n+1 M → M/I n+1 M → M/I n M → 0 induce
an exact sequence of R(I)-modules
Π
0 −→ GI (M ) −→ LI (M ) −
→ LI (M )(−1) −→ 0.
(13.5.1)
We call (13.5.1) the first fundamental exact sequence. We use (13.5.1) also to relate
the local cohomology of GI (M ) and LI (M ).
13.6. Let x be M -superficial with respect to I and set N = M/xM and u = xt ∈
S1 . Notice LI (M )/uLI (M ) = LI (N ). Let I = (x1 , . . . , xm ). There exists ci ∈ A
Ps
Pl
such that x = i=1 ci xi . Set X = i=s ci Xi . For each n ≥ 1 we have the following
exact sequence of A-modules:
I n+1 M : x
M
M ψn
N
−−→ n+1
−→ n
−→ n+1 −→ 0,
I nM
I M
I
M
I
N
n
n+1
where ψn (m + I M ) = xm + I
M.
0 −→
This sequence induces the following exact sequence of S-modules:
(13.6.1)
Ψ
ρx
→ LI (M ) −→ LI (N ) −→ 0,
0 −→ B I (x, M ) −→ LI (M )(−1) −−X
where ΨX is left multiplication by X and
M (I n+1 M : M x)
.
B I (x, M ) =
I nM
n≥0
We call (13.6.1) the second fundamental exact sequence.
GORENSTEIN APPROXIMATION
35
13.7. Notice λ B I (x, M ) < ∞. A standard trick yields the following long exact
sequence connecting the local cohomology of LI (M ) and LI (N ):
0 −→ B I (x, M ) −→ H 0 (LI (M ))(−1) −→ H 0 (LI (M )) −→ H 0 (LI (N ))
−→ H 1 (LI (M ))(−1) −→ H 1 (LI (M )) −→ H 1 (LI (N ))
······
(13.7.1)
13.8. We will use the following well-known result regarding *-Artinian modules
quite often:
Let L be a *-Artinian S-module.
(a) If ψ : L(−1) → L is a monomorphism then L = 0.
(b) If φ : L → L(−1) is a monomorphism then L = 0.
13.9. Set L = LI (M ),
ξI (M ) :=
min { i | H i (L)−1 6= 0 or ℓ(H i (L)) = ∞}.
0≤i≤r−1
ampI (M ) := max{ |n| | H i (L)n−1 6= 0 for i = 0, . . . , ξI (M ) − 1}.
In [23, 7.5] we showed that
depth GI l (M ) = ξI (M ) for all l > ampI (M ).
13.10. Recall an ideal I is said to be asymptotically normal if I n is integrally closed
for all n ≫ 0. If I is a asymptotically normal m-primary ideal and dim A ≥ 2 then
by a result of Huckaba and Huneke [8, 3.8], depth GI l (A) ≥ 2 for all l ≫ 0(also see
[23, 7.3]). It also follows from [23, 9.2] that in this case H 1 (L)n = 0 for n < 0. In
particular ℓ(H 1 (L)) < ∞ (here L = LI (M )).
14. Proof of Itoh’s-conjecture
Theorem 14.1. Let (A, m) be a Cohen-Macaulay local ring of dimension d ≥ 1 and
let a be a normal m-primary ideal. If ea3 (A) = 0 then Ga (A) is Cohen-Macaulay.
Remark 14.2. The following assertion is well-known. If an is integrally closed for
all n then an is Ratliff-Rush for all n ≥ 1. It follows that depth Ga (A) ≥ 1.
We first show that
Lemma 14.3. Theorem 14.1 holds if dim A = 1, 2.
This Lemma is certainly well-known to the experts. However we provide a proof
for the convenience of the reader.
Proof of Lemma 14.3. If d = 1 then by 14.2 the result holds.
Now assume d = 2. Note that we may assume that A has an infinite residue
field. Let c be a minimal reduction of a. Set σi = λ(ai+1 /cai ). Then it follows from
a result of Huneke [9, 2.4] that the h-polynomial of A with respect to a is
h(z) = ℓ(A/a) + (σ0 − σ1 )z + (σ1 − σ2 )z 2 + · · · + (σs−2 − σs−1 )z s−1 + σs−1 z s .
36
TONY J. PUTHENPURAKAL
It follows that
ea1 (A) =
X
σi ,
i≥0
ea2 (A) =
X
iσi , and
i≥1
ea3 (A) =
X i
σi .
2
i≥2
As
= 0 we have σ2 = 0. So we have a3 = ca2 . As a is integrally closed we also
have a ∩ c = ca. The result follows from Valabrega-Valla Theorem [32, 2.3].
ea3 (A)
2
14.4. Recall that an ideal a is said to be asymptotically normal if an is integrally
closed for all n ≫ 0. The crucial result to prove Itoh’s conjecture is the following:
Lemma 14.5. Let (A, m) be a Cohen-Macaulay local ring of dimension 3 and let
aLbe an asymptotically normal m-primary ideal with ea3 (A) = 0. Set La (A) =
n+1
considered as a module over the Rees algebra S. Let M = m ⊕
n≥0 A/a
S+ be the maximal homogeneous ideal of S. Then the local cohomology modules
i
(La (A)) vanish for i = 1, 2
HM
We will also need to extend Lemma 14.5 to dimensions d ≥ 4.
14.6. Remark and a Convention: Note that all the relevant graded modules considered below are modules over the Rees algebra S = A[at]. Also all local cohomology
will taken over M = m⊕S+ the maximal homogeneous ideal of S. We also note that
if x is A-superficial with respect to A then the Rees algebra S ′ = A/(x)[a/(x)t] is
a quotient of S. As we are only interested in vanishing of certain local-cohomology
modules, by the independence theorem of local cohomology it does not matter if
we take local cohomology of an S ′ -module with respect to M′ or over M (and considering the module in question as an S-module). So throughout we will only write
i
H i (−) to mean HM
(−).
Lemma 14.7. Let (A, m) be a Cohen-Macaulay local ring of dimension d ≥ 3 and
let a be an asymptotically normal m-primary ideal with ea3 (A) = 0. Then the local
i
cohomology modules HM
(La (A)) vanish for 1 ≤ i ≤ d − 1.
Proof. We prove the result by induction on d ≥ 3. For d = 3 the result follows from
Lemma 14.5. Now assume d ≥ 4 and the result has been proved for d − 1. After
passing through a general extension (see 2.7(d)) we may choose x, an A-superficial
element with respect to a such that in the ring B = A/(x) the ideal b = a/(x)
is asymptotically normal. Furthermore note that eb3 (B) = 0. Set L = La (A) and
L = Lb (B). By induction hypothesis we have H i (L) = 0 for 1 ≤ i ≤ d − 2. By
13.7.1 we get a surjective map H 1 (L)(−1) → H 1 (L) and for 2 ≤ i ≤ d− 1 injections
H i (L)(−1) → H i (L). By 13.8 it follows that H i (L) = 0 for i ≥ 2. By 13.10 we get
that H 1 (L) has finite length. So the surjection H 1 (L)(−1) → H 1 (L) induces an
isomorphism H 1 (L)(−1) ∼
= H 1 (L) and this forces H 1 (L) = 0.
We now give a proof of Theorem 14.1 assuming Lemma 14.5.
Proof of Theorem 14.1. If d = dim A = 1, 2, the result follows from Lemma 14.3.
i
If d ≥ 3 then by Lemma 14.7 we get HM
(La (A)) = 0 for 1 ≤ i ≤ d − 1. Also
n
as a is normal, in-particular a is Ratliff-Rush for all n ≥ 1. So by 13.3 we get
GORENSTEIN APPROXIMATION
37
0
HM
(La (A)) = 0. By taking cohomology of the first fundamental sequence 13.5.1
i
we get that HM
(GI (A)) = 0 for 0 ≤ i ≤ d − 1. Thus GI (A) is Cohen-Macaulay.
14.8. Thus to prove Itoh’s conjecture all we have to do is to prove Lemma 14.5.
This requires several preparatory results. For the rest of the section we will assume
dim A = 3 and a is an asymptotically normal ideal with ea3 (A) = 0. We first show
Lemma 14.9. (with hypotheses as in 14.8.) Then Gan (A) is Cohen-Macaulay for
n ≫ 0 and red(an ) ≤ 2 for n ≫ 0.
Remark 14.10. Lemma 14.9 is certainly known to the experts. We provide a
proof for the convenience of the reader.
Proof of Lemma 14.9. We may assume that the residue field of A is infinite. Let
c = (x1 , x2 , x3 ) be a minimal reduction of a.
n
Notice that for all n ≥ 1 we have ea3 (A) = ea3 (A) = 0. Furthermore c[n] =
(xn1 , xn2 , xn3 ) is a minimal reduction of an . By a result of Huckaba and Huneke [8,
3.8] we have depth Gan (A) ≥ 2 for all n ≫ 0, say n ≥ n0 . As a is asymptotically
normal we may also assume an is integrally closed for all n ≥ n0
Fix n ≥ n0 . Set σi = ℓ((an )i+1 /c[n] (an )i ). Then by a result of Marley [12, 3.9],
it follows that that the h-polynomial of A with respect to an is
h(z) = ℓ(A/an ) + (σ0 − σ1 )z + (σ1 − σ2 )z 2 + · · · + (σs−2 − σs−1 )z s−1 + σs−1 z s .
It follows that
n
ea1 (A) =
X
σi ,
i≥0
n
ea2 (A) =
X
iσi , and
i≥1
n
ea3 (A) =
X i
σi .
2
i≥2
n
As ea3 (A) = 0 we have σ2 = 0. So we have (an )3 = c[n] (an )2 . As an is integrally
closed we also have (an )2 ∩ c[n] = c[n] an . So by Valabrega-Valla Theorem [32, 2.3]
we have that Gan (A) is Cohen-Macaulay.
Next we show
Lemma 14.11. (with hypotheses as in 14.8.) We have
(1) a(Ga (A)) < 0.
(2) H 2 (La (A)) = 0.
P2
i
i
(3)
i=0 (−1) ℓ(H (Ga (A))) = 0.
(4) For i = 0, 1, 2 we have H i (Ga (A))n = 0 for n < 0.
Proof. (1) We have redc[n] (an ) ≤ 2 for all n ≫ 0. By a result of Hoa,[10, 2.6], the
result follows.
(2) Set G = Ga (A) and L = La (A). Also let S = A[at] be the Rees algebra of
A with respect to a and let M be its maximal homogeneous ideal. Throughout let
i
(−). The first fundamental exact sequence 0 → G → L → L(−1) → 0
H i (−) = HM
yields an exact sequence in cohomology
H 2 (L)n → H 2 (L)n−1 → H 3 (G)n .
38
TONY J. PUTHENPURAKAL
As a(G) < 0 we get for n ≥ 0, surjections H 2 (L)n → H 2 (L)n−1 . As H 2 (L)n = 0
for n ≫ 0 we get H 2 (L)n = 0 for n ≥ −1.
After passing through a general extension we may choose x, an A-superficial
element with respect to a such that in the ring B = A/(x) the ideal b = a/(x) is
asymptotically normal. Also notice dim B = 2. Set L = Lb (B). By [23, 9.2] we get
that H 1 ((L))n = 0 for n < 0. By 13.7.1 we have an exact sequence
H 1 (L)n → H 2 (L)n−1 → H 2 (L)n
By setting n = −1 we get H 2 (L)−2 = 0. Iterating we get H 2 (L)n = 0 for all
n ≤ −2.
It follows that H 2 (L) = 0.
(3) We first note that as Gan (A) is Cohen-Macaulay for all n ≫ 0 the ring Ga (A)
is generalized Cohen-Macaulay. We also have H 0 (L) and H 1 (L) have finite length
(see 13.3 and 13.10.
The first fundamental exact sequence 0 → G → L → L(−1) → 0 yields an exact
sequence in cohomology
0 → H 0 (G) → H 0 (L) → H 0 (L)(−1)
→ H 1 (G) → H 1 (L) → H 1 (L)(−1)
→ H 2 (G) → H 2 (L) = 0.
Taking lengths, the result follows.
(4) As a is normal we have that H 1 (L)n = 0 for n < 0 (see [23, ?]). Also by 13.3
we get H 0 (L)n = 0 for n < 0. The result follows from the above exact sequence in
cohomology.
Remark 14.12. Till now we have not used our theory of Gorenstein approximation. We do it now. We may after taking a general extension of A and completing
it we may assume that (A, m) is a complete Cohen-Macaulay local ring of dimension three, a is an asymptotically normal ideal in A and that there exists x, an
A-superficial element with respect to a such that in the ring B = A/(x) the ideal
b = a/(x) is asymptotically
Lnormal. Let [T, t, q, ψ] be a Gorenstein approximation
of [A, m, a]. Let Rq (T ) = n∈Z qn be the extended Rees ring of T with respect to
q. Note that as Gq (T ) is Gorenstein we also have Rq (T ) is a Gorenstein ring. Let
N be its maximal homogeneous ideal. Also let Ra (A) be the extended Rees ring of
A with respect to a.
We first show
Lemma 14.13. (with hypotheses as in 14.12.) We have
(1) For any prime P in Rq (T ) with ht P ≤ 3 we have Ra (A)P is Cohen-Macaulay.
i
(2) HN
(Ra (A)) has finite length for 0 ≤ i ≤ 3.
Proof. As Rq (T ) is a Gorenstein ring the assertion (2) follows from (1) by local
duality.
(1) We first consider the case t−1 ∈
/ P . Then (Ra (A))P is a localization of
(Ra (A))t−1 = A[t, t−1 ] which is Cohen-Macaulay.
Next consider the case when t−1 ∈ P . Set Q = P/(t−1 ) a prime ideal of height
≤ 2 in Gq (T ). As Ga (A) is generalized Cohen-Macaulay and since Gq (T ) is Gorenstein, by local duality we have that (Ga (A))Q is Cohen-Macaulay. It follows that
(Ra (A))P is Cohen-Macaulay.
GORENSTEIN APPROXIMATION
39
As a consequence we get
Corollary 14.14. (with hypotheses as in 14.12.) We have an exact sequence
t−1
3
4
4
0 → HN
(Ga (A)) → HN
(Ra (A))(+1) −−→ HN
(Ra (A)) → 0.
t−1
4
Proof. The short exact sequence 0 → Ra (+1) −−→ HN
(Ra ) → Ga (A) → 0, yields
an exact sequence of local cohomology modules
t−1
0
0
0
0 → HN
(Ra (A))(+1) −−→ HN
(Ra (A)) → HN
(Ga (A))
t−1
1
1
1
→ HN
(Ra (A))(+1) −−→ HN
(Ra (A)) → HN
(Ga (A))
t−1
π
2
2
2
→ HN
(Ra (A))(+1) −−→ HN
(Ra (A)) −
→ HN
(Ga (A))
t−1
3
3
(Ra (A))
→ HN
(Ra (A))(+1) −−→ HN
i
Let K be the cokernel of π. We note that HN
(Ga (A)) ∼
= H i (Ga (A)). As all the
module in the above exact sequence we have
ℓ(K) =
2
X
(−1)i ℓ(H i (Ga (A)) = 0
by Lemma 14.11(3).
i=0
Thus we have an inclusion
t−1
3
3
HN
(Ra (A))(+1) −−→ HN
(Ra (A))
3
3
3
As ℓ(HN
(Ra (A))) is finite we have an isomorphism HN
(Ra (A))(+1) ∼
(Ra (A)).
= HN
3
3
Again as ℓ(HN (Ra (A))) is finite this implies that HN (Ra (A)) = 0. The result
follows.
The following result is a crucial ingredient in proving Lemma 14.5.
Theorem 14.15. (with hypotheses as in 14.12.) Set ωA = HomT (A, T ) be the
canonical module of A and let EG be the injective hull of k considered as a Gq (T )module. Set W = HomGq (T ) (H 3 (Ga (A)), EG ). Then
(1) there exists an a-good filtration F on ωA such that GF (ω) ∼
= W upto a shift.
(2) W is Cohen-Macaulay.
Proof. (1)Let ER be the injective hull of k considered as a Rq (T )-module. Set
(−)∨ = HomRq (T ) (−, ER ). Dualizing the exact sequence in Lemma 14.14 we get a
sequence of Rq (T )-modules
t−1
4
4
(Ra (A))∨ (+1) → W → 0.
0 → HN
(Ra (A))∨ −−→ HN
As Rq (T ) is Gorenstein then by local duality we have an exact sequence
t−1
0 → HomRq (T ) (Ra (A), Rq (T )) −−→ HomRq (T ) (Ra (A), Rq (T ))(+1) → X → 0
where X ∼
= W upto shift. By Theorem 4.4 there exists an q-good filtration F on
ωA such that
R(F, ωA ) ∼
= HomRq (T ) (Ra (A), Rq (T ))
Also notice that as qA = a it follows that F is in fact an a-good filtration on ωA .
40
TONY J. PUTHENPURAKAL
(2) To prove this result we first L
Claim: The Hilbert series of W = n∈Z Wn is
HW (z) =
X
ℓ(Wn ) =
n∈Z
a1 z + a2 z 2
.
(1 − z)3
We first assume the claim and prove (2). By (1) there exists an a-good filtration G
on ωA such that GG (ωA ) ∼
= W (+1) We note that the Hilbert series of GG (ωA ) is
a1 + a2 z
.
(1 − z)3
Then by a result of Rossi and Valla, [26, 2.4.9] it follows that GG (ωA ) is CohenMacaulay. So W is Cohen-Macaulay.
It remains to prove the claim. Set G = Ga (A). Let V = H 3 (G). As a(G) < 0
we have that Vn = 0 for n ≥ 0. So Wm = 0 for m ≤ 0. By Grothendieck-Serre
formula we have
3
X
(−1)i ℓ(H i (G)n
Ha (n) − Pa (n) =
i=0
Where Ha (and Pa (z)) are Hilbert function(and Hilbert polynomial) of A with
respect to a. As H i (G)n = 0 for n < 0 for i = 0, 1, 2 we get that
ℓ(Vn ) = Pa (n)
for all n < 0.
It follows that ℓ(Wm ) = Pa (−m) for all m ≥ 1. So Pa (−z) is the Hilbert polynomial
of W (it should be remarked that deg(Pa (z)) = 2). As postulation number of W is
one it follows from [2, 4.1.12] the Hilbert series of W is
a0 + a1 z + a2 z 2
(1 − z)3
As W0 = 0 we get that a0 = 0. The result follows.
We finally give
Proof of Lemma 14.5. We may after taking a general extension of A and completing
it we may assume that (A, m) is a complete Cohen-Macaulay local ring of dimension three, a is an asymptotically normal ideal in A and that there exists x, an
A-superficial element with respect to a such that in the ring B = A/(x) the ideal
b = a/(x) is asymptotically normal. Set G = Ga (A), L = La (A), G = Gb (B) and
L = Lb (B).
We make the following
Claim-1: a(G) ≤ 0.
Claim-2: H 1 (L) = 0.
Let us now assume Claims 1, 2 for the moment. We prove our result as follows.
By 13.7.1 we have an exact sequence
H 1 (L)(−1) → H 1 (L) → H 1 (L) = 0.
As a is asymptotically normal we have by 13.10 that H 1 (L) has finite length.
Thus the surjective map H 1 (L)(−1) → H 1 (L) induces an isomorphism H 1 (L) ∼
=
H 1 (L)(−1). As H 1 (L) has finite length we get that H 1 (L) = 0. Thus all we have
to prove is Claims 1 and 2.
GORENSTEIN APPROXIMATION
41
We prove Claim-2 assuming Claim-1. As b is asymptotically normal we have
H 1 (L)n = 0 for n < 0. The first fundamental exact sequence 0 → G → L →
L(−1) → 0 yields an exact sequence in cohomology
H 1 (L)n → H 1 (L)n−1 → H 2 (G)n .
As a(G) ≤ 0 we get for n ≥ 1, surjections H 1 (L)n → H 1 (L)n−1 . As H 1 (L)n = 0
for n ≫ 0 we get H 1 (L)n = 0 for n ≥ 0. Thus H 1 (L) = 0. Thus we have proved
Claim-2 assuming Claim-1.
Proof of Claim-1: There exists exact sequences
x∗
(14.15.1)
0 → U → G(−1) −→ G → G/x∗ G → 0
(14.15.2)
0 → V → G/x∗ G → G → 0
where U, V are finite length G-modules. By 14.15.2 we get isomorphism’s H i (G) ∼
=
H i (G/x∗ G) for i > 0. By 14.15.1 we have an exact sequence
x∗
π
H 2 (G)(−1) −→ H 2 (G) −
→ H 2 (G) →
x∗
→ H 3 (G)(−1) −→ H 3 (G) → 0.
Set C = image π and K = coker π. As H 2 (G) has finite length we have that C has
finite length.
Let EG be the injective hull of k as G-module. Set (−)∨ = HomG (−, EG ). By
14.15 H 3 (G)∨ is Cohen-Macaulay of dimension 3. So K ∨ is Cohen-Macaulay of
dimension 2. Also note that H 2 (G)∨ has depth 2. The exact sequence 0 → C →
H 2 (G) → K → 0 induces an exact sequence 0 → K ∨ → H 2 (G)∨ → C ∨ → 0. As
∨
depth(K ∨ ⊕ G ) ≥ 2 we get that depth C ∨ ≥ 1. But C ∨ has finite length. So C ∨
(and hence C) is zero. Thus we have an exact sequence
x∗
0 → H 2 (G) → H 3 (G)(−1) −→ H 3 (G) → 0.
As H 3 (G)n = 0 for n < 0 we get that H 2 (G)n = 0 for n ≤ 0. The result follows.
Acknowledgments: I thank Lucho Avramov, Juergen Herzog, Srikanth Iyengar
and Jugal Verma for many useful discussions on the subject of this paper.
References
[1] M.P. Brodmann and R.Y. Sharp, Local cohomology: an algebraic introduction with geometric
applications, Cambridge Studies in Advanced Mathematics, vol. 60, Cambridge University
Press, Cambridge, 1998. MR MR1613627
[2] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, vol. 39, Cambridge University Press, Cambridge,(Revised edition) 1997. MR MR1251956
[3] A. Capani, G. Niesi, and L. Robbiano, CoCoA,a system for doing Computations in Commutative Algebra, 1995, available via anonymous ftp from cocoa.dima.unige.it.
[4] T. Cortadellas and S. Zarzuela, On the depth of the fiber cone of filtrations, J. Algebra 198
(1997), no. 2, 428–445. MR MR1489906
[5] D. R. Grayson and M. E. Stillman, Macaulay 2, a software system for research in algebraic
geometry, Available at http://www.math.uiuc.edu/Macaulay2/.
[6] U. Grothe, M. Herrmann, and U. Orbanz, Graded Cohen-Macaulay rings associated to equimultiple ideals, Math. Z. 186 (1984), 531–556. MR MR744964
[7] A. Guerrieri, On the depth of the associated graded ring of an m-primary ideal of a CohenMacaulay local ring, J. Algebra 167 (1994), 745–757. MR MR1287068
[8] S. Huckaba, and C. Huneke, Normal ideals in regular rings, J. Reine Angew. Math. 510
(1999), 63-82. MR MR1696091
42
TONY J. PUTHENPURAKAL
[9] C. Huneke, Hilbert functions and symbolic powers, Michigan Math. J. 34 (1987), no. 2, 293318. MR MR0894879
[10] L. T. Hoa, Reduction numbers and Rees algebras of powers of an ideal, Proc. Amer. Math.
Soc. 119 (1993), no. 2, 415–422. MR MR1152984
[11] S. Itoh, Coefficients of normal Hilbert polynomials, J. Algebra 150 (1992) 101-117.
MR MR1174891
[12] T. Marley, The coefficients of the Hilbert polynomial and the reduction number of an ideal,
J. London Math. Soc. (2) 40 (1989), no. 1, 1–8. MR MR1028910
, The reduction number of an ideal and the local cohomology of the associated graded
[13]
ring, Proc. Amer. Math. Soc. 117 (1993), no. 2, 335–341. MR MR1112496
[14] H. Matsumura, Commutative ring theory, Cambridge Studies in Advanced Mathematics,
vol. 8, Cambridge University Press, Cambridge, 1986, Translated from the Japanese by M.
Reid. MR MR879273
[15] S. Molinelli and G. Tamone, On the Hilbert function of certain rings of monomial curves, J.
Pure Appl. Algebra 101 (1995), no. 2, 191–206. MR MR1348035
[16] M. Nagata, Local rings, Interscience Tracts in Pure and Applied Mathematics, No. 13, Interscience Publishers a division of John Wiley & Sons New York-London, 1962. MR MR0155856
[17] M. Narita, A note on the coefficients of Hilbert characteristic functions in semi-regular local
rings, Proc. Cambridge Philos. Soc. 59 (1963), 269–275. MR MR0146212
[18] D. G. Northcott, A note on the coefficients of the abstract Hilbert function, J. London Math.
Soc. 35 (1960) 209-275. MR MR0110731
[19] A. Ooishi, On the associated graded modules of canonical modules, J. Algebra 141 (1991),
no. 1, 143–157. MR MR1118320
, On the Gorenstein property of the associated graded ring and the Rees algebra of an
[20]
ideal, J. Algebra 155 (1993), no. 2, 397–414. MR MR1212236
[21] T. J. Puthenpurakal, Hilbert-coefficients of a Cohen-Macaulay module, J. Algebra 264 (2003),
no. 1, 82–97. MR MR1980687
[22]
, The Hilbert function of a maximal Cohen-Macaulay module, Math. Z. 251 (2005),
551–573. MR MR2190344
, Ratliff-Rush filtration, regularity and depth of higher associated graded modules I,
[23]
J. Pure Appl. Algebra 208 (2007), no. 1, 159-176. MR MR2269837
[24]
, A short note on the non-negativity of partial Euler characteristics, Beiträge Algebra
Geom. 46 (2005), 559–560. MR MR2196937
[25] D. Rees, A theorem of homological algebra, Proc. Cambridge Philos. Soc. 52 (1956), 605–610.
MR MR0080653
[26] M. E. Rossi and G. Valla, Hilbert functions of filtered modules, Lecture Notes of the Unione
Matematica Italiana, 9. Springer-Verlag, Berlin; UMI, Bologna, 2010. MR MR2723038
[27] J. D. Sally, Tangent cones at Gorenstein singularities, Compositio Math. 40 (1980), no. 2,
167–175. MR MR563540
[28] J-P. Serre, Local algebra, Springer Monographs in Mathematics, Springer-Verlag, Berlin, 2000,
Translated from the French by CheeWhye Chin and revised by the author. MR MR1771925
[29] B. Singh, Effect of a permissible blowing-up on the local Hilbert functions, Invent. Math. 26
(1974), 201–212. MR MR0352097
[30] R. P. Stanley, Hilbert functions of graded algebras, Advances in Math. 28 (1978), no. 1, 57–83.
MR MR0485835
[31] N. V. Trung, Reduction exponent and degree bound for the defining equations of graded rings,
Proc. Amer. Math. Soc. 101 (1987), no. 2, 229–236. MR MR902533
[32] P. Valabrega and G. Valla, Form rings and regular sequences, Nagoya Math. J. 72 (1978),
93–101. MR MR514892
[33] H. Wang, Hilbert coefficients and the associated graded rings, Proc. Amer. Math. Soc. 128
(2000), 963–973. MR MR1628432
Department of Mathematics, Indian Institute of Technology, Bombay, Powai, Mumbai 400 076, India
E-mail address: [email protected]
| 0math.AC
|
Algorithms for `p Low-Rank Approximation
arXiv:1705.06730v1 [cs.DS] 18 May 2017
Flavio Chierichetti∗
Sapienza University
Rome, Italy
[email protected]
Silvio Lattanzi
Google
New York, NY
[email protected]
Sreenivas Gollapudi
Google
Mountain View, CA
[email protected]
Rina Panigrahy
Google
Mountain View, CA
[email protected]
Ravi Kumar
Google
Mountain View, CA
[email protected]
David P. Woodruff
IBM
Almaden, CA
[email protected]
Abstract
We consider the problem of approximating a given matrix by a low-rank matrix so as to minimize the entrywise
`p -approximation error, for any p ≥ 1; the case p = 2 is the classical SVD problem. We obtain the first provably good
approximation algorithms for this version of low-rank approximation that work for every value of p ≥ 1, including
p = ∞. Our algorithms are simple, easy to implement, work well in practice, and illustrate interesting tradeoffs
between the approximation quality, the running time, and the rank of the approximating matrix.
1
Introduction
The problem of low-rank approximation of a matrix is usually studied as approximating a given matrix by a matrix of
low rank so that the Frobenius norm of the error in the approximation is minimized. The Frobenius norm of a matrix
is obtained by taking the sum of the squares of the entries in the matrix. Under this objective, the optimal solution is
obtained using the singular value decomposition (SVD) of the given matrix. Low-rank approximation is useful in large
data analysis, especially in predicting missing entries of a matrix by projecting the row and column entities (e.g., users
and movies) into a low-dimensional space. In this work we consider the low-rank approximation problem, but under the
general entrywise `p norm, for any p ∈ [1, ∞].
There are several reasons for considering the `p version of low-rank approximation instead of the usually studied
`2 (i.e., Frobenius) version. For example, it is widely acknowledged that the `1 version is more robust to noise and
outliers than the `2 version [2, 13, 24]. Several data mining and computer vision-related applications exploit this insight
and resort to finding a low-rank approximation to minimize the `1 error [15, 16, 20, 22]. Furthermore, the `1 error is
typically used as a proxy for capturing sparsity in many applications including robust versions of PCA, sparse recovery,
and matrix completion; see, for example [2, 23]. For these reasons the problem has already received attention [10] and
was suggested as an open problem by Woodruff in his survey on sketching techniques for linear algebra [21]. Likewise,
the `∞ version (dubbed also as the Chebyshev norm) has been studied for the past many years [11, 12], though to the
best of our knowledge, no result with theoretical guarantees was known for `∞ before our work. Our algorithm is also
quite general, and works for every p ≥ 1.
Working with `p error, however, poses many technical challenges. First of all, unlike `2 , the general `p space is not
amenable to spectral techniques. Secondly, the `p space is not as nicely behaved as the `2 space, for example, it lacks
the notion of orthogonality. Thirdly, the `p version quickly runs into computational complexity barriers: for example,
even the rank-1 approximation in `1 has been shown to be NP-hard by Gillis and Vavasis [10]. However, there has been
∗ Work done in part while visiting Google. Supported in part by a Google Focused Research Award, by the ERC Starting Grant DMAP 680153,
and by the SIR Grant RBSI14Q743.
no dearth in terms of heuristics for the `p low-rank approximation problem, in particular for p = 1 and p = ∞: this
includes alternating convex (and, in fact, linear) minimization [14], methods based on expectation-maximization [19],
minimization with augmented Lagrange multipliers [26], hyperplanes projections and linear programming [1], and
generalizations of the Wiberg algorithm [6]. These heuristics, unfortunately, do not come with any performance
guarantees. Polynomial-time algorithms for the general problem of rank-k approximation has been stated as an open
problem [21]. While theoretical approximation guarantees have been given for the rank-1 version for the GF (2) and
the Boolean cases [3], to the best of our knowledge there have been no provably good (approximation) algorithms for
general matrices, or for rank more than one, or for general `p .
1.1
Our Contributions
In this paper we obtain the first provably good algorithms for the `p rank-k approximation problem for every p ≥ 1. Let
n × m be the dimensions of the input matrix. From an algorithmic viewpoint, there are three quantities of interest: the
running time of the algorithm, the approximation factor guaranteed by the algorithm, and the actual number of vectors
in the low-rank approximation that is output by the algorithm (even though we only desire k).
Given this setting, we show three main algorithmic results intended for the case when k is not too large. First,
we show that one can obtain a (k + 1)-approximation to the rank-k problem in time mk poly(n, m); note that this
running time is not polynomial once k is larger than a constant. To address this, next we show that one can get an
O(k)-approximation to the best k-factorization in time O(poly(nm)); however, the algorithm returns O(k log m)
columns, which is more than the desired k (this is referred to as a bi-criteria approximation). Finally, we combine
these two algorithms. We first show that the output of the second algorithm can further be refined to output exactly k
vectors, with an approximation factor of poly(k) and a running time of O(poly(n, m)(k log n)k ). The running time
now is polynomial as long as k = O(log n/ log log n). Next, we show that for any constant p ≥ 1, we can obtain
approximation factor (k log m)O(p) and a running time of poly(n, m) for every value of k.
Our first algorithm is existential in nature: it shows that there are k columns in the given matrix that can be
used, along with an appropriate convex program, to obtain a (k + 1)-approximation. Realizing this as an algorithm
would therefore naïvely incur a factor mk in the running time. Our second algorithm works by sampling columns and
iteratively “covering” the columns of the given matrix, for an appropriate notion of covering. In each round of sampling
our algorithm uniformly samples from a remaining set of columns; we note here that it is critical that our algorithm
is adaptive as otherwise uniform sampling would not work. While this is computationally efficient and maintains an
O(k)-approximation to the best rank-k approximation, it can end up with more than k columns, in fact O(k log m). Our
third algorithm fixes this issue by combining the first algorithm with the notion of a near-isoperimetric transformation
for the `p -space, which lets us transform a given matrix into another matrix spanning the same subspace but with small
`p distortion.
A useful feature of our algorithms is that they are uniform with respect to all values of p. We test the performance of
our algorithms, for p = 1 and p = ∞, on real and synthetic data and show that they produce low-rank approximations
that are substantially better than what the SVD (i.e., p = 2) would obtain.
1.2
Related Work
In [18], a low-rank approximation was obtained which holds for every p ∈ [1, 2]. Their main result is an (O(log m) poly(k))approximation in nnz(A) + (n + m) poly(k) time, for every k, where nnz(A) is the number of non-zero entries in A.
In our work, we also obtain such a result for p ∈ [1, 2] via very different sampling-based methods, whereas the results
in [18] are sketching-based. In addition to that, we obtain an algorithm with a poly(k) approximation factor which is
independent of m and n, though this latter algorithm requires k = O(log n/ log log n) in order to be polynomial time.
Another result in [18] shows how to achieve a k poly(log k)-approximation, in nO(k) time for p ∈ [1, 2]. For k
larger than a constant, this is larger than polynomial time, whereas our algorithm with poly(k) approximation ratio is
polynomial time for k as large as Θ(log n/ log log n).
Importantly, our results hold for every p ≥ 1, rather than only p ∈ [1, 2], so for example, include p = ∞.
In addition we note that there exist papers solving problems that, at first blush, might seem similar to ours. For
instance, [5] study a convex relation, and a rounding algorithm to solve the subspace approximation problem (an `p
generalization of the least squares fit), which is related to but different from our problem. Also, [7] offer a bi-criteria
1
solution for another related problem of approximating a set of points by a collection of flats; they use convex relaxations
to solve their problem and are limited to bi-criteria solutions, unlike ours. Finally, in some special settings robust
PCA can be used to solve `1 low-rank approximation [2]. However, robust PCA and `1 low-rank approximation have
some apparent similarities but they have key differences. Firstly, `1 low-rank approximation allows to recover an
approximating matrix of any chosen rank, whereas RPCA returns some matrix of some unknown (possibly full) rank.
While variants of robust PCA have been proposed to force the output rank to be a given value [17, 25], these variants
make additional noise model and incoherence assumptions on the input matrix, whereas our results hold for every
input matrix. Secondly, in terms of approximation quality, it is unclear if near-optimal solutions of robust PCA provide
near-optimal solutions for `1 low-rank approximation.
Finally, we mention concrete example matrices A for which the SVD gives a poor approximation factor for `p approximation error. First, suppose p < 2 and k = 1. Consider the following n × n block diagonal matrix composed of
two blocks: a 1 × 1 matrix with value n and an (n − 1) × (n − 1) matrix with all 1s. The SVD returns as a solution the
first column, and therefore incurs polynomial in n error for p = 2 − Ω(1). Now suppose p > 2 and k = 1. Consider the
following n × n block diagonal matrix composed of two blocks: a 1 × 1 matrix with value n − 2 and an (n − 1) × (n − 1)
matrix with all 1s. The SVD returns as a solution the matrix spanned by the bottom block, and so also incurs an error
polynomial in n for p = 2 + Ω(1).
2
Background
For a matrix M , let Mi,j denote the entry in its ith row and jth column and let Mi denote its ith column. Let M T
P
1/p
p
denote its transpose and let |M |p =
denote its entry-wise p norm. Given a set S = {i1 , . . . , it } of
i,j |Mi,j |
column indices, let MS = Mi1 ,...,it be the matrix composed of the
of M with the indices in S.
Pcolumns
m
Given a matrix M with m columns, we will use span M = { i=1 αi Mi | αi ∈ R} to denote the vectors spanned
by its columns. If M is a matrix and v is a vector, we let dp (v, M ) denote the minimum `p distance between v and a
vector in span M :
dp (v, M ) =
inf
|v − w|p .
w∈span M
Let A ∈ Rn×m denote the input matrix and let k > 0 denote the target rank. We assume, without loss of generality
(w.l.o.g.), that m ≤ n. Our first goal is, given A and k, to find a subset U ∈ Rn×k of k columns of A and V ∈ Rk×m
so as to minimize the `p error, p ≥ 1, given by
|A − U V |p .
Our second goal is, given A and k, to find U ∈ Rn×k , V ∈ Rk×m to minimize the `p error, p ≥ 1, given by
|A − U V |p .
Note that in the second goal, we do not require U be a subset of columns.
We refer to the first problem as the k-columns subset selection problem in the `p norm, denoted k-CSSp , and to the
second problem as the rank-k approximation problem in the `p norm, denoted k-LRAp .1 In the paper we often call
U, V the k-factorization of A. Note that a solution to k-CSSp can be used as a solution to k-LRAp , but not necessarily
vice-versa.
In this paper we focus on solving the two problems for general p. Let U ? V ? be a k-factorization of A that is optimal
in the `p norm, where U ? ∈ Rn×k and V ? ∈ Rk×m , and let optk,p (A) = |A − U ? V ? |p . An algorithm is said to be an
α-approximation, for an α ≥ 1, if it outputs U ∈ Rn×k , V ∈ Rk×m such that
|A − U V |p ≤ α · optk,p (A).
It is often convenient to view the input matrix as A = U ? V ? + ∆ = A? + ∆, where ∆ is some error matrix of
minimum `p -norm. Let δ = |∆|p = optk,p (A).
We will use the following observation.
1 k-LRA
2
is the classical SVD problem of finding U ∈ Rn×k , V ∈ Rk×m so as to minimize |A − U V |2 .
2
Lemma 1. Let U ∈ Rn×k and v ∈ Rn×1 . Suppose that there exists x ∈ Rk×1 such that δ = |U · x − v|p . Then, there
exists a polynomial time algorithm that, given U and v, finds y ∈ Rk×1 such that |U · y − v|p ≤ δ.
Proof. This `p regression problem is a convex program and well-known to be solvable in polynomial time.
3
An (mk poly(nm))-time algorithm for k-LRAp
In this section we will present an algorithm that runs in time mk poly(nm) and produces a (k + 1)-approximation
to k-CSSp (so also to k-LRAp ) of a matrix A ∈ Rn×m , m ≤ n, for any p ∈ [1, ∞]. The algorithm simply tries all
possible subsets of k columns of A for producing one of the factors, U , and then uses Lemma 1 to find the second factor
V.
3.1
The existence of one factor in A
For simplicity, we assume that |∆i |p > 0 for each column i. To satisfy this, we can add an arbitrary small random error
to each entry of the matrix. For instance, for any γ > 0, and to each entry of the matrix, we can add an independent
uniform value in [−γ, γ]. This would guarantee that |∆i |p > 0 for each i ∈ [m].
Recall that A = A? + ∆ is the perturbed matrix, and we only have access to A, not A? . Consider Algorithm 1
and its output S. Note that we cannot actually run this algorithm since we do not know A? . Thus, it is a hypothetical
algorithm used for the purpose of our proof. That is, the algorithm serves as a proof that there exists a subset of k
columns of A providing a good low rank approximation. In Theorem 3 we prove that the columns in A indexed by the
subset S can be used as one factor of a k-factorization of A.
Algorithm 1 Enumerating and selecting k columns of A.
Require: A rank k matrix A? and perturbation matrix ∆
Ensure: k column indices of A?
?
1: For each column index i, let Ã?
i ← Ai /|∆i |p .
2: Write Ã? = Ũ · Ṽ , s.t. Ũ ∈ Rn×k , Ṽ ∈ Rk×m .
3: Let S be the subset of k columns of Ṽ ∈ Rk×m that has maximum determinant in absolute value (note that the
subset S indexes a k × k submatrix).
4: Output S.
Before proving the main theorem of the subsection, we show a useful property of the matrix Ã? , that is, the matrix
having the vector A?i /|∆i |p as the ith column. Then we will use this property to prove Theorem 3.
P
Lemma 2. For each column Ã?i of Ã? , one can write Ã?i = j∈S Mi (j)Ã?j , where |Mi (j)| ≤ 1 for all i, j.
Proof. Fix an i ∈ {1, . . . , m}. Consider the equation ṼS Mi = Ṽi for Mi ∈ Rk . We can assume the columns in ṼS are
linearly independent, since w.l.o.g., Ã? has rank k. Hence, there is a unique solution Mi = (ṼS )−1 Ṽi . By Cramer’s
rule, the jth coordinate Mi (j) of Mi satisfies Mi (j) =
det(ṼSj )
,
det(ṼS )
where ṼSj is the matrix obtained by replacing the jth
column of ṼS with Ṽi . By our choice of S, | det(ṼSj )| ≤ | det(ṼS )|, which implies |Mi (j)| ≤ 1. Multiplying both
sides of equation ṼS Mi = Ṽi by Ũ , we have Ã?S Mi = Ã?i .
Now we prove the main theorem of this subsection.
Theorem 3. Let U = AS . For p ∈ [1, ∞], let M1 , . . . , Mm be the vectors whose existence is guaranteed by Lemma 2
T
and let V ∈ Rk×n be the matrix having the vector |∆i |p · (Mi (1)/|∆i1 |p , . . . , Mi (k)/|∆ik |p ) as its ith column. Then,
|Ai − (U V )i |p ≤ (k + 1)|∆i |p and hence |A − U V |p ≤ (k + 1)|∆|p .
3
Proof. We consider the generic column (U V )i .
(U V )i = |∆i |p
k
X
Mi (j)
|∆ij |p
j=1
=
|∆i |p
k
X
Mi (j)
|∆ij |p
j=1
Ai j
A?ij + ∆ij
k
X
∆ij
Mi (j)Ã?ij + Mi (j)
|∆ij |p
j=1
=
|∆i |p
=
|∆i |p Ã?i + |∆i |p
k
X
∆ij
|∆ij |p
Mi (j)
j=1
=
A?i +
k
X
|∆i |p · Mi (j)
j=1
∆
∆ij
|∆ij |p
∆
= A?i + Ei .
∆
Observe that Ei is the weighted sum of k vectors, |∆ii1|p , . . . , |∆iik|p , having unit `p -norm. Observe further that,
1
k
Pk
since the sum of their weights satisfies |∆i |p j=1 |Mi (j)| ≤ k|∆i |p , we have the `p -norm of Ei is not larger than
|Ei |p ≤ k|∆i |p . The proof is complete using the triangle inequality:
|Ai − (U V )i |p
3.2
≤
|A?i − Ai |p + |A?i − (U V )i |p
=
|∆i |p + |Ei |p
≤
(k + 1)|∆i |p .
An mk poly(nm)-time algorithm
In this section we give an algorithm that returns a (k + 1)-approximation to the k-LRAp problem in time mk poly(nm).
Algorithm 2 A (k + 1)-approximation to k-LRAp .
Require: An integer k and a matrix A
Ensure: U ∈ Rn×k , V ∈ Rk×m s.t. |A − U V |p ≤ (k + 1)optk,p (A).
[m]
1: for all I ∈ k do
2:
Let U = AI
3:
Use Lemma 1 to compute a matrix V that minimizes the distance dI = |A − U V |p
4: end for
[m]
5: Return U, V that minimizes dI , for I ∈ k
The following statement follows directly from the existence of k columns in A that make up a factor U having small
`p error (Theorem 3).
Theorem 4. Algorithm 2 obtains a (k + 1)-approximation to k-LRAp in time mk poly(nm).
4
A poly(nm)-time bi-criteria algorithm for k-CSSp
We next show an algorithm that runs in time poly(nm) but returns O(k log m) columns of A that can be used in place
of U , with an error O(k) times the error of the best k-factorization. In other words, it obtains more than k columns but
achieves a polynomial running time; we will later build upon this algorithm in Section 5 to obtain a faster algorithm for
4
the k-LRAp problem. We also show a lower bound: there exists a matrix A for which the best possible approximation
for the k-CSSp , for p ∈ (2, ∞), is k Ω(1) .
Definition 5 (Approximate coverage). Let S be a subset of k column indices. We say that column Ai is cp approximately covered by S if for p ∈ [1, ∞) we have minx∈Rk×1 |AS x − Ai |pp ≤ c
minx∈Rk×1 |AS x − Ai |∞ ≤ c(k + 1)|∆|∞ . If c = 1, we say Ai is covered by S.
We first show that if we select a set R columns of size 2k uniformly at random in
we cover a constant fraction of columns of A.
100(k+1)p |∆|p
p
,
n
[m]
2k
and for p = ∞,
, with constant probability
Lemma 6. Suppose R is a set of 2k uniformly random chosen columns of A. With probability at least 2/9, R covers at
least a 1/10 fraction of columns of A.
Proof. Let i be a column index of A selected uniformly at random and not in the set R. Let T = R ∪ {i} and let η be
the cost of the best `p rank-k approximation to AT . Note that T is a uniformly random subset of 2k + 1 columns of A.
Case: p < ∞. Since T is a uniformly random subset of 2k + 1 columns of A, ET [η p ] =
p
10(2k+1)|∆|p
p
”.
n
(2k+1)|∆|p
p
.
n
Let E1 denote
By a Markov bound, Pr[E1 ] ≥ 9/10.
the event “η ≤
By Theorem 3, there exists a subset L of k columns of AT for which minX |AL X − AT |pp ≤ (k + 1)p η p . Since
p p
η
i is itself uniformly random in the set T , it holds that Ei [minx |AL x − Ai |pp ≤ (k+1)
2k+1 . Let E2 denote the event
p p
η
“minx |AL x − Ai |pp ≤ 10(k+1)
”. By a Markov bound, Pr[E2 ] ≥ 9/10.
2k+1
Let E3 denote the event “i ∈
/ L”. Since i is uniformly random in the set T , Pr[E3 ] ≥
Clearly Pr[E1 ∧ E2 ∧ E3 ] ≥ 3/10. Conditioned on E1 ∧ E2 ∧ E3 , we have
min |AR x − Ai |pp
x
E3
≤
min |AL x − Ai |pp
E2
10(k + 1)p η p
2k + 1
100(k + 1)p |∆|pp
,
n
≤
E1
≤
k+1
2k
> 1/2.
x
which implies that i is covered by R. Note that the first inequality uses that L is a subset of R given E3 , and so the
regression cost using AL cannot be smaller than that of using AR
P
P
P 3
Let Zi be an indicator variable if i is covered by R and let Z = i Zi . We have E[Z] = i E[Zi ] ≥ i 10
=
9m
7
3m/10; hence E[m − Z] ≤ 7m
.
By
a
Markov
bound,
Pr[m
−
Z
≥
]
≤
.
10
10
9
Case p = ∞. Then η ≤ |∆|∞ since AT is a submatrix of A. By Theorem 3, there exists a subset L of k columns
of AT for which minX |AL X − AT |∞ ≤ (k + 1)η. Defining E3 as before and conditioning on it, we have
min |AR x − Ai |∞
x
≤ min |AL x − Ai |∞
x
≤ min |AL X − AT |∞
X
≤ (k + 1)|∆|∞ ,
i.e., i is covered by R. Again defining Zi to be the event that i is covered by R, we have E[Zi ] ≥
9m
5
7
E[m − Z] ≤ m
2 , which implies Pr[m − Z ≥ 10 ] ≤ 9 < 9 .
1
2,
and so
We are now ready to introduce Algorithm 3. We can without loss of generality assume that the algorithm knows
a number N for which |∆|p ≤ N ≤ 2|∆|p . Indeed, such a value can be obtained by first computing |∆|2 using the
SVD. Note that although one does not know ∆, one does know |∆|2 since this is the Euclidean norm of all but the top k
singular values of A, which one can compute from the SVD of A. Then, note that for p < 2, |∆|2 ≤ |∆|p ≤ n2−p |∆|2 ,
while for p ≥ 2, |∆|p ≤ |∆|2 ≤ n1−2/p |∆|p . Hence, there are only O(log n) values of N to try, given |∆|2 , one of
which will satisfy |∆|p ≤ N ≤ 2|∆|p . One can take the best solution found by Algorithm 3 for each of the O(log n)
guesses to N .
5
Algorithm 3 Selecting O(k log m) columns of A.
Require: An integer k, and a matrix A = A? + ∆.
Ensure: O(k log m) columns of A
S ELECT C OLUMNS (k, A)
if number of columns of A ≤ 2k then
return all the columns of A
else
repeat
Let R be uniform at random 2k columns of A
until at least (1/10)-fraction columns of A are cp -approximately covered
Let AR be the columns of A not approximately covered by R
return AR ∪ S ELECT C OLUMNS (k, AR )
end if
Theorem 7. With probability at least 9/10, Algorithm 3 runs in time poly(nm) and returns O(k log m) columns that
can be used as a factor of the whole matrix inducing `p error O(k|∆|p ).
Proof. First note, that if |∆|p ≤ N ≤ 2|∆|p and if i is covered by a set R of columns, then i is cp -approximately
covered by R for a constant cp ; here cp = 2p for p < ∞ and c∞ = 2. By Lemma 6, the expected number of
repetitions of selecting 2k columns until (1/10)-fraction of columns of A are covered is O(1). When we recurse on
S ELECT C OLUMNS on the resulting matrix AR , each such matrix admits a rank-k factorization of cost at most |∆|p .
Moreover, the number of recursive calls to S ELECT C OLUMNS can be upper bounded by log10 m. In expectation there
will be O(log m) total repetitions of selecting 2k columns, and so by a Markov bound, with probability 9/10, the
algorithm will choose O(k log m) columns in total and run in time poly(nm).
Let S be the union of all columns of A chosen by the algorithm. Then for each column i of A, for p ∈ [1, ∞), we
100(k+1)p 2p |∆|p
p
have minx |AS x − Ai |pp ≤
, and so minX |AS X − A|pp ≤ 100(k + 1)p 2p |∆|pp . For p = ∞ we instead
n
have minx |AS x − Ai |∞ ≤ 2(k + 1)|∆|∞ , and so minX |AS X − A|∞ ≤ 2(k + 1)|∆|∞ .
4.1
A lower bound for k-CSSp
In this section we prove an existential result showing that there exists a matrix for which the best approximation to the
k-CSSp is k Ω(1) .
Lemma 8. There exists a matrix A such that the best approximation for the k-CSSp problem, for p ∈ (2, ∞), is k Ω(1) .
Proof. Consider A = (k + 1)Ik+1 , where Ik+1 is the (k + 1) × (k + 1) identity matrix. And consider the matrix
B = (k + 1) · Ik+1 − E, where E is the (k + 1) × (k + 1) all ones matrix. Note that B has rank at most k, since the
sum of its columns is 0.
Case: 2 < p < ∞. If we choose any k columns of A, then the `p cost of using them to approximate A is (k + 1).
1/p
On the other hand, |A − B|∞ = 1, which means that `p cost of B is smaller or equal than (k + 1)2
.
Case: p = ∞. If we choose any k columns of A, then the `∞ cost of using them to approximate A is k + 1. On the
other hand, |A − B|∞ = 1, which means that `∞ cost of B is smaller or equal than 1.
√
Note also that in [18] the authors show that for p = 1 the best possible approximation is Ω( k) up to poly(log k)
factors.
5
A ((k log n)k poly(mn))-time algorithm for k-LRAp
In the previous section we have shown how to get a rank-O(k log m), O(k)-approximation in time poly(nm) to the
k-CSSp and k-LRAp problems. In this section we first show how to get a rank-k, poly(k)-approximation efficiently
6
starting from a rank-O(k log m) approximation. This algorithm runs in polynomial time as long as k = O
log n
log log n
.
We then show how to obtain a (k log m)O(p) -approximation ratio in polynomial time for every k.
Let U be the columns of A selected by Algorithm 3.
5.1
An isoperimetric transformation
The first step of our proof is to show that we can modify the selected columns of A to span the same space but to have
small distortion. For this, we need the following notion of isoperimetry.
Definition 9 (Almost isoperimetry). A matrix B ∈ Rn×m is almost-`p -isoperimetric if for all x, we have
|x|p
≤ |Bx|p ≤ |x|p .
2m
We now show that given a full rank A ∈ Rn×m , it is possible to construct in polynomial time a matrix B ∈ Rn×m
such that A and B span the same space and B is almost-`p -isoperimetric.
Lemma 10. Given a full (column) rank A ∈ Rn×m , there is an algorithm that transforms A into a matrix B such that
span A = span B and B is almost-`p -isoperimetric. Furthermore the running time of the algorithm is poly(nm).
Proof. In [4], specifically, Equation (4) in the proof of Theorem 4, the authors show that in polynomial time it is
possible to find a matrix B such that span B = span A and for all x,
√
|x|2 ≤ |Bx|p ≤ m|x|2 ,
for any p ≥ 1.
If p < 2, their result implies
√
√
|x|
√ p ≤ |x|2 ≤ |Bx|p ≤ m|x|2 ≤ m|x|p ,
m
and so rescaling B by
√
m makes it almost-`p -isoperimetric. On the other hand, if p > 2, then
√
|x|p ≤ |x|2 ≤ |Bx|p ≤ m|x|2 ≤ m|x|p ,
and rescaling B by m makes it almost-`p -isoperimetric.
Note that the algorithm used in [4] relies on the construction of the Löwner–John ellipsoid for a specific set of
points. Interestingly, we can also show that there is a more simple and direct algorithm to compute such a matrix B;
this may be of independent interest. We provide the details of our algorithm in the supplementary material.
5.2
Reducing the rank to k
The main idea for reducing the rank is to first apply the almost-`p -isoperimetric transformation to the factor U to
obtain a new factor Z 0 . For such a Z 0 , the `p -norm of Z 0 V is at most the `p -norm of V . Using this fact we show
that V has a low-rank approximation and a rank-k approximation of V translates into a good rank-k approximation of
U V . But a good rank-k approximation of V can be obtained by exploring all possible k-subsets of rows of V , as in
Algorithm 2. More formally, in Algorithm 4 we give the pseudo-code to reduce the rank of our low-rank approximation
from O(k log m) to k. Let δ = |∆|p = optk,p (A).
Theorem 11. Let A ∈ Rn×m , U ∈ Rn×O(k log m) , V ∈ RO(k log m)×m be such that |A − U V |p = O(kδ). Then,
Algorithm 4 runs in time O(k log m)k (mn)O(1) and outputs W ∈ Rn×k , Z ∈ Rk×m such that |A − W Z|p =
O((k 4 log k)δ).
7
Algorithm 4 An algorithm that transforms an O(k log m)-rank matrix decomposition into a k-rank matrix decomposition without inflating the error too much.
Require: U ∈ Rn×O(k log m) , V ∈ RO(k log m)×m
Ensure: W ∈ Rn×k , Z ∈ Rk×m
1: Apply Lemma 10 to U to obtain matrix W 0
2: Apply Lemma 1 to obtain matrix Z 0 , s.t. ∀i, |W 0 Zi0 − (U V )i |p is minimized
3: Apply Algorithm 2 with input (Z 0 )T ∈ Rn×O(k log m) and k to obtain X and Y
4: Set Z ← X T
5: Set W ← W 0 Y T
6: Output W and Z
Proof. We start by bounding the running time. Step 3 is computationally the most expensive since it requires to execute
a brute-force search on the O(k log m) columns of (Z 0 )T . So the running time follows from Theorem 4.
Now we have to show that the algorithm returns a good approximation. The main idea behind the proof is that U V
is a low-rank approximable matrix. So after applying Lemma 10 to U to obtain a low-rank approximation for U V we
can simply focus on Z 0 ∈ RO(k log m)×n . Next, by applying Algorithm 2 to Z 0 , we obtain a low-rank approximation in
time O(k log m)k (mn)O(1) . Finally we can use this solution to construct the solution to our initial problem.
We know by assumption that |A − U V |p = O(kδ). Therefore, it suffices by the triangle inequality to show
|U V − W Z|p = O((k 4 log k)δ). First note that U V = W 0 Z 0 since Lemma 10 guarantees that span U = span W 0 .
Hence we can focus on proving |W 0 Z 0 − W Z|p ≤ O((k 4 log k)δ).
We first prove two useful intermediate steps.
Lemma 12. There exist matrices U ∗ ∈ Rn×k , V ∗ ∈ Rk×m such that |W 0 Z 0 − U ∗ V ∗ |p = O(kδ).
Proof. There exist U ∗ ∈ Rn×k , V ∗ ∈ Rk×m such that |A − U ∗ V ∗ |p ≤ δ and, furthermore, |A − U V |p = |A −
W 0 Z 0 |p = O(kδ). The claim follows by the Minkowski inequality.
Lemma 13. There exist matrices F ∈ RO(k log m)×k , D ∈ Rk×n such that |W 0 (Z 0 − F D)|p = O(k 2 δ).
Proof. From Lemma 12, we know that |W 0 Z 0 − U ∗ V ∗ |p = O(kδ). Hence, from Theorem 3, we know that there exists
a matrix C ∈ Rn×k composed of k columns of W 0 Z 0 , and a matrix D ∈ Rk×m such that |W 0 Z 0 − CD|p = O(k 2 δ).
Furthermore, note that selecting k columns of W 0 Z 0 is equivalent to select the same columns in Z 0 and multiplying
them by W 0 . So we can express C = W 0 F for some matrix F ∈ RO(k log m)×k . Thus we can rewrite
|W 0 Z 0 − CD|p
= |W 0 Z 0 − W 0 F D|p
= |W 0 (Z 0 − F D)|p
≤ O(k 2 δ).
|y|
Now from the guarantees of Lemma 10 we know that for any vector y, |W 0 y|p ≤ k logp k . So we have |Z 0 − F D|p ≤
O((k 3 log k)δ), Thus |(Z 0 )T − DT F T |p ≤ O((k 3 log k)δ), so (Z 0 )T has a low-rank approximation with error at most
O((k 3 log k)δ). So we can apply Theorem 3 again and we know that there are k columns of (Z 0 )T such that the lowrank approximation obtained starting from those columns has error at most O((k 4 log k)δ). We obtain such a low-rank
approximation from Algorithm 2 with input (Z 0 )T ∈ Rn×O(k log m) and k. More precisely, we obtain an X ∈ Rn×k
and Y ∈ Rk×O(k log m) such that |(Z 0 )T − XY |p ≤ O((k 4 log k)δ). Thus |Z 0 − Y T X T |p ≤ O((k 4 log k)δ).
Now using again the guarantees of Lemma 10 for W 0 , we get |W 0 (Z 0 − Y T X T )|p ≤ O((k 4 log k)δ). So
0
|W (Z 0 − Y T X T )|p = |W 0 Z 0 − W Z)|p = |U V − W Z)|p ≤ O((k 4 log k)δ). By combining it with |A − U V |p =
O(kδ) and using the Minkowski inequality, the proof is complete.
8
5.3
Improving the Running Time
We now show how to improve the running time to (mn)O(1) for every k and every constant p ≥ 1, at the cost of a
poly(k log(m))-approximation instead of the poly(k)-approximation we had previously.
Theorem 14. Let A ∈ Rn×m , 1 ≤ k ≤ min(m, n), and p ≥ 1 be an arbitrary constant. Let U ∈ Rn×O(k log m) and
V ∈ RO(k log m)×m be such that |A − U V |p = O(kδ). There is an algorithm which runs in time (mn)O(1) and outputs
W ∈ Rn×k , Z ∈ Rk×m such that |A − W Z|p = (k log m)O(p) δ.
Proof. The proof of Theorem 11 shows there exists a rank-k matrix X for which |U XV T − A|p = O(k 4 log k)δ.
Instead of the enumeration algorithm used in the proof of Theorem 11 to find such an X, we will instead use `p -leverage
score sampling [4].
It is shown in Theorem 6 of [4] that given U one can in poly(mn) time and with probability 1 − o(1), find a
sampling and rescaling matrix S with (k log m)O(p) rows such that for all vectors w, |SU w|p = (1 ± 1/2)|U w|p .
Indeed, in their notation one can compute a well-conditioned-basis in poly(mn) time and then sample rows of U
according to the p-th power of the p-norms of the rows of the well-conditioned basis. Since S is a sampling and
rescaling matrix, we have E[|SY |pp ] = |Y |pp for any fixed matrix Y .
Let X ∗ be the rank-k matrix minimizing |U X ∗ V T − A|p . By the triangle inequality, for an arbitrary X we have
|SU XV T − SA|p = |SU X ∗ V T − SU XV T |p ± |SU X ∗ V T − SA|p .
∗
By a Markov bound, |SU X V
T
−
SA|pp
p
∗
≤ 100 |U X V
T
−
A|pp
(1)
p
with probability 1 − 1/100 , and so
|SU X ∗ V T − SA|p ≤ 100|U X ∗ V T − A|p
(2)
with this probability. Moreover, with probability 1 − o(1),
|SU X ∗ V T − SU XV T |p = (1 ± 1/2)|U X ∗ V T − U XV T |p
(3)
p
simultaneously for all X. By a union bound, both of these events occur with probability 1 − 1/100 − o(1). In this case,
it follows that if X 0 satisfies |SU X 0 V T − SA|p ≤ α minrank k B |SU BV T − SA|p , then also |SU X 0 V T − SA|p ≤
α|SU X ∗ V T − SA|p . Thus, using the triangle inequality, (2) and (3),
|U X 0 V T − A|p
≤
|U X 0 V T − U X ∗ V T |p + |U X ∗ V T − A|p
≤ (1 + 1/2)|SU X 0 V T − SU X ∗ V T |p + |U X ∗ V T − A|p
≤ (1 + 1/2)(|SU X 0 V T − SA|p + |SU X ∗ V T − SA|p ) + |U X ∗ V T − A|p
≤ (1 + 1/2)((α + 1)|SU X ∗ V T − SA|) + |U X ∗ V T − A|p
= O(α)|U X ∗ V T − A|p .
Now consider the problem |SU XV T − SA|p . We can compute a well-conditioned basis in poly(mn) time and
then sample columns of V according to the p-th power of the p-norms of the columns of the well-conditioned basis.
Let T denote this sampling matrix, which has (k log m)O(p) columns. We condition on analogous events to those in
(2) and (3) above, which hold again with probability 1 − 1/100p − o(1). Then if X 00 is a β-approximate minimizer to
|SU X 00 V T T − SAT |p , then analogously,
|SU X 00 V T − SA|p ≤ O(β)|SU X 0 V T − SA|p .
(4)
We thus have by several applications of the triangle inequality and the above,
|U X 00 V T − A|p
≤
|U X 00 V T − U X 0 V T |p + |U X 0 V T − A|p
≤
(1 + 1/2)|SU X 00 V T − SU X 0 V T |p + O(α)|U X ∗ V T − A|p
≤
(1 + 1/2)(|SU X 00 V T − SA|p + |SU X 0 V T − SA|p ) + O(αδ)
≤
O(β)|SU X 0 V T − SA|p + O(αδ)
≤
O(β)(|SU X 0 V T − SU X ∗ V T |p + |SU X ∗ V T − SA|p ) + O(αδ)
≤
O(β)(|U X 0 V T − U X ∗ V T |p ) + O((α + β)δ)
≤
O(β)(|U X 0 V T − A|p + δ) + O((α + β)δ)
≤
O(αβδ).
(5)
9
Ratio
200
0.84
45
0.82
40
0.8
35
Absolute Error
0.78
150
0.76
0.74
100
0.72
0.7
50
0
1
2
3
4
5
6
7
8
9
Alg 2
SVD
Ratio
1.2
1
30
0.8
25
0.6
20
15
0.68
10
0.66
5
0.64
0
0.4
Ratio of Alg 2 to SVD Errors
SVD
Absolute Error
Alg 2
Ratio of Alg 2 to SVD Errors
250
0.2
10
0
1
2
3
4
5
K
6
7
8
9
10
K
(a) FIDAP matrix (`1 )
(b) KOS matrix (`∞ )
Figure 1: Comparing the performance of Algorithm 2 with SVD on the real data sets.
Alg 2
SVD
0.8
Ratio
0.72
50
Absolute Error
0.74
100
Ratio of Alg2 to SVD Errors
0.76
150
Alg 2
SVD
1.2
Ratio
1
0.78
200
Absolute Error
1.2
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
Ratio of Alg2 to SVD Erros
250
0.7
0
0
0.68
1
2
3
4
5
6
7
8
9
0
1
2
3
4
10
5
6
7
8
9
10
K
K
(a) Random matrix (`1 )
(b) Random matrix (`∞ )
Figure 2: Comparing the performance of Algorithm 2 with SVD on the random matrix.
Finally, observe that since SU XV T T − SAT is a (k log m)O(p) × (k log m)O(p) matrix for any X, it follows that
its Frobenius norm is related up to a (k log m)O(p) factor to its entrywise p-norm. Consequently, the Frobenius norm
minimizer X 00 is a (k log m)O(p) -approximate minimizer to the entrywise p-norm, and so β = (k log m)O(p) in the
notation above. It then follows from (4) that α = O(β) = (k log m)O(p) as well. Consequently, by (5), we have that
|U X 00 V T − A|p = (k log m)O(p) δ.
Finally, note that the Frobenius norm minimizer X 00 to |SU X 00 V T T − SAT |p can be solved in time (k log m)O(p)
time, using the result in [9]. This completes the proof.
6
Experiments
In this section, we show the effectiveness of Algorithm 2 compared to the SVD. We run our comparison both on
synthetic as well as real data sets. For the real data sets, we use matrices from the FIDAP set2 and a word frequency
dataset from UC Irvine 3 . The FIDAP matrix is 27 × 27 with 279 real asymmetric non-zero entries. The KOS blog
entries matrix, representing word frequencies in blogs, is 3430 × 6906 with 353160 non-zero entries. For the synthetic
data sets, we use two matrices. For the first, we use a 20 × 30 random matrix with 184 non-zero entries—this random
matrix was generated as follows: independently, we set each entry to 0 with probability 0.7, and to a uniformly random
value in [0, 1] with probability 0.3. Both matrices are full rank. For the second matrix, we use a random ±1 20 × 30
matrix.
2 ihttp://math.nist.gov/MatrixMarket/data/SPARSKIT/fidap/fidap005.html
3 https://archive.ics.uci.edu/ml/datasets/Bag+of+Words
10
600
Alg 2
SVD
Ratio
0.94
500
Alg 2
SVD
Ratio
0.8
0.7
300
0.86
0.84
200
0.82
100
0.6
0.5
1.5
0.4
1
0.3
0.2
0.5
0.1
0.8
0
0
0.78
1
2
3
4
5
6
7
8
9
Ratio of Alg 2 to SVD Errors
0.88
Absolute Error
400
Ratio of Alg 2 to SVD Errors
2
0.9
Absolute Error
2.5
0.92
0
1
10
2
3
4
5
6
7
8
9
10
K
K
(a) ±1 matrix (`1 )
(b) ±1 matrix (`∞ )
Figure 3: Comparing the performance of Algorithm 2 with SVD on the ±1 matrix.
In all our experiments, we run a simplified version of Algorithm 2, where instead of running for all possible [m]
k
subsets of k columns (which would be computationally prohibitive), we repeatedly sample k columns, a few thousand
times, uniformly at random. We then run the `p -projection on each sampled set and finally select the solution with the
smallest `p -error. (While this may not guarantee provable approximations, we use this a reasonable heuristic that seems
to work well in practice, without much computational overhead.) We focus on p = 1 and p = ∞.
Figure 1 illustrates the relative performance of Algorithm 2 compared to the SVD for different values of k on the
real data sets. In the figure the green line is the ratio of the total error. The `1 -error for Algorithm 2 is always less than
the corresponding error for the SVD and in fact consistently outperforms the SVD by roughly 40% for small values of
k on the FIDAP matrix. On the larger KOS matrix, the relative improvement in performance with respect to `∞ -error is
more uniform (around 10%).
We observe similar trends for the synthetic data sets as well. Figures 2 and 3 illustrate the trends. Algorithm 2
performs consistently better than the SVD in the case of `1 -error for both the matrices. In the case of `∞ -error, it
outperforms SVD by around 10% for higher values of k on the random matrix. Furthermore, it consistently outperforms
SVD, between 30% and 50%, for all values of k on the random ±1 matrix.
To see why our `∞ error is always 1 for a random ±1 matrix A, note that by setting our rank-k approximation to be
the zero matrix, we achieve an `∞ error of 1. This is optimal for large values of n and m and small k as can be seen by
recalling the notion of the sign-rank of a matrix A ∈ {−1, 1}n×m , which is the minimum rank of a matrix B for which
the sign of Bi,j equals Ai,j for all entries i, j. If the sign-rank of A is larger than k, then for any rank-k matrix B, we
have kA − Bk∞ ≥ 1 since necessarily there is an entry Ai,j for which |Ai,j − Bi,j
√| ≥ 1. It is known that the sign-rank
of a random m × m matrix A, and thus also of a random n × m matrix A, is Ω( m) with high probability [8].
7
Conclusions
We studied the problem of low-rank approximation in the entrywise `p error norm and obtained the first provably good
approximation algorithms for the problem that work for every p ≥ 1. Our algorithms are extremely simple, which
makes them practically appealing. We showed the effectiveness of our algorithms compared with the SVD on real and
synthetic data sets. We obtain a k O(1) approximation factor for every p for the column subset selection problem, and
we showed an example matrix for this problem for which a k Ω(1) approximation factor is necessary. It is unclear if
better approximation factors are possible by designing algorithms that do not choose a subset of input columns to span
the output low rank approximation. Resolving this would be an interesting and important research direction.
11
References
[1] J.P. Brooks, J.H. Dulá, and E.L. Boone. A pure `1 -norm principal component analysis. Computational Statistics
& Data Analysis, 61:83–98, 2013.
[2] Emmanuel J. Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? JACM,
58(3):11:1–11:37, 2011.
[3] C. Dan, K. A. Hansen, H. Jiang, L. Wang, and Y. Zhou. On low rank approximation of binary matrices. Technical
Report 1511.01699v1, arXiv, 2015.
[4] Anirban Dasgupta, Petros Drineas, Boulos Harb, Ravi Kumar, and Michael W. Mahoney. Sampling algorithms
and coresets for lp regression. SICOMP, 38(5)(2060–2078), 2009.
[5] Amit Deshpande, Madhur Tulsiani, and Nisheeth K. Vishnoi. Algorithms and hardness for subspace approximation.
In SODA, pages 482–496, 2011.
[6] A. Eriksson and A. van den Hengel. Efficient computation of robust low-rank matrix approximations using the L1
norm. PAMI, 34(9):1681–1690, 2012.
[7] Dan Feldman, Amos Fiat, Micha Sharir, and Danny Segev. Bi-criteria linear-time approximations for generalized
k-mean/median/center. In SoCG, pages 19–26, 2007.
[8] Jürgen Forster. A linear lower bound on the unbounded error probabilistic communication complexity. J. Comput.
Syst. Sci., 65(4):612–625, 2002.
[9] Shmuel Friedland and Anatoli Torokhti. Generalized rank-constrained matrix approximations. SIAM Journal on
Matrix Analysis and Applications, 29(2):656–659, 2007.
[10] Nicolas Gillis and Stephen A. Vavasis. On the complexity of robust PCA and `1 -norm low-rank matrix approximation. Technical Report 1509.09236, arXiv, 2015.
[11] S. A. Goreinov and E. E. Tyrtyshnikov. The maximal-volume concept in approximation by low-rank matrices.
Contemporary Mathematics, 208:47–51, 2001.
[12] S. A. Goreinov and E. E. Tyrtyshnikov. Quasioptimality of skeleton approximation of a matrix in the Chebyshev
norm. Doklady Mathematics, 83(3):374–375, 2011.
[13] Peter J. Huber. Robust Statistics. John Wiley & Sons, New York„ 1981.
[14] Qifa Ke and Takeo Kanade. Robust L1 norm factorization in the presence of outliers and missing data by
alternative convex programming. In CVPR, pages 739–746, 2005.
[15] Cewu Lu, Jiaping Shi, and Jiaya Jia. Scalable adaptive robust dictionary learning. TIP, 23(2):837–847, 2014.
[16] Deyu Meng and Fernando. D. L. Torre. Robust matrix factorization with unknown noise. In ICCV, pages
1337–1344, 2013.
[17] Praneeth Netrapalli, U. N. Niranjan, Sujay Sanghavi, Animashree Anandkumar, and Prateek Jain. Non-convex
robust PCA. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information
Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 1107–1115, 2014.
[18] Zhao Song, David P. Woodruff, and Pelin Zhong. Low rank approximation with entrywise `1 -norm error. Technical
Report 1611.00898, arXiv, 2016.
[19] Naiyan Wang, Tiansheng Yao, Jingdong Wang, and Dit-Yan Yeung. A probabilistic approach to robust matrix
factorization. In ECCV, pages 126–139, 2012.
12
[20] Naiyan Wang and Dit-Yan Yeung. Bayesian robust matrix factorization for image and video processing. In ICCV,
pages 1785–1792, 2013.
[21] David P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations and Trends in Theoretical
Computer Science, 10(1-2):1–157, 2014.
[22] L. Xiong, X. Chen, and J. Schneider. Direct robust matrix factorization for anomaly detection. In ICDM, pages
844–853, 2011.
[23] H. Xu, C. Caramanis, and S. Sanghavi. Robust PCA via outlier pursuit. TOIT, 58(5):3047–3064, 2012.
[24] L. Xu and A. L. Yuille. Robust principal component analysis by self-organizing rules based on statistical physics
approach. IEEE Transactions on Neural Networks, 6(1):131–143, 1995.
[25] Xinyang Yi, Dohyung Park, Yudong Chen, and Constantine Caramanis. Fast algorithms for robust pca via gradient
descent. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural
Information Processing Systems 29, pages 4152–4160. Curran Associates, Inc., 2016.
[26] Y. Zheng, G. Liu, S. Sugimoto, S. Yan, and M. Okutomi. Practical low-rank matrix approximation under robust
L1 -norm. In CVPR, pages 1410–1417, 2012.
13
| 8cs.DS
|
RESEARCH ARTICLE
Computer-aided position planning of
miniplates to treat facial bone defects
Jan Egger1,2‡*, Jürgen Wallner3‡, Markus Gall1, Xiaojun Chen4, Katja SchwenzerZimmerer3, Knut Reinbacher3, Dieter Schmalstieg1
1 Institute for Computer Graphics and Vision, Faculty of Computer Science and Biomedical Engineering,
Graz University of Technology, Graz, Austria, 2 BioTechMed-Graz, Graz, Austria, 3 Department of Oral &
Maxillofacial Surgery, Medical University of Graz, Graz, Styria, Austria, 4 Institute of Biomedical
Manufacturing and Life Quality Engineering, School of Mechanical Engineering, Shanghai Jiao Tong
University, Shanghai, China
a1111111111
a1111111111
a1111111111
a1111111111
a1111111111
‡ These authors are joint first authors on this work.
* [email protected]
Abstract
OPEN ACCESS
Citation: Egger J, Wallner J, Gall M, Chen X,
Schwenzer-Zimmerer K, Reinbacher K, et al.
(2017) Computer-aided position planning of
miniplates to treat facial bone defects. PLoS ONE
12(8): e0182839. https://doi.org/10.1371/journal.
pone.0182839
Editor: Jose Manuel Garcia Aznar, University of
Zaragoza, SPAIN
Received: January 9, 2017
Accepted: July 25, 2017
Published: August 17, 2017
Copyright: © 2017 Egger et al. This is an open
access article distributed under the terms of the
Creative Commons Attribution License, which
permits unrestricted use, distribution, and
reproduction in any medium, provided the original
author and source are credited.
Data Availability Statement: All relevant data are
uploaded to Figshare at URL: https://figshare.com/
articles/Cranial_Defect_Datasets/4659565; DOI:
https://doi.org/10.6084/m9.figshare.4659565.v1.
Funding: The work received funding from
BioTechMed-Graz in Austria (“Hardware
accelerated intelligent medical imaging”) and the
6th Call of the Initial Funding Program from the
Research & Technology House (F&T-Haus) at the
Graz University of Technology (PI: DDr. Jan Egger).
Dr. Xiaojun Chen receives support by the Natural
In this contribution, a software system for computer-aided position planning of miniplates to
treat facial bone defects is proposed. The intra-operatively used bone plates have to be passively adapted on the underlying bone contours for adequate bone fragment stabilization.
However, this procedure can lead to frequent intra-operatively performed material readjustments especially in complex surgical cases. Our approach is able to fit a selection of common implant models on the surgeon’s desired position in a 3D computer model. This
happens with respect to the surrounding anatomical structures, always including the possibility of adjusting both the direction and the position of the used osteosynthesis material. By
using the proposed software, surgeons are able to pre-plan the out coming implant in its
form and morphology with the aid of a computer-visualized model within a few minutes. Further, the resulting model can be stored in STL file format, the commonly used format for 3D
printing. Using this technology, surgeons are able to print the virtual generated implant, or
create an individually designed bending tool. This method leads to adapted osteosynthesis
materials according to the surrounding anatomy and requires further a minimum amount of
money and time.
Introduction
The Reconstructions of facial deformations and defects due to bone fractures where two bone
fragments are operatively stabilized, so called osteosynthesis, is part of a surgeon’s daily life.
More precisely, these bone fractures occur as a result of applied outer forces, like happening
due to traffic accidents, results of tumor removal or deformation treatment [1]. One major
cause for facial fractures in regions where winter sports are frequently practiced is skiing and
snowboarding [2]. These two sports account for most of the facial injuries overall, since wearing a helmet is not mandatory thus resulting in a high percentage of practitioners not equipped
with appropriate protection like helmets, even though high velocities are reached not able to
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
1 / 19
Position planning of miniplates
Science Foundation of China (Grant No.:
81511130089) and the Foundation of Science and
Technology Commission of Shanghai Municipality
(Grants Nos.: 14441901002, 15510722200 and
16441908400).
Competing interests: The authors have declared
that no competing interests exist.
be absorbed by the skull in case of collision. According to Gassner et al. [3], the average age of
people transferred to the department of oral and maxillofacial surgery with facial injuries lies
by 26 years with 50% between age 16 and 38. The group with the highest rate of injuries is the
one of children between 7 and 12 years. Further, there is a high variation between the groups
of men and woman suffering from soft tissue damages or bone fractures in the facial area due
to skiing accidents, stating that men account for 65.3% of the cases whereas females show a
number of 34.7% injured patients. Moreover, the types of injuries range from facial bone fractures to dento-alveolar traumas and soft tissue injuries. There is also a wide range of different
mechanisms including simple falls, collisions with others or objects, a struck by equipment, lift
accidents and various more. For a better understanding of these data, Table 1 gives a detailed
overview of these findings.
However, accidents resulting in facial bone fractures require osteosynthesis where the application of varying osteosynthesis materials such as miniplates [4], [5] (Fig 1) is essential. As an
example, Fig 2 shows an osteosynthesis of a mandibular angle fracture. In this case two adaptive plates where used, one aligned to the upper mandibular border (Linea Oblique) and the
other on the lower border of the mandible. Additionally Fig 3 shows an x-ray image of a case
suffering from such a mandibular angle fracture. To the left, the defect is shown before treatment. On the right image the same patient after bone repositioning and miniplate application
is shown. These commonly used miniplates are straight titanium plates, consisting, depended
on the type, of at least two finishing ring sections at both ends, where screws for an implantbone fixation can be drilled through. The raw implant is available in different variations, consisting of further middle-ring sections, orthogonal bendings or circular ring section arrangements. Whereby, in this work, the focus lies on those which are commonly used in craniomaxillofacial surgery, e.g. in cases of mid-facial or also partly in lower jaw trauma treatment.
Table 1. Summary of results according to the trauma database due to facial injuries caused by ski
accidents in Innsbruck (Department of Oral and Maxillofacial Surgery, Innsbruck, Austria).
Variable
579 Skiers / 882 Incidents
Age
Mean
28.35
Standard Deviation
15.78
Min
2
Median
26
Max
81
Male
378
Female
201
Gender
Injury Type
Soft tissue injuries
336
Facial bone fracture
310
Dentoalveolar trauma
236
Collision with others
135
Mechanism of Injury
Struck by equipment
70
Falls
263
Collision with object
46
Lift accidents
34
Others
31
https://doi.org/10.1371/journal.pone.0182839.t001
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
2 / 19
Position planning of miniplates
Fig 1. The three figures show commonly available miniplates for bone fracture fixation. All models are of the MedArtis Trauma 2.0
series. The most left one, with article number M-4138, shows an overall length of 23 mm consisting of two end ring elements and two middle
ring elements. The model in the middle, with article number M-4320, with the same properties as the left one but an extended bar of 9 mm and
thus extended to 29 mm in the overall length. Last, the most right one, with article number M-4322, with two additional ring elements and an
overall length of 35 mm.
https://doi.org/10.1371/journal.pone.0182839.g001
The Modus 2.0 series from the MedArtis Group (www.medartis.com) were chosen as the
osteosynthesis materials used in this trial, (Fig 1).
For facial reconstructions, the implants are fixed perpendicular to the fracture on both fracture sites with the purpose of stabilizing the defect. In general, the miniplates cannot be
deployed in their stiff and straight initial form why it is necessary to adapt them along the fractured surface to gain a perfect fit using a bending tool. This can lead to intraoperatively performed time-consuming bone plate adapting procedures, especially in complex surgical cases.
Another method is to generate a stereolithographic (STL) model out of the Computed Tomography (CT) scan data and pre-bend the implant on those [6]. This method takes even more
time, especially in the pre-operative phase, also leading to higher expenses in overall costs [7].
Further, there are commercial software tools available, like offered by Materialise (www.
materialise.com). However, these software packages include complex user interfaces that are
Fig 2. a,b,c: Clinically used osteosynthesis of the left mandibular angle using two plate techniques.
X1, x2 plate position, N: Inferior alveolar nerve in the mandibular bone. Note: The nerve does not interfere with
the bone plates to avoid nerve damaging and loss of lip sensibility.
https://doi.org/10.1371/journal.pone.0182839.g002
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
3 / 19
Position planning of miniplates
Fig 3. Image of real patient x-ray data suffering from a mandibular angle fracture (left angle). Left image shows the untreated defect. The result of
the treatment with applied miniplates after bone repositioning is shown on the right part of the figure.
https://doi.org/10.1371/journal.pone.0182839.g003
difficult to handle in the practical routine and require high monetary costs since they are not
licensed free, personal resources for their use or further outsourced services to the industry.
Further, with this software it is not possible to pre-plan the application of commercial available
osteosynthesis materials, since the software is created to generate a patient specific implant
that is offered and manufactured by the industry as a high monetary service and sold to the
clinical center. Therefore, this contribution proposes a new method for computer aided planning of facial surgeries with miniplates using the medical image processing platform MeVisLab
(www.mevislab.de), which we have already successfully used in various medical applications,
like [8–13], and integrates and builds upon the common ITK and VTK libraries, similar to
(3D) Slicer [14], [15]. By using this software, the surgeon has the possibility to plan the implant
independent of its position on the facial bone without using a 3D model. Further, the planning
time can be performed in only a few minutes and readjustments of the material position are
easily possible to perform without any additional costs. This is achieved by first calculating the
surface curvature of an initially set marker on which the further implant is created. The generation includes the alignment and orientation of pre-constructed osteosynthesis models, for the
ring sections of the bone plate, and the generation of the bridge sections using Delaunay-triangulation [16]. Finally, the implant model can be 3D-printed [17–19], since it is stored in STLfile format, usable as bending and adapting tool or even as the final created implant.
Others working in the area of computer-aided treatment of facial bone defects are Zhao
et al. [20] who present the application of virtual surgical planning (VSP) with computer-assisted design (CAD) and manufacturing (CAM) technology to cranio-maxillofacial surgery. In
doing so, they illustrate the components, system and clinical management of the VSP and
CAD/CAM technology, which includes also the data acquisition, virtual surgical and treatment
planning, and fabrication, and outcome assessment. However, they focus more on individual
implant design, instead of common available implants. Further, Parthasarathy [21] gives a
review about recent trends in 3D modeling and custom implants in craniofacial reconstruction. Bell [22] presents and overview of computer planning and intraoperative navigation in
cranio-maxillofacial surgery. The author states also that in situ plate bending is time consuming, and that in situ plate bending is not practical if tumor grossly invades soft tissue on the lateral mandible. Moreover, the use of stereolithographic models for plate adaptation before
surgery has been used by the author for the past 5 years to aid in maxillo-mandibular reconstruction. Furthermore, Adamus and Lacki [23] investigated the titanium bending process.
Therefore, they carried out a numerical simulation based on the finite element method of the
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
4 / 19
Position planning of miniplates
bending of a Ti6Al4V ELI titanium alloy bar with the Adina System. Summarized, the influence of bar diameter, bending radius and bending angle on the strain and stress distribution in
the deformed element was analyzed. However, they state that problems occur when bent titanium elements have to be bent out or bent over during their exploitation, e.g. some titanium
body implants (like miniplates) need to be manually adjusted to the bone curvature during
surgical operations. The work closest to ours is from Burghart et al. [24] who present a planning system for the preoperative positioning of miniplates on a model of the patient’s skull.
However, they don’t report how fast the miniplate model calculation is and if it enables and
interactive placement in all steps. Furthermore, their tool is not available for other researchers
and it is not clear how many (medical) datasets they used for testing, and if they had typical
cases from the clinical routine. Finally, it seems they did not model real and conventional
available miniplates from a manufacture via their blueprints as it is performed in this study.
All this shortcomings will be addressed in this contribution and thus it is a consequent further
development of their research work.
This contribution is further organized as follows: Section 2 depicts details of the used material and the newly proposed integration, presents the theoretical background of the proposed
mechanism and provides sufficient detail to allow the work to be reproduced. In Section 3,
experimental results, including illustrations of generated implants, are presented. Section 4
gives a summary together with a short discussion to highlight the significance of the introduced achievements and lays the foundation for further work.
Material and methods
Datasets
Only high-resolution data sets (512x512 voxel in x- and y-directions) with slices not exceeding
1.0 mm with 0.25 mm pixel size in z-direction and providing physiological, complete mandibular bone structures without teeth were included in the selection process. Further, no difference was made between atrophic and non-atrophic mandibular bones—both were included
during the selection process. However, incomplete data sets consisting of mandibular structures altered by iatrogenic or pathological factors or fractured mandibles as well as data sets
showing osteosynthesis materials in the lower jaw were excluded in this trial. The used data
sets were provided as DICOM-image files [25]. The proposed software, however, uses the STL
file format why a conversion was performed using a simple MeVisLab network. This operation
also holds the positive side effect of patient information stored in DICOM tags is lost leading
to anonymized datasets. In addition, teeth and alveolar processes were removed due to imagebased artefacts to create an objective collection of data samples. For a further assessment or
own research purposes some of the data used in this trial can freely be downloaded, but we
kindly asked to cite our work [26]: https://figshare.com/articles/Cranial_Defect_Datasets/
4659565
Note: when more datasets get available over time or other researchers provide us new cases,
we will add them to the collection.
Ethics statement: All patient data was anonymized prior to analysis by the authors. In addition, the physicians and doctors who tested our newly developed tool have consented to their
results of testing being published.
Miniplate models
Purchasable miniplates from the Modus 2.0 series from the MedArtis Group (www.medartis.
com) were chosen as the osteosynthesis materials used in this trial [27]. Therefore, we created
CAD models reverse engineered from the miniplates blueprints. Three CAD models for
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
5 / 19
Position planning of miniplates
common miniplates for bone fracture fixation from the Trauma 2.0 series are shown in Fig 1.
The most left one, with article number M-4138, shows an overall length of 23 mm consisting
of two end ring elements and two middle ring elements. The model in the middle, with article
number M-4320, with the same properties as the left one but an extended bar of 9 mm and
thus extended to 29 mm in the overall length. Last, the most right one, with article number M4322, with two additional ring elements and an overall length of 35 mm.
Workflow overview
In this contribution, we implemented a MeVisLab network that generates and applies osteosynthesis materials such as commonly used miniplates. The materials, based on a seed point,
can be placed by the user, in a two-step approach. First, the user has to load the dataset into the
application to be able to set an initial point defining the implant’s center position, followed by
setting up the direction and choosing the appropriate implant model necessary for calculating
a baseline. This baseline shows the implant’s centerline [28], [29] and serves as a simplified
model saving computational time but still giving the user a clear and accurate response. If satisfaction is reached, the user visualizes the implant in a second step, which is further able to be
saved for parallel implant generation, or exported, for local and permanent storage. In the next
section this workflow is described in more detail, explaining the network and its single
modules.
Network
Fig 4 shows the constructed network, created within the MeVisLab platform. MeVisLab provides basic and advanced algorithms for medical image processing and visualization and also
includes an environment for individual module generation under C++ as well as an environment for implementing own user interfaces with the MDL-script (MeVisLab Description Language). Each block of the modular pipeline network represents a module with individual
functions. Three different colors represent the different types of available modules, where blue
represents the procession of WEM data, green modules mark an Open Inventor module,
working with visual scene graphs (3D), and last, the brown modules named macro modules
Fig 4. Constructed network using the MeVisLab prototyping environment, including selfimplemented modules named CurvatureCalc for baseline calculation and ImplantGen for generating
the implant models.
https://doi.org/10.1371/journal.pone.0182839.g004
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
6 / 19
Position planning of miniplates
hiding sub networks for preserving a clear network structure. Connections between modules
use two different connector types: Undirected lines, for a data connection, and directed ones,
indicated by arrows, for a parameter connection. Using data connections, it is important to
connect the output of one module, with an input of the same type of another module (note:
the data flow is bottom-up). Also, in- and outputs are distinguishable by three different types:
The squared ones for Base-objects, pointing to data structures, the triangulated ones for ML
images, and half circled ones for Inventor scenes. With the introduced modules and connections, the network was set up as follows: Starting at the bottom of the network, the WEMLoad
module loads the dataset into the network which is immediately rendered with the SoWEMRender1 module and visualized in the main viewer panel by the viewAll SoCustomExaminViewer. The self-implemented CurvatureCalc module holds the implementation for generating
the baseline. Therefore, in addition to the dataset itself, the network uses the user set point,
marking the implants center. This point is obtained by the So3DMarkerEditor, handling user
interactions in the viewer panel. The two outputs of the CurvatureCalc module hold once the
3D-position values for the baseline on the left, and once the according normal vectors at each
point on the right. Since the baseline is visualized continuously we use another So3DMarkerEditor named So3DMarkerEditor1. Next, the self-implemented ImplantGen module derives the
three-dimensional position of each baseline point and the according normals for generating
the final implant. Further, necessary do accomplish this task, the pre-constructed models of
the ring section, loaded with the additional WEMLoad modules, are connected to the input.
The additional SoWEMRenderer2 renders the generated implants which are passed on and
visualized together with the dataset. With the WEMSave module, it is possible to store the generated miniplates locally in STL format for further use or for 3D printing. Yet, three modules
are not described. SoMouseGrabber, SoOrientationInset and SoOrientationModel handle the
rotation of the mouse wheel which’s value is passed to the CurvatureCalc module using a
parameter connection. In addition, the two self-implemented modules are connected with a
parameter connection, which is important to further pass on the selected implant type.
User interface (UI)
As unintended user interaction with the network may lead to network errors, we packed the
constructed network into a single macro module. The module also holds a MDL script file
implementing the user interface (UI) as illustrated in Fig 5. The UI is separated in two main
parts: The control panel, including all interaction fields which actual change the data, and the
viewer panel including all visualizing interactions.
Generating pipeline
Fig 6 illustrates the workflow pipeline which can be separated in three groups according to the
two procedural steps, implant setup and implant generation, and the initial data acquisition,
all described in detail in the following paragraphs:
1. Data acquisition—Orange. First, the user has to load the patient’s dataset by choosing
a path to the locally stored WEM file by clicking on the browse button of the Loadfile field in
the user interface, resulting in already visualizing the data set in the user interface’s viewer
panel. The viewer implements already some camera controls, like zooming and rotating the
camera.
2. Implant set up—Grey. Second, the user sets up the implant’s properties by defining the
center point of the implant through placing an initial point on the skull. This is accomplished
by holding the ALT key and clicking on the desired location on the 3D models surface. The
Baseline Calculation algorithm (described in section Module: CurvatureCalc) starts now
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
7 / 19
Position planning of miniplates
Fig 5. User interface, separated in controls and viewer panels. The Viewer box shows the data set (white skull) attached with a
miniplate (gold) and the baseline (green).
https://doi.org/10.1371/journal.pone.0182839.g005
calculating the Baseline (green dotted line) with an initial orientation that depends on the
mouse wheel value and the position of the central point. This baseline shows the location,
direction and the approximated length of the later generated implant including the underlying
curvature and serves as a simplified version of the resulting implant. Hence, giving already an
idea on how the implant will be applied. In Fig 5, the resulting baseline (green dotted line) is
illustrated where the center position is placed on the forehead. Further, the length of this line
depends on the selection of the implant model, by marking one of the radio buttons in the user
interface’s control box on the left side. Since they vary in length, different models set the baseline’s length to different values according to the model’s dimensions. Additionally, by turning
the mouse wheel, the implant can be rotated, thus the direction of the baseline alters to match
the users desired placement and orientation. The field wheel step lets the user choose the accuracy of this rotation for high precision positioning. Also, the initial point, and thus the implant
itself, can be readjusted at any time by holding the ALT key and clicking on the new position
on the data set’s surface. We chose the intermediate step of visualizing the baseline first rather
than calculating the full implant directly to gain a more comfortable user experience since the
generation of the model utilizes several seconds of time resources, resulting in unwanted latencies that we want to avoid. With this compromise, the user gains a fast visual response of the
implant’s placement and adjustments are realized close to real-time. Moreover, this method
does not lack any characteristics a fully generated implant model would provide, as the baseline holds all the necessary information for optimal placement of a miniplate. If satisfaction
concerning the setup is reached, the generation of the implant follows, as described in the subsequent section.
3. Implant generation—Apricot. This step uses the baseline that is setup in the previous
section to generate the final implant model. Actuating the ON/OFF toggle button in the UI initializes this operation. In the overview workflow diagram this interaction accords to the green,
diamond shaped block named Visualize ON/OFF. In addition to the baseline properties
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
8 / 19
Position planning of miniplates
Fig 6. Flow chart of a miniplate generation process separated in three main parts: Orange marks the
data acquisition, grey the implant set-up including the baseline calculation and apricot the implant
generation and visualization.
https://doi.org/10.1371/journal.pone.0182839.g006
position, direction and implant model, this algorithm uses the implant’s ring sections, namely
the end and middle elements for constructing the final implant (details can be found in section
Module: ImplantGen). Depending on the implant model’s dimensions the positions and orientations of each ring element are sequenced calculated followed by further generating of the
bridge sections, also according to the model’s specifications. The ring elements have been preconstructed in Autodesk’s Inventor (http://www.autodesk.com/products/inventor/overview).
Again, Fig 5 shows a final result of this procedural step where the implant is visualized as a
golden structure. Currently, for the proof of concept, the selection of available implants is limited to different models from the Medartis Modus 2.0 series, which are characterized in the
product catalogue. Despite the fact that the implant’s setup can be altered during every step, it
has to be kept in mind, that rotating the implant or adjusting the position in visualizing mode
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
9 / 19
Position planning of miniplates
takes a lot of computational afford and thus may result in latencies. For optimal usage, it is
therefore recommended to adjust these characteristics only during inactivated visualization. In
the case of observing a satisfying result, the user can save the implant to continue setting up
another implant while visualizing the previous generated and saved ones. This is accomplished
by using the Save button, resulting in saving the implant for the current session only. This
method allows the generation and visualization of multiple implants, or in other words, this
function allows visualizing more than one implant at a time and is important in medical cases
where fracture treatment uses more than one miniplate as it is the case e.g. by suffering from
LeFort fractures [30]. However, it is important to understand the difference between the Save
and Export buttons. The Export button does export all implants, including the current (may
not be saved yet) and all saved ones to the path, given in the Savefile field, as STL file. Again,
export actually also means to save, but saves the implants permanently on the local drive in
STL format, rather than the Save button which does save the currently processed implant for
this session to the array of other already saved implants. This means, that as long as an implant
is not saved, it is possible to be readjusted. Another option, which is permanently able to be
used, as long as the baseline is already generated, is to change the baseline’s marker points in
any arbitrary direction by just changing its position using drag and drop interactions.
Within this section the reader should have gained an understanding on how the software
tool works and is used for planning the adaption of osteosynthesis material in the face. However, together with the description of the self-implemented algorithms of the CurvatureCalc
and the ImplantGen modules in the following sections, the reader gets a comprehensive
description on how the underlying processes are assembled.
Module: CurvatureCalc—The baseline that is visualized as a very basic structure in form of
markers is calculated in the CurvatureCalc module. The baseline is a simplified model of the
final implant but already showing the most important characteristics like position, direction
and curvature. Since the calculation of the baseline is less consuming in computational costs,
the generation and adjustment of the baseline happens much faster than a direct visualization
of the final implant. The method is based on the work of Möller and Trumbore [31], [32] and
their calculation method of ray-triangle intersections. The method casts a ray from the origin
O of the cast ray in direction D and then changes the base of that vector that yields a vector
(t u v)T, where t represents the distance to the plane in which a triangle lies and u,v as the coordinates inside the triangle. The big advantage is that the computation of the plane equation is
not performed resulting in significant memory savings. This method is used to determine the
baseline where not only one ray but also a cascade of parallel rays is cast, like shown in Fig 7.
The object’s surface, or in other words the patient’s skull, is represented by the light blue triangulated mesh. The red star in the figure illustrates the initially set point as a starting value.
From this point, first the normal vector N (purple) according to the nearest triangle is elicited.
Next, the direction vector D1 is determined by calculating the cross product of the normal vector N (which is the same for the implant and the baseline position) and the baselines direction
vector D, which are not necessarily perpendicular to each other. The origin of D2 is then calculated according to the implant model’s measures, parallel to D1. Starting from this vector’s origin, a cascade of rays is cast along vector D2’s direction, which means that the origins of the
rays, used for ray-triangle intersection, are located along vector D2. The next step is to cast
rays from the mentioned origins in direction of the negative normal vector -N. Now, according
to Möller and Trumbore, the ray-triangle intersections can be located to set marker at those
points (light blue stars), representing the final baseline. As a result, the baseline points’ positions in the three-dimensional space are applied to one output of the module as a vector where
each entry holds the three-dimensional vector to the points’ position. Additionally, the
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
10 / 19
Position planning of miniplates
Fig 7. Illustration of the baseline determination. The direction vector D (green) is set up parallel to the
initial point’s (red star) direction vector (green). The direction vector is perpendicular to the initial point’s
normal vector (purple). The origins of the cast rays (brown) are translated along the direction vector and
checked for intersections.
https://doi.org/10.1371/journal.pone.0182839.g007
according normal vectors to each baseline point are connected to the second output, again as a
vector where each entry holds a three-dimensional vector.
Module: ImplantGen–This module uses the information derived from the CurvatureCulc
module described in the previous section to set up the final miniplate model in a piecewise
manner. Therefore, first two ring sections are brought into position followed by building the
connecting bridge section rather than building the first bridge element after the first ring element. The generation process is executed not only with respect to the baseline characteristics,
but also includes the information of the basepoint’s normal vectors, the selected implant type
and the pre-constructed ring section models. The specific implant model distance from the
center (initial marker position) along the baseline is calculated for the exact placement of the
first ring element. Then, this point serves as a starting point where the first ring is located. Following ring sections are placed again along the baseline but in opposing direction where the
previous located element is used as starting point for distance calculation. Correct alignment of
orientation and direction of each section is ensured by using the normal vector marker list,
where the ring element’s normal vector N is adjusted to align the baseline’s normal vector N at
the position of correct placement, thus ensuring perfect alignment. Since the ring sections are
not symmetric due to at least one plane side for bridge section attachment, this “open” sides
need to be orientated correctly as well. Therefore, the direction vectors D of both, the ring element and the current baseline position are utilized where the direction vector of the baseline
shows the direction to the next marker and the implant’s direction vector describes the connecting end of the ring. Fig 8 illustrates this idea by showing the baseline (green) on the
patient’s skull, together with a ring element and described vectors N and D. For correct orientation of the ring elements, the two direction vectors are aligned in only two dimensions. Important to note is, that this is not a crucial task, since the baseline’s normal vector and direction
vector are not perpendicular to each other, which is different to the ring section where this is
always the case. Therefore, the direction vectors are only aligned in two dimensions since it is
important to keep the normal vectors aligned in all three dimensions. A final alignment shows
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
11 / 19
Position planning of miniplates
Fig 8. The patient’s skull (white) with an applied baseline (green) are shown on the left side, where the
normal vector N and the direction vector D are drawn for the placement of the ring elements. The ring
element itself is shown in the right image, where the normal vector N and direction vector D (red), which have
to be aligned with the vectors from the baseline, are illustrated.
https://doi.org/10.1371/journal.pone.0182839.g008
all four vectors in the same plane which is spanned by the two normal vectors, congruent with
each other, the perpendicular direction vector of the ring model and the baseline’s direction
vector. This is how each ring element is aligned and orientated correctly depending on the
underlying surface geometries. However, to finish an implant, the bridge sections between
those rings elements need to be generated as well: For building these connections, we use at
least two triangulated blocks per sections. These blocks are connected to the corners of the
plane sides of each ring element, illustrated as green crosses in Fig 9. Thus, the diameter shows
a rectangular shape. First, the four corner positions of where the bridge elements are connected
to the ring elements are determined. Since the ring’s geometries are known, it is sufficient to
forward the rings center position as well as the normal and the direction vector each, to the
determining method, which applies distance measurements using the L2-norm. Next, at each
baseline point between those two plane sides (green dots in Fig 9) a rectangle with the same
dimensions as the ring element’s plane rectangle is spanned (light blue rectangles in Fig 9). The
Fig 9. Bridge building principle with triangulation. Green crosses mark the corner of the ring element’s
plane side. To each baseline point (green points) a rectangle (light blue lines) is constructed. Rectangle
triangulation, spanned by the two starting nodes (N1 and N2) and two opposing nodes of the previous
constructed rectangle (N3 and N4). Node N1, N3 and N4 build a triangle T2, and Nodes N1, N2 and N3 build
the second triangle T1.
https://doi.org/10.1371/journal.pone.0182839.g009
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
12 / 19
Position planning of miniplates
orientation is defined by the basepoint’s direction and normal vector. As a result, the basis for
finalizing the bridge element using triangulation is built. In a next step, the starting, ending
and the corner positions of the rectangles are used for the triangulation task. To do so, first, a
surface consisting of four nodes has to be defined. As an example, one uses two neighboring
nodes from the starting rectangle (N1 and N2) and the opposing ones of the first constructed
rectangle (N3 and N4) where further triangulation is performed. This operation is performed
for all four surfaces of each block resulting in the final implant.
Results
The presented software was evaluated using ten CT data sets performed out of diagnosis and
treatment reasons in the clinical routine at a department of Oral and Maxillofacial Surgery.
The study participants for our software were clinical experts from the Department of Oral and
Maxillofacial Surgery, Medical University of Graz, Austria and technical experts from the
Institute of computer Graphics and Vision from the Faculty of Computer Science and Biomedical Engineering of Graz, University of Technology, Austria (TU Graz). We included physicians and computer scientists into the study to have different points of view on the software
solution and its interface. However, all experts are familiar with the state of the art techniques
in facial reconstructions and the performed evaluation consisted of three main steps. First,
during the so-called training phase, the developer of the prototype explained the software by
demonstrating an example planning case and explaining the functionality of the software. In
the next step, the physicians performed the planning of an implant on their own, where we
described the case to be solved with the following words:
Load the given data set and perform the placement of a mandibular media fracture using at
least two implants of different model type and export the data to the desktop, if satisfaction is
reached.
Subsequently, the testers were asked to answer a questionnaire containing eleven questions
on a six point Likert-scale with increasing consent from one to six [33] to rate the overall software solution. Generally, the questionnaire evaluates the software ergonomic formulated in
general terms without reference to situations of use, application, environment or technology
[34]. However, questions like “The software does not need a lot of training time” are seen in
context of the current user and his/her experience/knowledge in miniplate planning (in
respect to other tools or their medical experiences). One question (question 11) had to be
answered with yes or no. The questions are listed below and are adopted from the software
evaluation questionnaire of ISO Norm 9241/110 [35].
1. The software does not need a lot of training time
2. The software is adjusted well to achieve a satisfying result
3. The software provides all necessary functions to achieve the goal
4. The software is not complicated to use
5. How satisfied are you with the UI surface?
6. How accurate was the placement of the implant?
7. How satisfied are you with the presented result?
8. Was it easy to adjust the implant? (position, orientation, model, ...)
9. How satisfied have you been with the time consumed? (no training)
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
13 / 19
Position planning of miniplates
Fig 10. Result of the questionnaire derived from Table 2. The x-axis represents the questions and the
time. The y-axis shows the rating score of each question together with the superimposed standard error.
https://doi.org/10.1371/journal.pone.0182839.g010
10. How is your Overall impression
11. Would you use the software in a daily routine?
(Assumed that the 3D printed implant would fit with just a few adjustments)
12. Time used for gaining a satisfying result
The result of this questionnaire is shown in Fig 10. The x-axes represents the questions
from one to eleven (Q1—Q11) and the measured time T. The y-axis represents the median
score for Q1 to Q10. In addition, the median value for Q11 is shown but represents a yes or no
answer, where yes equals a value of six points and no equals a value of zero points. The time is
represented in minutes. For all data the superimposed standard error is represented by the
white lines. In Table 2 the underlying data is represented, where subject two and three represent the surgeons’ results.
Table 2. Results from the evaluation questionnaire for each subject using a six point Likert rating.
Median is the median value of the four subjects and Error describes the superimposed standard error. Q11 is
a yes/no question where no equals a value of zero and yes equals a value of 6. The time T is measured in
minutes.
No.
Subject
1
2
3
Median
Error
0
4
Q1:
6
6
6
6
6.00
Q2:
6
6
6
6
6.00
0
Q3:
6
5
6
6
5.75
0.21
Q4:
4
6
6
6
5.50
0.43
Q5:
6
6
6
6
6.00
0
Q6:
6
6
6
6
6.00
0
Q7:
6
6
6
6
6.00
0
Q8:
5
6
6
5
5.50
0.25
Q9:
6
6
6
6
6.00
0
Q10:
6
6
6
6
6.00
0
Q11:
6
6
6
6
6.00
0
T:
4
5.5
3.5
4.2
4.00
4.30
https://doi.org/10.1371/journal.pone.0182839.t002
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
14 / 19
Position planning of miniplates
Fig 11. Baseline (green) on more complex surfaces. On the left image a wave like surface is shown, and
on the right side the baseline adjusts to a ninety degree angle.
https://doi.org/10.1371/journal.pone.0182839.g011
Moreover, the baseline determination is very robust and is designed to capture the curvature of the skull’s surface independent of the surface curvature. Fig 11 shows that the baseline
does adjust itself also on very difficult surfaces like a wave form or even 90 degree angles. In
Fig 12, an example result is shown, containing three available implants, each located on a position where they are generally applied (left). For a better illustration, the single implants are
extended portrayed on the right side as well. Additionally, Fig 13 shows the same outcome
from its sagittal and coronal perspective.
Discussion
In the proposed work, a workflow and software tool for position planning of miniplates to
treat facial bone defects and reconstructions was implemented, which can be used in the clinical routine. Therefore, a modular network with the medical imaging platform MeVisLab was
Fig 12. Example result containing all three available implants each located at a position where they generally
are applied (left). For better illustration, the single implants are extended portrayed as well on the right side.
The plate could easily be positioned according to the underlying bone contours by the used software module.
https://doi.org/10.1371/journal.pone.0182839.g012
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
15 / 19
Position planning of miniplates
Fig 13. Result from Fig 12 in its sagittal (left) and coronal (right) perspectives.
https://doi.org/10.1371/journal.pone.0182839.g013
constructed that is able to load a patient’s CT-data set and plan the positioning of commonly
used miniplates.
The application of miniplates in the facial area can be challenging especially in complex surgical cases. This lacks not only on appropriate (free) software, but also state of the art techniques show some space for improvement. Furthermore, commercial software products do
not provide an option for the application of commonly used osteosynthesis material but rather
focus on the generation of patient specific implants manufactured and sold by the industry. In
this contribution, we have shown that with the use of MeVisLab it is possible to generate a custom network allowing preoperatively an easy and precise virtual placement of commonly used
miniplates. Furthermore, it is possible to use the outcome together with 3D printing technologies since the planned implant can be exported as an STL-file [17]. The 3D printed implant
can be used for an intraoperatively performed direct adaption of the bone plate according to
its function as a bending tool. In summary, the proposed contribution achieved the following
research highlights:
• The successful interactive planning and reconstruction of facial 3D implants;
• A precise capture of the surface curvature as a basis for any type of implant;
• Flexibility in model types as well as in positioning and orientating the miniplates;
• Evaluation with real patient Computed Tomography (CT) data from the clinical routine;
• Enabling 3D implant export as Computer-Aided Design (CAD) file format for 3D printing
and in-depth pre-surgical assessment;
• Providing clinical datasets and source code to the research community for own usage.
The intention of this contribution was to work out a workflow and establish, and evaluate a
software for the position planning of miniplates. A next step will be to test the software in the
clinical routine by using the planned miniplate positions in operations. Future works include
also the implementation of further implant types and even patient specific ones [36] to provide
to the user an even higher flexibility. In addition, bone repositioning [37] is of high priority in
future research, since fractures in the facial area mostly suffer from dispositions where
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
16 / 19
Position planning of miniplates
miniplates may not be deployed without further ado. Moreover, performing the miniplate generation on the GPU [38], [39] to enable an interactive placement of the implant and thus skipping the baseline step. Finally, it is necessary to perform a more comprehensive evaluation of
the software/workflow and the planning could be performed in Virtual Reality (VR) [40]
whereas the intraoperative deployment of the miniplates could be performed in an Augmented
Reality (AR) setting [41].
Acknowledgments
Blueprints of the miniplate dimensions were provided by Jan Stanzel from the MedArtis AG,
Hochbergerstraße 60E, 5057 Basel, Switzerland. A tutorial video demonstrating the interactive
planning of miniplates can be found under the YouTube channel: https://www.youtube.com/
c/JanEgger/videos
Author Contributions
Conceptualization: Jan Egger, Jürgen Wallner, Markus Gall, Knut Reinbacher.
Data curation: Jürgen Wallner, Xiaojun Chen, Knut Reinbacher.
Formal analysis: Jan Egger, Jürgen Wallner, Markus Gall, Xiaojun Chen, Knut Reinbacher.
Funding acquisition: Katja Schwenzer-Zimmerer, Dieter Schmalstieg.
Methodology: Markus Gall.
Project administration: Jan Egger, Xiaojun Chen, Katja Schwenzer-Zimmerer, Dieter
Schmalstieg.
Resources: Jan Egger, Jürgen Wallner, Xiaojun Chen, Katja Schwenzer-Zimmerer, Dieter
Schmalstieg.
Software: Markus Gall.
Supervision: Jan Egger, Xiaojun Chen, Katja Schwenzer-Zimmerer, Dieter Schmalstieg.
Validation: Jürgen Wallner, Markus Gall, Knut Reinbacher.
Visualization: Markus Gall.
Writing – original draft: Jan Egger, Markus Gall.
Writing – review & editing: Jan Egger.
References
1.
Ritacco LE, Milano FE, Chao E (editors). Computer-Assisted Musculosketal Surgery. Springer Press,
2015.
2.
Tuli T. Facial trauma: How dangerous is skiing and snowboarding. American Association of Oral and
Maxillofacial Surgeons 2010; 68(2): 293–299.
3.
Gassner R, Ulmer H, Tuli T, Emshoff R. Incidence of oral and maxillofacial skiing injuries due to different
injury mechanisms. J Oral Maxillofac Surg. 1999; 57(9): 1068–1073. PMID: 10484107
4.
Hidalgo DA. Titanium miniplate fixation in free flap mandible reconstruction. Annals of Plastic Surgery
1989; 23(6):469–507.
5.
Rasse M. Traumatologie des Mund-, Kiefer-, Gesichtsbereich. Chapter: Grundlagen der Traumatologie,
Springer Verlag, 2014; 27–82.
6.
Bell RB. A comparison of fixation techniques in oro-mandibular recon- struction utilizing fibular free
flaps. Journal of Oral and Maxillofacial Surgery 2007; 65(9): 39.
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
17 / 19
Position planning of miniplates
7.
Li P, Tang W, Liao C, Tan P, Zhang J, Tian W. Clinical evaluation of computer-assisted surgical technique in the treatment of comminuted mandibular fractures. Journal of Oral and Maxillofacial Surgery,
Medicine, and Pathology 2015; 27(3): 332–336.
8.
Egger J, Tokuda J, Chauvin L, Freisleben B, Nimsky C, Kapur T, et al. Integration of the OpenIGTLink
network protocol for image-guided therapy with the medical platform MeVisLab. Int J Med Robot. 2012;
8(3): 282–90. https://doi.org/10.1002/rcs.1415 PMID: 22374845
9.
Greiner K, Egger J, Großkopf S, Kaftan JN, Dörner R, Freisleben B. Segmentation of aortic aneurysms
in CTA images with the statistic approach of the active appearance models. Bildverarbeitung für die
Medizin (BVM) 2008; 51–55.
10.
Egger J, Lüddemann T, Schwarzenberg R, Freisleben B, Nimsky C. Interactive-cut: real-time feedback
segmentation for translational research. Comput. Med. Imaging Graphics 2014; 38(4): 285–295.
11.
Lu J, Egger J, Wimmer A, Großkopf S, Freisleben B. Detection and visualization of endoleaks in CT
data for monitoring of thoracic and abdominal aortic aneurysm stents. SPIE Medical Imaging 2008;
69181F.
12.
Schwarzenberg R, Freisleben B, Nimsky C, Egger J. Cube-Cut: Vertebral Body Segmentation in MRIData through Cubic-Shaped Divergences. PLoS ONE 9(4): e93389. https://doi.org/10.1371/journal.
pone.0093389 PMID: 24705281
13.
Egger J. PCG-Cut: Graph Driven Segmentation of the Prostate Central Gland. PLoS ONE 8(10):
e76645. https://doi.org/10.1371/journal.pone.0076645 PMID: 24146901
14.
Egger J, Kapur T, Fedorov A, Pieper S, Miller JV, Veeraraghavan H, et al. GBM Volumetry using the 3D
Slicer Medical Image Computing Platform. Sci Rep. 2013; 3: 1364. https://doi.org/10.1038/srep01364
PMID: 23455483
15.
Egger J, Kapur T, Nimsky C, Kikinis R. Pituitary Adenoma Volumetry with 3D Slicer. PLoS ONE 7(12):
e51788. https://doi.org/10.1371/journal.pone.0051788 PMID: 23240062
16.
Gopi M, Krishnan S, Silca CT. Surface reconstruction based on lower dimensional localized Delaunay
triangulation. Computer Graphics Forum 2000; 19(3): 467–478.
17.
Chua CK, Leong KF, Lim CS. Rapid Prototyping Principles and Applications. World Scientific 2000, 3.
Edition.
18.
Zhang YS. 3D bio printing for tissue and organ fabrication. Annals of Biomedical Engineering 2016;
1–16.
19.
Little SH. 3D printed modeling for patient-specific mitral valve intervention repair with a clip and a plug.
JACC: Cardiovascular Interventions 2016; 9(9): 973–975. https://doi.org/10.1016/j.jcin.2016.02.027
PMID: 27151611
20.
Zhao L, Patel PK, Cohen M. Application of Virtual Surgical Planning with Computer Assisted Design
and Manufacturing Technology to Cranio-Maxillofacial Surgery. Arch Plast Surg. 2012; 39(4): 309–
316. https://doi.org/10.5999/aps.2012.39.4.309 PMID: 22872832
21.
Parthasarathy J. 3D modeling, custom implants and its future perspectives in craniofacial surgery. Ann
Maxillofac Surg. 2014; 4(1): 9–18. https://doi.org/10.4103/2231-0746.133065 PMID: 24987592
22.
Bell RB. Computer planning and intraoperative navigation in cranio- maxillofacial surgery. Oral and
Maxillofacial Surgery Clinics of North America 2010; 22(1): 135–156. https://doi.org/10.1016/j.coms.
2009.10.010 PMID: 20159483
23.
Adamus J, Lacki P. Forming of the titanium elements by bending. Computational Materials Science
2011; 50: 1305–1309.
24.
Burghart CR, Neukirch K, Hassfeld S, Rembold U, Woern H. Computer aided planning device for preoperative bending of osteosynthesis plates. Stud Health Technol Inform. 2000; 70: 46–52. PMID:
10977582
25.
Mildenberger P, Eichelberg M, Martin E. Introduction to the DICOM standard. European Radiology
2002; 12(4): 920–927. https://doi.org/10.1007/s003300101100 PMID: 11960249
26.
Gall M, Li X, Chen X, Schmalstieg D, Egger J. Cranial Defect Datasets. ResearchGate 2016.
27.
Medartis Modus 2.0 product catalogue. Website (last accessed in January 2017): http://www.medartis.
com/uploads/MANDIBLE-08000000_v0_d.pdf
28.
Egger J, Mostarkic Z, Grosskopf S, Freisleben B. A fast vessel centerline extraction algorithm for catheter simulation. Twentieth IEEE International Symposium on Computer-Based Medical Systems
(CBMS’07) 2007; IEEE Press, 177–182.
29.
Egger J, Mostarkic Z, Grosskopf S, Freisleben B (2007) Preoperative Measurement of Aneurysms and
Stenosis and Stent-Simulation for Endovascular Treatment. IEEE International Symposium on Biomedical Imaging (ISBI): From Nano to Macro 2007; 392–395.
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
18 / 19
Position planning of miniplates
30.
Rasse M. Traumatologie des Mund-, Kiefer-, Gesichtsbereich. Chapter: Spezielle Traumatologie 2014;
Springer Verlag, 27–82.
31.
Möller T, Trumbore B. Fast, Minimum storage ray/triangle intersection. Journal of Graphics Tools 1997;
2(1): 21–28.
32.
Egger J, Mostarkić Z, Maier F, Kaftan JN, Grosskopf S, Freisleben B. Fast Self-Collision Detection and
Simulation of Bifurcated Stents to Treat Abdominal Aortic Aneurysms (AAA). 29th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society 2007; IEEE Press, 6231–6234.
33.
Trochim WM. Likert Scaling. Research Methods Knowledge Base, 2nd Edition 2006.
34.
Prümper J. Der Benutzungsfragebogen ISONORM 9241/10: Ergebnisse Zur Reliabilität und Validität.
In: Liskowsky R. u.a. (Hg.): Software-Ergonomie ’97, Stuttgart.
35.
DIN EN ISO 9241/110:2006 Ergonomics of Human System Interaction. Website (last accessed in January 2017): http://www.ergo-online.de/site.aspx?url=html/software/verfahren_zur_beurteilung_der/
fragebogen_isonorm_online.htm
36.
Chen X, Xu L, Yang Y, Egger J. A semi-automatic computer-aided method for surgical template design.
Sci. Rep. 2016; 6: 20280. https://doi.org/10.1038/srep20280 PMID: 26843434
37.
Olsson O, Nysjö F, Hirsch JM, Carlbom IB. A Haptics-Assisted Cranio-Maxillofacial Surgery Planning.
Int J Comput Assist Radiol Surg. 2013; 8(6): 887–894. https://doi.org/10.1007/s11548-013-0827-5
PMID: 23605116
38.
Smistad E, Falch TL, Bozorgi M, Elster AC, Lindseth F. Medical image segmentation on GPUs—A comprehensive review. Medical Image Analysis 2015; 20: 1–18. https://doi.org/10.1016/j.media.2014.10.
012 PMID: 25534282
39.
Röttger D, Egger J. GPU-Based Isosurface Calculation for Real-time Ablation Zone Navigation.
Research Disclosure 2013; ID 591066, Published in the Research Disclosure Journal 7/13, Published
digitally 20 June 2013.
40.
Egger J, Gall M, Wallner J, Boechat P, Hann A, Li X, et al. HTC Vive MeVisLab integration via OpenVR
for medical applications. PLoS ONE 12(3): e0173972. https://doi.org/10.1371/journal.pone.0173972
PMID: 28323840
41.
Chen X, Xu L, Wang Y, Wang H, Wang F, Zeng X, et al. Development of a surgical navigation system
based on augmented reality using an optical see-through head-mounted display. J Biomed Inform.
2015; 55: 124–31. https://doi.org/10.1016/j.jbi.2015.04.003 PMID: 25882923
PLOS ONE | https://doi.org/10.1371/journal.pone.0182839 August 17, 2017
19 / 19
| 5cs.CE
|
Closed-Loop Statistical Verification of Stochastic Nonlinear Systems
Subject to Parametric Uncertainties
arXiv:1709.06645v2 [cs.SY] 1 Oct 2017
John F. Quindlen1
Ufuk Topcu2
Abstract— This paper proposes a statistical verification
framework using Gaussian processes (GPs) for simulationbased verification of stochastic nonlinear systems with parametric uncertainties. Given a small number of stochastic simulations, the proposed framework constructs a GP regression
model and predicts the system’s performance over the entire
set of possible uncertainties. Included in the framework is a
new metric to estimate the confidence in those predictions
based on the variance of the GP’s cumulative distribution
function. This variance-based metric forms the basis of active
sampling algorithms that aim to minimize prediction error
through careful selection of simulations. In three case studies,
the new active sampling algorithms demonstrate up to a 35%
improvement in prediction error over other approaches and are
able to correctly identify regions with low prediction confidence
through the variance metric.
I. I NTRODUCTION
Ever increasing demands for high performance necessitates the adoption of advanced, nonlinear control techniques
in a wide variety of engineering systems. The complexity of
these controllers, coupled with the increasing complexity of
the open-loop systems themselves, complicates verification
that the closed-loop system actually satisfies the performance
requirements at off-nominal conditions, called parametric uncertainties. Control system verification analyzes the closedloop system and attempts to identify at which parametric
uncertainties the resulting trajectory fails to meet the requirements.
Traditional analytical and numerical verification techniques [1], [2] provably verify the system’s performance
at various uncertainties, but require modeling assumptions
that restrict their suitability to complex systems with highly
nonlinear dynamics and hybrid controllers. Recent work
[3]–[5] addressed some of these concerns, but are limited
by the size of the state and parameter spaces. Stochastic
dynamics further challenge these approaches and restrict
their applicability. For comparison, statistical verification
techniques [6], [7] replace provable guarantees with looser
probabilistic bounds, but are capable of handling a much
broader range of problems with fewer restrictions. Statistical
verification methods rely upon extensive simulations of the
system to test the likelihood of requirement satisfaction at
various conditions and are well-suited to stochastic systems.
1 Ph.D.
Candidate, Department of Aeronautics and Astronautics, Massachusetts Institute of Technology (MIT).
2 Assistant Professor, Department of Aerospace Engineering and Engineering Mechanics, University of Texas at Austin
3 Assistant Professor, Departments of Agricultural & Biological Engineering
and Aerospace Engineering, University of Illinois Urbana-Champaign
4 Richard C. Maclaurin Professor of Aeronautics and Astronautics, Laboratory for Information and Decision Systems (LIDS), MIT
Girish Chowdhary3
Jonathan P. How4
Recent work developed data-driven statistical verification
techniques [8], [9] to efficiently verify deterministic closedloop systems without relying upon exhaustive simulations.
These data-driven methods construct machine learning models from small sets of simulation trajectories and predict the
response at unobserved parameter settings. The latest work
[10] introduced Gaussian process (GP) prediction models and
novel active sampling procedures to improve the accuracy of
predictions given limited amounts of simulation data. The deterministic GP-based approach also quantifies the prediction
confidence online without the need for external validation
datasets and carefully selects simulations to maximize the
prediction confidence.
This paper extends data-driven methods to stochastic
verification problems and provides three contributions: a
GP-based statistical verification framework for stochastic
systems, a metric to estimate the accuracy of the framework’s
predictions, and active sampling algorithms specifically developed for closed-loop statistical verification. Given a small
set of stochastic simulations, the first two contributions will
predict the probability of requirement satisfaction at all
possible parametric uncertainties and compute the confidence
in those predictions. The third contribution exploits this
prediction confidence to select subsequent simulations and
minimize prediction error. Three case studies are provided to
demonstrate the effectiveness of the approach to maximize
accuracy given a constraint on the number of simulations.
II. P ROBLEM D ESCRIPTION
Consider the following nonlinear system
ẋ(t) = f (x(t), u(t), θ, w(t)),
n
(1)
m
where x(t) ∈ R is the state vector, u(t) ∈ R is the control
input vector, θ ∈ Rp represents the parametric uncertainties,
and w(t) ∈ Rn is the stochastic noise. A candidate controller
generates control inputs according to some predetermined
function u(t) = g(x(t)). Given this controller, the resulting
closed-loop dynamics reduce to
ẋ(t) = fcl (x(t), θ, w(t)),
(2)
and the trajectory Φ(x(t)|x0 , θ) is a function of nominal
initial state x0 , parametric uncertainties θ, and stochastic
noise w(t). We assume the nominal initial state x0 is known
and fixed.
The parametric uncertainties θ can arise from a variety of
sources. For instance, these parameters may include system
properties such as vehicle mass and inertia that will typically
vary across different operational scenarios. The main impact
of the parametric uncertainties is as variable initial conditions
that determine the evolution of the subsequent state trajectory
Φ(x(t)|x0 , θ). These parametric uncertainties may not be
known during real-world execution of the system, but are
assumed to fall within some bounded, compact set Θ that is
known in advance.
Assumption 1 The set of all possible parametric uncertainties θ ∈ Θ is a known, compact set Θ ∈ Rp .
Assumption 1 ensures the system has known feasible
bounds on the uncertain conditions for the verification procedure to examine. Most physical systems will naturally have
bounds on the set of allowable operating conditions. For
instance, aircraft will typically fly within a maximum take-off
weight set by structural limitations and a minimum (empty)
weight.
The major difference between (2) and the deterministic
systems found in earlier problems [8]–[10] is the addition
of stochastic noise w(t) in the dynamics. This noise term
encapsulates any source of randomness, such as process
and/or measurement noise and ultimately breaks the previous
works’ fundamental assumption of deterministic (noise-free)
dynamics. The randomness will cause multiple trajectories
starting from the same x0 and θ to obtain different results.
The noise could have a large effect on the satisfaction of the
performance requirements at a given x0 and θ.
A. Measurements of Trajectory Performance
A trajectory must meet certain performance requirements
in order for it to be considered “satisfactory.” As before
[10], these requirements are supplied by relevant certification authorities and may include a wide range of potential
specifications with varying complexity. Commonly, the requirements can be written as signal temporal logic (STL)
[11] specifications, which also offers a convenient method
to quantify the satisfaction of the requirements. Central to
STL is the computation of a scalar robustness degree ρϕ ∈ R
that quantifies the minimum robustness of the trajectory with
respect to the requirement specification ϕ. Therefore, the
satisfaction of the requirements along the entire trajectory
is captured with a single scalar variable. Building upon
that concept, this work assumes the robustness of a trajectory Φ(x(t)|x0 , θ) to a given performance requirement is
measured with a single measurement variable y(θ), where
y(θ) = ρϕ in problems with STL specifications. Other
applications will specify the performance requirements and
quantify their satisfaction through different methods, as will
be seen in the task allocation example from Section V-B.
Assumption 2 The simulation model provides a single,
scalar output y(θ) ∈ R that quantifies the minimum robustness of trajectory Φ(x(t)|x0 , θ) to a specified performance
requirement. The sign of y(θ) indicates satisfaction of the
requirements, where y > 0 signifies “satisfactory” performance and y ≤ 0 signifies “unsatisfactory” performance.
Since the closed-loop trajectory is a function of nominal
(fixed) x0 and θ, we write the trajectory robustness measurements as an explicit function of θ to emphasize the
effect of the parametric uncertainties upon the robustness
of the trajectory. As θ changes, so will the satisfaction of
the performance requirement.
B. Probabilistic Satisfaction of Requirements
Although a closed-loop trajectory and its corresponding
robustness measurement are functions of parametric uncertainties, stochasticity will cause multiple trajectories at the
same θ setting to obtain different y(θ). If a large number
of simulations are performed at this setting, then the corresponding y(θ) measurements will produce a continuous
distribution of possible robustness values. The underlying
true distribution of y(θ) may take a variety of different
forms, but this work addresses the case with a Gaussian
distribution.
Assumption 3 The distribution of robustness measurements
y(θ) at every θ follows a Gaussian distribution y(θ) ∼
N (ȳ(θ), 2y ) with spatially-varying mean ȳ(θ) and constant
standard deviation y .
While this assumption does restrict the class of problems,
it simplifies the implementation details to improve reader
clarity and better highlight the contributions of the approach.
This paper is intended to introduce the statistical verification
framework and discuss the challenges and contributions on
clearer, and still relevant, verification problems. Upcoming
extensions will relax Assumption 3 to consider spatiallyvarying y [12] and non-Gaussian distributions [13], but also
require significantly more complex and sensitive statistical
inference techniques to model those distributions. The implementation details will change in these extensions, but the
fundamental approach remains the same.
From the verification perspective, where each trajectory
either satisfies the requirement or does not, the Gaussian
distribution of y(θ) creates a Bernoulli distribution for
the likelihood an arbitrary trajectory at θ will satisfy the
requirement.
Definition 1 The satisfaction probability function psat (θ) ∈
[0, 1] defines the probability an arbitrary trajectory initialized
with θ will satisfy the performance requirement.
This probability of satisfaction is the cumulative distribution of positive y(θ) measurements,
ȳ(θ)
1 1
. (3)
psat (θ) = P(y(θ) > 0) = + erf q
2 2
22y
While psat (θ) defines the probabilistic satisfaction of the
requirements, the challenge for stochastic verification is
psat (θ) will generally be unknown in advance. Instead,
the statistical verification goal is to compute an estimated
pbsat (θ) for all possible conditions in Θ and minimize the
difference between the predictions and psat (θ).
Simultaneously, statistical verification must also contend
with computational costs. In most applications, external
constraints restrict the computational budget allocated to
the statistical verification process. For instance, verification
typically relies upon high-fidelity models of complex systems
and subsystems that drive up the computational time for a
single trajectory. It is infeasible to simply saturate Θ with
multiple simulations at every θ because each simulation trajectory requires a certain amount of time and computational
resources. Regardless of the source, each application has
some feasible limit on the amount of time, resources, or
money allocated to the statistical verification process. This
work models the feasible limit as an upper bound Ntotal on
the number of simulation trajectories.
III. S TATISTICAL V ERIFICATION F RAMEWORK
Given the verification objective, there are multiple possible approaches to compute pbsat (θ) from a finite set of
simulation-based observations. The most basic form is to
treat each trajectory as a Bernoulli trial with binary evaluations {“satisfactory”, “unsatisfactory”} from the sign of
y(θ). While this approach can address all types of systems,
including those that do not meet Assumptions 2 and 3, a
binary-based approach requires multiple simulations at every
θ to construct a binomial distribution. Each repetition counts
against the sampling budget |L| ≤ Ntotal and means one
few simulation at another location. Although the cost of
each simulation is less of a problem for applications with
simple simulation models, the cost of training a prediction
model given a large number of simulations at each location
is non-negligible. Therefore, the large number of simulations
required for practical predictions with binomial distributions
restricts the tractability of those approaches.
Instead, this work entirely avoids the need for expensive binomial distributions in the relevant class of systems
through the direct use of y(θ) and its Gaussian probability
density function (PDF). The main idea is to exploit a
single stochastic trajectory at each training location and its
noisy measurement y(θ) to infer the underlying PDF, which
defines the cumulative distribution psat (θ). Although various
computational methods are possible [14], [15], this problem
is particularly well-suited to Gaussian process regression [16]
due to the Gaussian distribution of y(θ).
A. GP-based Prediction Model
The Gaussian process regression model follows a similar
format to the earlier work in verification of deterministic
systems [10]. A finite collection of N total simulation trajectories forms a training dataset L = {D, y} consisting of their
parameter settings D = {θ 1 , θ 2 , . . . , θ N } and corresponding
robustness values y = [y(θ 1 ), . . . , y(θ N )]T . Unlike the deterministic approach, these stochastic measurements require
the introduction of a Gaussian likelihood model, where the
GP does not model y(θ) directly, but infers the latent mean
ȳ(θ).
The training process constructs the GP regression model
from the information provided by the training dataset L.
More details are found in [16], but fundamentally the
training procedure uses Bayesian inference to place a posterior probability distribution on latent mean ȳ given L.
Assuming a zero-mean prior N (ȳ|0, K) and likelihood
model N (y|ȳ, 2y I), the posterior predictive distribution at
an arbitrary location θ ∗ follows a Gaussian distribution
P(ȳ(θ ∗ )|L, θ ∗ ) = N (µ(θ ∗ ), Σ(θ ∗ )). The posterior predictive mean µ(θ ∗ ) and covariance Σ(θ ∗ ) are given by
−1
µ(θ ∗ ) = KT∗ K + 2y I
y
(4)
−1
T
K∗ ,
Σ(θ ∗ ) = κ(θ ∗ , θ ∗ ) − K∗ K + 2y I
where κ(θ i , θ j ) is the scalar kernel function, K∗ is the N ×1
vector of κ(θ ∗ , θ i ) ∀i = 1 : N and K is the N × N matrix
for κ(θ i , θ j ) ∀i, j = 1 : N . Different choices are possible,
but this work uses the common squared exponential kernel
with automatic relevance determination (SE-ARD) [16] for
the kernel function.
The choice in GP hyperparameters may drastically change
the predictive distribution, even with the same training set
L, since these hyperparameters control the kernel function κ
and the likelihood model. Unfortunately, the optimal choice
of hyperparameters that perfectly replicates ȳ(θ) will not be
known in advance and must be estimated online. This work
uses maximum likelihood estimation (MLE) [16] to optimize
the kernel hyperparameters given only the information provided by L. The use of the SE-ARD kernel also enables the
MLE procedure to adjust the sensitivity of the kernel to each
element of θ as ȳ(θ) may have a higher sensitivity to certain
elements than others. Additionally, the likelihood model
contains a hyperparameter to estimate standard deviation
y as this term may also be completely unknown. If y is
unknown, then an estimate b
y may be computed in the same
MLE optimization process as the kernel hyperparameters or
separately by repeated sampling at a training locations in D.
1) Expected Probability of Satisfaction: Although µ(θ ∗ )
and Σ(θ ∗ ) define the predictive PDF for ȳ(θ ∗ ), (4) only
completes half the goal. Definition 1 requires the cumulative
distribution function (CDF) for y(θ ∗ ), not the PDF. The CDF
in (3) defined the true psat (θ), but this computation requires
perfect knowledge of ȳ(θ) and y . Instead, the predicted
satisfaction probability function pbsat (θ) marginalizes the
CDF over the posterior predictive distribution of ȳ(θ),
h
i
pbsat (θ ∗ ) = Eȳ(θ∗ ) P y(θ ∗ ) > 0|ȳ(θ ∗ )
µ(θ)
1 1
(5)
.
= + erf q
2 2
2(Σ(θ ∗ ) + 2y )
B. Prediction Confidence
While (5) computes the expected probability of requirement satisfaction, pbsat (θ) will likely fail to perfectly model
psat (θ). Therefore, the statistical verification framework
must not only provide pbsat (θ), but also indicate where
its confidence in the accuracy of these predictions is low.
Unsurprisingly, the true prediction error pesat (θ) = psat (θ)−
pbsat (θ) is unknown, but offline validation methods using
external validation datasets [14] can estimate pesat (θ) by
comparing the predictions against the validation set’s known,
true values. However, these external validation methods are
wasteful as they generally require the removal of valuable
training data for the independent validation set.
For comparison, probabilistic inequalities provide
theoretically-justified bounds on pesat (θ) without the use of
validation sets. In particular, Chebyshev’s inequality bounds
the prediction error by the variance of the CDF,
h
i
Vȳ(θ∗ ) P y(θ ∗ ) > 0|ȳ(θ ∗ )
, (6)
P(|e
psat (θ)| ≥ a) ≤
a2
where a > 0. Lower CDF variance will translate into lower
probabilistic bounds on pesat (θ) and thus higher confidence
in the accuracy of pbsat (θ). The primary challenge with
(6) is the variance lacks an analytical closed-form solution.
Fortunately, the variance can be approximated using a 1st or
2nd order Taylor series expansion of the nonlinear random
function [17]. The 1st order approximation is
1 −µ(θ∗ )2 /2y
e
Σ(θ ∗ ),
Vȳ(θ∗ ) P(y(θ ∗ ) > 0|ȳ(θ ∗ )) ≈
2π2y
(7)
which also happens to conservatively upper bound the 2nd
order approximation. For simplicity, we subsequently refer
to the approximate CDF variance in (7) as V (θ|L).
Despite the convenience of (7), the accuracy of an approximation of a nonlinear random function is limited. This inaccuracy makes it inadvisable to blindly substitute (7) into (6)
without careful consideration. However, the approximation
in (7) does still provide a perfect metric to identify regions
of Θ where the confidence in pbsat (θ) is low and provide
intuition on the sensitivity of pesat (θ). The power of (7) is its
value as a computationally-efficient, online validation tool for
signifying prediction confidence. For instance, approximate
variance is highest in regions with small |µ(θ)| and large
Σ(θ), meaning the confidence in the predictions should be
quite low. The next section will exploit the approximate
CDF variance to derive a closed-loop verification process
that seeks to minimize prediction errors.
IV. C LOSED -L OOP S TATISTICAL V ERIFICATION
As discussed in Section II, statistical verification will have
some feasible limit on the amount of time or computational
resources allocated to the overall process. This limit is approximated as a cap on the number of simulation trajectories,
|L| ≤ Ntotal . Given this restriction on the size of training
dataset L, the ideal scenario would only perform informative
simulations and carefully select all θ ∈ D to minimize the
prediction error over all θ ∈ Θ. However, the informationmaximizing training dataset is not apparent until after all
the trajectories have been obtained. To address that issue,
this work applies active learning [18] to iteratively select
informative θ settings for future simulations and minimize
the expected prediction error. To emphasize the iterative,
feedback-based nature of the active learning procedure, we
label the process closed-loop statistical verification.
Active learning describes a wide variety of different procedures [18]–[23], each with their own definition for the “best”
sample to run next. Most of these procedures focus on a particular aspect of the Gaussian PDF. For instance, procedures
focused on the PDF mean [9], [19] favor points with ȳ(θ)
near zero, meaning the best location is θ = argmin|µ(θ)|.
Although originally intended for binary classification with
support vector machines, such an approach does correctly
label points with low |µ(θ)| as informative since the CDF
variance (7) is high there. Likewise, PDF variance-based
approaches and extensions [20]–[23] are significantly more
common for GP methods and aim to reduce the PDF variance
Σ(θ). These approaches favor points with high variance,
θ = argmax Σ(θ), which also correctly emphasizes points
with comparatively high CDF variance since Σ(θ) will
be large. Although they both correctly emphasize certain
aspects, neither of those two approaches explicitly minimizes
the posterior CDF variance.
A. Reduction in CDF Variance
Earlier work in deterministic closed-loop verification [10]
specifically developed new sample selection metrics to maximize prediction confidence. Even though the implementation
details have changed, this work utilizes the same underlying
motivation and attempts to minimize the approximate CDF
variance (7) in order to maximize the confidence in pbsat (θ).
The ideal selection criterion would minimize the maximum
posterior CDF variance V (θ|L+ ) after the new sample data
at θ was added, but this requires the posterior training set
L+ = L∪{θ, y(θ)} to be known apriori. Prior knowledge of
L+ is an impossible proposition since y(θ) cannot be known
before a simulation has actually been performed there. Even
if the expected posterior set Lb+ = L ∪ {θ, µ(θ)} replaces
infeasible L, the high computational cost of retraining the GP
at every prospective sample location, nominally an O(N 3 )
operation and O(N 2 ) at best, renders the approach computationally intractable.
A more computationally tractable approach maximizes the
local improvement in posterior CDF variance, effectively
selecting the θ setting which will experience the greatest
reduction in CDF variance if a simulation is performed at
that location. The local change in CDF variance is labeled by
Ve (θ|L) = V (θ|L)−V (θ|Lb+ ). Although the change Ve (θ|L)
requires expected posterior information Lb+ , it does not
require the GP model to be retrained, which was the source of
the previous computational intractability. The local posterior
CDF variance at location θ ∗ after a measurement there can be
written purely in terms of the current information. After the
Woodbury matrix identity, the expected posterior covariance
Σ(θ ∗ )+ reduces to
Σ(θ ∗ )
Σ(θ ∗ )+ = Σ(θ ∗ ) 1 −
(8)
Σ(θ ∗ ) + 2y
while mean µ(θ ∗ )+ = µ(θ ∗ ). Ultimately, the local change
in CDF variance is given by
1 −µ(θ∗ )2 /2y
Σ(θ ∗ )
e
V (θ ∗ |L) =
e
Σ(θ ∗ )
. (9)
2π2y
Σ(θ ∗ ) + 2y
Note that the CDF variance reduction criterion in (9) favors
points with low |µ(θ)| and high Σ(θ), just as in (7).
Algorithm 1 Sequential closed-loop stochastic verification
Input: initial training set L = {D, y}, available sample
locations U, max # of additional samples T
2: Initialize: train GP regression model
3: for i=1:T do
4:
Select θ = argmax Ve (θ|L)
1:
θ∈U
5:
6:
7:
8:
9:
Perform simulation at θ, obtain measurement y(θ)
Add {θ, y(θ)} to training set L, remove θ from U
Retrain GP model with updated L
end for
Return: expected pbsat (θ) and variance V θ|L
B. Sampling Algorithms
The section criterion from (9) forms the basis of the
closed-loop stochastic verification framework. Since the actual set Θ is infinite, the framework constructs a high
resolution lattice Θd to replicate Θ and then selects all
training locations from Θd rather than Θ directly (D ⊂ Θd ).
Set U = Θd \ D contains all remaining points not in training
set L available for future simulations. This paper presents
two versions of the closed-loop verification framework.
1) Sequential Sampling: The most straightforward implementation of closed-loop verification is sequential sampling,
described in Algorithm 1. The process starts with an initial
training set L of passively-selected samples, usually obtained
through random sampling or another open-loop design of
experiments approach. This initial training set of size |L| =
N0 produces the initial GP so active sampling can be
performed. The sequential procedure then selects one sample
from U according to (9) before it performs one simulation
at the selected θ vector, obtains y(θ), and updates the GP
prediction model. This iterative procedure repeats until the
number of remaining samples T = Ntotal − N0 has been
reached.
2) Batch Sampling: In comparison to Algorithm 1, batch
sampling [18] offers further computational efficiency by selecting multiple samples between retraining steps. Assuming
the same limit Ntotal , batch selection reduces the number
of GP retraining processes and can also fully exploit any
parallel computing capabilities of the simulation environment. Batch approaches select M samples at once and
perform their simulations in parallel. While this provides
computational savings, it also introduces the possibility of
redundant samples if the batch sample set S does not possess
adequate diversity.
First proposed for deterministic closed-loop verification in
[10], determinantal point processes (DPPs) [24] present efficient probabilistic methods for encouraging diversity within
S without severe computational overhead. The batch framework first converts Ve (θ|L) into a probability distribution
PV (θ) = Ve (θ|L)/ZV , where ZV is the normalization
constant. A relatively small number of MT samples are
obtained from Θd according to PV (θ) and used to construct
a DPP. This paper uses MT = 1000, but these are simply
Algorithm 2 Batch closed-loop stochastic verification framework using determinantal point processes
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
Input: initial training set L = {D, y}, available sample
locations U, # of iterations T , batch size M
Initialize: train GP regression model
for i=1:T do
Initialize: S = ∅
Transform Ve (θ|L) into distribution PV (θ)
Form k-DPP from MT random samples of PV (θ)
Generate M random samples from DPP, add to S
Run simulation ∀θ ∈ S, obtain measurements yS
Add {S, yS } to training set L, remove S from U
Retrain GP model with updated L
end for
Return: expected pbsat (θ) and variance V θ|L
samples of θ locations from the distribution and not actual
simulations are performed. The DPP randomly selects M
locations from U for S based upon a modified version of
PV that penalizes similarities in S and subsequently spreads
the datapoints out across high-valued regions in Θd with
significantly less redundancy. Algorithm 2 details the batch
closed-loop verification framework. Rather than T number
of additional simulations, Algorithm 2 operates in T batches
of M simulations, assuming M · T ≤ Ntotal − N0 .
V. E XAMPLES
This section examines statistical verification applied to
three stochastic systems and demonstrates the improved
performance of Algorithms 1 and 2 over existing techniques.
A. Example 1: Model Reference Adaptive Control System
The first example considers a stochastic version of the
concurrent learning model reference adaptive control (CLMRAC) system from earlier deterministic work [9], [10].
The CL-MRAC example examines a second order linear
system with two uncertain parameters θ = [θ1 , θ2 ]T . The
adaptive control system estimates these parameters online
and attempts to track a desirable reference trajectory. The
nonlinearities associated with adaptation result in a highly
nonlinear closed-loop system.
The verification goal is to predict the probability the actual
trajectory will remain within 1 unit of the reference trajectory
at all times between 0 ≤ t ≤ 40 seconds, given in STL
format as
ϕ = [0,40] (1 − |e1 [t]| ≥ 0)
(10)
where tracking error e1 (t) is the difference between actual
position x1 (t) and reference position xm1 (t). The STL
operator [0,40] refers to “for all times between t = 0 and
t = 40 seconds.” Measurement y(θ) is the STL robustness
degree ρϕ for the trajectory initialized with uncertainty
setting θ. The sampling lattice Θd covers θ1 : [−10, 10] and
θ2 : [−10, 10] with a grid of 40,401 points. In this example,
the open-loop dynamics are subject to additive process noise
with Gaussian distribution w(t) = N (0, I). The underlying
B. Example 2: Robust Multi-Agent Task Allocation
The second example addresses the robust multi-agent task
allocation problem from [25] with added stochasticity. The
task allocation problem attempts to assign a team of four
UAVs to complete fire surveillance tasks that will take
longer or shorter depending on the wind speed θ1 and
direction θ2 . Task durations are also corrupted by zero-mean
Gaussian noise. Unforeseen time delays will compound and
may potentially lead the UAVs to miss the completion of
tasks within their assigned window, thus lowering the overall
realized mission score. The verification goal is to determine
whether the team of UAV agents will sufficiently complete
the ordered tasking and achieve a minimum mission score
at different wind settings. Sampling grid Θd spans the set
Mean Absolute Error (%)
12
Open-Loop
PDF Mean
PDF Variance
CDF Variance
10
8
6
4
2
0
100
200
300
400
# Samples
Outperforms (%)
Fig. 1: (Example 1) Comparison of mean absolute error (MAE)
convergence for the four different sampling strategies. The standard
deviation intervals around the mean (solid lines) are given by the
0.5σ bound.
100
50
Open-Loop
PDF Mean
PDF Variance
0
100
200
300
400
# Samples
Fig. 2: (Example 1) Ratio of cases where Algorithm 2 directly
outperforms or matches the MAE of the indicated strategies.
24
Perecent Improvement (%)
true distribution was obtained by repeated sampling at each
of the points in Θd . To avoid any potential issues, Gaussian
distributions were fit to the raw data and the variance was
averaged across Θd to return y = 0.0372.
Figure 1 compares the performance of Algorithm 2 against
similar procedures using the existing selection metrics discussed in Section IV as well as an open-loop, random
sampling procedure. These procedures all start with an initial
training set of 50 simulations and select batches of M = 10
points until a total of 450 simulations has been reached.
Neither the kernel hyperparameters nor y are assumed to
be known so the procedures estimate these online using
maximum likelihood estimation. In order to fairly compare
each procedure, the algorithms all start from the same 100
randomly-chosen initial training sets with the same random
seed. At the conclusion of the process, Algorithm 2 demonstrates a 29%, 31%, and 35% improvement in average mean
absolute error (MAE) over the PDF mean, PDF variance, and
random sampling approaches. Additionally, the performance
of the algorithms in each of the 100 test cases can be
directly compared since they all start with the same L and
random seed. At this level, Figure 2 illustrates that Algorithm
2 will either eventually match or outperform the existing
sampling strategies nearly 100% of the time. Although the
exact numbers will change for different distributions, these
results highlight the value of the selection criteria from (9)
to further reduce prediction error given a fixed number of
samples.
Figure 3 demonstrates the utility of CDF variance to
identify regions of large prediction errors, regardless of the
particular sampling strategy. As discussed in Section III-B,
the true CDF variance is unavailable, but the approximate
CDF variance (7) still provides a meaningful metric to
compare prediction confidence across Θd . One of the best
uses of (7) is to rank points in Θd according to their CDF
variance in order to identify which predictions to trust the
least. Figure 3 displays the reduction in prediction error when
the points with the top 5% of CDF variance are removed. The
12-22% improvement in MAE proves that the CDF variance
did indeed correctly identify and remove regions with large
pesat (θ). When the top 10% is removed, the improvement
jumps to 40%.
22
20
Open-Loop
PDF Mean
PDF Variance
CDF Variance
18
16
14
12
10
100
200
300
400
# Samples
Fig. 3: (Example 1) Reduction in MAE after all points with the top
5% of CDF variance are removed.
of feasible wind conditions θ1 : [0◦ , 359◦ ], θ2 : [0, 40] km/hr
with 16,641 possible trajectory settings.
Figures 4 and 5 compare the performance of Algorithm 1
against the competing sampling strategies for 250 randomlyinitialized test cases. Ultimately, the MAE performance is
consistent with the last example. Algorithm 1 demonstrates
a 10-20% improvement in average MAE over the existing
strategies and either matches or exceeds the MAE of the
competing approaches when directly compared against one
another.
requirement expects the autopilot to maintain altitude x(t)
within a set threshold of the initial altitude when the autopilot
is engaged. This work explores a relaxed version of the
original requirement [26] in which the aircraft must remain
within a 55 foot window around the initial altitude x(0),
Mean Absolute Error (%)
20
Open-Loop
PDF Mean
PDF Variance
CDF Variance
15
ϕ = [0,50] (55 − |x[t] − x[0]| ≥ 0).
10
5
20
30
40
50
60
70
80
90
# Samples
Outperforms (%)
Fig. 4: (Example 2) Comparison of MAE convergence for the four
different sampling strategies over 250 randomly selected initializations.
100
50
Open-Loop
PDF Mean
PDF Variance
0
20
30
40
50
60
70
80
90
# Samples
Perecent Improvement (%)
Fig. 5: (Example 2) Ratio of cases where Algorithm 1 directly
outperforms or matches the MAE of the indicated strategies.
18
16
14
12
Open-Loop
PDF Mean
PDF Variance
CDF Variance
10
8
6
20
30
40
50
60
70
80
90
# Samples
Fig. 6: (Example 2) Reduction in MAE after all points with the top
5% of CDF variance are removed.
Figure 6 demonstrates a similar reduction in MAE as
witnessed in Figure 3. Once the points with the top 5% of
CDF variance are removed, the MAE of the predictions in
the remaining 95% of the data reduces by up to 12-16%.
If the top 10% are removed, the average reduction reaches
at least 28% for all the sampling strategies. This supports
the conclusion that CDF variance identifies points with low
prediction confidence where pesat (θ) may be large.
C. Example 3: Lateral-Directional Autopilot
Lastly, the third example considers the altitude-hold requirement of a lateral-directional autopilot [26]. In this
problem, a Dryden wind turbulence model [27] augments
the original nonlinear aircraft dynamics model to introduce
stochasticity into the closed-loop dynamics. The performance
(11)
The performance measurement is the STL robustness degree
ρϕ . This example tests the satisfaction of the requirement
against initial Euler angles for roll θ1 : [−60◦ , 60◦ ], pitch
θ2 : [4◦ , 19◦ ], and yaw θ3 : [75◦ , 145◦ ] and longitudinal
inertia θ4 : [5430, 8430](kg · m2 ), with a constant reference
heading of 112◦ . The 4D grid Θd consists of 937,692
possible parameter settings for the simulations.
Figures 7 and 8 illustrate the MAE performance of the
four different sampling strategies given a batch size of
M = 10. The equivalent of Figures 3 and 6 is skipped but
displays the same exact behavior. This example begins with
100 simulations in the initial training set L and compares
the performance over 120 random initializations. As before,
Algorithm 2 outperforms the PDF variance-focused and random sampling procedures with a 27% reduction in average
MAE. Unlike the previous two examples, Algorithm 2 only
slightly edges out the competing mean-focused strategy. This
new result is due to the comparatively low y for noisy
y(θ) from the low-altitude turbulence model. The resulting
tighter distribution only experiences changes in psat (θ) near
ȳ(θ) = 0, which favors the mean-focused selection criteria.
If the turbulence is increased then the distribution widens
dramatically and the performance of the PDF mean-focused
metric degrades significantly like the previous two examples.
These case studies all highlight the improved performance of
the novel CDF variance algorithms compared to the existing
sampling strategies. Just as important, Algorithms 1 and 2
will remain the best-performing sampling strategy even as the
distribution changes, whereas the performance of the other
strategies will vary according to changes in y .
VI. C ONCLUSION
This work introduced machine learning methods for
simulation-based statistical verification of stochastic nonlinear systems. In particular, the paper presented a GP-based
statistical verification framework to efficiently predict the
probability of requirement satisfaction over the full space
of possible parametric uncertainties given a limited amount
of simulation data. Additionally, Section III-B developed
new criteria based upon the variance of the cumulative
distribution function to qualify confidence in the predictions.
This metric provides a simple validation tool for online
identification of regions where the prediction confidence
is low. Section IV builds upon this metric and introduces
sequential and batch sampling algorithms for efficient closedloop verification. These new verification procedures demonstrate up to a 35% improvement in prediction error over
competing approaches in the three examples in Section V.
The examples also serve to highlight the utility of the CDF
Mean Absolute Error (%)
14
Open-Loop
PDF Mean
PDF Variance
CDF Variance
12
10
8
6
4
100
200
300
400
500
# Samples
Fig. 7: (Example 3) Comparison of MAE convergence for the four
different sampling strategies over 120 random initializations.
Outperforms (%)
100
50
Open-Loop
PDF Mean
PDF Variance
0
100
200
300
400
500
# Samples
Fig. 8: (Example 3) Ratio of cases where Algorithm 2 directly
outperforms or matches the MAE of the indicated strategies.
variance to correctly identify low-confidence regions in the
parameter space without external validation datasets.
While the paper only considers Gaussian distributions for
the performance measurements, this work lays the foundation
for more complex statistical verification frameworks capable
of handing a wider range of distributions. Upcoming work
will adapt the CDF variance metric and closed-loop verification procedures to recent developments in modeling of
spatially-varying standard deviations [12] and non-Gaussian
distributions [13]. Additionally, this work can be extended to
high-dimensional GP representations [28]. While the implementation considers Gaussian distributions, the fundamental
concepts have broader utility.
ACKNOWLEDGMENTS
This work is supported by the Office of Naval Research
and Air Force Office of Scientific Research under grants
ONR MURI N00014-11-1-0688 and AFOSR FA9550-15-10146 .
R EFERENCES
[1] E. M. Clarke, O. Grumberg, and D. A. Peled, Model Checking. MIT
Press, 1999.
[2] S. Prajna, “Barrier certificates for nonlinear model validation,” Automatica, vol. 42, no. 1, pp. 117–126, 2006.
[3] U. Topcu, “Quantitative local analysis of nonlinear systems,” Ph.D.
dissertation, University of California, Berkeley, 2008.
[4] A. Majumdar, A. A. Ahmadi, and R. Tedrake, “Control and verification
of high-dimensional systems with DSOS and SDSOS programming,”
in IEEE Conference on Decision and Control, 2014.
[5] J. Kapinski, J. Deshmukh, S. Sankaranarayanan, and N. Arechiga,
“Simulation-guided Lyapunov analysis for hybrid dynamical systems,”
in Hybrid Systems: Computation and Control, 2014.
[6] P. Zuliani, C. Baier, and E. M. Clarke, “Rare-event verification for
stochastic hybrid systems,” in Hybrid Systems: Computation and
Control, 2012.
[7] E. M. Clarke and P. Zuliani, “Statistical model checking for cyberphysical systems,” in International Symposium for Automated Technology for Verification and Analysis, 2011.
[8] A. Kozarev, J. F. Quindlen, J. How, and U. Topcu, “Case studies in
data-driven verification of dynamical systems,” in Hybrid Systems:
Computation and Control, 2016.
[9] J. F. Quindlen, U. Topcu, G. Chowdhary, and J. P. How, “Active
sampling-based binary verification of dynamical systems,” in AIAA
Guidance, Navigation, and Control Conference, 2018. [Online].
Available: http://arxiv.org/abs/1706.04268
[10] ——. (2017) Active sampling for closed-loop statistical verification
of uncertain nonlinear systems. Preprint. [Online]. Available:
https://arxiv.org/abs/1705.01471
[11] O. Maler and D. Nickovic, Monitoring Temporal Properties of Continuous Signals. Springer Berlin Heidelberg, 2004, pp. 152–166.
[12] M. Lázaro-Gredilla and M. Titsias, “Variational heteroscedastic gaussian process regression,” in The 28th International Conference on
Machine Learning, Bellevue, Washington, July 2011.
[13] D. Seiferth, G. Chowdhary, M. Muhlegg, and F. Holzapfel, “Online
gaussian process regression with non-gaussian likelihood,” in American Control Conference, 2017.
[14] C. M. Bishop, Pattern Recognition and Machine Learning (Information Science and Statistics), 1st ed. Springer, 2007.
[15] M. E. Tipping, “Sparse bayesian learning and the relevance vector
machine,” Journal of Machine Learning Research, vol. 1, pp. 211–
244, June 2001.
[16] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for
Machine Learning, T. Dietterich, Ed. MIT Press, 2006.
[17] A. H.-S. Ang and W. H. Tang, Probability Concepts in Engineering,
2nd ed. Wiley, 2007.
[18] B. Settles, Active Learning, R. J. Brachman, W. W. Cohen, and T. G.
Dietterich, Eds. Morgan and Claypool, 2012.
[19] J. Kremer, K. S. Pedersen, and C. Igel, “Active learning with support
vector machines,” Data Mining and Knowledge Discovery, vol. 4,
no. 4, pp. 313–326, July 2014.
[20] G. Chen, Z. Sabato, and Z. Kong, “Active learning based requirement
mining for cyber-physical systems,” in IEEE Conference on Decision
and Control, 2016.
[21] T. Desautels, A. Krause, and J. W. Burdick, “Parallelizing explorationexploitation tradeoffs in gaussian process bandit optimization,” Journal
of Machine Learning Research, vol. 15, pp. 4053–4103, 2014.
[22] A. Gotovos, N. Casati, and G. Hitz, “Active learning for level set estimation,” in International Joint Conference on Artificial Intelligence,
2013, pp. 1344–1350.
[23] Y. Zhang, T. N. Hoang, K. H. Low, and M. Kankanhalli, “Nearoptimal active learning of multi-output gaussian processes,” in AAAI
Conference on Artificial Intelligence, 2016.
[24] A. Kulesza and B. Taskar, “k-DPPs: Fixed-size determinantal point
processes,” in International Conference on Machine Learning, 2011.
[25] J. F. Quindlen and J. P. How, “Machine Learning for Efficient
Sampling-based Algorithms in Robust Multi-Agent Planning under
Uncertainty,” in AIAA SciTech Conference, 2017.
[26] C. Elliott, G. Tallant, and P. Stanfill, “An example set of cyber-physical
V&V challenges for S5,” in Air Force Research Laboratory Safe and
Secure Systems and Software Symposium (S5) Conference, Dayton,
OH, July 2016.
[27] Department of Defense MIL-HDBK-1797, “Flying qualities of piloted
aircraft.”
[28] K. Kandasamy, J. Schneider, and B. Poczos, “High dimensional
bayesian optimization and bandits via additive models,” in International Conference on Machine Learning, 2015.
| 3cs.SY
|
Group Scissor: Scaling Neuromorphic Computing Design
to Large Neural Networks
Yandan Wang
Wei Wen
Beiye Liu
University of Pittsburgh
University of Pittsburgh
University of Pittsburgh
University of Pittsburgh
University of Pittsburgh
[email protected]
[email protected]
ABSTRACT
Synapse crossbar is an elementary structure in Neuromorphic Computing Systems (NCS). However, the limited size
of crossbars and heavy routing congestion impedes the NCS
implementations of big neural networks. In this paper, we
propose a two-step framework (namely, group scissor ) to
scale NCS designs to big neural networks. The first step
is rank clipping, which integrates low-rank approximation
into the training to reduce total crossbar area. The second
step is group connection deletion, which structurally prunes
connections to reduce routing congestion between crossbars.
Tested on convolutional neural networks of LeNet on MNIST
database and ConvNet on CIFAR-10 database, our experiments show significant reduction of crossbar area and routing area in NCS designs. Without accuracy loss, rank clipping reduces total crossbar area to 13.62% and 51.81% in the
NCS designs of LeNet and ConvNet, respectively. Following
rank clipping, group connection deletion further reduces the
routing area of LeNet and ConvNet to 8.1% and 52.06%,
respectively.
1.
INTRODUCTION
input
filter 1
…
filter n
…
…
…
…
…
The record-breaking classification performance of deep neural networks (DNNs) [1] in recent years has stimulated the
fast-growing research on hardware design of neuromorphic
computing systems (NCS) [2][3][4][5][6][7]. NCS utilize devices and circuit components to mimic the behaviors of neural networks to perform intelligent tasks, such as image classification, speech recognition and natural language processing. Circuit-level and architecture-level NCS designs using
emerging memristor devices [8] and traditional CMOS technologies [3] are being explored.
In software applications, the depth of DNNs rapidly grows
from several layers to hundreds or even thousands of layers [9]. However, the scale of hardware design of NCS falls
far behind. A major critical issue that obstructs the scalingup of NCS to big neural networks is the limited synaptic
connection (e.g., crossbar) in hardware implementation. Accordingly, it results in heavy wire congestion (e.g., the routing between crossbars). Taking the memristor-based NCS as
an example, under the impact of IR-drop and process variations, both reading and writing reliability will be severely
degraded when the size of a memristor-based crossbar is beyond 64 × 64 [10][11]. The similar scenario can also be observed in CMOS-based conventional designs. For example,
the IBM TrueNorth chip, as a pioneer in NCS design, limits
the size of neurosynaptic crossbars to 256 × 256 [3]. It is
inevitable to interconnect multiple crossbars to implement
modern big neural networks. The increasing scale of neural networks could quickly exhaust the resources of synapse
crossbars and deteriorate the wire congestion [12][13].
To solve those problems, [13] mapped logically-connected
cores to physically-adjacent cores to reduce spike communications. However, it only optimized the placement of cores
and cannot reduce the core consumption. The existing NCS
optimization based on traditional sparse neural networks can
alleviate the wire congestion [12]. However, they usually
separate the software sparsifying and hardware deployment,
which makes the optimization more challenging.
Unlike previous work, we propose a tow-step framework
– group scissor – to overcome above issues so as to scale
NCS designs to big neural networks. The first step is rank
clipping, which integrates low-rank approximation into the
training process of neural networks. It targets at reducing
the dimensions of connection arrays in a group-wise way and
therefore reducing the consumption of synapse crossbars in
NCS. The second step – group connection deletion – structurally deletes/prunes groups of connections. The approach
tends to learn hardware-friendly sparse neural networks to
directly delete the routing wires between crossbars, with
controllable low hardware cost. Unlike [12] which evaluated
NCS by Hopfield networks using less challenging database,
we evaluate our group scissor by state-of-the-art convolutional neural networks – LeNet and ConvNet – using MNIST
and CIFAR-10 database. Our experiments show, without
accuracy loss, rank clipping respectively reduces total crossbar area to 13.62% and 51.81% in LeNet and ConvNet, and
group connection deletion reduces the routing area to 8.1%
and 52.06%, respectively.
input
arXiv:1702.03443v2 [cs.NE] 17 Jun 2017
[email protected]
[email protected]
[email protected]
Donald Chiarulli
Hai (Helen) Li
…
…
convolutional layer
(a)
output
(b)
Figure 1: The NCS designs for (a) a small convolutional layer, and (b) a big layer.
PRELIMINARY
Algorithm 1: Principal Components Analysis (PCA)
3.
THE GROUP SCISSOR FRAMEWORK
In this work, we propose the Group Scissor framework
to improve the scalability of neuromorphic computing design. The framework contains two steps: rank clipping for
crossbar area occupation reductions and group connection
deletion for routing congestion reduction. The details of the
proposed design will be described in this section. Moreover,
the estimations of circuit area and routing wires for MBCbased neuromorphic design are formulated.
3.1
Rank Clipping
As discussed above, the high crossbar area occupation and
heavy routing congestion are the major obstacles in realizing big neural networks. In order to overcome these issues,
we propose to utilize low-rank approximation (LRA) to reduce the dimensions of weight (connection) matrices in big
neural networks. Low-rank approximation is a mathematical technology, which uses the product of smaller matrices with reduced rank to approximate a given large matrix.
Specifically, an original weight matrix W ∈ RN ×M can be
approximated as
f
W ≈ U · VT = W,
N ×K
T
(1)
K×M
where U ∈ R
,V ∈ R
, and K is the rank of the
approximation. When K << M , U and V are reduced to
skinny matrices. The total crossbar area occupation can be
reduced when the rank K satisfies
K<
NM
.
N +M
(2)
There are various LRA techniques. Without losing generality, commonly used principal components analysis (PCA)
[15] and singular value decomposition (SVD) [11] are adopted
as the representatives in this work.
The PCA approach is formulated in Algorithm 1. The
essence of PCA is a linear projection from a high dimensional
space (wn ∈ RM ) to a lower dimensional subspace (un ∈
RK , K M ) to minimize the reconstruction error of W,
where wn and un is the n-th row of W and U, respectively,
and V is the basis of the subspace. The reconstruction error
is
PM
f 2
||W − W||
m=K+1 λm
= P
,
(3)
eK =
M
||W||2
m=1 λm
where ||·|| is the Euclidean norm, namely Euclidean distance.
1
2
Input : N × M matrix W, and rank K
1 PN
Get mean of rows wn ∀n ∈ [1...N ]: µ = N
n=1 wn ;
Centralize the data: replace each wn with wn − µ;
T
3
4
5
W
;
Calculate the M × M covariance matrix: C = W
N −1
Calculate the eigenvectors vm and eigenvalues λm of
covariance matrix C: Cvm = λm vm , ∀m ∈ [1...M ];
Project to subspace: N × K matrix U = WV, where
V = [v1 , ...vK ] is a M × K matrix and v1...K are
eigenvectors corresponding to the largest K eigenvalues;
f =U·VT
Output: N × M approximation matrix W
Although LRA can approximately reconstruct the original weights, small perturbation of weights can deteriorate
the classification accuracy. Table 1 compares the performance of the original baseline design (“Original ”) and the
low-rank neural networks which are directly decomposed by
PCA (“Direct LRA”). The accuracy drops rapidly after applying “Direct LRA”. Fine-tuning (retraining) the low-rank
neural networks can recover accuracy, but the optimal ranks
in all layers are unknown. More importantly, it is very timeconsuming to explore the entire design space by decomposing and retraining a wide variety of neural networks. We
propose the LRA-based rank clipping, which can not only
successfully retain the accuracy but also can automatically
converge to the optimal low ranks in all layers. Low ranks in
Table 1 are actually obtained by our rank clipping method.
The key idea of rank clipping is illustrated in Figure 2.
Rather than direct LRA after training, we integrate LRA
into the training process and carefully clip some ranks with
small reconstruction errors after a fixed number of training iterations, say, S iterations. The gentle clipping induces
small reconstruction errors and thus slightly affect the classification accuracy. As such, the accuracy could be recovered
by the following S iterations. The iteration of clipping and
training not only avoids irremediable accuracy degradation
but also enables neural networks to gradually converge to
the optimal ranks for all layers.
Algorithm 2 describes the detailed operation of the rank
clipping. The tolerable clipping error ε is the allowed maximum reconstruction error after each rank clipping. A gentle
clipping can be enabled by setting a small ε, e.g., 0.01. Our
rank clipping starts with a full-rank LRA without reconstruction errors, and iteratively examines if the low-dimensional
U can be further projected to a lower -rank subspace with
only reconstruction error of ε. Note that PCA is used as
the representative of LRA in Algorithm 2. However, other
LRA methods like SVD can also be used. The only modification is to replace the approximation of weight matrix by
③
≈
W
② ①
③
×
✂
Figure 1(a) illustrates an implementation of a convolutional layer in neural networks using memristor-based crossbars (MBC), where memristors (a.k.a. synapses) in each column encode the weights of one filter [14]. The implementation of a fully-connected layer utilizes the similar structure,
but each column realizes the connections to one output neuron. As aforementioned, the sizes of crossbars are limited.
So when implementing big neural networks, a high volume
of the interconnection of crossbars are required.
Figure 1(b) depicts a circuit-level implementation of a big
layer by tiling and interconnecting MBC [12]. As the scale
of modern neural networks grows, the high crossbar area
occupation and heavy routing congestion become critical issues and seriously obstruct the scalability of the hardware
implementation.
✂
2.
U
②
①
VT
① = Clipped ② = Clipping ③ = Remained
Figure 2: Rank clipping for crossbar area occupation
reduction.
Table 1: Accuracy and ranks
Database
Net
Method
Accuracy
‡
conv1†
conv2
conv3
fc1†
fc2
MNIST
LeNet [16]
Original
Direct LRA
Rank clipping
99.15%
96.44%
99.14%
Rank K
20
5
5
50
12
12
–
–
–
500
36
36
10
10
10
CIFAR-10
ConvNet [1]
Original
Direct LRA
Rank clipping
82.01%
43.29%
82.09%
Rank K
32
12
12
32
19
19
64
22
22
10
10
10
–
–
–
†
‡
conv1 is the first convolutional layer, fc1 is the first fully-connected layer, and so forth
corresponding rank indicates the number of filters in convolutional layers or indicates the number of output
neurons in fully-connected layers.
Algorithm 2: Rank Clipping
7
8
9
10
11
12
13
14
the other LRA method.
Figure 3 plots the trends of rank reduction and accuracy
retention during the rank clipping of LeNet in Table 1 using
PCA. Rank clipping is examined every S = 500 (denoted as
5e2 in x-axis title) iterations with ε = 0.03. In the figure,
the rank ratio is defined as the remained rank over full rank,
i.e., K/M . The figure demonstrates that ranks are rapidly
clipped at the beginning of iterations and converge to optimal low ranks. During the entire process, the accuracy
changes are limited to small fluctuations.
As shown in Figure 3 and Table 1, the proposed rank
1
1
conv1
conv2
fc1
Accuracy
0.6
0.4
0.95
0.9
Accuracy
Rank ratio
0.8
0.85
0.2
0
0.8
0
10
20
30
40
50
60
Training iteration (5e2)
Figure 3: Rank ratio of each layer and accuracy during training with rank clipping.
input
N inputs
P
…
6
×
row
group
…
4
5
column
group
…
3
…
…
1
2
Input : Trained original neural network, tolerable clipping
error ε, maximum training iteration I, clipping
step S
for each layer l do
PCA of weight matrix Wl = Ul · VT
l with full rank
Kl = Ml ;
end
while i = 1; i < I; i = i + S do
for each layer l do
T
PCA of Ul = Ûl · V̂l using the minimum rank K̂
which satisfies eK̂ ≤ ε;
if K̂ < Kl then
T
T
Kl = K̂; Ul = Ûl ; VT
l = V̂l · Vl
else
continue;
end
end
Train the neural network for S iterations;
end
Output: Clipped low-rank neural network with
approximation Wl = Ul · VT
l for each layer l
Q
×
…
…
K outputs
output
Figure 4: The group connection deletion.
clipping successfully reduces the ranks in both convolutional
layers and fully-connected layers without accuracy loss. The
crossbar area occupation of the entire LeNet (ConvNet) reduces to 13.62% (51.81%). Instead of PCA, when SVD is applied, the whole crossbar area can also be reduced to 32.97%
(55.64%) for LeNet (ConvNet), which indicates SVD is inferior to PCA. Therefore, we mainly conduct experiments
using PCA approach. Note that the last layers of LeNet
and ConvNet are not clipped because the rank (M = 10) is
already very small and there is little improvement space.
3.2
Group Connection Deletion
The rank clipping can reduce the total number of required
crossbars, but a large number of crossbars will be still necessary to implement modern big neural networks. The second step of our group scissor framework – group connection
deletion aims at removing interconnections between synapse
crossbars so as to reduce the circuit-level routing congestion
and architecture-level inter-core communication for NCS.
Figure 4 gives the basic idea of group connection deletion.
An array of MBC are interconnected to implement a large
weight matrix U ∈ RN ×K . Suppose the elementary synapse
crossbar
has P inputs and Q outputs (P N, Q K), a
N lK m
× Q array of crossbars must be interconnected to imP
plement U as illustrated in Figure 4. The implementation of
another matrix V shall follow the similar method. As memristors can be densely manufactured in the crossbar and the
area of each memristor cell is feature-size level, the routing
wires dominate the circuit area [12]. Suppose a row group
of connections in Figure 4 all have zero weights, implying
that those connections are removable, we can delete/prune
g=1
g=1
where W is the set of weights in the whole neural network,
ED (W) is the original minimization function when training traditional neural networks. G(r) and G(c) respectively
denote the number of row groups and column groups, and
(r)
(c)
Wg and Wg are the sets of weights in the g-th row group
and column group, respectively. And
(r)
G
[
g=1
Wg(r) =
(c)
G
[
Wg(c) = W.
(5)
g=1
λ is the hyper-parameter to control the trade-off between
classification accuracy and routing congestion reduction. A
larger λ can result in lower accuracy but larger reduction of
routing wires. During the back-propagation training with
Eq. (4), each weight w will be updated as
!
∂ED (W)
λw
λw
+
, (6)
+
w ←w−η
(r)
(c)
∂w
||Wi ||
||Wj ||
where η is the learning rate, i ∈ [1...G(r) ], j ∈ [1...G(c) ],
(c)
(r)
w ∈ Wi and w ∈ Wj .
With group connection deletion, we disconnect all the
zero-weighted connections and prune all the routing wires
connecting to all-zero row groups or column groups. After
deletion, we fine-tune (retrain) the structurally-sparse neural networks to improve accuracy.
Figure 5 plots the trends of deleted routing wires (i.e.,
all-zero row/column groups) and the classification accuracy
versus the iterations of group connection deletion. The deletion process starts with the low-rank LeNet in Table 1 that
was already compressed by rank clipping. In Figure 5, we
only delete the matrices of U and V whose dimensions are
beyond the largest size of MBC. More design details shall be
presented in Section 3.3 and Section 4. Even for low-rank
neural networks, our method can delete the routing wires
dramatically, e.g., 93.9% interconnection wires are removed
in the crossbar array of fc1 v. Fine-tuning the deleted neural
networks attains the baseline accuracy (99.1%),
Note that compared with our method, it is more challenging to use traditional sparse neural networks to reduce
the routing wires. This is because its sparse weights are
100
1
80
0.95
60
0.9
40
conv2_u
fc1_u
fc1_v
fc2_u
Accuracy
0.85
20
0
Accuracy
% deleted routing wires
the wire routing to the input of this row group. Similarly,
the wire routing from the output of a column group can be
deleted when the column group of connections are all-zeros.
Our group connection deletion method actively deletes those
groups of connections during the learning of neural networks, meanwhile maintaining the classification accuracy at
the similar level.
We harness group Lasso regularization to delete groups
of connections. Group Lasso is an efficient regularization
in the study of structured sparsity learning [17][18]. With
group Lasso regularization on each group of weights, a high
percentage of groups can be regularized to all-zeros. In
our group connection deletion method, weights are split to
row groups and column groups as illustrated in the figure.
And group Lasso regularization is enforced on each group.
Mathematically, the minimization function for training neural network with group Lasso can be formulated as:
(r)
(c)
G
G
X
X
(r)
(c)
E(W) = ED (W) + λ ·
||Wg || +
||Wg || , (4)
1
0
11
10
21
20
31
30
41
40
51
50
61
0.8
60
Training iterations (5e2)
Figure 5: The percentage of deleted routing wires
and accuracy during group connection deletion.
fc1 u and fc1 v is the low-rank matrix U and V of
fc1 after rank clipping, and so forth.
randomly distributed in the crossbar arrays and the corresponding routing wire must be preserved as long as there
is one nonzero weight existing in the row group or column
group.
3.3
Area Estimation
In this section, we formulate the area estimation method
adopted for hardware evaluation in this work.
MBC area estimation: The use of MBC in NCS design has been extensively studied. As a critical component
in such a system, MBCs occupy a significant proportion of
whole design area. Each MBC is an ultra high density crosspoint structure formed by a set of memristors and wires. The
area of a memristor cell in MBC is 4F 2 under the state-ofthe-art technology [11], where F is the minimum feature size.
Restricted by the technology limitations, a feasible MBC implementation only considers MBCs that are not larger than
64×64 [10]. To ensure the system reliability and robustness,
we only consider MBCs with dimensions constrained within
64×64 in the standard library. For those large weight matrices in neural networks, their connections can be distributed
into many MBCs in the library as demonstrated in Figure 1.
Routing area estimation: Assume that the metal width
is Wm , the distance between two metals is Wd , and the
length of i-th wire between crossbars is Li . The total routing
area occupied by the wires can be roughly formulated as
Ar = (Wm + Wd )
Nw
X
Li .
(7)
i
Here Nw is total wire count including electrostatic shielding
wires. Suppose the average wire length is linearly proportional to Nw , the routing area is estimated as
Ar = αNw2 ,
(8)
where α is a scalar.
4.
EXPERIMENT
In this section, experiments are conducted to evaluate the
effectiveness of proposed rank clipping and group connection
deletion methods. All the experiments conducted in this
paper are based on the NCS implemented by MBC. The
related experiment parameters on memristor and MBC are
summarized in Table 2. We mainly implement two neural
networks – LeN et on MNIST and ConvN et on CIFAR-10.
The detailed network structures are summarized in Table 1.
4.1
MBC Area Reduction
100%
Parameter
value
memristor cell area
4F 2
maximum crossbar size
64×64
Wire length between two memristors
2F
conv1
conv2
fc1
total
(a)
Experiment Parameters
80%
Crossbar area
Table 2:
60%
conv1
conv2
conv3
total
(b)
40%
20%
0%
0.8%
1.4%
2.0%
2.6%
17.5%
18.5%
Error
In our experiments, we clip all the convolutional and fullyconnected layers, except the last classifier layer. The original
rank in the last layer is determined by the number of classes
so the further reduction is meaningless. The rank clipping
method compresses each large weight matrix to two skinny
matrices by reducing the rank. Figure 6 shows the final
remained ranks with respect to the accuracy and tolerable
clipping error ε for convolutional layers in LeNet. Here the
original rank of conv1 and conv2 is 20 and 50, respectively, as
denoted by upper markers on the stems. For each layer, the
rank decreases as increases, and finally reaches to a very
small value. It can be seen that the corresponding accuracy
is well maintained. We also observe similar results in fc1.
More specifically, the layer-wise ranks are reduced to 5, 12
and 36 without accuracy loss, and to 4, 6 and 6 with merely
1% loss.
Figure 7(a) and (b) respectively plot the percentage of
remained MBC area with respect to the classification error
for LeN et and ConvN et. Routing area is excluded in this
evaluation. The area of each layer is the sum of the areas of
U and V. Total area includes the area of the last classifier
layer, i.e., fc2 in Lenet or fc1 in ConvNet. For both networks,
the layer-wise areas of both convolutional layers and fullyconnected layers rapidly reduce with small accuracy loss.
In summary, the rank clipping can reduce the total crossbar area of LeNet to 13.62% without sacrificing any accuracy
loss. The crossbar area can be further reduced to 3.78% with
merely 1% accuracy loss. For more challenging ConvN et,
no accuracy loss is observed when the crossbar area is reduced to 51.81%. And under an accuracy loss of 1%, the
total crossbar area can be reduced to 38.14%.
Routing Area Reduction
To evaluate the routing congestion alleviated by group
connection deletion method, we use the number of routing
wires and remained routing area of Eq. (8) as our metrics.
Although the estimation of routing area in the real circuit
can be more complex, the real routing area reduction in the
conv1
conv2
Table 3: MBC Sizes and remained routing wires in
big layers.
Net
type
conv1 u
conv2 u
conv3 u
fc1 u
fc1 v
fc last
LeNet
sizes
% wires
–†
–
50 × 12
47.5
–
–
50 × 36
24.8
36 × 50
6.7
50 × 10
18.0
ConvNet
sizes
% wires
25 × 12
83.3
50 × 19
40.5
50 × 22
74.4
–
–
–
–
64 × 10
81.9
†
The weight matrix can be implemented by one crossbar.
conv1 v, conv2 v and conv3 v are omitted for the same reason.
hardware must be positively correlated to our results.
As aforementioned in Section 3.3, our standard library
contains all types of memristor crossbars with dimensions
constrained within 64 × 64. When implementing a N × K
weight matrix U, the MBC sizes are selected based on the
following criteria: (1) Implement U in a N × K MBC, when
N ≤ 64 and K ≤ 64; (2) Implement U by an array of MBCs
when N > 64 or K > 64, with the largest available MBC size
P × Q, where N and K is divisible by P and Q, respectively.
In the experiments, the group connection deletion starts
with the rank-clipped LeNet or ConvNet without accuracy
loss as presented in Table 1. Based on the MBC selection
criteria, the sizes of MBC utilized in big layers are shown
in Table 3. Matrices with sizes constrained by 64 × 64 are
omitted in the table, and no group Lasso regularization is
enforced on those small matrices.
The experimental results of the remained routing wires after applying the group connection deletion without allowing
accuracy loss are also presented in Table 3. The results for
LeNet are remarkable. We can achieve the same accuracy
of the baseline, with routing wires being only 47.5%, 24.8%,
6.7% and 18.0% of the original ones in respective layer. This
can reduce the layer-wise routing area to 8.1%, on average.
Table 3 also shows that, in ConvNet, our method on average reduces layer-wise routing wires to 70.03% and thus reduce layer-wise routing areas to 52.06%, meanwhile achievconv1
50
Rank
30
20
10
0
1
0.2
0.15
0.95
0.1
0.05
Accuracy
0.9
0
Figure 6: The remained ranks in convolutional layers of Lenet. fc1 is omitted for better visualization
as its original rank 500 is out of chart.
Remained routing wires
100%
40
20.5%
Figure 7: The MBC area for (a) Lenet and (b) ConvNet, after applying the rank clipping.
conv2
conv3
fc1
conv1
100%
(a)
75%
50%
25%
0%
0.175
conv2
conv3
fc1
(b)
Routing area
4.2
19.5%
Error
75%
50%
25%
0%
0.18
0.185
0.19
0.195
Classification error
0.2
0.175
0.18
0.185
0.19
0.195
0.2
Classification error
Figure 8: The (a) routing wire and (b) routing area
w.r.t. the classification error in ConvNet.
Figure 9: Weight matrices (transposed) after group connection deletion. The deletion starts from the rankclipped ConvNet in Table 1. Matrices are plotted in scale in the order of conv1 u, conv2 u, conv3 u and fc1.
White regions have no connections. And connections in each blue/red block are implemented in a crossbar.
ing the same accuracy as the baseline. With an acceptable accuracy loss, the routing congestion can also be significantly alleviated. Figure 8 comprehensively studies the
remained routing wires and routing area under different classification errors. With merely 1.5% accuracy loss, the routing area in each layer is reduced to 56.25%, 7.64%, 21.44%
and 31.64%, respectively.
At last, Figure 9 shows the sparse weight matrices after
group connection deletion for ConvNet in Table 3 without
accuracy loss. Each blue/red block stands for a collection of
weights, which are implemented by one crossbar in the NCS
design. White regions indicate that there are no connections.
After applying the group connection deletion, the connections in crossbars become sparse. More importantly, the
sparsity is structural instead of being randomly distributed
in traditional sparse neural networks. In the figure, a high
ratio of column groups in crossbars are regularized to allzeros, such that interconnection wires routing from those
crossbar columns can be removed. Impressively, as conv2 u
and fc1 in the figure show, some blocks have no connections
in the whole region, indicating that the entire crossbar can
be removed in the NCS implementation. It is significant because not only routing congestion can be alleviated, but also
crossbar area can be reduced. We also note that a crossbar
with some zero columns/rows can be replaced by a smaller
but dense crossbar after removing those zero groups, which
can further reduce the crossbar area.
5.
CONCLUSIONS
In this work, we propose a framework named group scissor to alleviate the impact of hardware limitations on the
NCS implementation of big neural networks. Specifically,
rank clipping and group connection deletion methods are
proposed to reduce area consumption of synapse crossbars
and routing area between crossbars, respectively. Final experiments show that our methods can reduce crossbar area
(routing area) to 13.62% (8.1%) with no accuracy loss for
LeNet. Moreover, no accuracy loss is observed for more challenging ConvNet when crossbar area is reduced to 51.81%
and routing area is reduced to 52.06%. The proposed framework can significantly save hardware area and improve system scalability.
6.
REFERENCES
[1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet
classification with deep convolutional neural networks,” in
NIPS, pp. 1097–1105, 2012.
[2] S. H. Jo, T. Chang, I. Ebong, B. B. Bhadviya,
P. Mazumder, and W. Lu, “Nanoscale memristor device as
synapse in neuromorphic systems,” Nano letters, vol. 10,
no. 4, pp. 1297–1301, 2010.
[3] S. K. Esser, A. Andreopoulos, R. Appuswamy, P. Datta,
D. Barch, A. Amir, J. Arthur, A. Cassidy, M. Flickner,
P. Merolla, et al., “Cognitive computing systems:
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Algorithms and applications for networks of neurosynaptic
cores,” in IJCNN, pp. 1–10, 2013.
C. Xu, X. Dong, N. P. Jouppi, and Y. Xie, “Design
implications of memristor-based rram cross-point
structures,” in DATE, pp. 1–6, 2011.
M. Hu, H. Li, Y. Chen, Q. Wu, G. S. Rose, and R. W.
Linderman, “Memristor crossbar-based neuromorphic
computing system: A case study,” IEEE transactions on
neural networks and learning systems, vol. 25, no. 10,
pp. 1864–1878, 2014.
B. Li, Y. Wang, Y. Wang, Y. Chen, and H. Yang, “Training
itself: Mixed-signal training acceleration for
memristor-based neural network,” in ASP-DAC,
pp. 361–366, 2014.
W. Wen, C. Wu, Y. Wang, K. Nixon, Q. Wu, M. Barnell,
H. Li, and Y. Chen, “A new learning method for inference
accuracy, core occupation, and performance
co-optimization on truenorth chip,” in DAC, pp. 1–6, 2016.
D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S.
Williams, “The missing memristor found,” nature, vol. 453,
no. 7191, pp. 80–83, 2008.
K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual
learning for image recognition,” arXiv:1512.03385, 2015.
J. Liang and H.-S. P. Wong, “Cross-point memory array
without cell selectorsâĂŤdevice characteristics and data
storage pattern dependencies,” IEEE Transactions on
Electron Devices, vol. 57, no. 10, pp. 2531–2538, 2010.
B. Liu, H. Li, Y. Chen, X. Li, T. Huang, Q. Wu, and
M. Barnell, “Reduction and ir-drop compensations
techniques for reliable neuromorphic computing systems,”
in ICCAD, pp. 63–70, 2014.
W. Wen, C.-R. Wu, X. Hu, B. Liu, T.-Y. Ho, X. Li, and
Y. Chen, “An eda framework for large scale hybrid
neuromorphic computing systems,” in DAC, p. 12, 2015.
F. Akopyan, J. Sawada, A. Cassidy, R. Alvarez-Icaza,
J. Arthur, P. Merolla, N. Imam, Y. Nakamura, P. Datta,
G.-J. Nam, et al., “Truenorth: Design and tool flow of a 65
mw 1 million neuron programmable neurosynaptic chip,”
IEEE Transactions on Computer-Aided Design of
Integrated Circuits and Systems, vol. 34, no. 10,
pp. 1537–1557, 2015.
L. Song, X. Qian, H. Li, and Y. Chen, “PipeLayer: A
pipelined ReRAM-based accelerator for deep learning,”
HPCA, 2017.
I. Jolliffe, Principal component analysis. Wiley Online
Library, 2002.
Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner,
“Gradient-based learning applied to document recognition,”
Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324,
1998.
M. Yuan and Y. Lin, “Model selection and estimation in
regression with grouped variables,” Journal of the Royal
Statistical Society., vol. 68, no. 1, pp. 49–67, 2006.
W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li, “Learning
structured sparsity in deep neural networks,” in NIPS,
pp. 2074–2082, 2016.
| 9cs.NE
|
Partitioned Multiprocessor Fixed-Priority Scheduling
of Sporadic Real-Time Tasks
Jian-Jia Chen
of Informatics, TU Dortmund University, Germany
E-mail: [email protected]
arXiv:1505.04693v3 [cs.DS] 22 Jun 2016
1 Department
Abstract—Partitioned multiprocessor scheduling has been
widely accepted in academia and industry to statically assign
and partition real-time tasks onto identical multiprocessor systems. This paper studies fixed-priority partitioned multiprocessor
scheduling for sporadic real-time systems, in which deadlinemonotonic scheduling is applied on each processor. Prior to this
paper, the best known results are by Fisher, Baruah, and Baker
2
1
with speedup factors 4 − M
and 3 − M
for arbitrary-deadline
and constrained-deadline sporadic real-time task systems, respectively, where M is the number of processors. We show that
1
when
a greedy mapping strategy has a speedup factor 3 − M
considering task systems with arbitrary deadlines. Such a factor
holds for polynomial-time schedulability tests and exponentialtime (exact) schedulability tests. Moreover, we also improve the
speedup factor to 2.84306 when considering constrained-deadline
task systems. We also provide tight examples when the fitting
strategy in the mapping stage is arbitrary and M is sufficiently
large. For both constrained- and arbitrary-deadline task systems,
the analytical result surprisingly shows that using exact tests does
not gain theoretical benefits (with respect to speedup factors) if
the speedup factor analysis is oblivious of the particular fitting
strategy used.
Keywords: Sporadic real-time tasks, resource augmentation, approximation, schedulability analysis.
1
Introduction
The sporadic task model has been widely adopted as the
basic model for real-time systems with recurring executions
[31]. A sporadic real-time task τi is characterized by its
minimum inter-arrival time Ti , its timing constraint or relative
deadline Di , and its (worst-case) execution time Ci . A sporadic
task defines an infinite sequence of task instances, also called
jobs, that arrive with the minimum inter-arrival time constraint.
Under the minimum inter-arrival time constraint, any two
consecutive jobs of task τi should be temporally separated by
at least Ti . When a job of task τi arrives at time t, the job
should finish no later than its absolute deadline t + Di . If we
consider a task releases its jobs periodically, such a task model
is the well-known Liu and Layland task model [30], where Ti
is the period of the task. An input task set is said to have
1) implicit deadlines if the relative deadlines of sporadic tasks
are equal to their minimum inter-arrival times, 2) constrained
deadlines if the minimum inter-arrival times are no less than
their relative deadlines, and (3) arbitrary deadlines, otherwise.
Through this paper, we only consider implicitly preemptive
scheduling. That is, a job may be preempted by another job
on a processor. For scheduling sporadic tasks on a processor,
the preemptive earliest-deadline-first (EDF) policy is optimal
[30] to meet the timing constraints. However, EDF requires to
prioritize the jobs in the ready queue by using their absolute
deadlines, and the overhead is in general not negligible. The
industrial practice is to use fixed-priority scheduling, also
supported in most real-time operating systems, in which a task
is assigned with a fixed priority level. The seminal work by Liu
and Layland [30] shows that rate monotonic (RM) scheduling
is optimal for uniprocessor fixed-priority scheduling when
considering implicit-deadline task systems. Moreover, deadline
monotonic (DM) scheduling [29] is optimal for uniprocessor
fixed-priority scheduling for constrained-deadline task systems. For arbitrary-deadline task systems, Audsley et al. [2]
provide an optimal priority assignment algorithm to define the
priority levels of the sporadic tasks for uniprocessor fixedpriority scheduling.
Testing whether a task set can be feasibly scheduled by
a scheduling algorithm is called a schedulability test. Even
though RM and DM are known to be optimal for uniprocessor fixed-priority scheduling, the exact schedulability tests
for uniprocessor fixed-priority scheduling requires pseudopolynomial time by using the exact tests by Lehoczky, Sha, and
Ding [28] for constrained-deadline systems and in exponential
time by Lehoczky [27] for arbitrary-deadline systems. Specifically, computing the worst-case response time of one (lowestpriority) task is shown N P-hard by Eisenbrand and Rothvoß
[20]. There have been extensive results about testing the
schedulability of uniprocessor fixed-priority scheduling. The
more efficient strategy is to provide only sufficient conditions
that can be verified in polynomial-time, like the utilization
bound [10], [26], [30], the quadratic utilization bound [7],
the hyperbolic utilization bound [8], [16], the approximated
request bound functions [1], [11], [21].
To quantify the performance loss due to efficient schedulability tests and assigning tasks with fixed priority levels,
we will adopt the notion of speedup factors, (also known
as resource augmentation factors). A fixed-priority scheduling
algorithm with a speedup factor ρ guarantees to produce
feasible schedules by running (each processor) ρ times as fast
as in the original platform (speed), if there exists a feasible
schedule (under arbitrary policies) for the task system. The
speedup factors of DM scheduling, with respect to the optimal
uniprocessor EDF scheduling, are ln12 , 1.76322, and 2 for
implicit-deadline, constrained-deadline, and arbitrary-deadline
task sets [17], [19], respectively.
To schedule real-time tasks on multiprocessor platforms,
there have been three widely adopted paradigms: partitioned,
global, and semi-partitioned scheduling. A comprehensive survey of multiprocessor scheduling in real-time systems can
be found in [18]. In this paper, we consider partitioned
scheduling, in which the tasks are statically partitioned onto
processors and all the processors are identical. That is, all the
partitioned EDF
partitioned DM
implicit deadlines
1
4
3 − 3M [23]
PTAS [24]
7
4 [10]
1.5 [25]
constrained deadlines
1
3− M
[5]
1
2.6322 − M
[13]
1
3 − M [22]
2.84306 (this paper)
arbitrary deadlines
2
4− M
[4]
1
3 − M [13]
2
4− M
[22]
1
3 − M (this paper)
TABLE I: Summary of the speedup factors in the multiprocessor partitioned scheduling problem for sporadic task systems. For
more details of implicit-deadline cases, please refer to Table III in the survey [18].
jobs of a task are executed on a specific processor with fixedpriority scheduling.
tioned scheduling. We focus on fixed-priority multiprocessor
partitioned scheduling, and improve the best known results
by Fisher, Baruah, and Baker [22]. The deadline-monotonic
partitioning algorithm is explored in a great detail in this paper.
Our contributions are:
However, problems on multiprocessors become N Pcomplete (or worse) in the strong sense even in the simplest
possible cases. For example, deciding if an implicit-deadline
task set with the same period is schedulable on multiple
processors is already N P-complete in the strong sense [31].
To cope with these N P-hardness issues, one natural approach
is to focus on approximation algorithms, i.e., polynomial time
algorithms that produce an approximate solution instead of an
exact one. There have been many results for implicit-deadline
task systems, as summarized in the survey paper [18]. But, only
a few results are known for constrained-deadline and arbitrarydeadline task systems.
• We show that the deadline-monotonic partitioning algo1
rithm has a speedup factor 3 − M
when considering task
systems with arbitrary deadlines, where M is the number
of processors. Such a factor holds for polynomial-time
schedulability tests and exponential-time (exact) schedulability tests. Moreover, we also improve the speedup
factor to 2.84306 when considering constrained-deadline
task systems by using polynomial-time and pseudopolynomial-time (exact) schedulability tests.
• The existing results by adopting the deadline-monotonic
partitioning algorithm were analyzed based on approximated schedulability tests. One of our key contributions
is to answer the question: Will it be possible to further
reduce the speedup factors by using exact tests in the
deadline-monotonic partitioning algorithm? Our answer
to this question is NO!! Using exact tests in the above
algorithm does not have any chance to reduce the speedup
factors if the speedup factor analysis is oblivious of
the particular fitting strategy used. We show that all
the speedup factor analyses in this paper are asymptotically tight with polynomial-time schedulability tests and
exponential-time (or pseudo polynomial-time) schedulability tests if the speedup factor analysis is oblivious of the
particular fitting strategy used. As a result, to improve the
speedup factor, better fixed-priority scheduling strategies
or more precise analyses for concrete fitting strategies are
needed.
When considering sporadic task sets with constrained or
arbitrary deadlines, the problem becomes more complicated,
when EDF or fixed-priority scheduling is adopted on a processor. The recent studies in [3], [14] provide polynomialtime approximation schemes for some special cases when
speeding-up is adopted for EDF scheduling. For general cases,
Baruah and Fisher [4], [5] propose a simple method, denoted
as deadline-monotonic partitioning in this paper, which 1)
considers the tasks in a non-decreasing order of their relative
deadlines, and 2) assigns a task (in the above order) to a
processor if it can pass the schedulability condition. If there
are multiple processors that are feasible for assigning a task,
the deadline-monotonic partitioning algorithm by Baruah and
Fisher [4], [5] uses the first-fit strategy, but the analysis works
for any arbitrary fitting strategy. The (theoretical) advantage of
the first-fit strategy was not shown in the literature when we
consider constrained- or arbitrary-deadline task systems.
The deadline-monotonic partitioning strategy is simple, but
has been shown effective in the literature [4], [5], [13], [22].
When adopting speeding-up for resource augmentation, by
using EDF on a processor, the deadline-monotonic partitioning
proposed by Baruah and Fisher [4], [5] has been shown
1
with a 3 − M
speedup factor by Chen and Chakraborty
[13], where M is the given number of identical processors.
Prior to this paper, for fixed-priority multiprocessor partitioned scheduling for constrained- and arbitrary-deadline task
systems, the best known results are by Fisher, Baruah, and
2
1
Baker [22] with speedup factors 4 − M
and 3 − M
for
arbitrary-deadline and constrained-deadline sporadic real-time
task systems, respectively. All the above results are based
on a linear-approximation to efficiently and safely test the
schedulability under EDF or DM scheduling to decide whether
a task can be assigned on a processor.
2
2.1
System Models and Preliminary Results
Task and Platform Model
We consider a set T = {τ1 , τ2 , . . . , τN } of N independent
sporadic real-time tasks. A task τi is defined by (Ci , Ti , Di ).
That is, for task τi , Di is its relative deadline, Ti is its
minimum inter-arrival time (period), and Ci is its (worst-case)
execution time. We consider identical processors in the platform. Therefore, no matter which processor a task is assigned
to, the execution and timing property remains. According to
the relations of the relative deadlines and the minimum interarrival times of the tasks in T, the task set can be identified
to be with 1) implicit deadlines, i.e., Di = Ti , ∀τi ∈ T, 2)
constrained deadlines, i.e., Di ≤ Ti , ∀τi ∈ T, or 3) arbitrary
deadlines.
Our Contributions: Table I summarizes the related results
and the contribution of this paper for multiprocessor parti2
i
For brevity, the utilization of task τi is denoted by Ui = C
Ti .
Ci
Moreover, let ∆i be max{Ui , Di }. For a set X, its cardinality
is denoted by |X|.
are pretty straightforward. As we focus on arbitrary-deadline
and constrained-deadline sporadic task systems, we can also
quantify the necessary condition defined by the demand. Here,
we can release the first job of tasks synchronously (say, at
time 0), and the subsequent job arrivals should be as rapidly
as legally possible. A necessary condition to be schedulable is
to ensure that the total execution time of the jobs arriving no
earlier than a and with relative deadlines no later than d is no
more than M · (d − a) for any a < d. This is identical to the
well-known demand bound function dbf (τi , t), as in [6], of a
task τi within any time interval with length equal to t, defined
as
t − Di
dbf (τi , t) = max 0,
+ 1 × Ci .
(1)
Ti
We will consider preemptive fixed-priority scheduling
on each processor. Specifically, we will only use deadlinemonotonic (DM) scheduling on each processor to assign the
priority levels of the tasks. That is, task τi is with higher
priority than τj if Di < Dj , in which the ties are broken
arbitrarily. Therefore, for the rest of this paper, we index
the tasks from the shortest relative deadline to the longest,
i.e., Di ≤ Dj if i < j. Note that DM priority assignment
is an optimal fixed-priority scheduling for implicit-deadline
and constrained-deadline task sets [29]. Although DM priority
assignment is not an optimal fixed-priority assignment when
we consider arbitrary-deadline task systems, it has a constant
speedup factor, to be discussed in Section 2.3.
2.2
Therefore, as a necessary condition, to ensure the schedulability on M processors, if a task set is schedulable for an
algorithm on M processors, then
X
∀t > 0,
dbf (τi , t) ≤ M t.
(2)
Problem Definition
τi ∈T
Given task set T, a feasible task partition on M identical
processors is a collection of M subsets, says, T1 , T2 , . . . , TM ,
of T such that
With the above discussions, we can conclude the following
lemma for the necessary condition to be schedulable by
any algorithm, which has also been utilized by Chen and
Chakraborty [13].
• Tm ∩ Tm0 = ∅ for all m 6= m0 ,
• ∪M
m=1 Tm is equal to the input task set T, and
• set Tm can meet the timing constraints by DM scheduling
on a processor m.
Lemma 1: A task set is not schedulable by any multiprocessor scheduling algorithm by running the M processors at
any speed s if
P
P
Ui
τi ∈T dbf (τi , t)
max max
, τi ∈T , max ∆i > s.
t>0
τi ∈T
Mt
M
(3)
Without loss of generality, we can assume that Ui ≤ 100%
Ci
and D
≤ 100%, i.e., ∆i ≤ 100%, for any task τi ; otherwise,
i
there is clearly no feasible task partition.
2.3
Speedup Factors/Bounds
For the rest of the paper, we will focus ourselves on the
negation part of the speedup factor analysis. That is, we are
only interested in the failure cases of the partitioning algorithm
and use Lemma 1 to quantify s for showing the speedup
factors. Note that Lemma 1 is also the necessary condition
for global multiprocessor scheduling. It may seem that we are
more pessimistic by comparing to the necessary condition of
global multiprocessor scheduling instead of that of partitioned
multiprocessor scheduling. However, in our tightness analysis,
comparing to partitioned scheduling and global scheduling
does not differ very much.
This paper focuses on the case where the arrival times of
the sporadic tasks are not specified. Therefore, the approximation is for the worst cases by considering the worst-case
behaviour to be feasibly scheduled by DM. If an algorithm A
for the studied problem has a speedup factor ρ, it guarantees to
always produce a feasible solution by speeding each processor
up to ρ times of the original speed in the platform, if task set
T can be feasibly scheduled (not restricted to DM) on the
original M identical processors. In other words, by taking
the negation of the above statement, we know that if the
algorithm A fails to feasibly partition the task set T on M
identical processors, there is no feasible task partition when
each processor runs ρ1 times slower than the original platform
speed. For the rest of this paper, we use 1 to denote the original
platform speed. Therefore, running the platform at speed s
implies that the execution time of task τi becomes Csi . Note
that speedup factors are used for quantifying the behaviour
of the designed algorithm. This is useful, especially for the
negation part to quantify the error the algorithm makes when
it does not provide a feasible solution.
3
Deadline-Monotonic Partitioning
This section presents the deadline-monotonic partitioning
strategy, proposed by Baruah and Fisher [4], [5], [22], for the
multiprocessor partitioned scheduling problem. Note that such
a strategy works in general for fixed-priority scheduling (RM,
DM) and dynamic-priority scheduling (EDF), by adopting
proper schedulability tests. The speedup factor for EDF/DM
2
1
and 4− M
[4], [5], [22] for constrainedwas shown to be 3− M
deadline systems and arbitrary-deadline systems, respectively.
When considering EDF scheduling, Chen and Chakraborty
1
1
[15] improved the speedup factor to 2.6322− M
and 3− M
for
constrained-deadline systems and arbitrary-deadline systems,
respectively.
For fixed-priority scheduling, the speedup factors of DM
scheduling, with respect to the optimal uniprocessor EDF
scheduling, are ln12 , 1.76322, and 2 for implicit-deadline,
constrained-deadline, and arbitrary-deadline task sets [17],
respectively. To quantify the schedulability of the input task
set, we would need to know the necessary condition for
being schedulable at speed s on the M processors.
The
P
Ui
necessary conditions maxτi ∈T ∆i ≤ s and
≤
s
τi ∈T M
For completeness, we repeat the algorithm in [4], [5],
[22], in which the pseudo-code is presented in Algorithm 1.
Deadline-monotonic partitioning considers the given tasks
3
Algorithm 1 Deadline-Monotonic Partitioning
pseudo-polynomial-time complexity, as all the points that lie
in (0, Dk ] need to be checked for Eq. (4).
Input: set T of N tasks, M processors;
1: re-index (sort) tasks such that Di ≤ Dj for i < j;
2: T1 ← {τ1 };Tm ← ∅, ∀m = 2, 3, . . . , M ;
3: for k = 2 to N do
4:
if ∃m ∈ {1, 2, . . . , M } such that Tm ∪ {τk } is schedulable by DM
fixed-priority scheduling then
5:
choose m ∈ {1, 2, . . . , M } by preference such that Tm ∪ {τk }
is schedulable by DM fixed-priority scheduling;
6:
assign τk to processor m with Tm ← Tm ∪ {τk };
7:
else
8:
return “no feasible schedule is found”;
9:
end if
10: end for
11: return feasible task partition T1 , T2 , . . . , TM ;
Fisher, Baruah, and Baker [22] approximate the test in
Eq. (4) by testing only
X
t
1+
∃t with 0 < t ≤ Dk and Ck +
Ci ≤ t.
Ti
τi ∈Tm
(5)
Due to the linearity of the condition in Eq. (5), the test is
equivalent to the verification of whether
X
Dk
1+
Ck +
Ci ≤ Dk
(6)
Ti
τi ∈Tm
for constrained-deadline systems.
from the shortest relative deadline to the longest relative
deadline for assignment. When a task τk is considered, a
processor m with m ∈ {1, 2, . . . , M } is selected to assign
task τk , where Tm is the set of the tasks (as a subset of
{τ1 , τ2 , . . . , τk−1 }), which have been assigned to processor
m before considering τk . If there is no feasible m that can
feasibly schedule τk and Tm on the processor, we return that
no feasible solution is found by this algorithm.
3.1
We can also approximate the schedulability test by using
utilization-based analysis as follows: We classify the task set
Tm into two subsets:
• T1m consists of the higher-priority tasks with periods
smaller than Dk .
• T2m consists of the higher-priority tasks with periods
larger than or equal to Dk .
Fitting Strategy
The following theorem has been concluded recently by using
a utilization-based schedulability-test framework [16].
The fitting strategy when we consider to assign task τk on
a processor m can be
Theorem 1 (Chen, Huang, Liu [16]): Task τk in a sporadic task system with constrained deadlines is schedulable
by DM scheduling algorithm on processor m if
• the first-fit strategy: by choosing the minimum m that is
feasible;
• the arbitrary-fit strategy: by choosing any m that is
feasible (this is also the case if the speedup factor analysis
is oblivious of the particular fitting strategy used);
• the best-fit strategy: by choosing the index m that has the
maximum workload-index;
• the worst-fit strategy: by choosing the index m that has
the minimum workload-index.
(
0
Y
Ck,m
+ 1)
(Uj + 1) ≤ 2.
Dk
1
(7)
τj ∈Tm
0
where Ck,m
is Ck +
P
τi ∈T2m
Ci .
Arbitrary Deadline: For arbitrary-deadline systems, the exact
schedulability test is to use a busy-window concept to evaluate
the worst-case response time [27] by using TDA. The finishing
time Rk,h of the h-th job of task τk in the busy window is the
minimum t such that
X t
Ci ≤ t.
hCk +
Ti
The workload-index can be defined as the total utilization or
other means. The analysis in the literature [4], [5], [15], [22]
works in general by using any fitting strategy listed above, even
though in several cases only the first-fit strategy was mentioned
in the descriptions [4], [5], [22].
τi ∈Tm
3.2
Schedulability Tests for DM
Therefore, its response time is Rk,h − (h − 1)Tk . To test
the busy window length of task τk , the busy window of task
τk finishes on the h-th job if Rk,h ≤ hTk . The maximum
response time among the jobs in the busy window is the worstcase response time [27]. The schedulability test is exact, but
this takes exponential time complexity, since the busy window
length can be up to the hyper-period, which is exponential of
the input size.
Therefore, the remaining building block is to test whether
task τk can be feasibly scheduled on a processor m under
DM scheduling. This has been widely studied in the literature.
We will review some of these methods and explain their
corresponding speedup factors when they are adopted in Step
4 in Algorithm 1.
Constrained Deadline: To verify the schedulability of
constrained-deadline task τk under fixed-priority scheduling in
uniprocessor systems, the time-demand analysis (TDA) [28]
can be adopted. That is, if and only if
X t
∃t with 0 < t ≤ Dk and Ck +
Ci ≤ t, (4)
Ti
The approximation by Fisher, Baruah, and Baker [22] in
Eq. (5) remains feasible for analyzing the arbitrary-deadline
systems. This leads to test whether
X
Dk
Ci ≤ Dk and
(8a)
Ck +
1+
Ti
τi ∈Tm
X
Uk +
Ui ≤ 1.
(8b)
τi ∈Tm
then task τk is schedulable under DM scheduling, where Tm is
the set of tasks with higher priority than task τk since we sort
the tasks according to their relative deadlines. TDA requires
τi ∈Tm
Eq. (8b) is important in arbitrary-deadline systems to ensure
4
that the approximation in Eq. (8a) does not underestimate the
workload after Dk .
the violation of Eq. (8a), we know that
X X
Dk
|M1 |Ck +
1+
Ci > |M1 |Dk
Ti
m∈M1 τi ∈Tm
X X Ci
Ck
⇒|M1 |
+
+ Ui > |M1 |.
(10)
Dk
Dk
Moreover, Bini et al. [9] provide a tighter analysis than
Eq. (8a). They show that the worst-case response time of task
τk is at most
P
P
Ck + τi ∈Tm Ci − τi ∈Tm Ui Ci
P
.
1 − τi ∈Tm Ui
m∈M1 τi ∈Tm
By the violation of Eq. (8b), we know that
X X
Ui > |M2 |.
|M2 |Uk +
Therefore, the schedulability condition in Eqs. (8a) and (8b)
can be rewritten as
X
X
X
Ui Ci ≤ Dk , (9a)
Ci −
Ui ) +
Ck + Dk (
τi ∈Tm
τi ∈Tm
Ck
Recall that ∆k is defined as max{Uk , D
}. By Eqs. (10)
k
and (11), we know that
τi ∈Tm
Uk +
X
Ui ≤ 1. (9b)
|M1 |
τi ∈Tm
3.3
Time Complexity and Correctness
k−1
X
X
Ck
+ |M2 |Uk +
Ui +
Dk
i=1
m∈M
⇒ M ∆k +
k−1
X
Ui +
k−1
X
i=1
Since we use partitioned scheduling, as long as Algorithm 1
returns a task partition, it is guaranteed to be feasible if the test
in Step 4 is a sufficient schedulability test for task τk by using
DM scheduling. The time complexity of the algorithm depends
upon the time complexity of the schedulability test and the
fitting strategy. Suppose that the fitting strategy requires time
complexity O(F ) for one task and the time complexity to test
whether task τk is schedulable on processor m is O(H). Then,
the overall time complexity is O(N M H +N F ). All the fitting
strategies mentioned in Section 3.1 are in polynomial time.
Since H can be polynomial, pseudo polynomial, or exponential
of the input size, the time complexity is dominated by the
adopted schedulability test algorithm.
⇒ (M − 1)∆k +
i=1
k
X
i=1
X
1
τi ∈Tm
Ci
>M
Dk
Ci
> M.
Dk
Ui +
k
X
Ci
> M.
D
k
i=1
Pk
k
Ui X Ci
1
)∆k + i=1
+
> 1.
⇒ (1 −
M
M
M Dk
i=1
(
)
Pk
k
X
1
Ci
i=1 Ui
⇒ (3 −
) max ∆k ,
,
> 1.
M
M
M Dk
i=1
Pk
(12)
U
i
Therefore, we know that either ∆k > 3−1 1 , or i=1
>
M
M
P
k
Ci
1
1
1 , or
1 . Either of the former two
i=1 M Dk > 3− M
3− M
cases implies the unschedulability of any scheduling algorithm
with speed 3−1 1 . The demand bound function at time Dk
M
PN
Pk
is i=1 dbf (τi , Dk ) ≥ i=1 Ci . Therefore, by Eq. (2), we
Pk
Ci
1
implies the
know that the condition
1
i=1 M Dk > 3− M
unschedulability of any scheduling algorithm with speed 3−1 1 .
The main issue here is to answer what can be guaranteed
when Algorithm 1 returns failure in task partitioning. We will
quantify such failures by showing that s in Eq. (3) is also
sufficiently large to provide the speedup factor guarantee (by
using the negation arguments).
4
(11)
m∈M2 τi ∈Tm
M
Therefore, by using Lemma 1, we reach the conclusion.
The following corollaries show that the speedup factor 3 −
holds for any schedulability tests discussed in Section 3 for
arbitrary-deadline sporadic task systems under DM scheduling.
1
M
Analysis for Arbitrary-Deadline Systems
1
Corollary 1: The speedup factor of Algorithm 1 is 3 − M
when adopting Eqs. (9a) and (9b) for DM schedulability test
under any fitting strategy.
P
Proof: By not considering the term − τi ∈Tm Ui Ci in
Eq. (9a), the violation of Eq. (9a) leads to the same conclusion
1
.
in Eq. (10). Therefore, the speedup factor remains 3 − M
This section presents our analysis for arbitrary-deadline
systems. Our analysis is similar to the analysis by Fisher,
Baruah, and Baker [22], but is tighter. Here, we will mainly
analyze the property by using the schedulability condition
in Eqs. (8a) and (8b). At the end of this section, we will
explain why the analysis also works for arbitrary-deadline
TDA analysis by Lehoczky [27] and response time analysis
by Bini et al. [9].
1
Corollary 2: The speedup factor of Algorithm 1 is 3 − M
when adopting the exact schedulability test for DM scheduling
under any fitting strategy.
1
Theorem 2: The speedup factor of Algorithm 1 is 3 − M
when adopting Eqs. (8a) and (8b) for DM schedulability test
under any fitting strategy.
Proof: If task τk cannot pass the exact schedulability test,
it also does not pass the sufficient test by using Eqs. (8a) and
(8b). Therefore, we reach the same conclusion.
Proof: Suppose that Algorithm 1 fails to find a feasible
assignment for task τk due to the failure when testing Eq. (8a)
or Eq. (8b). Let M1 be the set of processors in which Eq. (8a)
fails. Let M2 be the set of processors in which Eq. (8a)
succeeds but Eq. (8b) fails. Since task τk cannot be assigned
on any of the M processors, we have |M1 | + |M2 | = M . By
4.1
Tightness Analysis
The following theorem shows that the analysis in Theorem 2 is asymptotically tight even for implicit-deadline systems
with first-fit strategy.
5
Theorem 3: The speedup factor of Algorithm 1 is at least
3 − M3+1 − γ when adopting Eqs. (8a) and (8b) for DM
schedulability test under the first-fit strategy, where γ is an
arbitrarily small positive number.
Proof: This theorem is proved by a concrete input task
system with 3M tasks. Let δ and be very small positive real
numbers, with δ . There are M tasks with execution time
1
3M , period ∞, and relative deadline 1 − δ. There are M tasks
with execution time 3 , period , and relative deadline 1. There
are M tasks with execution time 1+
3 , period ∞, and relative
deadline 1 + δ.
Proof: This theorem is proved by a concrete input task
system with N = 2M tasks. There are M light tasks with
1
execution time 3M
and M heavy tasks with execution time
1+
3 , in which is a small positive real, i.e., > 0. The M
light tasks are all with period 1 − δ and relative deadline 1 − δ
with arbitrarily small and positive δ . The M heavy tasks
are all with period 1 and relative deadline 1. Therefore, the
2M tasks are indexed such that
• Ci =
• Ci =
1
3M ,
1+
3 ,
Therefore, the 3M tasks are indexed such that
1
• Ci = 3M
, Ti = ∞, Di = 1 − δ, for i = 1, 2, . . . , M , and
• Ci = 3 , Ti = , Di = 1, for i = M + 1, M + 2, . . . , 2M .
• Ci = 1+
3 , Ti = ∞, Di = 1 + δ, for i = 2M + 1, 2M +
2, . . . , 3M .
Ti = Di = 1 − δ, for i = 1, 2, . . . , M , and
Ti = Di = 1, for i = M + 1, M + 2, . . . , 2M .
Again, the setting of δ is just to enforce the indexing.
We will directly take δ → 0 for the rest of the proof. Now, we
consider a feasible task assignment for the first 3M − 1 tasks,
in which
The setting of 0 < δ is just to enforce the indexing. We
will directly take δ → 0 for the rest of the proof.1
By using Algorithm 1 for the above task set when adopting
Eqs. (8a) and (8b) for DM schedulability test under the first-fit
strategy, the M light tasks are assigned on processor 1. Then,
when task τk with k > M is considered, the condition in
Eq. (8a) always fails for any of the first k − M processors.
Therefore, task τk is assigned to processor k − M + 1, for
k = M + 1, M + 2, . . . , 2M − 1. It is then clear that task τ2M
cannot be assigned on any of the M processors. Therefore,
Algorithm 1 returns “no feasible solution is found”.
P2M
By the above setting, we have i=1 Ui = 1+M3+M . By
using Lemma 1, we know that the speedup factor of the above
task set is at least
3
1
3 + 3M
=
3−
−γ
=3 −
1+M +M
M + 1 + M
M +1
3
• τ1 , τ2 , . . . , τM +1 are assigned on processor 1, and
• τi and τi+M −1 are assigned on processor i − M for i =
M + 2, M + 3, . . . , 2M .
By using Algorithm 1 for task τ3M , we know that task
τ3M cannot be feasibly assigned on any of the M processors since ∀0 <
l tm ≤ 1 and m = 1, 2, . . . , M , we have
P
1+
t
+
τi ∈Tm Ti Ci > t. Therefore, Algorithm 1 returns
3
“no feasible solution is found”.
P3M
By the above setting, we know that (1) i=1 dbf (τi , t) = 0
P3M
+1
for 0 < t < 1, (2) i=1 dbf (τi , 1) = 2M +M
, and (3) for
3
1 < t,
3M
X
1
1+
1
M+ M
dbf (τi , t) ≤ + (t − 1) M +
3
3
3
3
i=1
2
3M
in which the factor γ = (M +1)(M
+1+M ) becomes negligible
when is sufficiently small.
=
Such a factor can also be shown by a concrete partitioned
schedule. By the pigeonhole principle, the solution that minimizes the maximum utilization of a processor is to assign
a light task and a heavy task on a processor, in which the
1
utilization on the processor is 1+
3 + 3M . Therefore, the task
set is not schedulable on M processors only when the speed
1
is slower than 1+
3 + 3M , provided that δ is 0. As a result, we
reach the same conclusion.
2M + M t + 1
.
3
P3M
1+2+
dbf (τ ,t)
1
M
As a result, maxt>0 i=1 M t i =
3 P , when is small
3M Ui
1+
1
enough. Since maxτi ∈T ∆i = 3 and
i=1 M = 3 , by
Lemma 1, the speedup factor of the above task set is
1
1
3M
1
It may seem at first glance that the speedup factor 3 − M
in Corollary 2 is pessimistic, since we do not actually use
any property in the exponential-time exact schedulability test.
However, the following theorem shows that the speedup factor
1
3− M
is asymptotically tight for an arbitrary fitting strategy, for
any schedulability tests used in Theorem 2, Corollary 1, and
Corollary 2. As a result, to improve the speedup factor, better
fixed-priority scheduling strategies or more precise analysis for
concrete fitting strategies are needed.
+
2
3
+
1
3
3M
M + 2M + 1
3 + 6M
3
=3 −
=3−
− γ,
M + 1 + 2M
M +1
=
2
6M
in which the factor γ = (M +1)(M
+1+2M ) becomes negligible
when is sufficiently small. The above analysis also works
for the specific task partitioning which assigns three tasks
τi , τi+M , and τi+2M on one processor for i = 1, 2, . . . , M ,
and assigns the priority levels by using the optimal priority
assignment by Audsley et al. [2], i.e., τi+M has the lowest
priority on the processor.
Theorem 4: The speedup factor of Algorithm 1 is at least
3 − M3+1 − γ under an arbitrary fitting strategy, for any
schedulability tests used in Theorem 2, Corollary 1, and
Corollary 2, where γ is an arbitrarily small positive number.
5
Analysis for Constrained Deadlines
This section presents the analysis for constrained-deadline
sporadic real-time systems. By Theorem 3, we know that
the method by Fisher, Baruah, and Baker [22] leads to a
speedup factor 3 when M is sufficiently large even for implicitdeadline systems. The reason is mainly due to the pessimism
of Eq. (6) in the schedulability test. To get better results, we
1 That is, the tightness examples in Theorems 3, 4, and 9 hold even when δ
is 0 if the tasks are indexed in the specified order. Setting δ to an arbitrarily
small positive number is more precise mathematically, but this just complicates
the presentation of the proofs without any added value.
6
then Algorithm 1 by using TDA should succeed to assign task
τk on one of the M processors.
do need better tests. A more precise strategy is to simply use
the exact test for constrained-deadline systems by spending
pseudo-polynomial time complexity. We have already shown
(by Corollary 2 and Theorem 4) that spending more time
complexity does not help in arbitrary-deadline systems if the
analysis does not use the property of any specific fitting strategy. Is this also the same for constrained-deadline systems?
For completeness, we repeat the definition of the k2 U
framework and the key Lemma (with some simplifications to
remove individual coefficients for each task τi ) as follows.
Definition 1: A k-point effective schedulability test is a
sufficient schedulability test of a fixed-priority scheduling
policy by verifying the existence of tj ∈ {t1 , t2 , . . . tk } with
t1 ≤ t2 ≤ · · · ≤ tk such that
We will first present the analysis by using TDA as the
schedulability test in Step 4 in Algorithm 1. We will conclude
later that such high time complexity also does not help reduce
the speedup factor, compared to the results by using the
hyperbolic bound in Theorem 1.
Ck +
k−1
X
αti Ui +
i=1
5.1
Speedup Factor by Adopting TDA
X t
Ck +
Ci > t.
Ti
α
Ck
α
β +1
≤ Qk−1
− .
tk
β
j=1 (βUj + 1)
τi ∈Tm
By taking a summation of all the m = 1, 2, . . . , M inequalities
with respect to any t, we know that the unschedulability of task
τk by Algorithm 1 implies that
X t
∀t with 0 < t ≤ Dk ,
M Ck +
Ci > M t.
Ti
τi ∈T∗
(13)
(18)
Proof: This comes from Lemma 1 in [16].
By adopting the k2 U framework [16], we can conclude the
following theorem.
Theorem 5: If
Y
(1 +
τi ∈T∗1
Ui
2
,
)≤
C0
M
1 + Dkk
then task τk is schedulable under Algorithm 1 by using TDA.
(14)
Proof: In the proof, we will reindex the tasks to satisfy
the monotonicity of ti in Definition 1. That is, the κ−1 higherpriority tasks in T∗1 are reindexed to form the corresponding
sequence τ1 , τ2 , . . . , τκ−1 for ensuring that the arrival times of
the last jobs no later than Dk are
j ink a non-decreasing order.
k
For task τi in T∗1 , we set ti as D
Ti Ti . Now, we reindex the
κ − 1 higher-priority tasks such that t1 ≤ t2 ≤ · · · ≤ tκ−1 .
Moreover, let tκ be Dk .
τi ∈T
then Algorithm 1 by using TDA should succeed to assign task
τk on one of the M processors. This is basically very similar
to TDA with a minor difference by dividing the higher-priority
workload by M .
Testing the schedulability condition of task τk according
to Eq. (14) can be done by using the same strategy used in
the k2 U framework [16] to prove Theorem 1 as follows.
Since Ti < Dk for any task τi in T∗1 , we have ti ≥ Ti .
Therefore, for a given tj with j = 1, 2, . . . , κ, the demand
requested up to time tj in Eq. (16) is at most
l m
P
P
tj
τi ∈T∗2 Ci +
τi ∈T∗1 Ti Ci
Ck +
M
Pκ−1 l tj m
C
i
i=1
Ti
=Ck0 +
Pκ−1Mti
Pj−1
Ci
i=1 Ti Ci
0
≤Ck +
+ i=1 ,
M
M
where the inequality
comes from the indexing
l m
l m policy defined
t
t
above, i.e., Tji ≤ Ttii + 1 if j > i and Tji ≤ Ttii if j ≤ i.
We classify the k − 1 tasks in T∗ into two subsets.
• T∗1 consists of the tasks in T∗ with periods smaller than
Dk .
• T∗2 consists of the tasks in T∗ with periods larger than
or equal to Dk .
Now, let Ck0 be defined as follows:
X Ci
Ck0 = Ck +
.
M
∗2
(17)
Lemma 2 (Chen, Huang, and Liu [16]): For a given kpoint effective schedulability test, defined in Definition 1, of a
scheduling algorithm, in which 0 < α 6= ∞, and 0 < β 6= ∞,
task τk is schedulable by the scheduling algorithm if the
following condition holds
Therefore, we know that this leads to
Therefore, by taking the negation, we know that if
l m
X Tti Ci
≤ t,
∃t with 0 < t ≤ Dk , and Ck +
M
∗
βti Ui ≤ tj ,
i=1
where Ck > 0, α > 0, Ui > 0, and β > 0 are dependent upon
the setting of the task models and task τi .
Now, suppose that task τk is the first task that fails to
be assigned on any of the M processors by using TDA
schedulability analysis in Step 4 in Algorithm 1. For notational
brevity, let T∗ be the set {τ1 , τ2 , . . . , τk−1 } of the tasks
∀m, ∀t, with 0 < t ≤ Dk ,
j−1
X
(15)
τi ∈T
Suppose that we have κ − 1 tasks in T∗1 . Clearly, according
to the definition κ ≥ 1. Now, we can rewrite the condition in
Eq. (14) as follows: if
l m
t
X
Ti C i
∃t with 0 < t ≤ Dk and Ck0 +
≤ t, (16)
M
∗1
We only apply the test for these κ different ti values, which
is equivalent to thePtest of the existence of tj for j = 1, 2, . . . , κ
such that Ck0 +
τi ∈T
7
κ−1 ti
i=1 Ti
M
Ci
+
Pj−1
i=1
M
Ci
≤ tj . This satisfies
Definition 1 (when k is κ) with α =
by using Lemma 2, if
1
M
and β =
1
M.
Q
i
the other case with τi ∈T∗1 ( U
M + 1) < 2. Suppose that σ
2
is Q
− 1, in which σ > 0.
Ui
Therefore,
τi ∈T∗1 ( M
Ck0
1+1
− 1,
≤Q
Ui
Dk
τi ∈T∗1 ( M + 1)
If τk is not schedulable under Algorithm 1 by TDA, then
Pk−1
dbf (τi , Dk )
Ck
+ i=1
Dk
M
Dk
P
P
Ck
τi ∈T∗2 Ci
τi ∈T∗1 dbf (τi , Dk )
+
+
=
Dk
M Dk j k
M Dk
P
Dk
0
T
U
i
i
∗1
τi ∈T
Ti
C
≥ k +
Dk
M
D
k
Y
Ui
1+σ
1+
>1
=
,
M
2
∗1
then task τk is schedulable under Algorithm 1 by using TDA.
By reorganizing the above equation, we reach the conclusion.
The following corollary comes from the same proof of
Lemma 2 and Theorem 5, which shows that the schedulability condition
also implies a lower bound of the workload
P
Ck0
τi ∈T∗1 ti Ui
+
.
Dk
M Dk
τi ∈T
Corollary 3: Q
If the schedulability condition in Theorem 5
2
i
is violated, i.e., τi ∈T∗1 (1 + U
, then
C0
M) >
Ck
where >1 is by Corollary 3. Suppose that D
is x. Therefore,
k
we know that
X dbf (τi , Dk )
1+σ
>
− x.
(19)
M
2
∗
1+ Dk
k
Ck0
Dk
P
+
τi
j
∈T∗1
Dk
Ti
M Dk
k
Ti Ui
>
Y
τi ∈T∗1
1+
+1)
Ui
M
.
τi ∈T
Q
2
i
Moreover, with τi ∈T∗1 (1 + U
M ) = 1+σ and the fact that
τk is not schedulable under Algorithm 1 by using TDA, by
Lemma 4, we have
X Ui
2
> ln(
).
(20)
M
1
+
σ
∗1
Proof: This comes from the same proof as in Lemma 2
(Lemma 1 in [16]) by changing
Pk−1 the objective from minimizing
Ck∗ to minimizing Ck∗ + i=1 αti Ui to enforce the unschedulability. This property has been provided by Chen, Huang, and
Liu [16] to prove the speedup factor of deadline-monotonic
scheduling in uniprocessor systems. Moreover, this condition
has also been exploited by von der Brüggen, Chen, and
Huang [32] to prove the speedup factor of non-preemptive DM
scheduling with respect to non-preemptive EDF scheduling.
τi ∈T
For the rest of the proof, we consider two separate cases:2
Case 1 x ≥ σ: This is an easier case. We can conclude the
speedup factor by using Eq. (20)
(
)
2
Ck X Ui
≥ max σ, ln(
)
,
max
Dk
M
1+σ
τi ∈T∗
2
1
≥ min max{σ, ln(
)} =1
,
σ>0
1+σ
2.66793
The remaining proofs in this section require some mathematical tools, which are provided in the following lemmas.
P
Lemma 3: Suppose that τi ∈T1∗ Ui >
0 is fixed and Ui ≥
Q
i
0 for each task τi . Then, τi ∈T1∗ U
+
1
is maximized when
M
U1 = U2 = · · · = U|T1∗ | .
Proof: This can be easily proved by the fact that
Ui
τi ∈T1∗ M + 1 is a concave function with respect to the
(non-negative) values of Ui .
P
Ui
Lemma 4: The
infimum
to enforce
τi ∈T1∗ M
Q
Ui
+
1
>
x
is
ln(x).
1∗
τi ∈T
M
where =1 holds when eσ (1 + σ) = 2.
Q
Case 2 x < σ: There are two subcases
1+σ
• Case 2a: If x > 1+σ
4 , wenknow that x >
2 −
o x.
P
dbf
(τ
Ck
i ,Dk )
Therefore, by Eq. (19), max Dk , τi ∈T∗
≥
M
x > 1+σ
.
Hence,
4
(
)
Ck X Ui X dbf (τi , Dk )
max
,
,
Dk
M
M
τi ∈T∗
τi ∈T∗
1+σ
2
1
> max
, ln(
) ≥ min max y, ln
y≥0
4
1+σ
2y
1
=W (0.5) ≈
,
2.84306
where = holds when yey = 0.5.
1+σ
• Case 2b: If x ≤ 1+σ
4 , wen know that x ≤
2 −
o x.
P
dbf (τi ,Dk )
Ck
Therefore, by Eq. (19), max Dk , τi ∈T∗
>
M
Proof: This can be derived by using
Lagrange MultiP
Ui
plier Method to find the minimum
τi ∈T1∗ M such that
Q
Ui
τi ∈T1∗ M + 1 ≥ x. By Lemma 3, it is clear that the
worst case is to have all the tasks with the same utilization.
Suppose that |T1∗ | is n. We know that all the tasks
P are withi
1
utilization M (x n − 1), and the utilization bound τi ∈T1∗ U
M
1
is n(x n − 1). This converges to ln(x) when n approaches ∞.
With the above discussions, we can conclude the speedup
factor.
Theorem 6: The speedup factor of Algorithm 1 for
1
constrained-deadline task systems by using TDA is W (0.5)
≈
2.84306, where W (z) is the Lambert W function, i.e., the
unique solution of z = W (z)eW (z) .
Q
i
Proof: If τi ∈T∗1 ( U
M + 1) ≥ 2, we can already conclude
P
Ui
that
τi ∈T∗1 M ≥ ln 2 by using Lemma 4, and the speedup
factor is 1/ ln 2 < 2.84306 for such a case. We focus on
2 These cases are concluded by the following trick: Suppose that g (y) is
1
an increasing function and g2 (y) is a decreasing function with respect to y
when y ≥ 0. Then, we know that the lower bound on max{g1 (y), g2 (y)}
is the intersection of these two functions, i.e., g1 (y ∗ ), where g1 (y ∗ ) =
g2 (y ∗ ) if such a value y ∗ exists. That is, max{g1 (y), g2 (y)} ≥
miny≥0 max{g1 (y), g2 (y)} = g1 (y ∗ ).
8
1+σ
2
− x ≥ 1+σ
4 . Hence,
(
)
Ck X Ui X dbf (τi , Dk )
max
,
,
Dk
M
M
τi ∈T∗
τi ∈T∗
1+σ
2
1
> max
, ln
≥ min max y, ln( )
y≥0
4
1+σ
2y
1
=W (0.5) ≈
,
2.84306
where = holds when yey = 0.5.
By contrapositive, we reach the conclusion.
Theorem 8: The speedup factor of Algorithm 1 by using
1
the hyperbolic bound in Eq. (7) is W (0.5)
≈ 2.84306, where
W (z) is the Lambert W function, i.e., z = W (z)eW (z) .
Proof: Since the schedulability condition remains the
same as in Theorem 5, the speedup factor is also the same
as in Theorem 6.
5.3
Therefore, by all the above cases, we know that
max
We conclude this section by showing that the above
speedup factor analysis is tight when M is sufficiently large
under an arbitrary fitting strategy.
Ck X Ui X dbf (τi , Dk )
1
,
,
,
> W (0.5) ≈
Dk
M
M
2.84306
τ ∈T∗
τ ∈T∗
i
Theorem 9: For constrained-deadline task systems, the
1
speedup factor of Algorithm 1 is at least W (0.5)
≈ 2.84306
when adopting TDA or the hyperbolic bound in Eq (7) for DM
schedulability test under an arbitrary fitting strategy, when M
is sufficiently large.
i
which concludes the proof by applying Lemma 1.
5.2
Tightness Analysis
Speedup Factor by Hyperbolic Bound
Proof: We prove this theorem by providing a concrete
task system by assuming that M is sufficiently large. There
2
are N = 2M 2 + 1 tasks. Let f be
≈ 0.7034674, i.e.,
1
This subsection further presents the speedup factor of
Algorithm 1 when adopting the hyperbolic bound in Eq. (7)
for testing the schedulability of DM scheduling. The speedup
factor analysis in Theorem 6 for TDA schedulability analysis
relies only on the violation of the schedulability condition
in Theorem 5. We will show that adopting the hyperbolic
bound of Eq. (7) results in the same condition in Theorem 5.
Therefore, we can reach the same conclusion as in Theorem 6
by using the hyperbolic bound in polynomial time. We use the
same notations, e.g., T∗1 , T∗2 , κ, etc., as used in Section 5.1.
W (0.5)
ln( f1 ) = f2 . Let δ be an arbitrarily small positive number just
for enforcing the indexing:
1−f
1−f
• Ti = Di = f + ( Mi − 1) M
−1 , Ci = M −1 , for i =
2
1, 2, . . . , M ,
−1
• Ti = ∞, Di = 1 + δ, Ci = 1.5f
≈ 0.0552
M
M , for i =
2
2
2
M + 1, M + 2, . . . , 2M , and
• TN = ∞, DN = 1 + 2δ, CN = 0.5f + , with N =
2M 2 + 1, where is a positive small number.
Theorem 7: If
Y
Ui
2
1+
≤
,
C0
M
1 + Dkk
τi ∈T∗1
We again simply take δ to 0 for the rest of the proof. Moreover,
1
M is also consider negligible for the simplicity of computation,
since M is assumed to be sufficiently large.
then task τk is schedulable under Algorithm 1 P
by using the
∗2 Ci
0
hyperbolic bound in Eq. (7), where Ck is Ck + τi ∈T
.
M
For an arbitrary fitting algorithm, consider the following
task assignment by assigning task τi+jM to processor i with
i = 1, 2, . . . , M for every j = 0, 1, 2, . . . , 2M − 1. It is not
difficult to see that the above task assignment can be achieved
feasibly and results in a feasible task assignment for the first
2M 2 tasks. The set of the tasks assigned on processor m is
denoted as Tm . Now, consider task τ2M 2 +1 , i.e., τN to be
assigned on processor m. The overall execution time request
at time 0 is 0.5f + + 1.5f − 1 + 1 − f = f + on processor
m. Therefore, it can be easily seen that task Tm ∪ {τN } is
not schedulable under DM scheduling on processor m since
the TDA test in Eq. (6) fails. As a result, task τN cannot be
assigned on any processor.
Proof: We prove this by contrapositive. Suppose that
task τk is not schedulable under Algorithm 1 by using the
hyperbolic bound in Eq. (7) . Therefore, for m = 1, 2, . . . , M ,
we have
!
P
Y
Ck + τi ∈T2m Ci
+1
(Ui + 1) > 2.
Dk
1
τi ∈Tm
By multiplying the above M inequalities we reach
!!
P
M
Y
Y
C
+
C
2
k
i
τi ∈Tm
2M <
+1
(Ui + 1)
D
k
m=1
τi ∈T∗1
M
P
M
∗2 Ci
Y Ui
Ck + τi ∈T
M
,
≤1
+ 1
+1
Dk
M
∗1
In this input task set, the utilization of the individual task
Ci
are not more than f2 + for each task τi . Moreover, the
and D
i
PM −1 µ
1−f
overall task utilization is M i=0 f +iµ
, where µ = M
−1 .
Due to the assumption that M is sufficiently large,
the
above
PM −1 µ
total utilization is a left Riemann sum, i.e., M i=0 f +iµ
≈
R 1−f 1
f
1
1
M 0
f +x dx = M ln( f ). By the fact that ln( f ) = 2 , we
know that the total utilization is M f2 when M is sufficiently
large.
τi ∈T
where ≤1 comes from Lemma 3 (for the first part) and from
the fact (1+Ui ) ≤ (1+Ui /M )M when M is a positive integer
and Ui ≥ 0 (for the second part). Therefore, we conclude that
the unschedulability of task τk implies that
0
Y Ui
Ck
2<
+1
+ 1 .
Dk
M
∗1
maxt>0
τi ∈T
9
N
i=1
dbf (τi ,t)
Mt
>
PN
i=1 dbf (τi ,t)
. By definition,
t
PN M
−1
i=1 dbf (τi ,t)
maxt>0
. By the conMt
Now,Pwe examine maxt>0
2.2
0.355
dbf#(t)
2
0.35
1.8
1.6
0.345
1.4
0.34
1.2
1
0.335
0.8
0.33
0.6
0.325
0.4
0.2
1
2f
2
4f 3
5f
4 6f
5
8f
0.32
6
f/2
dbf#(t)/t
0
5
10
t
15
20
25
30
t
(a) dbf ] (t)
(a) 1 ≤ t ≤ 30
0.36
0.352
dbf#(t)/t
0.3515
0.34
0.351
0.32
0.3505
0.3
0.35
0.28
0.3495
0.26
0.24
0.349
1
2f
2
4f 3
5f
4 6f
5
8f
0.3485
6
f/2
#
dbf (t)/t
500
1000 1500 2000 2500 3000 3500
t
t
(b) 30 ≤ t ≤ 4000
(b) dbf ] (t)/t
]
]
Fig. 2: dbft (t) when t is a positive integer number (the curve
is just for visualization) for the proof of Theorem 9, and the
reference point f2 ≈ 0.3517337
Fig. 1: Functions dbf ] (t) and dbft (t) for the proof of Theorem 9 with t in the range of [1, 6].
struction of the task P
set, we have (1)
N −1
i=1
dbf (τi ,1)
M
PN −1
i=1
dbf (τi ,t)
Mt
6
= 0 if
f
2
0 < t < 1, and (2)
= if t = 1. Therefore,
we know that the speedup
factor
for
this task set is purely
PN −1
i=1 dbf (τi ,t)
.
dominated by maxt>0
Mt
PN −1
Concluding Remarks
This paper provides detailed analysis for the deadlinemonotonic partitioning algorithm proposed by Fisher, Baruah,
and Baker [22] for multiprocessor partitioned fixed-priority
scheduling, by using exact schedulability tests and approximated schedulability tests. It may seem at first glance that
using exact schedulability tests is more precise, but the proof
shows that such exact tests are with the same speedup factors
as approximated tests. We show that the deadline-monotonic
1
partitioning algorithm has a speedup factor 3− M
when considering task systems with arbitrary deadlines. Such a factor holds
for polynomial-time schedulability tests and exponential-time
schedulability tests. Moreover, we also improve the speedup
factor to 2.84306 when considering constrained-deadline task
systems.
dbf (τ ,t)
However, proving that i=1 M t i ≤ f2 if t > 1 is pretty
complicated. The proof involves quite some mathematical
derivations, and is left in Appendix A. It should be clear that
we can ignore the M − 1 duplicated copies of the tasks by
considering only the tasks assigned on one processor (before
considering τN ). The complete proof is Appendix A. The
proof strategy first makes
an over-approximation, denoted as
P
dbf ] (t), of the sum τi ∈Tm dbf (τi , t) of the demand bound
functions at time t (after removing the M − 1 duplicated
copies). Based on such an over-approximation, it can be shown
]
that maxt>0 dbft (t) happens when t is a positive integer for
]
Tm . It can then be proved that the maximum dbft (t) happens
]
when t = 1 or t = ∞, in which both leads dbft (t) to f2 .
The speedup factor analyses in this paper are asymptotically tight if the analysis is oblivious of the particular
fitting strategy used. In all the tightness analyses, we only
take Lemma 1, which also implicitly implies the reference to
optimal global scheduling. The tasks are designed on purpose,
e.g., M 2 tasks (instead of M tasks) with period ∞ in the
proof of Theorem 9, to show that such factors also hold
(asymptotically or with minor changes) for optimal partitioned
scheduling. However, this does not limit the potential to
have better speedup factors by adopting better fixed-priority
scheduling strategies or more precise analysis for concrete
fitting strategies.
]
Figure 1 draws the functions dbf ] (t) and dbft (t) in the
]
range of [1, 6]. Figure 2 provides an illustrative view of dbft (t)
from t = 1, 2, . . . , 4000. It can also be easily shown by
assigning τN to one processor alone, we can find a task
partition that requires a speedup factor asymptotically equal
to f2 when M is sufficiently large.
10
Acknowledgements. This paper is supported by DFG, as part
of the Collaborative Research Center SFB876 (http://sfb876.tudortmund.de/).
[26]
[27]
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[28]
K. Albers and F. Slomka. An event stream driven approximation for
the analysis of real-time systems. In ECRTS, pages 187–195, 2004.
N. Audsley, A. Burns, M. Richardson, K. Tindell, and A. J. Wellings.
Applying new scheduling theory to static priority pre-emptive scheduling. Software Engineering Journal, 8(5):284–292, 1993.
S. Baruah. The partitioned EDF scheduling of sporadic task systems.
In Real-Time Systems Symposium (RTSS), pages 116 –125, 2011.
S. K. Baruah and N. Fisher. The partitioned multiprocessor scheduling
of sporadic task systems. In RTSS, pages 321–329, 2005.
S. K. Baruah and N. Fisher. The partitioned multiprocessor scheduling
of deadline-constrained sporadic task systems. IEEE Trans. Computers,
55(7):918–923, 2006.
S. K. Baruah, A. K. Mok, and L. E. Rosier. Preemptively scheduling
hard-real-time sporadic tasks on one processor. In IEEE Real-Time
Systems Symposium, pages 182–190, 1990.
E. Bini. The quadratic utilization upper bound for arbitrary deadline
real-time tasks. IEEE Trans. Computers, 64(2):593–599, 2015.
E. Bini, G. C. Buttazzo, and G. M. Buttazzo. Rate monotonic analysis:
the hyperbolic bound. Computers, IEEE Transactions on, 52(7):933–
942, 2003.
E. Bini, T. H. C. Nguyen, P. Richard, and S. K. Baruah. A responsetime bound in fixed-priority scheduling with arbitrary deadlines. IEEE
Trans. Computers, 58(2):279–286, 2009.
A. Burchard, J. Liebeherr, Y. Oh, and S. H. Son. New strategies for
assigning real-time tasks to multiprocessor systems. pages 1429–1442,
1995.
S. Chakraborty, S. Künzli, and L. Thiele. Approximate schedulability
analysis. In IEEE Real-Time Systems Symposium, pages 159–168, 2002.
J. Chen.
Partitioned multiprocessor fixed-priority scheduling of
sporadic real-time tasks. Computing Research Repository (CoRR),
abs/1505.04693, http://arxiv.org/abs/1505.04693, 2015.
J.-J. Chen and S. Chakraborty. Resource augmentation bounds for
approximate demand bound functions. In IEEE Real-Time Systems
Symposium, pages 272 – 281, 2011.
J.-J. Chen and S. Chakraborty. Partitioned packing and scheduling for
sporadic real-time tasks in identical multiprocessor systems. In ECRTS,
pages 24–33, 2012.
J.-J. Chen and S. Chakraborty. Resource augmentation for uniprocessor
and multiprocessor partitioned scheduling of sporadic real-time tasks.
Real-Time Systems, 49(4):475–516, 2013.
J.-J. Chen, W.-H. Huang, and C. Liu. k2U: A general framework from
k-point effective schedulability analysis to utilization-based tests. In
Real-Time Systems Symposium (RTSS), 2015.
R. Davis, T. Rothvo, S. Baruah, and A. Burns. Quantifying the suboptimality of uniprocessor fixed priority pre-emptive scheduling for
sporadic tasksets with arbitrary deadlines. In Real-Time and Network
Systems (RTNS), pages 23–31, 2009.
R. I. Davis and A. Burns. A survey of hard real-time scheduling for
multiprocessor systems. ACM Comput. Surv., 43(4):35, 2011.
R. I. Davis, T. Rothvoß, S. K. Baruah, and A. Burns. Exact quantification of the sub-optimality of uniprocessor fixed priority pre-emptive
scheduling. Real-Time Systems, 43(3):211–258, 2009.
F. Eisenbrand and T. Rothvoß. Static-priority real-time scheduling:
Response time computation is NP-Hard. In Proceedings of the 29th
IEEE Real-Time Systems Symposium, RTSS 2008, Barcelona, Spain, 30
November - 3 December 2008, pages 397–406, 2008.
N. Fisher and S. K. Baruah. A fully polynomial-time approximation
scheme for feasibility analysis in static-priority systems with arbitrary
relative deadlines. In ECRTS, pages 117–126, 2005.
N. Fisher, S. K. Baruah, and T. P. Baker. The partitioned scheduling
of sporadic tasks according to static-priorities. In 18th Euromicro
Conference on Real-Time Systems, ECRTS’06, 5-7 July 2006, Dresden,
Germany, Proceedings, pages 118–127, 2006.
R. L. Graham. Bounds on multiprocessing timing anomalies. SIAM
Journal of Applied Mathematics, 17(2):416–429, 1969.
D. S. Hochbaum and D. B. Shmoys. Using dual approximation
algorithms for scheduling problems theoretical and practical results. J.
ACM, 34(1):144–162, 1987.
A. Karrenbauer and T. Rothvoß. A 3/2-approximation algorithm for
rate-monotonic multiprocessor scheduling of implicit-deadline tasks. In
[29]
[30]
[31]
[32]
Workshop of Approximation and Online Algorithms WAOA, pages 166–
177, 2010.
C.-G. Lee, L. Sha, and A. Peddi. Enhanced utilization bounds for qos
management. IEEE Trans. Computers, 53(2):187–200, 2004.
J. P. Lehoczky. Fixed priority scheduling of periodic task sets with
arbitrary deadlines. In RTSS, pages 201–209, 1990.
J. P. Lehoczky, L. Sha, and Y. Ding. The rate monotonic scheduling
algorithm: Exact characterization and average case behavior. In IEEE
Real-Time Systems Symposium, pages 166–171, 1989.
J. Y.-T. Leung and J. Whitehead. On the complexity of fixed-priority
scheduling of periodic, real-time tasks. Perform. Eval., 2(4):237–250,
1982.
C. L. Liu and J. W. Layland. Scheduling algorithms for multiprogramming in a hard-real-time environment. Journal of the ACM, 20(1):46–
61, 1973.
A. K. Mok. Fundamental design problems of distributed systems for the
hard-real-time environment. Technical report, Cambridge, MA, USA,
1983.
G. von der Bruggen, J.-J. Chen, and W. Huang. Schedulability
and optimization analysis for non-preemptive static priority scheduling
based on task utilization and blocking factors. In ECRTS, pages 90–101,
2015.
Appendix A
Proof of Theorem
9. The remaining part of the proof is
PN −1
dbf (τ ,t)
to show that i=1 M t i ≤ f2 for all t ≥ 1 in the input
instance. Since each task has M − 1 duplicated copies with
the same task characteristics, we will implicitly drop the M −
1 duplicated copies by considering only these 2M tasks in
Tm . We index these tasks according to their periods in a nondecreasing order. Due to the setting of the task periods, we
know that Ti = f + Mi−1
−1 for i = 1, 2, . . . , M and Ti = ∞ for
i = M + 1, M + 2, . . . , 2M .
For the rest of the proof, we will only analyze the demand
bound function of these 2M
tasks on one processor, and our
P2M
dbf (τ ,t)
objective is to show that i=1 t i
≤ f2 for all t ≥ 1.
Note that this is identical to the original N − 1 tasks on
M processors. Moreover, for the rest of the proof, we only
consider the arrival pattern that defines the demand bound
function of the system from time 0 on the processor.
P2M
dbf (τ ,1)
For t = 1, we know that i=1 1 i = f2 . For t > 1, we
need to identify the demand of the M periodic tasks, whose
periods are not ∞, more precisely. Here, we define some terms
for the simplicity of the explanations. We define a pile of jobs
as follows: The `-th job of task τi is placed in the `-th pile.
Therefore, by the definition, each pile has M jobs and has
1−f
total execution time equal to 1 − f + M
−1 ≈ 1 − f since M
is sufficiently large. Precisely, the absolute deadline and the
arrival time of the j-th job (from the earliest arrival) in the `th pile are ` · (f + j − 1) and (` − 1) · (f + j − 1), respectively.
Therefore, the contribution of the jobs in the `-th pile to the
demand bound function at time t is at most
if t < ` · f
0
dbf`] (t) = (t − ` · f ) · 1`
(21)
if ` · f ≤ t < `
1−f
if t ≥ `
We can now define a safe upper bound dbf ] (t) of the
demand bound function of the 2M tasks for t ≥ 1 as follows:
]
dbf (t) = 1.5f − 1 +
∞
X
`=1
11
dbf`] (t).
(22)
]
` ≥ 3. Therefore, if dbft (t) is an increasing function in the first
segment, it must be also an increasing
in theosecond
n function
]
]
(`+1)
when
segment. Hence, dbf ] (t)/t ≤ max dbf` (`) , dbf`+1
` ≤ t < `+1 for the case when there are two linear segments of
dbf ] (t) in [`, `+1). The same argument holds for the case with
three segments as well by examining the increasing slopes of
dbf ] (t) of these three segments. As a result, we can conclude
]
that maxt≥1 dbft (t) happens when t is a postive integer.
Based on such an over-approximation, dbf ] (t) is a piece-wise
linear function, which is differentiable. Figure 1 illustrates the
]
above definition of dbf ] (t) and dbft (t) .
]
We now prove that maxt≥1 dbft (t) happens when t is
a postive integer. For any positive number `, the function dbf ] (t) for t hin the
h `j +k 1) can have eij kinterval [`,
`
ther two segments `, ( f + 1)f , ( f` + 1)f, ` + 1 or
h j k
hj k
j k
three segments `, ( f` + 1)f , ( f` + 1)f, ( f` + 2)f ,
hj k
( f` + 2)f, ` + 1 of linearly continuous functions. For example, in Figure 1a, in interval [2, 3), there are three segments
and, in interval [3, 4), there are only two segments.
The remaining part of the proof is to show that the
]
maximum dbft (t) happens when t = 1 or t = ∞, in which
]
]
both lead dbft (t) to f2 . We can easily evaluate dbft (t) for
t = 1, 2, 3, 4, 5 as follows:
The case when ` is 1 is clear, since dbf ] (t) = 1.5f −1+1−
f = 0.5f when 1 ≤ t < 2f and dbf ] (t) = 0.5f + (t − 2f )/2
when 2f ≤ t < 2. Therefore, dbf ] (t)/t ≤ dbf ] (1) when
1 ≤ t < 2. Similarly, dbf ] (t)/t ≤ dbf ] (3)/3 when 2 ≤ t < 3,
as also illustrated in Figure 1b.
•
•
•
•
•
For a given t ≥ j3, by
k the definition that f ≈ 0.7034674,
t
we also know that f ≥ btc + 1. Consider the case that
there are two linear segments of the function dbf ] (t) in
[`, ` + 1) for han integer
` ≥
two
j k
h3.j Ink this case, these
`
`
segments are `, ( f + 1)f , ( f + 1)f, ` + 1 . When
j k
` ≤ t ≤ ( f` + 1)f , we know that
dbf ] (1)
1
dbf ] (2)
2
dbf ] (3)
3
dbf ] (4)
4
dbf ] (5)
5
= 1.5f − 1 + 1 − f = 0.5f ≈ 0.3517337.
= 1.5f −1+2−2f
≈ 0.3241.
2
+(3−4f )/4
= 1.5f −1+3−3f
≈ 0.3304.
3
1.5f −1+4−4f +(4−5f )/5
=
≈ 0.3344.
4
1.5f −1+5−5f +(5−6f )/6+(5−7f )/7
=
≈ 0.3357.
5
For any positive integer ` with ` ≥ 5, we can reformulate
Eq. (22) into the following equation:
`
X
dbf ] (`) =1.5f − 1 +
b`/f c
(1 − f ) +
i=1
dbf ] (t) = 1.5f − 1 + `(1 − f ) +
t
bX
fc
(t − i · f ) ·
i=`+1
=1.5f − 1 + ` − b`/f c f +
(23)
We can complete the proof by showing that
dbf ] (`)
`
k
j k
`
f
t − ( f` + 1)f
X
1
j k
.
dbf ] (t) = dbf ] (`) + (t − `) ·
+
`
i
+1
j
k
` +1
X
f
`
`
1
]
= dbf (
+ 1)f + t − (
+ 1)f ·
.
f
f
i
i=`+1
(24)
−
=−
For a linear segment started from t∗ , we have
dbf ] (t∗ )+σx
t∗ +x
dbf ] (t∗ +x)
t∗ +x
=
−
if x is no more than the length of the linear
segment, where σ is the slope of the linear function defined
in Eq. (23) or Eq. (24). Since the first order derivative of
∗
] ∗
dbf ] (t∗ )+σx
(t )
with respect to x is σt (t−dbf
, we know that the
∗ +x)2
t∗ +x
]
dbf ] (`+1)
`+1
−
dbf ] (`)
dbf ] (` + 1)
−
`+1
`
1.5f − 1 + ` + 1 − b(` + 1)/f c f
1.5f − 1 + ` − b`/f c f
=
−
`+1
`
f
j
(25)
≥ 0 for any positive integer ` ≥ 5. Such a fact can be
seen in Figure 2 based on numerical evaluations, but a formal
proof requires quite some effort due to the floor function. For
any positive integer ` with ` ≥ 5, we have
j k
When ( f` + 1)f ≤ t < ` + 1, we know that
i=`+1
X `
i
i=`+1
i=`+1
i=`+1
b`/f c
1
i
`
bX
fc
1
= dbf ] (`) + (t − `) ·
.
i
X `−i·f
i
1
` `+1
`
b(`+1)/f c
+
X
(` + 1) 1i
`+1
i=b`/f c+1
1.5f − 1 − b`/f c f
f (b(` + 1)/f c − b`/f c)
−
`(` + 1)
`+1
1
+
`+1
b(`+1)/f c
X
i=b`/f c+1
1
i
(26)
Since ` ≥ 5, the above equation is well-defined. Due to the
fact that f ≈ 0.7034674, we knowj that
k b(` + 1)/f c is either
`
`
b`/f c + 1 or b`/f c + 2. Let f = f + b where 0 ≤ b < 1.
j k
j k
`
1
1
`
If 0 ≤ b < 2 − f1 , then `+1
<
+
2
−
+
=
f
f
f
f
f + 2,
which implies that b(`
j +k 1)/f c = b`/f c+1jfork such a case. If
1
`
1
1
2 − ≤ b < 1, then f` + 1 + f1 > `+1
f ≥ f +2− f + f =
j kf
`
f + 2, which implies that b(` + 1)/f c = b`/f c + 2 for such
a case. We now analyze these two cases individually.
∗
function dbft∗(t+x+x) monotonically decreases, or monotonically
increases, or remains the same with respect to valid x values.
If, for contradiction,
there exists
a t with ]` ≤ t < ` + 1 such
]
]
]
]
(`+1)
that dbft (t) > dbf` (`) and dbft (t) > dbf`+1
, then dbft (t)
]
in the first segment must be increasing and dbft (t) in the
second segment must be decreasing. However, with Eq. (23)
and Eq. (24), it is rather clear that the slope of dbf ] (t) of the
second linear segment is larger than the slope of the first linear
segment when there are two linear segments in [`, ` + 1) and
Case 1: 0 ≤ b < 2 −
12
1
f:
In this case, b(` + 1)/f c =
b`/f c + 1. Therefore, Eq. (26) becomes
Y1 (b) =
( f` − b)f − 1.5f + 1
`(` + 1)
−
f +1
+
`+1
`
f
1
. (27)
−b+1
We take the first order derivative of Y1 (b) with respect
−f
1
1 (b)
to b. Since dYdb
= `(`+1)
+ ( ` −b+1)
is an increas2
f
−f
ing function with respect to b and `(`+1)
+ ( `+11−1)2 =
f
2
(`+1)−f 2 +f (`+1)2 −f (`+1)
f −(`+1) +2f`(`+1)(`+1−f
<
0,
for
a given
)2
1
positive `, the function Y1 (b) is at least Y1 (2 − f ). Since
Y1 (2 −
1
` − 3.5f + 2 f + 1
)=
−
+
f
`(` + 1)
`+1
1
−1
`+1
f
f
−f (` + 1) − 2.5f + 2
+
`(` + 1)
`+1−f
(` + 1)(f 2 + 2 − 3.5f ) + 2.5f 2 − 2f
=
`(` + 1)(` + 1 − f )
>0,
[due to ` ≥ 5.]
=
we know that Y1 (b) > 0 for any integer ` ≥ 5 and 0 ≤ b <
]
]
(`+1)
2 − f1 . Therefore, dbf`+1
− dbf` (`) > 0 for such a case.
Case 2: 2 − f1 ≤ b < 1: In this case, b(` + 1)/f c =
b`/f c + 2. Therefore, Eq. (26) becomes
Y2 (b) =
` − bf − 1.5f + 1 2f + 1
−
+
`(` + 1)
`+1
`
f
1
+
−b+1
`
f
1
.
−b+2
(28)
We take the first order derivative of Y2 (b) with respect to b.
−f
1
1
2 (b)
Since dYdb
= `(`+1)
+ ( ` −b+1)
2 + ( ` −b+2)2 is an increasing
f
f
−f
1
function with respect to b, and `(`+1)
+ ( `+11−1)2 + ( `+1
>
2
f
f )
−(`+1)+2f `
−f
2
> 0, for a given ` ≥ 3, the
`(`+1) + ( `+1 )2 = f
`(`+1)2
f
function Y2 (b) is at least Y2 (2 − f1 ). Since
Y2 (2 −
` − 3.5f + 2 2f + 1
1
1
1
)=
−
+ `+1
+ `+1
f
`(` + 1)
`+1
f −1
f
1
=Y1 (2 − ) > 0,
[due to ` ≥ 5.]
f
we know that Y2 (b) > 0 for any integer ` ≥ 5 and 2 −
b < 1. Therefore,
]
dbf (`+1)
`+1
−
]
dbf (`)
`
1
f
≤
> 0 for such a case.
]
With the above two cases, we can conclude that dbf` (`) ≤
]
dbf (∞)
= f2 for any positive integer ` ≥ 5, which concludes
∞
the proof.
13
| 8cs.DS
|
A Learning and Masking Approach to Secure Learning
Linh Nguyen, Sky Wang, Arunesh Sinha
University of Michigan, Ann Arbor
{lvnguyen,skywang,arunesh}@umich.edu
arXiv:1709.04447v4 [cs.CR] 27 Jan 2018
ABSTRACT
Deep Neural Networks (DNNs) have been shown to be vulnerable
against adversarial examples, which are data points cleverly constructed to fool the classifier. Such attacks can be devastating in
practice, especially as DNNs are being applied to ever increasing
critical tasks like image recognition in autonomous driving. In this
paper, we introduce a new perspective on the problem. We do so by
first defining robustness of a classifier to adversarial exploitation.
Next, we show that the problem of adversarial example generation
can be posed as learning problem. We also categorize attacks in
literature into high and low perturbation attacks; well-known attacks like FGSM [11] and our attack produce higher perturbation
adversarial examples while the more potent but computationally inefficient Carlini-Wagner [5] (CW) attack is low perturbation. Next,
we show that the dual approach of the attack learning problem can
be used as a defensive technique that is effective against high perturbation attacks. Finally, we show that a classifier masking method
achieved by adding noise to the a neural network’s logit output
protects against low distortion attacks such as the CW attack. We
also show that both our learning and masking defense can work
simultaneously to protect against multiple attacks. We demonstrate
the efficacy of our techniques by experimenting with the MNIST
and CIFAR-10 datasets.
KEYWORDS
adversarial examples; robust learning
1
INTRODUCTION
Recent advances in deep learning have led to its wide adoption in
various challenging tasks such as image classification. However,
the current state of the art has been shown to be vulnerable to
adversarial examples, small perturbations of the original inputs,
often indistinguishable to a human, but carefully crafted to misguide the learning models into producing incorrect outputs. Recent
results have shown that generating these adversarial examples are
inexpensive [11]. Moreover, as safety critical applications such as
autonomous driving increasingly rely on these tasks, it is imperative that the learning models be reliable and secure against such
adversarial examples.
Prior work has yielded a lot of attack methods that generate
adversarial samples, and defense techniques that improve the accuracy on these samples (see related work for details). However,
defenses are often specific to certain attacks and cannot adaptively
defend against any future attack and some general defense techniques have been shown to be ineffective against more powerful
novel attacks. More generally, attacks and defenses have followed
the cat-and-mouse game that is typical of many security settings.
Further, traditional machine learning theory assumes a fixed stochastic environment hence accuracy in the traditional sense is not
enough to measure performance in presence of an adversarial agent.
In this paper, with the goal of generality, we pursue a principled approach to attacks and defense. Starting from a theoretical
robustness definition, we present a attack and a defense that learns
to generate adversarial examples against any given classifier and
learns to defend against any attack respectively. Based on formal
intuition, we categorize known attacks into high and low perturbation attacks. Our learning attack is a high perturbation attack
and analogously our learning defense technique defends against
high perturbation attack. For low perturbation attacks, we provide
a masking approach that defends against such attacks. Our two
defense techniques can be combined to defend against multiple
types of attacks. While our guiding principle is general, this paper
focuses on the specific domain of adversarial examples in image
classification.
Our first contribution is a definition of robustness of classifiers
in presence of an adversarial agent. Towards the definition, we
define the exploitable space by the adversary which includes data
points already mis-classified (errors) by any given classifier and
any data points that can be perturbed by the adversary to force
mis-classifications. Robustness is defined as the probability of data
points occurring in the exploitable space. We believe our definition
captures the essence of the multi-agent defender-adversary interaction, and is natural as our robustness is a strictly stronger concept
than accuracy. We also analyze why accuracy fails to measure robustness. The formal set-up also provides an intuition for all the
techniques proposed in this paper.
Our second contribution is an attack learning neural network
(ALN). ALN is motivated by the fact that adversarial examples
for a given classifier C are subsets of the input space that the C
mis-classifies. Thus, given a data distribution with data points x
and a classifier C trained on such data, we train a feed forward
neural network A with the goal of generating output points A(x)
in the mis-classified space. Towards this end, we re-purpose an
autoencoder to work as our ALN A with a special choice of loss
function that aims to make (1) the classifier C mis-classify A(x) and
(2) minimize the difference between x and A(x).
Our third contribution are two defense techniques: defense learning neural network (DLN) and noise augmented classifier (NAC).
Following the motivation and design of ALN, we motivate DLN
D as a neural network that, given any classifier C attacked by an
attack technique A, takes in an adversarial example A(x) and aims
to generate benign example D(A(x)) that does not lie in the misclassified space of C. The DLN is prepended to the classifier C acting
as a sanitizer for C. Again, similar to the ALN, we re-purpose an autoencoder with a special loss function suited for the goal of the DLN.
For non-adversarial inputs the DLN is encouraged to reproduce the
input as well as make the classifier predict correctly. We show that
DLN allows for attack and defense to be set up as a repeated competition leading to more robust classifiers. Next, while DLN works
efficiently for attacks that produces adversarial examples with high
perturbation, such as fast gradient sign method [11] (FGSM), it is
not practical for low perturbation attacks (discussed in details in
Section 3.4) such as Carlini-Wagner [5] (CW). For low perturbation
attacks, we present NAC which masks the classifier boundary by
adding a very small noise at the logits output of the neural network
classifier. The small noise added affects classification in rare cases,
thereby ensuring original accuracy is maintained, but also fools
low perturbation attacks as the attack is mislead by the incorrect
logits. DLN and NAC can work together to defend simultaneously
against both high and low perturbation attacks.
We tested our approach on two datasets: MNIST and CIFAR-10.
Our ALN based attack was able to attack all classifiers we considered
and achieve performance comparable to other high perturbation
attacks. Our defense approach made the resultant classifier robust to
the FGSM and CW. Detailed experiments are presented in Section 4
and 5. Missing proofs are in an online appendix1 (see footnote).
2
ATTACK MODEL
Given the adversarial setting, it is imperative to define the capabilities of the adversary, which we do in this section. First, we use
inference phase of a classifier to mean the stage when the classifier
is actually deployed as an application (after all training and testing is done). The attacker attacks only in the inference phase and
can channel his attack only through the inputs. In particular, the
attacker cannot change the classifier weights or inject any noise in
the hidden layers. The attacker has access to the classifier weights,
so that it can compute gradients if required. The attacker’s goal
is to produce adversarial data points that get mis-classified by the
classifier. These adversarial examples should be legitimate (that is
not a garbage noisy image) and the true class and the predicted class
of the data point could be additional constraints for the adversary.
3
APPROACH
This section formally describes our approach to the adversarial
example generation and defense problem using the notion of robustness we define. We start by defining basic notations. Let the
function C : X → Y denote a classifier that takes input data points
with feature values in X and outputs a label among the possible
k labels Y = {1, . . . , k}. Further, for neural networks based classifiers we can define Cp : X → ∆Y as the function that takes
in data and produces a probability distribution over labels. Thus,
C = max{Cp (x)}, where max provides the maximum component
of the vector Cp (x). Let sim(x, x ′ ) denote the dissimilarity between
Í
x and x ′ . Let H (p, q) denote the cross entropy − i pi log(qi ). In
particular, let H (p) denotes the entropy given by H (p, p). For this
paper, we assume X is the set of legitimate images (and not garbage
images or ambiguous images). Legitimate images are different for
different domains, e.g., they are digits for digit classification. Given
a label y, let Cat(y) denote the categorical probability distribution
with the component for y set to 1 and all else 0. Let opsim(y, y ′ )
denote the dissimilarity between output distributions y, y ′ ∈ ∆Y .
3.1
Robustness
We first introduce some concepts from PAC learning [1], in order
to present the formal results in this section. It is assumed that data
points arise from a fixed but unknown distribution P over X . We denote the probability mass over a set Z ⊂ X as P(Z ). A loss function
l(yx , C(x)) captures the loss of predicting C(x) when the true label
for x is yx . As we are focused on classification, we restrict ourselves
to the ideal 0/1 loss, that is, 1 for incorrect classification and 0 otherwise. A classifier C is chosen that minimizes the empirical loss over
Í
the n training data points ni=1 l(yx i , x i ). Given enough data, PAC
learning
theory guarantees that C also minimizes the expected loss
∫
l(y
,
C(x))P(x).
Given, 0/1 loss this quantity is just P(MC (X )),
x
X
where MC (X ) ⊂ X denote the region where the classifier C misclassifies. Accuracy for a classifier is then just 1 − P(MC (X )). In
this paper we will assume that the amount of data is always enough
to obtain low expected loss. Observe that a classifier can achieve
high accuracy (low expected loss) even though its predictions in
the low probability regions may be wrong.
All classifier families have a capacity that limits the complexity of
separators (hypothesis space) that they can model. A higher capacity classifier family can model more non-smooth separators2 . Previous work [11] has conjectured that adversarial examples abound
due to the low capacity of the classifier family used. See Figure 1A
for an illustration.
Adversarial exploitable space: Define EC,ϵ (X ) = MC (X ) ∪
{x | sim(x, MC (X )) ≤ ϵ }, where sim is a dissimilarity measure
that depends on the domain and sim(x, MC (X )) denotes the lowest
dissimilarity of x with any data point in MC (X ). For
classiqimage
Í
′ 2
(x
−
fication sim can just be the l 2 (Euclidean) distance:
i i xi )
where i indexes the pixels. EC,ϵ (X ) is the adversarial exploitable
space, as this space includes all points that are either mis-classified
or can be mis-classified by a minor ϵ-perturbation. Note that we
assume that any already present mis-classifications of the classifier
is exploitable by the adversary without the need of any perturbation. For example, if a stop sign image in a dataset is mis-classified
then an adversary can simply use this image as is to fool an autonomously driven vehicle.
Robustness: Robustness is simply defined as 1 − P(EC,ϵ (X )).
First, it is easy to see that robustness is a strictly stronger concept
than accuracy, that is, a classifier with high robustness has higher
accuracy. We believe this property makes our definition more natural than other current definitions. Further, another readily inferable
property from the definition of EC,ϵ that we utilize later is that a
classifier C ′ with MC ′ (X ) ⊂ MC (X ) is more robust than classifier C
in the same setting. We call a classifier C ′ perfect if the robustness
is 100%.
There are a number of subtle aspects of the definition that we
elaborate upon below:
• A 100% robust classifier can still have MC ′ (X ) , ϕ. This is
because robustness is still defined w.r.t. P, for example, large
compact regions of zero probability with small sub-region
of erroneous prediction far away from the boundary can
still make robustness 100%. However, MC ′ (X ) = ϕ provides
2 While
1 https://drive.google.com/open?id=1CBaHsU6IL9jQ4UN_2yYUteUsbLP2aKMk
capacity is defined for any function class [1] (includes deep neural networks),
the value is known only for simple classifiers like single layered neural networks.
A Learning and Masking Approach to Secure Learning
100% robustness for any P. Thus, robustness based on just
MC ′ (X ) and not P is a stronger but much more restrictive
concept of robustness than ours.
• A perfect classifier (100% robust) is practically impossible
due to large data requirement especially as the capacity of
the classifier family grows. As shown in Figure 1 low capacity classifiers cannot model complex separators, thus, large
capacity is required to achieve robustness. On the other hand,
classifiers families with large capacity but not enough data
tend to overfit the data [1]. Thus, there is a delicate balance between the capacity of the classifier family used and
amount of data available. The relation between amount of
data and capacity is not very well understood for Dep Neural
Networks. In any case, perfect robustness provides a goal
that robust classifiers should aim to achieve. In this paper, for
the purpose of defense, we seek to increase the robustness
of classifiers.
• Robustness in practice may apparently seem to be computable by calculating the accuracy for the test set and the
adversarially perturbed test set for any given dataset, which
we also do and has been done in all prior work. However,
this relies on the fact that the attack is all powerful, i.e., it can
attack all perturb-able points. It is easy to construct abstract
examples with probability measure zero mis-classification set
(single mis-classified point in a continuous Euclidean space)
that is computationally intractable for practical attacks to discover. A detailed analysis of computing robustness is beyond
the scope of this paper and is left for future work.
• The definition can be easily extended to weigh some kinds
of mis-classification more, if required. For example, predicting a malware as benign is more harmful than the opposite
erroneous prediction. For our focus area of image classification in this paper, researchers have generally considered all
mis-classification equally important. Also the sim function
in the definition is reasonably well agreed upon in literature on adversarial learning as the l 2 distance; however, we
show later in experiments that l 2 distance does not capture
similarity well enough. Instantiating the definition for other
domains such as malware classification requires exploring
sim further such as how to capture that two malwares are
functionally similar.
Lastly, compared to past work [9, 27], our robustness definition
has a clear relation to accuracy and not orthogonal to it. Also, our
definition uses the ideal 0/1 loss function rather than an approximate loss function l (often used in training due to smoothness) as
used in other definitions [7? ]. We posit that the 0/1 loss measures
robustness more precisely, as these other approaches have specified
the adversary goal as aiming to perturb in order to produce the
maximum loss within an ϵ ball B(x, ∫ϵ) of any given point x, with the
defender expected loss defined as X maxz ∈B(x,ϵ ) l(yx , C(z))P(x).
However, this means that even if the class is same throughout the
ϵ ball, with a varying l the adversary still conducts a supposed “attack” and increases loss for the defender without flipping labels. For
example, the well-known hinge loss varies rapidly within one of the
classes and such supposed attacks could lead to an overestimation
of the loss for defender and hence underestimate robustness. Further, use of an approximation in the definition allows an adversary
to bypass the definition by exploiting the approximation by l when
the true loss is 0/1. It is an interesting question for future on what
kind of approximations can help in computing robustness within
reasonable error bounds.
Finally, we analyze if accuracy is ever suitable to capture robustness. First, we make a few mild technical assumptions that there
exists a density p(x) for the data distribution P over X , X is a metric
space with metric d and vol(X ) = 1. We have the following result:
Theorem 3.1. 1 − a accuracy implies at least 1 − (a + ν + Kϵ/T )
robustness for any output C if
• For all x ∈ X , sim(x, x ′ ) ≥ Td(x, x ′ ) for some T > 0.
• MC (X ) lies in a low density region, that is, for all x ∈ MC (X )
we have p(x) ≤ ν for some small ν .
• p(x) is K-Lipschitz, that is, |p(x) − p(x ′ )| ≤ Kd(x, x ′ ) for all
x, x ′ ∈ X .
The first two conditions in the above result are quite natural. In
simple words, the first two conditions says dissimilarity increases
with distance (high T ) and the regions that the output classifier
predicts badly has low amount of data in the data-set (low ν ).
However, the final condition may not be satisfied in many natural
settings. This condition states that the data distribution must not
change abruptly (low K). This is required as the natural behavior
of most classifiers is to predict bad in a low data density region
and if this region is near a high data density region, the adversary
can successfully modify the data points in the high density region
causing loss of robustness. But in high dimensional spaces, data
distribution is quite likely to be not distributed smoothly with many
pockets or sub-spaces of zero density as pointed out in a recent
experimental work [? ]. Thus, data distribution is an important
contributing factor that determines robustness.
3.2
ALN
Our goal is to train a neural network ALN to produce samples in the
misclassification region of a given neural network based classifier.
The ALN acts on a data point x producing x ′ . Thus, we choose the
following loss function for the ALN that takes into account the
output for the given classifier:
αsim(x, x ′ ) − opsim(Cat(yx ), Cp (x ′ )) ,
The input dissimilarity term in the loss function aims to produce
data points x ′ that are similar to the original input x while the
output dissimilarity term aims to maximize the difference between
the true label of x and prediction of C on x ′ . The α is a weight that
is tuned through a simple search. Observe that this loss function is
general and can be used with any classifier (by inferring Cp from
C in case of specific non neural network based classifiers). For the
image classification problem we use the l 2 distance for sim. For
opsim we tried a number of functions, but the best performance
was for the l 1 loss ||Cat(yx ) − Cp (x ′ )||1 .
Note that an alternate loss function is possible that does not
use the actual label yx of x, rather using Cp (x). This would also
work assuming that the classifier is good; for poor classifiers a lot
of the data points are as it is mis-classified and hence adversarial
example generation is interesting only for good classifiers. Further,
Figure 1: Intuition behind ALN and DLN. (A) shows a linear
classifier (low capacity) is not able to accurately model a nonlinear boundary. (B) shows the ALN as the distribution mapping function F . (C) shows that DLN does the reverse mapping of ALN.
this choice would allow using unlabeled data for conducting such
an attack, making attack easier for an attacker. However, in our
experiments we use the more powerful attack using the labels.
Next, we provide a formal intuition of what ALN actually achieves.
Any adversarial example generation can be seen as a distribution
transformer F such that acting on the data distribution P the resultant distribution F (P) has support mostly limited to MC (X ). The
support may not completely limited to MC (X ) as the attacks are
never 100% effective. Also, attacks in literature aim to find points in
MC (X ) that are close to given images in the original dataset. ALN
is essentially a neural network representation of such a function
F against a given classifier C. See Figure 1B for an illustration. We
return to this interpretation in the next sub-section to provide a
formal intuition about the DLN defense.
Lastly, we show in our experiments that ALN produces adversarial examples whose perturbations are roughly of the same order
as the prior attack FGSM. We categorize these as high perturbation attacks. On the other the attack CW produces adversarial
perturbation with very small perturbations, we call such attacks
low perturbation attacks. As mentioned earlier, we provide two
separate defenses for these two types of attacks. Both these defense
can be used simultaneously to defend against both types of attacks.
3.3
DLN
Our first defense approach is to insert a neural network DLN D
between the input and classifier so that D sanitizes the input enabling the classifier to correctly classify the input. Each data point
for training DLN has three parts: x ′ is the image to sanitize, x is
the expected output and yx is the correct label of x. x are images
from the provided dataset, and there are two possibilities for x ′ :
(1) x ′ = x so that DLN attempts to satisfy C(D(x)) = yx , even if
C(x) , yx , and (2) x ′ = A(x) so that DLN undoes the attack and
make the classifier C correctly classify x ′ .
We formulate a loss function for DLN that, similar to ALN, has
two terms: sim(x, D(x ′ )) that aims to produce output D(x ′ ) close
to x and opsim(Cat(yx ), Cp (D(x ′ ))) that aims to make the classifier
output on D(x ′ ) be the same as yx . Thus, the loss function is
αsim(x, D(x ′ )) + opsim(Cat(yx ), Cp (D(x ′ ))) .
In this paper we only use α = 1. Note that the attack A is used as a
black box here to generate training data and is not a part of the loss
function. After training the DLN, our new classifier is C ′ which is
C prepended by the DLN. The working of DLN can be interpreted
as an inverse map F −1 for the mapping F induced by the attack
A. See Figure 1C for an illustration. For the image classification
problem we use the l 2 distance for sim. For opsim we tried a number
of functions, but the best performance was for the cross-entropy
loss H (Cat(yx ), Cp (D(x ′ )).
An important point to note is that the original classifier C is unchanged. What this ensures is the mis-classification space MC (X )
does not change and allows us to prove an important result about
C ′ under certain assumptions. For the sake of this result, we assume that attacks A generate adversarial examples in a sub region
MC,A (X ) ⊂ MC (X ). We also assume a good DLN D, that is, C(D(x))
is correct for a non-empty subset Z ⊂ MC,A (X ) and C(D(x)) continues to be correct for all x < MC (X ). Then, we prove
Lemma 3.2. Assuming MC,A (X ) ⊂ MC (X ), DLN is good as defined
above, and MC,A (X ) , ϕ, then MC ′ (X ) ⊂ MC (X )
Proof. Since DLN does not decrease the performance of C on
points outside MC (X ), C ′ ’s prediction on inputs outside MC (X ) is
correct, hence MC ′ (X ) ⊆ MC (X ). Any data point not mis-classified
by a classifier does not belong to its mis-classification space. Good
sanitization by DLN makes C ′ predict correctly on Z ⊂ MC,A (X ),
which makes MC,A (X ) ∩ MC ′ (X ) ⊂ MC,A (X ). Thus, we can claim
the result in the lemma statement.
□
While the above proof is under ideal assumptions, it provides
an intuition to how the defense works. Namely, the reduction in
the adversarial exploitable space makes the new classifier more
robust (see robustness properties earlier). This also motivates the
generalization of this technique to multiple attacks presented in
the next sub-section.
3.4
Repeated DLN Against Multiple Attacks
The above DLN can be naturally extended to multiple attacks, say
A1 , . . . , An . The only change required is to feed in all possible adversarial examples A1 (x)’s, . . . ,An (x)’s. It is straightforward to see that
under assumptions of Lemma 3.2 for all the attacks, the resultant
classifier C ′ has an adversarial example space MC ′ (X ) that removes
subsets of MC,Ai (X ) for all Ai ∈ A from MC (X ). This provides, at
least theoretically under ideal assumptions, a monotonic robustness improvement property with increasing number of attacks for
the DLN based approach. In fact, if all the attacks combined act as
a generator for all the points in MC (X ), then given enough data
and perfect sanitization the resultant classifier C ′ tends towards
achieving MC ′ (X ) = ϕ which essentially would make C ′ a perfect
classifier. Perfect classifiers have no adversarial examples.
However, attacks rarely explore all of the mis-classified space,
which is why new attacks have defeated prior defense techniques.
Even for our approach, attacks successfully work against the DLN
that has been trained only once (accuracy numbers are in Experiments). However, DLN allows for easy retraining (without retraining the classifier) as follows: repeatedly attack and re-learn a DLN
in rounds, that is, conduct an attack on the classifier obtained in
every round and train a DLN in a round using the attacked training
data from all the previous rounds and the original training data.
More formally, at round i our training data consists of i copies of
A Learning and Masking Approach to Secure Learning
fitting problems could arise, which needs to be taken care of in
practice. Also, the training become more expensive over rounds
with increasing data size. In particular, low perturbation attacks are
not defeated with few rounds. We do observe improvement with
the low perturbation CW attack over rounds, but the improvement
is very small, as represented visually in Figure 2. The main reason is
that low perturbation attacks only exposes a very small volume of
misclassified space, thus, it would require a huge number of rounds
for repeated DLN to reduce the mis-classified space to such a small
volume that cannot be attacked. This motivates our next approach
of noise augmented classifier.
3.5
Figure 2: Intuition behind working of repeated DLN against
high and low perturbation attacks. (A),(B) shows a high perturbation attack causes a faster improvement in resultant
classifier. Further, beyond some rounds the attack does not
work as it can only find adversarial examples with high perturbation. (C),(D) shows a low perturbation attack causes a
slow improvement in resultant classifier.
original training data and i instances of attacked training data from
previous rounds. Observe that we add copies of the original training
data in each round, this is because the adversarial data swamps out
the original training data and accuracy suffers in regions where the
original training data is distributed. See Figure 2 for an illustration
of how repeated DLN works.
The following result provides formal intuition for this approach:
Lemma 3.3. Assume the following conditions hold for every round
i: MCi −1,A (X ) ⊂ MCi −1 (X ) and the DLN D i has good memory, which
means that given there exists a largest set Z i ⊂ MCi −1,A (X ) which the
DLN D i correctly sanitizes so that C(D i (x)) is correct for all x ∈ Z i
then Z i−1 ⊂ Z i . That is DLN D i can correctly sanitize data points that
the previous round DLN did plus possibly more data points. Further,
C(D i (x)) continues to be correct for all x < MC (X ). Then the classifier
Cn after n rounds satisfies MCn (X ) ⊂ MCn−1 (X ).
Proof. Arguing similarly to Lemma 3.2 we can show that MCn (X ) ⊆
MC (X ) due to the correct classification outside of MC (X ). Further,
it is easily inferable that MCn (X ) = MC (X )\Z n given Z n is a subset
of MC (X ) and given the largest such set condition on Z i . Then, the
good memory property leads to the required result.
□
The attack-defense competition technique is somewhat akin
to GANs [10]. However, there is a big difference, since in every
round the dataset used to train the DLN grows. Practically, this
requires DLN to have a large capacity in order to be effective; also
depending on the capacity and the size of dataset over or under
NAC
Figure 2 also provides a hint on how to overcome low perturbation
attacks. In order to achieve low perturbation, such attacks rely a
lot on the exact classifier boundary. Thus, masking the classifier
boundary can fool low perturbation attacks. We achieve this by
adding a small noise to the logits of the neural network calling the
resultant classifier a noise augmented classifier (NAC). This noise
should be small enough that it does not affect the classification
of original data points by much, but is able to mis-lead the low
perturbation attack. Also, following our intuition NAC should not
provide any defense against high perturbation attacks, which we
indeed observe in our experiments. However, observe that DLN
and NAC can be used simultaneously, thus, providing benefits of
both defense which we show in our experiments.
Further, a natural idea to bypass the defense provided by NAC is
to take the average of multiple logit outputs for the same given input
image (to cancel the randomness) and then use the average logits
as the logits required for the CW attack. We show experimentally
that this idea does not work effectively even after averaging over
many logit outputs.
4
EXPERIMENTS FOR ATTACKS
All our experiments, including for DLN and NAC, were conducted
using the Keras framework on a NVIDIA K40 GPU. The learning
problem can be solved using any gradient-based optimizer. In our
case, we used Adam with learning rate 0.0002. We use two wellknown datasets: MNIST digits and CIFAR-10 colored images.
We consider two classifiers one for MNIST and one for CIFAR-10:
we call them C M and CC . These classifiers are variants of wellknown architectures that achieved state-of-the-art performances
on their respective datasets. As stated earlier, we consider three
attacks: ALN, FGSM and CW. CW has been referred to in the literature [28] as one of the best attacks till date (at the time of writing
of this paper), while FGSM runs extremely fast. For the autoencoder we use a fourteen hidden layer convolutional architecture.
Our code is publicly available, but the github link is elided in this
submitted version for the review process. For our experiments we
pre-process all the images so that the pixels values lie between
[−0.5, 0.5], so all components (attacks, autoencoders, classifiers)
work in space [−0.5, 0.5]. We use FGSM with values of 0.03 and
0.01 for its parameter ϵ on MNIST and CIFAR, respectively.
Observe that all these attacks work against a given classifier
C, thus, we use the notation A(C, .) to denote the attack A acting
on an image x to produce the adversarial example A(C, x) (A can
Figure 3: Targeted attacks by ALN: target class on bottom
Accuracy
Distortion
OT D
FGSM(C M , OT D)
CW(C M , OT D)
ALNU(C M , OT D)
ALNT(C M , OT D)
99.45 %
0.72 %
0.03 %
1.65 %
0.0 %
−
14.99
1.51
4.43
4.34
Distortion
OT D
FGSM(CC , OT D)
CW(CC , OT D)
ALNU(CC , OT D)
ALNT(CC , OT D)
84.59 %
4.21 %
0%
6.16 %
8.21 %
−
10.03
0.18
2.57
3.01
Figure 4: Untargeted attack by ALN for CIFAR-10
Figure 5: Attack using FGSM for CIFAR-10
Table 1: Attacks on MNIST Dataset
5
Table 2 shows the result of untargeted attacks using ALN, FGSM
and CW on the CIFAR-10 dataset. We can see that ALN, just like
FGSM, produces slightly higher adversarial accuracy for MNIST,
but the distortion of FGSM is much higher. This does result in
a large difference in visual quality of the adversarial examples
produced—see Figure 4 for randomly chosen 25 perturbed images
using ALN and Figure 5 for randomly chosen 25 perturbed images
using FGSM. Also, we can attribute the higher distortion of ALN
for CIFAR as compared to MNIST partially to the CIFAR being a
higher dimensional space problems and the same capacity of the
autoencoder we used for CIFAR and MNIST.
Accuracy
Table 2: Attacks on CIFAR-10 Dataset
be any of the three attacks). A(C, Z ) denotes the set of adversarial
examples {A(C, x) | x ∈ Z }. We report accuracies on various test
sets: (1) original test dataset (OT D): these images are the original
test dataset from the dataset under consideration, (2) A(C, OT D) is
the adversarially perturbed dataset using attack A against classifier C, for example, this could be FGSM(C M , OT D). We also report
distortion numbers as has been done in literature [5]. Distortion
measures how much perturbation on average was added by an
attack and is meant to capture how visually similar the image is to
the original image. Distortion is measured as the average over all
test images of the l 2 distance between the original and perturbed
image.
Results: Untargeted attacks refers to attack that aim to produce
mis-classification but not with the goal of making the classifier output a specific label. Targeted attacks aim to produce an adversarial
example that gets classified as a given class y. It is also possible
to modify ALN to perform targeted attacks. This is achieved by
modifying the loss function to use a positive opsim term, like the
DLN loss function, but using the target class label y instead of the
original class label yx in the opsim term. Then, we can perform an
ALN attack in two ways: ALNU uses the ALN loss function as stated
and ALNT constructs a targeted attack per class label differing from
the original label and chooses the one with least distortion. Figure 3
shows an example of targeted attack with different target labels.
Table 1 shows this approach for MNIST with the targeted ALNT
version performance better than other attacks.
Test data type
Test data type
EXPERIMENTS FOR DEFENSE
For defense, we denote the new classifier using the DLN trained
′ (for MNIST)
against any attack A or NAC modified classifier as C M
′
or CC (for CIFAR). Also, we test accuracies on test data adversarially perturbed by attacks against the new classifiers, for example,
following our convention one such dataset would be denoted as
′ , OT D). For defense we focus on defending against known
A(C M
attacks in literature.
5.1
DLN Defense Against Single Attack
Table 3 shows the results when the DLN is trained to defense against
′ .
specific attacks using MNIST dataset to yield a new classifier C M
A Learning and Masking Approach to Secure Learning
Along expected lines, the accuracy on OT D drops slightly for all the
cases where DLN is trained against FGSM or CW. However, the new
′ is able to perform quite good on adversarial examples
classifier C M
that were produced by any attack on the original classifier C M , for
′ gets an accuracy of 95% of the adversarial examples
example, C M
CW(C M , OT D) that was produced by CW attacking C M . Lastly,
′ is still not resilient to
when attacked again the new classifier C M
′ , OT D) and
attacks as shown by the low accuracies for FGSM(C M
′
CW(C M , OT D).
One number that stands out is the success of the new classifier
′ in correctly classifying the adversarial example generated by
CM
CW for the original classifier C M . This supports our hypothesis
that CW is very sensitive to the exact classifier boundary, and a
newer classifier with a slightly different boundary is able to defeat
′
prior adversarial examples. Of course, CW is is able to attack C M
against quite successfully, which we later show can be defended
by our NAC approach. For FGSM, we show in later sections that
the performance of the classifier greatly improves when DLN is
repeatedly trained against FGSM—revealing that the DLN approach
is flexible enough to keep on improving its performance against
high perturbation attacks.
DLN
Trained
FGSM
FGSM
FGSM
CW
CW
CW
Test data type
OT D
FGSM(C M , OT D)
′ , OT D)
FGSM(C M
OT D
CW(C M , OT D)
′ , OT D)
CW(C M
Accuracy
96.77 %
88.5 %
13.75 %
98.6 %
95.42 %
0.14 %
−
4.55
6.98
−
5.77
3.5
Repeated DLN
In this section, we run DLN repeatedly as described earlier in Section 3.4. We cut off the experiments when a single round took more
than 24 hours to solve. We show the results for MNIST in Table 4
showing a clearly increasing trend in accuracy on adversarial examples produced by FGSM attacking the newer classifier, revealing
increasing robustness. For CIFAR, the approach becomes too computationally expensive within two rounds. Thus, while the DLN
approach is promising, as stated earlier it is computationally expensive, more so with larger and complex data sets. Further, as stated
earlier running DLN against low perturbation attacks like CW does
not show much improvement. However, we tackle that next using
the NAC defense approach.
5.3
Acc. OT D
Acc.
FGSM(Ci , OT D)
Distortion
0
1
2
3
4
5
99.36 %
97.70 %
97.61 %
97.95 %
97.79 %
97.77 %
0.72 %
13.70 %
24.86 %
43.39 %
52.88 %
56.57 %
14.99
13.63
14.58
14.73
14.57
14.52
Table 4: Classifier trained repeatedly against FGSM for
MNIST
unchanged. This can be seen in Table 5. The second line in that
table shows that the CW attack completely fails as the accuracy on
the adversarial examples in 93%. However, it can also be observed
that the new classifier is not resilient to attack by FGSM, as shown
by the third line in that table. This follows the intuition we provided
in Figure 2. For CIFAR, Table 6 shows that NAC is able to overcome
CW to a large extent.
Attack
Test data type
Accuracy
Distortion
CW
FGSM
OT D
′ , OT D)
CW(C M
′ , OT D)
FGSM(C M
99.36 %
93.60 %
0.74 %
1.49
14.99
Distortion
′ for MNIST
Table 3: New DLN prepended classifier C M
5.2
Round
NAC defense
Recall that the NAC defense works by adding noise to the logits
layer of the neural network classifier to produce a new classifier
′ for MNIST and C ′ for CIFAR. We use Gaussian noise with 0
CM
C
mean and variance 1. In this section, we show that the NAC defense
is able to produce classifiers that are resilient to CW attack. Further,
the new classifier’s accuracy on the original test data-set is nearly
′ for MNIST
Table 5: Accuracy of NAC Classifier C M
Attack
Test data type
Accuracy
Distortion
CW
FGSM
OT D
′ , OT D)
CW(C M
′ , OT D)
FGSM(C M
84.67 %
77.70 %
4.19 %
0.17
10.04
′ for CIFAR-10
Table 6: Accuracy of NAC Classifier C M
As stated earlier, a natural idea to attack NAC would be to query
an image n times and then average the logits before using it for the
CW attack. This augmented attack does make CW more effective
but not by much. Table 7 shows that the accuracy on the adversarial
′ remains high. Moreover, more queries
example generated for C M
make it more difficult to conduct the CW attack in practice (as
the adversary may be query limited), while also causing a small
increase (1% with 5000 sample) in the already high runtime of CW.
n
Adversarial accuracy
Distortion
500
5000
95.14 %
82.07 %
1.51
1.51
′ against improved
Table 7: Accuracy of NAC Classifier C M
CW for MNIST
5.4
Defense Against Multiple Attacks
Finally, we show that DLN and NAC can work together. We show
this by presenting the accuracy on the adversarial example generated in each round of DLN repetition when the classifier Ci after
each round is augmented with NAC and attacked by FGSM and
CW both. See Table 8. One observation is that NAC’s performance
decreases slightly over rounds stabilizing at 79%, while the accuracy for original test set and FGSM perturbed test set stays almost
exactly same as Table 4.
Round
Acc. OT D
Acc.
FGSM(Ci , OT D)
Acc.
CW(Ci , OT D)
0
1
2
3
4
5
99.36 %
97.70 %
97.73 %
97.60 %
97.64 %
97.73 %
0.72 %
13.72 %
24.28 %
43.20 %
53.17 %
56.45 %
94.2
93.7
84.7
83.3
79
79.3
Table 8: Classifier trained repeatedly against FGSM for
MNIST and augmented with NAC in each round
6
RELATED WORK
A thorough survey of security issues in machine learning is present
in surveys [25] and some of the first results appeared in [8, 19].
Here we discuss the most closely related work.
Attacks: Most previous attack work focuses on adversarial examples for computer vision tasks. Multiple techniques to create such
adversarial examples have been developed recently. Broadly, such attacks can be categorized as either using costs gradients [3, 11, 14, 20]
or the forward gradient of the neural network [24] and perturbing
along most promising direction or directly solving an optimization
problem (possibly using gradient ascent/descent) to find a perturbation [5, 21]. In addition, adversarial examples have been shown
to transfer between different network architectures, and networks
trained on disjoint subsets of data [22, 26]. Adversarial examples
have also been shown to translate to the real world [15], that is,
adversarial images can remain adversarial even after being printed
and recaptured with a cell phone camera. Attacks on non-neural
networks have also been explored in literature [3]. Our approach
is distinctly different from all these approaches as we pose the
problem of generating adversarial samples as a generative learning problem, and demonstrate generation of adversarial examples
given access to any given classifier. Our approach also applies to
any classifier that output class probabilities and not just neural
networks.
Defense: Also, referred to as robust classification in many papers, defense techniques can be roughly categorized into techniques
that do (1) adversarial (re)training, which is adding back adversarial
examples to the training data and retraining the classifier, often repeatedly [17], or modifying loss function to account for attacks [14];
(2) gradient masking, which targets that gradient based attacks by
trying to make the gradient less informative [23]; (3) input modification, which are techniques that modify (typically lower the
dimension) the feature space of the input data to make crafting
adversarial examples difficult [28]; (4) game-theoretic formulation,
which modifies the loss minimization problem as a constrained
optimization with constraints provided by adversarial utility in performing perturbations [16], and (5) filtering and de-noising, which
aims to detect/filter or de-noise adversarial examples (cited below).
Our defense approach differs from the first four kinds of defense
as our DLN approach never modify the classifier or inputs but add a
sanitizer (DLN) before the classifier. First, this increases the capacity
of the resultant classifier C ′ , so that it can model more complex
separators, which is not achieved when the classifier family stays
the same. Further, our defense is agnostic to the type of attack and
does not utilize properties of specific types of attacks. Interestingly,
the DLN approach can be used with any classifier that output class
probabilities and not just neural networks. Further, NAC is a very
minor modification to the classifier that, distinct from other randomized approaches [? ] that randomize over multiple classifiers,
aims to mask the classifier boundary. Also, NAC can work with
other defenses unlike techniques that modify inputs to try and
defend against CW [28].
More closely related to our work are some defense techniques
that have focused on detecting and filtering out adversarial samples [12, 18] or de-noising input [13]; here the filter or de-noiser
with the classifier could be considered as a larger neural network.
However, unlike these work, our goal for DLN is targeted sanitization. Moreover, recent attack work [4] have produced attack
techniques to defeat many known detection techniques. Our technique provides the flexibility to be resilient against more powerful
attacks by training the DLN with such an attack for high perturbation attacks or using NAC for low perturbation attacks.
Lastly, two concurrent unpublished drafts (available online) have
independently and simultaneously proposed an attack [2] similar to
ALN and a defense [6] apparently similar to DLN. The difference
for the attack work is in using the class label vs classifier output
in opsim term for the attack. For the defense work, we differ as we
show how DLN technique extends to multiple attacks and can be
repeatedly used in an attack-defense competition. Moreover, unlike
these drafts, we provide another defense technique NAC that works
against CW, define robustness and show that our defense techniques
approximately aims to achieve our definition of robustness. Further,
our formal reasoning reveals the underlying nature of attacks and
defenses.
7
CONCLUSION AND FUTURE WORK
Our work provides a new learning perspective of the adversarial
examples generation and defense problems with a formal intuition
of how these approaches work, using which we were able to defend
against multiple attacks including the potent CW. Further, unlike
past work, our defense technique does not claim to a catchall or
specific to any attack; in fact, it is flexible enough to possibly defend against any attack. Posing the attack and defense as learning
problems allows for the possibility of using the rapidly developing
research in machine learning itself to make the defense more effective in future, for example, by using a different specialized neural
A Learning and Masking Approach to Secure Learning
network architecture rather than an autoencoder. A number of variations of our theory and and tuning of the application framework
provides rich content for future work.
REFERENCES
[1] Martin Anthony and Peter L. Bartlett. 2009. Neural Network Learning: Theoretical
Foundations (1st ed.). Cambridge University Press, New York, NY, USA.
[2] Shumeet Baluja and Ian Fischer. 2017. Adversarial Transformation Networks:
Learning to Generate Adversarial Examples. CoRR abs/1703.09387 (2017). http:
//arxiv.org/abs/1703.09387
[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić,
Pavel Laskov, Giorgio Giacinto, and Fabio Roli. 2013. Evasion attacks against
machine learning at test time. In Joint European Conference on Machine Learning
and Knowledge Discovery in Databases. Springer, 387–402.
[4] Nicholas Carlini and David Wagner. 2017. Adversarial Examples Are Not Easily
Detected: Bypassing Ten Detection Methods. arXiv preprint arXiv:1705.07263
(2017).
[5] Nicholas Carlini and David Wagner. 2017. Towards evaluating the robustness
of neural networks. In Security and Privacy (SP), 2017 IEEE Symposium on. IEEE,
39–57.
[6] Xinyun Chen, Bo Li, and Yevgeniy Vorobeychik. 2016. Evaluation of Defensive
Methods for DNNs against Multiple Adversarial Evasion Models. (2016). https:
//openreview.net/forum?id=ByToKu9ll¬eId=ByToKu9ll.
[7] Moustapha Cisse, Piotr Bojanowski, Edouard Grave, Yann Dauphin, and Nicolas
Usunier. 2017. Parseval Networks: Improving Robustness to Adversarial Examples.
arXiv preprint arXiv:1704.08847 (2017).
[8] Nilesh Dalvi, Pedro Domingos, Sumit Sanghai, Deepak Verma, et al. 2004. Adversarial classification. In Proceedings of the tenth ACM SIGKDD international
conference on Knowledge discovery and data mining. ACM, 99–108.
[9] Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. 2015. Fundamental limits on
adversarial robustness. In Proc. ICML, Workshop on Deep Learning.
[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial
nets. In Advances in neural information processing systems. 2672–2680.
[11] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014. Explaining and
Harnessing Adversarial Examples. CoRR abs/1412.6572 (2014). http://arxiv.org/
abs/1412.6572
[12] Kathrin Grosse, Praveen Manoharan, Nicolas Papernot, Michael Backes, and
Patrick McDaniel. 2017. On the (Statistical) Detection of Adversarial Examples.
arXiv preprint arXiv:1702.06280 (2017).
[13] Shixiang Gu and Luca Rigazio. 2014. Towards deep neural network architectures
robust to adversarial examples. arXiv preprint arXiv:1412.5068 (2014).
[14] Ruitong Huang, Bing Xu, Dale Schuurmans, and Csaba Szepesvári. 2015. Learning
with a strong adversary. arXiv preprint arXiv:1511.03034 (2015).
[15] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. 2016. Adversarial examples
in the physical world. arXiv preprint arXiv:1607.02533 (2016).
[16] Bo Li and Yevgeniy Vorobeychik. 2014. Feature cross-substitution in adversarial
classification. In Advances in neural information processing systems. 2087–2095.
[17] Bo Li, Yevgeniy Vorobeychik, and Xinyun Chen. 2016. A General Retraining
Framework for Scalable Adversarial Classification. arXiv preprint arXiv:1604.02606
(2016).
[18] Xin Li and Fuxin Li. 2016. Adversarial Examples Detection in Deep Networks
with Convolutional Filter Statistics. arXiv preprint arXiv:1612.07767 (2016).
[19] Daniel Lowd and Christopher Meek. 2005. Adversarial learning. In Proceedings
of the eleventh ACM SIGKDD international conference on Knowledge discovery in
data mining. ACM, 641–647.
[20] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. 2015.
DeepFool: a simple and accurate method to fool deep neural networks. CoRR
abs/1511.04599 (2015). http://arxiv.org/abs/1511.04599
[21] Seyed Mohsen Moosavi Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal
Frossard. 2017. Universal adversarial perturbations. In Proceedings of 2017 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR).
[22] Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. 2016. Transferability
in machine learning: from phenomena to black-box attacks using adversarial
samples. arXiv preprint arXiv:1605.07277 (2016).
[23] Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik,
and Ananthram Swami. 2016. Practical black-box attacks against deep learning
systems using adversarial examples. arXiv preprint arXiv:1602.02697 (2016).
[24] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik,
and Ananthram Swami. 2016. The limitations of deep learning in adversarial
settings. In Security and Privacy (EuroS&P), 2016 IEEE European Symposium on.
IEEE, 372–387.
[25] Nicolas Papernot, Patrick McDaniel, Arunesh Sinha, and Michael Wellman. 2016.
Towards the Science of Security and Privacy in Machine Learning. arXiv preprint
arXiv:1611.03814 (2016).
[26] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan,
Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks.
arXiv preprint arXiv:1312.6199 (2013).
[27] Beilun Wang, Ji Gao, and Yanjun Qi. 2016. A Theoretical Framework for Robustness of (Deep) Classifiers Under Adversarial Noise. arXiv preprint arXiv:1612.00334
(2016).
[28] Weilin Xu, David Evans, and Yanjun Qi. 2017. Feature Squeezing: Detecting
Adversarial Examples in Deep Neural Networks. arXiv preprint arXiv:1704.01155
(2017).
| 1cs.CV
|
arXiv:1304.7526v2 [math.AC] 9 Sep 2013
Factorization of integer-valued polynomials with square-free
denominator
Giulio Peruginelli
April 9, 2018
Dedicated to Marco Fontana on the occasion of his 65th birthday
Abstract
We describe an algorithm to compute the different factorizations of a given image
primitive integer-valued polynomial f (X) = g(X)/d ∈ Q[X], where g ∈ Z[X] and
d ∈ N is square-free, assuming that the factorizations of g(X) in Z[X] and d in Z are
known. We translate this problem into a combinatorial one.
Keywords: Integer-valued polynomial, Factorization, Fixed divisor, Primary components.
MSC Classification codes: 13B25, 13F20, 13P05.
1
Introduction
It is well known that the ring of integer-valued polynomials Int(Z) = {f ∈ Q[X] | f (Z) ⊂ Z}
is far from being a unique factorization domain. We know that the ring Int(Z) is atomic
(every non-zero non-unit of Int(Z) admits a factorization into irreducibles) and every nonzero non-unit has only has only finitely many factorizations into irreducibles (see [7]). In
particular, this implies that Int(Z) is a bounded factorization domain (the length of the
different factorizations of a given element is bounded, see [2, Prop. VI.3.2]). Moreover,
in [3] it is shown that the ring has infinite elasticity, where the elasticity of a domain is
defined as the supremum of the set of ratios between length of factorizations of non-zero
non-units. We recall that the length of a factorization is the number of irreducible elements
which appear in the factorization itself. Two factorizations into irreducibles of an element
x in a commutative ring R, say x = r1 · . . . · rn and x = s1 · . . . · sm , are essentially the
same if n = m and after possibly re-indexing, ri is associated to si , for i = 1, . . . , n (that
is, there exists a unit ui ∈ R such that ri = ui si ). Otherwise the two factorizations are
essentially different (see [7]).
1
More recently, in [7] the following result is proved. Given a finite set S = {n1 , . . . , nr }
of (non necessarily distinct) positive integers greater than 1, there exists an integervalued polynomial f (X) with r essentially distinct factorizations into irreducibles of length
n1 , . . . , nr , respectively. Hence, there are elements in the ring of integer-valued polynomials
which admit distinct factorizations into irreducibles of arbitrary lengths.
We propose here a new method to describe the essentially different factorizations into
irreducibles of a given integer-valued polynomial, under the assumption that the denominator is square-free (we will treat the general case in a future work). We remark that in all
the examples produced in [7] to exhibit polynomials with prescribed sets of lengths, only
polynomials with square-free denominator appear (in [7, Thm. 10] there is a polynomial
with more than one prime in the denominator, in all the other results there is just one
prime factor in the denominator). So, a treatment of this case has a certain interest. We
begin by recalling some classical definitions.
P
Definition 1.1. The content of a polynomial g(X) = k=0,...,n ak X k ∈ Z[X] is defined
as the g.c.d. of its coefficients ak . We denote the content of g(X) by c(g). A polynomial
g ∈ Z[X] is called primitive if its content is equal to 1. Given f ∈ Int(Z), we denote
by d(f ) the fixed divisor of f , that is the g.c.d. of the set of values {f (n) | n ∈ Z}. An
integer-valued polynomial f (X) is said to be image primitive if its fixed divisor is equal
to 1. Let p ∈ Z be a prime. If g ∈ Z[X] we say that g(X) is p-primitive if p does not divide
c(g), that is, at least one of the coefficient of g(X) is not divisible by p. If f ∈ Int(Z), we
say that f (X) is p-image primitive if p does not divide d(f ).
Given a polynomial g ∈ Z[X], the content of g(X) is in general a proper divisor of
the fixed divisor of g(X): consider for example g(X) = X(X − 1) which is primitive but
its fixed divisor is equal to 2. Already in [4] it is shown the important role played by the
fixed divisor in the study of the factorizations of an integer-valued polynomial (see the
results that we recall below). For example, the polynomial g(X) = X 2 + X + 2, which is
irreducible in Z[X] (and consequently in Q[X] by Gauss Lemma), has fixed divisor equal
to 2, so that in Int(Z) we have the non trivial factorization g(X) = 2 · g(X)
2 .
We recall the following facts:
- Int(Z), Z[X] and Z share the same group of units, {±1} ([3, Lemma 1.1]).
- An irreducible integer p stays irreducible in Int(Z) ([2, Lemma VI.3.1]).
- Int(Z) has no prime elements ([1, Prop. 3.2]).
- If g ∈ Z[X] is p-primitive for some prime p and p divides the fixed divisor of g(X),
then p ≤ deg(g) (this is due to Polya, see [8, Thm. 3.1] for a modern treatment).
- Let g ∈ Z[X] be primitive. Then g(X) is irreducible in Int(Z) if and only if it is
image primitive and irreducible in Z[X] (Chapman-McClain [4, Thm. 2.6]). Hence,
2
an irreducible factor in Int(Z) of an irreducible polynomial g ∈ Z[X] is either a
constant c which is a divisor of d(g) or g(X)
d(g) .
- Given two integer-valued polynomials f and g, we have d(f g) ⊂ d(f )d(g) and in
general we may not have an equality (see the above example X(X − 1)). This is
the main difference between fixed divisor and content: in fact, the content of the
product of two polynomials g1 (X) and g2 (X) is equal to the product of the contents
of g1 (X) and g2 (X) by Gauss Lemma. This is equivalent to the fact that a primitive
polynomial g ∈ Z[X] is irreducible if and only if it is irreducible in Q[X]; this sentence
is no more true if we substitute the ring Q[X] with Int(Z) (see the above example
g(X) = X 2 + X + 2). By the above cited theorem of Chapman-McClain, we have to
add the assumption that g(X) is also image primitive. We notice that a factor of an
image primitive polynomial is image primitive ([4]).
Given a polynomial f ∈ Q[X], we have f (X) = g(X)/d, for some uniquely determined
g ∈ Z[X] and d ∈ N such that (d, c(g)) = 1 (we essentially use the fact that Z is UFD).
For short, we call d the denominator of f (X) and g(X) the numerator of f (X).
We can further express f (X) in the following way:
Q
ei
g(X)
i∈I gi (X)
= Q
(1)
f (X) =
ek
d
k∈K pk
Q
where g(X) = i∈I gi (X)ei is the unique irreducible factorization in Z[X] (the gi (X)
Q
may be possibly constant) and d = k∈K pfkk is the factorization of d in Z. Obviously,
f (X) is integer-valued if and only if d divides the fixed divisor of g(X), that is, for each
k = 1, . . . , m, pekk divides d(g). Since Int(Z) ⊂ Q[X] and Q[X] is a UFD, any irreducible
factor h(X) of f (X) in Int(Z) is a rearrangement of the irreducible factors gi (X) of g(X)
and the prime factors pk of d, in such a way that we still have an integer-valued polynomial,
that is:
Q
e′i
g1 (X)
i∈J gi (X)
= Q
h(X) =
f′
d1
pk
k∈T
e′i
ei , fk′
k
where J ⊆ I, T ⊆ K, ≤
≤ fk for each i ∈ J and k ∈ T and h(X) is in Int(Z), that
is d1 divides d(g1 ). It is already not clear how a general irreducible polynomial in Int(Z)
looks like (for polynomials g ∈ Z[X] which are irreducible in Int(Z) see the above Theorem
of Chapman-McClain). To our knowledge, the only characterization of such irreducibles
is given by [4, Cor. 2.9], which largely relies on the problem of establishing the fixed
divisor of a polynomial with integer coefficients. We will give a new characterization of the
irreducible elements of Int(Z) in the case of square-free denominator.
As we observed above, being image primitive is a necessary condition for an integervalued polynomial to be irreducible. We will give a chacterization of image primitive
3
polynomials. In particular, being image primitive implies that the numerator g(X) is
primitive (since we assume that the denominator d is coprime with the content of g(X)).
Notice that the converse of the previous statement does not hold, namely, if an integervalued f (X) is image primitive then it is not true in general that f (X) is irreducible in
Int(Z). Consider for example f (X) = X(X − 1)2 /2 which has the irreducible factor X − 1
(by [3, Cor. 2.2 & Example 2.3], monic linear polynomials are irreducible in Int(Z)).
Let p ∈ Z be a fixed prime. We set
Ip + pInt(Z) ∩ Z[X] = {g ∈ Z[X] | p|d(g)},
which is the ideal of polynomials in Z[X] whose fixed divisor is divisible by p. From [9]
(but see also [5, Chapt. 2, 18, p. 22]) we know that
Ip = (p, X p − X) = p,
Y
p−1
\
(p, X − j).
(X − i) =
i=0,...,p−1
j=0
The last intersection is precisely the primary decomposition of the ideal Ip (see [9, Lemma
2.2]). For j = 0, . . . , p − 1, we set
Mp,j + (p, X − j) = {g ∈ Z[X] | p|g(j)}.
The above intersection is actually equal to a product of ideals, since Mp,j , for j = 0, . . . , p−
1, are p distinct maximal ideals in Z[X]. More in general, if n is a positive integer, we set
Ipn + pn Int(Z) ∩ Z[X]
which is the ideal of polynomials whose fixed divisor is divisible by pn . Clearly, a polynomial
f ∈ Q[X] like in (1) is in Int(Z) if and only if for every prime factor pk of the denominator
d, the numerator g(X) is in Ipek .
k
In the next section we will introduce the notion of prime covering for the set of irreducible factors of the numerator of an integer-valued polynomial f (X). For each prime
p which appears in the denominator and for each irreducible polynomial g ∈ Z[X] which
appears in the numerator, we look for the primary components of Ip which contain g(X).
A subset of the irreducible factors {gi (X)}i∈I of the numerator of f (X) whose elements
are contained in all the primary components of Ip is called a p-covering. A p-covering is
minimal if, whenever we remove an element, one of the primary components of Ip is not
covered by any of the polynomials
left in the p-covering itself. In the case of prime denomiQ
g(X)
i∈I gi (X)
, f (X) is irreducible in Int(Z) if and onlyQif {gi (X)}i∈I
nator, say f (X) = p =
p
form a minimal p-covering. In the same way, if for a subset J ( I we have i∈J gi (X) in
Ip , then f (X) is reducible in Int(Z). If that choice is minimal in the above sense, then that
factor is irreducible.
In the subsequent section, we generalize the previous results to the case of an integervalued polynomial with square-free denominator. Finally, as an explicit example, we consider the case of an integer-valued polynomial with denominator equal to the product of
two distinct primes.
4
2
2.1
Integer-valued polynomials with prime denominator
Prime covering
Definition 2.1. Let g ∈ Z[X] and p ∈ Z be a prime. We set
Cp,g + {j ∈ {0, . . . , p − 1} | p|g(j)}.
Notice that the elements of Cp,g correspond precisely to the primary components Mp,j =
(p, X − j) of Ip which contain g(X). Observe that the set Cp,g can be empty: for instance,
take g(X) = X 2 + 1 and p = 3. Obviously, #Cp,g ≤ p. Equivalently, we may consider the
polynomial g ∈ (Z/pZ)[X] obtained by reducing the coefficients of g modulo p. A residue
class j ∈ Z/pZ is a root of g(X) if and only if the primary component Mp,j of Ip contains
g(X).
We have the following result, which involves the family of sets {Ci }i∈I just defined. We
omit the proof, which follows directly from the definitions.
Q
Lemma 2.1. Let g(X) = i∈I gi (X) be a product of polynomials in Z[X] and let p be a
prime. For each i ∈ I, let Ci = Cp,gi . Then
[
g ∈ Ip ⇔
Ci = {0, . . . , p − 1}.
i∈I
In particular, g(X) is p-image primitive if and only if there exists j ∈ {0, . . . , p − 1} such
that no Ci contains j.
Notice that the condition g(X) is p-image primitive is equivalent to g ∈
/ Ip . Obviously
we don’t need to factor a given integer coefficient polynomial g(X) in Z[X] in order to
establish whether it is p-image primitive or not (just consider it modulo p as we said
above). By Polya’s Theorem we cited in the introduction, it is sufficient to consider only
those primes p which are less or equal to the degree of g(X). Anyway, for the study of the
problem of the factorization in the ring Int(Z) it is useful to write the statement as it is.
We give now the following definition.
Definition 2.2. Let G = {gi (X)}i∈I be a set of polynomials in Z[X]. Let p be a prime.
For each i ∈ I we set Ci = Cp,gi . A p-covering for G (or just prime covering, if the
prime p is understood) is a subset J of I such that
[
Ci = {0, . . . , p − 1}.
i∈J
We say that J is minimal if no proper subset J ′ of J has the same property. We will
always assume that a given prime covering J is proper, that is, for each i ∈ J we have
Ci 6= ∅.
Two p-covering J1 , J2 ⊂ I for G are disjoint if J1 ∩ J2 = ∅.
5
Notice that from a prime covering we can always extract a minimal prime covering, by
discarding Q
the redundant sets Ci . We may rephrase Lemma 2.1 by saying that f (X) =
g(X)/p = i∈I gi (X)/p belongs to Int(Z) if and only if I contains a p-covering for {gi }i∈I .
A minimal p-covering can have 1 element, for example consider the irreducible polynomial
X p − X + p. It has at most p elements. The problem to find such p-coverings has a
combinatorial flavour.
The next example shows that given a minimal p-covering J, it does not follow that
{Ci }i∈J forms a family of disjoint subsets of the residue classes modulo p. In fact, a
polynomial gi (X) may belong to different primary components of Ip . If this is the case the
degree of gi (X) has to be greater than one.
Example 2.1.
(X 2 − X + 3)(X 2 + 2)
3
2
2
if we set g1 (X) = X − X + 3, g2 (X) = X + 2 we immediately see that
f (X) =
(2)
• C3,g1 = {0, 1}, C3,g2 = {1, 2}
• C2,g1 = ∅, C2,g2 = {0}
the second line implies that 2 does not divide the fixed divisor of the numerator, that is
f (X) is 2-image primitive (by Polya’s Theorem, we check only those primes p which are
less or equal to the degree of f (X)). We have that C3,g1 and C3,g2 covers the residue classes
modulo 3 and they have non trivial intersection. In particular, I = {1, 2} is a minimal
3-covering.
2.2
Integer-valued polynomials which are p-image primitive
We characterize now p-image primitive integer-valued polynomials, when the denominator
is exactly divisible by a prime p ( we denote this by p k d ).
Suppose that for a polynomial f (X) as in (1) the denominator d is equal to a prime p.
If f (X) is p-image primitive, then there exists i ∈ I such that ei = 1, otherwise the fixed
2
divisor of the numerator g(X) is divisible by pn , for some n > 1. For example, X(X−1)
2
2
2
is not (the numerator has fixed divisor equal to 4).
is 2-image primitive, while X (X−1)
2
However, this condition on the exponents of the irreducible factors gi (X) is not sufficient
to ensure that f (X) is p-image primitive, as the next example shows:
f (X) =
(X 2 + 4)(X 2 + 3)
2
(3)
The polynomial f (X) is not 2-image primitive since the numerator has fixed divisor equal
to 4 (modulo 2, each factor at the numerator has a double root in 0 and 1, respectively).
6
Moreover, under the above assumption, the next
Q lemma shows that all the minimal
p-coverings must intersect in one spot. For g(X) = i∈I gi (X) ∈ Z[X] and J ⊆ I, we set
gJ (X) +
Y
gi (X).
i∈J
For each i ∈ I we set Ci = Cp,gi . By Lemma 2.1, for any subset J ⊆ I we have gJ ∈ Ip ⇔ J
is a p-covering.
Lemma 2.2. Let
Q
gi (X)
d
be in Int(Z). Let p be a prime factor of d such that p k d. Then f (X) is p-image primitive
if and only if the following condition holds: there exists a primary component Mp,j of Ip ,
for some j ∈ {0, . . . , p − 1}, such that gi ∈ Mp,j \ M2p,j for some i ∈ I and for all i ∈ I,
f (X) =
i∈I
i 6= i, we have gi ∈
/ Mp,j .
If that condition holds, then for every minimal p-covering J ⊆ I, we have i ∈ J.
Proof : Suppose f (X) is p-image primitive. If for every j ∈ {0, . . . , p − 1} there exist
i1 (j) 6= i2 (j) in I such that gi1 , gi2 ∈ Mp,j , then we can form two disjoint p-coverings
Jt = {it (j)}j=0,...,p−1 , for t = 1, 2. By Lemma 2.1 the polynomials gJ1 and gJ2 belong to
Ip , thus their fixed divisor is divisible by p; since gI is divisible by gJ1 · gJ2 , it has fixed
divisor divisible by p2 , contradiction. So there exists j ′ ∈ {0, . . . , p − 1} for which only one
/ M2p,j ′ we are done. Suppose that is not the
irreducible factor gi′ (X) is in Mp,j ′ . If gi′ ∈
case. If for all the other j’s we have either more than one factor gi (X) in Mp,j or a factor
gi (X) which belongs to M2p,j we get again to the same contradiction as before. Hence,
there must be some j ∈ {0, . . . , p − 1} for which the corresponding primary component
Mp,j of Ip contains only one factor gi (X). Moreover, gi ∈
/ M2p,j .
Conversely, suppose there exists j ∈ {0, . . . , p − 1} as in the statement. If, for each
/ Ci for all i 6= i. Let J ⊆ I be a minimal
i ∈ I, we set Ci = Cp,gi we have that j ∈
p-covering S
for {gi }i∈I (we know that such a prime covering exists by Lemma 2.1). Since by
definition i∈J Ci = {0, . . . , p − 1}, and for all i ∈ I, i 6= i, we have Ci 6∋ j, it follows that
i is contained in J. Notice that this proves the last statement of the Lemma. So there are
/ M2p,j and gi is the only factor of the numerator of
no two disjoint p-coverings. Since gi ∈
f (X) in Mp,j we have that gJ ∈
/ Ip2 . Since this holds for every minimal p-covering J, this
concludes the proof of the lemma.
Remark 2.1. Under the assunptions of Lemma 2.2, f (X) is p-image primitive if and only
if there exists a primary component Mp,j of Ip which contains one and only one irreducible
/ M2p,j . In particular, this means that only
factor gi (X) of the numerator of f (X) and gi ∈
Ci contains j.
7
The last statement of Lemma 2.2 cannot be reversed, see for example (3). We have to
add the hypothesis that for each minimal p-covering J ⊆ I there exists at least one i ∈ J
such that gi ∈ Mp,j \ M2p,j for some j ∈ J. Equivalently, by the remarks after Definition
2.1, we can say that for at least one residue classes j modulo p, there is one and only one
irreducible factor gi (X) which has a simple root modulo p in j.
We can have more than one minimal p-covering, say J1 , J2 ⊆ I, provided they are not
disjoint, as Lemma 2.2 says. For instance, consider the polynomial:
f (X) =
X(X − 1)(X − 2)
2·3
(4)
which is known to be irreducible ([3, Example 2.8]; in particular, f (X) is image primitive).
We set gi+1 (X) = X − i, for i = 0, 1, 2. Then J1 = {1, 2} and J2 = {2, 3} are different
minimal 2-coverings, which are not disjoint.
Example 2.2.
X 2 · (X − 1) · (X 2 + 4)
2
in this example only X − 1 belongs to M2,1 and moreover it does not belong to M22,1 .
Hence, the polynomial is 2-image primitive.
f (X) =
Example 2.3.
X · (X 2 − 2X + 5) · (X + 6)
2
2
in this example, only g(X) = X − 2X + 5 belongs to M2,1 . Moreover, g ∈ M22,1 . No
irreducible polynomial in the numerator belongs to M22,0 , but there are two distinct factors,
namely X and X + 6, which belong to M2,0 . Hence, f (X) is not 2-image primitive, since
the fixed divisor of the numerator is 4. So it is not sufficient to have a unique i ∈ I such
that gi ∈ Mp,j . We must also take care of the exact power of the maximal ideal Mp,j to
which each polynomial gi (X) belongs to.
f (X) =
2.3
Irreducible integer-valued polynomials
Suppose that an integer-valued polynomial f (X) is of the form
Q
gi (X)
g(X)
f (X) =
= i∈I
p
p
(5)
where, for i ∈ I, gi ∈ Z[X] is irreducible. The fact that f ∈ Int(Z) is image primitive
amounts to saying that d(g) is equal to p. Since f ∈ Int(Z), by Lemma 2.1 there exists a
p-covering J ⊆ I for {gi (X)}i∈I . The next lemma establishes that f (X) is irreducible in
Int(Z) if and only if I is a minimal p-covering.
8
Lemma 2.3. An image primitive polynomial f (X) = g(X)/p in Int(Z) as inS(5) is irreducible in Int(Z) if and only if there is no proper subset J of I such that j∈J Cj =
{0, . . . , p − 1} (that is, I is a minimal p-covering).
Proof : Suppose there exists J ( I such that J is a p-covering. Then
f (X) =
gJ (X)
· gI\J (X)
p
is a non-trivial factorization of f (X) in Int(Z), because the first factor is integer-valued by
Lemma 2.1 and the second one is in Z[X] ⊂ Int(Z).
Conversely, if f (X) is reducible in Int(Z), then there exist non-constant g, h ∈ Int(Z)
such that f (X) = h1 (X)h2 (X) (because we are assuming f (X) to be image primitive).
Since p must appear in the denominator of one of the two factors, say h1 (X), then for
some ∅ =
6 J ( I we have h1 (X) = gJ (X)/p and consequently h2 = gI\J ∈ Z[X]. Since
h1 ∈ Int(Z), by Lemma 2.1 J is a p-covering (notice that h1 ∈ Int(Z) ⇔ gJ ∈ Ip ).
Notice that Lemma 2.3 does not hold without assuming f (X) to be image primitive, as
example (3) shows. By the arguments we have just given, we deduce that every factorization
is
of an image primitive integer-valued polynomial with prime denominator f (X) = g(X)
p
of the form f (X) = gJ (X)
· gI\J (X), for some J ⊆ I minimal p-covering. Notice that the
p
number of irreducible factors of the previous factorization in Int(Z) is 1 + #(I \ J). The
assumption that f (X) is image primitive implies that for each such a minimal p-covering
J, the set I \ J does not contain a p-covering.
3
Integer-valued polynomials with square-free denominator
The main problem in the general case of more than one prime factor in the denominator d
of an integer-valued polynomial f (X) is that each irreducible factor gi (X) of the numerator
of f (X) may belong to different primary components Mpk ,j of Ipk , where {pk }k∈K are the
different prime factors of d.
As already remarked in [6], this phenomenon has the effect that if p,q are two distinct
primes, then it does not follow that Ip · Iq = Ipq : for example, g(X) = X(X − 1)(X − 2)
belongs to I2·3 (see (4)), but it cannot be expressed as a product of a polynomial in I2 and
a polynomial in I3 . This is due to the fact that the only minimal 3-covering J = {1, 2, 3}
is equal to the set I itself, so in particular it has non-zero intersection with any possible
2-covering (we saw that there are only two of them). Hence, in the next subsection, we are
lead to give this globalizing definition.
9
3.1
Family of minimal P-coverings
Definition 3.1. Let G = {gi (X)}i∈I be a set of polynomials in Z[X] and let P = {pk }k∈K
be a set of distinct prime integers. A family of minimal P-coverings for G is a family
of sets {Jk }k∈K such that for each k ∈ K, Jk ⊆ I is a minimal pk -covering for G.
Let f ∈ Q[X] be as in (1). If f (X) is an integer-valued polynomial, then by Lemma
2.1 there exists a family of minimal P = {pk }k∈K -coverings for G = {gi (X)}i∈I .
We can now formulate a proposition, which gives a criterion for an integer-valued
polynomial to be irreducible, in the case that the denominator is square-free. This is a first
step to determine explicitly all the factorizations of a given element in the ring Int(Z).
Firstly we set some notations. Let
Q
g(X)
i∈I gi (X)
(6)
f (X) =
= Q
d
k∈K pk
be a polynomial in Q[X], with pk distinct prime integers, gi ∈ Z[X] irreducibleQ
polynomials.
Notice
that
the
condition
that
f
(X)
is
integer-valued
is
equivalent
to
g
(X)
=
I
i∈I gi (X) ∈
T
k∈K Ipk . We set G =Q{gi (X)}i∈I and P = {pk }k∈K . As in the previous section, given
J ⊆ I we set gJ (X) + i∈J gi (X). Notice that if J1 ⊆ J2 ⊆ I we have that gJ1 (X) divides
gJ2 (X) in Z[X] (and so in Int(Z)). Similarly, for a subset T ⊆ K we set
Y
pk
dT +
k∈T
(dK = d). With these notations, a factor of f (X) is of the form:
h(X) =
gJ (X)
dT
for some J ⊆ I and T ⊆ K.
Finally, if T ⊆ K and J = {Jk }k∈K is a family of minimal P-coverings for G, we set
[
IJ ,T +
Jk .
k∈T
Notice that, if T1 , T2 ⊆ K are two disjoint subsets, then IJ ,T1 ∪T2 = IJ ,T1 ∪ IJ ,T2 .
3.2
Irreducible integer-valued polynomials
Theorem 3.1. Let
Q
g(X)
i∈I gi (X)
= Q
f (X) =
d
k∈K pk
10
be an image primitive integer-valued polynomial. Let P = {gi (X)}i∈I and G = {pk }k∈K .
We suppose that the polynomials gi (X) are irreducible in Z[X] and that the pk are distinct
prime integers. Then f (X) is irreducible in Int(Z) if and only if the following holds: for
every family J = {Jk }k∈K of minimal P-coverings for G we have
i) I = IJ ,K .
˙ 2 such that IJ ,K1 ∩ IJ ,K2 = ∅.
ii) there is no non-trivial partition K = K1 ∪K
Notice that condition i) implies that for each i ∈ I there exists k ∈ K such that
Cpk ,gi 6= ∅, so that each of the gi ’s belongs to at least one of the primary components
Mpk ,j of some of the ideals Ipk . Moreover, condition ii) says that the union of the elements
Jk of the family J cannot be partitioned (in a sense we will make precise soon). We will
treat the case P = {p1 , p2 } as an example in section 3.4.
Proof : Suppose f ∈ Int(Z) irreducible. Let J = {Jk }k∈K be a family of minimal Pcoverings for G (it exists because of Lemma 2.1). If I strictly contains IJ ,K then there
exists t ∈ I which is not contained in any Jk (equivalently, Jk ⊆ I \ {t} for every k ∈ K).
This means that gt (X) divides f (X) in Int(Z), because we have
f (X) = gt (X) ·
gI\{t} (X)
d
and the second factor is integer-valued, since for each k ∈ K we have gJk (X) ∈ Ipk (see
Lemma 2.1). Hence, for all such k’s, we have gI\{t} (X) ∈ Ipk , since Jk ⊆ I \ {t}. This is a
contradiction, hence condition i) holds.
˙ 2 such that I1 + IJ ,K1 and I2 + IJ ,K2 are
If we have a non-trivial partition K = K1 ∪K
disjoint, then
gI (X) gI2 (X)
·
.
f (X) = 1
dK1
dK2
Notice that for every k1 ∈ K1 we have gJk1 (X) ∈ Ipk1 (again by Lemma 2.1) and gJk1 (X)
divides gI1 (X) in Z[X], since Jk1 ⊂ I1 . This implies that gI1 (X)/dK1 is integer-valued.
Similarly, the second factor is integer-valued, too. That would be a non-trivial factorization
of f (X), which is a contradiction.
Conversely, suppose that for every family J = {Jk }k∈K of minimal P-coverings for G
conditions i) and ii) hold. Since f (X) is image primitive, there is no non-unit in Z which
divides f (X) in Int(Z). If f (X) is reducible in Int(Z) we have f (X) = h1 (X)h2 (X), where
h1 , h2 ∈ Int(Z) are not constant. Since Int(Z) ⊂ Q[X] we have
hi (X) =
gIi (X)
dKi
for some Ii ⊆ I and Ki ⊆ K, for i = 1, 2. Necessarily, I1 , I2 are disjoint and I1 ∪ I2 = I.
Similarly, K1 and K2 are disjoint and K1 ∪ K2 = K. Suppose that one of the Ki , say K2 ,
11
is empty. Then, by Lemma 2.1 for each k ∈ K1 = K there exists a minimal pk -covering
Jk ⊆ I1 . We set J = {Jk }k∈K . By definition, the family J is a minimal P-coverings for
G. In particular, IJ ,K ⊆ I1 , because each of the Jk ’s is a subset of I1 . Because of i) we
have that I = IJ ,K , so that I = I1 and consequently I2 = ∅, since I1 and I2 are disjoint.
This means that h2 (X) is a unit.
Suppose now that Ki 6= ∅, for i = 1, 2. This fact also leads us to a contradiction. In
fact, by Lemma 2.1, for each i = 1, 2 and for each ki ∈ Ki there exists a minimal pki covering Jki ⊆ Ii . We set J = {Jk }k∈K , which is a family of minimal P-coverings for G.
In particular, IJ ,Ki ⊆ Ii . By condition i) on J we have that
˙ J ,K2 .
I = IJ ,K = IJ ,K1 ∪I
Since I1 ∪ I2 = I, we get IJ ,Ki = Ii for i = 1, 2, which is in contradiction with condition
ii).
Example 3.1. It is not sufficient that conditions i) and ii) of Theorem 3.1 hold only for
one family {Jk }k∈K of minimal P-coverings. For instance, let us consider
f (X) =
(X − 1) · (X − 2) · (X − 3) · (X − 9)
2·3
then if gi (X) = X − i, for i = 1, 2, 3, g4 (X) = X − 9 and I = {1, 2, 3, 4}, we have that
- J2 = {2, 1}, J2′ = {2, 3} and J2′′ = {2, 4} are the minimal 2-coverings.
- J3 = {1, 2, 3} and J3′ = {1, 2, 4} are the minimal 3-coverings.
We have that J = {J2′′ , J3 } is a family of minimal P-coverings for G which satisfies both
conditions i) and ii) but the polynomial is not irreducible, since X − 9 divides f (X) in
Int(Z). In fact, the family J ′ = {J2′ , J3 } of P-coverings for G does not satisfy condition i)
of the proposition.
Remark 3.1. From Theorem 3.1 we see that each family of minimal P-coverings for G
determines a (possibly trivial, like for J in Example 3.1) factorization for f (X) in Int(Z).
Conversely, every non-trivial factorization determines a family of minimal P-coverings for
G which can be partitioned in the following sense:
Definition 3.2. We say that a family Ṡ
J of minimal P-coverings for G is partitionable
if
S
there exist a partition for K, say K = j∈I Kj such that the sets {IJ ,Kj = k∈Kj Jk | j ∈
I} are disjoint.
However, notice that different families of minimal P-coverings may give the same factorization for f (X). For instance, in the Example 3.1, there are six possible such families (we
have to pair each minimal 2-covering with a minimal 3-covering). The family J ′′ = {J2 , J3 }
gives the same factorization as J ′ . This depends on the fact that IJ ′ ,K and IJ ′′ ,K are equal.
12
Corollary 3.1. Let f (X) be as in the assumptions of Theorem 3.1. If there exists k ∈ K
such that I is a minimal pk -covering, then f (X) is irreducible in Int(Z).
Proof : We retain the notations of Theorem 3.1. Let J be a family of minimal P-coverings
for G. Notice that I is the only minimal pk -covering, so I ∈ J and consequently I = IJ ,K
and J is not a partitionable family. Hence, the conditions i) and ii) of Theorem 3.1 are
satisfied for every family of minimal P-coverings, so f (X) is irreducible in Int(Z).
In particular, this corollary shows again that the polynomial in (4) is irreducible. The
condition of the previous corollary is not necessary, see (10) below for an example.
3.3
The algorithm of factorization in Int(Z)
The next corollary shows explicitly how to obtain a non-trivial factorization of an integervalued polynomial f (X) as in (6) from a partitionable family of minimal P-coverings for
G. We know from the proof of Theorem 3.1 that every such factorization is obtained in
this way.
We recall that we are assuming f (X) to be image primitive and the denominator of
f (X) to be square-free.
Schematically we are doing the following steps:
i) For each k ∈ K and for each i ∈ I we determine the sets Cpk ,gi .
ii) Afterwards for each k ∈ K we find all the minimal pk -coverings Jk , by grouping
together the sets Cpk ,gi .
iii) Then for each k ∈ K we choose one of the minimal pk -coverings we found at point
ii) and we define the family J = {Jk }k∈K of minimal P-coverings for G.
Corollary 3.2. Let
Q
gI (X)
i∈I gi (X)
=
f (X) = Q
p
dK
k∈K k
be an image primitive, integer-valued polynomial, where pk are distinct prime integers,
gi ∈ Z[X] distinct and irreducible.
Every factorization of f (X) in Int(Z) is obtained in the following way:
letṠJ = {Jk }k∈K be a family of minimalSP-coverings for G which is partitionable, say
K = j∈I Kj , so that the sets Ij + IJ ,Kj = k∈Kj Jk , for j ∈ I, are disjoint and for each
j ∈ I the integer-valued polynomial gIj (X)/dKjSsatisfies the conditions of Theorem 3.1 (so
that each of them is irreducible). We set I ′ + j∈I Ij . Then
f (X) = gI\I ′ (X) ·
Y gIj (X)
dKj
j∈I
13
is a factorization of f (X) in Int(Z) and every one of them is obtained in that way. Notice
that in the previous factorization we have #(I \ I ′ ) + #I irreducible factors.
3.4
Case d = p1 · p2
Let p1 , p2 ∈ Z be distinct primes. We consider an image primitive integer-valued polynomial
of the following form:
Q
gi (X)
g(X)
f (X) =
= i∈I
(7)
p1 p2
p1 p2
This amounts to saying that the fixed divisor d(g) is equal to p1 p2 . By Lemma 2.1, for
k = 1, 2, there exists a pk -covering Jk for {gi (X)}i∈I . For each i ∈ I and for each k = 1, 2
we consider the sets Cpk ,gi as defined in section 2. With the notation we introduced so far,
we can have two different kind of factorization of f (X). One possible factorization is:
Y
gJ (X) gJ2 (X)
gi (X)
(8)
·
·
f (X) = 1
p1
p2
i∈I\J1 ∪J2
for some J1 , J2 ⊆ I, where, for k = 1, 2, gJk (X)/pk ∈ Int(Z) is irreducible. By Lemma 2.3,
this corresponds to the fact that, for k = 1, 2, Jk is a minimal pk -covering. Obviously, J1
and J2 are disjoint.
Another possible factorization is
gJ (X) Y
·
gi (X)
(9)
f (X) =
p1 p2
i∈I\J
for some J ⊆ I. In this factorization gJ (X)/(p1 p2 ) ∈ Int(Z) is irreducible.
By Lemma 2.1, since gJ (X)/(p1 p2 ) is integer-valued then for each k = 1, 2, J contains
a minimal pk -covering Jk . By Theorem 3.1, the fact that gJ (X)/(p1 p2 ) is irreducible in
Int(Z) is equivalent to saying that J = J1 ∪ J2 (otherwise, we can factor out some gi (X)
from it) and J1 ∩ J2 6= ∅ (otherwise we fall in the previous case (8)). It is not true that for
some k = 1, 2 we must have I = Jk , like example (10) below shows.
In [4, Example 3.6] the authors construct an integer-valued polynomial which has two
distinct factorizations as in (8) and (9). Now we give other two explicit examples: in the
first one only the factorization as in (8) occurs, in the second one we give an irreducible
polynomial in Int(Z) of the form g(X)/(p1 p2 ).
Example 3.2.
(X 2 + 12)(X 2 + 2)(X 2 + 10)(X 2 + 16)(X 2 + 4)
3·5
(X 2 + 12)(X 2 + 2) (X 2 + 10)(X 2 + 16)(X 2 + 4)
=
·
3
5
f (X) =
14
the second line is the only factorization in Int(Z) that f (X) can have, since if we put
g1 (X) = X 2 + 12, g2 (X) = X 2 + 2, g3 (X) = X 2 + 10, g4 (X) = X 2 + 16, g5 (X) = X 2 + 4 we
have:
C3,g1 = {0}, C3,g2 = {1, 2}, C3,g3 = ∅,
C3,g4 = ∅,
C3,g5 = ∅
C5,g1 = ∅,
C5,g2 = ∅,
C5,g3 = {0}, C5,g4 = {2, 3}, C5,g5 = {1, 4}
so in I = {1, . . . , 5} we only have one 3-covering J3 = {1, 2} and only one 5-covering
J5 = {3, 4, 5}, and they are disjoint. It is easy to check that 2 and 7 do not divide the fixed
divisor of the numerator of f (X).
Example 3.3.
f (X) =
X(X 2 + 2)(X 2 + 16)(X 2 + 4)
3·5
(10)
so if g1 (X) = X, g2 (X) = X 2 + 2, g3 (X) = X 2 + 16, g4 (X) = X 2 + 4 we have:
C3,g1 = {0}, C3,g2 = {1, 2}, C3,g3 = ∅,
C3,g4 = ∅
C5,g1 = {0}, C5,g2 = ∅,
C5,g3 = {2, 3}, C5,g4 = {1, 4}
Then by Theorem 3.1 f (X) is irreducible in Int(Z) since J3 = {1, 2} is the only minimal
3-covering, J5 = {1, 3, 4} is the only minimal 5-covering, I = J3 ∪ J5 and J3 ∩ J5 6= ∅.
Notice that J3 ( I, J5 ( I. It is easy to check that 2 and 7 do not divide the fixed divisor
of the numerator g(X) of f (X). In particular, f (X) is image primitive, that is d(g) = 3 · 5.
Example 3.4. As another application of Theorem 3.1 we consider the polynomial:
f (X) =
X · (X 2 + 1) · (X 2 + X + 1) · (X 2 + 2X + 4)
2·3
and let g1 (X) = X, g2 (X) = X 2 + 1, g3 (X) = X 2 + X + 1, g4 (X) = X 2 + 2X + 4. Then
- J2 = {1, 2} and J2′ = {2, 4} are the minimal 2-coverings.
- J3 = {1, 3, 4} is the only minimal 3-covering.
So J = {J2 , J3 } is a family of minimal P-coverings of G such that J2 ( I, J3 ( I. The
same holds for J ′ = {J2′ , J3 }. The polynomial is irreducible by Theorem 3.1: if we consider
J we have I = J2 ∪ J3 and J2 ∩ J3 6= ∅. The same holds for J ′ .
Our method can be easily generalized to the case of denominator divisible by prime
powers pn such that n ≤ p, since in this case, by [9, Proposition 3.1], the primary components of the ideal Ipn are just the n-th power of the maximal ideals Mp,j , for j = 0, . . . , p−1.
In general, a further study of the primary components of the ideal Ipn is needed.
15
Acknowledgments
I would like to thank the referee for his/her valuable suggestions. The author was supported
by the Austrian Science Foundation (FWF), Project Number P23245-N18 and also by
INDAM - GNSAGA.
References
[1] D. F. Anderson, P.-J. Cahen, S.T. Chapman and W. Smith, Some factorization properties of the ring of integer-valued polynomials, Zero-dimensional commutative rings
(Knoxville, TN, 1994), 125-142, Lecture Notes in Pure and Appl. Math., 171, Dekker,
New York, 1995.
[2] J.-P. Cahen, J.-L. Chabert, Integer-valued polynomials, Amer. Math. Soc. Surveys and
Monographs, 48, Providence, 1997.
[3] J.-P. Cahen, J.-L. Chabert, Elasticity for integral-valued polynomials., J. Pure Appl.
Algebra 103 (1995), no. 3, 303-311.
[4] S.T.Chapman, B. McClain. Irreducible polynomials and full elasticity in rings of
integer-valued polynomials, J. Algebra 293 (2005), no. 2, 595-610.
[5] L. E. Dickson. Introduction to the theory of numbers. Chicago 1929.
[6] D. J. Lewis. Ideals and Polynomial Functions, Amer. J. Math. 78 (1956), 71-77.
[7] S. Frisch. A construction of integer-valued polynomials with prescribed sets of lengths
of factorizations, Monatsh. Math. 171 (2013), 341-350.
[8] W. Narkiewicz. Polynomial mappings, Lecture Notes in Mathematics, 1600. SpringerVerlag, Berlin, 1995. viii+130 pp. ISBN: 3-540-59435-3.
[9] G. Peruginelli. Primary decomposition of the ideal of polynomials whose fixed divisor
is divisible by a prime power. Accepted for publication by Journal of Algebra.
16
| 0math.AC
|
1
L ATTICE E RASURE C ODES OF L OW R ANK
WITH
N OISE M ARGINS
Vinay A. Vaishampayan
arXiv:1801.04466v1 [cs.IT] 13 Jan 2018
Dept. of Engineering Science and Physics
City University of New York-College of Staten Island
Staten Island, NY USA
Abstract
We consider the following generalization of an (n, k) MDS code for application to an erasure
channel with additive noise. Like an MDS code, our code is required to be decodable from any k
received symbols, in the absence of noise. In addition, we require that the noise margin for every
allowable erasure pattern be as large as possible and that the code satisfy a power constraint. In this
paper we derive performance bounds and present a few designs for low rank lattice codes for an additive
noise channel with erasures.
Index terms Lattices, Erasure Codes, MDS Codes, Compound Channel.
I. I NTRODUCTION
YS1
Z
EncoderModulator
X
Y
+
Erasure
Network
YS2
YSi
YSm
Fig. 1. Coding and Modulation for the Erasure Network.
We consider low rank lattice codes for transmission over a noisy erasure channel as illustrated
in Fig. 1. In this figure k information symbols are mapped by an encoder/modulator to a vector
x = (x1 , x2 , . . . , xn ) ∈ Λ, where Λ is a rank-k lattice in Rn . The output of the additive noise
channel is y = x + z, where z = (z1 , z2 , . . . , zn ) is a noise vector independent of x and with
independent components. Components of y are then erased by an erasure network, whose outputs
2
are obtained by retaining only those symbols of y indexed by subsets S ⊂ {1, 2, . . . , n} in a
given sub-collection of subsets; thus yS coincides with y is the positions identified by S. As
an example, with n = 4, S = {2, 4} and y = (a, b, c, d), yS = (b, d). A decoder estimates
the source symbols based on yS with a probability of error denoted Pe (S). The objective is to
minimize Pe (S) for each S by designing a single codebook which satisfies a power constraint
E[X t X] ≤ nP , where E denotes expectation with respect to a uniform distribution on the
codebook. Here we consider as our sub-collection S, all k-subsets of {1, 2, . . . , n}. Our paper is
organized as follows. Prior work and lattice background is in Sec. II. Two performance bounds
are presented in Sec. III, constructions for codes in dimension n = 4 are presented and compared
to the derived bounds in Sec. IV. A summary is in Sec. V.
We use the acronym w.l.o.g to mean ‘without loss of generality’.
II. P RIOR W ORK AND R EVIEW OF L ATTICE T ERMINOLOGY
The problem considered here may be viewed as a code design problem for a special case of
the compound channel, see e.g. [1], [5]. This work was motivated by a study on cross layer
coding that appeared in [4]. For prior contributions on the Gaussian erasure channel, please refer
to [8] and the references therein.We now develop notation and some basic definitions for low
rank lattices in Rn . Let {φi , i = 1, 2, . . . , k} be a collection of k ≤ n orthonormal column
vectors in Rn and let Φ = (φi , i = 1, 2, . . . , k) denote the associated n × k orthonormal matrix.
Let V denote a k × k generator matrix of full rank for a lattice ΛV = V Zk := {V u, u ∈ Zk }.
We will refer to ΛV as the mother lattice. Let
Λ = ΦΛV := ΦV Zk .
(1)
Λ is a rank-k lattice in Rn . Let G(ΛV ) = V t V denote the Gram matrix of ΛV (t is the transpose
operator).
The determinant of a lattice det Λ is defined in terms of the determinant of its Gram matrix by
det Λ := det(G(Λ)). Let ρ(Λ) denote the radius of the largest inscribed sphere in a Voronoi cell
of Λ. The packing density of Λ is defined in terms of Vk the volume of a unit-radius Euclidean
ball in Rk by
√
∆k (Λ) = Vk ρk / det Λ.
(2)
3
We denote by ∆k (opt), the largest packing density that can be acheived by any lattice in Rk . The
problem of finding lattices that maximize the packing density is a classical problem in number
theory and geometry, with several excellent references [3], [6].
The following definitions are from [6]. A body captures the notion of a solid subset of Rn ,
specifically, B ⊂ Rn is a body if it has nonempty interior and is contained in the closure
of its interior [6]. A body B ⊂ Rn is said to be centrally symmetric if B = −B, where
−B = {−x : x ∈ B}. A closed body B with the property that for any x ∈ B, the point
λx ∈ B for every 0 ≤ λ < 1 is called a star body. While convex bodies are star bodies, the
converse is not true. A simple example, and one directly relevant to us is the star body formed by
the union of centrally symmetric ellipsoids in Rn . A lattice Λ is said to be admissible for B ⊂ Rn ,
p
or B-admissible, if no non-zero point in Λ lies in B. The greatest lower bound of ( det Λ)
over all B−admissible lattices is called the lattice constant of B, denoted ∆(B) (which is set
to ∞ is there are no B− admissible lattices). A B-admissible lattice Λ with det Λ = ∆(B)2 is
said to be a critical lattice for B.
A lattice Λ is said to be a packing lattice for a body B if the sets B and B + λ are disjoint for
all non-zero λ ∈ Λ. It is known, Thm.1, Ch. 3, Sec 20 [6], that Λ is a packing lattice for centrally
symmetric, convex body B if and only if it is admissible for 2B. Thus, for a convex body, the
problem of finding a packing lattice for B is equivalent to that of finding an admissible lattice
for 2B. The connection between packings and admissibility for non-convex bodies is messier.
The distinction arises because for a centrally symmetric body B, Λ is a lattice packing of B
if and only if it is admissible for B + B, where + denotes the set sum or Minkowski sum. If
the centrally symmetric body is also convex, then B + B = 2B and thus packing problems and
admissibility problems are closely related. On the other hand, if B is centrally symmetric but
non-convex, in order to solve a packing problem for B one must solve an admissibility problem
for B + B, and this set may not be as easily described as B.
In our application, we need to index body B by subset S in a given sub-collection of subsets
S
and our problem is one of packing S∈S B(S), which is non-convex. While admissibility for
non convex centrally symmetric body 2C says nothing in general about packings for C, it turns
S
out that our design problem is equivalent to finding a critical lattice for 2 S∈S B(S) because the
decoder knows S. Thus it is possible to draw on the theory of admissible lattices for star bodies.
This theory provides several key ingredients to help find a solution to this problem. Most notably,
in the chapter on Mahler’s compactness theorem [2], Theorem VII states that every critical lattice
4
for a bounded star body S has n linearly independent points on the boundary of S.
III. B OUNDS
Let In = {1, 2, . . . , n}. For S ⊂ In , |S| = k let ΛS be the lattice obtained be retaining
only those coordinates that are in S or equivalently ΛS is the projection of Λ into the subspace
CS := Span{ei , i ∈ S} where ei = (0, ..., 0, 1, 0, ..., 0)t is the ith unit vector in Rn . For any
k-subset S ⊂ {1, 2, . . . , n}, we denote by ΦS the k × k submatrix obtained by extracting from Φ
the k rows identified by S. The generator matrix for ΛS is ΦS V and its Gram matrix G(ΛS ) =
V t ΦtS ΦS V .
Define the (packing volume) contraction ratio
βS = (ρ(ΛS )/ρ(ΛV ))k
(3)
let ρmin = minS ρ(ΛS ) and let βmin = minS βS .
A. Determinant Upper Bound
We will use symbols x̄, x# to denote the arithmetic mean and geometric mean, respectively,
of the real numbers xi over some index set I. When a k-dim mother lattice ΛV is set in IRn
using a basis Φ, the projections on the nk subsets S, cannot all be simultaneously good. There
are two important factors that measure the ‘goodness’ of the projections—the packing density
and the scale of the lattices ΛS . The following theorem develops one of two bounds presented
in this paper.
Theorem 1. (Determinant Bound) Given a mother lattice ΛV and orthonormal basis Φ, let
β # and ∆# be respectively, the geometric mean of the volume contraction ratios and packing
densities of the child lattices ΛS , taken over all k-subsets of {1, 2, . . . , n}. Then
(β # ∆(ΛV ))2 ≤
(∆# )2
.
n
(4)
k
Equality holds if and only if all child lattices have equal determinants.
Proof. The packing densities of the mother lattice ΛV and child lattice ΛS are related by the
following identity
∆2 (ΛV )βS2 = ∆2 (ΛS )
det ΛS
.
det ΛV
(5)
5
Compute the geometric mean of both sides over the collection of k-subsets S to get
! 1
Y det ΛS (nk)
∆2 (ΛV )(β # )2 = (∆# )2
.
det ΛV
S
(6)
From the arithmetic-geometric mean inequality it follows that
∆2 (ΛV )(β # )2 ≤ (∆# )2
1
n
k
X det ΛS
det ΛV
S
!
(7)
and equality holds if and only if det ΛS is a constant with respect to S. However
X
X
det ΛS =
det(G(ΛS ))
S
S
=
X
det((ΦS V )t ) det(ΦS V )
S
(a)
= det((ΦV )t (ΦV ))
= det ΛV ,
(8)
where in (a) we have used the Cauchy-Binet formula, see e.g. [7]. The remainder of the proof
follows directly.
The following corollary is immediate.
Corollary 1. Given mother lattice ΛV and orthonormal basis Φ, let βmin be the minimum volume
contraction ratio of the child lattices ΛS , taken over all k-subsets of {1, 2, . . . , n}, and ∆# the
geometric mean of the packing densities. Then
(βmin ∆(ΛV ))2 ≤
(∆# )2
∆k (opt)2
.
≤
n
n
k
(9)
k
Equality holds in the left inequality iff all the contraction ratios are equal and all the child lattices
have equal determinants. Equality holds in the right inequality iff all child lattices achieve the
optimal packing density in dimension k.
B. Trace Upper Bound
Theorem 2. (Trace Bound) For an (n, k) code, the compaction ratio is bounded as
2/k
2/k
βmin ≤ βS
≤
k
.
n
(10)
Equality holds if the shortest vector of each lattice ΛS is the image of the shortest vector in ΛV .
6
Proof. Upon summing over all k-subsets S we obtain
X
X
G(ΛS ) =
V t ΦtS ΦS V
S
S
!
X
= Vt
ΦtS ΦS
V
S
n−1 t
Φ ΦV
= V
k−1
n−1
=
V t V.
k−1
t
(11)
By definition the smallest packing radius of any child lattice ρmin satisfies
ρmin 2 ≤ ρ(ΛS )2 ≤ (1/2)ut G(ΛS )u
(12)
for any non-zero u ∈ Zk and any k-subset S. Upon averaging over subsets S we obtain the
upper bound
ρ(ΛS )2 ≤
=
2
1 X t
u G(ΛS )u
n
k
S
1
2
t
n u
k
X
G(ΛS )u
S
n−1
k−1
2 nk
=
ut G(ΛV )u.
(13)
Equality holds if ΦS V u is the shortest vector in ΛS for all S. Thus
ρ(ΛS )2 ≤
k 2
ρ (ΛV )
n
(14)
and (10) follows immediately.
IV. A NALYSIS OF S OME (4, k) C ODES
We construct Φ for n = 4 for various values of k and various mother lattices ΛV . Numerical
2/k
results for the (4, k), k = 2, 3 are presented in Fig. 2, in which βmin is plotted as a function of
the packing density of the mother lattice. We have plotted the determinant bound using both the
optimal and the cubic lattice for the child lattices. We have also plotted the trace bound.
Observe that in the (n, k) = (4, 2) case there is a significant gap between the best possible
construction and the upper bounds. In the (4, 3) case performance close to the determinant bound
is achieved by setting the mother lattice to be the cubic lattice. Also with the cubic lattice as
the mother lattice, since the trace bound is lower than the determinant bound, this is proof that
7
it is impossible to simultaneously achieve the packing density of D3 when the mother lattice is
the cubic lattice.
(n,k)=(4,3)
(n,k)=(4,2)
0.9
0.6
0.8
0.5
0.7
0.6
- 2/k
min
- 2/k
min
0.4
0.3
0.5
0.4
0.3
0.2
Det. Bound (A 2)
Trace Bound
Det Bound (Z 2)
$ M=Z2
0.1
$ M=A2
Lower Bound
0
0.75
0.8
0.85
0.9
0.2
Det. Bound: D 3
Trace Bound
0.1
Det Bound: Z 3
Z3
FCC
BCC
0
0.4
0.95
0.5
0.6
"($ M)
0.7
0.8
"($ M)
Fig. 2. Bounds on β 2/k and values obtained from the construction.
4
4
4
2
2
2
0
0
0
-2
-2
-2
-4
-5
0
5
-4
-5
0
5
-4
-5
4
4
4
2
2
2
0
0
0
-2
-2
-2
-4
-5
0
5
-4
-5
0
5
0
0.5
0
-4
-5
5
0
5
(a)
2
1.5
1
0.5
0
-0.5
-1
-1.5
-2
-2
-1.5
-1
-0.5
1
1.5
2
(b)
Fig. 3. Packings (a) and the star body (b) derived from noise spheres in each of the six subspaces for the (4, 2) code with
ΛV = Z2 as described in the text.
Geometrically, the ability to decode correctly, post-erasure, with iid Gaussian noise is determined by the largest noise sphere which can be packed by the projected (or child) lattice ΛS
in each of the nk subspaces CS . When each noise sphere is projected back onto the subspace
spanned by the columns of Φ, a noise sphere is transformed into an ellipsoid. To see this
consider the noise sphere kxk2 ≤ r2 in subspace CS . Setting x = ΦS y, this leads to the noise
8
ellipsoid BS (r) = {y : kΦS yk2 ≤ r2 }. The packing of the noise ellipsoids BS (ρmin ), (recall
that ρmin is the half the length of the shortest non-zero vector in ΛS ), by the mother lattice
ΛV is shown in the six panels in Fig. 3(a), one for each subspace. Fig. 3(b) shows the star
S
body B(2ρmin ) = s∈S BS (2ρmin ). This illustrates the star body, which is the union of the six
ellipses, two of which are circles. Also shown are points of the lattice ΛV , which in this case
is an admissible lattice for this star body and illustrates also the interpretation as a code design
problem for the compound channel. Observe that Z2 is simultaneously good as a packing for
all six erasure configurations, i.e. for each of the bodies BS (ρmin ). Also, ΛV is simultaneously
critical for five of the six bodies BS (2ρmin ) (notice that one circle does not touch any of the
lattice points).
A. (4, 1)
t
2
= 1/4. The trace and determinant upper
Let Φ = a a a a , a = 1/2. We obtain βmin
bounds yield β 2 ≤ 1/4. Hence this construction is optimal.
B. (4, 2)
√
With ΛV = Z2 , a = 1/ 3 and
Φt =
a
a
a 0
a −a 0 a
we obtain six child lattices with Gram matrices
2a2 0
a2 a2
,
,
2
2
2
0 2a
a 2a
2
2
2
2
a
−a
2a −a
,
,
2
2
2
2
−a 2a
−a
a
2a2 a2
a
2
a2
0
2
,
a
0
.
2
a
(15)
All six child lattices are similar to Z2 . The first one has shortest vector of square length 2a2 and all
2/k
the others have square length a2 . This code achieves βmin = βmin = 1/3, β # = 21/6 /3 = 0.374.
The trace upper bound is βmin ≤ βS ≤ 1/2, regardless of the mother lattice, while the
√
√
determinant upper bound depends on the mother lattice, and is 2/3 and 1/ 6 = 0.408 for
ΛV = A2 (hexagonal lattice) and Z2 , resp. Thus the determinant bound is tighter than the trace
bound but greater than 1/3. This construction does not meet the trace bound or the determinant
9
bound with equality. However with ΛV = Z2 , the following theorem shows this to be the best
βmin possible. A computer-based search has also failed to reveal any improvements in β # .
Theorem 3. Let Φ be a 4 × 2 matrix with orthonormal columns, so that Φt Φ = I. Let ΛV = Z2 .
Let ΛS = ΦS ΛV where S is a 2-subset of {1, 2, 3, 4}. Let rS be the length of the shortest non-zero
vector in ΛS . Then rS2 ≤ 1/3 for at least one of the six 2-subsets of {1, 2, 3, 4}.
Proof. For this proof we will write
x 1 x2 x3 x4
.
Φt =
y1 y2 y3 y4
(16)
Our proof is by contradiction. Suppose that the shortest non-zero vector in all ΛS has square
length r2 > 1/3. Now x2i < r2 /2, in at most one position i, else the shortest would be smaller
than r2 . The same is true for yi2 , (xi − yi )2 and (xi + yi )2 . Hence there exists one position i,
w.l.o.g. i = 1, such that x2i ≥ r2 /2 , yi2 ≥ r2 /2 and (xi − yi )2 ≥ r2 /2. It follows that (i) x1 and
y1 must be of opposite sign and (ii) (x1 + y1 )2 < r2 /2. (i) is true because if x1 , y1 are of the
√
√
√
same sign and (x1 − y1 )2 ≥ r2 /2 then |y1 | ≥ 2r or |x1 | ≥ 2r. Assuming x1 ≥ 2r, we have
r2 ≤ x23 + x24 = 1 − x21 − x22 ≤ 1 − 2r2 which contradicts the hypothesis that r2 > 1/3. (ii) is
√
√
true for if not |x1 | ≥ 2r or |y1 | ≥ 2r, and by the same argument as in (i) r2 cannot exceed
1/3.
Since (x1 + y1 )2 < r2 /2, it follows that (xi + yi )2 ≥ r2 /2 for positions i = 2, 3, 4. In at least
one of these positions, say i = 2, x2i ≥ r2 /2 and yi2 ≥ r2 /2. Now x2 and y2 must be of the
same sign and (x2 − y2 )2 < r2 /2, by a proof similar to that used before. Thus for i = 3, 4,
(xi + yi )2 ≥ r2 /2 and (xi − yi )2 ≥ r2 /2. Again for i = 3, 4, both x2i ≥ r2 /2 and yi2 ≥ r2 /2
cannot hold for the same i, hence, either x23 < r2 /2, y32 ≥ r2 /2 and x24 ≥ r2 /2, y42 < r2 /2 or
x24 < r2 /2, y42 ≥ r2 /2 and x23 ≥ r2 /2, y32 < r2 /2. We assume the first case. The proof for the
other case is similar.
We have already proved that x2 and y2 are of the same sign. Assume they are both positive
(if not reverse signs of all elements of Φ). Further, assume that y2 > x2 and consider positions
i = 2, 3 (if x2 > y2 , then the same proof applies but for positions i = 2, 4). We now break up
the proof into two cases:
Case 1: (y3 and x3 of the same sign): Either (a) (y3 − x3 )2 ≥ r2 or (b) (y3 − x3 )2 < r2 . If (a)
then |y3 | − |x3 | ≥ r and y22 + y32 ≥ x22 + (|x3 | + r)2 ≥ x22 + x23 + r2 ≥ 2r2 . Thus r2 ≤ y12 + y42 =
10
1 − (y22 + y33 ) ≤ 1 − 2r2 , hence r2 ≤ 1/3.
If (b) then let (y3 − x3 )2 = r2 − 2 (0 < 2 ≤ r2 /2). Thus |y3 | − |x3 | =
√
r2 − 2 and it follows
that y32 ≥ x23 + r2 − 2 . Since (y2 − x2 )2 + (y3 − x3 )2 ≥ r2 it follows that (y2 − x2 )2 ≥ 2 and
that |y2 | ≥ |x2 | + and thus y22 ≥ x22 + 2 . Thus y22 + y32 ≥ x22 + 2 + x23 + r2 − 2
= x22 + x23 + r2 ≥ 2r2 . But then r2 ≤ y12 + y42 = 1 − (y22 + y32 ) ≤ 1 − 2r2 and it follows that
r2 ≤ 1/3.
Case 2: (y3 and x3 are of opposite signs): Either (a) (y3 + x3 )2 ≥ r2 or (b) (y3 + x3 )2 < r2 .
If (a) then |y3 + x3 | ≥ r and due to opposite signs |y3 | − |x3 | ≥ r from which y32 ≥ x23 + r2 . It
follows that y22 + y32 ≥ x22 + x23 + r2 ≥ 2r2 . This implies r2 ≤ 1/3. If (b) then let
(y3 + x3 )2 = r2 − 2 . Since y3 and x3 have opposite signs (|y3 | − |x3 |)2 = r2 − 2 and thus
y32 ≥ x23 + r2 − 2 . Since (y2 + x2 )2 + (y3 + x3 )2 ≥ r2 it follows that (y2 + x2 )2 + r2 − 2 ≥ r2
and hence (y2 + x2 )2 ≥ 2 which implies that y22 ≥ x22 + 2 . Thus
y22 + y32 ≥ x22 + 2 + x23 + r2 − 2 ≥ 2r2 . Once again this means r2 ≤ 1/3.
C. (4, 3)
It is checked by direct evaluation that for
a
t
Φ = a
a
the four child lattices have Gram matrices
1/4
3/4 1/4
1/4 3/4 −1/4 ,
1/4 −1/4 3/4
3/4 1/4 −1/4
1/4 3/4 1/4
−1/4 1/4 3/4
a = 1/2, ΛV = Z3 and
−a −a −a
a −a a ,
−a a
a
(17)
3/4 −1/4 1/4
−1/4 3/4 1/4
1/4
1/4 3/4
3/4 −1/4 −1/4
−1/4 3/4 −1/4 .
−1/4 −1/4 3/4
√
All child lattices are congruent to the body-centered cubic lattice with packing density π 3/8.
This construction achieves the trace bound r2 = 3/4 and is optimal.
11
Set ΛV to be the face-centered cubic lattice, the densest lattice packing in R3 , packing density
√
∆V = π/ 18 = 0.7408, so that det ΛV = 4,
0
−1 1
2 0 1
V = −1 −1 1 , G(ΛV ) = 0 2 1 .
(18)
0
0 −1
1 1 2
With Φ such that
1
0 0
2/3 1 1
Φ∗V =
,
−2/3 1 0
1/3 0 1
(19)
we get β 2/k = (ρmin /ρ(ΛV ))2 = 0.5. Each child lattice has unit determinant and shortest vector
of length equal to unity. All four child lattices have an identical packing density of π/6, which
is the packing density of the cubic lattice Z3 . Also observe that the determinant bound (4) holds
with equality, which implies that the selected Φ is optimal among rotations for which all child
lattices achieve the packing density of the cubic lattice.
√
Let ΛV the body-centered cubic lattice, ∆(ΛV ) = π 3/8, ρ2 = 3/4,
3 −1 −1
1 −1 1
V = −1 1 1 , G(ΛV ) = −1 3 −1 .
−1 −1 3
−1 −1 1
√
√
Upon setting a = 1/2, b = a + a, c = a − a and
√
a a
0
−c c
b
√
0
−c b
a − a
c
Φ=
,Φ ∗ V =
√
0
−b c
a
a
b
√
a −a
0
b −b −c
we obtain child lattices with Gram matrices given in terms of d = 2c2 + b2 , e = 2b2 + c2 ,
f = −c2 − 2bc, g = c2 + 2bc, h = 3bc, i = −b2 − 2bc,
d f −g d −g f
f d h , −g e −h ,
−g h e
f −h d
12
e
−g
i
−g d −h ,
i −h e
e
−i
−g
−i e −h .
−g −h d
(20)
√
Each child lattice has determinant 4 and shortest vector with square length d = (9/4 − 1/ 2) =
1.5429 and packing density ∆ = πd3/2 /12 = 0.5017. This results in β 2/k = d/3 = 0.5143.
V. S UMMARY
We considered the design of power-constrained rank-k lattice codes in Rn , with the property
that error free recovery is possible for any k code symbols and the minimum distance for each of
the nk lattices obtained by projection onto k-dim subspaces spanned by k coordinate vectors are
bounded from below. The potential application is to erasure channels with additive noise. Bounds
on the performance are derived and the performance of specific constructions are investigated
for n = 4.
R EFERENCES
[1] D. Blackwell, L. Breiman, and A. Thomasian. The capacity of a class of channels. The Annals of Mathematical Statistics,
pages 1229–1241, 1959.
[2] J. W. S. Cassels. An introduction to the geometry of numbers. Springer Science & Business Media, 2012.
[3] J. H. Conway and N. J. A. Sloane. Sphere Packings, Lattices and Groups. Springer-Verlag, New York, 1988.
[4] T. A. Courtade and R. D. Wesel. Optimal allocation of redundancy between packet-level erasure coding and physical-layer
channel coding in fading channels. IEEE Transactions on Communications, 59(8):2101–2109, 2011.
[5] I. Csiszár and P. Narayan. Capacity of the gaussian arbitrarily varying channel. IEEE Transactions on Information Theory,
37(1):18–26, 1991.
[6] P. M. Gruber and C. G. Lekkerkerker. Geometry of numbers. North-Holland, 1987.
[7] R. A. Horn and C. R. Johnson. Matrix analysis. Cambridge university press, 2012.
[8] A. Ozçelikkale, S. Yuksel, and H. M. Ozaktas. Unitary precoding and basis dependency of mmse performance for gaussian
erasure channels. Information Theory, IEEE Transactions on, 60(11):7186–7203, 2014.
| 7cs.IT
|
ON THE RANKS OF THE THIRD SECANT VARIETY OF
SEGRE-VERONESE EMBEDDINGS
arXiv:1701.06845v2 [math.AG] 17 Jan 2018
EDOARDO BALLICO AND ALESSANDRA BERNARDI
Abstract. We give an upper bound for the rank of the border rank 3 partially symmetric
tensors. In the special case of border rank 3 tensors T ∈ V1 ⊗ · · · ⊗ Vk (Segre case) we can show
that all ranks among 3 and k − 1 arise and if dim Vi ≥ 3 for all i’s, then also all the ranks between
k and 2k − 1 arise.
Introduction
In this paper we deal with the problem of finding a bound for the minimum integer r(T ) needed
to write a given tensor T as a linear combination of r(T ) decomposable tensors. Such a minimum
number is now known under the name of rank of T . In order to be as general as possible we will
consider the tensor T to be partially symmetric, i.e.
T ∈ S d1 V1 ⊗ · · · ⊗ S dk Vk
(1)
where the di ’s are positive integers and Vi ’s are finte dimensional vector spaces defined over an
algebraically closed field K. The decomposition that will give us the rank of such a tensor will be
of the following type:
r(T )
X
⊗dk
⊗d1
⊗ · · · ⊗ vk,i
λi v1,i
(2)
T =
i=1
where λi ∈ K and vj,i ∈ Vj , i = 1, . . . , r(T ) and j = 1, . . . , k.
Another very interesting and useful notion of “ rank ” is the minimum r(T ) such that a tensor
T can be written as a limit of a sequence of rank r(T ) tensors. This last integer is called the border
rank of T (Definition 1.5) and clearly it can be strictly smaller than the rank of T (Remark 1.6).
It has become a common technique to fix a class of tensors of given border rank and then study all
the possible ranks arising in that family (cf. [7, 3, 10, 14, 6]). The rank of tensors of border rank 2
is well known (cf. [7] for symmetric tensors, [2] for tensors without any symmetry, [4] for partially
symmetric tensors). The first not completely classified case is the one of border rank 3 tensors.
In [7, Theorem 37] the rank of any symmetric order d tensor of border rank 3 has been computed
and it is shown that the maximum rank reached is 2d − 1. In the present paper, Theorem 1.7, we
prove that the rank of partially symmetric tensors T as in (1) of border rank 3 can be at most
r(T ) ≤ −1 +
k
X
2di .
i=1
In [10, Theorem 1.8] J. Buczyński and J.M. Landsberg described the cases in which the inequality
in Theorem 1.7 is an equality: when k = 3 and d1 = d2 = d3 = 1 they show that there is an
element of rank 5. All ranks for border rank 3 partially symmetric tensors are described in [9]
when k = 3, d1 = d2 = d3 = 1 and ni = 1 for at least one integer i. Therefore our Theorem 1.7
is the natural extension of the two extreme cases (tensors without any symmetry where di = 1 for
all i = 1, . . . , k and totally symmetric case where k = 1).
In the special case of tensors without any symmetry, i.e. T ∈ V1 ⊗ · · · ⊗ Vk , we will be able to
show, in Theorem 1.8, that all ranks among 3 and k − 1 arise and if dim Vi ≥ 3 for all i’s then also
all the ranks between k and 2k − 1 arise, therefore this result is sharp (cf. Remark 3.11). In the
proof of this theorem we will describe the structure of our solutions: they are all obtained from
(P2 )k taking as a border scheme a degree 3 connected curvilinear scheme. The two critical cases
are rank k − 1 on (P1 )k and rank 2k − 1 on (P2 )k and the other cases can be deduced from one of
these two.
1
2
E. BALLICO AND A. BERNARDI
In [5] we defined the notion of curvilinear rank for symmetric tensors to be the minimum length
of a curvilinear scheme whose span contains a given symmetric tensor. We can extend some of the
ideas in [5] and some of those used in our proof of Theorem 1.7 to the case of partially symmetric
tensors and prove that, if a partially symmetric tensor is contained in the span of aspecial degree c
Pk
curvilinear scheme with α components, the rank of this tensor is bounded by 2α+c −1 + i=1 di
(cf. Theorem 1.11).
1. Notation, Definitions and Statements
In this section we introduce the basic geometric tools that we will use all along the paper.
Notation 1.1. We indicate with
νn11 ,...,n1k : P
n1
× ···× P
nk
M
→ P , where M =
k
Y
!
(ni + 1)
i=1
−1
the Segre embedding of the multi-projective space Pn1 × · · · × Pnk , i.e. the embedding of Pn1 ×
· · · × Pnk by the complete linear system |OPn1 ×···×Pnk (1, . . . , 1)|.
For each i ∈ {1, . . . , k} let
πi : Pn1 × · · · × Pnk → Pni
denote the projection onto the i-th factor.
Let
τi : Pn1 × · · · × Pnk → Pn1 × · · · × P̂ni × · · · × Pnk
denote the projection onto all the factors different from Pni .
Let εi ∈ Nk be the k-tuple of integers εi = (0, . . . , 1, . . . , 0) with 1 only in the i-th position. We
say that a curve C ⊂ Pn1 × · · · × Pnk has multi-degree (a1 , . . . , ak ) if for all i = 1, . . . , k the line
bundle OC (εi ) has degree ai .
We say that a morphism h : P1 → Pn1 × · · · × Pnk has multi-degree (a1 , . . . , ak ) if, for all
i = 1, . . . , k:
h∗ (OPn1 ×···×Pnk (εi )) ∼
= OP1 (ai ).
Let
νnd1 ,...,ndk : P
1
n1
× ··· × P
nk
N
→ P , where N =
k
!
k
Y
di + ni
i=1
ni
−1
denote the Segre-Veronese embedding of Pn1 × · · · × Pnk of multi-degree (d1 , . . . , dk ) and define
X := νnd1 ,...,ndk (Pn1 × · · · × Pnk )
1
k
to be the Segre-Veronese variety.
The name “ Segre-Veronese ” is classically due to the fact that when the di ’s are all equal to 1,
then the variety X is called “ Segre variety ”; while when k = 1 then X is known to be a “ Veronese
variety ”.
Remark 1.2. An element of X is the projective class of a decomposable partially symmetric
tensor T ∈ S d1 V1 ⊗ · · · ⊗ S dk Vk where P(Vi ) = Pni . More precisely p ∈ X if there exists T ∈
⊗dk
⊗d1
⊗ · · · ⊗ vk,i
] with [vj,i ] ∈ Pni .
S d1 V1 ⊗ · · · ⊗ S dk Vk such that p = [T ] = [v1,i
Definition 1.3. The s-th secant variety σs (X) of X is the Zariski closure of the union of all
s-secant Ps−1 to X. The tangential variety τ (X) is the Zariski closure of the union of all tangent
lines to X.
Observe that
X = σ1 (X) ⊂ τ (X) ⊂ σ2 (X) ⊂ · · · ⊂ σs−1 (X) ⊂ σs (X) ⊂ · · · ⊂ PN .
(3)
THIRD SECANT VARIETY OF SEGRE-VERONESE
3
Definition 1.4. The X-rank rX (p) of an element p ∈ PN is the minimum integer s such that
there exist a Ps−1 ⊂ PN which is s-secant to X and containing p.
We indicate with S(p) the set of sets of points of Pn1 × · · · × Pnk “ evincing ” the X-rank of
p ∈ PN , i.e.
S(p) :=
{x1 , . . . , xs } ⊂ P
n1
× ··· ×P
nk
| rX (p) = s and p ∈ hνnd1 ,...,ndk (x1 ), . . . , νnd1 ,...,ndk (xs )i .
1
1
k
k
Definition 1.5. The X-border rank brX (p) of an element p ∈ PN is the minimum integer s such
that p ∈ σs (X).
Remark 1.6. For any p ∈ PN = P(S d1 V1 ⊗ · · · ⊗ S dk Vk ) we obviously have that brX (p) ≤ rX (P ).
In fact p ∈ PN of rank r is such that there exist a tensor T ∈ S d1 V1 ⊗ · · · ⊗ S dk Vk that can be
minimally written as in (2); while an element p ∈ PN has border rank s if and only if there exist a
sequence of rank r tensors Ti ∈ S d1 V1 ⊗ · · · ⊗ S dk Vk such that limi→∞ Ti = T and p = [T ].
The first result that we prove in Section 2 is an upper bound for the rank of points in σ3 (X).
Pk
Theorem 1.7. The rank of an element p ∈ σ3 (X) is rX (p) ≤ −1 + i=1 2di .
In the case di = 1 for all i = 1, . . . , k, i.e. if X is the Segre variety, we fill in all low ranks with
points of σ3 (X) \ σ2 (X). In Section 3 we prove the following result.
Theorem 1.8. Assume k ≥ 3 and let X ⊂ PM be the Segre variety of k factors as in Notation
1.1. Let α be the cardinality of the set {i ∈ {1, . . . , k} | ni ≥ 2}. For each x ∈ {3, . . . , α + k − 1}
there is p ∈ σ3 (X) \ σ2 (X) with rX (p) = x.
Remark 1.9. If α = k, i.e. if ni ≥ 2 for all i’s, Theorems 1.7 and 1.8 give the ranks of the points
of σ3 (X) \ σ2 (X). In Remark 3.11 we discuss the reason why we do not know the rank of a specific
p ∈ σ3 (X) \ σ2 (X).
Moreover, in the case of Segre varieties where factors have dimension ni ≥ 2, Theorem 1.8 says
that all ranks from 3 to 2k − 1 can be attained. Therefore the above result is sharp.
As remarked in the Introduction, we can extend some of the ideas of [5] on the notion of
curvilinear rank to some tools used in our proof of Theorem 1.7 to the case of partially symmetric
tensors.
Definition 1.10. A scheme Z ⊂ Pn1 × · · · × Pnk is curvilinear if it is a finite union of disjoint
schemes of the form OCi ,Pi /mepii for smooth points pi ∈ Pn1 × · · · × Pnk on reduced curves Ci ⊂
Pn1 ×· · ·×Pnk . Equivalently Z is curvilinear if the tangent space at each of its connected component
supported at the pi ’s has Zariski dimension ≤ 1. We define the curvilinear rank Cr(p) of a point
p ∈ PN as:
o
n
Cr(p) := min deg(Z) | νnd1 ,...,ndk (Z) ⊂ X, Z curvilinear, p ∈ hνnd1 ,...,ndk (Z)i .
1
k
1
k
In Section 4 we prove the following result.
Theorem 1.11. If there exists a degree c curvilinear scheme Z ⊂ Pn1 × · · · × Pnk such that p ∈
hνnd1 ,...,ndk (Z)i and Z has α connected components, each of them mapped by νnd1 ,...,ndk into a
1
k
1 Pk
k
linearly independent zero-dimensional sub-scheme of PN , then rX (p) ≤ 2α + c −1 + i=1 di .
2. Proof of Theorem 1.7
Remark 2.1. Fix a degree 3 connected curvilinear scheme E ⊂ P2 not contained in a line and
a point u ∈ P1 . The scheme E is contained in a smooth conic. Hence there is an embedding
f : P1 → P2 with f (P1 ) = C and f (3u) = E.
Remark 2.2. For any couple of points u, o ∈ P1 , there is an isomorphism f : P1 → P1 such that
f (u) = o. For any such f we have f (3u) = 3o.
Remark 2.3. Fix two points u, o ∈ P1 . There is a morphism f : P1 → P1 such that f (u) = o, f
is ramified at u and deg(f ) = 2, i.e. f ∗ (OP1 (1)) ∼
= OP1 (2). Since deg(f ) = 2, f has only order 1
ramification at u. Thus f (3u) = 2o (as schemes).
4
E. BALLICO AND A. BERNARDI
We recall the following lemma proved in [4, Lemma 3.3].
Lemma 2.4 (Autarky). Let p ∈ hXi with X being the Segre-Veronese variety of Pn1 × · · · × Pnk
embedded in multidegree (d1 , . . . , dk ). If there exist Pmi , i = 1, . . . , k, with mi ≤ ni , such that
p ∈ hνmd1 ,...,mdk (Pm1 × · · · × Pmk )i, then the X-rank of p is the same as the Y -rank of p where Y
1
k
is the Segre-Veronese νmd1 ,...,mdk (Pm1 × · · · × Pmk ).
1
k
Corollary 2.5. Let Γ ⊂ Pn1 × · · · × Pnk be a 0-dimensional scheme of minimal degree such that
p ∈ hνnd1 ,...,ndk (Γ)i, then the X-rank of p is equal to its Y -rank where Y is the Segre-Veronese
1
k
embedding of Pm1 × · · · × Pmk where each mi = dimhπi (Γ)i − 1 ≤ deg(πi (Γ)) − 1 (πi as in Notation
1.1). If there exists an index i such that deg(πi (Γ)) = 1, then we can take Y to be the SegreVeronese embedding of Pm1 × · · · × P̂mi × · · · × Pmk .
Proof. Consider the projections πi : Pn1 × · · · × Pnk onto the i-th factor Pni as in Notation 1.1. It
may happen that deg(πi (Γ)) can be any value from 1 to deg(Γ).
By the just recalled Autarky Lemma (cf. Lemma 2.4), we may assume that each πi (Γ) spans
the whole Pni . Therefore if there is an index i ∈ {1, . . . , k} such that deg(πi (Γ)) = 1 we can take
p ∈ Pn1 × · · · × P̂ni × · · · × Pnk . Moreover the autarkic fact that we can assume Pni to be hπi (Γ)i
implies that we can replace each Pni with Pdimhπi (Γ)i−1 and clearly dimhπi (Γ)i ≤ deg(πi (Γ)).
Proof of Theorem 1.7: Because of the filtration of secants varieties (3), for a given element p ∈
σ3 (X), it may happen that either p ∈ X, or p ∈ σ2 (X) \ X or p ∈ σ3 (X) \ σ2 (X). We distinguish
among these cases.
(1) If p ∈ X, then rX (p) = 1.
(2) If p ∈ σ2 (X)\ X then either p lies on a honest bisecant line to X (and in this case obviously
rX (p) = 2) or p belongs to certain tangent line to X. In this latter case, the minimum
number h ≤ k of factors containing such a tangent line is the minimum integer such that
p ∈ hνnd1 ,...,ndh (Pn1 ×· · ·×Pnh )i (maybe reordering factors). In [4, Theorem 3.1] we proved
1
h
P
that, if this is the case, then rX (p) = hi=1 di .
(3) From now on we assume that p ∈ σ3 (X) \ σ2 (X). By [10, Theorem 1.2] there is short list of
zero-dimensional schemes Γ ⊂ Pn1 × · · · × Pnk such that p ∈ hνnd1 ,...,ndk (Γ)i, therefore, in
1
k
order to prove Theorem 1.7, it is sufficient to bound the rank of the points in hνnd1 ,...,ndk (Γ)i
1
k
for each Γ in their list.
Since p ∈ σ3 (X) \ σ2 (X), The possibilities for Γ are only the following: either Γ is a
smooth degree 3 zero-dimensional scheme (case (3a) below), or it is the union of a degree
2 scheme supported at one point and a simple point (case (3b)), or it is a curvilinear
degree 3 scheme (case (3c)) or, finally, a very particular degree 4 scheme with 2 connected
components of degree 2 (case (3d)).
(3a) If Γ is a set of 3 distinct points, then obviously rX (p) = 3 ([10, Case (i), Theorem
1.2]).
(3b) If Γ is a disjoint union of a simple point a and a degree 2 connected scheme ([10,
Case (ii), Theorem 1.2]), then there is a point q on a tangent line to X such that
Pk
p ∈ h{νnd1 ,...,ndk (a), q}i. Hence rX (p) ≤ 1 + rX (q) ≤ 1 + i=1 di (for the rank
1
k
on the tangential variety of X see [2]). Since di > 0 for all i’s and k ≥ 2, then
Pk
Pk
1 + i=1 di ≤ −1 + i=1 2di .
(3c) Assume deg(Γ) = 3 and that Γ is connected ([10, Case (iii), Theorem 1.2]) supported
at a point {o} := Γred . Since the case k = 1 is true by [7, Theorem 37], we can prove
the theorem by using induction on k, with the case k = 1 as the starting case.
Since deg(Γ) = 3, by Corollary 2.5, we can assume that p belongs to a Segre-Veronese
variety of k factors all of them being either P1 ’s or P2 ’s, i.e., after having reordered
the factors,
p ∈ ν1d1 ,...,1da ,2da+1 ,...2dk (P1 )a × (P2 )b .
The P1 ’s correspond to the cases in which either deg(πi (Γ)) = 3 and dimhπi (Γ)i = 1
(i.e. πi (Γ) is contained in a line of the original Pni ), or deg(πi (Γ)) = 2 (notice that
in this case πi|Γ is not an embedding). The P2 ’s correspond to the cases in which
THIRD SECANT VARIETY OF SEGRE-VERONESE
5
dimhπi (Γ)i = 2, deg(πi (Γ)) = 3. Finally we can exclude all the cases in which
deg(πi (Γ)) = 1 because, again by Corollary 2.5, we would have that p belongs to a
Segre-Veronse variety of less factors and then this won’t give the highest bound for
the rank of p.
Now fix a point u ∈ P1 . By Remarks 2.1, 2.2 and 2.3 there is
fi : P1 → Pni with fi (3u) = πi (Γ).
(4)
Consider the map
f = (f1 , . . . , fk ) : P1 → Pn1 × · · · × Pnk .
We have f (u) = {o} and πi (f (3u)) = fi (3u) = πi (Γ). Since πi (f (3u)) = πi (Γ) for all
i’s, the universal property of products gives f (3u) = Γ. The map f has multi-degree
(a1 , . . . , ak ) where ai = 1 if ni = 1 and deg(πi (Γ)) = 3, and ai = 2 in all other
cases. Notice that fi is an embedding if deg(πi (Γ)) 6= 2. Since deg(πi (Γ)) = 2 if and
only if πi−1 (oi ) contains the line spanned by the degree 2 sub-scheme of Γ, we have
deg(πi (Γ)) = 2 for at most one index i. Since k ≥ 2, f is an embedding. Set
C := νnd1 ,...,ndk f P1 and Z := νnd1 ,...,ndk (Γ).
1
1
k
k
Pk
P
The curve C is smooth and rational of degree δ := i=1 ai di . Note that δ ≤ ki=1 2di .
Hence to prove Theorem 1.7 in this case it is sufficient to show that rC (p) ≤ δ − 1
because clearly rC (p) ≥ rX (p) since C ⊂ X.
By assumption p ∈ hZi. Since p ∈
/ σ2 (X), hZi is not a line of PN . Hence hZi is a
plane because deg(Z) = deg(Γ) = 3. Since C is a degree δ smooth rational curve,
we have dimhCi ≤ δ. By [14, Proposition 5.1] we have rC (p) ≤ dimhCi. Hence it
is sufficient to prove the case δ = dimhCi, i.e. we may assume that C is a rational
normal curve in its linear span.
If δ ≥ 4, since Z is connected and of degree 3, by Sylvester’s theorem (cf. [11]) we
have p has C-border rank 3 and rC (p) = δ − 1, concluding the proof in this case.
If δ ≤ 3, we have σ2 (C) = hCi and hence p ∈ σ2 (X), contradicting p ∈ σ3 (X) \
σ2 (X).
(3d) Assume that Γ has degree 4 ([10, Case (iv), Theorem 1.2]). J. Buczyński and J.M.
Landsberg show that p belongs to the span of two tangent lines to X whose set
theoretic intersections with X span a line which is contained in X. This means that
Γ = v⊔w with v, w being two degree 2 reduced zero-dimensional schemes with support
contained in a line L ⊂ Pn1 × · · · × Pnk and moreover that the multi-degree of L is
εi for some i = 1, . . . , k (cfr. Notation 1.1). This case occurs only when di = 1, i.e.
when νnd1 ,...,ndk (L) = νn11 ,...,n1k (L) = L̃ is a line.
1
k
Observe that ṽ := νnd1 ,...,ndk (v) and w̃ := νnd1 ,...,ndk (w) are two tangent vectors to
1
1
k
k
X. In [2, Theorem 1] we prove that the X-rank of a point p ∈ To (X) for a certain
point o = (o1 , . . . , ok ) ∈ X, is the minimum number ηX (p) for which there exist
E ⊆ {1, . . . , k} such that ♯(E) = ηX (p) and To (X) ⊆ h∪i∈E Yo,i i where Yo,i is the
ni -dimensional linear subspace obtained by fixing all coordinates j ∈ {1, . . . , k} \ {i}
equal to oj ∈ Pni . Let I and J be the sets playing the role of E for hṽi and hw̃i
respectively and set I ′ = I \ {i} (meaning that I ′ = I if i ∈
/ I and I ′ = I \ {i}
′
otherwise) and J = J \ {i} . Now take
X
X
α :=
dj +
dj + di
j∈I ′
j∈J ′
Pk
and note that α ≤ −1 + h=1 2dh , therefore if we prove that rX (p) ≤ α we are done.
Let Dj ⊂ Pn1 × · · · × Pnk , j ∈ I ′ , be the line of multi-degree εJ containing πj (v),
′
andlet Tj , j ∈
εjcontaining πj (w). The curve
J , be the line of X of multi-degree
S
S
L∪
j∈J ′ Tj contains w. Hence the curve
j∈I ′ Dj contains v and the curve L ∪
[
[
T := L ∪
Dj ∪
Tj
j∈I ′
j∈J ′
6
E. BALLICO AND A. BERNARDI
is a reduced and connected curve containing Γ. Since p ∈ hνnd1 ,...,ndk (Γ)i, we have
1
k
that if we call T̃ := νnd1 ,...,ndk (T ) then p ∈ hT̃ i and rX (p) ≤ rT̃ (p). The curve T̃
1
k
is a connected curve whose irreducible components are smooth rational curves and
with deg(T̃ ) = α. Hence dimhT̃ i ≤ α. Since T̃ is reduced and connected, as in [14,
Proposition 4.1] and in [11], we get rT̃ (p) ≤ α. Summing up rX (p) ≤ rT̃ (p) ≤ α ≤
P
−1 + kh=1 2dh .
3. Proof of Theorem 1.8
Autarky Lemma (proved in [4, Lemma 3.3] and recalled here in Lemma 2.4) is true also for the
border rank ([9, Proposition 2.1]). This allows to formulate the analog of Corollary 2.5 for border
rank. Therefore, in order to prove Theorem 1.8 and x ≤ k − 1, we can limit ourselves to the study
of the case ni = 1 for all i’s. This is the reason why in the first part of this section we will always
work with the Segre variety of P1 ’s. Let
ν1(k) : (P1 )k → Pr , r = 2k − 1
1
(5)
1 k
be the Segre embedding of k copies of P ’s and X := ν1(k) ((P ) ); and let
′
ν1(k−1) : (P1 )k−1 → Pr , r′ = 2k−1 − 1
1
′
(6)
1 k−1
be the the Segre embedding of k − 1 copies of P ’s and X := ν1(k−1) ((P )
).
Proposition 3.1. Assume k ≥ 3. Let Γ ⊂ (P1 )k be a degree 3 connected curvilinear scheme
such that deg(πi (Γ)) = 3 for all i’s, and let β be the only degree 2 sub-scheme of Γ. For all
p ∈ hν1(k) (Γ)i \ hν1(k) (β)i we have that
(a) if k = 3, then 2 ≤ rX (p) ≤ 3 and rX (p) = 2 if p is general in hν1(k) (Γ)i;
(b) if k ≥ 4, then rX (p) = k − 1.
Proof. Since Γ ⊂ (P1 )k is connected, it has support at only one point; all along this proof we set
o := Supp(Γ) ∈ (P1 )k .
(7)
First of all recall that in step (3a) of the proof of Theorem 1.7 we obtained an embedding
f = (f1 , . . . , fk ) with fi : P1 → P1 an isomorphism (see (4)); moreover we can fix a point u ∈ P1
such that f (u) = o and Γ = f (3u). We proved that
C := ν1(k) (f (P1 ))
is a degree k rational normal curve in its linear span. Obviously
rX (p) ≤ rC (p).
If k ≥ 4 Sylvester’s theorem implies rC (p) = k − 1.
Now assume k = 3. Since a degree 3 rational plane curve has a unique singular point, for any
q ∈ hCi there is a unique line L ⊂ hCi = P3 with deg(L∩C) = 2. Thus rC (p) = 2 (resp. rC (p) = 3)
if and only if p ∈
/ τ (C) (resp. p ∈ τ (C), cfr. Definition 1.3). Since τ (C) is a degree 4 surface, by
Riemann-Hurwitz, we see that both cases occur and that rC (p) = 2 (and hence rX (p) = 2 if p is
general in hν1(k) (Γ)i).
Claim 1. Let the point o ∈ (P1 )k be, as in (7), the support of Γ. Fix any F ∈ |O(P1 )k (εk )| such
that o ∈
/ F . Then hν1(k) (Γ)i ∩ hν1(k) (F )i = ∅.
Proof of Claim 1. It is sufficient to show that h0 (IF ∪Γ (1, . . . , 1)) = h0 (IF (1, . . . , 1)) − 3, i.e.
h0 (IΓ (1, . . . , 1, 0)) = h0 (O(P1 )k (1, . . . , 1, 0)) − 3. This is true because f1 , . . . , fk−1 (recalled at
the beginning of the proof this Proposition 3.1 and introduced in (4)) are isomorphisms.
(a) Assume k = 3. Since rX (p) ≤ rC (p) ≤ 3 and rC (p) = 2 for a general p in hν1(3) (Γ)i, we only
need to prove that rX (p) > 1. The case rX (p) = 1 corresponds to a completely decomposable
tensor: p = ν1(3) (q) for some q ∈ (P1 )3 . Clearly rX (ν1(3) (o)) = 1 but o ∈ hβi then, since we
took p ∈ hν1(3) (Γ)i \ hν1(3) (β)i, we have p 6= ν1(3) (o) and in particular q 6= o. In this case we
can add q to Γ and get that h1 (Iq∪Γ (1, 1, 1)) > 0 by [1, Lemma 1]. Since deg(fi (Γ)) = 3,
for all i’s, every point of hβi \ {o} has rank 2. Since q := (q1 , q2 , q3 ) 6= o := (o1 , o2 , o3 )
THIRD SECANT VARIETY OF SEGRE-VERONESE
7
we have qi 6= oi for some i, say for i = 3. Take F ∈ |O(P1 )3 (ε3 )| such that q ∈ F and
o∈
/ F . Hence F ∩ (Γ ∪ {q}) = {q}. We have h1 (F, Iq,F (1, 1, 1)) = 0, because O(P1 )3 (1, 1, 1) is
spanned. Claim 1 gives h1 (IΓ (1, 1, 0)) = 0. The residual exact sequence of F in (P1 )3 gives
h1 (IΓ∪{q} (1, 1, 1)) = 0, a contradiction.
(b) From now on we assume k ≥ 4 and that Proposition 3.1 is true for a smaller number of factors.
Since X ⊃ C, we have rX (p) ≤ k − 1 (in fact, as we already recalled above, rC (p) = k − 1
by Sylvester’s theorem). We need to prove that we actually have an equality, so we assume
rX (p) ≤ k − 2 and we will get a contradiction.
Take a set of points S ∈ S(p) of (P1 )k evincing the X-rank of p (see Definition 1.4) and
consider v = (v1 , . . . , vk ) ∈ S ⊂ (P1 )k to be a point appearing in a decomposition of p. We can
always assume that, if o = (o1 , . . . , ok ), then vk 6= ok : such a v ∈ S ⊂ S(p) exists because, by
Autarky (here recalled in Lemma 2.4), no element of S(p) is contained in (P1 )k−1 × {ok }.
Consider the pre-image
D := πk−1 (vk ).
Clearly by construction o ∈
/ D hence for any q ∈ (P1 )k \ D we have h1 (Iq∪D (1, . . . , 1)) =
1
h (Iq (1, . . . , 1, 0)) = 0, because O(P1 )k (1, . . . , 1, 0) is globally generated. This implies that
hν1(k) (D)i intersects X only in ν1(k) (D).
Now consider
k
k−1
ℓ : P2 −1 \ hν1(k) (D)i → P2 −1
the linear projection from hν1(k) (D)i. Since p ∈
/ hν1(k) (D)i (Claim 1), ℓ is defined at p. Moreover
the map ℓ induces a rational map ν1(k) ((P1 )k \ D) → ν1(k−1) ((P1 )k−1 ) which is induced by the
projection τk : (P1 )k → (P1 )k−1 defined in Notation 1.1. We have
ℓ ◦ ν1(k) = ν1(k−1) ◦ τk .
Since o ∈
/ D, we have ℓ(hΓi) = hν1(k−1) (Γ′ )i, where Γ′ = τk (Γ). Hence p′ := ℓ(p) ∈ hν1(k−1) (Γ′ )i.
By [2] every element of hν1(k−1) (β)i \ ν1(k−1) (o′ ), with o′ := τk (o), has X ′ -rank k − 1. Since
deg(πi (Γ)) = 3 for all i’s, we have deg(πi (β)) = 2 for i = 1, . . . , k − 1. This implies that the
minimal sub-scheme α of Γ′ such that p′ ∈ hν1(k−1) (α)i is such that α 6= β where β is the
degree 2 sub-scheme of Γ′ . Now let S ′ ⊂ (P1 )k−1 be the projection by τk of the set of points of
S ⊂ S(p) that are not in D, i.e. S ′ := τk (S \ S ∩ D). Since ♯(S ′ ) ≤ k − 2 and p′ ∈ hν1(k−1) (Γ′ )i,
the inductive assumption gives α 6= Γ′ (it works even when k = 4). Hence α = {o′ }. Thus
p ∈ hν1(k) ({o} ∪ D)i. Hence dim(hν1(k) (Γ ∪ D)i) ≤ dim(hν1(k) (D)i) + 2, contradicting Claim 1.
We need the following lemma, which is the projective version of an obvious linear algebra
exercise.
Lemma 3.2. Fix two linear spaces L1 ( L2 ⊂ Pm and a finite set E ⊂ L2 spanning L2 . Let
ℓ : Pm \ L1 → Pz , z := m − 1 − dim L1 , be the linear projection from L1 . Then ℓ(L2 \ L1 ) is a
linear space spanned by the set ℓ(E \ E ∩ L1 ).
Notation 3.3. Fix (a, b) ∈ N2 \ {(0, 0)}. Let ∆a,b be the set of all pairs (f, o), where o ∈ P1 ,
f : P1 → (P2 )a × (P1 )b , each πi ◦ f , 1 ≤ i ≤ a, is a degree 2 embedding and, for a + 1 ≤ i ≤ b,
πi ◦ f is an isomorphism.
Lemma 3.4. Set G̃ = Aut(P2 )a ×Aut(P1 )b , G := G̃×Aut(P1 ). Let G acts on ∆a,b via (g, h)(f, o) =
(g ◦ f ◦ h−1 , h(o)). Then this action is transitive, i.e., for (f, o), (f ′ , o′ ) we have (g, h) ∈ G such
that h(o) = o′ and g ◦ f ◦ h−1 = f ′ .
Proof. Fix any h ∈ Aut(P1 ) such that h(o) = o′ and write f˜ := f ◦ h−1 .
′
′
Write f˜ = (f˜1 , . . . , f˜a , f˜a+1 , . . . , f˜a+b ) and f ′ = (f1′ , . . . , fa′ , fa+1
, . . . , fa+b
) with f˜i := πi ◦ f˜ and
′
′
fi := πi ◦ f . We need to find g = (g1 , . . . , ga , ga+1 , . . . , ga+b ) ∈ G̃ such that g ◦ f˜ = f ′ , i.e. by the
universal property of maps to products, we need to find g = (g1 , . . . , ga , ga+1 , . . . , ga+b ) ∈ G̃ such
that gi ◦ f˜i = fi′ for all i.
−1
If a + 1 ≤ i ≤ a + b take gi := fi′ ◦ f˜i .
Now we fix i such that 1 ≤ i ≤ a. We have two degree 2 embeddings fi′ : P1 → P2 and
f˜i : P1 → P2 . Any two such maps are equivalent, up to an automorphism of P2 , because these
8
E. BALLICO AND A. BERNARDI
embeddings are induced by the complete linear system of the anticanonical line bundle of P1 . Thus
there is gi ∈ Aut(P2 ) such that gi ◦ f˜i = fi′ .
Notation 3.5. Take Y = (P2 )a × (P1 )b and let ν2(a) ,1(b) : Y → PN , N := 3a 2b − 1, be the Segre
embedding of Y . Let Γa,b (resp. Γ′a,b ) be the set of all p ∈ PN , such there is (f, o) ∈ ∆a,b with
p ∈ hν2(a) ,1(b) (f (3o))i (resp. and p ∈
/ hν2(a) ,1(b) (f (2o))i).
Since the image of an algebraic set by a morphism is constructible, Γa,b and Γ′a,b of Notation
3.5 are constructible sets. The closure of Γa,b in PN is irreducible. Therefore we are allowed to
inquire about the rank of a general element of Γa,b . If either a > 0 or b ≥ 2, then Γ′a,b 6= ∅ and the
closures in PN of Γa,b and Γ′a,b are the same.
Lemma 3.6. For all k ≥ 3 we have rX (p) = 2k − 1 for a general p ∈ Γk,0 as in Notation 3.5.
Proof. We use induction on k, the case k = 3 being true by [10, Theorem 1.8].
Now assume k ≥ 4. Call ν2(k) : (P2 )k → Pr , r := 3k − 1, the Segre embedding. Fix a ∈ P1 . For
each 1 ≤ i ≤ k let fi : P1 → P2 be a degree 2 embedding. Let f = (f1 , . . . , fk ) : P1 → (P2 )k
be the embedding with fi = πi ◦ f for all i. As in step (3c) of the proof of Theorem 1.7 we see
that the curve C := ν2(k) (f (P1 )) is a rational normal curve of degree 2k in its linear span. Fix
a ∈ P1 and set o := (o1 , . . . , ok ) := f (a) and A := f (3a). The scheme ν2(k) (A) has degree 3 and it
is curvilinear. Fix a general p ∈ hν2(k) (A)i \ hν2(k) (2o)i. Since p has border rank 3 with respect to
the rational normal curve C, Sylvester’s theorem gives rC (p) = 2k − 1. Hence rX (p) ≤ 2k − 1. To
prove the lemma for the integer k it is sufficient to prove that rX (p) ≥ 2k − 1.
Assume rX (p) ≤ 2k − 2 and fix B ∈ S(p).
(a) In this step we assume the existence of a line L ⊂ Pnk such that ok ∈
/ L and ♯(Y ′ ∩ B) ≥ 2,
nk−1
′
′
n1
× L. We have Y ∈ |OY (εk )|. Since ok ∈
/ L, we have o ∈
/ Y′
where Y := P × · · · × P
′
′
′
′
and hence A ∩ Y = ∅. Set B := B \ B ∩ Y . Set A := τk (A) where τk is defined in Notation
1.1. Since k ≥ 3 and (f1 , f2 ) : P1 → P2 × P2 is an embedding, we have deg(A′ ) = 3. Let
ν2(k−1) : (P2 )k−1 → Ps , s = 3k−1 − 1, be the Segre embedding of (P2 )k−1 . Note that the
linear projection from L of P2 \ L sends P2 \ L onto a point. Set E := hν2(k) (Y ′ )i. We have
dim E = 2 · 3k−1 − 1. Let ℓ : PM \ E → Ps denote the linear projection from E. Since
A ∩ Y ′ = ∅, ℓ(ν2(k) (A)) is a well-defined zero-dimensional scheme. Note that ν21 ,21 (f1 , f2 )(P1 )
is not a line of the Segre embedding of P2 × P2 . Since k ≥ 3, we get that ν2(k−1) (A′ ) spans a
plane. Hence ℓ(ν2(k) (A)) = A′ is linearly independent, i.e. hν2(k) (A)i ∩ E = ∅. Hence p′ := ℓ(p)
is well-defined and in particular it is well-defined its rank with respect to the Segre variety
X ′ := ν2(k−1) ((P2 )k−1 ). Since dimhν2(k) (A)i = dimhν2(k) (A′ )i and p is general in hν2(k) (A)i,
p′ is general in hν2(k−1) (A′ )i. By the inductive assumption (case k ≥ 5) or by [10, Theorem
1.8] (case k = 4), we have rX ′ (p′ ) = 2k − 3. Since p ∈ hν2(k) (B)i, Lemma 3.2 applied to
E := ν2(k) (B), m = 3k −1 and L1 = E, gives p′ ∈ hν2(k−1) (B ′ )i. Since ♯(B ′ ) ≤ ♯(B)−2 < 3k−3,
we get a contradiction.
(b) Assume the non-existence of a line L ⊂ Pnk such that ok ∈
/ L and ♯(Y ′ ∩ B) ≥ 2. By Autarky
2 k−1
we have B * (P )
× {ok }. Hence the assumption of this step is equivalent to assuming the
existence of b ∈ B such that πk (b) 6= ok , but πk (B) is contained in the line R ⊂ Pn2 spanned
by ok and πk (b). Hence B ⊂ (P2 )k−1 × R, contradicting Autarky, because nk = 2 and fk (3a)
spans P2 .
Lemma 3.7. Let ν2(1) ,1(1) (Y ) be the Segre embedding of P2 × P1 . We have Γ1,1 * ν2(1) ,1(1) (Y ).
Proof. We have τ (ν2(1) ,1(1) (Y )) ) ν2(1) ,1(1) (Y ). Since a general tangent vector of Y is of the form
f (2o) with (f, o) ∈ ∆1,1 , we get Γ1,1 * ν2(1) ,1(1) (Y ).
Definition 3.8. Let X ⊂ PN be any variety, Z a zero-dimensional scheme and H an effective
Cartier divisor. We define the scheme ResH (Z) ⊂ PN to be the residue scheme of Z with respect
to H, namely the subscheme of PN whose ideal sheaf is IZ : IH .
Lemma 3.9. Take Y = (P2 )2 . For every p ∈ Γ′2,0 we have rX (p) > 2 (cf. Notation 3.5).
THIRD SECANT VARIETY OF SEGRE-VERONESE
9
Proof. Assume the existence of a set B ⊂ Y such that ♯(B) ≤ 2 and p ∈ hν2(2) (B)i. Since
B ∈ S(p), we have p ∈
/ hν2(2) (B ′ )i for any B ′ ( B. Take (f, o) ∈ ∆2,0 such that p ∈ hν2(2) (f (3o))i
and p ∈
/ hf (2o)i. Set A := f (3o). By assumption we have p ∈
/ hν2(2) (A′ )i for any A′ ( A. In
1
particular B 6= {o}. By [1, Lemma 1] we have h (IA∪B (1, 1)) > 0. Since ♯(B) ≤ 2, there is a
line R ⊂ P2 such that π1 (B) ⊂ R. Set H := R × P2 ∈ |OY (1, 0)| and call ν ′ : H → P5 the
Segre embedding of H. We have ResH (A ∪ B) ⊆ A. Since π2 (A) spans P2 by the definition of
Γ2,0 , π2|ResH (A∪B) is an embedding and π2 (A ∪ B) is linearly independent. The residual exact
sequence of H in Y gives h1 (H, IH∩(A∪B),H (1, 1)) > 0. Hence hν ′ (H ∩ A)i ∩ hν ′ (H ∩ B)i 6= ∅.
Since π1 (A) spans P2 , we have A * H. Thus H ∩ A ( A. By the definition of Γ′2,0 we have
p ∈
/ hν2(2) (H ∩ A)i. Set J := hν2(2) (A)i ∩ ν2(2) (Y ). Since the only linear subspaces of ν2(2) (Y )
are the ones contained in a ruling of Y and (f, o) ∈ ∆2,0 , the plane hν2(2) (A)i is not contained in
ν2(2) (Y ). Hence J * hν2(2) (A)i. Since ν2(2) (Y ) is scheme-theoretically cut out by quadrics, J is
cut out by plane conics. Write J = ν2(2) (I) with I ⊂ Y . J is not a reducible conic or a double
line or a line, because πi (A) spans P2 , i = 1, 2, while all linear subspaces of ν2(2) (Y ) are contained
in a ruling of Y . If J were a smooth conic we would have that either π1 (I) spans P2 and π2 (I)
is a point, or π2 (I) spans P2 and π1 (I) is a point or π1 (I) and π2 (I) are lines, contradicting the
assumption that each πi (A) spans P2 . Thus J is a zero-dimensional scheme of degree ≤ 4. Since
A ∪ B ⊆ I, we get that either B = {o} (and we excluded this case) or B = {o, q} for some q ∈ A
with q 6= o. Thus deg(A ∪ B) = 4. We have h1 (IA∪B (1, 1)) 6= 0 ([1, Lemma 1]). Since p has not
rank 2 with respect to ν2(2) (C), we have q ∈
/ C. Thus there is M ∈ |OY (1, 1)| with M ⊃ C and
q∈
/ M . Thus M ∩ (A ∪ B) = A and ResM (A ∪ B) = {q}. Thus h1 (IQ ) = 0. Since h1 (IA (1, 1)) = 0,
the residual exact sequence of M in Y gives a contradiction.
Lemma 3.10. Fix integers a ≥ 0 and b ≥ 0 with a + b ≥ 3. We have rX (p) = 2a + b − 1 for a
general p ∈ Γa,b (cf. Notation 3.5).
Proof. The case a = 0 is true by Proposition 3.1. The case b = 0 is true by Lemma 3.6. Thus we
may assume that a > 0 and b > 0. Set k := a + b. Take (f, o) ∈ ∆a,b such that p is a general
element of hν2(a) ,1(b) (A)i with A := 3o. Set C := f (P1 ), fi := πi ◦ f and oi := πi (f (o)). Since
ν2(a) ,1(b) (C) is a degree 2a + b rational normal curve in its linear span and 2a + b ≥ 4, Sylvester’s
theorem gives rν2(a) ,1(b) (C) = 2a + b − 1. Thus rX (p) ≤ 2a + b − 1. Assume rX (p) ≤ 2a + b − 2
and take B ∈ S(p). By Autarky we have B * (P2 )a × (P1 )b−1 × {ok }. Take z ∈ B such that
bk := πk (z) 6= ok . Set Y ′ := (P2 )a × (P1 )b−1 × {bk }. Let ν2(a) ,1(b−1) := (P2 )a × (P1 )b−1 → Ps ,
s := −1 + 3a 2b−1 , be the Segre embedding of (P2 )a × (P1 )b−1 . Set E := hν2(a) ,1(b) (Y ′ )i. We have
dim E + 1 = 2 · 32 . Let ℓ : PM \ E → Ps the linear projection from E. Set A′ := τk (A) (as in
Notation 1.1). As in the proof of Lemma 3.6 we get E ∩ hν2(a) ,1(b) (A)i = ∅, ν2(a) ,1(b−1) (A′ ) = ℓ(A),
p′ := ℓ(p) is a general element of hν2(a) ,1(b−1) (A′ )i.
(a) Assume (a, b) = (1, 2). Since ν21 ,12 (Y ) * Γ1,2 , p is general in Γ1,2 and ♯(B) ≤ 2, we have
♯(B) = 2. Thus ♯(A′ ) = 1 and so p′ ∈ ν2(1) ,1(1) (P2 × P1 ). Hence a general element of Γ1,1 has
rank 1, contradicting Lemma 3.7.
(b) Assume (a, b) = (2, 1). We use Lemma 3.9.
(c) By the previous steps we may assume a + b ≥ 4, a > 0, b > 0 and use induction on the integer
a + b. (and hence by the inductive assumption applied to (a, b − 1) it has rank 2a + b − 2),
while p′ ∈ hν2(a) ,1(b−1) (B \ B ∩ Y ′ )i with ♯(B \ B ∩ Y ′ ) ≤ x − 2 (because bk ∈ πk (Y ′ ) ∩ πk (B)),
a contradiction.
Proof of Theorem 1.8: First assume x ≤ k − 1. If x = 3, then we may take as p a general point
of σ3 (X). Now assume x ≥ 4 and hence k ≥ 5. Apply Proposition 3.1 to (P1 )x+1 and then use
Autarky (Lemma 2.4). Now assume k ≤ x ≤ 2k − 1. For x = 2k − 1 use Lemma 3.6 and Autarky.
For each x ∈ {4, . . . , 2k − 2} use the case a = x + 1 − k and b = k − a of Lemma 3.10 and then
apply Autarky.
Remark 3.11. Take the set-up of Theorem 1.8. If ni ≥ 2 for all i, then Theorem 1.8 gives all
ranks of points of σ3 (X) \ σ2 (X), but it does not say the rank of each point of σ3 (X) \ σ2 (X).
One problem is that in Lemma 3.6 we do not check all ranks of points of Γ′1,1 . A bigger problem
is that the inductive proof should be adapted and the induction must start. These problems may
10
E. BALLICO AND A. BERNARDI
be not deal-breakers, but there is a class of points of σ3 (X) \ σ2 (X) (occurring even if ni = 1 for
some i) for which we do not have a good upper bound for the rank (except that rX (p) ≤ 2k − 1).
These are the points p ∈ hνn11 ,...,n1k (A)i with A ⊂ Y a connected curvilinear scheme of degree 3
and deg(πi (A)) = 2 for some i, because in this case A * C with C ⊂ Y and νn11 ,...,n1k (C) a rational
normal curve in its linear span. We have no idea about the rank of these points.
4. Proof of Theorem 1.11
Lemma 4.1. Fix an integer c > 0 and u ∈ P1 . Let E = cu ⊂ P1 be the degree c effective divisor of
P1 with u as its support. Let g : E → Pn be any morphism. Then there is a non-negative integer
e ≤ c and a morphism h : P1 → Pn such that h∗ (OPn (1)) ∼
= OP1 (e) and h|E = g.
Proof. Every line bundle on E is trivial. We fix an isomorphism between g ∗ (OPn (1)) and OE (c).
After this identification, g is induced by n + 1 sections u0 , . . . , un of OE (c) such that at least one
of them has a non-zero restriction at {u}. The map H 0 (OP1 (c)) → H 0 (OE (c)) is surjective and
its kernel is the section associated to the divisor cu. Hence there are v0 , . . . , vn ∈ H 0 (OP1 (c)) with
vi|E = ui for all i. Not all sections v0 , . . . , vn vanish at 0. If they have no common zero, then
they define a morphism P1 → Pn extending g and we may take e = c. Now assume that they
have a base locus and call F the scheme-theoretic base locus of the linear system associated to
v0 , . . . , vm . We have deg(F ) ≤ c. Set e := c − deg(F ) and S := Fred . The sections v0 , . . . , vn
induce a morphism f : P1 \ S → Pn with f|E = g. See v0 , . . . , vn as elements of |OP1 (c)| and set
ui := u−F ∈ |OP1 (e)|. By construction the linear system spanned by u0 , . . . , un has no base points,
hence it induces a morphism h : P1 → Pn such that h∗ (OPn (1)) ∼
= OP1 (e). We have h|P1 \S = f
and hence h|E = g.
Proof of Theorem 1.11: Let Z ⊂ Pn1 × · · · × Pnk such that p ∈ hνnd1 ,...,ndk (Z)i and Z has Z1 ,
1
k
. . . , Zα connected components, By assumption there is pi ∈ hνnd1 ,...,ndk (Zi )i such that p ∈
1
k
h{p1 , . . . , pα }i. Note that if Theorem 1.11 is true for each (Zi , pi ), then it is true for Z. Hence it
is sufficient to prove Theorem 1.11 under the additional assumption that Z is connected, so from
now on we assume
• Z connected.
P
Moreover, since rX (p) = 1 ≤ 2 − 1 + i di if c = 1, we may also assume that
• deg Z = c ≥ 2.
Pk
Finally, since the real-valued function x 7→ x −1 + i=1 di is increasing for x ≥ 1, with no loss
of generality we may assume that, for any G ( Z,
• p∈
/ hνnd1 ,...,ndk (G)i.
1
k
Fix u ∈ P1 and let E = cu ⊂ P1 be the degree c effective divisor of P1 with u as its support.
Since Z is curvilinear and deg(Z) = c, we have Z ∼
= E as abstract zero-dimensional schemes.
Let g : E → Pn1 × · · · × Pnk be the composition of an isomorphism E → Z with the inclusion
Z ֒→ Pn1 × · · · × Pnk :
g : E → Z ֒→ Pn1 × · · · × Pnk .
Set gi := πi ◦ g. If we apply Lemma 4.1 to each gi , we get the existence of an integer ci ∈ {0, . . . , c}
and of a morphism hi : P1 → Pni such that hi|Z = gi and hi ∗ (OPni (1)) ∼
= OP1 (ci ). The map
h = (h1 , . . . , hk ) : P1 → Pn1 × · · · × Pnk
(8)
has multi-degree (c1 , . . . , ck ). The curve
D := h(P1 )
is an integral rational curve containing Z. Since p ∈ hνnd1 ,...,ndk (Z)i, we have
1
rX (p) ≤ rν
d
d
n1 1 ,...,n k
k
k
(D) (p).
P
Thus it is sufficient to prove that, if we call D̃ := νnd1 ,...,ndk (D), then rD̃ (p) ≤ 2+c −1 + ki=1 di .
1
k
Pk
Since ci ≤ c for all i, it is sufficient to prove that rD̃ (p) ≤ 2 − c + i=1 ci di .
THIRD SECANT VARIETY OF SEGRE-VERONESE
11
Set Z̃ := νnd1 ,...,ndk (Z), m := dim(hD̃i) and
1
k
f = νnd1 ,...,ndk ◦ h : P1 → PN .
1
k
By assumption Z̃ is linearly independent in hD̃i ∼
= Pm and in particular c ≤ m + 1.
(a) Assume that the map h defined in (8) is birational onto its image. The curve D̃ ⊂ PN just
Pk
defined is a rational curve of degree a := i=1 ci di contained in the projective space Pm := hD̃i
and non-degenerate in Pm . Note that a ≥ m and that p ∈ hZ̃i.
(1) First assume that a = m. In this case D̃ is a rational normal curve of Pm . If c ≤
P
⌈(a + 1)/2⌉, then Sylvester’s theorem implies that rD̃ (p) = a + 2 − c = 2 − c + ki=1 ci di .
Now assume c > ⌈(a + 1)/2⌉. Since Z̃ is connected and curvilinear and p ∈
/ hGi for any
G ( Z̃, Sylvester’s theorem implies rD̃ (p) ≤ c.
(2) Now assume m < a. There is a rational normal curve C ⊂ Pa and a linear subspace
W ⊂ Pa such that dim(W ) = a − m − 1, C ∩ W = ∅ and h is the composition of the degree
a complete embedding j := P1 ֒→ Pa and the linear projection ℓ : Pa \ W → Pm from W .
The scheme E ′ := j(E) is a degree c curvilinear scheme and ℓ maps E ′ isomorphically onto
Z̃. Since Z̃ is linearly independent, then hE ′ i∩W = ∅ and ℓ maps isomorphically hE ′ i onto
hZ̃i. Thus there is a unique q ∈ hE ′ i such that ℓ(q) = p. Take any finite set S ⊂ j(P1 ) with
q ∈ hSi. Since C ∩W = ∅, ℓ(S) is a well-defined subset of D̃ with cardinality ≤ ♯(S). Hence
rD̃ (p) ≤ rC (q). As in step (a1) we see that either rC (q) = a + 2 − c (case c ≤ ⌈(a + 1)/2⌉)
or rC (q) ≤ c (case c > ⌈(a + 1)/2⌉).
(b) Now assume that h is not birational onto its image, but it has degree k ≥ 2. Note that k divides
Pk
ci for all i. In this case we will prove that rD̃ (p) ≤ 2 − c + i=1 ci di /k. Let h′ : P1 → h(P1 )
denote the normalization map. There is a degree k map h′′ : P1 → P1 such that h is the
composition of h′ ◦ h′′ and the inclusion h(P1 ) ⊂ Pn1 × · · · × Pnk . We have Z = h′ (E ′ ), where
E ′ = cu′ and u′ = h′′ (u). We use E ′ and h′ instead of E and h and repeat verbatim step (a).
Acknowledgements
We want to thank the anonymous referee and the Handling Editor Jan Draisma for their careful
jobs that improved the presentation of this paper. A special thank to the referee for her/his very
interesting questions that encouraged us in giving a better version of Theorem 1.8.
References
[1] E. Ballico and A. Bernardi, Decomposition of homogeneous polynomials with low rank, Math. Z., 271 (2012)
1141–1149.
[2] E. Ballico and A. Bernardi, Tensor ranks on tangent developable of Segre varieties, Linear Multilinear Algebra,
61 (2013) 881–894.
[3] E. Ballico and A. Bernardi, Stratification of the fourth secant variety of Veronese varieties via the symmetric
rank, Advances in Pure and Applied Mathematics, 4 (2013 ), 215–250.
[4] E. Ballico and A. Bernardi, A uniqueness result on the decompositions of a bi-homogeneous polynomial, Linear
and Multilinear Algebra, 65 (2017), 677–698.
[5] E. Ballico and A. Bernardi, Curvilinear schemes and maximum rank of forms, Matematiche (Catania) 72
(2017), no. 1, 137–144.
[6] A. Bernardi, N.S. Daleo, J.D. Hauenstein and B. Mourrain. Tensor decomposition and homotopy continuation,
Differential Geom. Appl. 55 (2017), 78–105.
[7] A. Bernardi, A. Gimigliano and M. Idà, Computing symmetric rank for symmetric tensors, J. Symbolic. Comput., 46 (2011) 34–53.
[8] J. Brachat, P. Comon, B. Mourrain and E.P. Tsigaridas, Symmetric Tensor Decomposition, Linear Algebra
Appl., 433 (2010) 1851–1872.
[9] J. Buczyński and J.M. Landsberg, Ranks of tensors and a generalization of secant varieties, Linear Algebra
Appl., 438 (2013), no. 2, 668–689.
[10] J. Buczyński and J.M. Landsberg, On the third secant variety, J. Algebraic Combin., 40 (2014) 475–502.
[11] G. Comas and M. Seiguer, On the Rank of a Binary form, Found. Comput. Math., 11 (2011) 65–78.
[12] A. Iarrobino and V. Kanev, Power sums, Gorenstein algebras, and determinantal loci, Lecture Notes in Mathematics, vol. 1721, Springer-Verlag, Berlin, 1999, Appendix C by Iarrobino and Steven L. Kleiman.
[13] J.M. Landsberg, Tensors: Geometry and Applications Graduate Studies in Mathematics, Amer. Math. Soc.
Providence, 128 (2012).
12
E. BALLICO AND A. BERNARDI
[14] J.M. Landsberg and Z. Teitler, On the ranks and border ranks of symmetric tensors, Found. Comput. Math.,
10 (2010) 339–366.
(Edoardo Ballico) Dipartimento di Matematica, Univ. Trento, Italy
E-mail address: [email protected]
(Alessandra Bernardi) Dipartimento di Matematica, Univ. Trento, Italy
E-mail address: [email protected]
| 0math.AC
|
Matroid Theory and Storage Codes: Bounds and
Constructions
arXiv:1704.04007v2 [cs.IT] 5 Sep 2017
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
Abstract Recent research on distributed storage systems (DSSs) has revealed interesting connections between matroid theory and locally repairable codes (LRCs).
The goal of this chapter is to introduce the reader to matroids and polymatroids,
and illustrate their relation to distributed storage systems. While many of the results
are rather technical in nature, effort is made to increase accessibility via simple
examples. The chapter embeds all the essential features of LRCs, namely locality,
availability, and hierarchy alongside with related generalised Singleton bounds.
1 Introduction to Locally Repairable Codes
In this chapter, we will discuss the theoretical foundations of locally repairable
codes (LRCs), which were introduced in Chapter 14. While our main interest is in
the codes and their applicability for distributed storage systems, significant parts
of our machinery comes from matroid theory. We will develop this theory to the
extent that is needed for the applications, and leave some additional pointers to
interpretations in terms of graphs and projective geometries.
The need for large-scale data storage is continuously increasing. Within the past
few years, distributed storage systems (DSSs) have revolutionised our traditional
ways of storing, securing, and accessing data. Storage node failure is a frequent obstacle in large-scale DSSs, making repair efficiency an important objective. A bottleneck for repair efficiency, measured by the notion of locality [27], is the number of
contacted nodes needed for repair. The key objects of study in this paper are locally
repairable codes (LRCs), which are, informally speaking, storage systems where a
Ragnar Freij-Hollanti, Camilla Hollanti, Thomas Westerbäck
Aalto University, Department of Mathematics and Systems Analysis, P.O. Box 11100, FI-00076
Aalto, Finland. e-mail: {ragnar.freij, camilla.hollanti, thomas.westerback}@aalto.fi.
The authors gratefully acknowledge the financial support from the Academy of Finland (grants
#276031 and #303819), as well as the support from the COST Action IC1104.
1
2
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
small number of failing nodes can be recovered by boundedly many other (closeby) nodes. Repair-efficient LRCs are already in use for large-scale DSSs used by,
for example, Facebook and Windows Azure Storage [41].
Another desired attribute, measured by the notion of availability [31], is the
property of having multiple alternative ways to repair nodes or access files. This
is particularly relevant for nodes containing so-called hot data that is frequently and
simultaneously accessed by many users. Moreover, as failures are often spatially
correlated, it is valuable to have each node repairable at several different scales.
This means that if a node fails simultaneously with the set of nodes that should
normally be used for repairing it, then there still exists a larger set of helper nodes
that can be used to recover the lost data. This property is captured by the notion of
hierarchy [32, 10] in the storage system.
Network coding techniques for large-scale DSSs were considered in [7]. Since
then, a plethora of research on DSSs with a focus on linear LRCs and various localities has been carried out, see [13, 27, 30, 33, 39] among many others. Availability
for linear LRCs was defined in [31]. The notion of hierarchical locality was first
studied in [32], where bounds for the global minimum distance were also obtained.
Let us denote by (n, k, d, r, δ ,t), respectively, the code length, dimension, global
minimum distance, locality, local minimum distance, and availability. Bold-faced
parameters (n, k, d, t) will be used in the sequel to refer to hierarchical locality and
availability. It was shown in [41] that the (r, δ = 2)-locality of a linear LRC is a
matroid invariant. The connection between matroid theory and linear LRCs was examined in more detail in [47]. In addition, the parameters (n, k, d, r, δ ) for linear
LRCs were generalised to matroids, and new results for both matroids and linear
LRCs were given therein. Even more generally, the parameters (n, k, d, r, δ ,t) were
generalised to polymatroids, and new results for polymatroids, matroids and both
linear and nonlinear LRCs over arbitrary finite alphabets were derived in [46]. Similar methods can be used to bound parameters of batch codes [49], as discussed in
Chapter 16. For more background on batch codes, see e.g. [16, 23]. Moreover, as
certain specific LRCs and batch codes1 belong to the class of private information retrieval (PIR) codes as defined in [9, Def. 4], the related LRC and batch code bounds
also hold for those PIR codes. See Section 5.3 and Chapter 16 for more discussion.
The main purpose of this chapter is to give an overview of the connection between matroid theory and linear LRCs with availability and hierarchy, using examples for improved clarity of the technical results. In particular, we are focusing on
how the parameters of a LRC can be analysed using the lattice of cyclic flats of an
associated matroid, and on a construction derived from matroid theory that provides
us with linear LRCs. The matroidal results on LRCs reviewed here are mostly taken
from [47, 46, 10].
The rest of this chapter is organised as follows. In Sections 1.1–1.2, we introduce
distributed storage systems and how they can be constructed by using linear codes.
In particular, we consider locally repairable codes with availability. Section 2 gives a
brief introduction to the concepts and features related to matroids relevant to LRCs.
1
To this end, we need to make specific assumptions on the locality and availability of the LRC [9,
Thm. 21], which also implies restrictions on the query structure of the batch code.
Matroid Theory and Storage Codes: Bounds and Constructions
3
In Section 3, we summarise the state-of-the-art generalised Singleton bounds on the
code parameters for linear codes, as well as discuss existence of Singleton-optimal
linear codes and matroids. Section 4 reviews some explicit (linear) code constructions. In Section 5, we go beyond linear codes and consider polymatroids and related
generalised Singleton bounds, which are then valid for all LRCs over any finite alphabet, and also imply bounds for PIR codes when restricted to systematic linear
codes. Section 6 concludes the chapter and discusses some open problems. Further
results in matroid theory and especially their representability are given in Appendix
for the interested reader. The following notation will be used throughout the paper:
F:
Fq :
E:
G:
G(X) :
a field;
the finite field of prime power size q;
a finite set;
a matrix over F with columns indexed by E;
the matrix obtained from G by restricting to the
columns indexed by X, where X ⊆ E;
C(G) :
the vector space generated by the columns of G;
R(G) :
the vector space generated by the rows of G;
C:
linear code C = R(G) over F generated by G;
CX :
the punctured code of C on X, i.e.,
CX = R(G(X)), where X ⊆ E;
2E :
the collection of all subsets of a finite set E;
[ j] :
the set {1, 2, . . . , j} for a positive integer j;
n, k, d, r, δ ,t, h : code length, dimension, minimum distance, locality,
failure tolerance, availability, hierarchy, respectively;
[n, k, d] , (n, k, d) : parameters of a linear/general code, respectively;
(n, k, . . .)i :
parameter values when we consider information
symbols, e.g., information symbol locality;
(n, k, . . .)a :
parameter values when we consider all code
symbols, e.g., all symbol locality;
(n, k, . . .)s :
parameter values when we consider systematic code
symbols, e.g., systematic symbol locality;
(n, k, d, t) :
parameter values for different hierarchy levels.
Remark 1. Here, d denotes the minimum (Hamming) distance of the code, rather
than the number of nodes that have to be contacted for repair, as is commonplace in
the theory of regenerating codes. In Chapter 14, the minimum distance of the code
was denoted by dH .
The motivation to study punctured codes arises from hierarchical locality; the
locality parameters at the different hierarchy levels correspond to the global parameters of the related punctured codes. The puncturing operation on codes corresponds
to the so-called restriction (or deletion) operation on matroids.
We also point out that G(E) = G and CE = C. We will often index a matrix G by
[n], where n is the number of columns in G.
4
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
1.1 Distributed Storage Systems from Linear Codes
A linear code C can be used to obtain a DSS, where every coordinate in C represents
a storage node in the DSS, and every point in C represents a stored data item. While
one often assumes that the data items are field elements in their own right, no such
assumption is necessary. However, if C is a code over the field F and the data items
are elements in an alphabet A, then we must be able to form the linear combinations
f1 a1 + f2 a2 for f1 , f2 in F and a1 , a2 in A. Moreover, if we know the scalar f , we
must be able to read off a from f a. This is achieved if A ∼
= Fα is a vector space over
F, wherefore we must have |A| ≥ |F|. Thus, the length of the data items must be at
least the number of symbols needed to represent a field element. In particular if the
data items are measured in, e.g., kilobytes, then we are restricted to work over fields
of size not larger than about 28000 . Beside this strict upper bound on the field size,
the complexity of operations also makes small field sizes — ideally even binary
fields — naturally desirable.
Example 1. Let C be the linear code generated by the following matrix G over F3 :
1
1
G= 0
0
0
2
0
1
0
0
3
0
0
1
0
4
0
0
0
1
5
1
1
0
0
6
1
0
1
0
7
1
1
1
0
8
1
2
0
1
9
1
2
0
2
Then, C corresponds to a 9 node storage system, storing four files (a, b, c, d), each
of which is an element in Fα3 . In this system, node 1 stores a, node 5 stores a + b,
node 9 stores a + 2b + 2d, and so on.
Two very basic properties of any DSS are that every node can be repaired by
some other nodes and that every node contains some information2 . We therefore
give the following definition.
Definition 1. A linear [n, k, d]-code C over a field is a non-degenerate storage code
if d ≥ 2 and there is no zero column in a generator matrix of C.
The first example of a storage code, and the motivating example behind the notion
of locality, is the notion of a maximum distance separable (MDS) code. It has several
different definitions in the literature, here we list a few of them.
Definition 2. The following properties are equivalent for linear [n, k, d] storage
codes:
(i) n = k + d − 1.
(ii) The stored data can be retrieved from any k nodes in the storage system.
2
We remark that if one takes into account queueing theoretic aspects, then data allocation may
become less trivial (some nodes may be empty). Such aspects are discussed, especially in a wireless
setting, in Chapters 12 and 13. However, these considerations are out of the scope of this chapter.
Matroid Theory and Storage Codes: Bounds and Constructions
5
(iii) To repair any single erased node in the storage system, one needs to contact k
other nodes.
A code that satisfies one (and therefore all) of the above properties is called an MDS
code.
By the Singleton bound, n ≥ k + d − 1 holds for any storage code, so by property
(i), MDS codes are “optimal” in the sense that they have minimal length for given
storage capacity and error tolerance. However, (iii) is clearly an unfavourable property in terms of erasure correction. This is the motivation behind constructing codes
with small n − k − d, where individual node failures can still be corrected “locally”.
1.2 Linear Locally Repairable Codes with Availability
The very broad class of linear LRCs will be defined next. It is worth noting that,
contrary to what the terminology would suggest, a LRC is not a novel kind of code,
but rather the “locality parameters” (r, δ ) can be defined for any code. What we
call a LRC is then only a code that is specifically designed with the parameters
(r, δ ) in mind. While the locality parameters can be understood directly in terms of
the storage system, it is more instructive from a coding theoretic point of view to
understand them via punctured codes. Then, the punctured codes will correspond
exactly to the restrictions to “locality sets”, which can be used to locally repair a
small number of node failures within the locality set.
Definition 3. Let G be a matrix over F indexed by E and C the linear code generated
by G. Then, for X ⊆ E, CX is a linear [nX , kX , dX ]-code where
nX = |X|,
kX = rank(G(X)),
dX = min{|Y | : Y ⊆ X and kX\Y < kX }.
Alternatively, one can define the minimum distance dX as the smallest support of
a non-zero codeword in CX = R(G(X)). We use Definition 3, as it has the advantage
of not depending on the linearity of the code.
Example 2. Consider the storage code C from Example 1. Let Y1 = {1, 2, 3, 5, 6, 7},
X1 = {1, 2, 5} and X2 = {2, 6, 7}. Then CY1 , CX1 and CX2 are storage codes with
[nY1 , kY1 , dY1 ] = [6, 3, 3] ,
[nX1 , kX1 , dX1 ] = [3, 2, 2] ,
[nX2 , kX2 , dX2 ] = [3, 2, 2] .
The parameter dX is the minimum (Hamming) distance of CX . We say that C is
an [n, k, d]-code with [n, k, d] = [nE , kE , dE ].
We choose the following definition for general (n, k, d, r, δ ,t)-LRCs (i.e., both
linear and nonlinear), which we will compare to known results for linear LRCs.
6
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
Definition 4. An (n, k, d)-code C over A is a nonempty subset C of An , where A is
a finite set of size s, k = logs (|C|), and d the minimum (Hamming) distance of the
code. For X = {i1 , . . . , im } ⊆ E, the puncturing CX is defined as
CX = {(ci1 , . . . , cim ) : c ∈ C}.
The code C is non-degenerate, if d ≥ 2 and |C{i} | > 1 for all coordinates i ∈ [n].
Definition 5. A locally repairable code over A is a non-degenerate (n, k, d)-code C.
A coordinate x ∈ [n] of C has locality (r, δ ) and availability t if there are t subsets
R1 , . . . , Rt of [n], called repair sets of x, such that for i, j ∈ [t]
(i) x ∈ Ri ,
(ii) |Ri | ≤ r + δ − 1,
(iii) d(CRi ) ≥ δ ,
(iv) i 6= j ⇒ Ri ∩ R j = {x}.
If every element x ∈ X ⊆ E has availability with parameters (n, k, d, r, δ ,t) in
C, then we say that the set X has (n, k, d, r, δ ,t)-availability in C. We will often
talk about codes with (n, k, d, r, δ )-locality, by which we mean a code that has
(n, k, d, r, δ , 1)-availability, so that symbols are not required to be included in more
than one repair set. If the other parameters are clear from the context, we may shortly
say that X has “locality (r, δ )” or “availability t”, along the lines of the above definition.
An information set of a linear [n, k, d]-code C is defined as a set X ⊆ E such that
kX = |X| = k. Hence, X is an information set of C if and only if there is a generator
matrix G of C such that G(X) equals the identity matrix, i.e., C is systematic in
the coordinate positions indexed by X when generated by G. In terms of storage
systems, this means that the nodes in X together store all the information of the
DSS.
Example 3. Two examples of an information set of the linear code C generated by
G in Example 1 are {1, 2, 3, 4} and {1, 2, 6, 8}.
More formally we define:
Definition 6. Let C be an (n, k, d)-code and X a subset of [n]. Then X is an information set of C if logs (|CX |) = k and logs (CY ) < k for all Y ( X. Further, X is
systematic if k is an integer, |X| = k and CX = Ak . Also, X is an all-coordinate set if
X = [n].
Definition 7. A systematic-symbol, information-symbol, and all-symbol LRC, respectively, is an (n, k, d)-LRC with a systematic, information set and all-coordinate
set X, such that every coordinate in X has locality (r, δ ) and availability t. These are
denoted by
(n, k, d, r, δ ,t)s −LRC, (n, k, d, r, δ ,t)i −LRC, and (n, k, d, r, δ ,t)a −LRC,
Matroid Theory and Storage Codes: Bounds and Constructions
7
respectively. Further, when availability is not considered (t = 1), we get natural
notions of (n, k, d, r, δ )s , (n, k, d, r, δ )i , and (n, k, d, r, δ )a -LRCs.
2 Introduction to Matroids
Matroids were first introduced by Whitney in 1935, to capture and generalise the
notion of linear dependence in purely combinatorial terms [48]. Indeed, the combinatorial setting is general enough to also capture many other notions of dependence
occurring in mathematics, such as cycles or incidences in a graph, non-transversality
of algebraic varieties, or algebraic dependence of field extensions. Although the
original motivation comes from linear algebra, we will see that a lot of matroid
terminology comes from graph theory and projective geometry. More details about
these aspects of matroid theory are relegated to the appendix.
2.1 Definitions
We begin by presenting two equivalent definitions of a matroid.
Definition 8 (Rank function). A (finite) matroid M = (ρ, E) is a finite set E together with a rank function ρ : 2E → Z such that for all subsets X,Y ⊆ E
(R.1) 0 ≤ ρ(X) ≤ |X|,
(R.2) X ⊆ Y ⇒ ρ(X) ≤ ρ(Y ),
(R.3) ρ(X) + ρ(Y ) ≥ ρ(X ∪Y ) + ρ(X ∩Y ).
An alternative but equivalent definition of a matroid is the following.
Definition 9 (Independent sets). A (finite) matroid M = (I , E) is a finite set E and
a collection of subsets I ⊆ 2E such that
(I.1) 0/ ∈ I ,
(I.2) Y ∈ I , X ⊆ Y ⇒ X ∈ I ,
(I.3) For all pairs X,Y ∈ I with |X| < |Y |, there exists
y ∈ Y \ X such that X ∪ {y} ∈ I .
The subsets in I are the independent sets of the matroid.
The rank function ρ and the independents sets I of a matroid on a ground set E
are linked as follows: For X ⊆ E,
ρ(X) = max{|Y | : Y ⊆ X and Y ∈ I },
and X ∈ I if and only if ρ(X) = |X|. It is an easy (but not trivial) exercise to
show that the two definitions are equivalent under this correspondence. Another
8
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
frequently used definition is in terms of the set of bases for a matroid, which are the
maximal independent sets. Further, we will also use the nullity function η : 2E → Z,
where η(X) = |X| − ρ(X) for X ⊆ E.
Any matrix G over a field F generates a matroid MG = (ρ, E), where E is the
set of columns of G, and ρ(I) is the rank over F of the induced matrix G(I) for
I ⊆ E. Consequently, I ∈ I precisely when I is a linearly independent set of vectors.
It is straightforward to check that this is a matroid according to Definition 8. As
elementary row operations preserve the row space R(G(I)) for all I ⊆ E, it follows
that row equivalent matrices generate the same matroid.
Two matroids M1 = (ρ1 , E1 ) and M2 = (ρ2 , E2 ) are isomorphic if there exists a
bijection ψ : E1 → E2 such that ρ2 (ψ(X)) = ρ1 (X) for all subsets X ⊆ E1 .
Definition 10. A matroid that is isomorphic to MG for some matrix G over F is said
to be representable over F. We also say that such a matroid is F-linear.
Two trivial matroids are the zero matroid where ρ(X) = 0 for each set X ⊆ E,
and the one where ρ(X) = |X| for all X ⊆ E. These correspond to all-zeros matrices
and invertible n × n-matrices respectively. The first non-trivial example of a matroid
is the following:
Definition 11. The uniform matroid Unk = (ρ, [n]), where [n] = {1, 2, · · · , n}, is given
by the rank function ρ(X) = min{|X|, k} for X ⊆ [n].
The following straightforward observation gives yet another characterisation of
MDS codes.
Proposition 1. G is the generator matrix of an [n, k, n − k + 1]-MDS code if and only
if MG is the uniform matroid Unk .
2.2 Matroid Operations
For explicit constructions of matroids, as well as for analysing their structure, a few
elementary operations are useful. Here, we will define these in terms of the rank
function, but observe that they can equally well be formulated in terms of independent sets. The effect of these operations on the representability of the matroid
is discussed in Appendix. In addition to the operations listed here, two other very
important matroid operations are dualisation and contraction. As these are not explicitly used here to understand locally repairable codes, we leave their definition to
Appendix.
Definition 12. The direct sum of two matroids M = (ρM , EM ) and N = (ρN , EN ) is
M ⊕ N = (τ, EM t EN ),
where t denotes the disjoint union, and τ : 2EM tEN → Z is defined by τ(X) =
ρM (X ∩ EM ) + ρ(X ∩ EN ).
Matroid Theory and Storage Codes: Bounds and Constructions
9
Thus, all dependent sets from M and N remain dependent in M ⊕ N, whereas there is
no dependence between elements in M and elements in N. If M and N are graphical
matroids3 , then M ⊕ N is graphical, and obtained from the disjoint union of the
graphs associated to M and N.
Definition 13. The restriction of M = (ρ, E) to a subset X ⊆ E is the matroid M|X =
(ρ|X , X), where
ρ|X (Y ) = ρ(Y ), for Y ⊆ X.
(1)
Obviously, for any matroid M with underlying set E, we have M|E = M. The
restriction operation is also often referred to as deletion of E \ X, especially if E \ X
is a singleton. Given a matrix G that represents M, the submatrix G(X) represents
M|X .
Definition 14. The truncation of a matroid M = (ρ, E) at rank k ≤ ρ(E) is Mk =
(ρ 0 , E), where ρ 0 (X) = min{ρ(X), k}.
Geometrically, the truncation of a matroid corresponds to projecting a point configuration onto a generic k-dimensional space. However, this does not imply that
truncations of F-linear matroids are necessarily F-linear, as it may be the case that
there exists no k-space that is in general position relative to the given point configuration. However, it is easy to see that Mk is always representable over some field
extension of F. In fact, via a probabilistic construction,
one sees that the field exten
sion can be chosen to have size at most q nk [17].
The relaxation is the elementary operation that is most difficult to describe in
terms of rank functions. It is designed to destroy representability of matroids, and
corresponds to selecting a hyperplane in the point configuration, and perturbing it
so that its points are no longer coplanar. To prepare for the definition, we say that a
circuit is a dependent set, all of whose subsets are independent. For any nonuniform
matroid M = (ρ, E), there are circuits of rank ρ(E) − 1. This is seen by taking any
dependent set of rank ρ(E) − 1, and deleting elements successively in such a way
that the rank does not decrease. We mention that a matroid that has no circuits of
rank < ρ(E) − 1 is called a paving matroid. It is conjectured that asymptotically (in
the size) almost all matroids are paving [24]. Recent research shows that this is true
at least on a “logarithmic scale” [28].
Definition 15. Let M = (I , E) be a matroid with rank function ρ, and let C be a
circuit of rank ρ(E) − 1. The relaxation of M at C is the matroid (I ∪ {C}, E).
Example 4. The first example of a matroid constructed by relaxation is the nonPappus matroid of Figure 1. This is constructed by relaxing the circuit {4, 5, 6}
from the representable matroid MG , where
12 3 4
1 0 -1 1/2
G=
11 1 0
11 1 1
3
5 6
0 -1/2
0 0
1 1
See Appendix for the definition of graphical matroids
7 8
1 0
-1 -1
1 1
9
-1
-1
1
10
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
and can be defined over any field of odd (or zero) characteristic, other than F3 .
Fig. 1 The non-Pappus matroid, which is not representable over any field. If
there were a line between
the three points in the middle, then the figure would
illustrate the matroid MG in
Example 4. Relaxation of the
circuit {4, 5, 6} corresponds
to deletion of this line in the
figure.
2.3 Matroid Invariants of Codes
There is a straightforward connection between linear codes and matroids. Indeed, let
C be a linear code generated by a matrix G. Then C is associated with the matroid
MG = (ρ, E). As two different generator matrices of C have the same row space, they
will generate the same matroid. Therefore, without any inconsistency, we denote the
associated linear matroid of C by MC = (ρC , E). In general, there are many different
codes C 6= C0 with the same matroid structure MC = MC0 . In the appendix, we will
see how this phenomenon can be interpreted as a stratification of the Grassmannian
over a finite field.
A property of linear codes that depends only on the matroid structure of the code
is called matroid invariant. For example, the collection of information sets and the
parameters [n, k, d] of a code are matroid invariant properties. This is the content of
the following easy proposition.
Proposition 2. Let C be a linear [n, k, d]-code and X ⊆ E. Then for MC = (ρC , E),
(i) nX = |X|,
(ii) kX = ρC (X),
(iii) dX = min{|Y | : Y ⊆ X, ρC (X \Y ) < ρC (X)},
(iv) X is an information set of C ⇐⇒
X is a basis of MC ⇐⇒ ρC (X) = |X| = k.
In addition to the parameters [n, k, d] of a linear code C, we are also interested
in the length, rank and minimum distance of the punctured codes, since these correspond to the locality parameters at the different hierarchy levels, which we will
discuss in more detail in Section 5.
Matroid Theory and Storage Codes: Bounds and Constructions
11
A punctured code can be analysed using matroid restrictions, since MC|X = MC |X
for every coordinate subset X. Thus, the parameters [nX , kX , dX ] of CX are also matroid invariant properties for C.
Example 5. Let C denote the [n, k, d]-code generated by the matrix G given in Example 1. Then [n, k, d] = [9, 4, 3], where the value of d arises from the fact that
ρC ([9] \ {i, j}) = 4 for i, j = 1, 2, . . . , 9, and ρC ([9] \ {4, 8, 9}) = 3. Two information
sets of C are {1, 2, 3, 4} and {1, 2, 6, 8}, as we already saw before in Example 3.
It is rather easy to see that two different linear codes can have the same associated
matroid. As a consequence, not every property of a linear code is matroid invariant.
An example of a code invariant that is not matroid invariant is the covering radius
[5, 37]. Indeed, an [n, k, d]-MDS code, i.e., a realisation of the uniform matroid Unk ,
generically has covering radius d − 1 = n − k, yet there exist MDS codes with lower
covering radii. An explicit example is given in [5].
2.4 The Lattice of Cyclic Flats
One matroid invariant that has singled out as essential for describing the repairability
of storage codes is the lattice of cyclic flats. To define this, remember that X ⊆ E is a
circuit in M = (I , E) if X is dependent, but all proper subsets of X are independent.
A cyclic set is a (possibly empty) union of circuits. Equivalently, X is cyclic if for
every x ∈ X
ρ(X \ {x}) = ρ(X).
Let us define the operation cyc : 2E → 2E by
cyc(X) = {x ∈ X : ρ(X \ {x}) = ρ(X)}.
Then X is cyclic if and only if cyc(X) = X. We refer to cyc as the cyclic core
operator.
Dually, we define the closure of X to be
cl(X) = {y ∈ E : ρ(X ∪ {y}) = ρ(X)},
and notice that X ⊆ cl(X) by definition. We say that X is a flat if cl(X) = X. Therefore, X is a cyclic flat if
ρ(X \ {x}) = ρ(X) and ρ(X ∪ {y}) > ρ(X)
for all x ∈ X and y ∈ E \ X. The set of flats, cyclic sets, and cyclic flats of M are
denoted by F (M), U (M), and Z (M), respectively.
It is not entirely obvious that the set of cyclic flats is nonempty. However, it
follows from the matroid axioms that the closure operator cl preserves cyclicity, and
that the cyclic core operator cyc preserves flatness. Thus we can consider cl and cyc
as maps
12
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
(
cl :
and
2E → F (M)
U (M) → Z (M)
(
cyc :
2E → U (M)
F (M) → Z (M)
,
.
In particular, for any set X ⊆ E, we have cyc ◦ cl(X) ∈ Z (M) and cl ◦ cyc(X) ∈
Z (M).
Let M[G] = (ρ, E) be a linear matroid, generated by G. Then X ⊆ E is a cyclic
flat if and only if the following two conditions are satisfied
(i) C(G(X)) ∩ C(G(E \ X)) = 0
(ii) x ∈ X ⇒ C(G(X \ {x})) = C(G(X)).
In terms of storage codes, a cyclic flat is thus a set X ⊆ E of storage nodes such that
every node in X can be repaired by the other nodes in X, whereas no node outside
X can be repaired by X. This observation shows the relevance of cyclic flats for
storage applications. The strength of using Z (M) as a theoretical tool comes from
its additional lattice structure, which we will discuss next.
A collection of sets P ⊆ 2E ordered by inclusion defines a partially ordered set
(poset) (P, ⊆). Let X and Y denote two elements of P. Z is the join X ∨Y if it is
the unique maximal element in {W ∈ P : X ⊆ W , Y ⊆ W }. Dually, Z is the meet
X ∧Y if it is the unique minimal element in {W ∈ P : X ⊇ W , Y ⊇ W }.
A pair of elements in an arbitrary poset does not need to have a join or a meet. If
(P, ⊆) is a poset such that every pair of elements in P has a join and a meet, then
P is called a lattice. The bottomWand top elements of
a finite lattice (P, ⊆) always
V
exist, and are denoted by 1P = X∈P X and 0P = X∈P X, respectively.
Two basic properties of cyclic flats of a matroid are given in the following proposition.
Proposition 3 ([4]). Let M = (ρ, E) be a matroid and Z the collection of cyclic
flats of M. Then,
(i) ρ(X) = min{ρ(F) + |X \ F| : F ∈ Z }, for X ⊆ E,
(ii) (Z , ⊆) is a lattice, X ∨Y = cl(X ∪Y ) and
X ∧Y = cyc(X ∩Y ) for X,Y ∈ Z .
Proposition 3 (i) shows that a matroid is uniquely determined by its cyclic flats
and their ranks.
Example 6. Let MC = (ρC , E) be the matroid associated to the linear code C generated by the matrix G given in Example 1. The lattice of cyclic flats (Z , ⊆) of MC is
given in the figure below, where the cyclic flat and its rank are given at each node.
Matroid Theory and Storage Codes: Bounds and Constructions
13
([9], 4)
1Z
({1, 2, 3, 5, 6, 7}, 3)
({1, 2, 4, 5, 8, 9}, 3)
Y1
Y2
({1, 2, 5}, 2)
X1
({3, 5, 7}, 2)
X2
X3
({2, 6, 7}, 2)
({4, 8, 9}, 2)
X4
X5
({1, 3, 6}, 2)
0Z
(0,
/ 0)
An axiom scheme for matroids via cyclic flats and their ranks was independently
given in [4] and [35]. This gives a compact way to construct matroids with prescribed local parameters, which we have exploited in [47].
Theorem 1 (see [4] Th. 3.2 and [35]). Let Z ⊆ 2E and let ρ be a function ρ : Z →
Z. There is a matroid M on E for which Z is the set of cyclic flats and ρ is the rank
function restricted to the sets in Z , if and only if
(Z0) Z is a lattice under inclusion,
(Z1) ρ(0Z ) = 0,
(Z2) X,Y ∈ Z and X ( Y ⇒
0 < ρ(Y ) − ρ(X) < |Y | − |X|,
(Z3) X,Y ∈ Z ⇒ ρ(X) + ρ(Y ) ≥
ρ(X ∨Y ) + ρ(X ∧Y ) + |(X ∩Y ) \ (X ∧Y )|.
For a linear [n, k, d]-code C with MC = (ρC , E) and Z = Z (MC ), and for a coordinate x, we have
(i) d ≥ 2 ⇐⇒ 1Z = E,
(ii) C{x} 6= {0F } for every x ∈ E ⇐⇒ 0Z = 0.
/
Hence, by Definition 1, we can describe non-degeneracy in terms of the lattice of
cyclic flats, as follows.
Proposition 4. Let C be a linear [n, k, d]-code and Z denote the collection of cyclic
flats of the matroid MC = (ρC , E). Then C is a non-degenerate storage code if and
only if 0Z = 0/ and 1Z = E.
Proposition 5. Let C be a non-degenerate storage code and MC = (ρC , E). Then,
for X ⊆ E, CX is a non-degenerate storage code if and only if X is a cyclic set of
MC .
14
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
As cyclic sets correspond to non-degenerate subcodes, and hence to systems
where every symbol is stored with redundancy, we will use these as our “repair
sets”. Therefore, we want to determine from the lattice of cyclic flats, whether a set
is cyclic or not, which we achieve through the following theorem.
Theorem 2. Let M = (ρ, E) be a matroid with 0Z = 0/ and 1Z = E where Z =
Z (M). Then, for any X ⊆ E, X ∈ U (M) if and only if the cyclic flat
FX =
{F ∈ Z : X ⊆ F}
^
is such that
ρ(F) + |X \ F| > ρ(F X ).
for all F ( F X in Z .
If this is indeed the case, then it is easy to verify that F X as defined in Theorem
2 is indeed the closure cl(X) as defined earlier. In order to analyze the parameters
[nX , kX , dX ] of a punctured code CX , we will use the lattice of cyclic flats of MC|X .
Theorem 3. Let M = (ρ, E) be a matroid with 0Z = 0/ and 1Z = E where Z =
Z (M). Then, for X ∈ U (M),
(i) Z (M|X) = {X ∩ F ∈ U (M) : F ∈ Z , F ⊆ F X },
(ii) Y ∈ Z (M|X) ⇒ ρ|X (Y ) = ρ(F Y ).
We remark that if X is a cyclic flat of a matroid M, then Z (M|X) = {F ∈ Z (M) :
F ⊆ X}.
Example 7. Let Z = Z (MC ) be the lattice of cyclic flats given in Example 6, where
MC = (ρC , E) is the matroid associated to the linear LRC C generated by the matrix
G given in Example 1. Then, F X = F Y = Y1 for X = {1, 2, 3, 7} and Y = {1, 2, 3}.
Further, X is a cyclic set but Y is not a cyclic set. The lattice of cyclic flats (ZX , ⊆)
for MC|X = MC |X is shown in the following figure.
({1, 2, 3, 7}, 3)
1ZX
0ZX
(0,
/ 0)
The very simple structure of ZX shows that CX has the very favourable property
of being an MDS code. Indeed, the following proposition is immediate from the
definitions of the involved concepts.
Matroid Theory and Storage Codes: Bounds and Constructions
15
Proposition 6. Let C be a linear code of length n and rank k. The following are
equivalent:
(i) C is an [n, k, n − k + 1]-MDS code.
(ii) MC is the uniform matroid Unk .
(iii) Z = Z (MC ) is the two element lattice with 1Z = [n] and 0Z = 0.
/
For linear LRCs we are also interested in when a coordinate set is an information
set, or equivalently, if it is a basis for the matroid. This property is determined by
the cyclic flats as follows.
Proposition 7. Let C be a linear [n, k, d]-code with 0Z = 0/ and 1Z = E where Z
is the collection of cyclic flats of the matroid MC = (ρC , E). Then, for any X ⊆ E, X
is an information set of C if and only if the following two conditions are satisfied,
(i) |X| = ρC (1Z ),
(ii) |X ∩ F| ≤ ρC (F) for every F ∈ Z .
Example 8. Let C be the linear [n, k, d]-code generated by the matrix G given in Example 1. Then, by the lattice of cyclic flats for MC given in Example 6, C is a linear
LRC with all-symbol (2, 2)-locality. We notice that, by Proposition 7, {1, 2, 3, 4} is
an information set of C. This follows as it is not contained in either of Y1 and Y2 ,
while all its subsets are. On the other hand, {1, 2, 8, 9} is not an information set of
C, as it is itself a subset of Y2 .
The parameters [n, k, d] of a linear LRC C and [nX , kX , dX ] of a punctured code
CX that is a non-degenerate storage code can now be determined by the lattice of
cyclic flats as follows.
Theorem 4. Let C be a linear [n, k, d]-LRC, where Z = Z (MC ) for the matroid
MC = (ρC , E). Then, for any X ∈ U (MC ), CX is a linear [nX , kX , dX ]-LRC with
(i) nX = |X|,
(ii) kX = ρ(F X ),
(iii) dX = nX − kX + 1 − max{η(Y ) : Y ∈ Z (MC |X) \ X}.
Example 9. Let Z = Z (MC ) be the lattice of cyclic flats given in Example 6, where
MC = (ρC , E) is the matroid associated to the linear [n, k, d]-LRC C generated by the
matrix G given in Example 1. Then by Example 7, X = {1, 2, 3, 7} is a cyclic set,
and by Theorem 4 CX is a linear [nX , kX , dX ]-LRC with parameters nX = 4, kX = 3
and dX = 4 − 3 + 1 − 0 = 2. Moreover, n = nE = 9, k = kE = 4 and d = dE =
9 − 4 + 1 − 3 = 3.
3 Singleton-type Bounds
Many important properties of a linear code C are due to its matroid structure, which
is captured by the matroid MC . By the results in [47] and [10], matroid theory seems
16
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
to be particularly suitable for proving Singleton-type bounds for linear LRCs and
nonexistence results of Singleton-optimal linear LRCs for certain parameters.
Even though matroids can be associated to linear codes to capture the key properties for linear LRCs, this cannot be done in general for nonlinear codes. Fortunately,
by using some key properties of entropy, any code (either linear and nonlinear) C
can be associated with a polymatroid PC so that PC captures the key properties of the
code when it is used as a LRC. A polymatroid is a generalisation of a matroid. For
any linear code C the associated polymatroid PC and matroid MC are the same object.
We will briefly discuss the connection between polymatroids and codes in Section5.
Singleton-type bounds for polymatroids were derived in [46], and polymatroid theory for its part seems to be particularly suitable for proving such bounds for general
LRCs. In Section 5 we will also review Singleton-type bounds for polymatroids and
general codes with availability and hierarchy.
3.1 Singleton-type Bounds for Matroids and Linear LRCs
Matroid theory provides a unified way to understand and connect several different
branches of mathematics, for example linear algebra, graph theory, combinatorics,
geometry, topology and optimisation theory. Hence, a theorem proven for matroids
gives “for free” theorems for many different objects related to matroid theory. As
described earlier, the key parameters (n, k, r, d, δ ,t) of a linear LRCs C are matroid
properties of the matroid MC and can therefore be defined for matroids in general.
Definition 16. Let M = (ρ, E) be a matroid and X ⊆ E. Then
(i) nX = |X|,
(ii) kX = ρ(X),
(iii) dX = min{|Y | : Y ⊆ X, ρ(X \Y ) < ρ(X)},
(iv) X is an information set of M if ρ(X) = kE and ρ(Y ) < kE for all Y ( X,
(v) M is non-degenerate if ρ(x) > 0 for all x ∈ E and dE ≥ 2.
Further, n = nE , k = kE and d = dE , and the definitions of repair sets, (r, δ )locality and t-availability for elements x ∈ E, as well as the concepts of (n, k, d, r, δ ,t)i matroids and (n, k, d, r, δ ,t)a -matroids are directly carried over from Definitions 5
and 7.
Before stating Theorem 5 below, it is not at all clear that the Singleton-type
bounds, already proven for linear LRCs, also hold for matroids in general. Especially, one could doubt this generality of the bound because of the wide connection
between matroids and a variety of different mathematical objects, as well as for the
sake of the recently proven result, stated in Theorem 13 later on, that almost all
matroids are nonrepresentable. However, Theorem 5 gives a Singleton-type bound
that holds for matroids in general. This implies, as special cases, the same bound on
linear LRCs and other objects related to matroids, e.g., graphs, almost affine LRCs,
and transversals. For the bound to make sense for various objects, a description of
Matroid Theory and Storage Codes: Bounds and Constructions
17
the parameters (n, k, d, r, δ ) has to be given for the objects in question. To give an
example, for a graph,
• n equals the number of edges,
• k equals the difference of the number of vertices and the number of connected
components,
• d is the smallest number of edges in an edge-cut (i.e., a set of edges whose removal increases the number of connected components in the graph).
Recall that the Singleton bound [36] states that for any linear [n, k, d]-code we
have
d ≤ n − k + 1.
(2)
In what follows, we state generalised versions of this bound, accounting for the
various parameters relevant for storage systems. We start with the general one for
matroids.
Theorem 5 ([47] Singleton-type bound for matroids). Let M = (ρ, E) be an
(n, k, d, r, δ )i -matroid. Then
k
− 1 (δ − 1).
(3)
d ≤ n−k+1−
r
Theorem 5 was stated for all-symbol locality in [47]. However, the proof given
in [47] implies also information-symbol locality. As an illustration on how matroid
theory and cyclic flats can be useful for proving Singleton-type of bounds we will
here give a proof of Theorem 5.
Proof. We know from Theorem 4 that d = n − k + 1 − max{η(Z) : Z ∈ Z \ E}.
Hence to prove the
theorem
we need to show that there exists a cyclic flat Z 6= E in
M with η(Z) ≥ kr − 1 (δ − 1).
Let B be an information set of M, i.e, B is a basis of M, such that M is an
(n, k, d, r, δ )i -matroid. For x ∈ B let Rx denote the repair set of x. Since Rx is a cyclic
set of we obtain that Zx = cl(Rx ) is a cyclic flat of M with
ρ(Zx ) = ρ(Rx ) ≤ r and η(Zx ) ≥ η(Rx ) ≥ dRx − 1 ≥ δ − 1.
As ρ(B) = k we can choose a subset of cyclic flats {Z1 , . . . , Zm } ⊆ {Zx : x ∈ B}
such that we obtain a chain of cyclic flats
0/ = Y0 ( Y1 ( . . . ( Ym = E
with Yi = Yi−1 ∨ Zi for i = 1, . . . , m. Since ρ(Y0 ) = η(Y0 ) = 0 and ρ(Ym ) = k, the
theorem will now be proved if we can prove that ρ(Yi ) − ρ(Yi−1 ) ≤ r and η(Yi ) −
η(Yi−1 ) ≥ δ − 1 for i = 1, . . . , m.
First, by the use of Axiom (R.3) and Proposition 3,
ρ(Yi ) = ρ(cl(Yi−1 ∪ Zi )) = ρ(Yi−1 ∪ Zi ) ≤ ρ(Yi−1 ) + ρ(Zi ) − ρ(Yi−1 ∩ Zi )
≤ ρ(Yi−1 ) + r.
18
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
Second, by Axiom (R.3), η(X) + η(Y ) ≤ η(X ∩ Y ) + η(X ∪ Y ) for X,Y ⊆ E.
Further, we observe that, cyc(Yi−1 ∩ Zi ) and Zi are cyclic flats of M|Zi and that
cyc(Yi−1 ∩ Zi ) ( Zi . Hence,
η(Yi ) = η(cl(Yi−1 ∪ Zi )) ≥ η(Yi−1 ∪ Zi ) ≥ η(Yi−1 ) + η(Zi ) − η(Yi−1 ∩ Zi )
= η(Yi−1 ) + η(Zi ) − η(cyc(Yi−1 ∩ Zi ))
= η(Yi−1 ) + |Zi | − ρ(Zi ) − η(cyc(Yi−1 ∩ Zi )) ≥ η(Yi−1 ) + dZi − 1
≥ η(Yi−1 ) + δ − 1.
That dZi ≥ δ follows from the fact that Zi = cl(Rx ) for some x ∈ B and therefore
dZi = dcl(Rx ) = min{|Y | : Y ⊆ cl(Rx ), ρ(cl(Rx ) \Y ) ≤ ρ(cl(Rx )}
≥ min{|Y | : Y ⊆ Rx , ρ(Rx \Y ) ≤ ρ(Rx )}
= dRx .
t
u
The Singleton bound given in (2) was generalised by Gopalan et al. in [13] as
follows. A linear (n, k, d, r)i -LRC satisfies
k
−1 .
(4)
d ≤ n−k+1−
r
The bound (4) shows that there is a penalty for requiring locality. That is, the smaller
the locality r the smaller the upper bound on d. By the definition of LRCs, any linear
[n, k, d]-code with locality r is also a linear [n, k, d]-code with locality k. Hence, by
letting the locality be k, the bound (4) implies (2).
The bound (4) was generalised in [30] as follows. A linear (n, k, d, r, δ )i -LRC
satisfies
k
d ≤ n−k+1−
− 1 (δ − 1).
(5)
r
The bound (5) again shows that there is a penalty on the upper bound for d depending on the size of the local distance δ . This is, the bigger the local distance δ the
smaller the upper bound on the global distance d. However, we must remark that any
linear (n, k, d, r, δ )i -LRC satisfies d ≥ δ , and this property also holds more generally
for matroids [47]. The bound (4) follows from the bound (5) by letting δ = 2.
A bound including availability was proven in [44]. This bound states that a linear
(n, k, d, r,t)i -LRC satisfies
t(k − 1) + 1
d ≤ n−k+1−
−1 .
(6)
t(r − 1) + 1
Again, the bound (4) follows from (6) above by letting t = 1.
The bounds (3)-(6) are stated assuming information-symbol locality. However,
since every matroid contains an information set, this implies that the bound is also
valid under the stronger assumption of all-symbol locality.
Matroid Theory and Storage Codes: Bounds and Constructions
19
3.2 Stronger Bounds for Certain Parameter Values
A linear LRC, or more generally a matroid, that achieves any of the Singleton-type
bounds given above will henceforth be called Singleton-optimal.
Any (n, k, d, r, δ )i -matroid M satisfies that δ ≤ d. Hence, by the bound (5),
k ≤ n − (δ − 1)d kr e for M. Thus, regardless of the global minimum distance d, any
(n, k, d, r, δ )−LRC with either information or all-symbol locality, has parameters
n, k, r, δ in the set
k
4
. (7)
P(n, k, r, δ ) = (n, k, r, δ ) ∈ Z : 2 ≤ δ and 0 < r ≤ k ≤ n − (δ − 1)
r
A very natural question to ask then is for which parameters (n, k, r, δ ) ∈ P(n, k, r, δ )
there exists a Singleton-optimal matroid or linear LRC, regarding both information and all-symbol locality. We remark that existence results on Singleton-optimal
linear LRCs imply existence results on Singleton-optimal matroids. Conversely,
nonexistence results on Singleton-optimal matroids implies nonexistence results on
Singleton-optimal linear LRCs.
When considering information-symbol locality it is known that the upper bound
for d given in (5) is achieved for all parameters (n, k, r, δ ) ∈ P(n, k, r, δ ) by linear
LRCs over sufficient large fields. This follows from [14], where a new class of codes
called pyramid codes was given. Using this class of codes, Singleton-optimal linear
(n, k, d, r, δ )i -LRCs can be constructed for all parameters in P(n, k, r, δ ).
It is well known that Singleton-optimal linear (n, k, d, r, δ )a -LRCs exist when
r = k. Namely, the LRCs in these cases are linear [n, k, n − k + 1] MDS-codes.
However, existence or nonexistence results when r < k are in general not that easy
to obtain. In [38], existence and nonexistence results on Singleton-optimal linear
(n, k, d, r, δ )a -LRCs were examined. Such results were given for certain regions of
parameters, leaving other regions for which the answer of existence or nonexistence of Singleton-optimal linear LRCs is not known. The results on nonexistence
were extended to matroids in [47]. All the parameter regions for the nonexistence
of Singleton-optimal linear LRCs in [47] were also regions of parameters for the
nonexistence of Singleton optimal matroids for all-symbol locality. Further, more regions of parameters for nonexistence of Singleton-optimal matroids with all-symbol
locality were given in [47]. This implies new regions of parameters for nonexistence
of Singleton-optimal linear LRCs with all-symbol locality.
The nonexistence results for Singleton-optimal matroids were proven via the following structure result in [47]. Before we state the theorem we need the concept of
nontrivial unions. Let M = (ρ, E) be a matroid with repair sets {Rx }x∈E . For Y ⊆ E,
we say that
RY =
[
Rx
and
RY is a nontrivial union if Rx * RY \{x} for every x ∈ Y .
x∈Y
Theorem 6 ([47] Structure theorem for Singleton-optimal matroids). Let M =
(ρ, E) be an (n, k, d, r, δ )a -matroid with r < k, repair sets {Rx }x∈E and
20
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
d = n−k+1−
k
− 1 (δ − 1).
r
Then, the following properties must be satisfied by the collection of repair sets and
the lattice of cyclic flats Z of M:
(i) 0Z = 0,
/
(ii) for each x ∈ E,
a) Rx is an atom of Z ,
(i.e., Rx ∈ Z , 0Z < Rx and @Z ∈ Z such that 0Z < Z < Rx ),
b) η(Rx ) = δ − 1,
(iii) for each Y ⊆ E with RY being a nontrivial union,
c) |Y | < d kr e ⇒ RY ∈ Z ,
d) |Y | < d kr e ⇒ ρ(RY ) = |RY | − |Y |(δ − 1),
T
e) |Y | ≤ d kr e ⇒ |Rx (RY \{x} )| ≤ |Rx | − δ , for each x ∈ Y,
f ) |Y | ≥ d kr e ⇒ {Z ∈ Z : Z ⊇ RY } = 1Z = E.
Conditions (i) and (ii) in the structure theorem above for Singleton-optimal matroids show that each repair set Rx must correspond to a uniform matroid with |Rx |
elements and rank |Rx | − (δ − 1). Further, condition (iii) gives structural properties on nontrivial unions of repair sets. This can be viewed as structural conditions on how nontrivial unions of uniform matroids need to be glued together in
a Singleton-optimal matroid, with the uniform matroids corresponding to repair
sets. For Singleton-optimal linear (n, k, d, r, δ )a -LRCs, the property of the repair
sets being uniform matroids corresponds to the repair sets being linear [|Rx |, |Rx | −
(δ − 1), δ ]-MDS codes. We remark that structure theorems when r|k for Singletonoptimal linear (n, k, d, r, δ )a -LRCs and the special case of (n, k, d, r, δ = 2)a -LRCs
are given in [18] and [13], respectively. These theorems show that the local repair
sets correspond to linear [r + δ − 1, r, δ ]-MDS codes that are mutually disjoint. This
result is a special case of Theorem 6.
Example 10. By Theorem 1, the poset with its associated subsets of E = [16] and
rank of these subsets in the figure below defines the set of cyclic flats Z and the rank
function restricted to the sets in Z of a matroid M on E. From Theorem 4, we obtain
that (n, k, d) = (16, 7, 6) for the matroid M. Choosing repair sets R1 = · · · = R5 = X1 ,
R6 = · · · = R9 = X2 , R10 = · · · = R13 = X3 and R14 = · · · = R16 = X4 , we obtain that
M is an (n = 16, k = 7, d = 6, r = 3, δ = 3)a -matroid. It can easily be checked that
all the properties (i)-(iii) are satisfied by the matroid M and the chosen repair sets.
Further, we also have that M is Singleton-optimal as
k
n−k+1−
− 1 (δ − 1) = 6 = d.
r
Matroid Theory and Storage Codes: Bounds and Constructions
21
([16], 7)
1Z
({1 − 9}, 5)
Y1
({1 − 5, 13 − 16}, 5)
Y3
Y2
({1, 5 − 9, 13 − 16}, 6)
Y5
Y4
({1 − 5, 9 − 13}, 6)
({5 − 13}, 5)
({5 − 9}, 3)
X1
({1 − 5}, 3)
Y6
({1, 9 − 16}, 5)
({1, 13 − 16}, 3)
X3
X2
X4
({9 − 13}, 3)
0Z
(0,
/ 0)
4 Code Constructions
4.1 Constructions of (n, k, d, r, δ )a -matroids via Cyclic Flats
In [47], a construction of a broad class of linear (n, k, d, r, δ )a -LRCs is given via
matroid theory. This is generalised in [46] and [10] to account for availability and
hierarchy, respectively.
A construction of (n, k, d, r, δ )a -matroids via cyclicSflats:
Let F1 , . . . , Fm be a collection of finite sets and E = m
i=1 Fi . Assign a function ρ :
{Fi } ∪ {E} → Z satisfying
(i) 0 < ρ(Fi ) < |Fi | for i ∈ [m],
(ii) ρ(Fi ) < ρ(E) for i ∈ [m],
(iii) ρ(E) ≤ |E| − ∑i∈[m] η(Fi ),
(iv) j ∈ [m] ⇒ |F[m]\{ j} ∩ Fj | < ρ(Fj ),
where
η(Fi ) = |Fi | − ρ(Fi ) for i ∈ [m] and
FI =
[
i∈I
Extend ρ to {FI } → Z by
Fi for I ⊆ [m].
(8)
22
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
ρ(FI ) = min{|FI | − ∑ η(Fi ), ρ(E)}
(9)
i∈I
and let Z be the following collection of subsets of E,
Z = {FI : I ⊆ [m] and ρ(FI ) < ρ(E)} ∪ E.
(10)
Theorem 7 ([47] Construction of S
(n, k, d, r, δ )a -matroids). Let F1 , . . . , Fm , be a
collection of finite sets with E = m
i=1 Fi and ρ : {Fi }i∈[m] → Z satisfying (8).
Then the pair (ρ, Z ), defined in (9) and (10), defines a (n, k, d, r, δ )a -matroid
M(F1 , . . . , Fm ; ρ) on E for which Z is the collection of cyclic flats, ρ is the rank
function restricted to the cyclic flats in Z , F1 , . . . , Fm , are the repair sets and
(i) n = |E|,
(ii) k = ρ(E),
(iii) d = n − k + 1 − max{∑i∈I η(Fi ) : FI ∈ Z \ E},
(iv) δ = 1 + min{η(Fi ) : i ∈ [m]},
(v) r = max{ρ(Fi ) : i ∈ m}.
That M(F1 , . . . , Fm ; ρ) defines a matroid follows from a proof given in [47] that
the pair (ρ, Z ) satisfies the axiomatic scheme of matroids via cyclic flats and their
ranks stated in Theorem 1. The correctness of the parameters (n, k, d, r, δ ) when
F1 , . . . , Fm are considered as the repair sets also follows from [47].
We remark, that the matroids constructed in Theorem 7 satisfy, for all unions of
repair sets FI with ρ(FI ) < ρ(E), that
(i) FI is a cyclic flat,
(ii) the nullity η(FI ) of FI is as small as possible.
(11)
Properties (i) and (ii) above are trivially seen to be fulfilled by uniform matroids
Uk,n , where Z = {0,
/ E}, ρ(0)
/ = 0 and ρ(E) = k. However, uniform matroids cannot
be constructed by Theorem 7, since all constructed matroids by this theorem have
r < k and uniform matroids have r = k. Though both uniform matroids and the
matroids constructed in Theorem 7 satisfy properties (i) and (ii) in (11), we will
consider them in terms of a class of matroids M , defined as follows:
M = {M = M(F1 , . . . , Fm ; ρ) : M is constructed in Theorem 7} ∪ {Uk,n }.
(12)
By the structure Theorem 6, the properties (i) and (ii) in (11) are necessary (but not
sufficient) for Singleton-optimal (n, k, d, r, δ )a -matroids.
Example 11. Let E = [12] and let F1 = {1, . . . , 4}, F2 = {3, . . . , 6}, F3 = {7, . . . , 10},
F4 = {10, . . . , 12} with ρ(F1 ) = ρ(F2 ) = ρ(F3 ) = 3, ρ(F4 ) = 2, and ρ(E) = 7.
Then, by Theorem 7, M(F1 , . . . , Fm ; ρ) is an (n, k, d, r, δ )a -matroid over E with
(n, k, d, r, δ ) = (12, 7, 3, 3, 2) and the following lattice of cyclic flats and their ranks.
Matroid Theory and Storage Codes: Bounds and Constructions
23
ρ(1Z ) = 7
ρ =6
ρ =4
ρ =6
ρ =5
ρ =6
ρ =5
ρ(F1 ) = 3
ρ(F2 ) = 3
ρ(F3 ) = 3
ρ(F4 ) = 2
ρ =4
ρ(0)
/ =0
Further, the matroid is not Singleton-optimal since
k
− 1 (δ − 1) = 4.
d = 3 < n−k+1−
r
4.2 A Matroidal Construction of Linear All Symbol LRCs
As will be explained below, all matroids constructed in Theorem 6 are contained in
a class of matroids called gammoids. These matroids are linear, which especially
implies that all (n, k, d, r, δ )a -matroids constructed by Theorem 6 are matroids associated with linear (n, k, d, r, δ )a -LRCs.
Definition 17. Any (finite) directed graph Γ = (V, D) and vertex subsets E, T ⊆ V
define a gammoid M(Γ , E, T ), where M(Γ , E, T ) = (I , E) is a the matroid with
I = {X ⊆ E : ∃ a set of |X| vertex-disjoint paths from X to T }.
Theorem 8 ([20]). Every gammoid M(Γ , E, T ) is Fq -linear for all prime powers
q ≥ 2|E| .
In [47], it is proven that the matroids constructed in Theorem 7 are indeed gammoids, and hence representable. This is achieved by explicitly constructing a triple
(Γ , E, T ) whose associated matroid is M(F1 , . . . , Fm ; ρ). The details of the construc-
24
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
tion are left to Theorem 14 in the appendix. The essence of the argument is to construct a graph of depth three, whose sources correspond to the ground set of the
matroid, and whose middle layer corresponds to the repair sets, with multiplicities
to reflect the ranks of the repair sets.
Example 12. The following directed graph Γ = (V = E ∪ H ∪ T, D) is constructed
in Theorem 14 from the matroid M(F1 , . . . , Fm , E; k; ρ) given in Example 11.
T:
H:
R1
E:
1
R{1,2}
2
3
R{1,2}
5
4
R3
R2
7
6
R3
9
8
R{3,4}
R4
10
12
11
In general it is extremely hard to prove that a matroid is linear (or the converse).
There is no known deterministic algorithm to solve this problem in general. However, by combining the results given in Theorems 7–14, we obtain the following
result.
Theorem 9 ([47] A matroidal construction of (n, k, d, r, δ )a -LRCs). For every
(n, k, d, r, δ )a -matroid M(F1 , . . . , Fm , E; k; ρ) given by Theorem 7 and every prime
power q ≥ 2|E| there is a linear (n, k, d, r, δ )a -LRC C over Fq with repair sets
F1 , . . . , Fm such that M(F1 , . . . , Fm , E; k; ρ) = MC .
Example 13. The (12, 7, 3, 3, 2)a -matroid M(F1 , . . . , Fm , E; k; ρ) given in Example 11
equals the matroid MC = M[G], where G equals the following matrix over F5 :
1
1
0
0
G= 0
0
0
0
2
1
2
3
0
0
0
0
3
0
1
0
0
0
0
0
4
0
0
1
0
0
0
0
5
0
0
0
1
0
0
0
6
0
1
3
2
0
0
0
7
0
0
0
0
1
0
0
8
0
0
0
0
0
1
0
9
0
0
0
0
1
2
3
10
0
0
0
0
0
0
1
11
1
1
1
1
1
1
1
12
3
3
3
3
3
3
1
Hence, the code C generated by the rows of G is a linear (12, 7, 3, 3, 2)-LRC over
F5 with repair sets F1 = {1, 2, 3, 4}, F2 = {3, 4, 5, 6}, F1 = {7, 8, 9, 10} and F1 =
{10, 11, 12}.
Note that the bound q ≥ 2|E| given in Theorem 9 is a very rough bound. There
are many matroids M(F1 , . . . , Fm , E; k; ρ) = MC for linear LRCs C over Fq where
q 2|E| . In Example 13, for instance, we constructed a code over F5 , while the field
Matroid Theory and Storage Codes: Bounds and Constructions
25
size predicted by Theorem 9 was 212 = 4096 5. To construct an explicit linear
(n, k, d, r, δ )a -LRC from a matroid M(F1 , . . . , Fm , E; k; ρ), one can use the directed
graph representation of the matroid given in Theorem 14, together with results on
how to construct a generator matrix from this representation [20].
As we saw earlier, it is known that there exists a Singleton-optimal linear
(n, k, d, r, δ )i -LRC for all parameters (n, k, r, δ ) ∈ P(n, k, r, δ ) (cf. (7) for a definition
of P(n, k, r, δ )). Further, it is also known that if r = k, then all Singleton-optimal linear LRCs are linear [n, k, n − k + 1]-MDS codes. In [38] existence and nonexistence
of Singleton-optimal linear (n, k, d, r, δ )a -LRCs were examined. The parameter regions for existence given in [38] were both obtained and extended in [47] by the construction of linear LRCs via matroid theory given in Theorem 9. Hence, the results
in [47] about nonexistence and existence of Singleton-optimal linear (n, k, d, r, δ )a LRCs settled large portions of the parameter regions left open in [38] leaving open
only a minor subregion. Some improvements of the results in [38] were also given
ofr δ = 2 in [45] via integer programming techniques.
For (n, k, r, δ ) ∈ P(n, k, r, δ ) it is also very natural to ask what is the maximal
value of d for which there exist an (n, k, d, r, δ )a -matroid or a linear (n, k, d, r, δ )a LRC. We denote this maximal value by dmax (n, k, r, δ ). In [47] it was proven that
k
(δ − 1)
dmax (n, k, r, δ ) ≥ n − k + 1 −
r
for linear LRCs. For
matroids, this result is straightforward, as a matroid with
d = n − k + 1 − kr (δ − 1) can be constructed as a truncation of the direct sum
of d r+δn−1 e uniform matroids of size si ≤ r + δ − 1 and rank si − δ + 1 ≤ r. As
representability (over some field) is preserved under direct sums and truncation, the
result follows for linear LRCs. However, with this straightforward argument, and
with the bound on the field size of truncated matroids from [17], the field size required could be as large as
r·d r+δn−1 e
(r + δ − 1) ·
∏
i=k
n
.
i
Significant work is needed in order to bound the field size even in this special case.
This result was improved in [47] and further in [29]. Also, the parameter region of Singleton-optimal linear (n, k, d, r, δ )-LRCs was also extended in [29]. The
existence of Singleton-optimal linear LRCs obtained by the matroidal construction
described here depends mainly on the relation between the parameters a and b where
a = d kr er − k and b = d r+δn−1 e(r + δ − 1) − n. Thus, we can easily get Singletonoptimal linear LRCs for all possible coding rates.
26
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
4.3 Random Codes
An alternative way to design (n, k, d, r, δ )-LRCs with prescribed parameters is by
exploiting the fact that independence is a generic property for r- and k-tuples of
vectors over large fields. This allows us to use randomness to generate (n, k, d, r, δ )LRCs in a straightforward way, once the matroid structure of the code is prescribed.
This is the key element in [8]. As opposed to in the gammoid construction from the
last section, we will now consider the field size q to be fixed but large. Indeed, a
sufficiently large field will be Fq with
r
q > (rδ )4 r
r
n + (rδ )(r−1)4
.
k−1
For given (n, k, r, δ ), we will construct (n, k, d, r, δ )-LRCs where
k
d ≥ n−k+1−
(δ − 1).
r
Comparing this to the generalised Singleton bound (5), we notice that the codes we
construct are “almost Singleton-optimal”.
The underlying matroid will again be a truncation of
d r+δn−1 e
G
Ussii −δ −1 .
i=1
However, rather than first representing this direct sum, which has rank
n
∑(si − δ + 1) = n − d r + δ − 1 e(δ − 1),
i
we will immediately represent its truncation as an n × k matrix. The random construction proceeds as follows. Divide the columns [n] into locality sets Si of size si .
For each Si , we first generate the ri = si − δ + 1 first columns uniformly at random
from the ambient space Fk . This gives us an ri × k-matrix Gi . After this, we draw
δ − 1 vectors from Fri , and premultiply these by Gi . The resulting ri + δ − 1 = si
vectors will be in the linear span of the ri first vectors, and so have rank ≤ ri as a
point configuration in Fk . We arrange the si vectors into a matrix G0i of rank ≤ ri in
Fsi ×k . Let Ai be the event that all ri -tuples of columns in Gi are linearly independent.
It is easy to see that, if the field size grows to infinity, the probability of Ai tends to
one.
Juxtaposing the matrices G0i for i = 1, · · · , d r+δn−1 e, we obtain a generator matrix
G for a code of length n. Let B be the event that G has full rank. Again, assuming the field size is large enough, the probability of B can be arbitrarily close to
one. Now, the random matrix G generates an (n, k, d, r, δ )-LRC if all the events
A1 , . . . , Ad n e , B simultaneously occur. A simple first moment estimate shows
r+δ −1
Matroid Theory and Storage Codes: Bounds and Constructions
that, if
r4r
q > (rδ )
27
r
n + (rδ )(r−1)4
,
k−1
then the probability of this is positive, so there exists an (n, k, d, r, δ )-LRC.
4.4 Constructing LRCs as Evaluation Codes
As suggested in the previous sections, there are several assumptions that can be
made in order to give more explicit code constructions for optimal LRCs. Next, we
will follow [41, 39] in assuming that n is divisible by r + δ − 1 and k is divisible by
r. Then, an optimal LRC with d = n − k − ( kr − 1)(δ − 1) + 1 exists for any choice
of k. We will also assume that n = q is a prime power, although this assumption can
easily be removed at the price of a more technical description of the code.
We will construct a Singleton-optimal code in this case as an evaluation code,
generalising the construction of MDS codes as Reed-Solomon codes. The main philosophy goes back to [41], but due to a technical obstacle, [41] still required exponential field size. This technicality was overcome by the construction in [39], which
we will present next. Evaluation codes have a multitude of favourable properties,
not least that the field size can often be taken to be much smaller than in naı̈ve random code constructions. Moreover, the multiplicative structure used for designing
evaluation codes can also be exploited when one needs to do computations with the
codes in question.
Let A be a subgroup of Fq of size r + δ − 1 and let g = ∏i∈A (x − i) be the polynomial of degree r + δ − 1 that vanishes on A. We will construct a storage code whose
nodes are the elements of Fq and whose locality sets are the cosets of A. Thus, there
are r+δn−1 locality sets, each of size r + δ − 1. The codewords will be the evaluations over Fq of polynomials of a certain form. As the rank of the code that we are
designing is k = r · kr , we can write the messages as a r × kr matrix
a0,0 · · · ar−1,0
..
.
..
a = ..
.
.
a0, k −1 · · · ar−1, k −1
r
r
over Fq . Now consider the polynomial function
k
fa = 1 g(x) g(x)2 · · · g(x) r −1 ·
Consider the code
1
x
x2
..
.
a0,0 · · · ar−1,0
..
..
..
·
.
.
.
.
a0, k −1 · · · ar−1, k −1
r
r
r−1
x
28
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
r× kr
C = { fa (x) : x ∈ Fq , a ∈ Fq
}.
By design, fa has degree
deg fa ≤ (r + δ − 1)
k
k
− 1 + r − 1 = k − 1 + (δ − 1)
−1 ,
r
r
and can therefore be computed for every point in Fq by evaluation on any k + (δ −
1)( kr − 1) points. Therefore, the code C protects against
k
d − 1 = n − k − (δ − 1)
−1 +1
r
errors. It remains to see that it has locality (r, δ ).
To this end, note that the row vector
a0,0 · · · ar−1,0
..
k
.
..
.
1 g(x) g(x)2 · · · g(x) r −1 · ..
.
a0, k −1 · · · ar−1, k −1
r
r
of polynomials is constant over the subgroup A ⊆ Fq and thus on all of its cosets by
construction of g. It follows that when restricted to any such coset, the function fa is
a polynomial of degree ≤ r − 1, and so can be extrapolated to all points in the coset
from any r such evaluation points. This proves the (r, δ )-locality.
As discussed, this construction depends on a collection of assumptions on the
divisibility of parameters that are needed for the rather rigid algebraic structures to
work. Some of these assumptions can be relaxed, using more elaborate evaluation
codes, such as algebraic geometry codes over curves and surfaces [3, 2]. While
this field of research is still very much developing, it seems that the rigidity of the
algebraic machinery makes it less suitable for generalisations of the LRC concept,
for example when different nodes are allowed to have different localities.
5 Beyond Linear Storage Codes
In this section we will introduce the notion of hierarchical codes, which are natural
generalisations of locally repairable codes. After this, we will briefly describe the
connection between (n, k, d, r, δ ,t)-LRCs and polymatroids given in [46].
5.1 Hierarchical Codes
Definition 18. Let h ≥ 1 be an integer, and let
Matroid Theory and Storage Codes: Bounds and Constructions
29
(n, k, d, t) = [(n1 , k1 , d1 ,t1 ), . . . , (nh , kh , dh ,th )]
be a h-tuple of integer 4-tuples, where ki ≥ 1, ni , di ≥ 2, and ti ≥ 1 for 1 ≤ i ≤ h.
Then, a coordinate x of a linear [n, k, d] = [n0 , k0 , d0 ]-LRC C indexed by E has h-level
hierarchical availability (n, k, d, t) if there are t1 coordinate sets X1 , . . . , Xt1 ⊆ E such
that
(i) x ∈ Xi for i ∈ [t1 ],
(ii) i, j ∈ [t1 ], i 6= j ⇒ Xi ∩ X j = {x},
(iii) nXi ≤ n1 , kXi = k1 and dXi ≥ d1 for the punctured
[nXi , kXi , dXi ]-code CXi , for i ∈ [t1 ],
(iv) for i ∈ [t1 ], x has (h − 1)-level hierarchical availability [(n2 , k2 , d2 ,t2 ), . . . , (nh , kh , dh ,th )] in CXi .
The code C above as well as all the related subcodes CXi should be non-degenerate.
For consistency of the definition, we say that any symbol in a non-degenerate storage
code has 0-level hierarchical availability.
Example 14. Let C be the code generated by the matrix G in Example 1 and let
x = 2. Then x has 2-level hierarchical availability
(n, k, d, t) = [(6, 3, 3, 1), (3, 2, 2, 2)] .
This follows from Example 2 where CY1 implies the (6, 3, 3, 1)-availability, and the
(3, 2, 2, 2)-availability is implied by CX1 and CX2 .
The most general Singleton bound for matroids with hierarchy in the case t = 1
are the following given in [32, 10]:
ki
−1 ,
di (M) ≤ ni − ki + 1 − ∑ (d j − d j+1 )
kj
j>i
where we say dh+1 = 1.
5.2 General Codes from Polymatroids
Definition 19. Let E be a finite set. A pair P = (ρ, E) is a (finite) polymatroid on
E with a set function ρ : 2E → R if ρ satisfies the following three conditions for all
X,Y ⊆ E:
(R1) ρ(0)
/ = 0,
(R2) X ⊆ Y ⇒ ρ(X) ≤ ρ(Y ) ,
(R3) ρ(X) + ρ(Y ) ≥ ρ(X ∪Y ) + ρ(X ∩Y ) .
Note that a matroid is a polymatroid which additionally satisfies the following
two conditions for all X ⊆ E:
(R4) ρ(X) ∈ Z ,
(R5) ρ(X) ≤ |X| .
30
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
Using the joint entropy and a result given in [11] one can associate the following
polymatroid to every code.
Definition 20. Let C be an (n, k)-code over some alphabet A of size s. Then PC =
(ρC , [n]) is the polymatroid on [n] with the set function ρC : 2[n] → R where
|C|
|{c ∈ C : cX = zX }|
logs
ρC (X) = ∑
|C|
|{c ∈ C : cX = zX }|
zX ∈CX
and ρC (0)
/ = 0.
We remark that for linear codes MC = PC . Using the above definition of PC , one
can now prove the following useful properties.
Proposition 8. Let C be an (n, k)-code over A with |A| = s. Then for the polymatroid
PC = (ρC , [n]) and any subsets X,Y ⊆ [n],
(i) PC (X) ≤ |X|,
(ii) |CX∪Y | > |CX | ⇐⇒ ρC (X ∪Y ) > ρC (X),
(iii) |C| = sρC ([n]) ,
(iv) |C|/|An | = sρC ([n])−n .
We remark that, even though |C| = sρC ([n]) for nonlinear codes and |CX | = sρC (X)
for all X for linear codes, it is not true in general that |CX | = sρC (X) for X ( [n]
for nonlinear codes. This stems from the fact that, for non-linear codes, the uniform
distribution over the code does not necessarily map to the uniform distribution under
coordinate projection.
After scaling the rank function of a finite polymatroid P = (ρ, E) by a constant c
such that cρ(X) ≤ |X| for all X ⊆ E, we obtain a polymatroid satisfying axiom (R5).
We will assume that such a scaling has been performed, so that all polymatroids
satisfy axiom (R5).
We are now ready to define a cyclic flat of a polymatroid P = (ρ, E), namely
X ⊆ E is a cyclic flat if
ρ(X ∪ {e}) > ρ(X) for all e ∈ E \ X and ρ(X) − ρ(X \ {x}) < 1 for all x ∈ X.
Let P = (ρ, E) be a polymatroid and X ⊆ E. The restriction of P to X is the
polymatorid P|X = (ρ|X , X) where ρ|X (Y ) = ρ(Y ) for Y ⊆ X. We can now define
the distance of P|X as
d(P|X) = min{|Y | : Y ⊆ X, ρ|X (X \Y ) < ρ|X (X)}.
Let Z denote the family of cyclic flats of the polymatroid P. Assuming that
E ∈ Z , we can define the parameters n, k, d of P via the cyclic flats and their ranks,
namely
n = |E|, k = ρ(E) and d = bn − k + 1 − max{|X| − ρ(X) : X ∈ Z \ E}c.
Matroid Theory and Storage Codes: Bounds and Constructions
31
The definitions of (n, k, d, r, δ ,t)i and (n, k, d, r, δ ,t)a -polymatroids are carried
over directly from Definition 16. In addition, the parameters (n, k, d, r, δ ,t)i and
(n, k, d, r, δ ,t)a of a LRC C are the same as the corresponding parameters for PC .
Using the cyclic flats and similar methods as for matroids, Singleton-type bounds
can be proven for polymatroids in general, which then imply bounds on all objects
related to polymatroids, e.g., matroids, linear and nonlinear LRCs, and hypergraphs.
This is the content of the next section.
5.3 Singleton-type Bounds for Polymatroids and General LRCs
It is not clear whether the Singleton-type bounds given for linear LRCs in (2)–(6)
also hold for general LRCs — in general the upper bound on d might have to be
larger. As we will describe briefly in Section 5, any general LRC can be associated
with a polymatroid that captures the key properties of the LRC. Using this connection we are able to define the (n, k, d, r, δ ,t)-parameters and information-symbol,
systematic-symbol, and all-symbol locality sets for polymatroids in general.
The class of polymatroids is much bigger than the class of the polymatroids
arising from general LRCs. Hence, it is also not clear whether the Singleton-type
bounds given in (2)–(6) also hold for polymatroids in general. However, from [46],
we obtain a Singleton-type bound for polymatroids in Theorem 10 below. This theorem shows that all the Singleton-type bounds given in (2)–(6) are polymatroid properties. Further, the polymatroid result also extends all these bounds by including all
the parameters (n, k, d, r, δ ,t) at the same time.
The methods used to prove the Singleton-type bound given for polymatroids in
Theorem 10 are similar to those used for proving the Singleton-type bound for matroids in Theorem 5. Especially, the notion of cyclic flats is generalised to polymatroids and used as the key tool in the proof. However, some obstacles occur since
we are dealing with real-valued rank functions in the case of polymatroids instead
of integer-valued rank functions, which was the case for matroids. As a direct consequence of Theorem 10, the Singleton-type bounds given in (2)–(6) are valid for
all objects associated to polymatroids.
Theorem 10 ([46] Singleton-type bound for polymatroids). Let P = (ρ, E) be an
information-set (n, k, d, r, δ ,t)i -polymatroid. Then
t(dke − 1) + 1
− 1 (δ − 1).
(13)
d ≤ n − dke + 1 −
t(r − 1) + 1
Theorem 10 is stated for information-symbol locality. This implies that the bound
(13) is also valid for systematic-symbol and all-symbol locality. Hence, as a direct
corollary, the bounds (2)–(13) hold for information-symbol, systematic-symbol, and
all-symbol locality for all objects associated to polymatroids, e.g., entropy functions,
general LRCs, hypergraphs, matroids, linear LRCs, graphs and many more. If we
restrict to systematic linear codes, then the bound also holds for PIR codes [9, Def.
32
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
4]. The connection is not as straightforward in the nonlinear case, since the definitions of a repair group are then slightly different for LRCs (as defined here) and PIR
codes, while coinciding in the linear case.
The bound (4), for all-symbol LRCs (as subsets of size |B|K of Bαn , where B
is a finite set, A = Bα is the alphabet, and α and K are integers), follows from a
result given in [27]. The bound (5), for all-symbol LRCs (as a linear subspace of
α
Fαn
q with the alphabet A = Fq ), is given in [33]. This result is slightly improved for
information-symbol locality in [18]. The bound (6), for (n, k, d, r,t)s -LRCs where k
is a positive integer, follows from a result given in [31]. The following bound for
(n, k, d, r,t)a -LRCs with integral k was given in [40],
t
d ≤ n−k+1− ∑
i=1
k−1
.
ri
One parameter which has not been included above is the alphabet size. Small
alphabet sizes are important in many applications because of implementation and
efficiency reasons. The bound (14) below takes the alphabet size into account, but is
only inductively formulated. Before stating this bound we introduce the following
notation:
(q)
kopt (n, d) = max{k : C is an (n, k, d)-code over an alphabet of size q}.
By [6], an all-symbol (n, k, d, r)-LRC over a finite alphabet A of size q satisfies
(q)
k ≤ min (sr + kopt (n − s(r + 1), d)).
s∈Z+
(14)
It is a hard open problem in classical coding theory to obtain a value for the param(s)
eter kopt (n, d) for linear codes. This problem seems to be even harder for codes in
general. However, by using other known bounds, such as the Plotkin bound or Greis(s)
mer bound, it is possible to give an explicit value for kopt (n, d) for some classes of
parameters (s, n, d). This has been done for example in [34].
We remark that when considering nonlinear LRCs, some extra care has to be
taken in terms of how to define the concepts associated with the LRCs. Two equivalent definitions in the linear case may differ in the nonlinear case. In this chapter,
we have chosen to consider δ as a parameter for the local distance of the repair sets,
i.e., any node in a repair set R can be repaired by any other |R| − δ + 1 nodes of R.
The condition used in [6, 27, 31, 40] is for δ = 2 only assuming that a specific node
in a repair set R can be repaired by the rest of the nodes of R. It is not assumed that
any node in R can be repaired by the other nodes of R, i.e., that the local distance is
2. A Singleton bound using the weaker condition of guaranteeing only repair of one
node in each repair set implies directly that the same upper bound on d is true for
the case with local distance 2.
Matroid Theory and Storage Codes: Bounds and Constructions
33
6 Conclusions and Further Research
We have shown how viewing storage codes from a matroidal perspective helps
our understanding of local repairability, both for constructions and for fundamental
bounds. However, many central problems about linear LRCs boil down to notoriously hard representability problems in matroid theory.
A famous conjecture, with several consequences for many mathematical objects,
is the so called MDS-conjecture. This conjecture states that, for a given finite field Fq
and a given k, every [n, k, d]-MDS code over Fq has n ≤ q + 1, unless in some special
cases. Currently, the conjecture is known to hold only if q is a prime [1]. Linear
Singleton-optimal LRCs may be seen as a generalisation of linear MDS codes. An
interesting problem would therefore be to consider an upper bound on n for linear
Singleton-optimal LRCs over a certain field size q with fixed parameters (k, r, δ ,t).
In this setting, a sufficiently good upper bound on n would be a good result.
Instead of fixing the Singleton-optimality and trying to optimise the field size, we
could also fix the field Fq , and try to optimise the locality parameters. This would
give us bounds on the form
k
− 1 (δ − 1) − p(q, n, k, r, δ ),
d ≤ n−k+1−
r
where the dependence on the field size q is isolated to a “penalty” term p(q, n, k, r, δ ).
Partial results in this direction are given by the Cadambe-Mazumdar bound [6], and
LRC versions of the Griesmer and Plotkin bounds [34]. However, the optimality of
these bounds is only known for certain ranges of parameters. Further research in this
direction is definitely needed, but seems to lead away from the most obvious uses
of matroid theory.
Finally, it would be interesting to characterise all Singleton-optimal LRCs up to
matroid isomorphism. The constructions discussed in this paper appear to be rather
rigid, and unique up to shifting a few “slack” elements between different locality sets. However, it appears to be difficult to prove that all Singleton-optimal matroids must have this form. Once a complete characterisation of Singleton-optimal
matroids has been obtained, this could also be taken as a starting point for possibly finding Singleton-optimal nonlinear codes in the parameter regimes where no
Singleton-optimal linear codes exist.
References
1. Ball, S.: On sets of vectors of a finite vector space in which every subset of basis size is a
basis. Journal of the European Mathematical Society 14, 733–748 (2012)
2. Barg, A., Haymaker, K., Howe, E., Matthews, G., Várilly-Alvarado, A.: Locally recoverable
codes from algebraic curves and surfaces (2017). ArXiv: 1701.05212
3. Barg, A., Tamo, I., Vlăduţ, S.: Locally recoverable codes on algebraic curves (2016). ArXiv:
1603.08876
34
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
4. Bonin, J.E., de Mier, A.: The lattice of cyclic flats of a matroid. Annals of combinatorics 12,
155–170 (2008)
5. Britz, T., Rutherford, C.G.: Covering radii are not matroid invariants. Discrete Mathematics
296, 117–120 (2005)
6. Cadambe, V., Mazumdar, A.: An upper bound on the size of locally recoverable codes. In:
International Symposium on Network Coding, pp. 1–5 (2013)
7. Dimakis, A., Godfrey, P.B., Wu, Y., Wainwright, M.J., Ramchandran, K.: Network coding
for distributed storage systems. IEEE Transactions on Information Theory 56(9), 4539–4551
(2010)
8. Ernvall, T., Westerbäck, T., Freij-Hollanti, R., Hollanti, C.: Constructions and properties of
linear locally repairable codes. IEEE Transactions on Information Theory 62, 5296–5315
(2016)
9. Fazeli, A., Vardy, A., Yaakobi, E.: Pir with low storage overhead: coding instead of replication.
arXiv:1505.06241 (2015)
10. Freij-Hollanti, R., Westerbäck, T., Hollanti, C.: Locally repairable codes with availability and
hierarchy: matroid theory via examples. In: International Zürich Seminar on Communications,
pp. 45–49. IEEE/ETH (2016)
11. Fujishige, S.: Polymatroidal dependence structure of a set of random variables. Information
and control 39(1), 55–72 (1978)
12. Geelen, J., Gerards, B., Whittle, G.: Solving Rota’s conjecture. Notices of the American
Mathematical Society 61(736–743) (2014)
13. Gopalan, P., Huang, C., Simitci, H., Yekhanin, S.: On the locality of codeword symbols. IEEE
Transactions on Information Theory 58(11), 6925–6934 (2012)
14. Huang, C., Chen, M., Lin, J.: Pyramid codes: Flexible schemes to trade space for access efficiency in reliable data storage systems. In: International Symposium on Network Computation
and Applications, pp. 79–86. IEEE (2007)
15. Ingleton, A., Main, R.: Non-algebraic matroids exist. Bulletin of the London Mathematical
Society 7(144–146) (1975)
16. Ishai, Y., Kushilevitz, E., Ostrovsky, R., Sahai, A.: Batch codes and their applications. In: The
36th ACM Symposium on Theory of Computing (STOC) (2004)
17. Jurrius, R., Pellikaan, R.: Truncation formulas for invariant polynomials of matroids and geometric lattices. Mathematics in Computer Scinece 6, 121–133 (2012)
18. Kamath, G.M., Prakash, N., Lalitha, V., Kumar, P.V.: Codes with local regeneration and erasure correction. IEEE Transactions on Information Theory 60(8), 4637–4660 (2014)
19. Knuth, D.: The asymmetric number of geometries. Journal of Combinatorial Theory, Series
A 16, 398–400 (1974)
20. Lindström, B.: On the vector representations of induced matroids. Bulletin of the London
Mathematical Society 5, 85–90 (1973)
21. Lindström, B.: On p-polynomial representations of projective geometries in algebraic combinatorial geometries. Mathematica Scandinavica 63, 36–42 (1988)
22. Lindström, B.: On algebraic matroids. Discrete Mathematics 111(357–359) (1993)
23. Lipmaa, H., Skachek, V.: Linear batch codes. In: The 4th International Castle Meeting on
Coding Theory and Applications (4ICMCTA) (2015)
24. Mayhew, D., Newman, M., Welsh, D., Whittle, G.: On the asymptotic proportion of connected
matroids. European Journal of Combinatorics 32(6), 882–890 (2011)
25. Mayhew, D., Newman, M., Whittle, G.: Yes, the missing axiom of matroid theory is lost
forever (2015). ArXiv: 1412.8399
26. Nelson, P.: Almost all matroids are non-representable. ArXiv: 1605.04288
27. Papailiopoulos, D., Dimakis, A.: Locally repairable codes. In: International Symposium on
Information Theory, pp. 2771–2775. IEEE (2012)
28. Pendavingh, R., van der Pol, J.: On the number of matroids compared to the number of sparse
paving matroids. Electronic Journal of Combinatorics 22, 17pp. (2015)
29. Pöllänen, A., Westerbäck, T., Freij-Hollanti, R., Hollanti, C.: Improved singleton-type bounds
for locally repairable codes. In: International Symposium on Information Theory, pp. 1586–
1590. IEEE (2016)
Matroid Theory and Storage Codes: Bounds and Constructions
35
30. Prakash, N., Kamath, G.M., Lalitha, V., Kumar, P.V.: Optimal linear codes with a local-errorcorrection property. In: International Symposium on Information Theory, pp. 2776–2780.
IEEE (2012)
31. Rawat, A.S., Papailiopoulos, D., Dimakis, A., Vishwanath, S.: Locality and availability in
distributed storage (2014). ArXiv: 1402.2011v1
32. Sasidharan, B., Agarwal, G.K., Kumar, P.V.: Codes with hierarchical locality (2015). ArXiv:
1501.06683v1
33. Silberstein, N., Rawat, A.S., Koyluoglu, O., Vishwanath, S.: Optimal locally repairable codes
via rank-metric codes. In: International Symposium on Information Theory, pp. 1819–1823.
IEEE (2013)
34. Silberstein, N., Zeh, A.: Optimal binary locally repairable codes via anticodes (2015). ArXiv:
1501.07114v1
35. Sims, J.A.: Some problems in matroid theory. Ph.D. thesis, Oxford University (1980)
36. Singleton, R.C.: Maximum distance q-nary codes. IEEE Transactions on Information Theory
10(2), 116–118 (1964)
37. Skorobogatov, A.: Linear codes, strata of grassmannians, and the problems of segre. In: International Workshop on Coding Theory and Algebraic Geometry, pp. 210–223 (1992)
38. Song, W., amd C. Yuen, S.H.D., Li, T.J.: Optimal locally repairable linear codes. IEEE Journal
on Selected Areas in Communications 32(5), 1019–1036 (2014)
39. Tamo, I., Barg, A.: A family of optimal locally recoverable codes. IEEE Transactions on
Information Theory 60(8), 4661–4676 (2014)
40. Tamo, I., Barg, A., Frolov, A.: Bounds on the parameters of locally recoverable codes. IEEE
Transactions on Information Theory 62(6), 3070–3083 (2016)
41. Tamo, I., Papailiopoulos, D., Dimakis, A.: Optimal locally repairable codes and connections
to matroid theory. In: International Symposium on Information Theory, pp. 1814–1818. IEEE
(2013)
42. Tutte, W.: A homotopy theorem for matroids, I, II. Transactions of the Amarican Mathematical
Society 88, 148–178 (1958)
43. Vámos, P.: The missing axiom of matroid theory is lost forever. Journal of the London Mathematical Society 18, 403–408 (1978)
44. Wang, A., Zhang, Z.: Repair locality with multiple erasure tolerance. IEEE Transactions on
Information Theory 60(11), 6979–6987 (2014)
45. Wang, A., Zhang, Z.: An integer programming-based bound for locally repairable codes. IEEE
Transactions on Information Theory 61(10), 5280–5294 (2015)
46. Westerbäck, T., Freij, R., Hollanti, C.: Applications of polymatroid theory to distributed storage systems. In: Allerton Conference on Communication, Control, and Computing, pp. 231–
237 (2015)
47. Westerbäck, T., Freij-Hollanti, R., Ernvall, T., Hollanti, C.: On the combinatorics of locally
repairable codes via matroid theory. IEEE Transactions on Information Theory 62, 5296–
5315 (2016)
48. Whitney, H.: On the abstract properties of linear dependence. American Journal of Mathematics 57, 509–533 (1935)
49. Zhang, H., Skachek, V.: Bounds for batch codes with restricted query size. In: IEEE International Symposium on Information Theory (ISIT) (2016)
Appendix: More about Matroid Theory
A matroid realisation of an F-linear matroid M has two geometric interpretations.
Firstly, we may think of a matrix representing M as a collection of n column vectors
in Fk . As the matroid structure is invariant under row operations, or in other words
36
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
under change of basis in Fk , we tend to think of M as a configuration of n points in
abstract projective k-space.
The second interpretation comes from studying the row space of the matrix, as
an embedding of Fk into Fn . Row operations correspond to a change of basis in Fk ,
and hence every matroid representation can be thought of as a k-dimensional subspace of Fn . In other words, a matroid representation is a point in the Grassmannian
Gr(n, k; F), and Gr(n, k; F) has a stratification as a union of realisation spaces R(M),
where M ranges over all F-representable matroids of size n and rank k. This perspective allows a matroidal perspective also on the subspace codes discussed in Chapter
1–4, where the codewords themselves are matroid representations. However, so far
this perspective has not brought any new insights to the topic.
Another instance where matroids appear naturally in mathematics is graph theory. Let Γ be a finite graph with edge set E. We obtain a matroid MΓ = (I , E),
where I ⊆ E is independent if the subgraph ΓI ⊆ Γ induced on I ⊆ E is a forest, i.e.,
has no cycles. A matroid that is isomorphic to MΓ for some graph Γ is said to be a
graphical matroid.
Example 15. The matrix G and the graph Γ given below generate the same matroid,
regardless of the field over which G is defined.
123
100
G= 0 1 0
001
000
4 5 6
0 0 1
0 0 1
0 -1 -1
1 1 0
7
1
1 ,
0
-1
Γ=
1
3
4
2
6
5
7
Some examples of independent sets in G and Γ are {3, 4, 6}, {1, 2, 3, 5}, {2, 3, 4, 6}.
The set X = {5, 6, 7} is dependent in MΓ as these edges form a cycle, and it is
dependent in MG as the submatrix
5
0
G(X) = 0
-1
1
6
1
1
-1
0
7
1
1
0
-1
has linearly dependent columns.
Indeed, graphical matroids are representable over any field F. To see this, for a
graph Γ with edge set E, we will construct a matrix G(Γ ) over F with column set
E as follows. Choose an arbitrary spanning forest T ⊆ E in Γ , and index the rows
of G(Γ ) by T . Thus G(Γ ) is a T × E-matrix. Choose an arbitrary orientation for
each edge in the graph. For e ∈ T ⊆ E and uv ∈ E, the entry in position (e, {uv}) is
1 (respectively −1) if e is traversed forward (respectively backward) in the unique
Matroid Theory and Storage Codes: Bounds and Constructions
37
path from u to v in the spanning forest T . In particular, the submatrix G(Γ )(T ) is an
identity matrix. It is straightforward to check that the independent sets in G(Γ ) are
exactly the noncyclic sets in Γ .
Example 16. The matrix G in Example 15 is G(Γ ) where Γ is the graph in the same
example, and the spanning forest T is chosen to be {1, 2, 3, 4}.
The restriction to X ⊆ E of a graphical matroid MΓ is obtained by the subgraph
of Γ containing precisely the edges in X.
A third example of matroids occurring naturally
in mathematics are algebraic matroids [22]. These
are associated to field extensions F : K together
with a finite point sets E ⊆ K, where the independent sets are those I ⊆ E that are algebraically independent over F. In particular, elements that are algebraic over F have rank zero, and in general ρ(I)
is the transcendence degree of the field extension
F(I) : F.
It is rather easy to see that every F-linear ma- Fig. 2 The Vamos matroid of size
troid is also algebraic over F. Indeed, let X1 , · · · , Xk 8 and rank 4, which is not algebraically representable.
be indeterminates, and let
g : Fk → F(X1 , · · · , Xk )
be given by ei 7→ Xi for i = 1, · · · k. Then J ⊆ E is linearly independent over F if
and only if {g( j) : j ∈ J} is algebraically independent over F. Over fields of characteristic zero the converse also holds, so that all algebraic matroids have a linear
representation. However, in positive characteristic there exist algebraic matroids that
are not linearly representable. For example, the non-Pappus matroid of Example 4
is algebraically representable over F4 , although it is not linearly representable over
any field [21]. The smallest example of a matroid that is not algebraic over any field
is the Vamos matroid, in Figure 2 [15].
Definition 21. The dual of M = (ρ, E) is M ∗ = (ρ ∗ , E), where
ρ ∗ (X) = |X| + ρ(E \ X) − ρ(E).
The definition of the dual matroid lies in the heart of matroid theory, and has profound interpretations. In geometric terms, let M be represented by a k-dimensional
subspace V of Fn . Then, the matroid dual M ∗ is represented by the orthogonal complement V ⊥ ⊆ Fn . Surprisingly and seemingly unrelatedly, if Γ is a planar graph
and M = MΓ is a graphical matroid, then M ∗ = MΓ̄ , where Γ̄ is the planar dual of
Γ . Moreover, the dual MΓ∗ of a graphical matroid is graphical if and only if Γ is
planar.
Definition 22. The contraction of X ⊆ E in the matroid M = (ρ, E) is M/X =
(ρ 0 , M \ X), where ρ 0 (Y ) = ρ(Y ∪ X) − ρ(X).
38
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
∗
Contraction is the dual operation of deletion, in the sense that M/X = (M|E\X
)∗ .
The terminology comes from graphical matroids, where contraction of the edge
e ∈ E corresponds to deleting e and identifying its endpoints in the graph. Notice
that it follows directly from submodularity of the rank function that ρM/X (Y ) ≤
ρM|E\X (Y ) for every Y ⊆ E \ X. In terms of subspace representations, contraction of
e ∈ E corresponds to intersecting the subspace that represents M with the hyperplane
{xe = 0}.
As matroids are used as an abstraction for linear codes, it would be desirable to
have a way to go back from matroids to codes, namely to determine whether a given
matroid is representable, and when it is, to find such a representation. Unfortunately,
there is no simple criterion to determine representability [43, 25]. However, there
are a plethora of sufficient criteria to prove nonrepresentability, both over a given
field and over fields in general. In recent years, these methods have been used to
prove two long-standing conjectures, that we will discuss in Sections 6.1 and 6.2
respectively.
6.1 Rota’s Conjecture
While there is no simple criterion to determine linear representability, the situation
is much more promising if we consider representations over a fixed field. It has
been known since 1958, that there is a simple criterion for when a matroid is binary
representable.
Theorem 11 ([42]). Let M = (ρ, E) be a matroid. The following two conditions are
equivalent.
1. M is linearly representable over F2 .
2. There are no sets X ⊆ Y ⊆ E such that M|Y /X is isomorphic to the uniform
matroid U42 .
In essence, this means that the only obstruction that needs to be overcome in order to be representable over the binary alphabet, is that no more than three nonzero
points can fit in the same plane. For further reference, we say that a minor of the
matroid M = (ρ, E) is a matroid of the form M|Y /X, for X ⊆ Y ⊆ E. Clearly, if M is
representable over F, then so is all its minors. Let L(F) be the class of matroids that
are not representable over F, but such that all of their minors are F-representable.
Then the class of F-representable matroids can be written as the class of matroids
that does not contain any matroid from L(F) as a minor. Gian-Carlo Rota conjectured in 1970 that L(F) is a finite set for all finite fields F. A proof of this conjecture
was announced by Geelen, Gerards and Whittle in 2014, but the details of the proof
still remain to written up [12].
Theorem 12. For any finite field F, there is a finite set L(F) of matroids such that
any matroid M is representable if and only if it contains no element from L(F) as a
minor.
Matroid Theory and Storage Codes: Bounds and Constructions
39
Since the 1970’s, it has been known that a matroid is representable over F3 if and
only if it avoids the uniform matroids U52 , U53 , the Fano plane P2 (F2 ), and its dual
P2 (F2 )∗ as minors. The list L(F4 ) has seven elements, and was given explicitly in
2000. For larger fields, the explicit list is not known, and there is little hope to even
find useful bounds on its size.
6.2 Most Matroids are Nonrepresentable
For a fixed finite field F, it follows rather immediately from the minor-avoiding
description in the last section that the fraction of n-symbol matroids that is Frepresentable goes to zero as n → ∞. It has long been a folklore conjecture that
this is true even when representations over arbitrary fields are allowed. However, it
was only in 2016 that a verifiable proof of this claim was announced [26].
Theorem 13.
lim
n→∞
#linear matroids on n elements
= 0.
#matroids on n elements
The proof is via estimates of the denominator and enumerator of the expression
in 13 separately. Indeed, it is shown in [19] that the number of matroids on n nodes
n
is at least Ω (2(2−ε) ) for every ε > 0. The proof of Theorem 13 thus boiled down to
3
proving that the number of representable matroids is O(2n ). This is in turn achieved
by bounding the number of so called zero-patterns of polynomials.
6.3 Gammoid Construction of Singleton-Optimal LRCs
For completeness, we end this appendix with a theorem that explicitly presents the
matroids constructed in Theorem 7 as gammoids. As discussed in Section4.2, this
proves the existence of Singleton-optimal linear LRCs whenever a set system satisfying (8) exists.
Theorem 14 ([47], M(F1 , . . . , Fm , E; k; ρ)-matroids are gammoids). Let M(F1 , . . . , Fm ; ρ)
be a matroid given by Theorem 7 and define s : E → 2[m] where s(x) = {i ∈ [m] :
x ∈ Fi }. Then M(F1 , . . . , Fm , E; k; ρ) is equal to the gammoid M(Γ , E, T ), where
Γ = (V, D) is the directed graph with
40
Ragnar Freij-Hollanti, Camilla Hollanti, and Thomas Westerbäck
(i) V = E ∪ H ∪ T where E, H, T are pairwise disjoint,
(ii) T = [k],
(iii) H equals the union of the pairwise disjoint sets
H1 , . . . , Hm , H≥2 , where
|Hi | = ρ(Fi ) − |{x ∈ Fi : |s(x)| ≥ 2}| for i ∈ [m],
H≥2 = {hy : y ∈ E, |s(y)| ≥ 2},
(iv) D = DS
1 ∪ D2 ∪ D3 , where
D1 = i∈[m] {(−
x,→y) : x ∈ E, s(x) = {i}, y ∈ Hi },
−−→
D2 = {(x, hy ) : x ∈ E, hy ∈ H≥2 , s(x) ⊆ s(y)},
D3 = {(−
x,→y) : x ∈ H, y ∈ T }.
| 7cs.IT
|
Exponentially Stabilizing Continuous-Time
Controllers for multi-domain hybrid systems with
application to 3D bipdeal walking
Chunbiao Gan, Haihui Yuan, Shixi Yang and Yimin Ge
Abstract—This paper presents a systematic approach to exponentially stabilize the
periodic orbits of multi-domain hybrid systems arising from 3D bipedal walking. Firstly,
the method of Poincaré sections is extended to the hybrid systems with multiple domains.
Then, based on the properties of the Poincaré maps, a continuous piecewise feedback
control strategy is presented, and three methods are furthermore given to design the
controller parameters based on the developed theorems. By those design methods, the
controller parameters in each continuous phase can be designed independently, which
allows the strategy to be applied to hybrid systems with multiple domains. Finally, the
proposed strategy is illustrated by a simulation example. To show that the proposed
strategy is not limited to bipedal robots with left-right symmetry property which is
assumed in some previous works, an underactuated 3D bipedal robot with asymmetric
walking gait is considered.
Index Terms—3D Bipedal Walking, Nonlinear Control, Hybrid Systems, Multiple
Domains, Orbital Stability
I. Introduction
This paper addresses the problem of exponentially stabilizing the periodic orbits
of multi-domain hybrid systems arising from 3D bipedal walking [1-5]. Although
there is a considerable body of work on the exponentially stable walking control for
planar bipeds [6-10], the exponentially stable walking of the 3D bipeds is still far
from being well solved. One of the key features of the problem is that, compared with
the planar biped, the 3D bipedal walking system is not only hybrid [6, 9-11], but of
multiple domains (or continuous phases) [1, 12].
For a planar biped, a common “trick” in the field is to develop a model of the
biped with one of the legs in contact with the ground, and then to “swap” or relabel
the joint angles after the swing right leg impact with the ground in order to propagate
the dynamics for the ensuing step [1]. In this way, the number of the domains can be
reduced by half. Thus, despite the fact that the dynamical model of a planar biped has
two continuous phases, the planar walking is often modeled with only a single domain.
However, when it comes to the dynamical model of a 3D biped, one must “flip” the
sign on the hip width and “flip” the sign convention for all joint angles that are not in
the sagittal plane [1]. As a result, the dynamical models in the left and right support
phase are not absolutely symmetric. In other words, the number of the domains in the
dynamical model can’t be reduced. Therefore, the 3D bipeds are naturally modeled as
multi-domain hybrid systems. However, to the best of our knowledge, few results
have been focused on exponentially stabilizing the periodic orbits of multi-domain
hybrid systems.
In the early work [8], the optimal control method is successfully applied to
achieve the exponetial stability of a planar running gait with two continuous phases.
However, this method requires a high computational cost since the optimization
procedure designs the gait and the controller simultaneously [8, 13]. During the
optimization procedure, all the feedback control parameters within each continuous
phase as well as the virtual constraint parameters, i.e., the gait design parameters,
have to be taken as optimization variables. In [8], the running gait of a five-link planar
biped with two continuous phases has 23 independent optimization variables.
Compared with the planar bipeds, the 3D ones have more degrees of freedom and
more complicated dynamical equations. Considering that the amount of computational
cost is exponentially proportional to the number of optimization variables, it would
result in a huge computational cost. Therefore, technical limits exist in extending this
optimal control method to the 3D bipedal walking with multiple continuous phases.
In the later works [5, 14, 15], the well-known event-based feedback control
method is successfully applied to achieve the exponetial stability of a 3D bipedal
walking gait with two continuous phases. In these studies, two partial maps (or
one-step maps [15]) are firstly constructed corresponding to the two continuous
phases. Then based on the left-right symmetry property of the walking gait, one of the
two partial maps is transformed into a linear relationship while the event-based
feedback control is imposed on the other partial map. In this way, the two-domain
system can be dealt with just like a single-domain system. In [16], the event-based
feedback control method is further developed into a time-invariant hybrid control
strategy on the basis of symmetry. However, this strategy is performed mainly for
symmetric walking. Moreover, the work does not give a clue on how to extend the
strategy into the hybrid systems with multiple domains.
Recently, A BMI optimization method is applied to exponentially stabilizing the
periodic orbits of hybrid systems [17]. Firstly, a parameterized family of continuous
time controller is designed for the hybrid system. By assuming that the periodic orbit
is invariant under the choice of controller parameters, the problem of exponential
stabilization was translated into a set of Bilinear Matrix Inequalities (BMIs). Then, a
BMI optimization problem is set up to tune the parameters of the continuous-time
controllers so that the Jacobian of the Poincaré map has its eigenvalues in the unit
circle. By introducing continuous-time feedback laws in each continuous phase of a
hybrid system, this method can be applicable even for a hybrid system with multiple
domains. Unfortunately, during this method, the controller parameters in each
continuous phase need to be designed simultaneously, which would become
challenging when the number of continuous phases significantly increases. In addtion,
the assumption of invariant periodic orbits for the closed-loop hybrid system would be
a potential limitation of this method [17].
In this work, to address the problem of exponentially stabilizing the periodic
orbits of multi-domain hybrid systems arising from 3D bipedal walking, we develop a
continuous piecewise feedback control strategy. In section II, the method of Poincaré
sections is extended to the hybrid systems with multiple domains. By exploiting the
properties of the Poincaré maps, a continuous piecewise feedback control strategy is
presented in secion III, and three methods are furthermore given to design the
controller parameters based on the design theorems developed in this paper. By those
design methods, the controller parameters in each continuous phase can be designed
independently. In this way, it takes almost no additional computational cost even
when the number of continuous phases significantly increases. In section IV, the
proposed strategy is illustrated by a simulation example. To show that the proposed
strategy is not limited to bipedal robots with left-right symmetry property which is
assumed in some previous works [5, 16, 17], an underactuated 3D bipedal robot with
asymmetric walking gait is considered. In the final section, the results of this work are
summarized and discussed.
II. Technical background
As stated above, the 3D bipeds are naturally modeled as multi-domain hybrid
systems. In addtion, it often occurs that there are three or more continuous phases in
bipedal walking gaits, for example, bipedal walking with several continuous
subphases [18] and bipedal running with nontrival feet. In those cases, bipedal robots
would surely be modeled as multi-domain hybrid systems. In this section, we will
review the hybrid models for multi-domain hybrid systems as well as the primary
analysis tool that is used in this paper: the Poincare’s method.
A. Multi-domain hybrid models
Hybrid systems with a single-domain in bipedal locomotion have been studied
extensively [9-11]. Here, we are going to address systems with N 2 continuous
domains and discrete transitions between the domains. Without loss of generality,
suppose that phases are executed in a fixed order 1 2 … N 1 , as shown in Fig.
1. Considering the cyclic behavior in bipedal locomotion, we will use the
notation N 1 1 . For compactness of notation, the multi-domain model will be
denoted as a 5-tuple ( X , U , FG, S , )
X X i N : X i R mi
i 1
N
U U ii 1 : U i R n i
N
FG f i, g ii1 : x i f i( xi) g i( xi)u i ,
i 1
i 1 N
i 1
S S i i 1 : x i | H i 0, H i 0
i 1 N
i 1
i i 1 : x i 1 i ( x i )
(1)
where X i is an open connected subset of R mi , upon which the autonomous differential
x n f n ( x n) g n ( x n )u n
S
( S 1n )
n
n 1
x1 f 1( x1) g 1( x1)u1
(S
n
n 1
)
1
Sn
2
S1
( S 12)
( S 32)
S
3
2
x 2 f 2 ( x 2 ) g 2( x 2)u 2
Fig. 1. The multi-domain hybrid model in periodic walking motion. S ii 1 (i 1,..., n 1) are the
switching hyper-surfaces, S ii 1 (i 1,..., n 1) denote the hyper-surfaces after each impact event or the
initial hyper-surfaces of the next domain after a nonsmooth process.
equation x i f i ( x i ) g i ( x i )u i of phase-i is defined, u i U i R ni is the vector of
actuator torques, S ii 1 is an embedded submanifold of co-dimension one in the state
space X i that determines when a transition from X i to X i 1 takes place according to
the reset map ii 1 , x i is final state of phase-i , x i1 is the initial state of phase-i+1.
Under assumptions analogous to those for the single-domain model [9-11], a
unique, maximal solution of the multi-domain model can be constructed by piecing
together trajectories of the flows FG i in such a way that a transition occurs when a
flow intersects a switching hyper-surface, S ii 1 , and at each transition, the new initial
condition is determined by the reset maps ii 1 . To avoid chattering, it is assumed that
a solution through a domain must have a non-zero duration [13].
B. Periodic Orbits and the Poincare’s Method
Let X X 1 X 2... X N , A solution (t ) of (1) is periodic if there exists a
finite T such that (t T ) (t ) for all t [0, ) . A set O X is a periodic orbit of
(1) if O { (t ) | t t 0} for for some periodic solution (t ) .A periodic orbit O is
transversal to S ii 1 if its closure intersects
i 1
Si
in exactly one point, and
for x *i O S ii 1 , H ii 1( x *i ) 0 . A periodic orbit O is transversal if it is transversal
to S ii 1 for all i. In the case of a bipedal robot, a nontrivial, transversal, periodic orbit
will also be referred to as periodic locomotion [13].
The definitions of orbital stability in the sense of Lyapunov, orbital asymptotic
stability, and orbital exponential stability are analogous to those systems with
single-domain. The Poincaré return map remains the mathematical tool for
determining the existence and stability properties of periodic orbits. Define the
phase-i time-to-impact function, T I ,i : x i R ,by
inf{t 0 | i (t , x 0) S ii 1}
T I ,i ( x 0) :
where
i (t , x 0 )
is
an
integral
curve
if t such that i(t , x 0) S ii 1
otherwise,
of
the
differential
,
(2)
equation
x i f i ( x i ) g i ( x i )u i defined in (1) corresponding to i (0, x 0) x 0 . From [13],
T I ,i ( x 0) is continuous at points x 0 where 0 T I ,i ( x 0 ) and the intersection with
i 1
i 1
S i is transversal. Hence, X i : {x i | 0 T I ,i ( x 0) and H i ( i (T I ,i ( x i ), x i )) 0 } is
1
i
open, and consequently, S i 1 : ii 1 ( X i ) is an open subset of S ii 1 . Under the
hypotheses in [13], the generalized Poincaré phase-i map P i : S ii 1 S ii 1 can be
defined as a partial map
S
S
S
i
i
P i ( x i 1) : i (T I ,i ( i 1( x i 1)), i 1( x i 1)) ,
(3)
where x iS1 is a state defined in S ii 1 , and P i is C 1 in a neighborhood of x iS1 . It is noted
that, P i : S ii 1 S ii 1 is often expressed as P i : S ii 1 S ii 1 for simplicity. The Poincaré
return map can be defined as the composition of the generalized Poincaré phase-i
maps, starting at any point in the cycle 1 2 … N 1 . Without lose of generality,
we start it at i = 1, so that
P P N P N 1... P1 .
(4)
Theorem 1(Connecting Multi-domain Models to Single-domain Models [1, 13]):
Let P be the Poincaré return map defined in (4) for the multi-domain model in (1).
P is also the Poincaré return map for the single-domain system
x1 f 1( x1) g 1( x1)u1 x1 S 12
,
x ( x )
1
x1 S 2
1
1
(5)
where : 1N P N ... P 2
This theorem is very important because it shows that results developed for
single-domain models can be applied to models with multiple domains. In particular,
suppose that the multi-domain hybrid model (1) is C 1 in each phase and has a
transversal periodic orbit O . Then, results in [13] show that : 1N P N ... P 2 is C 1 in
a neighborhood of x * O S , and thus P is C 1 in a neighborhood of x * .
Exponential stability the hybrid system can be checked by evaluating eigenvalues of
the Jacobian of P at x * , and the system is exponential stable if the Jacobian has
spectral radius smaller than 1. Moreover, according to the result in [11], the
exponential stability can also be by evaluated in a reduced-dimensional state space.
Theorem 2: Suppose that the multi-domain hybrid model (1) is C 1 in each phase and
has a transversal periodic orbit O . Let A
P *
( x1) be the Jacobian matrix of the
x1S
Poincaré return map P . Let Ai PSi ( x *i ) be the Jacobian matrix of P i . Then
xi
A A N A N 1... A1 .
Proof. From Theorem 1, we can know that the Poincaré return map P is C 1 in a
neighborhood of x * O S . Then, according to the chain rule for computing
the derivative of the composition of two or more functions, the Jacobian A for P can
be expressed as
P * ( P N ... P 2 P1) * ( P N ... P 2)
( )
( x1 )
( P1( x1*)). PS1 ( x1*)
S x1
S
S
x1
x1
( P1( x1 ))
x1
( ... )
( ... 3)
P N S P 2 ( x*2). PS1 ( x1*) P N P
( P 2( x*2)). PS2 ( x*2). PS1 ( x1*)
S
x2
x1
( P 2( x 2 ))
x2
x1
, (6)
PSN ( x*N )... PS2 ( x*2). PS1 ( x1*)
xN
x2
x1
A N A N 1... A1
which completes the proof.
III. Exponentially stabilizing the periodic orbits of multi-domain hybrid systems
As stated before, the BMI optimization method can be applied to exponentially
stabilizing the periodic orbits of multi-domain hybrid systems. However, during this
method, the controller parameters in each continuous phase need to be designed
simultaneously, which would result in huge computational cost when the number of
continuous phases significantly increases. In addtion, the assumption of invariant
periodic orbits for the closed-loop hybrid system would be a potential limitation of
this method. To address these two problems, in this section, a continuous piecewise
feedback control strategy is proposed, and two design theorems are furthermore given
to design the controller parameters in each phase independently.
A.The continuous piecewise feedback control strategy
For hybrid systems with periodic orbits, the design of the controller is equivalent
to the design of the Jacobian of the Poincaré return map. From theorem 2, one can
clearly see that the Jacobian of the Poincaré return map can be expressed as the
successive multiplies of the Jacobians of partial maps. This leads us to think that the
design of the Poincaré return map’s Jacobian can be realized by imposing continuous
feedback in each continuous phase to design the Jacobian of each partial map.
Consider the multi-domain hybrid system with periodic orbit defined in (1). For
each phase-i, suppose that the continuous-time controller can be expressed as the
following parameterized control law
u i , i ( x i, i ) ,
(7)
Where i i and i R p represent the finite-dimensional parameter vector in an
open neighborhood of 0 and set of admissible parameters, respectively, for some
positive integer p. Moreover i ( x i , i ) : X i i U i is a C n p map, where n p is some
positive integer larger than 1. In contrast with the BMI method in [17], where the
continuous-time controller i ( x i, i ) is designed to meet i ( x i, i ) i 0 , so that
the periodic orbit will be invariant under the choice of the control parameters i , in
this paper, we consider i ( x i, i ) i 0 . In particular, i ( x i, i ) will be designed
such that
C1) lim i ( x i, i ) u i ( x i )
i 0
C2) lim i ( x i, i ) x i u i ( x i ) x i .
i 0
Conditions C1 and C2 are required so that the periodic orbit of the hybrid system
will not be significantly changed by the parameter vector i which is designed to be
small enough, see (42) and (43). In order to have a more specific understanding about
the continuous-time controller i ( x i, i ) , the following lemma is presented.
Lemma 1. Suppose that u i ( x i ) defined in (1) can be expressed by a general
form u i ( x i ) 1( 2( x i )) 3( x i ) , where 1 , 2 , and 3 are all C 1 functions with
respect to x i . If there exists a continuous function ( x i , i ) which is C n p functions
with respect to ( x i , i ) such that lim ( x i, i ) 0 and lim ( x i, i ) x i 0 . Then
0
0
i
i
both 1( x i ) 1( 2( x i ) ( x i , i )) 3( x i ) and 2( x i ) 1( 2( x i )) 3( x i ) ( x i, i )
meet the conditions C1 and C2.
Proof. The proof can be straightly obtained.
Since the designed controller i ( x i, i ) is C n p with respect to i , integrating the
differential equation x i f i ( x i ) g i ( x i ) i ( x i , i ) , we can easily know that the integral
curve of the differential equation, i , (t , x 0, i ) , is also C n p with respect to i .
Therefore,
the
generalized
Poincaré
phase-i
map P i , ( x iS1, i ) : i (T I ,i ( ii 1( x iS1)), ii 1( x iS1), i ) is C n p with respect to i . Recall
that P i , is continuously differentiable with respect to x iS1 . Therefore, the total
derivative of P i , can be expressed as
d P i ,
P i , S P i,
d x i 1
di .
x iS1
i
(8)
In the next step, if we design a continuous feedback law such that
i K i ( x iS1 x *i 1) ,
(9)
where K i R pn is the feedback parameter matrix. Replacing (9) in (8) yields
Ai Ai , F i , K i ,
where
S
Ai d P i , d x i 1
is
the
designed
(10)
Jacobian
of
partial
map P i , , Ai , P i , x iS1 is the partial derivative of P i , with respect to x i ,
and F i , P i , i is the partial derivative of P i , with respect to i .
Suppose that we have designed the Jacobians Aid for all i independently by
following the above procedures, which will be done in the next subsection. Then,
according to theorem 2, the Jacobian of the Poincare return map can be designed as
A AN AN 1... A1 .
(11)
Let ( AN AN 1... A1 ) be the spectral radius of AN AN 1... A1 . From theorem 1, the
closed-loop hybrid system is exponentially stable, if ( AN AN 1 ... A1 ) 1 , which is the
ultimate goal of this work.
B.The picewise feedback controller
As stated before, in some previous works, the controller parameters in each
continuous phase need to be designed simultaneously, which would become
challenging when the number of continuous phases significantly increases. In this
subsection, we present two design theorems so that we can design the feedback
parameters in each contiunous phase independently.
Taking the designed controlle i ( x i, i ) and the feedback law (9) in the above
subsection into account, the multi-domain hybrid model defined in (1) will be updated
as 6-tuple cl ( X ,U , , FG, S , )
X X i N : X i R mi
i 1
N
U U ii 1 : U i R ni
N
ii 1 : i K i ( x iS1 x*i 1)
,
N
cl
FG f i, g ii 1 : x i f i ( x i ) g i ( x i ) i ( x i, i )
i 1
i 1 N
i 1
S S i i 1 : x i | H i 0, H i 0
i 1 N
i 1
i i 1 : x i 1 i ( x i )
(12)
N
where ii1 is piecewise defined to impose feedbacks, for convenience, we call it
the piecewise feedback controller.
According to Appendix A, when we design the Jacobian of each partial map, (10)
can be expressed in a simple form
d
Ai Ai F i K i ,
(13)
which is only dependent on the feedback parameter matrix K i . Therefore, according
N
to theorem 2, if we have designed the piecewise feedback controller ii1 , the
Jacobian of the Poincare return map would be designed as
d d
d
d
A AN AN 1... A1 .
(14)
Let ( ANd ANd 1 ... A1d ) be the spectral radius of ANd ANd 1... A1d . Then our goal is to design the
N
piecewise feedback controller ii1 so that ( ANd ANd 1 ... A1d ) 1 .
To satisfy the condition ( ANd ANd 1... A1d ) 1 , the feedback parameters in the
N
piecewise feedback controller ii1 can be designed simultaneously, as in [17].
However, it would result in huge computational cost when the number of continuous
phases significantly increases. Hence, we hope to find ways to design the feedback
parameters in each continuous phase independently. For this goal, the following
design theorems are presented.
Theorem 3: For a N-domain hybrid closed-loop system cl ( X ,U , , FG, S , ) ,
N
if there exists a piecewise feedback controller ii1 , under which the Jacobian of
each phase-i , Aid R nn ,is designed such that
H1) Aid R nn is symmetric;
H2) The spectral radius of Aid meets ( Aid ) 1 .
Then the closed-loop system is exponentially stable.
Proof. See Appendix A.
Remark 1: The Item 1 is very important. If the symmetric condition is not met, we
will have no idea whether the spectral radius of ANd ... A2d A1d is acceptable, even when we
have designed a set of Jacobians that all look perfect. Here, we will show you an
interesting example. For a two-domain hybrid system, suppose that we have designed
0
0.7 0.65 d
0.7
two Jacobians for the two partial maps such as, A1d
, A2
0.5
0
0.65 0.5
Both the spectral radiuses of them are 0.75, which is quite acceptable. However, the
spectral radius of A2d A1d is 1.0453 , which indicates an unstable closed-loop system.
Theorem 4: For a N-domain hybrid closed-loop system cl ( X ,U , , FG, S , ) ,
N
if there exists a piecewise feedback controller ii1 , under which the Jacobian of
each phase-i, Aid R nn , satisfies mi max Aid ( j , k ) 1 , then the closed-loop
n
1 j , k n
system is exponentially stable.
Proof. See Appendix B.
From the above two theorems, we can know that the feedback parameter
matrix K i in each continuous pahse can be designed independently to achieve the
exponential stability of the hybrid system, as long as the designed Jacobian of each
partial map satisfy some conditions. In the following subsection, three methods are
presented to design the feedback parameter matrix K i based on the above theorems.
B.Methods for Designing the Piecewise feedback controller
The implement of the continuous piecewise feedback control strategy requires
N
the design of the piewise feedback controller ii1 . And the essetial part is to obtain
the feedback parameter matrix K i in each pahse independently. In the following, three
different methods are given to obtain the feedback parameters.
1) Symmetric matrix method. The method is based on Theorem 3. To achieve
the exponential stability of a N-domain hybrid system, a constant symmetric
matrix M sym with spectral radius smaller than 1 is firstly picked. Then, according to
Theorem 3, the hybrid system would be exponentilly stable, if the Jacobian of each
phase-i is designed to be M sym , i.e.,
Aid M sym , 1 i N .
(15)
M sym Ai F i K i .
(16)
According to (13), we obtain
In general, F i is not a square matrix. Hence, the feedback parameter matrix K i
from the above equation can be calculated as
K i pinv( F i )( Ai M sym) ,
(17)
where pinv ( F i ) denotes the Moore-Penrose pseudoinverse of F i .
During the symmetric matrix method, the feedback parameters can be
independently and directly obtained with no optimization process, which allows the
method to be used for real-time application.
2) Scale factor method. The Scale factor method is based on Theorem 4.
Let mi max Ai ( j, k ) denotes the largest element of Ai , and let c i 1 ( n.m i ) be the
1 j , k n
corresponding constant scale factor. Then, from Theorem 4, it’s obvious that the
N-domain hybrid system is exponentially stable, if the Jacobian of each phase-i, Aid , is
designed to be
Aid c i Ai , 1 i N .
(18)
Then, combining (13) and (18), we obtain
c i Ai Ai F i K i .
(19)
Similar to (17), the feedback matrix K i can be calculated as
K i pinv( F i )(1 c i ) Ai .
(20)
In the Scale factor method, different Jacobians are designed with different
coefficients, i.e. c i .Therefore, compared with the symmetric matrix method, the scale
factor method accounts for the differences between the Jacobians of partial maps.
3) DLQR-based method. The conventional DLQR method has been extensively
used to design controllers. Here, it wil be extended to the multi-domain hybrid system.
Starting by defining Y i x iS1 x *i 1 , the linearization of equation (13) can be expressed
as
Y i 1 AiY i F i i .
(21)
During the DLQR method, the feedback gain matrix K i is calculated via a
discrete linear quadratic regulator (DLQR) so that the feedback control
law i K iY i minimizes the cost function k (Y i Q iY i i R i i ) subject to the
state dynamics (21), where Q i is a real positive semi-definite weighting matrix
associated with the robustness of the system, and R i
is a real positive
definite weighting matrix associated with the input energy, i.e., i . The two weighting
matrices are often decided using trial and error method. For simplicity, they are often
designed to be a unit matrix multiplied by a positive constant number. Now that the
gain matrix K i has been obtained, the Jacobian of each phase-i, Aid , would be
designed as
Aid Ai F i K i , 1 i N .
(22)
By the conventional DLQR method, the spectral radius of Jacobian Aid would
be smaller than 1. However, since the Jacobians Aid designed by the DLQR method
are generally not symmetric, we can then know from Remark 1 that the inequality
condition ( ANd ... A2d A1d ) 1 may not be met, even though each Jacobian Aid has
spectral radius smaller than 1.
From Theorem 4, we know that the spectral radius of ( ANd ... A2d A1d ) would be
smaller than 1 if the largest element of Aid is smaller than 1 . Fortunately, this can be
n
done by tuning the weighting matrix Q i , which is associated with the robustness of the
system. Increasing the Q i matrix leads to a smaller Aid . Therefore, we can choose a
large enough Q i so that the largest element of Aid is smaller than 1 .
n
Compared with the symmetric matrix method as well as the scale factor method,
the DLQR-based method provides us with optimal feedback gain matrices in terms of
the trade off between the robusness and the input energy. Hence this method may be
more energy efficient. But on the other hand, the DLQR-based method would take
manual efforts in picking appropiate weighting matrices Q i .
IV. Application to underactuated 3D bipedal locomotion
The objective of this section is to illustrate the continuous piecewise feedback
control strategy to efficiently achieve the exponential stability of periodic 3D bipedal
walking with multiple continuous phases. To show the result in a simple and clear
q7
q4
q6
x4
z5
z4
x5
z7
z6
q5
x4
x5
x7 x6
z4
z5
q5
q6
q4
z7
q7
z6
q8
z3
q3
x3
x7 x6
z8
x8
z3
z 1sw
q3
x3
q 2sw
x 0sw
z 0sw
q 0sw
q8
zI
x 2sw x1sw
I
z 2sw
q1sw
xI
x 2 x1
q2
xR
x0
x0 z 0 z R
( x st , y st , z st )
q1
q0
x8
z2
xR
yI
( x st , y st , z st )
z2
z8
z1
q2
zR z0
z1
zI
x 2 x1
q0
q1
I
x 2sw x1sw
z 0sw
xI
yI
z 1sw
q 2sw
x 0sw
z 2 sw
q1sw
q 0sw
Fig. 2. The model and coordinate systems of the 3D five-link biped with point feet. In (a) and (b), the
robot is supporting on the left and right legs, respectively. The structural parameters of the biped are
chosen to be the same as those in [5]. The coordinate systems are established by the DH method. The
virtual feet represented by the dashed box are used to show the yaw orientation of the biped.
way, a two-phase hybrid system is studied in this section. In addtion, to show that the
proposed strategy is not limited to bipedal robots with left-right symmetry property
which is assumed in [5, 16, 17], an underactuated 3D bipedal robot with asymmetric
walking gait is considered in this section.
A.Hybrid models
The 3D biped studied here is assumed to be underactuated with point feet,
consisting of five links: a torso and two legs with knees, as shown in Fig.2. For
simplicity, each link is modeled by a point mass at its center, and the stance leg end is
assumed to act as a passive pivot. There are 9 DOF in total for the biped in the single
support phase, i.e. ( q 0, q1,.., q 8) , including three underacuated DOF, i.e. ( q 0, q1, q 2) .
In particular, the yaw angle, q 0 ,is assumed to be inhibited by friction during the whole
single support phase. Therefore, the configuration coordinates of the robot in single
support can be denoted simplely by q ( q1,.., q 8) .
The periodic walking gait studied in this section consists of two continuous
phases, i.e., the two single support phases with support on the left and right legs,
respectively. Let leg 1 be the left leg, and let leg 2 be the right leg. Without loss of
generality, assume leg 1 is now the support leg. Based on the Lagrangian method, the
dynamic model for the robot in the single support phase can be obtained as
D1(q)q C 1(q, q )q G1(q) Bu1 ,
(22)
where D1(q ) is the positive-definite (8 8) mass-inertia matrix, C 1(q, q ) is the
(8 8) Coriolis matrix, G1(q) is the (8 1) gravity vector, B is an (8 6) full-rank,
constant matrix indicating whether a joint is actuated or not, and u 1 is the
(6 1) vector of the input torques.
Under assumptions analogous to those in [5], the impact model in the double
support phase can be obtained as
1
T
q e D1,e E 1,2
D1,e q e
,
F 2 E 1,2 0 4 4 0 4 4
where q e
(23)
and q e are the extended velocities before and after the impact
instantaneously, respectively, F 2 is the impulsive reaction force on leg 2 at the
contact
point,
D1,e
is
the
extended
mass–inertia
matrix,
and E 1,2 q e [ x1, sw, y1, sw, z 1, sw, q 0 sw] is the Jacobian for the position of the swing foot
and its orientation.
The support-swing leg swapping and relabeling take place immediately after the
impact of the swing foot. This leads to the exchange between the notations of the
angles, i.e., [ q 3, q 4, q 5, q 6, q 7, q 8] [ q 8, q 7, q 6, q 5, q 4, q 3] [5]. At the same time, the
angles ( q 0, q1, q 2) exchange the roles with (q 0 sw, q1sw, q 2 sw) . Combining the impact
model (23) and the coordinate relabeling, the dynamical model for the double support
can be written as
q 1,q(q )
,
(
,
)
q
q
q
1,
q
(24)
where (q , q ) is the final state of the current step, and (q , q ) is the initial state of the
next step. Hence the double support model acts as a reset map.
When leg 2 is the support leg, the dynamic models can be obtained in the same
way. Based on the above equations, the two-domain hybrid model for the biped robot
can be established as
X X i2 : X i R16
i 1
2
U U ii 1 : U i R 6
2
FG f i, g ii1 : x i f i( x i) g i( x i)u i ,
i 1 2
S S i i 1 : x i z i , sw( q) 0, x i , sw(q) 0
i 1 2
i 1
i i 1 : x i 1 i ( x i )
(25)
where
q
0
, g 1( x) 1
f i ( x ) 1
D i (C i q G i )
Di B
,
(
)
q
i
,
q
x i 1 i( x i )
(
,
)
q
q
i
,
q
(26)
B. Asymmetric walking gait design and stability analysis
In order to implement the control strategy of this paper on a robot, a periodic and
asymmetric walking gait was obtained as shown in Fig.3. Analogous to [5], the design
of the gait was cast as a constrained nonlinear optimization problem, and the virtual
constraints approach was also used here to design the output of the control for each
phase-i, i.e.,
y i h i ( q ) q a h i , d ( ) , i 1, 2 .
(27)
where y i denotes the output of control design, q a [q 3, q 4, q 5, q 6, q 7, q 8] is the vector
of the actuated coordinates, (q) is a quantity strictly monotonically increasing
along the walking cycle, and h i ,d ( ) is the desired evolution of the actuated variables
as a function of . Specifically, h1,d ( ) and h 2,d ( ) are designed as the Bezier
polynomials with 5 and 3 degrees, respectively. The whole optimization process was
performed in MATLAB with the Patternsearch function of the optimization toolbox.
0.5
1
times(s)
q7 (deg)
q6 (deg)
0
-20
0
0.5
1
times(s)
-50
-100
0
u(3) (Nm)
0
0
0.5
1
times(s)
50
0
0.5
1
times(s)
0
0.5
1
times(s)
-20
0
0.2
0
0
0.2
0
0
0.2
0.4 0.6
times(s)
-5
0
0.5
1
times(s)
5
0
-5
0
0.5
1
times(s)
10
0
-10
0
0.5
1
times(s)
10
0
-10
0
0.5
1
times(s)
0
0.2
0.4 0.6
times(s)
0
0.2
0.4 0.6
times(s)
0
0.2
0.4 0.6
times(s)
0
0.2
0.4 0.6
times(s)
10
0
-10
0
0.5
1
times(s)
5
0
-20
-40
0
-5
0
0.5
1
times(s)
10
0
-10
0
0.5
1
times(s)
(c)
20
Fst(2) (N)
-1
Fst(1) (N)
0.5
1
times(s)
0
0
0.2
150
0
-10
0.06
100
50
0
10
-20
0.4 0.6
times(s)
Zsw (m)
0
0
0.4 0.6
times(s)
0
-5
20
1
Fst(3) (N)
5
-2
dq2/dt (rad/s)
0.5
1
times(s)
0
dq5/dt (rad/s)
0
2
0.2
(b)
dq8/dt (rad/s)
0
dq1/dt (rad/s)
1
dq4/dt (rad/s)
2
dq7/dt (rad/s)
dq6/dt (rad/s)
dq3/dt (rad/s)
d/dt (rad/s)
(a)
0
5
0.4 0.6
times(s)
50
-50
0
-50
0.4 0.6
times(s)
10
-10
50
u(2) (Nm)
0
u(4) (Nm)
0.5
1
times(s)
20
-20
0.5
1
times(s)
u(1) (Nm)
q2 (deg)
0
0
0
20
u(6) (Nm)
0
20
-5
0.5
1
times(s)
-50
-100
0
u(5) (Nm)
20
0
0
q4 (deg)
q3 (deg)
40
0
-20
0.5
1
times(s)
5
q5 (deg)
0
0
q8 (deg)
0
-0.2
20
q1 (deg)
(deg)
0.2
0.04
0.02
0
0
0.2
0.4 0.6
times(s)
-0.02
(d)
Fig. 3. The periodic and asymmetric walking motion of the 3D biped over two steps, where (a) is the
joint angular profiles of the periodic motion, (b) is the required torque profiles of the periodic motion,
(c) is the joint angular velocity profiles of the periodic motion, and (d) is the reaction force on the
stance foot and the evolution of the swing leg tip.
To save the computational cost, the numerical integration of the single support phase
dynamics was carried out on a lower-dimensional system, which is called the swing
phase zero dynamics, see [5]. It is noted that, during the optimization, the walking gait
is feasible only when the reaction forces F st on the support leg meets the no-take-off
and friction constraints. The calculation of F st is given in Appendix B.
For a single-domain hybrid system, the exponential stability of the full hybrid
model can be evaluated by the corresponding reduced hybrid zero dynamics.
According to Theorem 1, we can also apply this method to the two-domain hybrid
model in this section. Firstly, a zero dynamics of the full hybrid model has to be
created. To achieve that, a correction term hi,c was introduced to the output in each
phase-i.
y i h i ( q ) q a h i , d ( ) h i ,c , i 1, 2 .
(28)
where hi,c was designed so that the initial errors of the outputs and their derivatives,
which are caused by the impact event, can be smoothly joined to the original virtual
constraint at the middle of the step, see [5].
After the HZD for the two-domain hybrid model had been created, two restricted
partial maps were then constructed, i.e., P 1 : S 12 Z S 12 Z , P 2 : S 12 Z S 12 Z
where Z {( q , q ) | y i ( q ) 0, y i ( q ) 0} . Following the technique in [5], the Jacobians
for P1 and P 2 can be calculated as
2.0824 0.3380 0.2596
A1 2.2540 1.2701 4.5384
7.1221 2.8911 1.1735
(29)
2.1280 0.2878 0.3763
A2 2.2854 1.3484 4.7298
7.3238 2.6283 0.5894
(30)
respectively. According to Theorem 2, the Jacobian of the Poincaré return map P can
be obtained as
7.7603 2.1727 1.1953
A 25.8883 11.1895 0.0243
16.9777 4.1095 9.3347
(31)
with the three eigenvalues:
1 2.8271, 2,3 7.7955 2.2193i
2,3 8.1053
(32)
Since ( A) 8.1053 1 , the designed asymmetric gait is unstable. Therefore, we
need to design a feedback controller to stabilize the hybrid system.
C. Achieving exponential stability by the continuous piecewise feedback control
strategy
The objective of this subsection is to stabilize the unstable asymmetric walking
gait by the continuous piecewise feedback control strategy. Firstly, a continuous time
controller i ( x i, i ) is designed for each continuous phase. Secondly, the scale factor
method is applied to design the feedback parameters to stabilize the hybrid system.
In the first step, a continuous time controller i ( x i, i ) is designed for each
continuous phase based on the result of Lemma 1. From the result in [13], we know
that the nominal input torque u i is determined by the output y i , and thus u i can be
expressed as u i ( y i ) . Therefore, from Lemma 1, i ( x i, i ) can be designed as
i ( x i, i ) u i ( y i ( x i, i )) , i 1, 2 ,
(33)
where ( x i, i ) satifies the conditions lim ( x i, i ) 0 , and lim ( x i, i ) x i 0 .
0
0
i
i
Specifically, ( x i, i ) is designed in the form of ( , i ) . It follows that the ouput
(28) will be updated as
y i h i ( q ) q a h i ,d ( ) h i ,c ( , i ) , i 1, 2 .
(34)
It is noted that one of the other important requirements for the design of
( , i ) is to preserve the HZD that have been created in the above section.
Here, ( , i ) is designed as a fifth-order polynomial.
In the second step, the scale factor method is applied to design the feedback
parameters to stabilize the hybrid system. Firstly, Similar to the calculation of A1 and
A2 , the Jacobians of P1 and P 2 with respect to 1 and 2 are obtained as
-0.0326 -0.0369 -0.1358 0.0604 0.0396 0.0229
F 1 0.0890 0.1037 -0.0422 0.0306 0.4580 0.1079 ,
0.1596 0.2808 0.3050 0.6396 0.3171 0.1466
(35)
0.0483 0.0397 0.1208 0.0589 0.0445 0.0259
F 2 0.0868 0.0880 0.0529 0.0243 0.4689 0.1057 ,
0.2465 0.2798 0.2130 0.6212 0.3389 0.1624
(36)
Let m1 max A1( j , k ) , m 2 max A2( j , k ) . Then, the scale factors for the two partial
1 j , k n
1 j , k n
maps can be designed as c1 1 ( n.m1) and c 2 1 ( n.m 2) . From (20), we have
-1.4696
-3.8313
-12.1720
K1
4.0627
-2.8061
0.3861
3.1276
4.6644
-12.4826
K2
4.9906
2.8635
-0.7739
-0.0063 -1.5780
2.2786 3.9165
0.5296 -2.5886
,
-3.9668 -5.5839
1.7885 7.7203
0.1001 1.6090
0.3810 1.3404
2.2536 -3.5804
-0.2466 -3.0383
,
3.9753 -5.8621
2.1413 -8.3819
0.0753 -1.6155
(37)
(38)
And the corresponding Jacobians are designed as
0.0975 0.0158 0.0122
A 0.1055 0.0594 0.2124 ,
0.3333 0.1353 0.0549
(39)
0.0131 0.0171
0.0969
A 0.1040 0.0614 0.2153 ,
0.3333 0.1196 0.0268
(40)
d
1
d
2
Finally, according to (14), the Jacobian of the Poincaré return map P are designed as
0.0166 0.0046 0.0025
0 ,
A 0.0551 0.0238
0.0363 0.0088 0.0199
d
(41)
which has a spectral radius of 0.0173. Hence, the closed-loop system is exponential
stable.
To illustrate the effect of the continuous piecewise feedback control strategy, the
HZD model of the 3D biped in a closed-loop is simulated with the initial state
perturbed away from the fixed point x 1* , and the initial errors of 0.01, 0.01 and 0.01
are introduced to the three components of x 1* , respectively. Fig. 4 shows the phase
plots of the uncontrolled variables and q 2 , from which the asymmetric walking
gait is apparently stabilized by the continuous piecewise feedback control strategy.
1.8
0.8
1.6
0.6
0.4
dq2/dt (rad/s)
d/dt (rad/s)
1.4
1.2
1
0.8
0
-0.2
-0.4
0.6
0.4
-0.2
0.2
-0.6
0
(rad)
0.2
-0.8
-0.05
0
q2 (rad)
0.05
Fig. 4. Phase plots for and q 2 , where the straight lines correspond to two impact phases in one
walking cycle.
V. Conclusion
This paper addresses the problem of exponentially stabilizing the periodic orbits
of multi-domain hybrid systems arising from 3D bipedal walking. By exploiting the
properties of the Poincaré maps regarding the multi-domain hybrid systems, a
continuous piecewise feedback control strategy is presented, and three methods are
furthermore given to design controller parameters based on the developed theorems.
In contrast with previous methods that require the controller parameters in each
continuous phase to be designed simultaneously, the proposed method is able to
design the controller parameters in each continuous phase independently, which
allows the strategy to be applied to hybrid systems with multiple domains. The
validity of this approach was illustrated in an underactuated 3D bipedal robot with
asymmetric walking gait.
The strategy presented in this paper can be extended to more general forms of 3D
bipedal locomotions, including running and human-like walking. In future research,
we will investigate the robustness of this strategy against all sources of uncertainties
in 3D bipedal walking as we have done in the previous works [19, 20].
Appendix
A. Approximation of Aid Ai , F i , K i
Let f op f i ( x i ) g i ( x i )u i ( x i ) , and let f cl f i ( x i ) g i ( x i ) i ( x i, i ) . Then, under
the conditions C1 and C2, we can easily prove that
lim f op f cl ,
(42)
i 0
And
lim f cl ( x i, i ) x i f op ( x i, i ) x i .
(43)
i 0
From (42) and (43), we can know that, when the control parameter vector i is
close to zero, the nominal trajectory generated by x i f cl equals that of x i f op . In
other words, the periodic orbit is invariant when i is close to zero. Combining the
result in [17], where the mathematical expression of P i , x iS1 is given, we will
have
lim P i , x iS1 P i x iS1 .
(44)
i 0
Therefore, for a small enough vector i , we think that P i x iS1 is approximately
equal to P i , x iS1 , i.e., Ai , Ai ,where Ai P i x iS ( x *i ) .
When
it
comes
to
the
estimation
of F i , ,
it
is
much
easier.
Since P i , ( x iS1, i ) : i (T I ,i ( ii 1( x iS1)), ii 1( x iS1), i ) is C n p with respect to i , we can
then know that F i , P i , i is continuous. Hence, we have
lim F i , ( i ) F i , (0) .
i 0
(45)
For convenience, F i , (0) will be denoted as F i . Replacing (44) and (45) in (10) , we
have
d
Ai Ai F i K i ,
(46)
where Aid is independent of i , Ai and F i can be calculated by following the techniqie
in [5].
B. Proof of Theorem 3
In the first step of the proof, we will prove that, for any symmetric
matrix A R nn , its spectral norm is equivelent to its spectral radius, i.e. A 2 ( A) .
From the definition of spectral norm, we have
A 2 max| i (( A) H A) | ( AH A) .
(47)
i
Since A is symmetric, the above equation can be rewritten as
A 2 ( AH A) ( A2) .
(48)
From the matrix theory, we can know that if 1 , 2 ,…, s are the eigenvalues of A ,
then ( 1) k , ( 2) k ,…, ( s ) k would be the eigenvalues of Ak . Hence, we have
k
k
k
( Ak ) max| i | max| i | ( A) .
i
i
(49)
Combining (48) and (49), we have
A 2 ( A2) 2( A) ( A) .
(50)
The next step is to prove that, for two designed symmetric Jabocians, such as
Adj and Akd ,where 1 j , k n , if ( Adj ) 1 as well as ( Akd ) 1 , then ( Adj Akd ) 1 .
According to Theorem 5.6.9 in [21], we have
( Adj Akd ) Adj Akd
2
Adj
2
Akd .
(51)
2
d
Since Aj and Akd are symmetrix matrixes, from (50), we have ( Adj ) Adj
2
,
( Akd ) Akd 2 . Hence, (51) can be rewritten as
( Adj Akd ) Adj
2
Akd
2
( Adj ) ( Akd ) 1 ,
(52)
Based on (52), Theorem 3 can be proved in a further step.
C. Proof of Theorem 4
Firstly,
we will
prove that,
Adj and Akd ,where 1 j , k n ,
for two
if
designed
Jabocians, such as
d
m j max Aj (i, k ) 1 n
1 i , k n
as
well
as
d d
d
m k max Ak (i, j ) 1 n , then ( Aj Ak ) 1 .
1 i , j n
From matrix theory, for any A R nn , define a norm A max a ij , and
1 i , j n
then N ( A) n A is a matrix norm. According to Theorem 5.6.9 in [21], we can then
have
( Adj Akd ) N ( Adj Akd ) .
d
(53)
d
Let Acom Aj Ak . Then, the (l 1, l 2) entry of Acom can be expressed as Acom (l 1, l 2)
n
A
d
j
(l 1, m) Akd (m, l 2)
,
where
.
1 l 1, l 2 n
Since
m 1
that m j max Adj (i, k ) 1 , m k max Akd (i, j ) 1 , we can then easily prove that,
n
n
1 i , j n
1 i , k n
Acom(l 1, l 2) 1 n , for any l 1, l 2 , i.e.,
Acom
max Acom(l 1, l 2) 1 .
n
1 l 1,l 2 n
(54)
Hence N ( Acom) n. Acom 1 .Therefore, (53) can be rewritten as
( Adj Akd ) N ( Adj Akd ) N ( Acom) 1 .
(55)
Based on (55), Theorem 4 can be proved in a further step.
D. Calculation of the ground reaction forces on the stance leg
Suppose the biped is supporting on leg 1. Then, based on the Lagrangian method,
the dynamical model of the single support phase is developed by the full set of
generalized coordinates q e [ x st , y st , z st , q 0, q1,..., q 8] , i.e.
041 F st
D1,e(q e)qe C 1,e(q e, q e)q e G1,e(q e) ,
Bu1 081
(56)
where D 1,e ( q e ) is the positive-definite (12 12) mass-inertia matrix, C 1,e( q e, q e) is
the (12 12) Coriolis matrix, G 1,e( q e ) is the (12 1) gravity vector, and F st is the
(4 1) ground reaction force vector on the stance leg. When we partition the
matrices D 1,e ( q e ) , C 1,e( q e, q e) , G 1,e( q e ) into the corresponding submatrices, (56) can be
rewritten by
D1,e11 D1,e12 q pos C 1,e11 C 1,e12 q pos
D1,e 21 D1,e 22 q C 1,e 21 C 1,e 22 q
G 1,e11 0 41 F st
G1,e 21 Bu1 081
,
(57)
where q pos [ x st , y st , z st , q 0] , q [q1,..., q 8] . From the first row of (57), we obtain
D1,e11q pos D1,e12q C 1,e11q pos C 1,e12q G1,e11 F st .
(58)
Since the values of the components of q pos [ x st , y st , z st , q 0] are always zero in the
single support phase, the ground reactive forces F st can be calculated from (58) and
we have
F st D1,e12q C 1,e12q G1,e11 .
(59)
REFERENCES
[1] J. W. Grizzle, C. Chevallereau, R. W. Sinnet, and A. D. Ames, “Models, feedback control, and
open problems of 3D bipedal robotic walking,” Automatica, vol. 50, no. 8, pp. 1955-1988, Aug, 2014.
[2] R. D. Gregg, and L. Righetti, “Controlled Reduction With Unactuated Cyclic Variables:
Application to 3D Bipedal Walking With Passive Yaw Rotation,” Ieee Transactions on Automatic
Control, vol. 58, no. 10, pp. 2679-2685, Oct, 2013.
[3] R. D. Gregg, A. K. Tilton, S. Candido, T. Bretl, and M. W. Spong, “Control and planning of 3-D
dynamic walking with asymptotically stable gait primitives,” Robotics, IEEE Transactions on, vol. 28,
no. 6, pp. 1415-1423, 2012.
[4] R. D. Gregg, and M. W. Spong, “Reduction-based control of three-dimensional bipedal walking
robots,” The International Journal of Robotics Research, vol. 29, no. 6, pp. 680-702, 2010.
[5] C. Chevallereau, J. W. Grizzle, and C.-L. Shih, “Asymptotically stable walking of a five-link
underactuated 3-D bipedal robot,” Robotics, IEEE Transactions on, vol. 25, no. 1, pp. 37-50, 2009.
[6] A. D. Ames, K. Galloway, K. Sreenath, and J. W. Grizzle, “Rapidly Exponentially Stabilizing
Control Lyapunov Functions and Hybrid Zero Dynamics,” Ieee Transactions on Automatic Control, vol.
59, no. 4, pp. 876-891, Apr, 2014.
[7] Y. Orlov, Y. Aoustin, and C. Chevallereau, “Finite time stabilization of a perturbed double
integrator—part I: continuous sliding mode-based output feedback synthesis,” Automatic Control,
IEEE Transactions on, vol. 56, no. 3, pp. 614-618, 2011.
[8] C. Chevallereau, E. Westervelt, and J. Grizzle, “Asymptotically stable running for a five-link,
four-actuator, planar bipedal robot,” The International Journal of Robotics Research, vol. 24, no. 6, pp.
431-464, 2005.
[9] E. R. Westervelt, J. W. Grizzle, and D. E. Koditschek, “Hybrid zero dynamics of planar biped
walkers,” Automatic Control, IEEE Transactions on, vol. 48, no. 1, pp. 42-56, 2003.
[10] J. W. Grizzle, G. Abba, and F. Plestan, “Asymptotically stable walking for biped robots: Analysis
via systems with impulse effects,” Automatic Control, IEEE Transactions on, vol. 46, no. 1, pp. 51-64,
2001.
[11] B. Morris, and J. W. Grizzle, “Hybrid invariant manifolds in systems with impulse effects with
application to periodic locomotion in bipedal robots,” Automatic Control, IEEE Transactions on, vol.
54, no. 8, pp. 1751-1764, 2009.
[12] R. Vasudevan, A. Ames, and R. Bajcsy, “Persistent homology for automatic determination of
human-data based cost of bipedal walking,” Nonlinear Analysis-Hybrid Systems, vol. 7, no. 1, pp.
101-115, Feb, 2013.
[13] E. R. Westervelt, J. W. Grizzle, C. Chevallereau, J. H. Choi, and B. Morris, Feedback control of
dynamic bipedal robot locomotion: CRC press Boca Raton, 2007.
[14] C.-L. Shih, J. Grizzle, and C. Chevallereau, “From stable walking to steering of a 3D bipedal
robot with passive point feet,” Robotica, vol. 30, no. 7, pp. 1119-1130, 2012.
[15] C. Shih, J. Grizzle, and C. Chevallereau, “Asymptotically stable walking and steering of a 3D
bipedal robot with passive point feet,” IEEE Transactions on Robotics, 2009.
[16] K. A. Hamed, and J. W. Grizzle, “Event-Based Stabilization of Periodic Orbits for Underactuated
3-D Bipedal Robots With Left-Right Symmetry,” Robotics, IEEE Transactions on, vol. 30, no. 2, pp.
365-381, 2014.
[17] K. A. Hamed, B. G. Buss, and J. W. Grizzle, “Exponentially Stabilizing ContinuousTime
Controllers for Periodic Orbits of Hybrid Systems: Application to Bipedal Locomotion with Ground
Height Variations,” The International Journal of Robotics Research, 2015.
[18] D. Tlalolini, C. Chevallereau, and Y. Aoustin, “Human-like walking: Optimal motion of a bipedal
robot with toe-rotation motion
” Mechatronics, IEEE/ASME Transactions on, vol. 16, no. 2, pp.
310-320, 2011.
[19] C.-B. Gan, C.-T. Ding, and S.-X. Yang, “Dynamical analysis and performance evaluation of a
biped robot under multi-source random disturbances,” Acta Mechanica Sinica, vol. 30, no. 6, pp.
983-994, 2014.
[20] C.-T. Ding, S.-X. Yang, and C.-B. Gan, “Input torque sensitivity to uncertain parameters in biped
robot,” Acta Mechanica Sinica, vol. 29, no. 3, pp. 452-461, 2013.
[21] R. A. Horn, and C. R. Johnson, Matrix analysis: Cambridge university press, 2012.
| 3cs.SY
|
∗
A look at the inner structure of the 2-adic ring C -algebra
and its automorphism groups
♯
†
♮
arXiv:1604.06290v4 [math.OA] 22 Jun 2017
Valeriano Aiello , Roberto Conti , Stefano Rossi *
†
Dipartimento di Matematica e Fisica
Università Roma Tre
Largo S. Leonardo Murialdo 1, 00146 Roma, Italy.
♯
Dipartimento di Scienze di Base e Applicate per l’Ingegneria
Sapienza Università di Roma
Via A. Scarpa 16, I-00161 Roma, Italy.
♮
Dipartimento di Matematica,
Università di Roma Tor Vergata
Via della Ricerca Scientifica 1, I–00133 Roma, Italy.
Abstract
∗
We undertake a systematic study of the so-called 2-adic ring C -algebra Q2 . This is the
∗
2
universal C -algebra generated by a unitary U and an isometry S2 such that S2 U = U S2 and
∗
∗ ∗
∗
S2 S2 + U S2 S2 U = 1. Notably, it contains a copy of the Cuntz algebra O2 = C (S1 , S2 )
through the injective homomorphism mapping S1 to U S2 . Among the main results, the rela∗
′
tive commutant C (S2 ) ∩ Q2 is shown to be trivial. This in turn leads to a rigidity property
enjoyed by the inclusion O2 ⊂ Q2 , namely the endomorphisms of Q2 that restrict to the
identity on O2 are actually the identity on the whole Q2 . Moreover, there is no conditional
expectation from Q2 onto O2 . As for the inner structure of Q2 , the diagonal subalgebra D2
∗
and C (U ) are both proved to be maximal abelian in Q2 . The maximality of the latter allows
a thorough investigation of several classes of endomorphisms and automorphisms of Q2 . In
particular, the semigroup of the endomorphisms fixing U turns out to be a maximal abelian
subgroup of Aut(Q2 ) topologically isomorphic with C(T, T). Finally, it is shown by an
explicit construction that Out(Q2 ) is uncountable and non-abelian.
Contents
1 Introduction
2
2 First results
2.1 The gauge-invariant subalgebra . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 The canonical representation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
5
6
3 Structure results
3.1 Two maximal abelian subalgebras . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Irreducible subalgebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 The relative commutant of the generating isometry . . . . . . . . . . . . . . . .
9
9
12
13
*
E-mail: [email protected], [email protected], [email protected]
1
4 Extending endomorphisms of the Cuntz algebra
4.1 Uniqueness of the extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Extensible Bogoljubov automorphisms . . . . . . . . . . . . . . . . . . . . . . .
15
15
18
5 Outer automorphisms
5.1 Gauge automorphisms and the flip-flop . . . . . . . . . . . . . . . . . . . . . . .
5.2 A general result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
21
22
6 Notable endomorphisms and automorphism classes
6.1 Endomorphisms and automorphisms α such that α(S2 ) = S2 . . . . . . . . . . .
6.2 Automorphisms α such that α(U ) = U . . . . . . . . . . . . . . . . . . . . . .
6.3 Automorphisms α such that α(U ) = zU . . . . . . . . . . . . . . . . . . . . . .
24
24
25
28
2
2
A The functional equation f (z ) = f (z) on the torus
2
2
30
1 Introduction
Ever since their formal debut in the most cited paper [17], the Cuntz algebras have received a great
deal of attention. The reasons are so many they resist any attempt to be only briefly accounted,
and this introduction will be no exception. Therefore, we cannot but draw a rather quick and
incomplete outline of the later developments until the present state of the art, if only to better
frame the scope of our work. For many authors who have focused their interest on more and more
general constructions inspired by the Cuntz algebras, there are as many authors who have devoted
themselves to as thorough as possible a study of the concrete Cuntz algebras. This study includes,
in particular, an in-depth investigation of endomorphisms and automorphisms. Cuntz is among
those who have undertaken both the tasks. As for the first, he and other authors have written a long
∗
series of works where increasingly broad classes of C -algebras associated with algebraic objects
∗
such as rings are contrived. In particular, in [18] he introduced a C -algebra QN associated with
the ax + b-semigroup over the natural numbers. A few years later, Larsen and Li [23] considered
its 2-adic version which, accordingly, they denoted by Q2 . The main object of our interest in the
∗
present paper, this novel C -algebra is in fact naturally associated with the semidirect product
semigroup of the additive group Z acted upon by multiplication with non-negative powers of 2. It
did appear before elsewhere, cf. [23] and the references therein, but it is in the above-mentioned
work of Larsen and Li that it was studied systematically for the first time. After recalling that Q2
∗
is a nuclear C -algebra, they prove, among other things, that Q2 is also a purely infinite simple
∗
C -algebra. They give two proofs of this fact. Notably, one is a straightforward application of
Q2 being a Cuntz-Pimsner algebra, to which general results of Exel, an Huef and Raeburn [20]
apply. From our viewpoint, this lucky circumstance is well worth mentioning. Indeed, very little
is known about the general structure of endomorphisms or automorphisms for general Pimsner
algebras, cf. [28, 19, 7]. Therefore, as should follow from some of the main results announced in
the abstract, a good way to look at Q2 might be to regard it as a felicitous example of a Pimsner
algebra for which a far-reaching study is not that prohibitive. Far be it from us, however, to
allege we have done all that could be done. Rather, our hope is that further research may stem
from this work. That for a Pimsner algebra a thorough comprehension of the properties of all
its endomorphisms is a virtually impossible task should be no surprise. Indeed, already for the
Cuntz algebras On the problem, at least in its full generality, has turned out to be well beyond
the reach of current research. For instance, all endomorphisms of On are known to come from
unitary elements of On via a correspondence first pointed out by Takesaki, see e.g. [16], and yet
it is notoriously difficult to find non-tautological and effective characterizations of those unitaries
that yield automorphisms, although a number of intriguing if partial results about the structure of
Aut(On ) have recently been achieved in [14, 11, 5, 12, 6, 8, 10], see also [9, 13] for an informative
2
account. Without further ado, we can now move on to the basic definitions needed throughout the
∗
paper. While being a Pimsner algebra, the 2-adic ring C -algebra is perhaps best described as the
∗
universal C -algebra Q2 generated by a unitary U and an isometry S2 such that
2
S2 U = U S2
and
∗
∗
∗
S2 S2 + U S2 S2 U = 1
The reader interested in its description in terms of Pimsner algebras is again referred to [23].
However, in this paper we will never need to resort to that picture, which is why we may as well
dispense with it. What we do need to observe is that Q2 contains a copy of the Cuntz algebra O2 .
∗
Indeed, the latter is by definition the universal C -algebra generated by two isometries X1 and X2
∗
∗
such that X1 X1 + X2 X2 = 1. Therefore, the map taking X1 to U S2 and X2 to S2 extends by
universality to a homomorphism from O2 to Q2 , which is injective thanks to the simplicity of O2 .
Accordingly, as of now it will be convenient to think of O2 as being a subalgebra of Q2 . To us
the rather explicit description of the inclusion O2 ⊂ Q2 was in fact among the strongest motivations to carry out the present study of Q2 , especially as far as the extension problem is concerned.
This asks whether an endomorphism of O2 extends to Q2 . It turns out that this is not always the
case. For instance, as soon as Bogoljubov automorphisms are looked at, easy examples are found
of non-extensible automorphisms. More precisely, we find that the only extensible Bogoljubov
automorphisms are the flip-flop, the gauge automorphisms and their products. In addition, facing
the extension problem in general leads to an interesting rigidity property enjoyed by the inclusion
O2 ⊂ Q2 , namely if an endomorphism of Λ of Q2 restricts to O2 trivially, then it is the identity
automorphism. To the best of our knowledge, the pair (O2 , Q2 ) is the only known example of a
non-trivial inclusion of Pimsner algebras that fulfills the rigidity condition. In this respect, it is
also worth mentioning that there is no conditional expectation from the larger onto the smaller of
the two.
Once these questions have been answered, it is natural to go on to study endomorphisms and
automorphisms of Q2 irrespective of whether they leave O2 globally invariant or not. Asking questions of this sort is of course motivated by the overwhelming literature written on similar issues for
the Cuntz algebras. However, this entails a preliminary study of the inner structure of Q2 . In this
∗
regard, we prove that both the C -algebra generated by U and the diagonal subalgebra D2 ⊂ O2
are maximal abelian. It came as a surprise to us to learn that not as many results as one would
∗
expect are known on maximal abelian subalgebras for general C -algebras. Apparently, that of
maximal abelian subalgebra is a notion far more relevant to von Neumann algebras. Yet Q2 seems
∗
to be one of the few exceptions, for its theory does benefit from C (U ) being such a subalgebra.
∗
Indeed, we exploit the maximality of C (U ) to derive a number of results on the general form
of selected classes of automorphisms, many of which are, incidentally, quasi-free in the sense of
Dykema-Shlyakhtenko and Zacharias, see [19, 28]. Notably, we show that the semigroup of the
endomorphisms of Q2 that fix U is in fact a maximal abelian subgroup of Aut(Q2 ) isomorphic
with C(T, T), the group of all continuous T-valued functions defined on the one-dimensional torus
T understood as the spectrum of U . These results are in fact in same spirit as those expounded
in [16]. Moreover, they indicate that it is not unconceivable to regard C(T, T) ⊂ Aut(Q2 ) as a
generalized maximal torus, although not connected, for which a kind of infinite-dimensional Weyl
theory might well worth attempting, as is done in [6, 8] for Cuntz algebras.
Since the group of equivalence classes of outer automorphisms of O2 is known to be so large
as to contain most locally compact groups, our investigation also addresses Out(Q2 ). We have
partial evidence to hold that Out(Q2 ) is not as large, not least because, as a drawback of the
aforementioned result on Bogoljubov automorphisms, we no longer have a general procedure for
embedding locally compact groups into Out(Q2 ) as we would do with Out(O2 ). At any rate, we
prove that Out(Q2 ) is still an uncountable non-abelian group. This is done in two steps. First,
3
we prove that both the flip-flop and the gauge automorphisms are mutually non-equivalent outer
automorphisms. Second, we provide a broad class of outer automorphisms that do not commute
in Out(Q2 ) with the flip-flop. Even so, the non-commutativity thus exhibited is admittedly of a
rather mild form. We do believe that it is an interesting, albeit difficult, problem to say to what
extent Out(Q2 ) is non-abelian.
A few words on the organization of the material are in order. The various results of the paper
are scattered throughout several sections, which more or less follow the order in which the topics
developed have been introduced above, as to allow the reader to find them more easily. For convenience, here follows a description of the content of the several sections, to be also understood as
a short guide to the main results. Section 2 is preparatory in character, as it sets the stage for our
subsequent considerations. Indeed, all the needed definitions and basic properties are to be found
∗
here. In Section 3.1 both C (U ) and D2 are shown to be maximal abelian subalgebras of Q2 ,
see Theorems 3.4 and 3.9, respectively. In Section 3.2 the extended canonical endomorphism is
proved to be a shift on Q2 , Theorem 3.14. Moreover, there is no conditional expectation from Q2
onto O2 , Theorem 3.16. The main result of Section 3.3 is Theorem 3.20, where the relative com∗
′
mutant C (S2 ) ∩ Q2 is shown to be trivial. Section 4 is focused on the uniqueness of extensions
of automorphisms from O2 to Q2 , which is proved in Theorem 4.5, and the non-extendability
of general Bogoljubov automorphisms, which is proved in Theorem 4.14. Section 5 deals with
the outerness of the gauge automorphisms and the flip-flop, but it also includes a general result,
see Theorems 5.1, 5.8, 5.9, respectively. Finally, Section 6 provides a complete description of
AutC ∗ (U ) (Q2 ), see Theorem 6.13. Moreover, this group is shown to be isomorphic with C(T, T)
and maximal abelian in Aut(Q2 ), see Theorems 6.14 and 6.16, respectively. The last two results
along with Theorem 6.9, which states that the outer automorphism group is non-abelian, should
be regarded as the main results of the present paper.
Throughout the paper, all endomorphisms are assumed to be unital and ∗-preserving. Finally,
for endomorphisms of the Cuntz algebra O2 we adopt the well-established notations to be found
in the wide literature of the field (see the beginning of Section 4 for a very short description of the
Cuntz-Takesaki correspondence). This is certainly the case for the symbols αz , ϕ, λf introduced
in full detail in the next section.
2 First results
∗
As we observed in the introduction, the 2-adic ring C -algebra contains a copy of the Cuntz al∗
gebra O2 , as the C -subalgebra generated by S2 and S1 ≐ U S2 . Since the theory of the latter
has been enriched by a deeper and deeper knowledge of distinguished classes of endomorphisms
as well as automorphisms, problems to do with their extensions to Q2 are undeniably among the
most natural things to initiate a study with. We will see in Section 4.1 that as soon as too much
generality is allowed, these problems begin to be intractable for all practical purposes. If one asks
a bit more specific questions, a great many partial results do start cropping out. At any rate, we
will also obtain a general result, namely that whenever extensions exist they are unique, which
is yet another way to state the rigidity property of the inclusion O2 ⊂ Q2 we explained in the
introduction. The precise statement of this fact, too, is contained in Section 4.1. In the present
section, we limit ourselves to three remarkable examples that are easily dealt with. The first is the
canonical shift. The second is the flip-flop. The third are the gauge automorphisms.
∗
∗
The canonical shift is explicitly defined on every x ∈ O2 as ϕ(x) = S1 xS1 +S2 xS2 , therefore
if we set
∗ ∗
∗
ϕ(x)
̃
= U S2 xS2 U + S2 xS2 for any x ∈ Q2
4
we still define an endomorphism of Q2 , which restricts to O2 as the usual shift. The intertwining rules Si x = ϕ(x)S
̃
i for any x ∈ Q2 with i = 1, 2 still hold true. Moreover, a straight2
forward computation shows that ϕ(U
̃ ) = U . Since the continuous functional calculus of a
normal operator commutes with any endomorphism, the above equality can also be rewritten as
2
ϕ(f
̃ (U )) = f (U ), which is true for any continuous function f . It goes without saying that the
same equality retains its validity with any Borel function whenever Q2 is represented on some
Hilbert space. We shall avail ourselves of this useful fact later on.
As is well known, the flip-flop is the involutive automorphism λf ∈ Aut(Q2 ) that switches
S1 and S2 with each other. The flip-flop extends as well, although the proof is less obvious and
needs an argument. This is done here below.
Proposition 2.1. ([15]) The flip-flop automorphism of O2 extends to an automorphism of Q2 .
′
∗
′
′
′∗
′ ′
′∗
∗
Proof. If we set U ≐ U and S2 ≐ U S2 , then the identity S2 S2 +U S2 S2 U = 1 is immediately
̃ f ∈ End(Q2 ) such that
checked. By universality of Q2 , there exists a unique endomorphism λ
̃λf (U ) = U ′ = U ∗ and λ̃ f (S2 ) = S2′ = U S2 . This endomorphism is necessarily injective as Q2
̃ f (U ∗ ) = U and λ̃ f (U S2 ) = S2 , the image of α must be the whole Q2 , that is
is simple. Since λ
to say λ̃ f is an automorphism. Finally, it is obviously an extension of the flip-flop.
□
As of now, the above extension will be referred to simply as the flip-flop of Q2 and will be
̃f .
denoted by λ
It is also well known that the Cuntz algebra O2 is acted upon by T through the so-called gauge
automorphisms αz given by αz (Si ) = zSi , with z ∈ T. We can also prove the following extension
result concerning gauge automorphisms.
Proposition 2.2. The gauge automorphisms of O2 can all be extended to automorphisms of Q2 .
′
′
Proof. Now we set U ≐ U and S2 ≐ zS2 , where z is any complex number of absolute value equal
′ ′ ′∗ ∗
′ ′∗
̃ z ∈ Aut(Q2 )
to 1 . As we still have S2 S2 + U S2 S2 U = 1, there exists an automorphism α
such that α
̃ z (U ) = U and α
̃ z (S2 ) = zS2 . To conclude, all that we are left to do is note that
α̃ z (S1 ) = α̃ z (U S2 ) = α
̃ z (U )α̃ z (S2 ) = U zS2 = zS1 .
□
With a slight abuse of terminology, the automorphisms α̃ z obtained above will be referred
to as the gauge automorphisms. To conclude, it is worth noting that the flip-flop and the gauge
automorphisms commute.
2.1 The gauge-invariant subalgebra
The gauge-invariant subalgebra of O2 , usually denoted by F2 , is known to be isomorphic with
the CAR algebra. The corresponding gauge-invariant subalgebra of Q2 , which throughout this
∗
T
paper will be denoted by Q2 , can no longer be identified with such a remarkable C -algebra.
However, it can be described far more conveniently as the closure of a suitable linear span. To do
so, we need to point out the following simple but useful result. In order to state it as clearly as
possible, let us first set some notation. As in [17], we denote by W2 the set of all multi-indices
µ = (µ1 , µ2 , . . . , µn ) with µi ∈ {1, 2} and n ∈ N; the integer n is commonly referred to as the
length of the multi-index µ and is denoted by ∣µ∣. For any such multi-index µ = (µ1 , µ2 , . . . , µn ),
we denote by Sµ the monomial Sµ1 Sµ2 . . . Sµn .
∗
k
Proposition 2.3. Q2 = span{Sµ Sν U ∶ µ, ν ∈ W2 , k ∈ Z}.
5
Proof. In order to prove the equality above all we have to do is observe that the following relations
∗
∗
allow us to take both U and U from the left to the right side of any monomial of the form Sµ Sν .
• U S1 = S2 U
• U S2 = S1
∗
∗
∗
∗
• U S1 = S2 U
• U S2 = S2 U
2
∗
• U S1 = S2
∗
• U S2 = S1 U
∗
• U S1 = S1 (U )
∗ ∗
∗
∗ ∗
∗
• U S2 = S1 U
∗ 2
∗
The relations themselves are immediately verified by direct computation instead.
□
̃ from Q2 onto QT2 by averaging
The gauge automorphisms yield a conditional expectation E
̃(x) = ∫ α̃ z (x)dz, with dz being the
the action itself on T, that is for any x ∈ Q2 we have E
T
∗ k
̃ µ Sν U ) = Sµ Sν∗ U k ∫02π ei(∣µ∣−∣ν∣)θ dθ , we also
normalized Haar measure of T. Now since E(S
2π
̃(Sµ Sν∗ U k ) = 0 if and only if ∣µ∣ ≠ ∣ν∣. This helps to prove the description alluded to
have E
above.
Proposition 2.4. The equalities below hold:
Q2 = span {Sµ Sν U ∶ µ, ν ∈ W2 , ∣µ∣ = ∣ν∣, k ∈ Z} = C (U, F2 ) ⊂ Q2
T
∗
∗
k
Proof. The second equality is obvious. We focus then on the first, for which we only have to worry
T
∗ k
about the inclusion Q2 ⊂ span{Sµ Sν U ∶ ∣µ∣ = ∣ν∣}, the other being immediately checked. If
T
̃(x). Now pick a sequence {xn } in the algebraic linear span of the set
x ∈ Q2 , then x = E
∗ k
{Sµ Sν U ∶ µ, ν ∈ W2 , k ∈ Z} such that ∥xn − x∥ tends to zero. As Ẽ is a bounded map,
̃(xn ) − E
̃(x)∥ = ∥E
̃(xn ) − x∥ tends to zero as well. The conclusion follows easily now
∥E
̃ n ) ∈ span{Sµ Sν∗ U k ∶ ∣µ∣ = ∣ν∣} by the remark we made above.
because E(x
□
We should also mention that C (F2 , U ) = C (D2 , U ) is the Bunce-Deddens algebra of type
∞
2 , see [3, Remark 2.8].
∗
∗
2.2 The canonical representation
In this section we gather as much information as we need about a distinguished representation of
Q2 , which will actually play a major role in most of what follows here and in the next sections.
As far as we know, it was first exhibited in [23], where it is called the canonical representation.
Therefore, from now on it will always be referred to as the canonical representation. After a brief
review of its main properties, we discuss a number of results where the canonical representation
proves to be rather useful.
The canonical representation acts on ℓ2 (Z) through the operators S2 , U ∈ B(ℓ2 (Z)) given by
S2 ek ≐ e2k and U ek ≐ ek+1 , where {ek ∶ k ∈ Z} is the canonical orthonormal basis of ℓ2 (Z), i.e.
ek (m) = δk,m . The very first thing to note is that 1 is the only eigenvalue of S2 , corresponding
to the one-dimensional eigenspace generated by e0 . This simple observation enables us to give a
short proof that the canonical representation is irreducible. Since we do not know of any reference
where this possibly known fact is explicitly pointed out, we do include an independent proof for
the reader’s convenience.
6
Proposition 2.5. The canonical representation of Q2 is irreducible.
Proof. Let M ⊂ ℓ2 (Z) be a Q2 -invariant closed subspace. If P is the associated orthogonal
′
projection, then P ∈ Q2 . In particular, S2 P = P S2 , and so S2 P e0 = P e0 . As the eigenspace of
S2 corresponding to the eigenvalue 1 is spanned by e0 , we must have either P e0 = e0 or P e0 = 0.
∗
In the first case, e0 ∈ M , and therefore C (U )e0 ⊂ M , which says that M = ℓ2 (Z) because e0
∗
⊥
⊥
is a cyclic vector for C (U ). In the second, e0 ∈ M instead. As above, M being Q2 -invariant
⊥
too, we have M = ℓ2 (Z), i.e. M = 0. Note, however, that O2 does not act irreducibly on ℓ2 (Z),
for the closed span of the set {ek ∶ k = 0, 1, . . . , } is obviously a proper O2 -invariant subspace. □
However, the canonical representation restricts to O2 as a reducible representation, which we
denote by π. More precisely, it is a direct sum of two inequivalent irreducible representations of
O2 . To see this, let us define H+ , H− ⊂ ℓ2 (Z) as the closed subspaces given by
H+ ≐ span{ek ∶ k ≥ 0}
and
H− ≐ span{ek ∶ k < 0}
The Hilbert space ℓ2 (Z) is immediately seen to decompose into the direct sum of these subspaces,
i.e. ℓ2 (Z) = H+ ⊕ H− . Furthermore, both H+ and H− are O2 -invariant, and finally they may be
checked to be O2 -irreducible too. This last statement should be a well-known fact. Even so, we
give the proof for the sake of self-containedness.
Proposition 2.6. The subspaces H± are both O2 -irreducible.
Proof. We only need to worry about H+ , for H− is dealt with in much the same way. Exactly
as above, if M ⊂ H+ is an O2 -invariant subspace, then either M or its orthogonal complement
⊥
M must contain e0 . The proof is thus complete if we can show that an O2 -invariant subspace
containing e0 , say N , is the whole H+ , and this is proved once we show ek ∈ N for every k ≥ 0.
This is in turn easily achieved by induction on k. Suppose we have proved {el ∶ l = 0, 1, . . . , k} ⊂
N . For the inductive step we have two cases, according as k + 1 is even or odd. If it is even, then
ek+1 = S2 e k+1 ; if it is odd, then ek+1 = S1 e k . In either cases we see that ek+1 is in N , as wished.
2
2
□
Denoting by π± the restriction of π to H± respectively, the decomposition into irreducible
representations π = π+ ⊕ π− has just been proved to hold. Now, as what we are really interested
′
in is the commutant π(O2 ) , we also need to observe that π+ and π− are disjoint. This is done
here below.
Lemma 2.7. If π+ and π− are the irreducible representations defined above, then π+ ⫰ π− .
Proof. It is enough to note that π+ (S2 ) has 1 in its point spectrum, whereas π− (S2 ) does not. □
To state the next result as clearly as possible some notation is needed, so let us denote by E±
the orthogonal projections onto H± respectively.
Proposition 2.8. The commutant of π(O2 ) is given by π(O2 ) = CE+ + CE− .
′
Proof. According to the decomposition described above we have π(O2 ) = (π+ (O2 )⊕π− (O2 )) .
But because π+ and π− are disjoint, we can go a bit further and write
′
′
(π+ (O2 ) ⊕ π− (O2 )) = (π+ (O2 )) ⊕ (π− (O2 )) = CE+ ⊕ CE−
′
′
where the last equality is due to the irreducibility of π± .
7
′
□
This immediately leads to the following corollary, which needs no proof.
Corollary 2.9. In the canonical representation of Q2 the bicommutant of O2 is given by
π(O2 ) = {T ∈ B(ℓ2 (Z)) ∶ T = T+ ⊕ T− ∶ T± ∈ B(H± )}
′′
The information we have gathered about π will actually turn out to be vital in tackling the
∗
′
problem as to whether C (U ) ∩ O2 ⊂ Q2 is trivial. This is the case indeed, as anyone would
expect. However, the relative proof is not as obvious as the statement. In fact, we still lack some
basic ingredients. In particular, we need to observe that the basis vectors ek are all cyclic and
∗
∗
separating for U . Therefore, the W -algebra W (U ) generated by U is a maximal abelian von
∗
Neumann algebra of B(ℓ2 (Z)). Furthermore, it is common knowledge that W (U ) is isomorphic
∞
with L (T, µ), where µ is the Haar measure of T.
In passing, we also take the opportunity to exploit the canonical representation to show that
⊂ B(ℓ2 (Z)) is a maximal abelian subalgebra as well. This will in turn be vital to conclude that
D2 is a maximal abelian subalgebra of Q2 . Henceforward we shall denote by ℓ∞ (Z) the atomic
MASA of B(ℓ2 (Z)) acting through diagonal operators with respect to the canonical basis.
′′
D2
Proposition 2.10. In the canonical representation we have D2 = ℓ∞ (Z).
′
Proof. Since ℓ∞ (Z) is a MASA, it is enough to prove that D2 = ℓ∞ (Z), which will be immediately checked once we have proved that the projections Ek onto Cek all belong to the strong
∗ n
n
closure of D2 . To begin with, we note that the sequence {S2 (S2 ) } ⊂ D2 strongly converges to
k n
∗ n −k
E0 . But then the sequence {U S2 (S2 ) U ∶ n ∈ N} strongly converges to Ek . The conclusion
now follows from the fact that D2 is globally invariant under ad(U).
□
′′
We now have all the necessary tools to get to prove that C (U ) ∩ O2 is trivial. This will in
turn result from a straightforward application of the next proposition, where much more is proved.
∗
Proposition 2.11. We have W (U ) ∨ π(O2 ) = B(ℓ2 (Z)).
∗
′
′
Proof. Let P be an orthogonal projection in the commutant of W (U ) ∨ π(O2 ) . From P E+ =
E+ P and P E− = E− P , H± are straightforwardly seen to be both P -invariant. In particular, P e0
must take the form P e0 = ∑k≥0 ak ek . For the same reason, P e−1 is in H− , but it is also given by
∗
∗
P e−1 = P U e0 = U P eo = ∑k≥0 ak ek−1 , which means ak = 0 for every k > 0. In other words,
e0 must be an eigenvector of P . As such, we have either P e0 = 0 or P e0 = e0 . In the first case
∞
P = 0, whilst in the second P = 1, because P f (U )e0 = f (U )P e0 for every f ∈ L (T).
□
∗
′
Corollary 2.12. In the canonical representation W (U ) ∩ π(O2 ) = C1. As a consequence,
∗
′
C (U ) ∩ O2 = C1.
∗
′′
Remark 2.13. We have included an elementary proof of 2.11 for the reader’s convenience. How∗
ever, note that the rank-one orthogonal projections onto Cen , with n ∈ Z, all belong to W (U ) ∨
′
n
∗ n
π(O2 ) : indeed, we have U E+ (U ) = Espan{ek ∶ k≥n} for any n ∈ Z. In light of this, the above
proposition is well worth comparing with the more far-reaching classical result that for any discrete
group Γ the von Neumann algebra on ℓ2 (Γ) generated by λ(Γ) and the multiplication operators
Mf , with f ∈ c0 (Γ), is the whole B(ℓ2 (Γ)). Obviously, our case corresponds to Γ = Z.
∗
At this stage, there is another step to take to improve our knowledge of the C -algebra gener∗
ated by U . Indeed, we are yet to prove that C (U ) is a maximal abelian subalgebra of Q2 as well.
Since this task requires some technical work, we postpone the proof to the next section.
8
3 Structure results
3.1 Two maximal abelian subalgebras
The goal of the present section is to tackle two structure problems for Q2 , namely that both C (U )
∗
and D2 are maximal abelian subalgebras. We start with C (U ). The relative result is easily
guessed, and yet its proof is unfortunately far from being straightforward, in that it needs some
more refined tools such as conditional expectations from B(H) onto a maximal subalgebra. As
is known, the proof of the existence of such conditional expectations can be traced back to the
classic work of Kadison and Singer [21], where the authors first described a general procedure to
obtain them. Nowadays, the existence of conditional expectations of this sort is preferably seen
as an immediate consequence of the injectivity of abelian von Neumann algebras. Even so, we do
sketch the original procedure, not least because we make use of it in the proof of Theorem 3.6.
∣P
This runs as follows. Given T ∈ B(H), set T ≐ P T P + (I − P )T (I − P ) for any projection P
∗
∗
in W (U ). If {Pi } is a generating sequence of projections of W (U ), then every cluster point of
∣P ∣P ∣...∣Pn
∗
′
∗
the sequence {T 1 2
} lies in W (U ) = W (U ), as explained in [21]. This enables us to
∗
define a conditional expectation from B(H) onto W (U ) associated with each ultrafilter p ∈ βN
simply by taking the strong limit of the subnet corresponding to that ultrafilter.
∗
We next show as a key lemma to achieve our result that, for any conditional expectation E
∗
∗
from B(H) onto W (U ), we have that E[Sα Sβ ] is at worst a monomial in U . For the sake of
clarity, our proof is in turn divided into a series of preliminary lemmata.
Lemma 3.1. With the notations set above, the equalities E[S2 ] = E[(S2 ) ] = 0 hold for k ∈ N.
∗ k
k
Proof. First note that the second equality is a straightforward consequence of the first thanks
∗
∗
to the fact that E[T ] = E[T ] , which holds for every T ∈ B(H). The commutation rules
S2 U = U S2 give E[S2 ]U = U E[S2 ] = E[S2 ]U . If we now set f (U ) ≐ E[S2 ], we see
k
k
2
k
k
k
2
k
k
k
2
k
that f (z)(z − z ) = 0, hence f (z) = 0 for every z ∈ T. This says E[S2 ] = 0 and we are done.
□
k
2
k
Lemma 3.2. We have that E[S2 (S2 ) ] = 2
−k
∗ k
k
for every non-negative integer k.
Proof. To begin with, we observe that
∣α∣
∗ ∣α∣
E[Sα Sα ] = U E[S2 (S2 )
∗
h
]U
−h
∣α∣
for some positive integer h. Since 1 = ∑∣α∣=k Sα Sα we have that
∗
∣α∣
∗ ∣α∣
1 = ∑ E[Sα Sα ] = 2 E[S2 (S2 )
∗
k
∣α∣=k
∗ ∣α∣
= E[S2 (S2 )
]
].
This implies that E[S2 (S2 ) ] = 2 .
k
−k
∗ k
□
Lemma 3.3. We have that E[S2 (S2 ) ] = 0 for k, m ≠ 0, k ≠ m.
∗ m
k
Proof. Thanks to the last two lemmas, it suffices to show the statement for k > m > 0. By using
the commutation rules S2 U = U S2 we get U E[S2 (S2 ) ]U
k
k
2
k
2
k
k
∗ m
m
−2
= E[S2 (S2 ) ]. If we
k
∗ m
− 1)f (z) = 0, which
now set f (U ) ≐ E[S2 (S2 ) ], we obtain the functional equation (z
k
∗ m
clearly implies f (z) = 0 for every z ∈ T. This shows that E[S2 (S2 ) ] = 0 and we are done. □
k
∗ m
k
m
2 −2
We now have all the necessary information to carry out our proof of C (U ) being a maximal
abelian algebra.
∗
9
Theorem 3.4. C (U ) is a maximal abelian C -subalgebra of Q2 .
∗
∗
Proof. As C (U ) ∩ Q2 = W (U ) ∩ Q2 , it is enough to prove that given f ∈ L (T) with
f (U ) in Q2 , then f is in fact a continuous function. Now if f (U ) belongs to Q2 , it is also the
∗ h
norm limit of a sequence {xk } ⊂ Q2 with each xk taking the form ∑α,β,h cα,β,h Sα Sβ U . If
∗
E ∶ B(H) → W (U ) is any of the conditional expectations considered above, we have f (U ) =
h+k
k
E(f (U )) = limk E(xk ). But then each E(xk ) is of the form ∑α,β,h cα,β,h U α,β , where U α,β
∗
is nothing but E(Sα Sβ ). Consequently, there exists a sequence of Laurent polynomials pk such
∗
that ∥f (U ) − pk (U )∥ tends to zero, that is f (U ) ∈ C (U ), as maintained.
□
∗
′
∗
∞
Among other things, it is interesting to note that the former proof yields a distinguished condi∗
tional expectation from Q2 onto C (U ), which is simply obtained by restricting any of the afore∗
said conditional expectations to Q2 . Although there are conditional expectations onto W (U )
aplenty, as proved in [21], the above computations also show that E is in fact unique, a fact worth
a statement of its own.
Theorem 3.5. The conditional expectation E ∶ Q2 → C (U ) is unique.
∗
To complete the picture, we next show that E is faithful. This is actually a straightforward
consequence of a general well-known result due to Tomiyama [27], whose proof in our setting
is nevertheless included for the sake of completeness, being utterly independent of Tomiyama’s
work to boot.
Proposition 3.6. The unique conditional expectation E above is faithful.
Proof. By uniqueness it is enough to make sure that the conditional expectation yielded by the
2
Kadison-Singer procedure is faithful. Now if T is an α-coercive operator, i.e. (T x, x) ≥ α∥x∥
∣P
with α > 0, then T is α-coercive as well regardless of the projection P . In particular, if T ∈
B(H) is a coercive operator, then E[T ] cannot zero, being by definition a weak limit of coercive
positive operators all with the same constant as T . If now T is any non-zero positive operator
and ε > 0 is any real number with ε < ∥T ∥, the spectral theorem provides us with an orthogonal
decomposition H = Mε ⊕ Nε with Mε and Nǫ both T -invariant and such that the restriction
T ↾Nε is ε-coercive. The remark we started our proof with allows to conclude that E[T ] is not
zero either.
□
We can now move on to D2 . Again, the techniques we employ make a rather intensive use of
conditional expectations. Before we start, it is worth mentioning that this result can be understood
as a generalization of the well-known property of D2 being maximal in O2 . We start attacking the
problem with the following couple of lemmas, for which we first need to set some notation. We
still denote by E the unique faithful conditional expectation from B(H) onto ℓ∞ (Z). As known,
this is simply given by (E[T ]ei , ej ) = (T ei , ej )δi,j .
Lemma 3.7. The following relations hold:
• E[U ] = δk,0 I,
k
• E[Sα Sα ] = Sα Sα ,
∗
∗
• if ∣α∣ ≠ ∣β∣, E[Sα Sβ U ] is either 0 or Ei ,
∗
h
• if ∣α∣ = ∣β∣, E[Sα Sβ U ] is either 0 or Sα Sβ U .
∗
∗
h
10
h
Proof. The first two equalities need no proof. For the third, without harming generality, we may
suppose that ∣α∣ < ∣β∣
E[Sα Sβ U ] = E[U
∗
h
= E[U
=U
h(α)
h(α)
h(α)
(S2 )
(S2 )
(S2 )
∣α∣
∣α∣
∣α∣
∗ ∣β∣
(S2 )
∗ ∣α∣
(S2 )
U
U
h−h(β)
−h(α)
]
U
h(α)
∗ ∣β∣−∣α∣
(S2 )
U
h−h(β)
]
∗ ∣α∣ −h(α)
h(α)
∗ ∣β∣−∣α∣ h−h(β)
(S2 ) U
E[U
(S2 )
U
].
where h(α) and h(β) are positive integers. Accordingly, we are led to compute E[U (S2 ) U ],
h
k l
k
where h ∈ Z. Now the condition U (S2 ) U ei = ei implies that i = 2 (i + l) + h, and because
the former equation has a unique solution, we get the thesis. Finally, for the fourth we have that
h
E[Sα Sβ U ] = E[U
∗
h
= E[U
=U
h(α)
h(α)
h(α)
(S2 )
(S2 )
(S2 )
∣α∣
∣α∣
∣α∣
∗ ∣β∣
(S2 )
∗ ∣α∣
(S2 )
∗ ∣α∣
(S2 )
= δh−h(β)+h(α),0 U
h(α)
= δh−h(β)+h(α),0 U
h(α)
∗
U
h−h(β)
U
U
−h(α)
−h(α)
(S2 )
(S2 )
∣α∣
∣α∣
h
= δh−h(β)+h(α),0 Sα Sβ U .
]
U
h−h(β)+h(α)
E[U
h−h(β)+h(α)
∗ ∣α∣
(S2 )
∗ ∣α∣
(S2 )
U
U
−h(α)
k
l
]
]
h−h(β)
□
In order to make our proof work, we also need to take into account the conditional expectation Θ from Q2 onto D2 described in [23]. We recall that this is uniquely determined by
Θ((S2 ) U f U S2 ) ≐ δi,i′ δl,l′ (S2 ) U f U S2 , where f is in F2 . Moreover, it is there shown to
be faithful too.
∗ i
−l
l
′
i
′
∗ i
−l
l i
Lemma 3.8. If ∣α∣ = ∣β∣, Θ[Sα Sβ U ] is either 0 or Sα Sβ U . In particular, Θ and E coincide
∗ h
on monomials Sα Sβ U with ∣α∣ = ∣β∣.
∗
∗
h
h
Proof. By direct computation. Indeed, we have
Θ[Sα Sβ U ] = Θ(U
∗
h
h(α)
(S2 )
∣α∣
= δh(α),−h+h(β) U
∗ ∣α∣
(S2 )
h(α)
(S2 )
U
h−h(β)
∣α∣
)
∗ ∣α∣ h−h(β)
(S2 ) U
□
Theorem 3.9. The diagonal subalgebra D2 ⊂ Q2 is a maximal abelian subalgebra.
Proof. As usual, all we have to do is make sure that the relative commutant D2 ∩Q2 = ℓ∞ (Z)∩Q2
reduces to D2 . Let x ∈ ℓ∞ (Z) ∩ Q2 , then there exists a sequence {xk } converging normwise to
∗ h
x with each of the xk of the form ∑ cα,β,h Sα Sβ U . As above, x = E(x) = limk E(xk ). Thanks
to the former lemmata, we can rewrite E(xk ) as dk + fk , where dk ∈ D2 and fk are all diagonal
finite-rank operators. Now, being dk = Θ(xk ), we see that dk must converge to some d ∈ D2 . But
then fk converge normwise to a diagonal compact operator, say k, which means k = x − d is in
Q2 , hence k = 0, being K(H) ∩ Q2 = {0}, and x = d ∈ D2 .
□
′
The former result readily leads to D2 being a Cartan subalgebra of Q2 . An anonymous referee
pointed out an alternative approach to show this by using the works of N. Larsen and X. Li, [23],
and J. Renault, [26]. We believe this approach is very elegant and we give a brief account of it.
∗
∗
In the first place, one should see the 2-adic ring C -algebra Q2 as a reduced groupoid C -algebra
11
(see [23, Section 5]). Indeed, if G = (Z[ 21 ]⋊2 )⋉Q2 is the transformation groupoid, where Z[ 12 ]
Z
and 2 act on Q2 by addition and multiplication, respectively, and F ∶= G∣Z2 , then Cr (F ) ≃ Q2 .
Now, the claim follows at once by using [26, Proposition 3.1] after showing that the set of isotropy
points of Z2 is countable, and thus that the set of points with trivial isotropy is dense (that is, the
corresponding action is topologically free [26, Sec. 6.1]). The canonical abelian subalgebra of
a crossed product with respect to a group is considered as the most typical example of MASA,
provided that the action is topologically free. In the present case, although Q2 is only a crossed
product by a semigroup, the situation is almost the same because it is just a corner of a crossed
product by a group. One advantage of this description of Q2 as a crossed product makes more
natural to recognize D2 as the canonical Cartan subalgebra (and the canonical expectation).
As for O2 , later on we will prove that there is no conditional expectation from Q2 onto O2 .
∗
Z
3.2 Irreducible subalgebras
In order to take a step further towards the study of Q2 , especially as far as the properties of the
inclusion O2 ⊂ Q2 are concerned, it is worthwhile to recall a useful result proved by Larsen
and Li in their aforementioned paper [23]. It says that a representation ρ of O2 extends to a
representation of Q2 if and only if the unitary part of the Wold decomposition of ρ(S1 ) and ρ(S2 )
are unitarily equivalent. Accordingly, once the unitary parts of the Wold decompositions have
been proved to be unitarily equivalent, the isometries are unitarily equivalent too because of the
relation U S1 = S2 U . As remarked by the authors themselves, this allows us to think of Q2 as a
sort of symmetrized version of O2 . Notably, the result applies to those representations π of O2 in
which both π(S1 ) and π(S2 ) are pure. Moreover, in such cases the sought extension is unique, as
pointed out in [23, Remark 4.2]. For the reader’s convenience, though, we do single out this as a
separate statement.
Proposition 3.10. Let π be a representation of O2 on the Hilbert space Hπ such that π(S1 )
̃ ∈ B(Hπ ) such that π(S2 )U
̃=
and π(S2 ) are both pure. Then there exists a unique unitary U
2
̃
U π(S2 ) and π(S1 ) = Ũ π(S2 ).
The recalled theorem by Larsen and Li would actually be enough to prove that the Cuntz
algebra O2 is irreducible in Q2 , as might be expected. Even so, this can also be derived from the
much stronger result that even the UHF subalgebra F2 has trivial relative commutant, which is
shown below.
′
Theorem 3.11. The UHF subalgebra F2 ⊂ Q2 is irreducible, i.e. F2 ∩ Q2 = C1.
′
′
Proof. As D2 is a subalgebra of F2 , we have F2 ∩ Q2 ⊂ D2 ∩ Q2 = D2 , where the last equality
′
depends upon D2 being a maximal abelian subalgebra of Q2 . Therefore, we find F2 ∩ Q2 =
′
′
□
F2 ∩ D2 ⊂ F2 ∩ F2 = Z(F2 ) = C1.
While immediately derived from the above theorem, the following couple of corollaries do
deserve to be highlighted.
′
Corollary 3.12. The relative commutant of O2 in Q2 is trivial, i.e. O2 ∩ Q2 = C1.
T
Corollary 3.13. The relative commutant of the gauge-invariant subalgebra Q2 is trivial, i.e.
T ′
(Q2 ) ∩ Q2 = C1.
Interestingly, the irreducibility of F2 also applies to the canonical shift, which turns out to
k
enjoy the so-called shift property, i.e. ⋂k ϕ̃ (Q2 ) = C1, whence its name. This important property should have first been singled out by R. T. Powers, who called it strong ergodicity, but we do
not have a precise reference for the reader. The canonical shifts of On are of course known to be
strongly ergodic, see [22] for a full coverage of the topic.
12
Theorem 3.14. The canonical endomorphism ϕ̃ of Q2 is a shift, i.e. ⋂k ϕ̃ (Q2 ) = C1.
k
Proof. Since ϕ̃ (x) = ∑γ∶∣γ∣=k Sγ xSγ , the equality ϕ̃ (x)Sα Sβ = Sα Sβ ϕ̃ (x), x ∈ Q2 , is
straightforwardly checked to hold true for every pair of multi-indices α and β with ∣α∣ = ∣β∣ = k.
k
k ′
k
∗
This says that ϕ̃ (Q2 ) is contained in (F2 ) ∩ Q2 for every k, where F2 ≐ span{Sµ Sν ∶ ∣µ∣ =
k
k ′
′
∣ν∣}. But then we have the chain of inclusions ⋂k ϕ̃ (Q2 ) ⊂ ⋂k (F2 ) ∩ Q2 ⊂ F2 ∩ Q2 = C1. □
∗
k
∗
k
∗ k
Remark 3.15. The former theorem says in particular that ϕ̃ is not surjective in a rather strong
sense. We can be more precise by observing that U is not in ϕ(Q
̃ 2 ). Indeed, by maximality of
∗
∗
∗
C (U ), any inverse image of U should lie in C (U ), but the restriction of ϕ̃ to C (U ) does not
yield a homeomorphism of T.
We can now go back to the announced result that conditional expectations onto the Cuntz
algebra O2 do not exist.
Theorem 3.16. There is no unital conditional expectation from Q2 onto O2 .
Proof. Suppose that such a conditional expectation does exist. We want to show that this leads
to E(U ) being U , which is obviously absurd. We shall work in any representation in which S1
and S2 are both pure, for instance the one described in [4]. If we compute E on the operator
n
∗
∗ n
n
n
U S1 S2 S2 (S1 ) by using the commutation rule S2 U = U S1 we easily get to the equality
E[U ]S1 S2 S2 (S1 ) = S2 S1 S2 (S1 )
n
∗
∗ n
∗
n
∗ n
But on the other hand we also have that U S1 S2 S2 (S1 ) = S2 S1 S2 (S1 ) . Accordingly, E(U )
n
∗
∗ n
and U must coincide on the direct sum of the subspaces Mn ≐ S1 S2 S2 (S1 ) Hπ , which can be
easily seen to be the whole Hπ .
□
∗
n
∗ n
n
∗
∗ n
3.3 The relative commutant of the generating isometry
This section is entirely devoted to proving that C (S2 ) ∩ Q2 is trivial. We first observe that this
∗
′
∗
∗
∗
is the same as proving that C (S1 ) ∩ Q2 is trivial, merely because ad(U )(C (S2 )) = C (S1 ).
For this, we still need some preliminary definitions and results.
∗
′
Given any k ∈ N, we set B2 ≐ span{SαSβ U ∣ ∣α∣ = ∣β∣ = k, h ∈ Z}. For every k we have
k
k+1
∗
∗
the inclusion B2 ⊂ B2 , which readily follows from the Cuntz relation 1 = S1 S1 + S2 S2 .
∗
k
h
k
Lemma 3.17. Let x ∈ B2 , then
1. (S1 ) xS1 ∈ C (U );
∗ k
k
∗
2. the sequence {(S1 ) xS1 } stabilizes to a scalar cx ∈ C.
∗ m
m
∗
h
Proof. Without loss of generality, suppose that x = Sα Sβ U . We have that
(S1 ) xS1 = δα,1 Sβ U S1 .
∗ k
∗
k
h k
If h > 0, then by using the relation S2 U = U S1 , we see that the r.h.s. of the last expression is
given by
∗
l(h)
l(h)
= δα,1 δβ,γ U
δα,1 Sβ Sγ U
where γ is a multi-index of length k. If h = 0, we obtain
∗ k
δα,1 Sβ S1 = δα,1 δβ,1 .
13
∗
∗
∗
When h < 0, by using the relations U S1 = S2 and U S2 = S1 U we obtain
∗
h k
δα,1 Sβ U S1 = δα,1 U
l(h) ∗
Sγ S1
= δα,1 δγ,1 U
l(h)
where γ is a multi-index of length k. We observe that in these cases we always have ∣l(h)∣ ≤ ∣h∣,
and this fact will be important in the sequel. For the second part of the thesis, we may suppose
that m > k + ∣h∣ + 1. The needed computations can be made faster in the canonical representation
(for brevity we write l instead of l(h)):
(S1 ) U S1 ej = (S1 ) U e2m j+2m −1 = (S1 ) e2m j+2m −1+l .
∗ m
l m
∗ m
∗ m
l
m
m
m
m
The expression above is non-zero if and only if 2 j + 2 − 1 + l = 2 i + 2 − 1 for some i, that
m
is to say l = 2 (i − j). But m > k + h + 1 and ∣l∣ ≤ ∣h∣, therefore we finally get i = j and l = 0.
□
Proposition 3.18. Let x ∈ Q2 = span{Sα Sβ U ∣∣α∣ = ∣β∣, h ∈ Z}. Then
∗
T
h
lim(S1 ) xS1 ∈ C
∗ h
h
h
f (k)
Proof. By hypothesis there exists a sequence xk ∈ B2 that tends normwise to x. Choose a pair
of natural numbers i and j. For any k ∈ N sufficiently larger than f (i) and f (j), by the former
∗ k
k
∗ k
k
lemma we have that (S1 ) xi S1 =∶ ci , (S1 ) xj S1 =∶ cj ∈ C. The sequence ci is convergent
since
∗ k
k
∗ k
k
∣ci − cj ∣ = ∥(S1 ) xi S1 − (S1 ) xj S1 ∥ ≤ ∥xi − xj ∥.
We denote by c the limit. Now the sequence (S1 ) xS1 tends to c as well.
∗ h
h
□
T
For any non-negative integer i we now define the linear maps Fi ∶ Q2 → Q2 given by
Fi (x) ≐ ∫ α
̃ z [x(S1 ) ]dz
∗ i
T
F−i (x) ≐ ∫ α
̃ z [S1 x]dz.
i
T
We observe that
Fi (x) = Fi (x)S1 (S1 )
i
F−i (x) =
∗ i
i
∗ i
S1 (S1 ) F−i (x).
(1)
(2)
Before proving the main result of the section, we also need to recall the following lemma, whose
proof can be adapted verbatim from the original [17, Proposition 1.10], where it is proved for the
Cuntz algebras instead.
Proposition 3.19. Let x ∈ Q2 be such that Fi (x) = 0 for all i ∈ Z. Then x = 0.
Now we have all the tools for completing our proof.
Theorem 3.20. Let w ∈ U (Q2 ) such that wS1 w = S1 . Then w ∈ T1.
∗
14
∗
Proof. First of all we observe that we also have wS1 w
Fi (w). Indeed, for i ≥ 0,
∗
= S1 . We have that S1 Fi (w)S1 =
∗
∗
S1 Fi (w)S1 = S1 (∫ α
̃ z [w(S1 ) ]dz) S1 = ∫ S1 α̃ z [w(S1 ) ]S1 dz
∗
∗
∗ i
∗
T
=∫
∗
S1 F−i (w)S1
T
=
∗
∗ i
α̃ z [S1 w(S1 ) S1 ]dz
∗
i
S1 (∫ α
̃ z [S1 w]dz) S1
T
∗ i
T
= ∫ α̃ z [w(S1 ) ]dz = Fi (w)
∗ i
T
= ∫ S1 α̃ z [S1 w]S1 dz
∗
i
T
= ∫ α̃ z [S1 S1 wS1 ]dz = ∫ α̃ z [S1 w]dz = F−i (w)
∗ i
i
T
T
By Proposition 3.18 we obtain that for each i ∈ Z one has
lim(S1 ) Fi (w)S1 = Fi (w) ∈ C.
∗ k
k
Equation (1)-(2) together imply that for i ≠ 0 we have Fi (w) = 0. Now Proposition 3.19 applied
to w − F0 (w) gives the claim.
□
Exactly as for the Cuntz algebras, we can also state a slight generalization of the former result,
which says an inner automorphism of Q2 cannot send Si to a scalar multiple of it.
Proposition 3.21. Let φ ∈ Aut(Q2 ) be such that φ(Si ) = zSi for some z ∈ T \ {1} and i = 1 or
i = 2. Then φ is an outer automorphism.
The proof is straightforwardly obtained by recasting Theorem 3.20 in terms of a unitary w such
∗
that wS1 w = zS1 . This is done for On in [24], from which some of the techniques deployed in
this section have actually been taken.
4 Extending endomorphisms of the Cuntz algebra
For what follows it may be convenient to recall that associated with any unitary V ∈ U (O2 )
there is an endomorphism λV of O2 , defined by λV (Si ) = V Si , i = 1, 2. Notably, the other
way round is also true, i.e. any endomorphisms λ ∈ O2 comes from a unique V . The bijective
correspondence thus obtained is often named after Cuntz and Takesaki. To take just one important
∗
∗
example, the flip-flop λf is nothing but the automorphism corresponding to f ≐ S1 S2 + S2 S1 .
4.1 Uniqueness of the extensions
This section is mostly concerned with the problem of extending endomorphisms of O2 to endomorphisms of Q2 . More precisely, we spot necessary and sufficient conditions for an extension
to exist. Before entering into the details, some comments are in order. Indeed, the equations
we get are hardly ever easy to verify save for the endomorphisms we already know of to extend. Notwithstanding their intrinsic difficulty, they do provide general information when applied
to idO2 . We can now go on with our discussion. To begin with, if V is a unitary of O2 such
that the associated endomorphism λV extends to an endomorphism λ̃ of Q2 , then one must have
V U S2 = V S1 = λV (S1 ) = λ̃(S1 ) = λ̃(U )λV (S2 ) = λ̃(U )V S2 . Therefore, λ̃(U )V S2 = V U S2
∗ ∗
∗ ∗
and thus U V λ̃(U )V S2 = S2 . Thus, setting W = U V λ̃(U )V , it holds W S2 = S2 and
∗
λ̃(U ) = V U W V . We now examine whether such extensions exist. We always have
∗
∗
∗
∗
∗
∗
∗
V S2 S2 V + V U W V V S2 S2 V V W U V
∗
15
= V (S2 S2 + U S2 S2 U )V
∗
∗
∗
∗
=VV
∗
=1
and we must have
V S2 λ̃(U ) = λ̃(U )V S2
2
or, equivalently,
V S2 V U W V
∗
= (V U W V ) V S2 = V U W V V U W V V S2 = V U W U W S2 = V U W U S2 .
∗ 2
∗
∗
We have thus shown the following result.
Proposition 4.1. Let V ∈ U (O2 ) and let λV ∈ End(O2 ) be the associated endomorphism. Then
λV extends to an endomorphism of Q2 if and only if there exists a unitary W ∈ Q2 such that
∗
W S2 = S2 and S2 V U W V = U W U S2 . For any such W , we have an extension λ̃ = λ̃V,W with
∗
λ̃(U ) = V U W V .
As shown later, the W defined above is uniquely determined in all the cases we have examined.
̃ is necessarily injective because of Q2 being simple. Moreover,
Furthermore, the endomorphism λ
if λV is an automorphism of O2 , then λ̃ is surjective if and only if the associated W is contained
in λ̃ (Q2 ). Moreover, for the extensions built above the following composition rule holds:
̃ λ (V ′ )V,W V ∗ λ̃ (W ′ )V
̃ V,W ◦ λ̃ V ′ ,W ′ = λ
λ
V
V,W
As an example, if ϕ̃ is the canonical shift introduced in Sect. 2, we have
ϕ̃ = λ̃ θ,U ∗ θU 2 θ
where θ = ∑i,j=1 Si Sj Si Sj ∈ U (F2 ) is the self-adjoint unitary flip.
2
∗ ∗
2
It is interesting to note that the extensions of the gauge automorphisms we have considered all
work with W = 1. This is not a case. In fact, the converse also holds true.
Proposition 4.2. Let V ∈ U (O2 ). If the associated endomorphism λV ∈ End(O2 ) extends to
λ̃V,1 , that is to say the choice W = 1 does yield an extension, then V = z1, for some z ∈ T.
∗
∗
2
Proof. If we put W = 1 in the equality S2 V U W V = U W U S2 , we get S2 V U V = U S2 . But
2
∗
∗
U S2 = S2 U , and so we must have S2 V U V = S2 U . Hence V U V = U , i.e. V commutes with
∗
′
U . Since V is a unitary, we also have V ∈ C (U ) ∩ O2 , which concludes the proof.
□
Extensions of the identity map of O2 , which obviously correspond to V = 1, may be looked
∗
at more closely. If we define W ≐ U λ̃(U ), we find that W is a unitary in Q2 such that λ̃(U ) =
∗
∗
∗ ∗
∗
∗ ∗
U W , W S2 = S2 and W S1 = S1 W . Indeed, S2 S2 + U W S2 S2 W U = S2 S2 + U S2 S2 U = 1
2
2
and S2 U W = (U W ) S2 , so that U S2 W = U W U W S2 and thus U S2 W = W U S2 . Hence,
S1 W = W S1 , as stated. Obviously, the trivial choice W = 1 corresponds to the trivial extension.
Proposition 4.3. If W ∈ U (Q2 ) is such that W S2 = S2 and W S1 = S1 W , then W = 1.
Proof. This is in fact a straightforward application of the fact that C (S1 ) ∩ Q2 = C1. However,
we can also give an alternative if longer proof that only depends on the theorem of Larsen and
Li we have quoted. The computations are easily made in the irreducible representation of O2
2
produced in [4]. This acts on the Hilbert space H = L ([−1, 1]) through the pure isometries
S1 , S2 ∈ B(H) given by the formulas:
∗
0
for −1 ≤ t ≤ 0
(S1 f )(t) = { √
2f (2t − 1) for 0 < t ≤ 1
√
2f (2t + 1) for −1 ≤ t ≤ 0
(S2 f )(t) = {
0
for 0 < t ≤ 1
16
′
√
√
) and (S2 f )(t) = 22 f ( t−1
) for every f ∈ H. The unitary
Note that (S1 f )(t) = 22 f ( t+1
2
2
operator U ∈ U (H) given by (U f )(t) = f (−t) for every t ∈ [−1, 1] is an intertwiner between
S1 and S2 , namely U S2 = S1 U . By virtue of the result of Larsen and Li we mentioned above,
we can then regard this representation as a representation of Q2 as well, which allows us to think
of Q2 as a subalgebra of B(H). In order to prove the proposition, we will actually show even
more: any unitary W ∈ B(H) such that W S2 = S2 and W S1 = S1 W must be the identity
operator I on H. To accomplish this task, we define a sequence of orthogonal projections given
n
∗
∗ n
by Pn ≐ (S1 ) S2 S2 (S1 ) for each n ∈ N. It is straightforwardly checked that W Pn = Pn for
n
every n and Pn Pm = Pm Pn = 0 for every m ≠ n. Therefore Qn ≐ ∑k=0 Pk is still an orthogonal
projection such that W Qn = Qn . Accordingly, the conclusion will be easily achieved once we
have proved that Qn converges to I in the strong operator topology. As easily recognized, we have
n
∗ n
the following explicit formulas for S1 and (S1 ) :
∗
∗
0
n
(S1 f )(t) = { √
and
1
for −1 ≤ t ≤ 1 − 2n−1
n
n
k
i
1
2 f (2 t − ∑i=0 2 ) for 1 − 2n−1
<t≤1
∗ n
((S1 ) f )(t)
√
n
t+2 −1
2
)
=( ) f(
2
2n
n
We can use them to see that Pn is the projection corresponding to the multiplication operator by
1
χ[1− 1 ,1− 1 ] , i.e. the characteristic function of the interval [1− 2n−1
, 1− 21n ]. As a consequence,
2n−1
2n
Qn is nothing but the projection associated with χ[−1,1−
topology, which was to be proved.
1
2n
] . Hence Qn → 1 in the strong operator
□
Remark 4.4. The representation of Q2 recalled in the proof of the above result is not equivalent
to the canonical representation, merely because its restriction to O2 is still irreducible, as proved
in [4], whereas the restriction of the canonical representation to O2 is not, as we remarked.
We are at last in a position to prove the following result that says that a non-trivial endomorphism of Q2 cannot fix O2 pointwise.
Theorem 4.5. If Λ ∈ End(Q2 ) is such that Λ ↾O2 = idO2 , then Λ = idQ2 .
Proof. A straightforward application of the former proposition.
□
Remark 4.6. Actually, the theorem just obtained strengthens the information that the relative
′
′
commutant O2 ∩ Q2 is trivial. For, if u ∈ O2 ∩ Q2 , then ad(u) is an automorphism fixing O2
pointwise. As such, ad(u) is trivial, hence u is a central element. Since Q2 is simple, u must be a
′
multiple of the identity, i.e. O2 ∩ Q2 = C1.
As a simple corollary, we can also get the following property of the inclusion O2 ⊂ Q2 .
Corollary 4.7. If Λ1 ∈ Aut(Q2 ) and Λ2 ∈ End(Q2 ) are such that Λ1 ↾O2 = Λ2 ↾O2 , then
Λ1 = Λ2 . In particular, Λ2 is an automorphism as well.
−1
Proof. Just apply the above theorem to the endomorphism Λ1 ◦ Λ2 , which restricts trivially to
O2 .
□
In particular, the extensions of both the flip-flop and the gauge automorphisms are unique.
Of course there are automorphisms of Q2 that do not leave O2 globally invariant. The most el∗
ementary example we can come up with is probably ad(U ). Indeed, ad(U )S1 = U S1 U = S2 =
∗
∗
∗
U S1 , ad(U )S2 = U S2 U = S1 U = U S2 Hence, ad(U )(O2 ) is not contained in O2 , be∗
cause S1 U is not in O2 . Even more can be said. Indeed, ad(U )(F2 ) is not contained in O2 either.
17
This is seen as easily as before, since for instance ad(U )(S1 S2 ) = U S1 S2 U does not belong to
∗
∗ ∗
∗ ∗
∗
∗
∗
O2 although S1 S2 belongs to F2 . Given that U S1 S2 U = S2 U S2 U = S2 U S1 U U = S2 U S1 ,
∗ ∗
∗
∗
if U S1 S2 U were in O2 , then U = S2 S2 U S1 S1 would in turn be in O2 , which is not. Even so,
ad(U ) does leave the diagonal subalgebra D2 globally invariant. This can be shown by means of
∗
k
easy computations involving the projections of D2 ≐ span{Sα Sα s.t. ∣α∣ = k} for every k ∈ N.
∗
∗
∗
We would like to end this section by remarking that for each Λ ∈ End(Q2 ) there still exists
a unique uΛ ∈ U (Q2 ) such that Λ(S2 ) = uΛ S2 and Λ(S1 ) = uΛ S1 , which is simply given by
∗
∗
uΛ = Λ(S1 )S1 + Λ(S2 )S2 . Furthermore, Λ leaves O2 globally invariant if and only if uΛ ∈ O2 .
This allows us to regard the map End(Q2 ) ∋ Λ → uΛ ∈ U (Q2 ) as a generalization of the wellknown Cuntz-Takesaki correspondence. Nevertheless, this map is decidedly less useful for Q2
than it is for O2 , not least because it is not surjective. In other words, there exist unitaries u in
U (Q2 ) such that the correspondence S1 → uS1 , S2 → uS2 do not extend to any endomorphism
of Q2 . Examples of such u are even found in U (O2 ), as we are going to see in the next section,
where we shall give a complete description of the extensible Bogoljubov automorphisms. For
the time being we observe that if a unitary u ∈ U (Q2 ) does give rise to an endomorphism Λu ,
∗
the equation uU S2 = Λu (U )uS2 must be satisfied. This says that Λu (U ) = uU W u for some
∗
W ∈ U (Q2 ) such that W S2 = S2 and S2 uU W u = U W U S2 . By the same computations as at
the beginning of the section, the converse is also seen to be true. Hence we obtain a complete if
hitherto unmanageable description of End(Q2 ). At any rate, our guess is that the above equations
∗
are hardly ever verified unless u is of a very special form, such as u = v ϕ(v
̃
) for any v ∈ U (Q2 ),
corresponding to inner automorphisms, u = z1, corresponding to the gauge automorphisms α̃z , or
∗ ∗
∗
u = S2 S2 U + U S2 S2 , corresponding to the flip-flop. In fact, this prediction is partly supported
by the result in the negative obtained in the next section. Moreover, it is still not clear at all whether
the map End(Q2 ) ∋ Λ → uΛ is injective, although its restriction to Aut(Q2 ) certainly is.
4.2 Extensible Bogoljubov automorphisms
We have seen some remarkable classes of automorphisms of O2 that extend to Q2 . However,
there is no a priori reason to expect every endomorphism of O2 to extend automatically to an
endomorphism of Q2 . In fact, we next give rather elementary examples of automorphisms of O2
that do not extend. Indeed, if we denote by ηα,β the automorphism of O2 defined by ηα,β (S1 ) =
αS1 and ηα,β (S2 ) = βS2 for any given α, β ∈ T, we have the following proposition.
Proposition 4.8. The automorphisms ηα,β ∈ Aut(O2 ) defined above extend to endomorphisms
of Q2 if and only if α = β.
Proof. Since S1 and S2 are unitarily equivalent in Q2 , their images αS1 and βS2 would be unitarily equivalent as well if an extension of ηα,β existed. In particular, we would find {α} =
σp (αS1 ) = σp (βS2 ) = {β}, where σp denotes the point spectrum. This is absurd unless α = β,
in which case the corresponding endomorphism does extend being but a gauge automorphism. □
It is no surprise that the same proof as above covers the case of the so-called anti-diagonal
automorphisms. These are simply given by ρα,β (S1 ) = αS2 and ρα,β (S2 ) = βS1 for any given
α, β ∈ T. Again, an automorphism ρα,β extends precisely when α = β. To complete the picture,
we shall presently determine which Bogoljubov automorphisms of O2 extend to endomorphisms
of Q2 . A suitable adaptation of some of the techniques developed by Matsumoto and Tomiyama
in [24] will be again among the ingredients to concoct the proof of the main result of this section.
This says that the extensible Bogoljubov automorphisms are precisely the flip-flop, the gauge
automorphisms, and their products, which altogether form a group isomorphic with the direct
18
product T × Z2 . To this aim, consider a unitary matrix
A=(
a b
2
) ∈ U (C ).
c d
and let α = λA be the corresponding automorphism of O2 , i.e. α(S1 ) = aS1 + cS2 = (aU + c)S2
2
and α(S2 ) = bS1 +dS2 = (bU +d)S2 . Set f (U ) = (bU +d) for short. The condition S2 U = U S2
2
implies that f (U )S2 α(U ) = α(U ) f (U )S2 . Suppose that α is extensible and denote by α
̃ such
an extension. Finally, set Ũ ≐ α(U
̃ ), S̃2 ≐ α(S2 ), S̃1 ≐ α(S1 ). From now on we shall always be
focusing on the case where a, b, c, d are all different from zero. That said, the first thing we need
to prove is the extension is unique provided that it exists.
Lemma 4.9. If λA extends, then its extension is unique.
Proof. By Proposition 3.10 all we need to check is that S̃1 is pure as an isometry acting on
n
∗ n
ℓ2 (Z). This entails ascertaining that ⋂n Ran[S̃1 (S̃1 ) ] = {0}. To this aim, let us set Mn ≐
n
∗ n
Ran[S̃1 (S̃1 ) ]. As Mn+1 ⊂ Mn , we have that E⋂n Mn = lim EMn strongly. Thus we are led
to show limn EMn = 0. For this it is enough to prove limn ∥EMn ek ∥ = 0 for every k. Now
n
n (α) n (α)
∗
the powers of S̃1 are given by S̃1 = ∑∣α∣=n cα Sα , where cα ≐ a 1 c 2
∈ C with n1 (α)
being the number of 1’s occurring in α and n2 (α) the number of 2’s occurring in α. We set
L ≐ max{∣a∣, ∣c∣} and observe that L < 1 by the hypotheses on the unitary matrix A. We have
that
∗ n
n
∥(S̃1 ) ek ∥ = ∣cα(k) ∣ ≤ L → 0 n → +∞
for a unique coefficient cα(k) that depends on k (this is actually a consequence of the fact that
∗
∑∣α∣=k Sα Sα = 1). This in turn implies the claim.
□
In light of the previous result, it is a very minor abuse of notation to denote by λA also its
extension to Q2 when existing.
T
Lemma 4.10. If λA extends, then Ũ ∈ Q2 .
T
Proof. Suppose that Ũ is not in Q2 . Then by definition there must exist a non-trivial gauge
automorphism α˜z such that α˜z (Ũ ) ≠ Ũ . By applying α˜z to both sides of the equalities S̃2 Ũ =
2
∗
∗ ∗
Ũ S̃2 and S̃2 S̃2 + Ũ S̃2 S̃2 Ũ = 1, we also get
α˜z (Ũ ) S̃2 = S̃2 α˜z (Ũ )
2
S̃2 S̃2 + α˜z (Ũ )S̃2 S̃2 α˜z (Ũ) = 1
∗
∗
∗
which together say there exists an endomorphism Λ ∈ End(Q2 ) such that Λ(S2 ) = S̃2 and
Λ(U ) = α˜z (Ũ ). Now Λ(S1 ) = Λ(U S2 ) = Λ(U )Λ(S2 ) = α˜z (Ũ )S̃2 = z̄ α˜z (Ũ S̃2 ) = z̄ α˜z (S̃1 ) =
S̃1 = λA (S1 ). A contradiction is thus arrived at because Λ and λ˜A are different maps.
□
Now we introduce some lemmas to prove that α̃(U ) is actually contained in C (U ).
∗
Lemma 4.11. For any x ∈ B2 , the element (S̃2 ) xS̃1 belongs to C (U ).
k
∗ k
k
∗
k
∗ k
∗ h
Proof. Suppose that x = Sα Sβ U , where ∣α∣ = ∣β∣ = k. If h ≥ 0, we have that (S̃2 ) xS̃1 =
∗ k
∗ h k
(S̃2 ) Sα Sβ U S̃1 is a polynomial in U . The case h ≤ 0 can be handled with similar computations.
□
k
∗ k
T
Lemma 4.12. Let x ∈ Q2 such that the sequence (S̃2 ) xS̃1 converges to an element z. Then
∗
z ∈ C (U ).
19
Proof. Let {yk }k≥0 be a sequence such that yk ∈ B2 and yk → x normwise. Then the thesis
follows from the following inequality
k
∥z − (S̃2 ) yk S̃1 ∥ ≤ ∥z − (S̃2 ) xS̃1 ∥ + ∥(S̃2 ) (x − yk )S̃1 ∥ .
∗ k
∗ k
k
∗ k
k
k
□
Lemma 4.13. We have that Ũ ∈ C (U ).
∗
k
k
k
k
̃
Proof. By applying λ
A to the identity U S1 = S2 U we get Ũ S̃1 = S̃2 Ũ . For all k ∈ N we have
∗ k
∗
k
that (S̃2 ) Ũ S̃1 = Ũ . Therefore, Ũ is in C (U ) thanks to Lemma 4.12, applied to x = Ũ .
□
We have verified that α(U ) = g(U ) for some g ∈ C(T), which turns out to be vital in proving
the following result.
Theorem 4.14. If α ∈ Aut(O2 ) is a Bogoljubov automorphism, then α extends to Q2 if and only
if α is the flip-flop, a gauge automorphism, or a composition of these two.
Proof. By the discussion at the beginning of this section it is enough to consider the case in which
a, b, c, d are all different from zero. All the computations are henceforth made in the canonical
2
representation. The condition f (U )S2 α(U ) = α(U ) f (U )S2 yields
f (U )S2 g(U ) = g(U ) f (U )S2
2
f (U )g(U )S2 = f (U )g(U ) S2 .
2
2
Since the point spectrum of U is empty, f (U ) is always injective, unless b = d = 0, in which
2
2
case A is not unitary. Thus g(U )S2 = g(U ) S2 . At the function level we must then have
2
2
l
g(z ) = g(z) for every z ∈ T. By continuity, we find that g(z) = z , for this see e.g. the
l
l+1
l
Appendix. Therefore g(U ) = U . We have that α(S1 ) = aS1 + cS2 = bU S2 + dU S2 . If we
compute the above equality on the vectors em , we get
ae2m+1 + ce2m = be2m+l+1 + de2m+l .
which is to be satisfied for each m ∈ Z. Therefore, there are only two possibilities to fulfill these
conditions:
1. l = 1, and a = d ≠ 0, b = c = 0;
2. l = −1, and b = c ≠ 0, a = d = 0.
The first corresponds to gauge automorphisms, whilst the second to the flip-flop and its compositions with gauge automorphisms.
□
5 Outer automorphisms
In this section the group Out(Q2 ) is shown to be non-trivial. More precisely, it turns out to be a
non-abelian uncountable group. A thorough description of its structure, though, is still missing.
As far as we know, it might well be chimerical to get.
20
5.1 Gauge automorphisms and the flip-flop
Below the flip-flop and non-trivial gauge automorphisms are proved to be outer. In fact, this
parallels analogue known results for O2 . Since gauge automorphisms are more easily dealt with,
we start our discussion focusing on them first. The next result (partly) follows from Proposition
3.21, however we give an alternative proof because it will shed further light on the properties of
the gauge automorphisms, see Remark 5.2.
Theorem 5.1. The extensions to Q2 of the non-trivial gauge automorphisms of O2 are still outer
automorphisms (and they are not weakly inner in the canonical representation).
Proof. We shall argue by contradiction. From now on Q2 will always be thought of as a concrete
subalgebra of B(ℓ2 (Z)) via the canonical representation. We will actually prove: no unitary
V ∈ B(ℓ2 (Z)) can implement a non-trivial gauge automorphism. Indeed, let z ∈ T different
from 1, and let Λz ∈ Aut(Q2 ) the corresponding gauge automorphism. If V is a unitary operator
on ℓ2 (Z) such that Λz = ad(V ) ↾Q2 , then in particular we must have Λz (U ) = ad(V )(U ),
∗
namely U = V U V . This shows that V commute with U . Since U generates a MASA, V must
∞
∗
take the form V = f (U ), for some f ∈ L (T); in particular it belongs to W (U ) too. But we
∗
also have zS2 = Λz (S2 ) = ad(V )(S2 ) = V S2 V , that is to say zS2 V = V S2 . If we now compute
this identity between operators on the vector e0 , we get zS2 V e0 = V e0 , i.e. S2 V e0 = z̄V e0 . As
a consequence, we also have V e0 = 0, merely because 1 is the only eigenvalue of S2 . Since
∗
∗
V ∈ W (U ) and e0 is a separating vector for W (U ) as well, we finally find that V is zero,
which is clearly a contradiction.
□
Remark 5.2. Actually, the proof given above says a bit more. Indeed, the non-trivial gauge
automorphisms of Q2 ⊂ B(ℓ2 (Z)) are not weakly continuous. For if they were, they should
clearly extend to an automorphism of B(ℓ2 (Z)), but this is absurd because the automorphisms of
B(ℓ2 (Z)) are all inner.
Among other things, we also gain the additional information that Out(Q2 ) is an uncountable
group, in that different gauge automorphisms give raise to distinct classes in Out(Q2 ). Indeed,
if α̃z and α̃w are two different gauge automorphisms, then by Proposition 3.21 there cannot exist
∗
any unitary u ∈ Q2 such that uα̃z (x)u = α̃w (x) for every x ∈ Q2 . For the sake of complete∗
ness we should also mention that every separable traceless C -algebra is actually known to have
uncountable many outer automorphisms [25].
Remark 5.3. Notably, the former result also provides a new and simpler proof of the well-known
fact that the gauge automorphisms on O2 are outer. However, the case of a general On cannot be
recovered from our discussion, and must needs be treated separately, as already done elsewhere.
As for the flip-flop, instead, we start our discussion by showing it is a weakly inner automorphism, which is the content of the next result.
Proposition 5.4. The extension of the flip-flop to Q2 is a weakly inner automorphism.
Proof. By definition, we only have to produce a representation π of Q2 such that λ̃f is im′′
plemented by a unitary in π(Q2 ) . The canonical representation does this job well. For if
∗
V ∈ U (ℓ2 (Z)) is the self-adjoint unitary given by V ek ≐ e−k−1 , the equalities V S1 V = S2
∗
and V S2 V = S1 are both easily checked. Since the canonical representation is irreducible, the
proof is thus complete.
□
This result should also be compared with a well-known theorem by Archbold [1] that the
flip-flop is weakly inner on O2 .
21
∗
Remark 5.5. The unitary V as defined above can be rewritten as V = PU = U P, where P is the
self-adjoint unitary given by Pek = e−k , k ∈ Z. Obviously, V is in Q2 if and only if P is. We shall
∗
∗ ∗
prove that P is not in Q2 in a while. At any rate, we observe the equality V (S1 V S1 +S2 V S2 ) =
∗
∗
∗
∗
S1 S2 + S2 S1 ≐ f ∈ O2 , which is immediately checked, and f V = S1 V S1 + S2 V S2 = V f .
Finally, it is worth noting that U = λZ (1), and that P is the canonical intertwiner between λZ and
∗
ρZ . In this picture, Q2 is thus the concrete C*-algebra on ℓ2 (Z) generated by Cr (Z) and the copy
of O2 provided by the canonical representation.
In spite of being weakly inner, λ̃f is an outer automorphism, as is its restriction to O2 . To
prove that, we first need to show that the unitary V above is up to multiplicative scalars the unique
operator in B(ℓ2 (Z)) that implements the flip-flop.
Proposition 5.6. If W ∈ B(ℓ2 (Z)) is such that ad(W ) ↾Q2 = λ̃f , then W = λV for some λ ∈ T.
Proof. First note that we must have ad(W ) = idB(ℓ2 (Z)) since the flip-flop is an involutive
′′
2
automorphism and Q2 = B(ℓ2 (Z)). Hence W is a multiple of 1, and therefore there is no loss
2
∗
of generality if we also assume that W = 1, i.e. W = W . From the relation W S1 W = S2 , we
2
get S1 W e0 = W e0 . Hence W e0 = λe−1 for some λ ∈ T. From e0 = W e0 = λW e−1 we get
∗
W e−1 = λ̄e0 . We now show that either λ = 1 or λ = −1. Indeed, from U W = W U it follows
∗
that W e−1 = W U e0 = U W e0 = U (λe−1 ) = λe0 = λ̄e0 , which in turn implies that λ is real. Of
course, we only need to deal with the case λ = 1. The conclusion is now obtained at once if we
∗
∗
∗
use the equality W U = U W inductively, for W ek+1 = W U ek = U W ek = U e−k−1 = e−k−2,
as maintained.
□
2
Remark 5.7. Of course, the uniqueness of V could also have been obtained faster merely by
irreducibility of Q2 . However, the proof displayed above has the advantage of showing how we
came across the operator V .
Here finally follows the theorem on the outerness of the flip-flop.
Theorem 5.8. The extension of the flip-flop is an outer automorphism.
Proof. Thanks to the former result, all we have to prove is that P is not in Q2 , which entails checking that P cannot be a norm limit of a sequence xk of operators of the form xk =
h
∗ h
∑k ck Sαk Sβk∗ U k . Indeed, if this were the case, we should have ε > ∥P −∑α,β,h cα,β,h Sα Sβ U ∥
for some finite sum of the kind ∑α,β,h cα,β,h Sα Sβ U , with ε > 0 as small as needed. If so, we
would also find the inequality
h
∥e−n − ∑ cα,β,h Sα Sβ U en ∥ = ∥Pen − ∑ cα,β,h Sα Sβ U en ∥ < ε
∗
∗
h
α,β,h
h
α,β,h
This inequality, though, becomes absurd as soon as ε < 1 and n is sufficiently large, i.e. n is
∗ h
2
bigger than the largest value of ∣h∣, for we would have ∥e−n − ∑α,β,h cα,β,h Sα Sβ U en ∥ =
1 + ∥ ∑α,β,h cα,β,h Sα Sβ U en ∥ ≥ 1, as ∑α,β,h cα,β,h Sα Sβ U en ∈ H+ .
∗
h
∗
2
h
□
Now, as we know λ̃f is outer, we would also like to raise the question whether there exists a
representation of Q2 in which λ̃f is not unitarily implemented. The answer would indeed complete
our knowledge of λ̃f itself.
5.2 A general result
We saw above that the flip-flop is an outer automorphism. This is not an isolated case, for every
automorphism that takes U to its adjoint must in fact be outer. This is the content of the next result.
22
Theorem 5.9. Every automorphism α ∈ Aut(Q2 ) such that α(U ) = U is an outer automorphism.
∗
∗
∗
Proof. All we have to prove is that there is no unitary W ∈ Q2 such that W U W = U . To
this aim, we will be working in the canonical representation. If W ∈ B(H) is a unitary operator
∗
∗
∗
∗
such that W U W = U , then we have W U W = PU P, hence PW U (PW ) = U , which
∞
says that PW commutes with U . Therefore, PW = f (U ) for some f ∈ L (T) with ∣f (z)∣ =
1 a.e. with respect to the Haar measure of T, hence W = Pf (U ). Then we need to show
that such a W cannot be in Q2 . If f is a continuous function, there is not much to say, for
∗
Pf (U ) ∈ Q2 would immediately imply that P = Pf (U )f (U ) is in Q2 as well, which we
know is not the case. The general case of an essentially bounded function is dealt with in much
∞
the same way apart from some technicalities to be overcome. Given any f ∈ L (T) and ε > 0,
thanks to Lusin’s theorem we find a closed set Cε ⊂ T such that µ(T \ Cε ) < ǫ and f ↾Cε is
continuous. This in turn guarantees that there exists a continuous function gε ∈ C(T, T) that
coincides with f on Cε by an easy application of the Tietze extension theorem. If Pf (U ) is in
∗
Q2 , then Pf (U )gε (U ) is also in Q2 . Note that f g¯ε = 1 + hε , where hε is a suitable function
∗
whose support is contained in T\Cε . In particular, we can rewrite Pf (U )gε (U ) as P +Phε (U ).
∗ h
If the latter operator were in Q2 , then we could find an operator of the form ∑α,β,h cα,β,h Sα Sβ U
such that ∥P + Phε (U ) − ∑α,β,h cα,β,h Sα Sβ U ∥ < ε. If N is any natural number greater than
the maximum value of ∣h∣ as h runs over the set the above summation is performed on, we should
∗ h
have ∥PeN + Phε (U )eN − ∑α,β,h cα,β,h Sα Sβ U eN ∥ < ε, namely
∗
h
∥e−N + Phε (U )eN − ∑ cα,β,h Sα Sβ U eN ∥ < ε
∗
h
α,β,h
But then we should also have
∥e−N +Phε (U )eN − ∑ cα,β,h Sα Sβ U eN ∥ ≥ ∥e−N − ∑ cα,β,h Sα Sβ U eN ∥−∥Phε (U )eN ∥
∗
∗
h
α,β,h
Hence
h
α,β,h
ε > ∥e−N + Phε (U )eN − ∑ cα,β,h Sα Sβ U eN ∥ ≥ 1 − ∥Phε (U )eN ∥
∗
h
α,β,h
The conclusion is now got to if we can show that ∥Phε (U )eN ∥ is also as small as needed. But the
norm ∥Phε (U )eN ∥ is easily computed in the Fourier transform of the canonical representation,
where it takes the more workable form (∫ ∣hε (z )z
−1
∣ dµ(z)) 2 and is accordingly smaller than
−N 2
1
2µ(T \ Cε ) 2 ≤ 2ε 2 . The above inequality becomes absurd as soon as ε is taken small enough. □
1
1
As a straightforward consequence, we can immediately see that the class of the flip-flop in
Out(Q2 ) do not coincide with any of the classes of the gauge automorphisms. In other terms,
∗
̃ −1
the automorphisms α̃ z ◦ λ
f are all outer, sending U in U . Furthermore, if we now denote
−1
∗
by χ−1 the automorphism such that χ−1 (S2 ) ≐ S2 and χ−1 (U ) ≐ U = U , then the above
result also applies to χz ≐ χ−1 ◦ α̃ z , which are outer as well by the same reason. Also note that
χz (S2 ) = zS2 . More interestingly, if z is not 1, the corresponding χz yields a class in Out(Q2 )
̃ f do
other than the one of the flip-flop. To make sure this is true, we start by noting that χ−1 and λ
not commute with each other. Even so, they do commute in Out(Q2 ), in that they even yield the
∗
∗
same conjugacy class. Indeed, we have ad(U ) ◦ χ−1 = λ̃ f , or equivalently λ̃ f ◦ χ−1 = ad(U ),
̃ f ◦ χ−1 has infinite order in Aut(Q2 ) while being the product of
which in addition says that λ
two automorphisms of order 2. From this our claim follows easily. For, if χz ◦ λ̃ f is an inner
̃ f = χ̃ z ◦ ad(U) implies at once that χ̃ z is inner
automorphism, the identity χz ◦ λ̃ f = χ̃ z ◦ χ−1 ◦ λ
̃ f ] do commute in
as well, which is possible for z = 1 only. However, the classes [χz ] and [λ
2
−1
Out(Q2 ), because χz ◦ λ̃ f ◦ χz ◦ λ̃ f = ad(U ). In order to prove that Out(Q2 ) is not abelian,
we still need to sort out a new class of outer automorphisms. This will be done in the next sections.
23
6 Notable endomorphisms and automorphism classes
6.1 Endomorphisms and automorphisms α such that α(S2 ) = S2
For any odd integer 2k + 1, whether it be positive or negative, the pair (S2 , U
) still satisfies the two defining relations of Q2 . This means that the map that takes S2 to itself and U
2k+1
to U
extends to an endomorphism of Q2 , which will be denoted by χ2k+1 . Trivially, this
∗
endomorphism extends the identity automorphism of C (S2 ). A slightly less obvious thing to
note is that these endomorphisms cannot be obtained as extensions of endomorphisms of O2 . In2k+1
2k+1
deed, χ2k+1 (S1 ) = χ2k+1 (U S2 ) = U
S2 , and U
S2 is not in O2 : if it were, we would
∗ 2k+1
∗ ∗
2k
∗
k
k
also find that S1 U
S2 = S2 U U U S2 = S2 S2 U = U would be in O2 , which is not.
In this way we get a class endomorphisms χ2k+1 , k ∈ Z with χ1 = id and χ−1 being clearly
an automorphism of order two. All these endomorphisms commute with one other and we have
χ2k+1 ◦ χ2h+1 = χ(2k+1)(2h+1) for any k, h ∈ Z. Phrased differently, the set {χ2k+1 ∶ k ∈ Z}
is a semigroup of proper endomorphisms of Q2 . One would like to know if the endomorphisms
singled out above give the complete list of the endomorphisms of Q2 fixing S2 . In other words,
the question is whether the set
U2 ≐ {V ∈ U (Q2 ) ∣ V S2 = S2 V, S2 S2 + V S2 S2 V
∗
2
∗
∗
2k+1
= 1}
2k+1
contains elements other than the U
with k ∈ Z above. As a matter of fact, answering this
question in its full generality is not an easy task. An interesting if partial result does surface,
though, as soon as we introduce an extra assumption. Going back to the endomorphisms χ2k+1 ,
we next show they are all proper apart from χ±1 . The proof cannot be considered quite elementary,
∗
in that it uses the maximality of C (U ).
Proposition 6.1. None of the endomorphisms χ2k+1 is surjective if 2k + 1 ≠ ±1.
Proof. Let A ⊂ Q2 the C -subalgebra of those x ∈ Q2 such that χ2k+1 (x) ∈ C (U ). We clearly
∗
have C (U ) ⊂ A. By simplicity of Q2 the endomorphism χ2k+1 is injective, which means A is
∗
∗
still commutative. Therefore, by maximality of C (U ), we must have A = C (U ). From this it
now follows that U is not in the range of χ2k+1 , for the restriction χ2k+1 ↾C ∗ (U ) is induced, at the
2k+1
spectrum level, by the map T ∋ z ↦ z
∈ T.
□
∗
∗
In addition, in the canonical representation the endomorphisms χ2k+1 cannot be implemented
by any unitary W ∈ B(ℓ2 (Z)). Indeed, we can state the following result.
Proposition 6.2. Let 2k + 1 be an odd integer different from 1. Then there is no unitary V in
∗
2k+1
B(ℓ2 (Z)) such that V S2 = S2 V and V U V = U
.
Proof. The proposition is easily proved by reductio ad absurdum. Let V be such a unitary as in
the statement. From V S2 e0 = S2 V e0 we deduce that V e0 is an eigenvector of S2 with eigenvalue
1. Without loss of generality, we may assume that V e0 = e0 . Now,
h
V eh = V U e0 = U
(2k+1)h
V e0 = U
(2k+1)h
e0 = e(2k+1)h , h = 0, 1, 2, . . .
and, similarly,
V e−h = e−(2k+1)h , h = −1, −2, . . . .
To conclude, it is now enough to observe that V is not surjective whenever 2k + 1 ≠ −1, whereas
the case of 2k + 1 = −1 leads to V being equal to P, which does not belong to Q2 .
□
Rather than saying what the whole U2 is, we shall focus on its subset U2 ∩ C (U ) instead,
which is more easily dealt with. This task is accomplished by the next result.
∗
24
Theorem 6.3. The set U2 ∩C (U ) is exhausted by the odd powers of U , i.e. U2
2k+1
{U
∶ k ∈ Z}.
∗
∩ C ∗ (U ) =
Proof. Let W ∈ U2 ∩C (U ). Then there exists a function f ∈ C(T, T) such that W = f (U ).
2
2
The condition S2 W = W S2 can be rewritten as S2 f (U ) = f (U ) S2 . On the other hand, we also
2
2
2
2
2
have S2 f (U ) = f (U )S2 . Therefore f (U )S2 = f (U ) S2 , that is f (U ) = f (U ) . Accord2
2
ingly, the function f must satisfy the functional equation f (z ) = f (z) . Being continuous, our
n
function f must be of the form f (z) = z for some integer n ∈ Z, see the Appendix. This means
n
∗
n
∗ −n
that W = U . If we also impose the condition on the ranges S2 S2 + U S2 S2 U = 1, we finally
find that n is forced to be an odd number, say n = 2k + 1.
□
∗
The result obtained above can also be stated in terms of endomorphisms of Q2 . With this in
∗
mind, we need to introduce a bit of notation. In particular, we denote by EndC ∗ (S2 ) (Q2 , C (U ))
∗
the semigroup of those endomorphisms of Q2 that fix S2 and leave C (U ) globally invariant.
Corollary 6.4. The semigroup EndC ∗ (S2 ) (Q2 , C (U )) identifies with {χ2k+1 ∶ k ∈ Z}. As a
result, we also have
∗
AutC ∗ (S2 ) (Q2 , C (U )) = {id, χ−1 } ≅ Z2
∗
The foregoing result might possibly be improved by dropping the hypothesis that our endo∗
morphisms leave C (U ) globally invariant also. This is in fact a problem we would like to go
back to elsewhere.
6.2 Automorphisms α such that α(U) = U
In this section we study those endomorphisms and automorphisms Λ ∈ End(Q2 ) such that
Λ(U ) = U . Of course, the problem of describing all of them amounts to determining the set
S2 ≐ {W ∈ Q2 ∶ W W = 1, W U = U W, W W + U W W U = 1} .
∗
2
∗
∗
∗
Curiously enough, it turns out that S2 can be described completely, which is what this section is
chiefly aimed at. We start our discussion by sorting out quite a simple class of automorphisms
f
of that sort. Given a function f ∈ C(T, T) we denote by β the automorphism of Q2 given by
f
f
β (U ) = U and β (S2 ) = f (U )S2 , which is well defined because the pair (f (U )S2 , U ) still
f
g
f ⋅g
satisfies the two defining relations of Q2 . Note that β ◦ β = β so that we obtain an abelian
subgroup of AutC ∗ (U ) (Q2 ) and that a constant function f (z) = w gives back the gauge automorphism α̃ w . Furthermore, we have the following result, which gives a sufficient condition on f
f
for the corresponding β to be outer. As the condition is not at all restrictive, the correspondence
f
f ↦ β , which is one to one, provides plenty of outer automorphisms.
Proposition 6.5. If f ∈ C(T, T) is such that f (1) ≠ 1, then β is an outer automorphism.
f
Proof. If V ∈ Q2 is a unitary such that β = ad(V ), then V commutes with U and therefore it
∗
f
is of the form g(U ) for some g ∈ C(T, T) by maximality of C (U ). The condition β (S2 ) =
∗
2 ∗
ad(V )(S2 ) yields the equation f (U )S2 = g(U )S2 g(U ) , that is f (U )S2 = g(U )g(U ) S2 . But
then g and f satisfy the relation f (z) = g(z)g(z 2 ) for every z ∈ T. In particular, the last equality
says that f (1) = g(1)g(1) = 1.
□
f
However, the condition spotted above is not necessary. This will in turn result as a consequence
of the following discussion. We will be first concerned with the problem as to whether an automorf
phism β may be equivalent in Out(Q2 ) to a gauge automorphism. If so, there exist z0 ∈ T and
∗
∗
W ∈ U (Q2 ) such that W U W = U and W f (U )S2 W = z0 S2 . As usual, the first relation says
2 ∗
that W = h(U ) for some h ∈ C(T, T), which makes the second into h(U )f (U )h(U ) = z0 ,
25
that is h satisfies the functional equation h(z)f (z)h(z 2 ) = z0 . The latter says in particular that z0
is just f (1). We next show that there actually exist many continuous functions f for which there
is no continuous h that satisfies
h(z)h(z ) = f (z)f (1) ≐ Ψ(z) .
2
(3)
Note that Ψ(1) = 1 and that there is no loss of generality if we assume h(1) = 1 as well. By
evaluating (3) at z = −1 we find h(−1) = 1 provided that f (−1) = f (1).
Remark 6.6. By density, the continuous solutions of the equation h(z)h(z ) = Ψ(z) are comn
pletely determined by the values they take at the 2 −th roots of unity. Furthermore, the value of
n
2
such an h at a point z with z = 1 is simply given by the interesting formula h(z) = n−1 1 2k .
2
The latter is easily got to by induction starting from the relation h(z ) = h(z)Ψ(z).
2
∏k=0 Ψ(z
)
f
Here is our result, which provides examples of β not equivalent with any of the gauge auiθ
iθ
tomorphisms. Let f ∈ C(T, T) be such that f (e ) = 1 for 0 ≤ θ ≤ π and f (e ) = −1 for
π
π + ε ≤ θ ≤ 2π − ε with 0 < ε ≤ 4 , then we have the following.
f
Proposition 6.7. If f ∈ C(T, T) is a function as above, then the associated β is not equivalent
to any gauge automorphism.
Proof. With the above notations, suppose that h(z) is a solution of (3) such that h(1) = h(−1) =
iπ
2
1, which is not restrictive. Since Ψ(i) = Ψ(e 2 ) = 1 and h(i ) = h(−1) = 1, we immediately see
i π
that h(i) = 1. As Ψ(z) = 1 for 0 ≤ θ ≤ π, by using the functional equation we find that h(e 2n ) =
5i π
2(5i π4 )
5i π
) = h(i) = 1,
1 for any n ∈ N. Consider then z = e 4 . We have Ψ(e 4 ) = −1 and h(e
which in turn gives h(e 4 ) = −1. By induction we also see h(e 2n+2 ) = −1. This proves that any
solution h of the functional equation with f as in the statement cannot be continuous.
□
5πi
5i π
We can now devote ourselves to answering the question whether Out(Q2 ) is abelian. It turns
f
out that it is not. Our strategy is merely to show that automorphisms β corresponding to suitable
f
functions f do not commute in Out(Q2 ) with the flip-flop. To begin with, if β does commute
̃ f in Out(Q2 ), then there must exist a unitary V ∈ Q2 such that λ
̃ f ◦ β f ◦ λ̃ f = ad(V ) ◦ β f .
with λ
Exactly as above, the unitary V is then a continuous function of U , say V = h(U ). In addition,
we also have
∗
∗
2 ∗
f (U )S2 = h(U )f (U )S2 h(U ) = h(U )f (U )h(U ) S2
and so we find that f and h satisfy the equation f (z̄) = h(z)f (z)h(z 2 ) for every z ∈ T, which
can finally be rewritten as f (z̄)f (z) = h(z)h(z 2 ), to be understood as an equation satisfied by
the unknown function h, with f being given instead. We next exhibit a wide range of continuous
functions f for which the corresponding h does not exist. To state our result as clearly as possible,
we fix some notation first. Let f ∈ C(T, T) be such that f (e
i 9π
8
i 9π
8
apart from a sufficiently small neighborhood of z = e
.
) = i, and f (z) = 1 everywhere
f
Proposition 6.8. If f ∈ C(T, T) is as above, then β does not commute with the flip-flop in
Out(Q2 ).
Proof. Repeat almost verbatim the same argument as in the foregoing proposition, now verifying
that h(e 2n ) = 1 first and then h(e 2n+3 ) = −1.
πi
9πi
□
Notably, this also yields the announced result on Out(Q2 ).
Theorem 6.9. The group Out(Q2 ) is not abelian.
26
Now we have got a better guess of what S2 might be, we can finally prove that it is in fact
exhausted by isometries of the form f (U )S2 , where f is a continuous function onto T. This
still requires some preliminary work. First observe that given any s ∈ S2 , a straightforward
∗
∗
∗
computation shows that both s S2 and s S1 commute with U , but then by maximality of C (U )
we can rewrite them as h(U ) and g(U ) respectively, with h and g being continuous functions.
Lemma 6.10. There exists a continuous function f such that s = f (U )S2 .
Proof. We start with the equality s = s (S1 S1 + S2 S2 ) = (s S1 )S1 + (s S2 )S2 , in which we
∗
∗
∗
∗
substitute the above expressions. This leads to s = g(U )S1 + h(U )S2 , that is s = S1 g(U ) +
∗
∗
2 ∗
2 ∗
2 ∗
S2 h(U ) = U S2 g(U ) + h(U ) S2 = (U g(U ) + h(U ) )S2 . Therefore, our claim is true
□
with f (z) = zg(z ∗ ) + h(z 2 ).
∗
∗
∗
∗
∗
∗
∗
∗
Lemma 6.11. With the notations set above, for every z ∈ T we have ∣h(z)∣ + ∣g(z)∣ = 1.
2
2
∗
Proof. It is enough to rewrite the equality s s = 1 in terms of h and g.
□
Lemma 6.12. With the notations set above, for every z ∈ T we have zh(z)g(z) + g(z)h(z) = 0.
∗
Proof. Once again it is enough to rewrite the equality s U s = 0, which is merely the orthogonality
relation between s and U s, in terms of h and g.
□
We are at last in a position to prove the main result on S2 .
Theorem 6.13. If s ∈ S2 , then there exists a f ∈ C(T, T) such that s = f (U )S2 .
Proof. At this stage, all we have to do is prove that ∣f (z)∣ = 1. But
2
∣f (z)∣ = (g(z 2 )z + h(z 2 )(g(z )z + h(z )) = 1 + zg(z 2 )h(z ) + z̄g(z )h(z ) = 1
2
2
2
2
2
2
□
As an immediate consequence, we finally gain full information on AutC ∗ (U ) (Q2 ).
Theorem 6.14. The equalities hold
EndC ∗ (U ) (Q2 ) = AutC ∗ (U ) (Q2 ) = {β ∶ f ∈ C(T, T)}
f
In particular, the semigroup EndC ∗ (U ) (Q2 ) is actually a group isomorphic with C(T, T).
f
Remark 6.15. The bijective correspondence f ↔ β is also a homeomorphism between C(T, T)
equipped with the uniform convergence topology and AutC ∗ (U ) (Q2 ) endowed with the norm
pointwise convergence.
We end this section by proving that AutC ∗ (U ) (Q2 ) is in addition a maximal abelian subgroup
of Aut(Q2 ).
Theorem 6.16. The group AutC ∗ (U ) (Q2 ) is maximal abelian in Aut(Q2 ).
Proof. We have to show that if α ∈ Aut(Q2 ) commutes with any element of AutC ∗ (U ) (Q2 )
then α is itself an element of the latter group. Now, the equality α ◦ ad(U ) = ad(U ) ◦ α gives
ad(α(U )) = ad(U ). Therefore, α(U ) = zU for some z ∈ T by simplicity of Q2 . The conclusion
is then achieved if we show that actually z = 1. Exactly as above, we also have ad(α(g(U ))) =
ad(g(U )) for any g ∈ C(T, T). Again, thanks to simplicity we see that g(zU ) = g(α(U )) =
α(g(U )) = λg(U ) for some λ ∈ T, possibly depending on g. In terms of functions we find the
equality g(zw) = λg g(w), which can hold true for any g ∈ C(T, T) only if z = 1. Indeed, when
n
z ≠ 1 is not a root of unity the characters w are the sole eigenfunctions of the unitary operator
2
Φz acting on L (T) as (Φz f )(w) ≐ f (zw). Finally, the case of a z that is a root of unity is dealt
with similarly.
□
27
Remark 6.17. The findings above are worth comparing with a result obtained in [16] that the
group of automorphisms of O2 fixing the diagonal D2 is maximal abelian too.
Moreover, the theorem enables to thoroughly describe the automorphisms that send U to its
adjoint, which have been shown to be automatically outer.
Theorem 6.18. If α is an automorphism of Q2 such that α(U ) = U , then α(S2 ) = f (U )S2 for
a suitable f ∈ C(T, T).
∗
̃ f ◦ α.
Proof. Just apply the former result to λ
□
f
Finally, the automorphisms β can also be characterized in terms of the Cuntz-Takesaki generalized correspondence we discussed at the end of Section 4.1. Indeed, they turn out to be precisely
∗ ∗
those Λ ∈ End(Q2 ) for which the corresponding W ≐ U u Λ(U )u equals 1, where u stands
∗
2
for uΛ for brevity. For W = 1 we find in fact the equality S2 uU u = U S2 = S2 U , whence
∗
uU u = U . Therefore by maximality there exists a function f ∈ C(T, T) such that u = f (U ),
f
that is Λ = β .
6.3 Automorphisms α such that α(U) = zU
The following discussion addresses the problem of studying those automorphisms Λ of Q2 such
that Λ(U ) = zU , with z ∈ T. We start tackling the problem by defining two operators acting on
′
′
k
ℓ2 (Z). The first is the isometry Sz , which is given by Sz ek ≐ z e2k . The second is the unitary Uz ,
k
which is given by Uz ek ≐ z ek . The following commutation relations are both easily verified:
• Uz U = zU Uz
′
• Uz S 2 = S z Uz
The first relation can also be rewritten as ad(Uz )(U ) = zU . We caution the reader that at this
level ad(Uz ) makes sense as an automorphism of B(ℓ2 (Z)) only, because we do not know yet
whether Uz sits in Q2 . If it does, the first relation says, inter alia, that Q2 also contains a copy
of the noncommutative torus Az in a rather explicit way, which is worth mentioning. In order to
decide what values of z do give a unitary Uz belonging to Q2 , the first thing to note is that if Uz is
in Q2 , then it must be in the diagonal subalgebra D2 , as shown in the following lemma.
Lemma 6.19. If Uz is in Q2 , then Uz ∈ D2 .
Proof. A straightforward application of the equality D2 ∩ Q2 = D2 , as Uz is in D2 = ℓ∞ (Z). □
′
′
The second thing to note is that the unitary representation T ∋ z ↦ Uz ∈ U (B(ℓ2 (Z))) is only
strongly continuous. This implies that not every Uz is an element of Q2 . For the representation
z ↦ Uz is only strongly continuous, which means the set {Uz }z∈T is not separable with respect
to the norm topology, whereas Q2 obviously is. The next result provides a first answer to the
question whether Uz belongs to D2 . More than that, it also gives an explicit formula for Uz .
n
Proposition 6.20. If z ∈ T satisfies z
2
= 1 for some natural number n, then Uz is in D2 .
k
i2π/2
Proof. Obviously only primitive roots have to be dealt with. But for such roots, say z = e
2
∑j=0
k−1
the unitary Uz may in fact be identified to the sum
,
j
z Pj , where the projection Pj belongs
to D2 , being more explicitly given by Pi1 i2 ...ik , where the multi-index (i1 , i2 , . . . , ik ) ∈ {1, 2} is
the j-th with respect to the lexicographic order in which 2 < 1 and the multi-index itself is read
from right to left.
□
28
k
The automorphisms obtained above are of course of finite order. More precisely, the order of
ad(Uz ) is just the same as the order of the corresponding z. In other words, what we know is
∗
n
that the automorphism of C (U ) induced by the rotation on T by a 2 -th root of unity extends to
n
an inner automorphism of Q2 , whose order is still finite being just 2 . Due to the lack of norm
continuity of the representation z → Uz , though, the case of a general z is out of the reach of the
foregoing proposition and must needs be treated separately with different techniques. To begin
with, we recall a result whose content should be well known. Nevertheless, we do include a proof
not only for the sake of completeness but also to set some notations we shall need in the following
considerations.
Lemma 6.21. Any projection P ∈ D2 is in the linear algebraic span of {Sα Sα }α∈W2 .
∗
∗
Proof. It is convenient to realize D2 as the concrete C -algebra C(K), with the spectrum K being
N
given by the Tychonoff product {1, 2} . If we do so, the projections of D2 are immediately seen
to identify with the characteristic functions of the clopens of K, and these are clearly the cylinder
sets in the product space. The conclusion now follows noting that for any multi-index α ∈ W2
the characteristic function of a cylinder Cα = {x ∈ K ∶ x(k) = αk for any k = 0, 1, . . . , ∣α∣}
∗
□
corresponds indeed to Sα Sα .
At this point, it remains to show that Uz does not belong to D2 for any other values of z ∈ T.
Although this could be done by means of explicit computations, as it was in an early version of
this paper, we prefer to present a rather elegant method suggested by the referee.
The Cantor set K = {0, 1} can also be realized as the ring of 2-adic integer numbers Z2 via
∞
n
∞
the bijective correspondence K ∋ x = {xn }n=0 ↔ ∑n=0 xn 2 ∈ Z2 . In this picture the former
digit 2 has to be replaced by 0. Accordingly, as of now we think of D2 as C(Z2 ). Furthermore, as
Z2 is by definition the completion of Z under the metric induced by the 2-adic absolute value, any
f ∈ C(Z2 ) is uniquely determined by its restriction to Z ⊂ Z2 . This gives an isometric inclusion
of D2 ≅ C(Z2 ) into C(Z) ⊂ ℓ∞ (Z), which is nothing but the canonical representation of D2
∗
on ℓ2 (Z). To see this, it is enough to note that the generating projections Sα Sα are indeed the
n
n−1
j
characteristic functions of the subsets {2 k + l ∶ k ∈ Z} ⊂ Z, where n = ∣α∣ and l = ∑j=0 αj 2 .
Phrased differently, we have obtained the following useful characterization.
N
Lemma 6.22. Let f ∈ ℓ∞ (Z) ⊂ B(ℓ2 (Z)). Then f is in D2 if and only if f ∶ Z → C extends to
a continuous function f˜ ∶ Z2 → C.
We are now in a position to state and prove the main result of the present subsection.
Theorem 6.23. Let z ∈ T. Then Uz ∈ D2 if and only if z is a root of unity of order a power of 2.
k
Proof. Thanks to Lemma 6.22, it is enough to make it plain when Z ∋ k ↦ z ∈ C extends to a
continuous function of Z2 . It is easily seen that this is the case if and only if z is a dyadic root of
unity.
□
For those z ∈ T such that Uz lies in Q2 we can say a bit more.
Proposition 6.24. Let z ∈ T be a dyadic root of unity and let α ∈ Aut(Q2 ) be such that α(U ) =
′
zU . Then there exists a f ∈ C(T, T) such that α(S2 ) = f (zU )Sz .
Proof. By its very definition ad(Uz −1 ) ◦ α(U ) = U . Therefore, we must have ad(Uz −1 ) ◦ α =
f
f
β for some f ∈ C(T, T). But then f (U )S2 = β (S2 ) = Uz −1 α(S2 )Uz , i.e. α(S2 ) =
′
□
Uz f (U )S2 Uz −1 = f (zU )Uz S2 Uz −1 = f (zU )Sz .
′
Remark 6.25. We have already seen that if Uz ∈ Q2 then Sz ∈ Q2 . The converse, too, is true. In
∗ ′
fact, one can easily observe that Uz = S2 Sz hence the claim follows. In particular, whenever Uz
is not in Q2 , the corresponding ad(Uz ) understood as an automorphism of the whole B(ℓ2 (Z))
does not even leave Q2 globally invariant.
29
A
2
2
The functional equation f (z ) = f (z) on the torus
2
2
This appendix presents a self-contained treatment of the functional equation f (z ) = f (z) , of
which we made an intensive use in the previous sections. Although the following facts might all
be well known, we do include complete arguments, because their proofs are not to be easily found
in the literature, however carefully examined.
2
2
Proposition A.1. Let f be a continuous function from T to T such that f (z ) = f (z) for every
n
z ∈ T. Then there exists a unique n ∈ Z such that f (z) = z .
2
2
Proof. Thanks to the compactness of T and the continuity of f , the winding number of f is a
well-defined integer n ∈ Z, for details see e.g. Arveson’s book [2, Chapter 4, pp 114-115]. The
−n
new function g(z) ≐ z f (z) still satisfies our equation. Furthermore, the winding number of g
2πih(z)
is zero by construction. Therefore, there exists h ∈ C(T, R) such that g(z) = e
for every
2
z ∈ T. Rephrasing the equation in terms of h, we find that h(z ) − 2h(z) must be an integer for
every z ∈ T. By connectedness, the function h is thus a constant. Obviously there is no lack in
2
generality if we also assume h(z ) − 2h(z) = 0 for every z ∈ T. Being bounded, the function h
n
is then forced to be identically zero, which finally leads to f (z) = z for every z ∈ T.
□
The above proposition can be regarded as a one-variable description of the characters of the
one-dimensional torus. It is worth pointing out, though, that it no longer holds true as soon as T is
replaced by the additive group R. In other words, there do exist continuous functions f ∶ R → T
2
itx
such that f (2x) = f (x) other than ft (x) ≐ e , which are obtained by exponentiating non-linear
continuous functions g ∶ R → R such that g(2x) = 2g(x) for every x ∈ R. However, any such g
cannot be everywhere differentiable with continuous derivative at 0.
The proof given above can be further simplified if we assume that f satisfies a stronger funcn
n
tional equation, i.e. f (z ) = f (z) for every n ∈ N and z ∈ T.
Proposition A.2. If f ∈ C(T, T) satisfies f (z ) = f (z) for every z ∈ T and n ∈ N, then there
k
exists a unique k ∈ Z such that f (z) = z .
n
n
Proof. Obviously, it is enough to prove that f (zw) = f (z)f (w) for any z, w ∈ T. Let z be a fixed
element of T, and let Cz ⊂ T be the set Cz ≐ {w ∈ T ∶ f (zw) = f (z)f (w)}. From the equality
n
n
n
f (z ) = f (z) we may note that Cz contains the set {z ∶ n = 0, 1, 2, . . .}. By continuity of f
2iπθ
we have in addition that Cz is closed. Hence Cz = T if z = e
, with θ being irrational. In other
words, for such z, we have f (zw) = f (z)f (w) for any w ∈ T. The full conclusion is now easily
got to by density.
□
Remark A.3. The continuity assumption cannot be left out in either the above propositions. To
see this, let R ⊂ T be the set ∪n Hn , where Hn ⊂ T is the subgroup of the nth roots of unity.
n
n
The function f that is 1 on R and f (z) = z on its complement still satisfies f (z ) = z for every
n ∈ N, as easily verified. Due to the density of R in the torus, this function is nowhere continuous.
Nevertheless, it is equal to the character z almost everywhere. This seems to indicate that any
measurable solution of the equation might equal a character almost everywhere. At any rate, it
n
n
is worth pointing up that the solutions of the equation f (z ) = f (z) do not enjoy automatic
continuity, unlike the solutions of the equation f (zw) = f (z)f (w), which are of course even
automatically differentiable.
Acknowledgments We would like to take this opportunity to thank László Zsidó for a fruitful
conversation about the functional equation discussed in the Appendix. We are also grateful to
Nicolai Stammeier for his valuable comments on the first draft of this paper. Finally, we owe the
30
referee a debt of gratitude for his or her particularly attentive perusal of the manuscript, which
resulted in many improvements not only to notation and presentation but even to the proof of the
main result in the Appendix.
References
[1] R. J. Archbold, On the ”flip-flop”automorphism of C (S1 , S2 ), Quart. J. Math. Oxford Ser.
(2) 30 (1979), 129-132.
∗
[2] W. Arveson, A Short Course on Spectral Theory, Springer-Verlag, Graduate Texts in Mathematics 209 (2002).
∗
[3] S. Barlak, T. Omland, and N. Stammeier, On the K-theory of C -algebras arising from
integral dynamics, arXiv:1512.04496, to appear in Ergodic Theory Dynam. Systems.
[4] M. D. Choi and F. Latrémolière, Symmetry in the Cuntz algebra on two generators, J. Math.
Anal. Appl. 387 (2012), 1050–1060.
[5] R. Conti, Automorphisms of the UHF algebra that do not extend to the Cuntz algebra, J.
Aust. Math. Soc. 89 (2010), 309–315.
[6] R. Conti, J. H. Hong, and W. Szymańsky, The restricted Weyl group of the Cuntz algebra and
shift endomorphisms, J. Reine Angew. Math. 667 (2012), 177-191.
[7] R. Conti, J. H. Hong, and W. Szymańsky, Endomorphisms of graph algebras, J. Funct. Anal.
263 (2012), 2529-2554.
[8] R. Conti, J. H. Hong, and W. Szymański, The Weyl group of the Cuntz algebra, Adv. Math.
231 (6) (2012), 3147-3161.
[9] R. Conti, J. H. Hong, and W. Szymański, Endomorphisms of the Cuntz Algebras, Banach
Center Publ. 96 (2012), 81–97.
[10] R. Conti, J. H. Hong, and W. Szymański, On conjugacy of maximal abelian subalgebras
and the outer automorphism group of the Cuntz algebra, Proc. Roy. Soc. Edinburgh 145 A
(2015), 269–279.
[11] R. Conti, J. Kimberley, and W Szymański, More localized automorphisms of the Cuntz algebras, Proc. Edinburgh Math. Soc. (Series 2) 53 (2010), 619–631.
[12] R. Conti, M. Rørdam and W. Szymański, Endomorphisms of On which preserve the canonical UHF-subalgebra, J. Funct. Anal. 259 (2010), 602–617.
[13] R. Conti and W. Szymański, Automorphisms of the Cuntz algebras, Progress in operator
algebras, noncommutative geometry, and their applications, 1–15, Theta Ser. Adv. Math.,
15, Theta, Bucharest, 2012.
[14] R. Conti and W. Szymański, Labeled Trees and Localized Automorphisms of the Cuntz Algebras, Trans. Amer. Math. Soc. 363 (2011) no. 11, 5847–5870.
[15] R. Conti and W. Szymański, Unpublished.
[16] J. Cuntz, Automorphisms of certain simple C*-algebras, in Quantum fields-algebrasprocesses, ed. L. Streit, Springer, 1980.
31
[17] J. Cuntz, Simple C*-algebras generated by isometries, Commun. Math. Phys. 57 (1977),
173–185.
∗
[18] J. Cuntz, C -algebras associated with the ax + b-semigroup over N, K-theory and noncommutative geometry, 201215, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich, 2008.
∗
[19] K. Dykema and D. Shlyakhtenko, Exactness of Cuntz-Pimsner C -algebras, Proc. Edinburgh Math. Soc. (Series 2) 44 (2001), 425–444.
∗
[20] R. Exel, A. an Huef, and I. Raeburn, Purely infinite C -algebras associated to integer dilation matrices, Indiana Univ. Math. J. 60 (2011), 1033–1058.
[21] R. V. Kadison and I. M. Singer, Extensions of pure states, Amer. J. Math. 81 (1959), 383–400.
[22] M. Laca, Endomorphisms of B(H) and Cuntz algebras, J. Operator Theory 30 (1993), 85–
108.
[23] N. S. Larsen and X. Li, The 2-adic ring C*-algebra of the integers and its representations, J.
Funct. Anal. 262 (4) (2012), 1392–1426.
[24] K. Matsumoto and J. Tomiyama, Outer automorphisms on Cuntz algebras, Bull. London
Math. Soc. 25 (1) (1993), 64–66.
∗
[25] J. Phillips, Outer Automorphisms of separable C -algebras, J. Funct. Anal. 70 (1987), 111–
116.
∗
[26] J. Renault, Cartan subalgebras in C algebras, Irish Math. Soc. Bull. (2008), 61, 29–63.
[27] J. Tomiyama, On Some Types of Maximal Abelian Subalgebras, J. Funct. Anal. 10 (1972),
373–386.
∗
[28] J. Zacharias, Quasi-free automorphisms of Cuntz-Krieger-Pimsner algebras, C -algebras
(Münster, 1999), 262-272, Springer, Berlin, 2000.
32
| 4math.GR
|
Explicit Parallel-in-time Integration of a Linear Acoustic-Advection System
D. Ruprechta,∗, R. Krausea
a Institute
of Computational Science, Università della Svizzera italiana, Via Giuseppe Buffi 13, 6906 Lugano, Switzerland
arXiv:1510.02237v1 [cs.CE] 8 Oct 2015
Abstract
The applicability of the Parareal parallel-in-time integration scheme for the solution of a linear, two-dimensional
hyperbolic acoustic-advection system, which is often used as a test case for integration schemes for numerical weather
prediction (NWP), is addressed. Parallel-in-time schemes are a possible way to increase, on the algorithmic level, the
amount of parallelism, a requirement arising from the rapidly growing number of CPUs in high performance computer
systems. A recently introduced modification of the ”parallel implicit time-integration algorithm” could successfully
solve hyperbolic problems arising in structural dynamics. It has later been cast into the framework of Parareal. The
present paper adapts this modified Parareal and employs it for the solution of a hyperbolic flow problem, where the
initial value problem solved in parallel arises from the spatial discretization of a partial differential equation by a
finite difference method. It is demonstrated that the modified Parareal is stable and can produce reasonably accurate
solutions while allowing for a noticeable reduction of the time-to-solution. The implementation relies on integration
schemes already widely used in NWP (RK-3, partially split forward Euler, forward-backward). It is demonstrated that
using an explicit partially split scheme for the coarse integrator allows to avoid the use of an implicit scheme while
still achieving speedup.
Keywords: numerical weather prediction, parallel-in-time integration, parareal, Krylov-subspace-enhancement,
acoustic-advection system
1. Introduction
Numerical weather prediction (NWP) is a classical application area for high performance computing and the list of
the 500 most powerful supercomputers in the world1 contains a number of systems operated by weather services and
meteorological research centers. As the growth of the frequency of individual processors has stopped several years
ago because of fundamental physical problems, increases in performance of supercomputers are now achieved by a
rapidly growing number of nodes and cores. Contemporary massively parallel computers already feature up to several
hundred thousand CPUs. This change towards massively parallel systems requires software to feature increasing
levels of parallelism in order to run efficiently on computers with more and more CPUs, see for example [1, 2] for
discussions of this point in general and [3] for a recent assessment specifically for weather and climate simulations
and also [4] for a much earlier one.
Parallelization by domain decomposition is by now a standard technique employed in NWP codes like for example
COSMO2 or WRF3 . It relies on the decomposition of the computational domain into subdomains and the assignment
of the computations for a particular subdomain to one processor. While this approach can yield very good parallel
scaling to a large number of cores, it nevertheless saturates at some critical number of processors beyond which the
subdomains become too small. In view of the fact that NWP is a time critical application, additional directions of
parallelization are required in order to achieve a maximized reduction of the time-to-solution on upcoming computer
∗ Corresponding
author.
Email address: [email protected] (D. Ruprecht)
1 www.top500.org
2 www.cosmo-model.org
3
www.wrf-model.org
Preprint submitted to Computers & Fluids
October 9, 2015
systems with quickly growing numbers of cores. One possibility is to modify codes on the implementation level, as
done for example in [5], where parts of WRF are ported to graphic processing units (GPU) thereby achieving significant acceleration. Investigating numerical algorithms inherently suitable for parallel computation is another important
direction. Parallel-in-time integration schemes, allowing for a parallelization of the time-stepping procedure, are a
possible approach to increase parallelism on the algorithmic level.
The Parareal algorithm for parallel-in-time integration has been introduced in [6]. It relies on the introduction of
a computationally cheap, coarse integrator, which is used to produce guesses of the solution at several later points
in time. Starting from these guesses, an accurate but computationally expensive fine integrator is run concurrently.
The results are then used to propagate a correction to the guesses by again running the coarse integrator serially
over all intervals. This procedure is iterated and converges to the solution that would be obtained by running the
fine propagator sequentially. In [7] the algorithm is modified and used to solve a nonlinear parabolic PDE arising in
financial mathematics. Since then, Parareal has been applied successfully to a broad range of problems. In [8] it is
applied to problems arising in quantum chemistry, in [9, 10, 11] to the Navier-Stokes equations, in [12] it is used for
the Princeton ocean model while in [13] Parareal is employed to successfully speed up simulations of fully turbulent
plasma. Finally, a hybrid approach coupling Parareal and spectral deferred correction methods is presented in [14].
Besides several applications, a number of theoretical results have been published as well. In [15], a comprehensive
mathematical analysis is provided, including the interpretation of Parareal as a multiple-shooting as well as multigridin-time method. Also super-linear convergence on short and linear convergence on long time intervals is shown. A
super-linear convergence theorem for Parareal when applied to nonlinear systems of ODEs is proven in [16]. Convergence for different numerical examples is demonstrated, among others Lorenz equations and a system arising from
the discretization of the viscous Burgers equation. Also, NWP is mentioned as a possible application where Parareal
could be beneficial.
Inspired by the original Parareal, a parallel-in-time scheme called ”parallel implicit time-integration algorithm”
(PITA) is introduced in [17] and its applicability to three model problems from fluid, structure and fluid-structure
interaction applications is investigated. While PITA is found to work well for parabolic and first order hyperbolic
problems, stability issues for second order hyperbolic systems are identified. Similar instabilities of Parareal applied
to hyperbolic equations are indicated in [18]. Stability criteria for Parareal are derived in [19] and it is found that
Parareal is unstable for problems with pure imaginary eigenvalues or eigenvalues with a dominant imaginary part and
hence that the original version of Parareal is not applicable to flow problems that are strongly dominated by advection.
Further analysis of the problems arising from imaginary eigenvalues can be found in [15]. Finally, it is shown in
[20] that Parareal can be efficient for advection on bounded domains, where the solution is mainly determined by
the boundary values and not by the initial value. It is also shown that in general Parareal is inapplicable to even the
one dimensional version of the wave equation. These limitations most likely forbid using the original Parareal to
solve the equations arising in NWP. Indeed, it is confirmed in the present paper that the original Parareal can develop
instabilities for the two-dimensional advection as well as the two-dimensional acoustic-advection problem, although
the sequentially run coarse and fine integrator both remain stable.
However, a modification of PITA is introduced in [21] that is demonstrated to work for second-order hyperbolic
problems arising in linear structural dynamics. It is extended and applied to nonlinear problems in [22]. In [23] it
is shown that for linear problems PITA and Parareal are equivalent, the new version of PITA is cast into the Parareal
framework and demonstrated to successfully solve second order ODEs. Also the name ”Krylov-subspace-enhanced
Parareal” is coined, which is adopted in the present paper and will be abbreviated KSE-Parareal from now on. To the
authors’ knowledge, no works exist that address the feasibility of KSE-Parareal for hyperbolic flow problems. Also,
the results obtained for the examples from structural dynamics relied on the use of an implicit coarse propagator in
order to allow for large coarse time steps. This poses a problem when considering a possible application of KSEParareal for NWP: While there are codes where implicit solvers are used to model atmospheric flows, see for example
[24], most NWP codes rely on explicit integration schemes. Adding an efficient nonlinear solver into such an existing
code would be at best a challenging endeavor.
As a first step to investigate the feasibility of using KSE-Parareal for NWP, the present paper investigates the
applicability to a linear two-dimensional acoustic-advection system. This system is often used as a simplified test
case for feasibility or stability studies of integration schemes for NWP applications, see for example [25, 26, 27,
28]. It is demonstrated that for this type of problem an explicit partially split scheme (see [25, 26, 29]) with strong
divergence damping can be employed as coarse propagator, avoiding the necessity for the implementation of an
2
implicit scheme. Partially split methods are widely used in codes for numerical weather prediction and a possible
implementation of parallel-in-time integration could rely on these already implemented schemes (compare for the
comments in subsection 2.1). Both efficiency and accuracy of the parallel-in-time scheme for the investigated example
problem are discussed. A comparison with the speedup obtained by switching to a partially split scheme is given and
it is found that KSE-Parareal yields larger speedup while maintaining comparable accuracy.
The present paper is motivated by addressing the potential feasibility of applying KSE-Parareal in codes for NWP
and the choice of parameters and methods is inspired by this goal. However, the investigated test problem, as a
reduced model emerging from the Euler equations, is of potential interest to a much broader field of applications. It
contains the two-dimensional wave equation as a special case, which is a model arising in numerous applications. It
is also mathematically equivalent to the linearized shallow water equations, which are another popular reduced model
in atmospheric as well as in ocean sciences.
2. The Parareal Algorithm
This section recapitulates the Parareal parallel-in-time algorithm, first in subsection 2.1 in its original version and
then in subsection 2.2 in the modified version, suitable for hyperbolic systems.
2.1. Original Version
The presentation of Parareal in this section follows the formulation of the algorithm as a predictor-corrector
method in [15]. The aim is to solve an initial value problem of the type
qt = f(q),
q(0) = q0 ∈ Rd ,
t ∈ [0, T ]
(1)
in parallel. In the following, f stems from the spatial discretization of a partial differential equation (”method-of-lines”)
and d is the total number of degrees-of-freedom. Let Fδt denote some integration scheme of appropriate accuracy,
using a constant time step δt. In order to employ Parareal, a second integration scheme G∆t is required, which has
to be much cheaper than Fδt in terms of computation time but which can also be much less accurate. Usually, G∆t
will employ a larger time step, that is ∆t ≫ δt, and be of lower order, so that less evaluations of the right hand side
f are required. Further, G∆t can also feature a less accurate spatial discretization, using lower order spatial operators
and/or a coarser spatial mesh (see for example [11]), leading to a different right hand side function fc , f for the case
of a semi-discrete PDE. As demonstrated in [30], the coarse propagator can even be based on simplified or averaged
model physics. Following common terminology, Fδt will be referred to as the fine propagator and G∆t as the coarse
propagator.
For a given ∆t assume that the time interval [0, T ] can be decomposed into Mc coarse intervals with endpoints
0 = t0 < t1 < . . . < t Mc = T
(2)
and constant lengths
∆t = ti+1 − ti ,
i = 0, . . . , Mc − 1.
(3)
Further let ∆t be a multiple of the fine time step δt, that is there exists an integer Nf ∈ N such that
∆t = Nf δt
(4)
and every coarse interval contains Nf fine intervals, hence
T = Mc ∆t = (Mc Nf ) δt.
(5)
Finally assume that Mc is a multiple of the number of available processors Np so that integration from t = 0 to t = T
can be done in Mp := Mc /Np parallel steps of length Np ∆t, each step distributing Nc := Mc /Mp = Np coarse intervals
onto Np processors. Note that in principal one can also assign multiple coarse intervals to a single processor, that is
let Nc be a multiple of Np , in order to perform larger parallel steps but below it is always assumed that Nc = Np . The
number of fine steps per parallel step is equal to Nt := Nc Nf .
3
Algorithm 1 Original version of Parareal
1: {Initialization:}
2: q00 = q0
3: for i = 0 to Nc − 1 do
4:
q0i+1 = G∆t (q0i , ti+1 , ti )
5: end for
6: {Iteration:}
7: k := 0
8: repeat
9:
{Parallel predictor step:}
10:
for i = 0 to Nc − 1 do
11:
q̃ki+1 = Fδt (qki , ti+1 , ti )
12:
end for
13:
{Sequential correction step:}
14:
for i = 0 to Nc − 1 do
k+1
15:
qi+1
= G∆t (qik+1 , ti+1 , ti ) + q̃ki+1 − G∆t (qki , ti+1 , ti )
16:
end for
17:
k := k + 1
18: until k = Nmaxit
Denote by
Fδt (q, t1 , t0 ),
G∆t (q, t1 , t0 )
(6)
the result of the fine or coarse propagator if used to integrate (1) from an initial value q at time t0 to a time t1 > t0
using time steps δt or ∆t respectively. Let a subscript n denote the approximation of q at time tn , that is
qn ≈ q(tn )
(7)
and let a superscript k denote quantities in the k-th iteration of Parareal. With this notation, the basic iteration of
Parareal performed in one parallel step reads
k+1
=G∆t qnk+1 , tn+1 , tn + Fδt qkn , tn+1 , tn
qn+1
− G∆t qkn , tn+1 , tn ,
(8)
compare for [15]. For k → Nc , the iteration converges to a a solution qn , n = 0, . . . , Nc , satisfying
qn+1 = Fδt (qn , tn+1 , tn ) ,
(9)
that is to an approximation of the exact solution with the accuracy of the fine propagator. After Nc iterations, the
algorithm will always provide the sequential solution, but in order to be efficient, it has to converge after Nit ≪ Nc
iterations. Note that as the values qkn in (8) are given from either the initialization or the previous iteration, the
expensive computations of the values Fδt (qki , ti+1 , ti ) in (8) are all independent and can be performed concurrently.
Algorithm 1 sketches the complete iteration for one parallel step in pseudo code. For simplified notation, the endpoints of the involved Nc coarse intervals have been re-indexed to t0 , . . . , tNc . If implementations of the propagators Fδt
and G∆t are available, implementing Parareal does not require much more than implementing this iteration, although
achieving efficiency might require modifications of the implementation of Fδt and G∆t . Parallelization of the loop
computing the values of the fine propagator can easily be done, for example, by adding the corresponding OpenMP
directives4. This makes Parareal a good candidate for a hybrid spatial/temporal parallelization approach, where spatial
4 www.openmp.org
4
subdomains are assigned to computational nodes containing several processors that share the same memory. Communication between subdomains would then be implemented by message-passing between nodes instead of cores while
integration inside a subdomain is performed in parallel employing the cores available inside the node. It is stressed,
however, that the algorithm is inherently parallel by design and not limited to a specific model of parallelization. A
combination of Parareal with domain decomposition is explored in [31]. Also note that only the loop in the lines
10–12 of Algorithm 1 has to be parallelized, so over the rest of the computation the idle CPUs could be assigned to
other tasks.
Using a stopping criterion in 1 instead of a fixed number of iterations is possible as well by defining an error
estimate of the form
rk = max kqik+1 − qki k.
(10)
i=1,...,Nc
The iteration would then be terminated if either k reaches the maximum number of iterations allowed or if the error
estimate drops below some prescribed threshold. Throughout this paper, to allow for easier comparison of results, a
fixed number of iterations is carried out without considering an error estimate.
2.2. Krylov-subspace-enhanced Parareal
While the original version of Parareal presented in subsection 2.1 has been used successfully for a large number
of different types of problems, it was found for example in [15, 17, 19] that instabilities can arise if it is applied to
hyperbolic problems. This subsection presents the modified version of PITA/Parareal introduced in [21, 22] which
proved to be stable when used to integrate hyperbolic problems in structural dynamics and which is applied to a
hyperbolic flow problem in the present paper. The presentation given here is different from the original one in [21, 22]
and follows again the interpretation of the algorithm in the Parareal framework in [23].
In the original version of Parareal sketched in Algorithm 1, the results q̃ni+1 from the fine propagator computed in
line 11 are used in the correction step in line 15 and then thrown away. The idea of the modified version is to keep
this information in order to successively enhance the coarse propagator by reusing values from the fine integrator
computed in previous iterations. Define by
n
o
Sk := span qij : i = 0, . . . , Nc − 1, j = 0, . . . k
(11)
the subspace spanned by all values on the coarse mesh from previous iterations. Further, denote by F (Sk ) the space
spanned by the results obtained by applying the fine propagator to these values, that is by
j
q̃i+1
= Fδt (qij , ti+1 , ti ), i = 0, . . . , Nc − 1, j = 0, . . . , k.
(12)
Finally, let Pk be the orthogonal projection onto Sk with respect to the standard Euclidean scalar product. The coarse
propagator is now replaced by
K∆t (q, ti+1 , ti ) := G∆t ((I − Pk )q, ti+1 , ti ) + Fδt (Pk q, ti+1 , ti )
(13)
resulting in Algorithm 2. By definition Pk q ∈ Sk , so letting s1 , . . . , sr denote an orthogonal basis of Sk , for a linear
problem Fδt (Pk q, ti+1 , ti ) can be computed from
r
X
k
Fδt (P q, ti+1 , ti ) = Fδt αi si , ti+1 , ti
i=1
=
r
X
(14)
αi Fδt (si , ti+1 , ti )
i=1
where the Fδt (si , ti+1 , ti ) are known from F (Sk ) and are thus available without running the fine propagator again. The
information generated in previous iterations successively increases the accuracy of the coarse propagator K∆t . As the
dimension of Sk increases, formally Pk → I and K(q) → F (q), see [23]. It is also shown there that the method
converges once the dimension of Sk is no longer increasing. For the nonlinear case the update step becomes more
complex, see [22] for the nonlinear version of PITA, and no formulation in the Parareal framework is available, yet.
5
Algorithm 2 Krylov-subspace-enhanced Parareal
1: {Initialization:}
2: y00 = q0 , S 0 = F (S 0 ) = ∅
3: for i = 0 to Nc−1 do
4:
q0i+1 = G∆t (q0i , ti+1 , ti )
5: end for
6: {Iteration:}
7: k := 0
8: repeat
9:
{Parallel predictor step:}
10:
for i = 0 to Nc − 1 do
11:
q̃ki+1 = Fδt (qki , ti+1 , ti )
12:
end for
13:
{Sequential correction step:}
14:
Update S k−1 → S k , F (S k ) and Pk using
qki−1 , q̃ki , i = 1, . . . , Nc
15:
for i = 0 to Nc − 1 do
k+1
16:
qi+1
= K∆t (qik+1 , ti+1 , ti ) + q̃ki+1 − K∆t (qki , ti+1 , ti )
17:
end for
18:
k := k + 1
19: until k = Nmaxit
Additionally, the nonlinear version relies on using an implicit scheme for G∆t and the adaption for an explicit scheme
is not straightforward.
In order to perform the update step Sk−1 → Sk in line 14 of Algorithm 2, an orthogonal basis of the space
n
o
Sk−1 ∪ qk0 , . . . , qkNc −1
(15)
has to be computed. In principle this can be done by using the Gram-Schmidt algorithm, keeping the basis of Sk−1
while successively adding new orthogonal basis vectors. However, the Gram-Schmidt algorithm is known to be
unstable, see [32], and applying it in the context of Parareal does indeed require frequent re-orthogonalization5.
Hence in the present paper the update step is performed by computing a full QR decomposition in every step using
the LAPACK routine ”DGEQP3” 6 . There are highly optimized implementations of this library available for basically
every computer architecture and the examples presented in section 3 show that the run time spent in the subspace
update is small compared to the time spent for the coarse and fine integrator (see especially figure 10). Nevertheless,
the subspace update does contribute to the sequential part of the algorithm, thus further restricting scalability. Also
note that the update of the subspace in line 14 introduces a ”synchronization point”, so that overlapping computation
as used in [14, 33] for the original version is not possible here.
2.3. Expected Speedup
The ratio of the sequential to the parallel execution time for a given number of processors is referred to as speedup.
Denote by τc the time required to complete one step of length ∆t of the coarse propagator, by τf the time for one step
of length δt of the fine propagator and by τQR (k) the time required to compute the subspace update in iteration k.
Recall that Nt = Nc Nf is the total number of δt-time steps in one parallel step, Nc the total number of ∆t-steps (as
defined in subsection 2.1) and Nit the number of iterations. As the dimension of Sk increases monotonically with the
number of iterations performed 7 , it is
τQR (k) ≤ τQR (Nit ).
(16)
5 Martin
J. Gander and Stefan Güttel, personal communication
6 www.netlib.org/lapack
7 Note that the QR decomposition is located in the sequential part of the code, so its run time could probably be reduced by using multithreaded
libraries.
6
The speedup obtainable in one parallel step by Algorithm 2 using Np processors can then be estimated by
Nt τf
s(Np ) ≈
Nc τc + Nit Nc τc +
=
(1 + Nit )
N
Nt
Np τf
1
c τc
Nt τ f
+
+ Nit τQR (Nit )
Nit τQR (Nit )
Nt τ f
+
Nit
Np
.
(17)
See also [14]. From (17), three different upper bounds for s(Np ) can be derived. First, it is
s(Np ) ≤
Np
Nit
(18)
and hence the efficiency of the algorithm depends critically on convergence within few iterations. Also note that according to (18) the parallel efficiency is bounded by 1/Nit , so that a perfect speedup is impossible by design. As spatial
parallelization by domain decomposition can scale almost perfectly until some critical number of processors where
communication becomes dominant, parallel-in-time integration should be considered as a possibility for additional
fine grain parallelism on top of an existing coarse grain domain decomposition, to be used if more processors are
available after the spatial parallelization has saturated.
A second bound for s emerging from (17) is
s(Np ) ≤
Nf τf
Nf τf 1
≤
.
Nc τc 1 + Nit
Nc τc
(19)
Hence running the coarse propagator once over all coarse intervals has to be sufficiently cheaper in terms of computation time than running the fine propagator. A reduction of the run time of the coarse propagator can be achieved either
by a larger coarse time step ∆t, leading to fewer coarse steps and improving the ratio Nf /Nc and/or by reducing the
run time for a single coarse step by using a lower order scheme, lower order spatial operators, etc. and thus improving
τf /τc . However, note that the bounds (18) and (19) are competing in the sense that a cheaper, less accurate coarse
propagator will improve the bound (19) but might require more iterations, thus degrading (18). For the algorithm to
be efficient, a reasonable balance has to be found.
The QR decomposition required in the subspace update in KSE-Parareal introduces another bound, namely
s(Np ) ≤
Nf τf
.
Nit τQR (Nit )
(20)
The computational complexity of the QR decomposition of a n × m rectangular matrix with m ≫ n is O(n2 m), see
[34]. In Algorithm 2, the number of columns is equal to the dimension of Sk while the number of rows is equal to
the number of degrees-of-freedom d. The former is bounded by Nit Nc , as in each iteration the dimension of Sk can
increase by a maximum of Nc if all added values qki , i = 0, . . . , Nc − 1, turn out to be linearly independent, so typically
d ≫ dim(Sk ) and thus
τQR (Nit ) ∼ dNc2 Nit2 .
(21)
Evaluating f requires at least some computations for every degree of freedom, so that an efficient implementation
should yield τf = O(d). Hence
Nf d
Nf τf
Nf 1
,
(22)
∼
∼
Nit τQR (Nit ) Nit dNit2 Nc2
Nc Nc Nit3
so that the computational cost for the subspace update should be independent of the problem size but will increase
rapidly with the number of performed iterations. Also (22) results in a stricter bound for the speedup as Nc increases,
even if Nf /Nc remains constant, so performing multiple parallel steps with one coarse interval per processor will yield
better speedup than fewer parallel steps where each processor handles multiple coarse intervals.
Assuming that both the coarse and the fine propagator use the same spatial mesh, (19) can be expressed in terms
of the CFL numbers of the two propagators, that is
Cf =
cδt
,
δx
Cc =
7
c∆t
,
δx
(23)
where c is the fastest wave speed arising in a considered problem. For the acoustic-advection system in section 3, c is
the speed of sound. Using (23) in (17) and neglecting the cost of the subspace update allows to derive the estimate
s(Np ) ≤
1
(1 + Nit ) CCcf
τc
τf
+
Np
Nit
,
(24)
and it is shown in Section 3 that (24) gives a reasonable estimate of the actually achieved speedups in the investigated
example when using an empirically computed ratio τc /τf . Note that if a coarsened spatial mesh would be used for G,
an additional factor corresponding to the ratio of the cell sizes in the coarse and fine mesh would appear in front of
Cf /Cc in (24).
3. Numerical Results
In this section, first the instability of the original Parareal and the stability of KSE-Parareal for the two-dimensional
advection equation
qt + U · ∇q = 0
(25)
is demonstrated. Then the performance of KSE-Parareal is addressed in detail for the linearized acoustic-advection
system
ut + U · ∇u + cs π x = 0
vt + U · ∇v + cs πy = 0
πt + U · ∇π + cs u x + vy = 0,
(26)
where cs denotes the speed of sound and U = (U, V) a constant-in-time advection velocity. The unknowns are the
perturbation velocity fields u = (u, v) and the perturbation pressure π. The system (26) contains two out of three
major processes that govern the stability of integration schemes used for atmospheric flows, namely advection and
acoustic waves. It does not include internal gravity waves whose propagation velocity ranges somewhere between
typical advection speeds and the speed of sound, depending on their wave length. Note that (26) is mathematically
equivalent to the linearized shallow water equations.
The employed spatial discretization is a finite difference scheme in conservation form on equally sized rectangular
cells. Let qi denote a cell centered value of some quantity for a cell of size ∆x × ∆y with index (i, j). Then the rate of
change at a given time is
Fi+1/2, j − Fi−1/2, j Gi, j+1/2 − Gi, j−1/2
−
,
(27)
∂t qi, j = −
∆x
∆y
where Fi+1/2, j , Fi−1/2, j are the fluxes across the interfaces in x-direction and Gi, j+1/2 , Gi, j−1/2 the fluxes across the
interfaces in y-direction. The fluxes are evaluated according to the stencils of order one to six given in [26, 35].
Applying the spatial discretization to a given PDE results in an ODE system of the form (1) which can then be solved
by standard integration schemes. The acoustic fluxes in (26) are always computed with a second order centered stencil
in both the coarse and fine propagator while the order of the advective fluxes ranges between a sixth order accurate
centered stencil to a first order upwind flux.
3.1. Instability of original Parareal
For a brief illustration of the instability that can arise for the original version of Parareal, consider the simple
two-dimensional advection problem (25) on the unit square [0, 1] × [0, 1], discretized with 40 × 40 cells of equal size
and with periodic boundary conditions. The advection velocity is set to
U = (1, 1),
(28)
and the simulation is run from T = 0 to T = 1. As initial value for u, a distribution also used for example in [35] as a
test of two-dimensional advection is employed,
u0 (x, y) =
1
cos (πr(x, y)) + 1
2
8
(29)
Figure 1: kqk∞ over time for the coarse (RK-3, first order upwind flux, Cc = 0.6) and fine propagator (RK-3, sixth order centered flux, Cf = 0.1) run
sequentially as well as the original Parareal (upper) and KSE-Parareal (lower) for pure advection. The lines corresponding to the fine propagator
and the parallel solution in the lower figure do essentially coincide.
with
and
r
2
2
(y
−
y
)
(x
−
x
)
0
0
+
r(x, y) = min 1, 4
2
2
0.5
0.5
x0 = y0 = 0.5.
(30)
(31)
The fine propagator is a Runge-Kutta-3 scheme with Cf = 0.1 and sixth order fluxes, the coarse propagator is a
Runge-Kutta-3 scheme with Cc = 0.6 and first order upwind flux. The parallel algorithm uses Np = 6 coarse intervals
but is run sequentially for testing purposes. Nit = 5 iterations are performed in every parallel step. Figure 1 shows
kqk∞ over time for the parallel integration scheme as well as for the coarse and fine propagator run sequentially from
T = 0 to T = 1. Both the fine and the coarse propagator are stable, the coarse propagator being strongly diffusive.
The solution computed by the original Parareal clearly becomes unstable at about T = 0.2 (upper figure). Switching
to the Krylov-subspace-enhanced version (lower figure) results again in a stable scheme.
Figure 2: kqk∞ over time for the coarse (RK-3, first order upwind flux, Cc = 0.6) and fine propagator (RK-3, sixth order centered flux, Cf = 0.1)
run sequentially as well as the original Parareal (upper) and KSE-Parareal (lower) for the acoustic-advection system. The lines corresponding to
the fine propagator and the parallel solution in the lower figure do essentially coincide.
The instability can also be demonstrated to arise for the acoustic-advection system (26). Again, use a RungeKutta-3 scheme with Cf = 0.1 and sixth order fluxes for the advective terms for F and a Runge-Kutta-3 scheme with
Cc = 0.6 and first order upwind advective fluxes for G. Both schemes use second order centered fluxes for the acoustic
terms. The sound speed is set to cs = 1.0 for this example and (29) is used as initial value for u while v and π are
set to zero at T = 0. All other parameters remain as in the pure advection example above. Figure 2 again shows the
maximum absolute value for the solution vector q, containing all the components, over time. Clearly, the original
Parareal is again unstable while KSE-Parareal is not.
Further tests not documented here indicate, however, that the occurrence of the instability does depend on different
factors, for example the actual schemes used for the coarse and fine propagators as well as the order of the underlying
spatial discretization. The stability of Parareal applied to semi-discrete PDEs is apparently more involved. Some
9
analytic results concerning the stability of Parareal for PDEs can be found in [18], but for the present paper the
example suffices to demonstrate the inapplicability of the original Parareal for the considered problem, confirming the
requirement to use the more complex KSE-Parareal variant.
3.2. KSE-Parareal for Acoustic-Advection
In the following, the accuracy and efficiency of KSE-Parareal for the integration of the acoustic-advection system
(26) is investigated. For the coarse propagator G a partially split forward Euler scheme is used, with a forwardbackward scheme for the acoustic steps. Partially split schemes rely on the separation of the modes into fast and slow,
in (26) the fast modes are sound waves and the slow mode is advection. For every update of the slow tendencies, Nsound
small time steps for the fast tendencies are performed. These schemes allow for larger time steps in the slow modes,
than the stability limit enforced by the meteorologically unimportant acoustic waves permits in non-split schemes.
See [25, 26, 29, 35] for details on partially split schemes. While the partially split forward Euler is found in [29] to
be unstable in the semi-discrete case, it can be stabilized by using diffusive low order spatial operators and strong
divergence damping, see [28, 36]. Divergence damping, introduced in [25], is generally required to stabilize partially
split schemes, see [28]. Although the strong dissipation generated by G would be problematic when using it as a
stand-alone method, it turns out that Parareal is quite effective in correcting for the too strong damping of the coarse
propagator, which is consistent with its reported good performance for diffusive problems. The momentum equations
are modified to
ut + U · ∇u + cs π x = α1 (∇ · u) x
vt + U · ∇v + cs πy = α2 (∇ · u)y ,
(32)
with u = (u, v) and mesh and time step dependent damping coefficients
α1 = νc/f
δx2
δy2
, α2 = νc/f
,
τ
τ
(33)
with τ = ∆t in the coarse and τ = δt in the fine propagator. In the coarse propagator, when a partially split scheme is
used, the coefficients become
δy2
δx2
, α2 = νc
(34)
α1 = νc
∆t/Nsound
∆t/Nsound
as the divergence damping is employed in every acoustic step8 . For the fine propagator F , a non-split Runge-Kutta-3
scheme is used. While no divergence damping is required to stabilize this scheme, it is found that when using a
partially split scheme for G, a small amount of divergence damping in F is required in order for the Parareal iteration
to converge reasonably fast. A value of νf = 0.005 is used throughout this paper. The advection velocity is set to
U = γ · (y − 0.5, −(x − 0.5)) ,
γ = π,
(35)
corresponding to a solid body rotation completing one full revolution in a non-dimensional time of T = 2. The computational domain is the unit square [0, 1] × [0, 1], resulting in a maximum value of U of about 1.5. The sound speed is
set to cs = 30, thereby roughly reproducing the typical ratio between advection and sound speed in atmospheric flows
of about 10 m s−1 to 300 m s−1 .
Parallelization in the employed implementation is done by using OpenMP directives. All computations are performed using a computer with two sockets and one quad-core Intel Xeon processor with a frequency of 2.26 GHz
per socket. Each socket has 8 GB attached and access to the memory of the other socket is slower, hence the system
features ”non-uniform memory access” (NUMA). Careful use of ”first touch policy” is required to ensure efficiency
of the multithreaded implementation, see [37].
8 Note that the use of a partially split scheme for G introduces a third time step size ∆t/N
sound besides the two time steps of the fine and coarse
propagator. No symbol is introduced for it in order to keep notation simple. The reader is cautioned, however, that ∆t and δt refer to the time steps
of G and F and not to the acoustic step size in the split scheme.
10
The initial value for u is again (29) but with y0 = 0.65, that is, the maximum is not located in the center of the
domain. The y-component of the velocity and the pressure π are set to zero initially. This setup is run with from T = 0
to T = 2 with fixed fine CFL number of Cf = 0.2 and different values for Cc , Nsound , Np , Nit and νf , νc and results are
presented and analyzed below.
3.2.1. Vorticity
Figure 3: Vorticity of the resulting velocity field at T = 2 after one complete revolution. The spatial resolution is 40 × 40 cells, the difference
between isolines is 2 and gray isolines correspond to negative values of vorticity. Shown are the sequentially computed reference solution (upper
left) and the parallel solution for Nit = 1 (upper right), Nit = 2 (lower left) and Nit = 3 (lower right). For all parallel solutions, a value of Np = 6
was used. The CFL number for the coarse propagator is Cc = 4.0 with Nsound = 4 acoustic steps per time step. The divergence damping for the fine
propagator is νf = 0.005 and for the coarse propagator νc = 0.1. Corresponding run times can be found in Table 2.
The two-dimensional vorticity is defined as
ω := uy − v x .
(36)
Cross differentation of (26) and using the special form of the advection velocity (35) yields
ωt + U · ∇ω = ∇ · u.
11
(37)
The divergence at the right hand side is associated with acoustic modes and turns out to be small in the present
example, so that at leading order the vorticity is simply advected by the flow U and has completed one full revolution
at T = 2.
Figure 3 shows the vorticity of the obtained velocity fields at the end of the simulation for a sequentially computed
reference solution and a parallel solution computed with different number of iterations per parallel time step. Even
with just one iteration per parallel step, the parallel integration scheme captures the advected vorticity quite well,
although some damping of the negative extremum results in a slightly asymmetric distribution. For the parallel
solutions with Nit = 2 and Nit = 3, the resulting vorticity fields are indistinguishable from the reference solution.
Hence KSE-Parareal not only remains stable but also does a good job in capturing the essential slow dynamics of
the simulated problem, that is the advection of vorticity. Corresponding run times can be found in Tables 1 and 2.
Naturally, the solution with Nit = 1 yields the largest speedup, but as seen introduces noticeable errors in the vorticity
field. Parallel solutions with two or three iterations, both showing very good agreement of the resulting vorticity fields
with the reference solution, still allow for a reduction of the time-to-solution by a factor of 1.9 and 1.3 respectively.
3.2.2. Time Series
Figure 4 shows plots of the horizontal velocity of the parallel computed solution for Nit = 1 and Nit = 2 at
(x, y) = (0.49, 0.34), that is, the point opposite to the center of the initial distribution. The agreement in the horizontal
velocity is very good, even if only one iteration per parallel step is used. Some small differences are visible in
the minimum passing through the point at T = 1. For Nit = 2, the parallel and sequential solution are basically
indistinguishable. The representation of the pressure π, shown in figure 5 is still surprisingly good, given that the
partially split scheme employed G does by design not represent sound waves very accurately. Increasing the number
of iterations provides a more accurate resolution of the acoustic modes, but also severely inhibits the efficiency of the
parallel scheme. For a highly accurate representation of the acoustic modes with a small number of iterations, a nonsplit scheme for G is most likely required, which would however reduce the efficiency of the time-parallel scheme.
Figure 4: Horizontal velocity u at (x, y) = (0.49, 0.34) over time, plotted after each completed parallel step. The solid line represents the parallel
solution after one (upper) or two (lower) iteration while the dashed, gray line represents the sequentially computed reference solution. The
simulation parameters are identical to the ones given for Figure 3.
3.2.3. Energy
The system (26) allows for the definition of the energy density
E = u 2 + v2 + π 2 .
(38)
By basic manipulations of (26) one can derive
Et + U · ∇E + cs ∇ · E = 0.
(39)
Integrating over the domain, using the divergence theorem as well as the assumed periodicity of the boundary values
yields that the total energy
Z
E dx
Etot =
[0,1]2
12
(40)
Figure 5: Pressure π at (x, y) = (0.49, 0.34) over time, plotted after each completed parallel step. The solid line represents the parallel solution after
one (left) or two (right) iteration while the dashed, gray line represents the sequentially computed reference solution. The simulation parameters
are identical to the ones given for Figure 3.
is conserved in the original system. However, because of the introduced divergence damping, this is no longer true
for the numerical solution. Figure 6 shows a comparison of the evolution of the total energy for the parallel and the
sequential solution. Even for the sequential solution, the divergence damping results in a significant reduction of the
total energy, although the damping parameter νf = 0.005 is quite small. The parallel solution computed with a single
iteration per parallel step is slightly more diffusive than the sequentially computed, as apparently one iteration is not
enough to correct for the strong diffusivity of the coarse propagator. Performing two iterations per parallel step yields
an energy evolution that is indistinguishable from the one of the sequential solution.
In order to assess how the parallel integrator performs in terms of energy conservation if no divergence damping
is present in the fine integrator and a much less diffusive G is used, an additional simulation is run with νf = 0. Also,
the divergence damping in the coarse propagator is reduced to a value of νc = 0.005, the CFL number reduced to
Cc = 2 and the first order advective flux is replaced by a less diffusive third order flux. To stabilize the coarse scheme,
an increase of the number of acoustic steps to Nsound = 8 is necessary. Figure 7 shows the total energy over time for a
mesh of 40 × 40 cells and also for the solution computed on refined mesh featuring 80 × 80 cells. The energy defect
at the end of the simulation is now down to 3 % on the coarse and 2 % on the fine mesh. Hence using a less diffusive
G leads, as can be expected, to a less diffusive parallel solution. However, this setup yields only small speedups and
if better conservation properties are desired, using a different scheme for G is most likely in order.
Figure 6: Total energy over time for the sequential solution and the parallel solution with Nit = 1 (upper) and Nit = 2 (lower). The parameters
of the simulation are identical to the ones in Figure 3. Note the divergence damping removes energy associated with acoustic waves, so the total
energy over time decreases.
3.2.4. Runtimes and Speedup
All run times indicated below are measured using the OpenMP function OMP GET WTIME and do not include
I/O operations. No averaging over an ensemble of different runs was performed, but the run times proved to be
very stable with variations only on the order of a few percent over different runs, so that the indicated values are
representative.
13
Figure 7: Total energy over time on a mesh with 40 × 40 cells (upper) and 80 × 80 cells (lower). There is no divergence damping for the fine
propagator, that is νf = 0, and the divergence damping for the coarse propagator is reduced to νc = 0.005. Also, coarse CFL number is reduced
to Cc = 2.0 and the massively diffusive first order upwind flux for the coarse propagator has been replaced by a somewhat less diffusive third
order flux. To maintain stability, the number of acoustic steps in the coarse scheme has been increased to Nsound = 8. In both cases, the parallel
computation used Np = 6.
Tables 1 and 2 show errors, run times, speedup and parallel efficiency for parallel solutions computed for coarse
CFL numbers of Cc = 2 and Cc = 4 with different numbers of iterations using Np = 4, 6, 8 processors. The error
indicates the relative l2 -norm of the difference between the solution qseq obtained by running Fδt sequentially and the
solution of the parallel scheme qpar , that is
kqpar − qseq k2
.
(41)
ǫ :=
kqseq k2
Tables 1, 2 show ǫ evaluated at the end of the simulation while Figure 8 shows ǫ evaluated after every parallel step.
Speedup is the ratio of the run time of the reference solution to the run time of the parallel solution and efficiency is
the speedup divided by the number of used processors. In all runs, the number of coarse intervals per parallel step is
equal to the number of processors so that runs with smaller values of Np perform more but shorter parallel steps while
those with larger values of Np perform fewer larger steps.
In general, every iteration reduces the error by about half an order of magnitude, exceptions being the cases in
Table 2 for Np = 4, Nit = 2 and Np = 8, Nit = 4 where only a marginal reduction of the error is observed. As
expected from the theoretical analysis in subsection 2.3, achieved speedups depend very strongly on the number
of iterations performed and noticeable acceleration requires the algorithm to terminate after two or at most three
iterations. Comparing the values in the ”error” as well as the ”speedup” columns in 1 and 2 reveals that a larger value
of Cc noticeably improves the obtained speedup at the expense of a less accurate solution.
Note, however, that there is no real benefit from approximating the sequential solution with an accuracy that is
higher than the discretization error of the sequential solution. In order to estimate the temporal discretization error,
the fine propagator is run with a ten times smaller time step and the difference between Fδt and Fδt/10 is used as
an estimate for the error of the temporal discretization. As the divergence damping inhibits proper convergence of
the RK-3 scheme, this estimate is generated by running both F with νf = 0, so that a fair comparison is obtained.
Figure 8 shows the difference measured according to (41) between the two sequential solutions, Fδt and Fδ/10 , without
damping as well as the difference between the parallel and the Fδt -sequential solution with νf = 0.005 for coarse
CFL numbers of Cc = 2.0 (upper) and Cc = 4.0 (lower). For Nit = 2, at the end of the simulation the estimated
discretization error is about equal to the iteration error of the parallel solution for Cc = 4.0 and noticeably smaller for
Cc = 2.0. Tables 1 and 2 indicate that for Nit = 2, depending on the number of processors, speedups ranging from
1.2 to 2.1 can be obtained. There is an interesting possibility of coupling Parareal with adaptive time step refinement
here: As an adaptive scheme has to provide estimates of the discretization error anyhow, this estimate could be reused
to adaptively modify the prescribed error tolerance for the Parareal iteration if a stopping criterion is used instead of a
fixed iteration number.
A more complete study of the obtained speedup is found in Figure 9. It shows the achieved speedup for a number
of processors ranging between one and eight for parallel solutions performing Nit = 1, 2, 3, 4 iterations. The upper
figure shows simulations on a grid consisting of 40×40 cells, the lower figure on a 80×80 cell grid. Again, the speedup
depends critically on the number of iterations and the simulations with four iterations produce nearly no speedup at
14
#it.
1
2
Error
1.3 × 10−1
2.8 × 10−2
Np = 6, Cc = 2.0
#it.
Error
Runtime Speedup
Efficiency
1
1.4 × 10−1
6.3 s
2.4
50 %
2
5.0 × 10−2
10.9 s
1.4
29 %
3
1.1 × 10−2
15.5 s
1.0
Np = 8, Cc = 2.0
Error
Runtime Speedup Efficiency
1.6 × 10−1
5.8 s
2.6
32 %
6.7 × 10−2
10.1 s
1.5
19 %
1.7 × 10−2
14.7 s
1.0
–
3.2 × 10−3
19.6 s
0.8
–
Np = 4, Cc = 2.0
Runtime Speedup
7.4 s
2.0
13.0 s
1.2
#it.
1
2
3
4
Efficiency
40 %
23 %
–%
Table 1: Results for a mesh using 40 × 40 cells, Cfine = 0.2, Nsound = 4 and Ccoarse = 2.0 for Np = 4 (upper), Np = 6 (middle) and Np = 8
(lower) coarse intervals per parallel time step distributed on an equal amount of processors. The divergence damping for the fine propagator is set
to νf = 0.005 while the coarse propagator used νc = 0.1. The column ”Error” indicates the difference (41) between the parallel solution and the
sequentially computed reference solution in all three fields, u, v and π, at T = 2 after one complete revolution. The run time for the sequential
integration is 14.9 s.
#it.
1
2
Error
1.8 × 10−1
1.7 × 10−1
Np = 6, Cc = 4.0
#it.
Error
Runtime Speedup
Efficiency
1
1.9 × 10−1
4.5 s
3.4
68 %
2
8.6 × 10−2
8.0 s
1.9
37 %
3
3.7 × 10−2
11.6 s
1.3
Np = 8, Cc = 4.0
Error
Runtime Speedup Efficiency
2.0 × 10−1
3.9 s
3.9
48 %
9.3 × 10−2
7.9 s
2.1
27 %
3.4 × 10−2
10.2 s
1.5
18 %
3.0 × 10−2
13.6 s
1.1
14 %
Np = 4, Cc = 4.0
Runtime Speedup
5.6 s
2.7
10.1 s
1.5
#it.
1
2
3
4
Efficiency
56 %
31 %
22 %
Table 2: Results for simulations with the same parameters as in Table 1, but now with a coarse CFL number Cc = 4.0, again for Np = 4 (upper),
Np = 6 (middle) and Np = 8 (lower) processors.
all, even if using eight processors. As a consequence of the inherently moderate scaling, gains in speedup quickly
deteriorate as the number of processors increases, again advocating the use of Parareal for adding fine grain parallelism
inside nodes. However, solutions with Nit = 2, providing reasonable accuracy, still allow for an acceleration of about
a factor two if using six processors.
The grey lines in Figure 9 indicate the estimated speedup according to (24) with the ratio τc /τf being computed
from sequential test runs. As the partially split scheme requires several acoustic steps, a single coarse step is even
slightly more expensive than a single step of F , resulting in a ratio of τc /τf ≈ 1.165. The reason is that evaluating the
discrete advective flux is relatively straightforward in the linear case and requires neither averaging of velocities to cell
interfaces nor any form of limiting. Including both procedures would increase the computational cost of evaluating
the advective tendencies, reduce the ratio τc /τf and improve the parallel scaling. Also, an implementation coupling
the space and time discretization more closely than the here used method-of-lines approach might allow for a more
efficient implementation of the partially split scheme. Note that estimate (24) neglects the run time required for the
subspace update, but nevertheless gives a quite good estimate of the speedups actually obtained, supporting the point
that in the presented examples the time required for the QR decomposition plays only a minor role.
For comparison, a partially split Runge-Kutta-3 / forward-backward scheme is run sequentially to integrate the
test problem with Csound = 1.2 and Nsound = 6 as well as Csound = 2.4 and Nsound = 12 and, as before, a divergence
damping of νf = 0.005, an end time of T = 2 and a mesh with 40 × 40 cells. Switching to the partially split scheme
reduces the run time to 8.3 seconds for Csound = 1.2 and further to 7.3 seconds for Csound = 2.4, corresponding to
a speedups of 1.8 and 2.0 compared to the non-split version. The relative l2 -differences (41) between the split and
15
Figure 8: Difference between the parallel and the sequential solution for Nit = 1, 2, 3 (grey lines) for Cc = 2.0 (upper) and Cc = 4.0 (lower)
and estimated temporal discretization error of F (black line) over time. The temporal discretization error is estimated by comparing the solution
obtained by running F sequentially to a reference solution obtained using a ten times smaller time step while disabling divergence damping in both
runs. The other parameters of the parallel runs are as indicated in Figure 3.
Figure 9: Measured speedup depending on the number of processors for Nit = 1, 2, 3, 4. The coarse CFL number is Cc = 4.0, using Nsound = 4
acoustic steps in every large step. The mesh consists of 40 × 40 cells (left) or 80 × 80 cells (right). All other parameters are as indicated in Figure
3. The grey lines show the speedup estimate obtained by (24) with an empirically determined ratio τf /τc .
non-split solutions in both cases is about 1.6 × 10−1 . Comparing this with the results in Table 1 shows that KSEParareal achieves this level of accuracy with only a single iteration and provides larger speedups in this case. As
mentioned above, the used method-of-lines approach is not optimal for the partially split scheme and a properly tuned
implementation might yield runtimes comparable to or better than the parallel scheme. An even larger reduction of
the overall time-to-solution of KSE-Parareal could be achieved by using a split scheme for F , too. However, one
would have to deal with two sources of error in this case: the splitting error as well as the error from the iteration of
the parallel scheme.
Finally, Figure 10 shows the distribution of the total run time of the parallel scheme over the three parts of the
iteration, that is the coarse and fine propagator and the subspace update, depending on the total number of degreesof-freedom. The run times are wall clock times, that is the run time for the fine propagator corresponds to the time
required to compute the fine integration step in parallel. The parameters, aside from the number of cells, are as
indicated in Figure 3. As it is to be expected from the asymptotic estimate (22), the fraction of the total run time spent
in the subspace update does not depend on the problem size. For the case with Nit = 3 iterations shown in Figure
10, about 5 % of the total time are spent for the subspace update, so the overhead for the Krylov-subspace-enhanced
version of Parareal compared to the original version is moderate, even when computing a full QR decomposition in
every iteration.
4. Discussion
The paper addresses the applicability of the Krylov-subspace-enhanced Parareal parallel-in-time integration scheme
to a two-dimensional, hyperbolic, linear acoustic-advection problem, which is often used as a test problem for integration methods for numerical weather prediction. First it is shown that the original version of Parareal, as expected
16
from the results of studies by other authors, can become unstable, even if used to integrate a comparatively simple
two-dimensional advection problem. The KSE-Parareal variant, by contrast, does not suffer from this instability. As
KSE-Parareal has up to now not been applied to hyperbolic flow problems, this was not necessarily to be expected.
In contrast to the original version of Parareal, the KSE variant requires a QR decomposition in every iteration. The
results in the present paper show that this can be done by using an implementation of the LAPACK library in a
straightforward way without severely increasing execution times.
The possibility of using a partially split scheme as coarse propagator and thus avoiding the necessity of implementing an implicit method is established. However, in order for the Parareal iteration to converge reasonably fast,
a small amount of divergence damping is required not only in the coarse but also in the non-split fine propagator.
It is demonstrated that the parallel-in-time scheme can capture important features like advection of vorticity and the
amount of diffusivity of the parallel scheme in comparison to the underlying sequential scheme is discussed. Obtained
speedups are indicated and it is found that reasonably accurate solutions can be computed while still allowing for accelerations in the range of factors of 1.5 to 2.1 using between four and eight CPUs. Comparing the speedups achieved
by KSE-Parareal with the acceleration obtained by switching to a partially split scheme shows that KSE-Parareal provides larger speedup while producing solutions of comparable accuracy, but the implementation of the partially split
scheme is probably not the most efficient.
Implementing Parareal essentially requires the iterated application of two integration schemes. As the numerical
examples employ integration schemes that are already widely used in contemporary codes for modeling atmospheric
flows (Runge-Kutta-3, forward-backward, forward Euler), a possible future implementation into an existing code
could likely rely heavily on already implemented features.
The successful application of KSE-Parareal to a nontrivial, meteorologically interesting test problem demonstrated
in this paper indicates that parallel-in-time schemes are promising candidates for increasing the level of parallelism
in codes for numerical weather prediction beyond the already employed spatial parallelization. Such increases will
be necessary to achieve a maximized reduction of the time-to-solution on existing and, probably more important, on
upcoming high performance computing systems featuring a rapidly growing number of CPUs. Also, it is stressed that
the Parareal algorithm allows for parallel computation by design and is not bound to a specific model of parallelization,
so its usefulness is not dependent on a specific architecture. Hence parallel-in-time schemes increase parallelism on
the algorithmic side, thus potentially offering lasting benefits beyond tuning an existing implementation.
Important issues that have to be addressed in future works include: (i) Addressing the performance of the scheme
to a model of increased complexity, for example the compressible Boussinesq system. While the parallel scaling will
most likely benefit from a computationally heavier problem, the ability of the scheme to produce accurate solutions
within a few iterations has to be demonstrated. (ii) Evaluating the performance of a combined spatial/temporal parallelization. Because of the limited scaling of the parallel-in-time scheme, the most promising approach seems the
use of a hybrid MPI/OpenMP parallelization, where subdomains are assigned to computational nodes that communicate by message passing while the time stepping within the subdomains is done in parallel using the shared memory
and the CPUs available inside a node. It remains to be demonstrated that such a hybrid approach can reduce the
time-to-solution below what a pure spatial parallelization can achieve. (iii) The extension of the Krylov-subspace
enhancement for PITA introduced in [22] relies on the use of an implicit scheme. An adaption for explicit schemes
would be desirable in order to avoid the necessity of implementing an implicit scheme into existing NPW codes. (iv)
Investigating alternatives to the used integration schemes with respect to their performance in the Parareal framework.
A simple modification would be to use two partially split schemes for both, the coarse and the fine propagator, in order
to further reduce run times. But also completely different schemes might be worth considering. As examples, improved split-explicit schemes as introduced in [27] or [38] are mentioned. Interesting candidates are also the explicit
peer-methods introduced in [39], that allow for a splitting between low and high frequency modes without requiring
artificial damping. Exploring the possibility of using a reduced-physics sound-proof model in the coarse propagator
might also be of interest.
17
Figure 10: Fraction of run time (wall clock time) spent in the coarse propagator, the fine propagator and the subspace update for increasingly fine
meshes, that is a growing number of degrees-of-freedom. All other parameters are as indicated in Figure 3.
Acknowledgments
This work is funded by the Swiss ”High Performance and High Productivity Computing Initiative” (HP2C). The
implementation uses the ”Automatically Tuned Linear Algebra Software” (ATLAS)9, the GCC compiler suit, the
FORTRAN Namelist Suite for MATLAB by S. Lazerson and the NetCDF library10 .
References
[1] Sutter H. The free lunch is over: A fundamental turn toward concurrency in software. Dr Dobb’s Journal 2005;30(3).
[2] Asanovic K, Bodik R, Demmel J, Keaveny T, Keutzer K, Kubiatowicz J, et al. A view of the parallel computing landscape. Commun ACM
2009;52:56–67.
[3] Wehner M, Oliker L, Shalf J. Towards ultra-high resolution models of climate and weather. International Journal of High Performance
Computing Applications 2008;22(2):149–65.
[4] Drake J, Foster I. Introduction to the special issue on parallel computing in climate and weather modeling. Parallel Computing 1995;21:1539–
44.
[5] Michalakes J, Vachharajani M. GPU acceleration of numerical weather prediction. Parallel Processing Letters 2008;18(4):531–48.
[6] Lions JL, Maday Y, Turinici G. A ”parareal” in time discretization of PDE’s. Comptes Rendus de l’Academie des Sciences Series I
Mathematics 2001;332:661–8.
[7] Bal G, Maday Y. A ”parareal” time discretization for non-linear PDE’s with application to the pricing of an american put. In: Lecture Notes
in Computer Science and Engineering; vol. 23. Springer Berlin; 2002, p. 426–32.
[8] Maday Y, Turinici G. Parallel in time algorithms for quantum control: Parareal time discretization scheme. Int J Quant Chem 2003;93(3):223–
8.
[9] Trindade JMF, Pereira JCF. Parallel-in-time simulation of the unsteady Navier-Stokes equations for incompressible flow. Int J Numer Meth
Fluids 2004;45:1123–36.
[10] Trindade JMF, Pereira JCF. Parallel-in-time simulation of two-dimensional, unsteady, incompressible laminar flows. Numerical Heat Transfer, Part B 2006;50:25–40.
[11] Fischer PF, Hecht F, Maday Y. A parareal in time semi-implicit approximation of the Navier-Stokes equations. In: Kornhuber R, Hoppe
RHW, Péeriaux J, Pironneau O, Widlund OB, Xu J, editors. Proceedings of the 15th International Domain Decomposition Conference; vol. 40
of Lect. Notes Comput. Sci. Eng. Berlin: Springer; 2003, p. 433–40.
[12] Liu Y, Hu J. Modified propagators of parareal in time algorithm and application to Princeton ocean model. International Journal for Numerical
Methods in Fluids 2008;57(12):1793–804.
[13] Samaddar D, Newman DE, Sánchez RS. Parallelization in time of numerical simulations of fully-developed plasma turbulence using the
parareal algorithm. J Comp Phys 2010;229:6558–73.
[14] Minion ML. A hybrid parareal spectral deferred corrections method. Comm App Math and Comp Sci 2010;5(2):265–301.
[15] Gander MJ, Vandewalle S. Analysis of the parareal time-parallel time-integration method. SIAM J Sci Comp 2007;29(2):556–78.
[16] Gander M, Hairer E. Nonlinear convergence analysis for the parareal algorithm. In: Langer U, Widlund O, Keyes D, editors. Proceedings of
the 17th International Domain Decomposition Conference. LNCSE; Springer; 2008, p. 45–56.
[17] Farhat C, Chandesris M. Time-decomposed parallel time-integrators: Theory and feasibility studies for fluid, structure, and fluid-structure
applications. Internat J Numer Methods Engrg 2005;58:1397–434.
[18] Bal G. On the convergence and the stability of the parareal algorithm to solve partial differential equations. In: Kornhuber R, Hoppe RHW,
Péeriaux J, Pironneau O, Widlund OB, Xu J, editors. Proceedings of the 15th International Domain Decomposition Conference; vol. 40 of
Lect. Notes Comput. Sci. Eng. Berlin: Springer; 2003, p. 426–32.
9 www.math-atlas.sourceforge.net
10 www.unidata.ucar.edu/software/netcdf/
18
[19] Staff GA, Rønquist EM. Stability of the parareal algorithm. In: Kornhuber R, Hoppe RHW, Péeriaux J, Pironneau O, Widlund OB, Xu J,
editors. Proceedings of the 15th International Domain Decomposition Conference. LNCSE; Berlin: Springer; 2003, p. 449–56.
[20] Gander MJ. Analysis of the parareal algorithm applied to hyperbolic problems using characteristics. Bol Soc Esp Mat Apl 2008;42:21–35.
[21] Farhat C, Cortial J, Dastillung C, Bavestrello H. Time-parallel implicit integrators for the near-real-time prediction of linear structural dynamic
responses. Int J Numer Meth Engng 2006;67:697–724.
[22] Cortial J, Farhat C. A time-parallel implicit method for accelerating the solution of non-linear structural dynamics. Int J Numer Meth Engng
2008;77:451–70.
[23] Gander M, Petcu M. Analysis of a Krylov subspace enhanced parareal algorithm for linear problems. ESAIM Proc 2008;25:114–29.
[24] Prusa JM, Smolarkiewicz PK, Wyszogrodzki AA. EULAG, a computational model for multiscale flows. Computers & Fluids 2008;37:1193–
207.
[25] Skamarock WC, Klemp JB. The stability of time-split numerical methods for the hydrostatic and non-hydrostatic elastic equations. Mon Wea
Rev 1992;120:2109–27.
[26] Wicker LJ, Skamarock WC. Time-splitting methods for elastic models using forward time schemes. Mon Wea Rev 2002;130:2088–97.
[27] Gassmann A. An improved two-time-level split-explicit integration scheme for non-hydrostatic compressible models. Meteorol Atmos Phys
2005;88:23–8.
[28] Baldauf M. Linear stability analysis of Runge-Kutta-based partial time-splitting schemes for the Euler equations. Mon Wea Rev
2010;138:4475–96.
[29] Wicker LJ, Skamarock WC. A time-splitting scheme for the elastic equations incorporating second-order Runge-Kutta time differencing.
Mon Wea Rev 1998;126:1992–9.
[30] Kaber SM, Maday Y. Parareal in time approximation of the Korteveg-deVries-Burgers’ equations. Proc Appl Math Mech 2007;7:1026403–.
[31] Maday Y, Turinici G. The parareal in time iterative solver: A further direction to parallel implementation. In: Kornhuber R, Hoppe RHW,
Péeriaux J, Pironneau O, Widlund OB, Xu J, editors. Proceedings of the 15th International Domain Decomposition Conference; vol. 40 of
Lect. Notes Comput. Sci. Eng. Berlin: Springer; 2003, p. 441–8.
[32] Trefethen LN, III DB. Numerical Linear Algebra. SIAM; 1997.
[33] Aubanel E. Scheduling of tasks in the parareal algorithm. Parallel Computing 2011;37:172–82.
[34] Deuflhard P, Hohmann A. Numerische Mathematik I. Eine algorithmisch orientierte Einführung. De Gruyter; 2002.
[35] Durran DR. Numerical Methods for Fluid Dynamics with Applications to Geophysics. Texts in Applied Mathematics; Springer; 2nd ed.;
2010.
[36] Baldauf M. Integration by time-splitting in mesoscale models. In: Doms G, Schättler U, editors. COSMO Newsletter No. 2. 2002, p. 101–4.
[37] Chapman B, Jost G, van der Pas R. Using OpenMP - portable shared memory parallel programming. Scientific and Engineering Computation
Series; The MIT Press; 2008.
[38] Wensch J, Knoth O, Galant A. Multirate infinitesimal step methods for atmospheric flow simulation. BIT Numer Math 2009;49:449–73.
[39] Jebens S, Knoth O, Weiner R. Explicit two-step peer methods for the compressible Euler equations. Mon Wea Rev 2009;137:2380–92.
19
| 5cs.CE
|
On the Fundamental Theorem of the Theory of
Relativity
arXiv:1603.07328v3 [math-ph] 1 Sep 2016
Marco Mamone-Capria
Dipartimento di Matematica – via Vanvitelli, 1 – 06123 Perugia - Italy
E-mail: [email protected]
March 8, 2018
Abstract
A new formulation of what may be called the “fundamental theorem of the
theory of relativity” is presented and proved in (3+1)-space-time, based on the full
classification of special transformations and the corresponding velocity addition
laws. A system of axioms is introduced and discussed leading to the result, and
a study is made of several variants of that system. In particular the status of the
group axiom is investigated with respect to the condition of the two-way isotropy
of light. Several issues which are ignored or misunderstood in the literature are
emphasized.
Keywords Relativity principle, worldlines, space-time homogeneity, spatial
anisotropy, two-way velocity of light.
Contents
1 Introduction
2
2 Basic assumptions (Axioms 1-7)
3
3 Inertial equivalence and physical velocities (Axioms 8-10)
7
3.1 Space-time ‘homogeneity’ . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4 Special transformations (Axiom 11)
12
4.1 Explicit formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5 Spatial Isotropy and the Fundamental Theorem (Axiom 12)
19
6 Two-way light isotropy (Axiom 6∗ )
22
6.1 Absolute simultaneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1
6.2 Two-way light isotropy and the group axiom . . . . . . . . . . . . . . . . 25
7 Appendix – Proof of Theorem 3.3
1
28
Introduction
The discovery of the equations of the special Lorentz transformation (SLT) prompted
several scholars, since the early days of the special theory of relativity, to look for general
assumptions, possibly less questionable than the postulate of constancy of the velocity
of light, from which those equations could be derived. In 1936 the Italian mathematician
Francesco Severi echoed a widespread concern when he qualified the “intervention of the
light-velocity” in special relativity as “the scandal of the theory” ([39], p. 260). The
search for alternative assumptions was on the whole successful, and in fact produced
several derivations of the SLT – and, at the same time, of the corresponding classical,
Galilean transformations – from general axioms on space and time taken together with
some ‘relativity’ condition. This search represents an important, if somewhat neglected,
undercurrent in the development and reception of special relativity, which has attracted
very many physicists and mathematicians since the pioneering contributions by von
Ignatowsky ([43, 44]) till the present day.
What I call the ‘Fundamental Theorem of the Theory of Relativity’ is an ideal
theorem which puts together the essential insights gained through this volume of work.
In the present paper I advance one proposal for such a theorem – the kind of statement
that, accompanied by physical and historical supporting arguments, should find its way
into textbooks. The basic idea is not new in itself: to create an appropriate formal
and conceptual environment which allows for the notion of special transformation to
develop into a structure group.
Too many articles on this subject have been published in more than a century for
a single scholar to be sure that nothing relevant escaped his or her attention. I made
a honest attempt not to miss apparently important texts, my list of references being a
representative sample of the (often repetitious) literature I checked. I have come to the
conclusion that many of the arguments found in the literature are flawed, unsatisfactory,
or incomplete. One reason for my dissatisfaction is that those arguments usually and
tacitly assume that the passage from the derivation in (1+1) space-time to the derivation
in (3+1) space-time is straightforward, which is not; also, in most treatments the notion
of a physical process or of signalling is not sufficiently clarified; and, to give a further
example, the status of the group assumption is often left hazy.
Still, in a classical field such as that of the foundations of relativity it would be
hazardous, if not arrogant, to claim an high degree of originality for one’s results. Several
authors have found themselves informed by colleagues in due time, and sometimes
shortly after publication, that what they had published was close, and sometimes very
close, to results already in the published record.1 I have done my best to avoid this
risk, by screening the literature for similarities with my main results.
1
See for instance the “Postscript” to [26] added in [27] or the “Note added in proof” of [21]. In some
cases, like [4] vs. [8], the ‘anticipated’ paper has come to be much more famous and cited than the
‘anticipating’ one.
2
In §2 the basic definitions and axioms are stated concerning the class of privileged,
or admissible, coordinate systems (space-time structure); they make, first of all, the
space-time into a differential manifold diffeomorphic to standard R4 . The concept
of a worldline is introduced, together with axioms guaranteeing that all admissible
coordinate systems share the same notion of causality, time orientation and spatial
orientation. One axiom which is rarely, if ever, explicitly stated in the literature is Axiom
6, which limits the richness of the space-time structure by requiring that mutually at
rest cs’s have the same spatial geometry and, up to an additive constant, the same time
coordinate. Of course this rules out all unit changes but, more importantly, prevents the
introduction of nonstandard synchronies. These are dealt with separately in terms of
the same theoretical framework, and their relationship with the group condition (Axiom
7) is examined in §6.
Inertial equivalence is carefully defined in §3, in a way which takes into account, as
is unfrequently done, the fact that in a theory which is designed to include as particular
cases both classical mechanics and special relativity not all velocities may be allowed a
priori for uniform motions. The link between inertial equivalence and affine equivalence
is stressed, with a criticism of the usual approach to it in terms of ‘homogeneity of
space-time’. Axiom 10 links the set of physical velocities with the admissible coordinate
systems. Section 4 is devoted to the concept of special transformation, which is at the
core of the vast majority of treatments of the ‘fundamental theorem’, and contains
a theorem which reaches back to Frank and Rothe’s 1911 classic paper ([15]), but
which seems never to have been fully stated and proved. It is also shown that not
all space-time structures satisfying axioms 1-9 satisfy Axioms 10 and 11, in particular
they may not admit any one-dimensional subgroups of special transformations. In
the next section the final condition, spatial isotropy (Axiom 12), and my proposal for
the ‘fundamental theorem’ are presented, the latter assuming all axioms 1-12. The full
family of trasformations satisfying two-way light velocity isotropy (Axiom 6∗ ) is derived
in §6, and the issue of the compatibilty of Axioms 6∗ and 7 is discussed, which motivates
the introduction of what I call the Reichenbach groups, with some comments on their
relationship to the issue of the conventionality of simultaneity.
2
Basic assumptions (Axioms 1-7)
The general framework adopted in this paper is essentially the traditional one, reconstructing space-time structures by means of classes of admissible (or privileged) coordinate systems. A space-time theory is specified when the rules for the selection of such
coordinate systems are formulated. This contrasts with the approach favored in purely
mathematical treatments, where geometric structures are superimposed from the start
to some set, and privileged coordinate systems are derived thence. In our approach, for
instance, the requirement that ‘space-time has a real affine structure’ is to be reformulated in terms of the existence of a class of affinely equivalent coordinate systems; and
a reasonably physical, if idealized, criterion must be provided for determining whether
a given coordinate system belongs or not to such a class. We denote by E the spacetime; its elements are called events. In the following only the more important items
(definitions, theorems) are numbered; axioms are listed and numbered separately.
3
Definition 2.1 A (global) coordinate system (cs) is a bijection φ : E → R4 . A
space-time structure is a nonempty class Φ of global coordinate systems, which are
variously termed as privileged or admissible. A cs in the space-time structure Φ is
denoted by
φ : E → R4 ≡ R3 × R, p 7→ (r(p), t(p)),
where r ≡ rφ : E → R3 is the position function and t ≡ tφ : E → R is the time
function of φ. A property or a quantity is absolute if it is the same with respect to
all φ ∈ Φ; it is relative otherwise. Two events p, q are synchronous with respect to
φ if t(p) = t(q).
By theory of relativity I mean the theoretical framework which is described by
axioms 1-5 only. Several other axioms will be stated in due course producing different
specializations of the theory. I will not insist on the considerable amount of idealization
in each of the axioms, starting with the first, which is a cardinality axiom.
Axiom 1 [space-time structure] Physical space-time has a space-time structure Φ.
Axiom 2 [topology and differential structure] For every φ, φ′ ∈ Φ the transition
function
′
r = r′ (r, t)
′
−1
4
4
φ ◦ φ : R → R , or equivalently
,
t′ = t′ (r, t)
is a C 2 diffeomorphism. Space-time E is endowed with the topological and differential
structures determined by any atlas consisting of a single global chart (E, φ) with φ
admissible and R4 standard.
That E be topologically and differentially modeled on standard 4-dimensional Euclidean space is a nontrivial assumption that is maintained, in local form, also in general
relativity. In particular, from the assumption that E is homeomorphic to R3 × R (with
the product topology) it does not follow that E is also diffeomorphic to R3 × R (with
the product differential structure).2
A crucial concept is that of a physical process in space-time, formally introduced as
follows.
Definition 2.2 A worldline is any regular curve in E which can be totally parametrized
by φ’s time coordinate, for every φ ∈ Φ.
This is to be understood as follows: a worldline Γ is such that for every φ ∈ Φ
there is a differentiable map r : I → R3 such that φ(Γ) is the inverted graph of r, i.e.
φ(Γ) = {(r(s), s) : s ∈ I}, where I is a (nondegenerate) interval of R; we say that r
represents Γ in the 3-space of φ. Given such a function, the 3-vector
2
I remind the reader of the highly remarkable fact that of all Rn s, R4 is the only one to have nondiffeomorphic differential structures; indeed, there is a continuum of such structures (for an introduction
to these results in differential topology, see for instance chapter I of [6]).
4
v :=
dr
,
ds
is the 3-velocity (i. e. the ordinary physical velocity) of the worldline at a fixed event
according to φ. 3
Definition 2.3 A point at rest in any admissible cs φ is the image of a map of the
form
γ : R → E, s 7→ φ−1 (r0 , s).
where r0 ∈ R3 .
Axiom 3 [causality] Any point at rest in an admissible cs is a worldline.
This is equivalent to requiring the following condition to hold:
∂t′
6= 0
∂t
(1)
for every pair of admissible φ, φ′ and for all (r, t) ∈ R4 ; when inequality (1) holds, φ
and φ′ may be seen as ‘causally compatible’.
Axiom 4 [time orientation] All admissible cs’s define the same time order on all
worldlines.
Given (1), this is equivalent to requiring the following condition to hold:
∂t′
>0
∂t
(2)
It is easy to verify, by exchanging roles between φ and φ′ , that as a consequence
the spatial Jacobian determinant must be nowhere zero:
det(
∂x′α
)α,β=1,2,3 6= 0.
∂xβ
(3)
Axiom 5 [spatial orientation] All admissible cs’s define the same spatial orientation
(i. e., the determinant in (3) is positive, for every pair of admissible cs’s).
Because of Axiom 3, spatial orientation is equivalent to space-time orientation, thus
for all admissible φ, φ′ , the transition function φ′ ◦ φ−1 is an orientation-preserving
diffeomorphism of R4 .
Here ends the list of the few basic axioms defining our most general version of theory
of relativity (which, of course, fails to encompass “general relativity”, with or without
the field equation).
3
For other purposes, which will not concern us here, one may admit nonregular curves as worldlines,
e.g. with a discrete subset of points at which 3-velocity is not well-defined.
5
The coordinates of admissible cs’s admit of a natural interpretation. If p and q are
φ-synchronous, or φ-simultaneous, events (i.e. t(p) = t(q)), their φ-distance is
dφ (p, q) := |r(p) − r(q)|.
On the other hand, for a point at rest in φ the difference of the φ-time coordinates
of two events in the worldline is interpreted as the (positive or negative) time lapse
between the events as measured by a clock described by that worldline.
Given admissible cs’s φ and φ′ , we say that φ′ is at rest with respect to φ if all
points at rest in φ′ are also at rest in φ. It is easy to see that if φ′ is at rest with
respect to φ, then also φ is at rest with respect to φ′ (so “reciprocity” – cf. §5 – holds
automatically for zero velocity), and also transitivity holds. Thus ‘being at rest’ is
an equivalence relation between admissible cs’s , partitioning Φ into static equivalence
classes.
Definition 2.4 The group of all 4 × 4 matrices of the form ΣS with S ∈ SO(3), where
ΣS =
S 0
0T 1
,
(4)
will be denoted by SO4(3). They are called spatial rotations.
The group of all (space-time) translations
Tb : R4 → R4 , Tb (x) = x + b
will be denoted by T (R4 ).
Axiom 6 [mutual rest] Admissible cs’s which are mutually at rest differ by a spatial
rotation and a space-time translation, i.e. for any two such cs’s φ, φ′ there exist S ∈
SO(3) and b ∈ R4 such that
φ′ = Tb ◦ ΣS ◦ φ.
Axiom 6 means that mutually at rest, admissible cs’s must share the same space
distance and, up to a translation, the same time coordinate. In particular they must
have the same synchrony relation. This is a standard simplifying requirement, ruling
out conventionalist ǫ-theory (§6.2), and should not be confused with its converse, which
requires that if φ is an admissible cs and S is any spatial rotation, then ΣS ◦ φ is also
admissible (this is “spatial isotropy”, our Axiom 12, to be introduced at a later stage).
The principle of relativity is often considered to have as a consequence, or as its
algebraic counterpart, the group property for the set of all transition functions between
admissible cs’s (e.g. [3]).4 Actually, the essence of “relativity”, historically speaking, is
the requirement that some fundamental physical laws be expressed in the same way in
4
In Einstein’s article [11] (p. 907) one finds a statement to the effect that the special transformations
must form a group, with no further elaboration or explanation.
6
all cs’s belonging to a certain, non-static5 fundamental class Φ. A group G , or more
exactly a subgroup of the transformation group of R4 , may be involved in the selection
of such a class in a natural way, that is, when the class of admissible cs’s happens to
be the orbit of a single cs under the action of G on Bi(E, R4). However that a group
arises (in this sense) from requiring invariance of a number of laws cannot in general be
taken for granted: it depends on the form of the law(s) (cf. §6).
I state the existence of a structure group as one of the axioms:
Axiom 7 [structure group] There is a subgroup G of Bi(R4 ) such that Φ = G · φ
where φ ∈ Φ and the “dot” at the righthand side denotes the natural left action of
Bi(R4 ) on Bi(E, R4) (i. e. composition of maps).
Since G is a group, it clearly does not matter which φ ∈ Φ is chosen in Axiom 6.
Equivalently, all the maps:
F φ : Φ → Bi(R4 ), φ′ 7→ φ′ ◦ φ−1
(5)
have the same image (i.e. G ) no matter how φ ∈ Φ is chosen.
From Axioms 6 and 7 it follows that the subset of all g ∈ G having indentically
vanishing 3-velocity function is in fact a subgroup of the Newton group GN = SO4 (3) ⋊
T (R4 ) (cf. [24]), and the static equivalence classes are its orbits in Φ. For future
reference I shall call this subgroup the rest subgroup of G .
3
Inertial equivalence and physical velocities (Axioms
8-10)
Classical and special relativity both assume the validity of the law of inertia, which
permits a drastic simplification in the form of the transition functions between cs’s.
This law, in the present context, can be interpreted as claiming that there is a class of
worldlines which are uniformly moving in an absolute sense, i. e. they have constant
3-velocities for all admissible cs’s. Dynamics adds an explanation why these worldlines
are so privileged, but in order for the inertiality requirement to be stated and put to
work, no dynamical complements are needed.
There is a classic theorem in the foundations of geometry (see e.g. [18]), according
to which every bijection F from Rn to itself (with n > 1) which maps all straightlines
into straightlines is an affinity, i. e. it is a map of the form
F : Rn → Rn , x 7→ Bx + b, with B ∈ GL(n, R), b ∈ Rn .
This theorem, however, cannot be used directly in the present setting to conclude to
inertiality, contrary to what some authors assume (e.g. [16], p. 750; more weakly, [9],
p. 29, [34], p. 9). In fact it is a priori unreasonable, and a posteriori inconsistent,
to assume that all straightlines according to some admissible cs are worldlines. This
is true neither in Newtonian physics nor in special relativity. In Newtonian physics
5
Incidentally, for systematic as well as for historical reasons, I think that room should be left in the
presentation of the principle of relativity also for ‘Aristotelian’ and ‘Newtonian’ static space-times.
7
the straightlines which are not worldlines are those (and only those) contained in any
simultaneity space; in special relativity, more embarrassingly, the straightlines which
are not worldlines cannot even be obtained as limits of sequences of worldlines, since
they are, as is well known, all straightlines passing through an event and lying outside
the lightcone with vertex at that event.6
Definition 3.1 A worldline is uniform with respect to φ ∈ Φ if its 3-velocity is
constant according to φ; it is uniform if it is uniform with respect to every φ ∈ Φ.
We need to assume something on how many uniform worldlines exist with respect to
every admissible cs. The following is a reasonable assumption:
Axiom 8 [physical velocities] For every admissible cs φ and p ∈ E , the subset
V̂φ,p ⊆ R3 of 3-velocities of worldlines through p which are uniform with respect to φ
is a star-shaped neighborhood of 0.
Under this assumption, not all speeds turn out to be physically possible in all directions according to any given φ: it is only assumed that for every direction, worldlines
with small enough constant velocity (how small possibly depending on the direction)
exist and that in every direction there is a half-interval of speeds starting at 0. We
shall see that this level of generality applies to concrete and not particularly ‘exotic’
examples (§6.2).
Notice that if φ and φ′ are affinely equivalent, i. e. φ′ ◦ φ−1 is an affinity, then we
can write uniquely (cf. (2) and Axiom 5)
′
x = Bx + b, where B =
A −AV
kT
α
, b ∈ R4 , and α > 0,
(6)
where A is a 3 × 3 matrix with det A > 0 and V is the (constant) velocity with respect
to φ of any point at rest in φ′ . We shall also say that V is the velocity of the matrix
B and write: V = VB . Moreover
∂t
det A
0< ′ =
= det
∂t
det B
I3 −V
kT α
−1
=
1
,
α+k·V
from which it follows that
α + k · V > 0.
(7)
Definition 3.2 Two admissible cs’s are inertially equivalent if all worldlines which
are uniform with respect to one are uniform also with respect to the other. A space-time
structure is inertial if all its members are inertially equivalent.
The theorem which contains what is needed about the relationship between inertiality and affinity is the following:
6
Weyl [47] (pp. 179, 313) sketches a proof which applies a weaker form of the theorem to the case
that the inertial worldlines belong to a “given, arbitrarily thin, cone”. Also Schwartz ([36]) recognizes
the difficulty, but his adaptation of the theorem is not satisfactory, since it assumes, unwarrantedly in
its context, that the equation of an hyperplane must be affine.
8
Theorem 3.3 If φ and φ′ are inertially equivalent, then they are affinely equivalent,
and conversely.
Proofs of similar statements can be found in the literature (e.g. [14]), but for completeness’ sake a proof fitting the present formalism and assumptions (and in particular
Axiom 8) is provided in the Appendix.
Axiom 9 [inertiality] All admissible cs’s are inertially equivalent, and any cs which
differs from an admissible cs by any space-time translation is also admissible.
From Axioms 7 and 9 and Theorem 3.3, we can conclude that G is a subgroup of
Af f (R4 ), which is the affine group of R4 .
Remember that a group G is a semi-direct product of two subgroups H and K if it
is generated by their union, if their intersection is trivial and if one of them is normal;
in case K is the normal subgroup we write G = H ⋊ K . The next proposition is a
special case of Proposition 2.4 of [24].
Proposition 3.4 Structure group G is of the form
G = G0 ⋊ T (R4 ),
(8)
where G0 is a subgroup of the general linear group GL(4, R).
Proof Given Theorem 3.3 and Axiom 9, we have to show that if
T (R4 ) ≤ G ≤ Af f (R4),
then G is of the form (8): but this follows easily from the fact that Af f (R4 ) =
GL(4, R) ⋊ T (R4 ).
We now introduce an axiom guaranteeing the ‘abundance’ of admissible cs’s and
specifying their relationship with physical velocities.
Axiom 10 [abundance of cs’s] There are φ ∈ Φ and p ∈ E such that for every V
lying in the interior of V̂φ,p , there exists a φ′ ∈ Φ such that the velocity of φ′ with
respect to φ is V ; conversely, for every φ′ ∈ Φ the velocity of φ′ with respect to φ lies
in the interior of V̂φ,p .
Remark In order to show the need for Axiom 10, one can take as G0 the subgroup of
the proper orthochronous Lorentz group L+
↑ formed by all matrices having (0, 0, 1, 0)
as their 3rd row. Then, clearly, the relative velocity of φ′ with respect to φ, for any
φ′ ∈ G0 ·φ, can never attain (or exceed) c, but the line represented in φ by the equations
x1 = 0, x2 = 0, x3 = tU , with U 6= 0 constant, is a worldline whose velocity in φ may
be as large as desired. Thus Axiom 10 would fail.
Notice, also, that Axioms 8 and 10 taken together rule out ‘static’ structure groups
such as, e.g., the Newton group GN .
♦
Proposition 3.5 The set V := Int(V̂φ,p ) is independent of both p and φ and starshaped.
9
Proof The set V̂φ,p does not depend on p: in fact if Γ is a uniform worldline through
the origin o = φ−1 (0) of φ and φ(p) = b , then φ−1 (φ(Γ) + b) is a uniform worldline
through p, as is easy to verify, and the 3-velocity of the latter at p according to φ
coincide with the 3-velocity of Γ at o. So V̂φ,p = V̂φ , but we also have that Int(V̂φ )
does not depend on φ either, because by Proposition 3.4 and Axiom 10
Int(V̂φ ) = {VB : B ∈ G0 },
where the righthand side clearly does not depend on φ (cf. (5)).
As to the last point, for every t ∈]0, 1[ the set t · V is open and contained in V̂φ , for
any φ (Axiom 8), therefore
t · V ⊆ Int(V̂φ ) = V,
and the thesis follows.
Remark The distinction between worldlines having a 3-velocity always in V and worldlines with a velocity possibly on the boundary of V is the distinction between (worldlines
of) particles (on which an instantaneous admissible cs can be based, at any fixed event)
and signals (on which, if they are not particles, there are events at which an instantaneous cs cannot be based). Nothing in our argument until and including Theorem 5.1
depends on the assumption that there are signals which are not particles.
♦
The following proposition states an intuitively plausible fact about velocities of cs’s.
Proposition 3.6 Let φ, φ1 , φ2 ∈ Φ, with Φ inertial, be such that the velocities of φ1
and φ2 with respect to φ are equal; then φ1 is at rest with respect to φ2 .
Proof Suppose that φ1 and φ2 are admissible cs’s, having the same velocity U with
respect to φ; we fix our notation by writing
x1 = B1 x + b1 , x2 = B2 x + b2 , with VB1 = VB2 = U.
Clearly x2 = Bx1 + b3 for a suitable b3 ∈ R4 and B := B2 B1−1 . Since B2 = BB1 , a
simple computation gives the velocity VB2 of φ2 with respect to φ as
T −1
−1
VB2 = (I3 − A−1
1 Vk1 ) (U + α1 A1 V).
(9)
−1
Equating the LHS to U we get U − (k1 · U)(A−1
1 V) = U + α1 A1 V , and therefore
(α1 + k1 · U)(A−1
1 V) = 0. Now, by (7), V must vanish, that is, φ1 and φ2 are mutually
at rest.
Corollary 3.7 Let B1 , B2 ∈ G0 ; if VB1 = VB2 , then there is a S ∈ SO(3) such that
B2 = ΣS B1 .
Proof Take φ ∈ Φ and define φi = Bi φ (i = 1, 2); then apply Axiom 6 to the previous
proposition.
10
3.1
Space-time ‘homogeneity’
The topic of inertiality has often been discussed in a somewhat misleading way, under
the label of “space-time homogeneity”, following the lead of Einstein’s 1905 article.7 Let
us first explain theoretically how this view has come about (cf. e.g. [19], [13]).
Any admissible cs defines its own associated affine structure on E , which is obtained
by introducing a vector space V (E, φ) of translations in the natural way:
T(φ,b) : E → E, p 7→ φ−1 (φ(p) + b), b ∈ R4 .
Now let φ′ be another admissible cs, and suppose that
V (E, φ′) = V (E, φ).
(10)
This means that for every b ∈ R4 there is a b′ ∈ R4 such that
T(φ,b) = T(φ′ ,b′ ) ,
or equivalently, denoting F := φ′ ◦ φ−1 , we must have that for every b ∈ R4 , F ◦
Tb ◦ F −1 is a translation; that is, F lies in the normalizing subgroup of T (R4 ) in the
homeomorphism group of R4 . That this subgroup coincides with the affinity group
Af f (R4 ) is easy to verify.8
Condition (10) means that the affine structures associated to φ and φ′ are the
same, but this is, of course, much more than simply asking for the space-time to be
“homogeneous”. In fact homogeneity is standardly defined as the property of a set-group
pair where the group acts on the set transitively; but not all homogeneous spaces are
affine, not even those of more evident geometric or physical interest.
For instance, we can define Mk as what we may call the generalized Minkowski
space-time Mk : 9
Mk = Sk × R, k = 0, 1, −1,
where Sk = R3 , S 3 , H 3 , which are the Euclidean, spherical and hyperbolic space forms
with curvature k = 0, 1, −1 respectively, dσk2 is the standard Riemannian metric on Sk ,
and the space-time metric is:
ds2 = dσk2 − c2 dt2 .
Clearly M0 is Minkowski space-time, but Mk is a homogeneous space also for k = ±1
with respect to its own isometry group; nevertheless it is not an affine space. Notice that
7
“First of all, it is clear that the equations must be linear because of the properties of homogeneity
which we attribute to space and time” ([11], p. 898; italics in the original). Cf. [20], [8], [21].
8
Details are given, for instance, in [4], p. 1519. Other arguments assume a differentiability condition
(cf. [13]).
9
These space-times are topologically, but not metrically, equivalent to the Friedmann space-times,
of well-known cosmological relevance.
11
M1 is the static Einstein space-time of 1917, with unit radius of the universe ([12]), and
that M−1 is also diffeomorphic to R4 . The reason we value particularly the affine group
is its link with the inertial motions (Theorem 3.3), not that it is the only way we can
conceive of space-time as being “homogeneous”. Therefore the emphasis on “space-time
homogeneity” as a supposedly more modest, or clarifying, assumption than inertiality
is in fact question-begging.10
4
Special transformations (Axiom 11)
The notion of classical coordinate systems related by an affine transformation, with
parallel spatial axes and same space-time origins:
r′ = a(r − tV)
,
t′ = bt
with a, b > 0, generalizes to a transformation of type (6) by requiring that: 1) V
is a positive eigenvector of A, 2) A restricted to [V]⊥ is a positive homothety, and
3) synchrony is preserved (at least) on 2-planes orthogonal to V . This leads to the
following definition.
Definition 4.1 A special transformation with velocity V = vu, |u| = 1, is a linear
map of R4 whose matrix is a special matrix, that is of the type:
λ
I3 + (a − 1)uuT −aV
bVT
α̂
.
(11)
with a, α̂, λ > 0. A pair of admissible cs’s (φ, φ′) is a special pair if φ′ ◦ φ−1 is a
special transformation.
The next axiom is implicit in most, if not all, treatments of special relativity which
do not start from light-velocity invariance. It remains hidden in standard (1+1)dimensional treatments because under the assumption of just one spatial dimension
all inertial tranformations must be ‘special’.
Axiom 11 [special transformations] There exists a unit vector u ∈ R3 and a
differentiable map defined on an open interval I ⊆ R
B ≡ Bu : I → G, with VB(v) = vu, B(0) = I4 ,
such that the image of B comprises all special transformations in the direction u.
Remark Even when Axiom 11 is satisfied for some u in a given theory of relativity,
this does not mean that in that theory the special transformations are the ‘typical’ form
10
The obscurity of the derivation of affinity from homogeneity is pointed out, for instance, in [28]
and [31]. In the present context proofs based on the invariance of d’Alembert equation or on light
signalling ([45], [48], [17]) are obviously not relevant.
12
of the coordinate changes, to which, that is, reduction is always possible via a spatial
rotation. In fact in the structure group there might not be enough spatial rotations. ♦
It is easy to check that the product of any two such transformations B(v1 )B(v2 ) is
a special transformation in the same direction, and therefore it can be written as B(w)
for a unique w , and the same is true for the inverse of every special matrix. Thus
the image of B is a one-dimensional Lie subgroup of G diffeomorphic to R, which we
denote by H(u) and which can be identified with a subgroup of GL(4, R).
We denote w , as just defined, by v1 ∗ v2 , that is11
v1 ∗ v2 := B −1 (B(v1 )B(v2 )).
(12)
Now by computing the product matrix B(v1 )B(v2 ) and by defining:
(13)
f (v) := α̂(v)/a(v), g(v) := b(v)/a(v),
we obtain the following relationships:
a(v1 ∗ v2 )
a(v1 ∗ v2 )v1 ∗ v2
b(v1 ∗ v2 )v1 ∗ v2
α̂(v1 ∗ v2 )
λ(v1 ∗ v2 )
= a(v1 )a(v2 )(1 − g(v2 )v1 v2 ),
= a(v1 )a(v2 )(f (v2 )v1 + v2 ),
= a(v1 )a(v2 )(g(v1 )v1 + f (v1 )g(v2)v2 ),
(14)
= a(v1 )a(v2 )(f (v1 )f (v2 ) − g(v1 )v1 v2 ),
= λ(v1 )λ(v2 ).
Dividing the second equation by the first we get:
v1 ∗ v2 =
f (v2 )v1 + v2
.
1 − g(v2)v1 v2
(15)
Notice that f is always positive and f (0) = 1. The following statement provides a full
characterization of the subgroup of special transformations in a given direction.
Theorem 4.2 The subgroup H(u) of G of the special transformations in the direction
u is given by all matrices of the form
Bu (v) = λ(v)
I3 + (a(v) − 1)uuT
−a(v)vu
T
ma(v)vu
a(v)(1 − lv)
,
(16)
where l, m are dimensional constants and m ≤ 0.
1) If m = 0, the domain of Bu is R if l = 0; and ] − ∞, c1[ or ] − c1 , +∞[, where
c1 := 1/|l|, if l 6= 0. The corresponding functions a(v) and λ(v) are, in the first case
(l = 0):
(
11
a(v) = ea1 v ,
λ(v) = eλ1 v ,
Of course in (12) B −1 stands for the inverse of the map, not the inverse of the matrix.
13
(17)
and in the second case (l 6= 0):
a(v) = (1 − lv)r1 ,
λ(v) = (1 − lv)r2 .
(18)
for arbitrary a1 , λ1 , r1 , r2 ∈ R.
2) If m < 0, let c0 > 0 such that m = −1/c20 . Then for arbitrary r1 , r2 ∈ R and
β = v/c0 ,
I = ] − c0 p, c0 p[
r
1 + pβ 1
1
(1 + pβ)r1 −1/2
p
a(v)
=
,
=
2
1 − pβ
(1 − pβ)r1 +1/2
r2 1 − 2ηβ − β
1 + pβ
,
λ(v) =
1 − pβ
p : = p1 + η 2 + η, p := p1 + η 2 − η, η := lc0 .
2
(19)
Moreover, if ΣS ∈ SO4 (3) is an admissible transition function, we have
ΣS Bu (v)ΣTS = BSu (v).
(20)
Proof Standard Lie group theory implies that H(u) is commutative (in fact the only
1-dimensional Lie groups are, up to isomorphism, S 1 – the unit circle group – and R).
The commutativity condition B(v1 )B(v2 ) = B(v2 )B(v1 ), is equivalent, given (14) 2 and
(14) 3 , to the following identities for all v1 , v2 :
1 − g(v2 )v1 v2 = 1 − g(v1 )v1 v2 , v1 + f (v1 )v2 = v2 + f (v2 )v1 .
From the first one it follows that g is constant (and we denote by m its value) and from
the second one it follows that l := 1−fv (v) (for v 6= 0) is another constant. It follows
that (cf. (13)) :
f (v) = 1 − lv, α̂(v) = a(v)(1 − lv).
(21)
By suitably exploiting the velocity addition law we shall now obtain stronger conditions
on the coefficients. In the new notation, (15) can be re-written as:
v1 ∗ v2 =
v1 + v2 − lv1 v2
,
1 − mv1 v2
(22)
where v1 , v2 range over I . Since (I, ∗) is a Lie group, the interval I must be invariant
under every left translation:
hv0 : I → I, v 7→ v0 ∗ v =
v0 + (1 − lv0 )v
,
1 − mv0 v
with v0 ∈ I . Now taking the derivative of hv0 we obtain:
14
(23)
h′v0 (v) =
mv02 − lv0 + 1
,
(1 − mv0 v)2
so the sign of h′v0 is constant (and nonzero), and since h′0 ≡ 1 > 0, by continuity it
follows that hv0 is a strictly increasing function mapping I onto itself for every v0 ∈ I .
In particular, the endpoints w1 , w2 must be fixed points for the continuous extension
to the ‘closure’ of I (possibly including ±∞) of every such hv0 , i.e.
lim hv0 (v) = w1 , lim hv0 (v) = w2 .
v→w2
v→w1
This means that w1 , w2 are either infinite (one of them or both) or (real) solutions of
the quadratic equation:
mv 2 − lv + 1 = 0.
(24)
Notice that, since 1 − lv = f (v) > 0 for all v ∈ I , m is necessarily negative or zero. If
m 6= 0, then we can introduce a positive constant c0 and write m = −1/c20 .
Let l = 0. If also m = 0, then it must be I = R because of the left invariance
under (23). If m 6= 0, then I =] − c0 , c0 [. Of course these conditions give the Galilean
and Lorentzian addition laws, respectively.
1) Let m = 0 and suppose that l 6= 0. According to whether l > 0 or l < 0, if
m = 0 then I =] − ∞, c1 [ or I =] − c1 , +∞[, respectively, where c1 = 1/|l|.
By substituting g(v) = m = 0 in (13) 2 we obtain immediately b = 0 and then from
(14) 1 and (14) 2 , for any l , we deduce the functional equations for λ and a:
a(v1 + v2 − lv1 v2 ) = a(v1 )a(v2 ),
(25)
λ(v1 + v2 − lv1 v2 ) = λ(v1 )λ(v2 ).
(26)
By partially differentiating (25) with respect to v1 and then putting v1 = 0, v2 = v ,
we obtain the first order differential equation:
da
−
dv
a′ (0)
1 − lv
a = 0,
from which it follows, if l = 0, a(v) = ea1 v , and if l 6= 0:
a(v) = (1 − lv)r1 , where r1 := −a′ (0)/l.
By the same argument for λ we get, respectively, λ(v) = eλ1 v if l = 0, and
λ(v) = (1 − lv)r2 , where r2 := −λ′ (0)/l
15
if l =
6 0.
2) Consider now m < 0; by solving (24) for m = −1/c20 , we get
I =] − c0 p, c0 p[, where p :=
p
1 + η 2 + η, p :=
p
lc0
1 + η 2 − η, η :=
.
2
(27)
Notice that η is a dimensionless parameter and that pp = 1; from (21) we have α̂(v) =
a(v)(1 − 2ηβ). By substituting the appropriate values in (14) we deduce the functional
equations for a and λ, respectively:
a(
v1 v2
v1 + v2 − lv1 v2
)
=
a(v
)a(v
)(1
+
),
1
2
1 + v1 v2 /c20
c20
(28)
v1 + v2 − lv1 v2
) = λ(v1 )λ(v2 ).
1 + v1 v2 /c20
(29)
λ(
By an argument perfectly similar to that used in 1) we get the first order differential
equation:
a′ (0) + v/c20
da
a = 0,
−
dv
1 − 2ηv/c0 − v 2 /c20
from which, after a few straightforward computations, the following explicit formula
follows:
r
1
1 + pβ 1
p
,
a(v) =
1 − pβ
1 − 2ηβ − β 2
′
(0)−η
. (The second equality in (19) 2 follows from the identity:
where β := v/c, r1 := ca p+p
2
1 − 2ηβ − β = (1 − pβ)(1 + pβ)). The same procedure applied to (29) leads us to:
λ(v) =
1 + pβ
1 − pβ
r2
, where r2 :=
c0 λ′ (0)
.
p+p
Finally, (20) is a direct consequence of (16).
By applying the results of the previous theorem to (15) we obtain:
(30)
Corollary 4.3 The velocity addition laws corresponding to the different values taken
by the constants m, l are:
v1 ∗ v2 =
v v
v1 + v2 − 2η
c0 1 2
v
v
1 + 1c2 2
0
v1 + v2 ± c11 v1 v2
v1 + v2
16
2η
1
,l = ,
2
c0
c0
for
m=−
for
1
c1
m = 0, l = 0.
for
m = 0, l = ∓
Remark From (29) it turns out that λ : (I, ∗) → (R+ , ·) is a group homomorphism,
and by composition with logarithm this gives a homomorphism from (I, ∗) to (R, +).
An isomorphism is obtained for every r 6= 0 as:
f : (I, ∗) → (R, +), v 7→ r ln
1 + pβ
.
1 − pβ
For instance, in the (conformal) Lorentzian case, i.e. for η = 0, we have (taking
r = 1) f (v) = tanh−1 (v/c0 ). Clearly, if one slightly modifies Axiom 10 by requiring the
existence of a ‘one-parameter subgroup’ (as standardly defined in Lie theory) of special
transformations, the ‘parameter’ should be s = f (v), not the speed itself. Similarly, in
(26) with l = 1/c1 , we should use s = ln(1 − v/c1 ) etc.
♦
4.1
Explicit formulae
Adapting (16) to the usual form for coordinate transformations we obtain, for the case
m = 0 (or c0 = +∞):
(
r′ = eλ1 v (r + (ea1 v (u · r − vt) − u · r)u)
t′ = ea1 +λ1 t,
(31)
if l = 0, and
(
r′ = (1 − lv)r2 (r + ((1 − lv)r1 (u · r − vt) − u · r)u)
t′ = (1 − lv)1+r1 +r2 t,
(32)
if l 6= 0. The transformations (31) (resp. (32)) reduce to the Galilean case
(
r′ = r − vtu = r − tV
t′ = t,
(33)
when λ1 = a1 = 0 (resp. l = 0).
In case u is the unit vector of the x1 -axis (u = e1 ), as in the usual presentation of
the special Lorentz transformations, we obtain what we call the generalized standard
special transformations for c0 = ∞, respectively for l = 0:
and for l 6= 0:
′1
x
x′2
x′3
′
t
= e(λ1 +a1 )v (x1 − vt)
= eλ1 v x2
= eλ1 v x3
= e(λ1 +a1 )v t,
17
(34)
′1
x
x′2
x′3
′
t
= (1 − lv)r1 +r2 (x1 + vt)
= (1 − lv)r2 x2
= (1 − lv) x3
(35)
= (1 − lv)1+r1 +r2 t.
As for the case c0 < ∞ we have
! !
r2
r1
pβ
pβ
1
+
u
·
r
−
vt
1
+
′
p
r+
−u·r u
r =
1 − pβ
1 − pβ
1 − 2ηβ − β 2
r1 +r2
pβ
(1 − 2ηβ)t − (v/c20 )u · r
1
+
′
p
,
t =
1 − pβ
1 − 2ηβ − β 2
(36)
which reduces to
! !
r2
r1
1
+
β
u
·
r
−
vt
1
+
β
′
p
r+
−u·r u
r =
1−β
1−β
1 − β2
r1 +r2
t − (v/c20 )u · r
1
+
β
′
p
,
t
=
1−β
1 − β2
(37)
when η = 0 (or, equivalently, l = 0). The generalized standard special trasformation
for c0 < ∞ is:
r1 +r2
1
+
pβ
x1 − vt
p
x′1 =
1 − pβ
1 − 2ηβ − β 2
r2
1 + pβ
x2
x′2 =
1
−
pβ
r2
1
+
pβ
x′3 =
x3
1 − pβ r1 +r2
(1 − 2ηβ)t − vx1 /c20
1 + pβ
′
p
,
t
=
1 − pβ
1 − 2ηβ − β 2
(38)
the inverse transformation being:
−(r1 +r2 )
pβ
(1 − 2ηβ)x′1 + vt′
1
+
1
p
x
=
1 − pβ
1 − 2ηβ − β 2
−r2
1 + pβ
x′2
x2 =
1 − pβ −r2
1 + pβ
x′3
x3 =
1
−
pβ
−(r1 +r2 )
t′ + vx′1 /c2
1 + pβ
p
.
t
=
1 − pβ
1 − 2ηβ − β 2
18
(39)
We can conclude that the special transformations, for a fixed direction and module of
the velocity, form a one-dimensional connected subgroup depending on 1) a constant c0
(finite or infinite) having the dimension of a velocity; 2) constants having the dimension
of the inverse of a velocity (l, a1 , λ1 ); 3) exponents (r1 , r2 ).
Remark Formulas (31)-(39) seem not to have appeared in print previously. However,
the functional equation approach (applied to a(v) only) was used by Frank and Rothe
in their 1911 treatment (with one space dimension), and in particular the expression
for a(v) (in particular (19)) can be found in their paper, with a different notation ([15],
pp. 850-4).
5
Spatial Isotropy and the Fundamental Theorem (Axiom 12)
The final axiom needed to arrive at the traditional bifurcation (‘either the Poincaré or
the Galilei group’) is:
Axiom 12 [spatial isotropy] The group G0 contains SO4 (3).
Equivalently, given Axiom 9, the rest subgroup of G coincides with the Newton group.
Because of (20), Axioms 11 and 12 together imply that a one-subgroup of special
transformations exists for every direction.
Let us list by their labels the axioms introduced so far: 1) space-time structure;
2) topology and differential structure; 3) causality; 4) time orientation; 5) spatial orientation; 6) mutual rest; 7) structure group; 8) physical velocities; 9) inertiality; 10)
abundance of cs’s; 11) special transformations; 12) spatial isotropy. The axioms that
mostly have to do with ‘relativity’ are (the second half of ) 9) and 12), which guarantee, respectively, that by space-time translations and by spatial rotations one stays
inside the space-time structure (cf. fn. 5), and 11), which ensures that, as in ‘Galilean’
physics, the space-time structure is not static.
We remember that all the matrices B in the (proper orthochronous) homogeneous
Galileo group (GG )0 can be obtained in the form:
B = ΣS1 BG (v)ΣS2 , where BG (v) =
I3 −ve1
0T
1
(40)
.
for all v ∈ R, and S1 , S2 ∈ SO(3), while all the matrices in the proper orthochronous
Lorentz group can be written in the form:
Λ = ΣS1 Λ(v)ΣS2 , where Λ(v) = Λ(ve1 ) =
I3 + (α − 1)e1 eT1 −αve1
− cα2 veT1
α
.
(41)
Factorizations (40) and (41) show both the structural kinship between the Galileo and
the Poincaré groups, and the similar role played as building blocks by standard special
transformations and spatial rotations. We can now state and prove:
19
Theorem 5.1 (Fundamental Theorem of the Theory of Relativity) Let Φ be a
space-time structure satisfying Axioms 1-12. Then the structure group G of Φ is either
the Galileo group or the proper orthochronous Poincaré group.
Proof Let φ, φ′ ∈ Φ be related by an affinity (Axiom 9) whose linear part is given by
the matrix B , with notation like in (6) (in particular VB ≡ V ). Because of Proposition
3.4 we can limit ourselves to dealing with G0 , that is, with matrices. By Axiom 12,
for every S ∈ SO(3) the map ΣS ◦ φ belongs to Φ; notice that the transition function
from ΣS ◦ φ to φ′ is, up to a translation, BΣS T , whose velocity is SV . It follows from
Axiom 10 and Proposition 3.5 that V is SO(3)-invariant and star-shaped.
Moreover, if Bu (v) is a subgroup of special transformations contained in G (Axiom 11),
then G must contain as well all matrices of the form ΣS1 Bu (v)ΣS2 for S1 , S2 ∈ SO(3);
in particular, for a suitable choice of S1 = S2T = S , we have, by (20), that also the
standard special transformations Be1 (v) are in G .
Now,12 if we take the following element of SO(3)
−1 0 0
S0 = 0 −1 0 ,
0
0 1
we see that the velocity of ΣTS0 Be1 (v)ΣS0 is −V = (−v, 0, 0). An easy examination of
the explicit formulas for the standard special transformations in the case m = 0 shows
that
ΣTS0 Be1 (v)ΣS0 = Be1 (−v)
(42)
applies to (34) (resp. (35)) if and only if a1 = λ1 = 0 (resp. l = 0); while (42) applies
to (38) if and only if η = r1 = r2 = 0. So we have, in the first case, Be1 (v) = BG (v),
and in the second case, Be1 (v) = Λ(v). Because of the characterizations (40) and (41),
we have that G contains, respectively, GG and P↑+ .
In order to show that in both cases not just inclusion but equality holds, we notice that
the only star-shaped open neighborhoods of 0 which are SO(3)-invariant are I) R3 and
II) the open balls of finite radius c0 centered in 0 (for every c0 > 0). Thus in both
cases all velocities in V are represented by matrices in GG and in P↑+ , respectively, and
Corollary 3.7 forbids any other matrices to be contained in the structure group.
Remark 1. In the proof of Theorem 5.1, the case in which c0 is a possible physical
speed and the case where it is just a supremum need not be separated. In fact at this
foundational level the hypothesis that there exist actual signals with speed c0 does not
make any difference. In particular the theorem, in itself, is perfectly compatible with
whatever assumption on this issue.
♦
2. An alternative route to the Galilei and Poincaré groups passes through the so-called
reciprocity principle, which many authors (e.g. [39], [49], [33], [40], [46], [10], [29], [30],
[1], [32]) introduced in order to obtain a further simplification after a transformation of
type (11) had been somehow (often not very transparently) arrived at.
12
What follows generalizes Poincaré’s argument in [35].
20
Let us call reciprocal velocity of the matrix B the velocity of B −1 , that is the velocity
of φ with respect to φ′ ; we denote it by W ≡ WB = VB−1 . The reciprocity principle
is the condition that the module of the velocity of φ′ with respect to φ be the same as
the module of the reciprocal velocity: |V| = |W|.
This requirement reduces to W = −V for special pairs (φ, φ′); however, for a general
inertial transformation (that is for arbitrary pairs linked, say, by a general Galilei or
Lorentz transformation) the latter version of ‘reciprocity’ does not hold, contrary to
what seems to be assumed in some treatments (see e.g. [50], p. 18; cf. [23], §5.1).
Now for special pairs we have:
W=−
a(V)
V,
α(V)
(43)
so (RP) reduces to the identity a = α. Thus (16) obeys the reciprocity principle if and
only if l = 0 (notice, however, that also (31) satisfies it). Under this condition in (38)
we have p = p = 1, and if we put r1 = 0, r2 = s we obtain the transformation:
s
1+β
′1
x =
1−β
s
1+β
′2
x =
1 − β s
1+β
x′3 =
1 − β s
1+β
′
t =
1−β
x1 − vt
p
1 − β2
x2
(44)
x3
t − vx1 /c2
p
.
1 − β2
This one-dimensional subgroup (for every fixed s ∈ R) is contained in the conformal
Poincaré group. It has a common null eigenvector and it can be shown to be contained
in a 8-dimensional subgroup G of the conformal Poincaré group with this property.
Group G has been the object of a series of articles by Bogoslovsky, starting in 1977
([7]). The Lie algebra of its linear factor (denoted by G0 in [23]), is generated by a basis
(σ0 , σ1 , σ2 , σ3 ) verifying the identities:
[σ0 , σ1 ] = 0 [σ0 , σ2 ] = σ3 , [σ0 , σ3 ] = −σ2 , [σ2 , σ3 ] = 0, [σ1 , σ2 ] = −σ2 , [σ1 , σ3 ] = −σ3 .
Here σ0 corresponds to a 1-parameter subgroup of spatial rotations and σ1 to the only
one-dimensional subgroup of special Lorentz transformations. It follows that G has
a 7-dimensional subgroup (where the Lie algebra of its linear factor is generated by
(σ0 , σ2 , σ3 )), which violates both Axioms 10 and Axiom 11. In particular, this subgroup
is an example of a non-static structure group which does not admit a one-dimensional
subgroup of special transformations.
A similar kind of anisotropy can be introduced in the classical setting, thus defining
an anisotropic Galilei group. A unified approach and a discussion of the properties
mentioned here, and others, of these anisotropic groups is contained in [23].
21
6
Two-way light isotropy (Axiom 6∗ )
The class of theories in which both spatial isotropy and reciprocity fail is particularly
interesting from the viewpoint of the conventionality of simultaneity issue ([22]). Let us
introduce a new axiom, which is also the first one which endowes the limiting, directiondependent speeds, with a concrete physical meaning, that of being the speeds of physical
signals (“light”).
Axiom 6∗ The two-way isotropy of the speed of light holds, with a constant c.
The two-way isotropy of the speed of light may be concisely expressed by the equation (cf. §2 in [22]) :
1
2
1
+
= .
c+ c−
c
(45)
Conventionalists hold that only the round-trip, or two-way, isotropy of light is a
law with an unambiguous physical content, and that the standard one-way version
of the law adds nothing to the former except for the choice of a convention – one
among infinitely many, justified essentially by formal expediency ([22],[24]). Now, from
the conventionalist viewpoint also the assumption that the transition functions among
admissible cs’s form a group (Axiom 7) must be classified as a convention of the same
kind. In fact if we adopt the round-trip law (instead of the ‘classical’ one-way law) as
our theory’s cornerstone, the space-time structure Φ does not come out naturally as
a group orbit, but as a union of infinitely many group orbits with respect to different
subgroups of Bi(R4 ), as we are going to see.
We shall now derive and briefly discuss the transformation between inertial cs’s in
which the round-trip isotropy of light is verified. The derivation has been done with
varying degrees of generality and explicitness by several authors (e.g. [50], [37], [2], [38]).
The following derivation strikes me as both simpler and more logically transparent than
others.
Recall that an inertial basis in R4 with the usual Lorentz structure:
gc : R4 × R4 → R, (x, y) 7→ x1 y 1 + x2 y 2 + x3 y 3 − c2 x4 y 4 ≡ xT Gc y,
is a basis b = (w1 , w2 , w3 , w4) where w1 , w2 , w3 are space-like and w4 ≡ u is a future
timelike vector (cf. [22], p. 788). An inertial cs is an affine coordinate system whose
associated basis is inertial. Let us start reformulating, for ease of reference, Proposition
3.13 of [24]:
Proposition 6.1 In Minkowski space-time, for every inertial cs φ such that the isotropy
of the two-way velocity of light holds, there exists a unique Minkowski coordinate system
φ such that the transition function from φ to φ is of the form:
x = λKx with λ > 0, K =
I3 0
kT 1
, |k| < 1/c.
(46)
Vice versa every φ related to a Minkowskian φ by (46) satisfies the two-way isotropy
of the velocity of light.
22
Let φ and φ′ be inertial cs’s for which the isotropy of the two-way speed of light
holds; we want to find the transition function φ′ ◦ φ−1 .
With reference to the previous proposition, we have for λ1 , λ2 > 0, x = λ1 K1 x, x′ =
′
λ2 K2 x′ , with K1 , K2 like in (46), where φ, φ are Minkowskian cs’s systems. Let Λ be
′
the Lorentzian matrix giving the linear part of the transition function from φ to φ . It
follows that
x′ = λBx + b, where B = K2 ΛK1−1
(47)
where b = (b, b4 ) ∈ R4 . Taking account of the form of the Lorentz matrices:
Λ=
A
−AV
− cα2 VT
α
, with AT A = I3 +
α2
VVT , α = (1 − β 2 )−1/2 ,
c2
(48)
′
where V is the velocity of φ with respect to φ, we have (with λ = λ2 /λ1 ):
B=λ
I3 0
kT2 1
A
−AV
α
T
− c2 V A
α
I
0
−k1 1
.
A straighforward computation leads to the following:
Proposition 6.2 The coordinate change between two cs’s obeying Axioms 1-5 and 6∗
is
(
r′ = λA((I3 + VkT1 )r − tV) + b
t′ = λ((AT k2 −
α
V
c2
− (α − kT2 AV)k1) · r + (α − kT2 AV)t) + b4 ,
(49)
with λ > 0, b ∈ R4 , V ∈ B(0, c), k1 , k2 ∈ B(0, 1/c) and A satisfying (48). System (49)
defines a set F of transformations depending on 17 independent parameters.
The set F of all transformations preserving equation (45) is not a group (under
ordinary map composition), although the inverse of every transformation in F is still in
F . This shows that loosely framed claims such as “the set of all one-one transformations
leaving [unchanged] any set of equations forms a group” must be handled with some
care.13 So this gives an example of a theory of relativity (Axioms 1-5) where the basic
physical law which is required to hold in all admissible cs’s does not produce in a natural
way a structure group for the theory.
By putting in (49):
V = (v, 0, 0), k1 = (a1 , 0, 0), k2 = (a2 , 0, 0),
13
The statement, italicized in the original, is quoted from [5], p. 100.
23
and by requiring furthermore that A be diagonal and with positive entries, it follows
by (48) that A = diag(α, 1, 1), and therefore:
′1
x
x′2
x′3
′
t
=
=
=
=
λα((1 + a1 v)x1 − vt)
λx2
λx3
λα((a2 − v/c2 − (1 − a2 v)a1 )x1 + (1 − a2 v)t).
(50)
The values for the free parameters which give the special Lorentz transformations are,
clearly, a1 = a2 = 0, λ = 1. These special transformations satisfy Axiom 6 if and only
if a1 = a2 , λ = 1 (v = 0 implies α = 1), so in general even this axiom is not satisfied.
Remark If we renounce Axiom 9, in Proposition 6.1 we can substitute the linear term
k · r with any function. More precisely, for a fixed Minkowski coordinate system φ any
coordinate transformation
(51)
r = λr, t = λ(t + g(r)),
where g : R3 → R is an arbitrary differentiable function such that |∇g| < 1/c (for
instance take g(r) := (sin x1 )/2c), defines a coordinate system φ which satisfies the
two-way light isotropy. This is easily proved (cf. [2], pp. 131-2). Suppose that Γ is any
lightlike curve and its projection C in the 3-space of φ is closed; then, if r : I → R3
represents C in φ, we have |v| ≡ c and therefore
I
C
ds
1
=
|v|
c
1
(1 + ∇g · v)ds =
c
C
I
I
ds +
C
I
∇g · ds =
ℓ(C)
,
c
where ℓ(C) is the length of C .
6.1
Absolute simultaneity
The absolute simultaneity condition (cf. [22], §9) is the requirement of proportionality
between the time coordinates, i.e. t′ ∝ t; for (49) this is equivalent to:
AT k2 −
α
V = (α − kT2 AV)k1 ,
c2
and the corresponding sub-family of (49) becomes:
r′ = λA((I + VkT1 )r − tV) + b
t′ = λ(α − kT2 AV)t + b4 .
If we scalarly multiply both sides of (52) by V , we get, after a few passages:
kT2 AV = α
β 2 + k1 · V
,
1 + k1 · V
and by substituting this into the time coordinate equation we get:
24
(52)
r′ = λA((I + VkT1 )r − tV) + b
λt
+ b4 ,
t′ =
α(1 + k1 · V)
(53)
which is a set of transformations – not a group, again – depending on 14 parameters.
In case φ = φ, we have k1 = 0 and therefore
(
r′ = λA(r − tV) + b
α
λt
, with AT A = I3 + 2 VVT ,
4
′
t =
+b .
c
α
(54)
which is the most general transformation from a Minkowskian cs to an inertial cs satisfying both the round-trip isotropy and the absolute simultaneity conditions. The “special”
version of (54) is:
′1
x
x′2
x′3
t′
= λα(x1 − vt),
= λx2 ,
= λx3 ,
λt
=
.
α
(55)
Notice that while |v| < c, the corresponding speed of light in a φ′ can be as big as
desired.
This set of transformations was discussed by several authors, particularly by Tangherlini ([41], (1.16), p. 9); but it is surely the three-part article by Mansouri and Sexl
([25]) which must be credited for its more recent revival (cf. [22], §9).
6.2
Two-way light isotropy and the group axiom
We have seen that Axiom 6∗ potentially conflicts with Axiom 7. Nevertheless it makes
sense to inquire whether one can enforce Axiom 6∗ along with Axiom 7, that is whether
there exist nontrivial invariance groups for the law of two-way constancy of the limiting
velocity (‘nontrivial’ means here violating Axiom 12). In fact there are infinitely many
such groups, as we are going to see in a moment.
Suppose G is a structure group of a space-time (E, Φ) satisfying Axiom 9, so that in
particular G = G0 ⋊ T (R4 ), and every B ∈ G0 is of the form (47). Let φ ∈ Φ be fixed,
so that Φ = G · φ; let us pick any other φ1 ∈ Φ. If Φ satisfies Axiom 6∗ with respect
to a fixed Minkowski structure Φ, we have from (47) that
B := (φ1 ◦ φ−1 )0 = λK1 ΛK −1 ,
where K can be considered as fixed, while λ > 0, Λ ∈ L+
↑ and K1 all depend on φ1 .
A simple computation shows that
VB =
V
,
1−k·V
25
(56)
which shows that Axiom 10 is satisfied if all V ∈ B(0, c) are allowed in the Lorentzian
factor. Suppose the conformal factor λ is set equal to 1. The following theorem holds.
Theorem 6.3 Let G0 be a subgroup of GL(4, R) with all elements of the form K ′ ΛK −1
with K ′ varying in some nonempty set and K fixed, both K, K ′ of the form (46), and
suppose that Λ varies in a subset X ⊆ L+
↑ such that
X → B(0, c), Λ 7→ VΛ
−1
is onto. Then X is a subgroup of L+
.
↑ and G0 = KX K
Proof It is easy to verify that the set X in the statement must be a subgroup of L+
↑ .
Consider now any product of the form:
(K2 Λ2 K −1 )(K1 Λ1 K −1 ) = K3 Λ3 K −1 ,
or, equivalently:
−1
−1
K4 Λ 2 K5 = Λ ∈ L +
K1 , B := Λ3 Λ−1
1
↑ , with K4 := K3 K2 , K5 := K
where K4 , K5 have the same form as in (46), except, possibly, for the condition on the
upper bound on |k|. Our claim is that K5 = I4 , or, equivalently, k5 = 0 (that is,
k1 = k). A simple computation, taking into account that Λ and Λ2 are of the form
(48) gives:
Λ=
A2 (I3 − V2 kT5 )
−A2 V2
α
(AT2 k4 − c22 V2 + (α2 − k4 · A2 V2 )k5 )T α2 − k4 · A2 V2
,
therefore
V ≡ VΛ =
V2
.
1 − k5 · V2
(57)
Since both Λ and Λ2 must satisfy the identities in (48), we have:
α22
1
V2 V2T − (V2 kT5 + k5 V2T
c2
+
V22 k5 kT5 )
1
= 2
c
α2 − k4 · V2
1 − k5 · V2
2
V2 V2T .
We can choose V2 arbitrarily in B(0, c), so let us take it as a nonzero vector orthogonal
to k5 , and multiply scalarly both sides by k5 ; we obtain:
|V2 |2 |k5 |2 k5 = |k5 |2 V2 ,
and since V2 ⊥ k5 , it must be k5 = 0 as required.
26
One may call any maximal group with the linear factor of the form described in
Theorem 6.3 a Reichenbach group if k 6= 0:
−1
GR (k) := KL+
⋊ T (R4 )
↑K
(58)
The corresponding Reichenbach function is (cf. [22], p. 791):
1
r
ǫ(r) = (1 + ck · ).
2
r
In terms of the Reichenbach groups we can express the spacetime structure Φ as
Φ=
[
k∈B(0,1/c)
R+ · GR (k) · Kφ
(59)
with φ any fixed cs in the Minkowski structure.
The linear factor GR (k)0 is the invariance group of the Lorentzian matrix:
G = (K
−1 T
) Gc (K
−1
)=
I3 − c2 kkT c2 k
c2 kT
−c2
,
and the corresponding space-time metric is:
ds2 = drT (I3 − c2 kkT )dr + 2c2 dr · kdt − c2 dt2 .
Therefore the set V of relative velocities of admissible cs’s – i.e. those cs’s belonging to
a fixed class Φk having GR (k) as structure group (cf. Proposition 3.5) – is
v
1
T
G
v
1
= vT (I3 − c2 kkT )v + 2c2 k · v − c2 < 0,
and its boundary is what may be called the velocity ellipsoid of GR (k):
|v|2 − c2 (1 − k · v)2 = 0.
(60)
In order to study this surface suppose k = ke1 . The equation becomes
(v 1 )2 + (v 2 )2 + (v 3 )2 − c2 (1 − kv 1 )2 = 0,
which is an ellipsoid with a 1-parameter group of rotations
√ around the bigger semi-axis
2 2
c/(1 − c k ), the other two semi-axes being equal to c/ 1 − c2 k 2 . Since the centre of
the ellipsoid is (−c2 k/(1 − c2 k 2 ), 0, 0), the interval of the speeds in the x1 -direction is
asymmetric with respect to 0:
c] −
1
1
,
[,
1 − ck 1 + ck
27
while in both the v 2 and the v 3 -directions (i. e. in the directions orthogonal to k) it
is the standard ] − c, c[.
The Reichenbach groups form a family of Lie groups parametrized by k, as k varies
in the open set 0 < |k| < 1/c in R3 . They are all isomorphic to one another, and
they are all isomorphic to the Poincaré group; explicitly: GR (k) = KP↑+ K −1 . The
Reichenbach groups satisfy Axiom 11 for the direction u = vers(k). In fact we have:
KΛ(vu)K
−1
=
I3 + (α(1 + kv) − 1)uuT
−αvu
−α( c12 − k 2 )vuT
α(1 − kv)
,
which verifies definition 4.1. Notice however that Axiom 6 is not satisfied: in fact for
every S ∈ SO(3) matrix B := KΣS K −1 has zero velocity (VB = 0), since
B=
S
0
T
k (S − I3 ) 1
,
therefore Axiom 6 would be satisfied if and only if k = 0, which would take us back
to the Poincaré group. Equivalently, the rest subgroup of GR (k) is KGN K −1 , which is
not contanined in GN unless k = 0.
Physically, to take a Reichenbach group for some k as the structure group of
physics corresponds to having a relativistic physics in which the privileged cs’s φ have
been all synchronized by choosing the same Reichenbach function ǫ with respect to
a Minkowskian cs depending on φ. As a consequence none of the elements of Φk is
Minkowskian: indeed, there is a unique Minkowskian structure Φ related to Φk by
Φ = K −1 Φk
for a suitable K . It must be stressed, however, that this construction goes against the
spirit of the Reichenbach’s approach, since his ǫ-argument was advanced only to show
that every single inertial notion of rest (denoted as Γ(u) in [22, 24]) could arbitrarily
choose its own synchrony with considerable latitude, irrespective of what other notions
of rest did. As we have seen (cf. (59)), the genuine mathematical counterpart of
Reichenbach’s view is to take the space-time structure Φ as a union of orbits of infinitely
many groups. The main interest of a theory based on GR (k) is that it is a theory of
relativity satisfying the group axiom and where the law of light propagation holds only in
the two-way version. We can re-state the conventionalist position concerning relativity
as holding that there is no physical fact of the matter which allows us to distinguish
between any of the Reichenbach groups and the Poincaré group. I have explained in [22]
and [24] why I think this claim cannot be maintained except in a rather uninteresting
sense.
7
Appendix – Proof of Theorem 3.3
It is easy to verify that if φ, φ′ are affinely equivalent cs’s, then they are also inertially
equivalent.
28
Let us consider the converse. Suppose that every uniform worldline for φ is also
uniform for φ′ . Since points at rest for φ are physical worldlines (Axiom 3), we have
for every r0 ∈ R3 :
∂r′
∂t′
(r0 , t) =
(r0 , t)W
∂t
∂t
(61)
with W depending at most on r0 . It follows that the transition function from φ to φ′
can be re-written as
r′ = t′ (r, t)W(r) + U(r)
t′ = t′ (r, t).
where U : R3 → R3 is a suitable differentiable function.
Claim 1: W is constant.
In fact let r(t) = r0 + (t − t0 )v be any uniform motion for φ with v physical and
R as its domain. If we put f (t) := t′ (r0 + (t − t0 )v, t) it is easy to see that f is a
diffeomorphism of R onto itself. (Let v′ the constant velocity of the same worldline
according to φ′ , and suppose that the interval f (R) had, say, a finite supremum b;
then, if φ′ ◦ φ−1 (r0 , t0 ) = (r′0 , t′0 ) the subset
φ ◦ φ′−1 ({(r′0 + (t′ − t′0 )v′ ) : t′ ∈ [t′0 , b]})
would have to be noncompact, which is absurd.) Now
v′ = W +
1
(f (v · ∇)W + (v · ∇)U)
f˙
(62)
must not depend on t. Putting t = t0 we get that there is a function v′ = K(r0 , v)
such that K(r(t), v) ≡ K(r0 , v), and
f˙(t0 ) (K(r0 , v) − W(r0)) − f (t0 )((v · ∇)W)r0 = ((v · ∇)U)r0 .
For every fixed r0 and physical velocity v this is an ordinary differential equation in
f (t) (after notation change from t0 to t) with vector coefficients:
f˙A + f B = C, where A := K − W, B := −(v · ∇)W, C := (v · ∇)U.
Now, if for all r0 we have B = 0 when we choose v in three linearly independent
directions, then ∇W = 0 and therefore W is constant. But for B there is no other
possible value. Suppose by contradiction, that B 6= 0 for some r0 and v. Then f must
be a solution of a (scalar) differential equation with constant coefficients of the form
af˙ + bf = c with b 6= 0. Now a 6= 0, since otherwise f would be constant; it follows
that f is of the type f (t) = ke−bt/a + c/b, which clearly is not onto R, no matter what
the values of a, b, c are. So claim 1 is proven.
29
It follows that r′ (r, t) = t′ (r, t)W + U(r) with W constant. Note that U(r) cannot
be constant, otherwise
det
∂r′
∂r
= det(W(∇t′ )T ) = 0,
since every 3 × 3 matrix of the form abT is singular, which would contradict Axiom 5.
Thus (62) becomes:
K(r0 , v) = W +
1
(v · ∇)U.
f˙
(63)
By comparing the functional dependence of the various terms of this equation we get
that f˙ must be independent of t; since
∂t′
f˙ = (∇t′ )(r0 , t) · v +
(r0 , t)
∂t
∂t′
(put v = 0), nor ∇t′ may depend on t. It follows that
∂t
t′ = g(r) + αt for some α > 0 and g : R3 → R differentiable.
Claim 2: Both functions U and g are affine.
Equation (63) can be re-written as
this means that neither
K(r0 , v) = W + (∇g(r0) · v + α)−1 ((v · ∇)U)(r0 ).
By substituting r0 with r(t) and differentiating with respect to t, we obtain zero; if we
neglect the irrelevant denominator and evaluate for t = t0 we get:
3
X
∂2U
∂2g
∂U
γ
(∇g(r0 ) · v + α) α β (r0 ) − α β (r0 ) γ (r0 )v v α v β .
0=
∂x ∂x
∂x ∂x
∂x
α,β=1
By re-arranging the terms we have:
3
X
∂g ∂ 2 U
∂ 2 g ∂U
∂2U α β
v v =−
− α β γ vαvβ vγ .
α
α
β
γ
α
β
∂x ∂x
∂x ∂x ∂x
∂x ∂x ∂x
α,β,γ=1
α,β=1
3
X
Now, for every fixed r0 at the lefthand side there is a quadratic vector polynomial
in the v α , while at the righthand side there is a cubic polynomial – both homogeneous
with respect to v α : this is possible, for every v in an open neighborhood of 0 (Axiom
8), if and only if both polynomials vanish for every r0 , that is, if all their coefficients
are identically zero:
∂2U
∂ 2 g ∂U
= 0,
= 0.
∂xα ∂xβ
∂xα ∂xβ ∂xγ
30
(64)
From the first equality we have that U is an affine map: U(r) = Xr + a, so we can
write
′
r = (g(r) + αt)W + Xr + a,
t′ = g(r) + αt + ℓ.
Since, as we have seen, U is nonconstant, at least one of the vectors
thus the second equation in (64) implies
∂U
is nonzero;
∂xγ
∂2g
≡ 0,
∂xα ∂xβ
which means that also g is an affine map, and Claim 2 is proven; this ends also the
proof of the theorem.
References
[1] H. Almström: “Derivation of the Lorentz transformation without use of light", J. Phys. A
(Proc. Phys. Soc.), ser. 2, 1, 1968, 331-3.
[2] R. Anderson, I. Vetharaniam, G. E. Stedman: “Conventionality of Synchronization, Gauge
Dependence and Test Theories of Relativity", Physics Reports, 295, 1998, 93-180.
[3] H. Bacry, J-M Lévy-Leblond: “Possible kinematics", J. Math. Phys., 9, 1968, 1605-14.
[4] V. Berzi, V. Gorini: “Reciprocity principle and the Lorentz transformations", J. Math.
Phys., 10, 1969, 1518-24.
[5] G. Birkhoff 1960: Hydrodynamics - A study in Logic, Fact and Similitude, Revised Edition,
Princeton University Press.
[6] H. Blaine Lawson, Jr. : The Theory of Gauge Fields in Four Dimensions, Providence
(Rhode Island), American Mathematical Society 1985.
[7] G. Yu. Bogoslovsky: “A Special-Relativistic Theory of the Locally Anisotropic SpaceTime. I: The Metric and Group of Motions of the Anisotropic Space of Events", Il Nuovo
Cimento, 40 B, 1977, 99-115; “II: Mechanics and Electrodynamics in the Anisotropic
Space", 40 B, 1977, 116-34.
[8] C. Cattaneo: “Sui postulati comuni alla cinematica classica e alla cinematica relativistica",
Lincei - Rend. Sc. fis. mat. e nat., 24, 1958, 526-32.
[9] R. D’Inverno: Introducing Einstein’s Relativity, Cambridge University Press, 1992.
[10] E. Drake: “Deduction from a Kinematic Principle of Relativity", Am. J. Phys., 34, 1966,
899-900.
[11] A. Einstein: “Zur Elektrodynamik bewegter Körper", Annalen der Physik, 17, 1905, 891921.
[12] A. Einstein: “Kosmologische Betrachtungen zur allgemeinen Relativitätstheorie", Preussische Akademie der Wissenschaften, Sitzungsberichte, 1917, P. 1, 142-52.
31
[13] L. J. Eisenberg: “Necessity of the Linearity of Relativistic Transformations between Inertial
Systems", Am. J. Phys., 35, 1967, 649.
[14] V. Fock: The Theory of Space, Time and Gravitation, New York, Pergamon Press, 1959
[15] P. Frank, H. Rothe: “Ueber die Transformation der Raum-Zeit Koordinaten von ruhenden
auf bewegte Systeme", Ann. Phys., 34, 1911, 825-55.
[16] P. Frank, H. Rothe: “Zur Herleitung der Lorentztransformation", Physik. Zeitschr., 13,
1912, 750-3.
[17] L. R. Gomes: “Sur la déduction des formules de Lorentz", Lincei - Rend. Sc. fis. mat. e
nat., 21, 1935, 433-7.
[18] J. Jeffers: “Lost Theorems of Geometry”, The American Mathematical Monthly, 107,
2000, 800-812.
[19] V. Lalan: “La cinématique et la théorie des groupes”, Compt. Rend., 203, 1936, 1491-3.
[20] E. Le Roy: “Sur les formules de Lorentz", Compt. Rend., 202, 1936, 794-5.
[21] J.-M. Lévy-Leblond: “One more derivation of the Lorentz transformation", Am. J. Phys.,
44 (3), 1976, 271-7.
[22] M. Mamone-Capria: “On the Conventionality of Simultaneity in Special Relativity",
Found. Phys., 31, 2001, 775-818.
[23] M. Mamone-Capria: “Spatial Directions, Anisotropy and Special Relativity”, Found. Phys,
31, 2011, 1375-97.
[24] M. Mamone-Capria: “Simultaneity as an Invariant Equivalence relation”, Found. Phys.,
42, 2012, 1365-83.
[25] R. Mansouri, R. Sexl: “A Test Theory of Special Relativity: I. Simultaneity and Clock
Synchronization", General Relativ. Gravit., 8, 1977, 497-513.
[26] N. D. Mermin: “Relativity without light”, Am. J. Phys., 52, 1984, reproduced in [27],
247-265 (with a Postscript).
[27] N. D. Mermin N. D.: Boojums All The Way Through, Cambridge University Press, 1990.
[28] Y. Mimura, T. Iwatsuki: “On the Linearity of the Lorentz Transformation", Journal of
science - Hiroshima University, A1, 1931, 111-6.
[29] V. Mitvalský: “Special Relativity without the Postulate of Constancy of Light", Am. J.
Phys., 34, 1966, 825.
[30] C. Moller: The Theory of Relativity, 2nd edition, Oxford University Press, 1972.
[31] V. V. Narliker: “The restriction to linearity of the Lorentz transformation", Proc. Camb.
Phil. Soc., 28, 1932, 460-2.
[32] P. B. Pal: “Nothing but relativity”, European Journal of Physics, 24, 2003, 315-9.
[33] L. A. Pars: “The Lorentz Transformation", Phil. Mag., 42, 1921, 249-58.
[34] W. Pauli: Theory of Relativity, Oxford, Pergamon Press, 1958 (New York, Dover, 1981).
32
[35] H. Poincaré: “Sur la dynamique de l’électron", Rend. Circ. Matem. Palermo, 21, 1906,
129-76.
[36] H. M. Schwartz: “Axiomatic Deduction of the General Lorentz Transformations", Am. J.
Phys., 30, 1962, 697-707.
[37] F. Selleri, “Noninvariant one-way velocity of light,” Found. Phys., 5, 1996, 641-64.
[38] F. Selleri, “Teorie alternative alla relatività e natura del tempo,” in La scienza e i vortici
del dubbio, L. Conti and M. Mamone Capria, eds. (ESI, Naples, 1999), pp. 213-47.
[39] F. Severi: “The principles of the relativity theory deduced from the common sense", Proc.
Phys-Math. Soc. Japan, 18, 1936, 257-67.
[40] K. D. Stiegler: “Sur le principe de la constance de la vitesse de la lumière", Compt. Rend.,
234, 1952, 1250-2.
[41] F. R. Tangherlini: “An Introduction to the General Theory of Relativity", Nuovo Cimento,
Supplemento, 20, ser. X, 1961, 1-86.
[42] Ya. P. Terletskii: Paradoxes in the Theory of Relativity, New York, Plenum, 1968.
[43] W. A. von Ignatowski: “Einige allgemeine Bemerkungen sum Relativitätsprinzip", Verh.
Desutsch. Phys. Ges., 12 1910, 788-96.
[44] W. A. von Ignatowski: “Einige allgemeine Bemerkungen sum Relativitätsprinzip", Physik.
Zeitschr., 11, 1910, 972-6.
[45] R. Weinstock: “Derivation of the Lorentz transformation equations without a linearity
assumption", Am. J. Phys., 32, 1964, 260-4.
[46] R. Weinstock: “New Approach to Special Relativity", Am. J. Phys., 33, 1965, 640-5.
[47] H. Weyl: Space-Time-Matter, 4th German edition, 1922; transl. New York, Dover, 1952.
[48] G. J. Whitrow: “A derivation of the Lontz formulae”, Quart. J. Math., 4, 1933, 161-72.
[49] E. Whittaker: From Euclid to Eddington, Cambridge University Press, 1949.
[50] Y. Z. Zhang: Special Relativity and Its Experimental Foundations, Singapore, World
Scientific, 1997.
33
| 4math.GR
|
SLAC: A Sparsely Labeled Dataset for Action Classification and Localization
Hang Zhao1∗, Zhicheng Yan2∗, Heng Wang2 , Lorenzo Torresani2,3 , Antonio Torralba1
1
Massachusetts Institute of Technology, 2 Facebook Applied Machine Learning, 3 Dartmouth College
arXiv:1712.09374v1 [cs.CV] 26 Dec 2017
{hangzhao, torralba}@mit.edu, {zyan3, hengwang, torresani}@fb.com
Abstract
enabled the successful application of progressively more sophisticated learning models. In image categorization, we
moved from the Caltech101 [6] which was introduced in
2004 and contained only 9,146 examples, to the ImageNet
dataset [3], which was introduced in 2011 and includes over
1.2M examples. In object detection, we have seen a similar
trend in scaling-up training set sizes, despite the additional
human annotation cost involved in collecting bounding-box
information. Pascal VOC [4] was first released in 2007 and
contained only 1,578 examples, while the recently introduced COCO dataset [17] consists of over 200K images and
500K object instance annotations.
This paper describes a procedure for the creation of
large-scale video datasets for action classification and localization from unconstrained, realistic web data. The
scalability of the proposed procedure is demonstrated by
building a novel video benchmark, named SLAC (Sparsely
Labeled ACtions), consisting of over 520K untrimmed
videos and 1.75M clip annotations spanning 200 action categories. Using our proposed framework, annotating a clip
takes merely 8.8 seconds on average. This represents a saving in labeling time of over 95% compared to the traditional
procedure of manual trimming and localization of actions.
Our approach dramatically reduces the amount of human
labeling by automatically identifying hard clips, i.e., clips
that contain coherent actions but lead to prediction disagreement between action classifiers. A human annotator
can disambiguate whether such a clip truly contains the hypothesized action in a handful of seconds, thus generating
labels for highly informative samples at little cost. We show
that our large-scale dataset can be used to effectively pretrain action recognition models, significantly improving final metrics on smaller-scale benchmarks after fine-tuning.
On Kinetics [14], UCF-101 [30] and HMDB-51 [15], models pre-trained on SLAC outperform baselines trained from
scratch, by 2.0%, 20.1% and 35.4% in top-1 accuracy, respectively when RGB input is used. Furthermore, we introduce a simple procedure that leverages the sparse labels in SLAC to pre-train action localization models. On
THUMOS14 [12] and ActivityNet-v1.3[2], our localization
model improves the mAP of baseline model by 8.6% and
2.5%, respectively.
In the video domain, the gap in scale between datasets
for action classification and those for action localization
has been widening. While action classification datasets created a few years ago consisted of a few thousands examples (6849 videos in HMDB51 [15], 13K in UCF101 [30],
3669 in Hollywood2 [18]), recent benchmarks have scaledup dataset sizes by up to two orders of magnitude
(Sports1M [13] contains over 1M videos, and Kinetics [14]
has 306K videos). But we have not seen a similar growth
in datasets for action localization. THUMOS [12] was created in 2014 and contained 2.7K trimmed videos and 1K
untrimmed videos with localization annotations. Today, the
largest benchmarks for action localization are only moderately larger. For example, ActivityNet [2] includes 20K
videos and 30K annotations, AVA [19] includes 58K clips,
and Charade [28] contains 67K temporally localized intervals. In Table 1, we give detailed comparisons between different video datasets.
Why are action localization datasets still so much smaller
than those for object detection? And, why are action localization datasets still more than one order of magnitude
smaller than those for action categorization? In this paper
we put forward the hypothesis that this is primarily due to
two reasons. First, gathering temporal annotations on video
is very time-consuming. According to our experiments with
professional annotators, manually annotating the start and
end of action segments in a video takes more than 4 times
the length of the video. In order to give accurate temporal annotations, annotators need not only to watch the entire
1. Introduction
The fields of image categorization and object detection have witnessed parallel outstanding progress in recent
years [11, 10]. Arguably, these advances are the result of
steady improvements and growths in datasets, which have
*Hang and Zhicheng contribute equally to this work.
1
Datasets
UCF101[30]
HMDB51[15]
Kinetics[14]
Sports1M[13]
THUMOS 15[12]
ActivityNet[2]
Charades[28]
AVA[19]
SLAC
Actions
101
51
400
487
101
200
157
80
200
Videos
13K
7K
305K
1.1M
24K
20K
10K
214
520K
Annotations
13K
7K
305K
1.1M
21K
23K
67K
197K
1.75M
Source
YouTube
YouTube/Movie
YouTube
YouTube
YouTube
YouTube
267 Homes
Movie
YouTube
Localization
No
No
No
No
Yes
Yes
Yes
Yes
Yes
Table 1. Comparisons between SLAC and other video datasets. Note that the annotations from Sports1M are produced automatically by
analyzing the text metadata surrounding the videos, and thus inaccurate.
only happen for a short period of time during the whole
video. This poses substantial difficulties in learning good
video representations for action classification as further discussed in Section 4. Compared to Kinetics, SLAC contains nearly 6 times more clip annotations (1.75M vs 305K),
which may explain the superior generalization performance
of deep models trained on our benchmark.
Finally, we demonstrate that the sparse clip annotations
in SLAC can also be used to pre-train action localization
models, which give dense predictions at each frame. On
THUMOS Challenge 2014 and ActivityNet-v1.3 datasets,
we are able to improve baseline models by 8.6% and 2.5%
in mAP.
The rest of the paper is organized as follows. We review past work on video dataset collection in Section 2. The
SLAC dataset and its collection procedure are elaborated in
Section 3. We evaluate the merits of the SLAC dataset for
action classification in Section 4 and for action localization
in Section 5. We conclude our work in Section 6.
sequence, but also to replay several parts of the video back
and forth to find exact boundaries. Second, marking temporal action boundaries is often ambiguous. While object
boundaries are defined by their physical extents, the action
boundaries are frequently blurry as a result of the smooth
continuity of human movements and the poor definition of
what constitutes an action.
In this work we propose to recast the temporal annotation task into a form that is more efficient and less ambiguous. The idea is to sample a small number of short clips
from each video for annotators to review. An active learning algorithm is used to select one easy clip and several hard
clips to label. The annotator is then asked to simply confirm
whether such clips truly contain the hypothesized action.
Our experiments suggest that providing a binary yes/no answer is easy and fast for annotators. This induces a saving in
labeling time of over 95% compared to the traditional procedure of exhaustive review of the entire video for manual
marking of action boundaries. The limited amount of human intervention allows us to build large-scale datasets with
high-quality consistent annotations. While our procedure
gives rise to only a sparse set of annotated clips, we demonstrate that models supervised by such annotations have superior generalization performance for both action classification and localization tasks.
2. Related Work
The motivation behind our work is the creation of an
annotated video collection that can serve as an instrumental training set for learning general spatiotemporal features,
i.e., features that can be successfully transferred to other
action recognition and detection domains and that can be
fine-tuned effectively on smaller benchmarks.
Over the last decade, several video datasets were created
for similar reasons. For example, in action recognition, the
HMDB51 [15] and the UCF101 [30] datasets were introduced to provide benchmarks with higher variety of actions
and more realistic capture conditions compared to precedent datasets, such as KTH [24]. These benchmarks have
indeed served our community well, by enabling the handdesign of very effective and robust features for action recognition, such as space-time interest points [16] or dense trajectories [34]. However, these datasets are not large enough
to support modern end-to-end training of deep models. The
For action classification, the large-scale of our dataset
can be leveraged to pre-train video models. We show that
fine-tuning these models on well-established action classification benchmarks (UCF101, HMDB51 and Kinetics)
yields significant gains over learning from scratch. On Kinetics, UCF101 and HMDB51, we are able to improve a
ResNet-based 3D ConvNet baseline [32] by 2.0%, 20.1%
and 35.4%, respectively. We also show that pre-training on
SLAC is more beneficial than pre-training on Sports1M or
Kinetics. Sports-1M annotations are generated by a tag prediction algorithm, which inevitably injects prominent noise
into the dataset. Also, the average length of Sports1M
videos is over 5 minutes and the tag predicted action may
2
large-scale Kinetics dataset [14] was recently introduced to
fill this gap and it was shown to enable the pre-training of
deep models that can be fine-tuned to achieve state-of-theart performance on smaller action recognition benchmarks,
such as HMDB51 and UCF101. Although Kinetics represents a remarkable resource for training deep action classifiers and for learning general video features, it cannot be
used for learning action detection models, as it consists of
only trimmed videos.
clips. Furthermore, since labeling on these clips is posed in
the form of simple yes/no questions (e.g., is action “kissing”
happening in this clip?), annotators can provide labels efficiently and consistently, without the need to stop or replay
the video.
Our approach has admittedly two downsides. The first
is that our annotation pipeline yields a sparse labeling of
clips, where other clips in the video remain unannotated
and thus not usable for supervised training. The second
disadvantage is that, since manual labels are sought only
for the clips deemed relevant by our method, the labeled
dataset is inherently biased rather than a uniform sampling
of the data source. We demonstrate empirically that the
benefits of our approach outweigh these two shortcomings.
Specifically, we propose a straightforward procedure for
pre-training action models from our sparse localization labels, and show that it improves model performances considerably after fine-tuning on fully-labeled data, compared
to learning from scratch or pre-training on other datasets.
Furthermore, we show that the dataset bias is small in practice through transfer learning experiments on four different
benchmarks for action classification and localization, where
our approach yields consistent metric gains.
Our objective is to introduce a benchmark that will enable not only the learning of action recognition models
but also the pre-training of deep models for action localization in untrimmed videos. Action detection in unconstrained videos is crucial to automatically understand Internet videos, which are typically several minutes long, and
include multiple actions or several people interacting. Recently, several datasets for untrimmed video analysis have
been presented. THUMOS Challenge 2014 [12] includes
2,765 trimmed training videos on 20 actions. It was subsequently extended into MultiTHUMOS [38] which contains
more action classes (65 instead of 20) and more temporal
and frame-level annotations. One shortcoming of THUMOS and MultiTHUMOS is that they only include examples of sport activities. Other datasets with fine granularity of classes but focused on narrow domains include
MPII Cooking [22] and MPII Cooking 2 [23], which depict human subjects preparing dishes. Models trained on
such domain-focused videos may not generalize optimally
to recognition of every-day activities. Conversely the Charades dataset [28] was purposefully designed to include
more general, daily activities. Since such uneventful activities are rarely shown in videos shared on the Internet, the
authors asked people to act out everyday routine activities
in their own homes in front of a camera, as previously done
in the Activities of Daily Living dataset [20]. ActivityNetv1.3 [2] includes over 600 hours of untrimmed videos and
about 30K temporal annotations of actions spanning many
common activities. Finally, the recently introduced AVA
dataset [9] provides person-centric annotations, where each
action annotations is a spatiotemporal box corresponding
to a person performing an atomic action, such as shaking
hands or kicking an object.
3. Dataset Collection Pipeline
3.1. The SLAC Dataset at a Glance
The SLAC dataset includes 200 action classes, which are
taken from ActivityNet-v1.3 dataset. It has a total of over
520K videos retrieved from YouTube. Each video is strictly
shorter than 4 minutes, and the average length is 2.6 minutes. Over 1.75M 2-second clips are sampled by a novel active sampling approach (section 3.6). 755K and 993K clips
are annotated as positive and negative samples, respectively,
by a team of 70 professional annotators working 6 hours a
day. The overall annotation takes effectively 10.5 working
days to complete, including 1 hour for training the annotators. We split SLAC into training, validation and testing
set with 502K, 6K and 12K videos, respectively. Dataset
webpage is hosted at http://slac.csail.mit.edu.
An overview of our data collection pipeline is shown in
Figure 1. Next, we elaborate in detail on the collection procedure.
One common trait of all the aforementioned datasets for
action detection is that localization in training videos was
obtained through exhaustive manual search of individual actions in each video sequence. To produce accurate temporal
localizations the annotator must typically stop, slow down,
play back and forth the video multiple times. This renders the process of annotation very time consuming. This
has hampered the scaling up in size of benchmarks for action detection. Our approach bypasses manual annotation
of temporal extents by automatically identifying the hard
clips that are hypothesized to be more informative than easy
3.2. Challenges in Data Collection
Collecting a large-scale action dataset is challenging due
to several reasons. First, compared with negative examples
of human action, the number of positive examples is arguably much smaller. If we uniformly sample video/clips
to annotate, we may end up with a large number of negative
examples which are far less useful than positive examples
for training video models. Second, not all negative examples are equally useful. We hypothesize that hard negative
3
Shot Detec3on and Person Detec3on
Video Crawling and De-duplica3on
Video Frame Classifier
Google Image Classifier
Annota3on: Is this ac3on happening? Skateboarding
CNN
CNN
Disagreement
Ac3ve Sampling
Figure 1. An overview of data collection procedure to obtain Sparsely Labeled Actions.
Positives
examples are more valuable than easy negative examples for
model learning. We must properly address both challenges
to make our data collection procedure cost-effective.
To address the first challenge and sample more positive
examples for annotation, we bias our sampling distribution towards videos that are likely to contain actions. This
is achieved by leveraging both video search engine (e.g.,
YouTube) to retrieve videos with actions of interest (Section 3.4), and person detectors to exclude videos that do not
contain humans (Section 3.5).
To address the second challenge, we review the different
types of negative examples, and their usefulness for training
models (Section 3.3).
Negatives
Action: Archery
No person
No context
No action
Figure 2. An illustration of negative samples on action datasets.
merely holding a bow and talking, as opposed to performing
archery.
As the first two types of negatives occur more often in
videos and are easier to obtain, we use pre-trained models
to automatically identify them, and spend most of our annotation budget on the third type.
3.3. Negative Sample Analysis
In general, there are three types of negative samples in
human action datasets, with increasing levels of mining difficulty:
No person. Since we are only interested in human actions, negative videos without human presence should be
excluded from annotation as much as possible.
No context. Another common type of negative samples is
represented by videos that contain people but lack the context of ongoing action. For instance, if the objective is to
collect videos of action Archery, we should exclude videos
if arrows or bows are not present in the videos.
No action. The hardest negative examples are videos that
include both people and context, but where the action of
interest is not being performed. We identify such hard negative examples by sending them to annotators for validation.
Concrete examples of these prototypical negative samples are shown in Figure 2, where Archery is the action to
localize. Examples of “No person” can be animations in
the video. Examples of “No context” are clips where people are present, but the clips are not related to archery. An
example of “No action” is represented by a person who is
3.4. Video Retrieval and De-duplication
SLAC
171,378
Kinetics
4,029
ActivityNet
1,925
UCF-101
660
HMDB-51
11
Table 2. De-duplication of SLAC within itself, and against the
test/validation sets of prior video datasets. Starting with an initial set of 1.1M videos, we show the number of duplicate videos
removed in each step.
Based on the taxonomy of ActivityNet-v1.3 where 200
human actions are defined, we retrieve 1.1M candidate
videos (1K to 6K videos per class) from YouTube by using the class label as query to find videos with matching
titles and descriptions.
Internal De-duplication. We de-duplicate videos within
SLAC, since YouTube may include several copies of the
same video, possibly differing in post-processing steps,
such as saturation/contrast enhancement or in duration.
Specifically, we resize each video to have short edge of
size 256 with aspect ratio preserved, and evenly sample
4
3.6. Proposal Models and Active Frame Sampling
224 × 224 central crops from frames at 0.2 fps. For each
video, we extract pool5 features of 2048-D dimensions using an ImageNet-pretrained ResNet-50 model for all sampled frames, and compute a mean feature vector. For each
action class, we treat its videos as nodes in a graph, and connect two video nodes if their cosine similarity is higher than
a threshold γ. After processing all pairs of videos in the
same action class, we compute the connected components
in the graph, and sample only one video from each connected component [14]. We choose γ = 0.935 by ensuring
through manual inspection that this thresholds removes all
duplicate videos in a large subset of video pairs with high
similarity scores. Through this de-duplication step, we remove about 171K videos.
Between the other two types of negatives, we hypothesize that “No Context” clips are easier to recognize and less
useful than “No Action” clips. Unfortunately both types of
videos can be returned by YouTube due to the video description matching the class label. We want to bias the distribution of our clip sampling for human annotation towards
“No Action” hard negative samples, and away from “No
Context” easy negative samples. To achieve this, we trained
two image classifiers, which are used to sample clips.
YouTube Frame-based Model. The first model is trained
on frames extracted from the top 500 videos retrieved by
YouTube for each action class. Only videos frames with
person detected are used as positive example for training.
This sums to a total of over 600K frames (3K frames per
action on average). We also randomly choose frames with
low person score as examples of the background class.
Google Image-based Model. The second model is trained
on images retrieved from Google Images using the class labels as queries. This strategy yields a total of over 300K
images (1.5K per class) after thresholding on person detection. We also use 1000 class labels of ImageNet as keywords to retrieve examples with Google Images and assign
these images to a background class.
We use ResNet-50 architecture for both proposal models and train them as 201-way classification, including 200
actions and 1 background class.
We use a measure of consensus between these two classifiers to bias the distribution of sampling clips for annotations towards “No Action” samples. In active learning,
this consensus method is known as query-by-committee
(QBC) [25]. First, for each video we randomly sample one
frame from each shot in the video. We denote with X the
set of these frames, representing all the candidate clips for
annotation in the video. Next, we want to define a sampling distribution R(x) over x ∈ X, in order to sample the
frames that are most informative for annotation. We define
the unnormalized sampling probability R(x) as follows:
R(x) = PcY (x) + PcG (x) ∗ DKL P Y (x), P G (x) ∗
X
1
0
S(f (x), f (x ))
|X| − 1 0
0
External De-duplication. We perform additional deduplication to ensure that SLAC does not overlap with the
validation and testing sets of other video datasets, namely
Kinetics, ActivityNet, UCF101 and HMDB51. As the taxonomy between SLAC and other datasets can be different,
for each action class A in SLAC we first manually compile a
list of similar action classes {Bi }i in the target dataset. For
example, we selected actions swimming backstroke, swimming breast stroke, and swimming butterfly stroke in Kinetics as the classes that may contain duplicates of SLAC
videos for action swimming. We compute the similarity between each SLAC video in class A and each target video of
actions {Bi }i . We remove a video from SLAC if its similarity score with a video from the target dataset is above
the threshold. We sequentially de-duplicate SLAC with respect to Kinetics, ActivityNet, UCF101 and HMDB51, and
remove 4, 029, 1, 925, 660 and 11 videos from SLAC.
3.5. Person detection
In this step we use a person detector to remove the first
kind of negative samples: videos that do not contain people.
To reduce the computation, shot detection is performed beforehand. We compute the dissimilarity score between color
histograms of consecutive frames, and threshold on both the
score and its gradient to find shot boundary. From each
shot, 2 frames are sampled and a person score is computed.
We run the Faster R-CNN [21] trained on the COCO person class to get the maximum score of person being present
over the two frames. A threshold of 0.5 on the person score
is used to find shots with people (nearly 49% of all video
shots). These shots are the initial proposals of human actions used for further processing. To verify the accuracy of
this procedure, we test the same person detector on the validation set of ActivityNet-v1.3, and get a recall of >97%.
If all shots in a video have scores below the threshold, we
remove such video. This removes 193K videos.
x ∈X,x 6=x
(1)
where P Y (x) and P G (x) are the 201-dimensional predicted probabilities from YouTube frame classifiers Y and
Google image classifier G, respectively. c denotes the action label that was used to retrieve the video. Therefore,
the first term (PcY (x) + PcG (x)) represents the consensus
of both classifiers on the probability of frame x containing
action c. The rationale behind this term is that it can reduce
the sampling probability for “No Context” frames. The sec5
Action: Clean and jerk
Definition: a two-movement weightlifting exercise in which a weight is raised
above the head following an initial lift to shoulder level. (Different from Snatch)
Positive
Negative in preparation stage
Figure 4. Annotation tool. The user can quickly flip the border
color of clips to mark positive/negative example by simple clicking. The action being annotated above is Kayaking. Due to space
constrain, we only show 3 clips here.
Figure 3. Annotation guideline, including definition of the action
class, as well as positive and negative examples.
ond term DKL (·) is the KL divergence, which represents
the disagreement of class probabilities between two image
classifiers. The third term computes the average similarity between x and all other frames in X. Frame similarity is
assessed as the cosine similarity between conv5 feature activation f (x) extracted from an Imagenet-trained ResNet-50
model. This third term biases our sampling towards representative frames in the video. Once we compute {R(x)}x
for all frame candidates in X, we then perform L1 normalization to obtain a proper sampling distribution. We sample at most 8 frames from X (i.e., per video) according to
this distribution but only retain samples for which the first
term (action probability) is greater than τ , which is chosen
as 0.01. In addition, we also sample one extra frame, corresponding to the frame having the highest value of under
the first term (consensus). This sample can be viewed as
an easy positive example. We select such samples as we
hypothesize that a small number of easy positive examples
can be useful for model training as well.
After sampling the video frames, we take one second on
either side of a frame to convert it into a 2-second clip. We
also ensure that such clips do not cross shot boundaries.
As a result, a total of 1.75M clips are sampled from 520K
videos.
are shown in green and negative clips in red.
Per clip
Per video
Dataset
Full annotation
N/A
11.3 min
est. 113.2K hr
Sparse annotation
8.8 sec
30.6 sec
4,390 hr
Table 3. Cost comparison for full and sparse annotations.
Annotation Quality Control. To ensure a high annotation quality, we first ensure clips from the same action label
are labeled by the same annotator. This removes the interannotator variance. Second, we personally labeled 10K
clips before kicking off the annotation. We refer to this set
of clips as the Golden Set which is mixed with other clips
to be annotated. We use the clips in the Golden Set to compute annotation accuracy per annotator, by comparing the
labels provided by us to those given by the annotators. We
monitor the accuracy and iteratively give customized clarifications when an annotator has low accuracy. We ensure
all annotators have over 90% accuracy. The final average
accuracy is 94%.
Annotation Cost. We hire 70 professional annotators
working 6 hours per day. The annotation task is finished
in 10.5 working days. In the last 2 days, several annotators have already completed the annotations and passed
the accuracy check, while others receive more clarifications
and re-annotate actions due to low Golden Set accuracy.
Through this procedure we gathered on average 4,390 annotator × hour. The average annotation cost (in time) is 8.8
seconds per clip, and 30.6 seconds per video. To compare
the cost with traditional annotation of action localization,
we take 20K videos from SLAC and ask the same group of
annotators to manually mark start and end of the actions in
each video. Statistics are shown in Table 3. The total saving
is over 95%. These numbers suggest that collecting localizations manually is unachievable in practice at the scale of
our dataset, due to the very long time that this would take
and the consequent monetary cost.
3.7. Video Clip Annotation
The video annotation task is aimed at labeling the 1.75M
clips of 2-second duration in a clear and consistent manner.
Annotation guideline We prepare a detailed annotation
guideline providing clear action definitions. Both text descriptions and positive/negative clip examples with clarifications are provided to annotators to reduce the ambiguity of action annotation. An illustration of our guideline is
shown in Figure 3.
Annotation Tool. Our annotation tool is visualized in Figure 4. Clips are organized into a grid layout, and automatically played. Annotators can quickly mark them as positive/negative samples by clicking on the clips. The border
color visualizes the current annotations where positive clips
6
4. Action Recognition Benchmark
each class, and finally compute the average of class-wise
top-1accuracies.
Results. As shown in Table 4, with RGB input, we achieve
54.3% and 76.5% Class-Agnostic and Class-Aware clip@1
accuracy on validation set. On test set, we obtain 53.7%
and 76.2% Class-Agnostic and Class-Aware clip@1 accuracy. We also experiment with using optical flow as input to
the video model. We adopt efficient Farneback’s approach
[5] to quickly compute optical flow. For optical flow, we
use the same number of frames and the same resolution
as RGB. Results of using optical flow alone are also reported in Table 4, and are about 7% and 12% worse than
those with RGB input alone for Class-Agnostic and ClassAware Clip@1 accuracy. We also experiment with combining RGB and optical flow by late fusion, where the final prediction score is a weighted sum of two separate prediction scores from RGB and optical flow. We empirically
set weights for RGB and optical flow to 0.6 and 0.4, where
RGB weight is bigger as RGB input alone gives higher accuracy than optical flow input. Late fusion of RGB and optical flow improves the results of RGB input alone by 8.1%
and 1.8% for Class-Agnostic and Class-Aware Clip@1 on
validation set, and 7.8% and 1.6% for Class-Agnostic and
Class-Aware Clip@1 on test set.
In this section we present results on action recognition.
First, we train 3D ConvNet on SLAC training set and evaluate on SLAC validation and test set. Second, we evaluate
the efficacy of SLAC for pre-training models, and compare
SLAC with respect to other large scale datasets (such as Kinetics and Sports1M). Meanwhile, we experiment with fusing RGB and optical flow features, and compare with the
state-of-the-art in the literature.
4.1. Evaluations on SLAC
Network Architecture. ResNet [11] has emerged as one of
the most successful models for image classification. Here
we experiment with its 3D version, namely Res3D [32],
which extends the convolutional operators from 2D to 3D.
The Res3D model adopted in this paper has 34 parametric
layers and uses an input of 32 adjacent frames with a fixed
resolution of 112 × 144. We denote it as Res3D-34.
Implementation Details. We implement our approach in
Caffe2 [1]. For training Res3D-34 model, we use both positive and negative clips in SLAC. There are 200 positive
classes, and we assign a single background label to all negative clips. Therefore, Res3D-34 is trained with 201-way
softmax loss. SLAC has highly imbalanced positive and
negative clips where 57.5% of training clips are negative,
and on average each positive class accounts for only 0.215%
of total clips. To prevent model training being biased by
negative clips, we use weighted sampling to assemble training mini-batches. We empirically use weight 0.02 for negative clips and 1.0 for positive clips. Weights serve as unnormalized multi-nomial distribution from which we sample clips to form mini-batches. Mini-batch size per GPU is
set to 16. To speed up the model training on large-scale
datasets, we use 16 hosts with 4 Nvidia M40 GPUs per
host for distributed training with parameter synchronization. Therefore, the total mini-batch size is 1024. This
yields a throughput of about 600 video clips per second.
We train on SLAC for 45 epochs, with the first 10 epochs
for learning rate warmup [8], and then reduce the learning
rate by a factor of 10 every 10 epochs. The base learning
rate per GPU is set to 0.005, and the effective learning rate
after warmup is 0.32. After training, we evaluate on the
SLAC validation set and test set.
Evaluation Metrics. For each clip in validation and test
set, we sample a single sequence of 32 frames in the beginning of the clip. Since both validation and test set have
highly imbalanced positive and negative clips as in training set, we report both class-agnostic and class-aware clip
accuracy. (1) Class-Agnostic Clip@1: we count the number of clips where the top predicted label is correct, and
divide it by the total number of clips in the validation/test
set. (2): Class-Aware Clip@1: we first group clips based
on groundtruth class label, then compute top-1 accuracy for
4.2. Evaluations on Transfer Learning Tasks
Given the trained Res3D-34 model, we can finetune it on
other target datasets. By comparing the performance on target datasets using finetuned models with that from models
trained from scratch, we can evaluate the efficacy of SLAC
for pre-training large models.
Evaluation Metrics. At test time, we measure performance
according to three metrics. (1) Clip@1: We sample a single
32-frame clip in the beginning of the video, and report top-1
accuracy. (2) Video@1: We evenly sample 10 clips of 32
frames in the video, average the clip predictions and report
top-1 accuracy. (3) Dense prediction w/ multi-crops@1.
We densely sample clips of 32 frames with a stride of 4
frames. For each clip, we also take 5 different crops (4 in
the corners, and 1 in the center) as well as their mirrored
versions. We average predictions over crops and clips, and
report top-1 accuracy.
Action recognition datasets. We use 4 other datasets
for the transfer learning of action recognition experiments.
UCF-101, and HMDB-51 are used as target datasets. Kinetics are used as both pre-training dataset and target dataset.
Sports-1M is used only as pre-training dataset. We finetune
models on UCF-101, HMDB-51 and Kinetics for 110, 50,
and 110 epochs. As there are many long videos in Sports1M, we cut them into shorter clips to better utilize the data
and end up with a training set of about 5M samples. For
Kinetics, we report the accuracy on the validation set as the
annotations of the testing set are not publicly available. For
7
Input
RGB
Flow
RGB + Flow
Validation Set
Class-Agnostic, Clip@1 Class-Aware, Clip@1
54.3
76.5
47.3
64.9
62.4
78.3
Test Set
Class-Agnostic, Clip@1 Class-Aware, Clip@1
53.7
76.2
46.3
63.6
61.5
77.8
Table 4. Results on SLAC. We train Res3D-34 model using 201-way softmax loss.
Input
RGB
Flow
Target Dataset
Training
w/o pre-training
w/ pre-training
w/o pre-training
w/ pre-training
UCF-101
Clip@1 Video@1
72.7
75.6
93.4
95.7
68.7
80.8
87.9
93.6
HMDB-51
Clip@1 Video@1
34.0
35.9
67.3
71.3
34.0
45.1
60.0
69.5
Kinetics
Clip@1 Video@1
57.0
68.2
58.5
70.2
42.2
59.2
45.0
62.6
Table 5. Evaluations on the transfer learning performance of SLAC using different types of input.
Target Dataset
Dataset pre-train
w/o pre-training
Kinetics
Sports1M
SLAC
UCF-101
Clip@1 Video@1
72.7
75.6
89.9
92.5
90.9
92.6
93.4
95.7
HMDB-51
Clip@1 Video@1
34.0
35.9
64.9
67.4
64.1
65.9
67.3
71.3
Kinetics
Clip@1 Video@1
57.0
68.2
57.0
68.2
57.4
68.4
58.5
70.2
Table 6. Comparisons of SLAC with other datasets for pre-training action recognition models. Test performance is evaluated by finetuning
on UCF-101, HMDB-51, Kinetics. All numbers are obtained by training the same model on RGB input.
Method
Long-term temporal convolutions [33]
ST-Multiplier Net + IDT [7]
Temporal Segment Networks [35]
Two-Stream I3D, Kinetics pre-training [14]
Two-Stream Res3D-34, SLAC pre-training, video@1
Two-Stream Res3D-34, SLAC pre-training, dense prediction, w/ multi-cropping@1
UCF-101
92.7
94.9
94.2
98.0
96.6
97.6
HMDB-51
67.2
72.2
69.4
80.7
75.6
76.7
Kinetics
N/A
N/A
N/A
75.7
73.5
74.8
Table 7. Comparing Res3D-34 model pre-trained on SLAC with the state-of-the-art models on UCF101, HMDB51 and Kinetics.
UCF101 and HMDB51, we only use split-1 for training and
testing.
Results. As shown in Table 5, with RGB input pre-training
on SLAC improves the Clip@1 by 20.7%, 33.3%, and
1.5%, and Video@1 by 20.1%, 35.4%, and 2.0% on UCF101, HMDB-51 and Kinetics, respectively. When optical flow input is used, pre-training on SLAC improves the
Clip@1 by 19.2%, 26.0%, and 2.8%, and Video@1 by
12.8%, 24.4%, and 3.4% on UCF-101, HMDB-51 and Kinetics, respectively.
Comparisons with other pre-training datasets. To put
the efficacy of SLAC as a pre-training dataset, we compare it with two other large scale datasets (Kinetics and
Sports1M) for transfer learning. Results are summarized in
Table 6. For pre-training, Kinetics is better than Sports1M
when finetuning models on UCF101 and HMDB51, as the
annotations from Sports1M are noisy despite being one order of magnitude more numerous than those of Kinetics. On
the other hand, pre-training on Sports1M still improves the
accuracy on Kinetics, albeit only slightly. SLAC reports the
best accuracy as a pre-training dataset for all three target
datasets due to its large-scale and its high-quality annotations. The gain is up to 3.9% for Video@1 when comparing to Kinetics or Sports-1M. In Table 6, SLAC achieves
the state-of-the-art performance for transfer learning.
Feature Fusion. The two-stream [29] model has been
shown to be successful in action recognition by combing
the appearance information from RGB and motion information from optical flow. Table 5 presents the results of
transfer learning on SLAC using RGB and optical flow input alone. Overall, we observe similar performance gain
on optical flow by pre-training on SLAC as RGB. When
8
comparing the accuracy between RGB and optical flow with
pre-training, RGB is always better than optical flow on all
three target datasets. As expected, combining RGB with
optical flow using two streams always improves the final
accuracy. The improvements in Video@1 metric on UCF101, HMDB-51 and Kinetics over RGB-only models are
0.9% (95.7% vs. 96.6%), 4.3% (71.3% vs. 75.6%) and 3.3%
(70.2% vs. 73.5%), respectively (Table 6 and Table 7).
Comparisons with other methods. Finally, we compare
our SLAC-pretrained Res3D-34 model with the state-ofthe-art in Table 7. On UCF-101 and HMDB-51, by simply using an off-the-shelf 3D ResNet of 34 layer and leveraging the power of a large-scale well annotated dataset,
we outperform three competitive methods [33, 7, 35] by at
least 2.7%. On Kinetics validation set, our model performs
slightly (0.9%) worse than the two-stream I3D model [14].
There are several factors contributing to such performance
gap. First, I3D is a more sophisticated model with parameters inflated from ImageNet pre-trainining. Second, I3D
uses the TV-L1 optical flow [39], which is more expensive
and accurate than the Farneback optical flow [5] adopted by
our model. Third, we train with 32 video frames with size
112 × 144 while I3D uses 64 frames of size 224 × 224. We
leave empirical explorations of these differences on performance as future work.
architecture can be found in Supplementary Materials.
Training. When pre-train on SLAC, we randomly take 32
frames from 2-second SLAC clips as training examples,
and give all frames a uniform label based on clip annotation. Similar to how we trained recognition models, a single background label is assigned to all negative examples,
and sampling weights of positive and negative samples are
1.0 and 0.02, respectively. The pre-trained model is further
finetuned on THUMOS14 and ActivityNet-v1.3. For each
action annotation, we expand it by 2 seconds at both the
beginning and the end of the segment. Then we randomly
choose 32 frames as training examples. Frames within the
original segment are assigned positive action label, while
frames outside of the original segments are assinged background label. The localization models are finetuned using
distributed training with 16 hosts of 4 GPU per host for 11
epochs, including 2 warm-up epochs, where learning rate
increases linearly from 0.005 to 0.32. Afterwards, learning
rate is discounted by 1/10 every 3 epochs.
Inference. Although localization model can produce
frame-level dense predictions, it is non-trivial to design robust algorithm to group frames to form contiguous labeling
of action segments based on prediction scores. On the other
hand, there are existing segment proposal methods to generate candidate segments with high recall. Therefore, for
THUMOS14, we adopt proposals from [27], and apply our
dense prediction model for localization; for ActivityNet, we
adopt the output from [36] as proposals, and then we follow
the method in [26] to refine proposal boundaries. Given a
segment proposal, we first extend it by a percentage α on
each side (α = 25%), and assign the label with the highest
mean prediction over the entire extended segment. Proposals assigned with background label are abandoned. Otherwise, we perform Gaussian density estimation to refine
proposal boundaries by computing the mean µ and standard
deviation σ of action scores. We scan the prediction score
of the predicted action label on each side of the extended
proposal, and keep shrinking the boundary until the action
score at a frame is higher than µ − β ∗ σ. β is tuned on a
small held-out set of videos taken from training set, and is
set to 2.0.
Comparison. We compare the performance of action localization models with and without SLAC pre-training, on the
THUMOS14 test set and ActivityNet-v1.3 validation set, in
Table 8 and Table 9, respectively. In general, pre-training
significantly improves performance of localization models,
in terms of mean average precision (mAP). Specifically, the
framework is good at improving hits at high IoU thresholds (IoU> 0.6), which can be explained by the accurate
action boundary prediction by frame-wise dense labeling.
Compared with our localization model based on Res3D34 and Conv-Deconv operators trained from scratch, the
model with the same architecture but pre-trained on SLAC
5. Action Localization Benchmark
In this section, we present the results of action localization on two popular benchmarks, THUMOS14 and
ActivityNet-v1.3.
Framework. For the task of action localization, we adopt
the framework from [26]. First, candidate video segments
are proposed from the untrimmed videos; second, dense
frame-wise labels are predicted for all the proposal segments to localize actions accurately. Note that we only use
RGB frames (without optical flow) as input in all experiments of action localization.
Network Architecture. For dense frame-wise action
prediction, the Convolutional-De-Convolutional (ConvDeconv) operators are used in our localization model.
Given the input frames of size 32 × 3 × 112 × 112, the
full localization model uses a Res3D-34 trunk, which produces feature maps of size 4 × 512 × 7 × 7, and 3 successive Conv-Deconv operators, which produce feature maps
of size 32 × 512 × 1 × 1. A 201-way softmax is used to generate predictions over 200 actions and 1 background class.
The Conv-Deconv operator, originally proposed in [26], implements convolution (and downsampling) in space and deconvolution (and upsampling) in time in a single step, and
it involves more parameters to learn than separate Conv operator and Deconv operator. The number of parameters in
Res3D-34 trunk is 63.4M, while the Conv-Deconv operators add extra 20.4M parameters. Details of the localization
9
mAP @ IoU
w/o pre-training
pre-train on Sports1M [26]
pre-train on SLAC
R-C3D [37]
SSN [40]
0.3
33.0
40.1
45.0
44.8
51.9
0.4
27.1
29.4
35.8
35.6
41.0
0.5
18.9
23.3
29.2
28.9
29.8
0.6
10.5
13.1
17.2
N/A
N/A
0.7
4.0
7.9
9.5
N/A
N/A
Average
18.7
22.8
27.3
N/A
N/A
Table 8. Comparison of action localization performance on the THUMOS14 test set with different pre-training datasets, and the state-ofthe-art methods.
mAP @ IoU
before refinement [36]
after refinement, w/o pre-training
after refinement, pre-train on Sports1M [26]
after refinement, pre-train on SLAC
0.5
43.3
43.0
45.3
42.9
0.55
39.9
39.5
N/A
38.9
0.60
36.7
36.0
N/A
34.7
0.65
16.8
32.0
N/A
31.0
0.70
8.7
28.6
N/A
28.0
0.75
3.9
24.6
26.0
24.6
0.80
1.9
13.5
N/A
20.9
0.85
0.8
5.4
N/A
16.9
0.90
0.2
1.4
N/A
9.7
0.95
0.0
0.2
0.2
1.4
Average
15.3
22.4
23.8
24.9
Table 9. Comparison of action localization performance on the ActivityNet-v1.3 validation set before and after refinement with different
pre-training datasets.
achieves the absolute gain of 8.6% on THUMOS14 and
2.5% on ActivityNet-v1.3. We also compare with the original Conv-Deconv work pre-trained on Sports-1M [31], our
model outperforms it by 4.5% on THUMOS14 and 1.1%
on ActivityNet-v1.3. State-of-the-art methods with different frameworks are shown in the last rows of Table 8, our
performance beats the competitive R-C3D [37] model, but
not SSN [40]. Exploring the best framework for action localization is beyond the focus of this paper, and we will
leave it for future work.
[3]
[4]
[5]
6. Conclusions
[6]
We present a novel and highly efficient data collection procedure to construct large-scale human action video
datasets. The proposed pipeline yields a 95% reduction
in annotation time. Our resulting dataset, SLAC, includes
1.75M clip annotations spanning over 520K untrimmed
videos. We demonstrate that SLAC consistently outperforms Sports-1M and Kinetics as a pretraining dataset for
action recognition models. We also show superior transfer learning performance on action localization on THUMOS14 and ActivityNet-v1.3 benchmarks. Dataset, pretrained models, and code will be released upon publication
of the paper.
[7]
[8]
[9]
Acknowledgement We thank Du Tran for insightful discussions along the development of this project.
[10]
References
[11]
[1] Caffe2: A new lightweight, modular, and scalable deep
learning framework. https://caffe2.ai/.
[2] F. Caba Heilbron, V. Escorcia, B. Ghanem, and J. Carlos Niebles. Activitynet: A large-scale video benchmark for
[12]
10
human activity understanding. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 961–970, 2015.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. FeiFei. Imagenet: A large-scale hierarchical image database.
In Computer Vision and Pattern Recognition, 2009. CVPR
2009. IEEE Conference on, pages 248–255. IEEE, 2009.
M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and
A. Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303–
338, June 2010.
G. Farnebäck. Two-frame motion estimation based on polynomial expansion. Image analysis, pages 363–370, 2003.
L. Fei-Fei, R. Fergus, and P. Perona. One-shot learning of
object categories. IEEE transactions on pattern analysis and
machine intelligence, 28(4):594–611, 2006.
C. Feichtenhofer, A. Pinz, and R. P. Wildes. Spatiotemporal
multiplier networks for video action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 4768–4777, 2017.
P. Goyal, P. Dollár, R. Girshick, P. Noordhuis,
L. Wesolowski, A. Kyrola, A. Tulloch, Y. Jia, and K. He.
Accurate, large minibatch sgd: Training imagenet in 1 hour.
arXiv preprint arXiv:1706.02677, 2017.
C. Gu, C. Sun, S. Vijayanarasimhan, C. Pantofaru, D. A.
Ross, G. Toderici, Y. Li, S. Ricco, R. Sukthankar, C. Schmid,
and J. Malik. AVA: A video dataset of spatio-temporally localized atomic visual actions. CoRR, abs/1705.08421, 2017.
K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask r-cnn.
arXiv preprint arXiv:1703.06870, 2017.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In The IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), June 2016.
Y. Jiang, J. Liu, A. R. Zamir, G. Toderici, I. Laptev, M. Shah,
and R. Sukthankar. Thumos challenge: Action recognition
with a large number of classes, 2014.
[13] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar,
and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014.
[14] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, et al.
The kinetics human action video dataset. arXiv preprint
arXiv:1705.06950, 2017.
[15] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre.
Hmdb: a large video database for human motion recognition. In Computer Vision (ICCV), 2011 IEEE International
Conference on, pages 2556–2563. IEEE, 2011.
[16] I. Laptev. On space-time interest points. International journal of computer vision, 64(2-3):107–123, 2005.
[17] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In European conference on computer
vision, pages 740–755. Springer, 2014.
[18] M. Marszalek, I. Laptev, and C. Schmid. Actions in context.
In Computer Vision and Pattern Recognition, 2009. CVPR
2009. IEEE Conference on, pages 2929–2936. IEEE, 2009.
[19] C. Pantofaru, C. Sun, C. Gu, C. Schmid, D. Ross,
G. Toderici, J. Malik, R. Sukthankar, S. Vijayanarasimhan,
S. Ricco, et al. Ava: A video dataset of spatio-temporally
localized atomic visual actions. 2017.
[20] H. Pirsiavash and D. Ramanan. Detecting activities of daily
living in first-person camera views. In Computer Vision
and Pattern Recognition (CVPR), 2012 IEEE Conference on.
IEEE, 2012.
[21] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards
real-time object detection with region proposal networks. In
Advances in neural information processing systems, pages
91–99, 2015.
[22] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A
database for fine grained activity detection of cooking activities. In 2012 IEEE Conference on Computer Vision and
Pattern Recognition, Providence, RI, USA, June 16-21, 2012,
pages 1194–1201, 2012.
[23] M. Rohrbach, A. Rohrbach, M. Regneri, S. Amin, M. Andriluka, M. Pinkal, and B. Schiele. Recognizing fine-grained
and composite activities using hand-centric features and
script data. International Journal of Computer Vision, pages
1–28, 2015.
[24] C. Schuldt, I. Laptev, and B. Caputo. Recognizing human
actions: a local svm approach. In Pattern Recognition, 2004.
ICPR 2004. Proceedings of the 17th International Conference on, volume 3, pages 32–36. IEEE, 2004.
[25] H. S. Seung, M. Opper, and H. Sompolinsky. Query by committee. In Proceedings of the fifth annual workshop on Computational learning theory, pages 287–294. ACM, 1992.
[26] Z. Shou, J. Chan, A. Zareian, K. Miyazawa, and S.-F. Chang.
CDC: Convolutional-de-convolutional networks for precise
temporal action localization in untrimmed videos. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
[27] Z. Shou, D. Wang, and S. Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In 2016
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
11
IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016,
pages 1049–1058, 2016.
G. A. Sigurdsson, G. Varol, X. Wang, A. Farhadi, I. Laptev,
and A. Gupta. Hollywood in homes: Crowdsourcing data
collection for activity understanding. In European Conference on Computer Vision, 2016.
K. Simonyan and A. Zisserman. Two-stream convolutional
networks for action recognition in videos. In Advances
in neural information processing systems, pages 568–576,
2014.
K. Soomro, A. R. Zamir, and M. Shah. UCF101: A dataset
of 101 human actions classes from videos in the wild. arXiv
preprint arXiv:1212.0402, 2012.
D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri.
Learning spatiotemporal features with 3D convolutional networks. In Proceedings of the IEEE international conference
on computer vision, pages 4489–4497, 2015.
D. Tran, J. Ray, Z. Shou, S.-F. Chang, and M. Paluri. Convnet architecture search for spatiotemporal feature learning.
arXiv preprint arXiv:1708.05038, 2017.
G. Varol, I. Laptev, and C. Schmid. Long-term temporal
convolutions for action recognition. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 2017.
H. Wang, A. Kläser, C. Schmid, and C.-L. Liu. Action recognition by dense trajectories. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages
3169–3176. IEEE, 2011.
L. Wang, Y. Xiong, Z. Wang, Y. Qiao, D. Lin, X. Tang, and
L. Val Gool. Temporal segment networks: Towards good
practices for deep action recognition. In ECCV, 2016.
R. Wang and D. Tao. UTS at ActivityNet 2016. In Computer
Vision and Pattern Recognition Workshop (CVPRW) on ActivityNet Large Scale Activity Recognition Challenge, 2016.
H. Xu, A. Das, and K. Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In The
IEEE International Conference on Computer Vision (ICCV),
Oct 2017.
S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori,
and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of
Computer Vision, pages 1–15, 2015.
C. Zach, T. Pock, and H. Bischof. A duality based approach
for realtime TV-L1 optical flow. Pattern Recognition, pages
214–223, 2007.
Y. Zhao, Y. Xiong, L. Wang, Z. Wu, X. Tang, and D. Lin.
Temporal action detection with structured segment networks.
In ICCV, 2017.
| 2cs.AI
|
Published as a conference paper at ICLR 2018
S KIP C ONNECTIONS E LIMINATE S INGULARITIES
A. Emin Orhan
Xaq Pitkow
[email protected]
[email protected]
Baylor College of Medicine & Rice University
arXiv:1701.09175v8 [cs.NE] 4 Mar 2018
A BSTRACT
Skip connections made the training of very deep networks possible and have become an indispensable component in a variety of neural architectures. A completely satisfactory explanation for their success remains elusive. Here, we present
a novel explanation for the benefits of skip connections in training very deep networks. The difficulty of training deep networks is partly due to the singularities
caused by the non-identifiability of the model. Several such singularities have
been identified in previous works: (i) overlap singularities caused by the permutation symmetry of nodes in a given layer, (ii) elimination singularities corresponding to the elimination, i.e. consistent deactivation, of nodes, (iii) singularities generated by the linear dependence of the nodes. These singularities cause degenerate
manifolds in the loss landscape that slow down learning. We argue that skip connections eliminate these singularities by breaking the permutation symmetry of
nodes, by reducing the possibility of node elimination and by making the nodes
less linearly dependent. Moreover, for typical initializations, skip connections
move the network away from the “ghosts” of these singularities and sculpt the
landscape around them to alleviate the learning slow-down. These hypotheses are
supported by evidence from simplified models, as well as from experiments with
deep networks trained on real-world datasets.
1
I NTRODUCTION
Skip connections are extra connections between nodes in different layers of a neural network that
skip one or more layers of nonlinear processing. The introduction of skip (or residual) connections
has substantially improved the training of very deep neural networks (He et al., 2015; 2016; Huang
et al., 2016; Srivastava et al., 2015). Despite informal intuitions put forward to motivate skip connections, a clear understanding of how these connections improve training has been lacking. Such
understanding is invaluable both in its own right and for the possibilities it might offer for further
improvements in training very deep neural networks. In this paper, we attempt to shed light on this
question. We argue that skip connections improve the training of deep networks partly by eliminating the singularities inherent in the loss landscapes of deep networks. These singularities are caused
by the non-identifiability of subsets of parameters when nodes in the network either get eliminated
(elimination singularities), collapse into each other (overlap singularities) (Wei et al., 2008), or
become linearly dependent (linear dependence singularities). Saad & Solla (1995); Amari et al.
(2006); Wei et al. (2008) identified the elimination and overlap singularities and showed that they
significantly slow down learning in shallow networks; Saxe et al. (2013) showed that linear dependence between nodes arises generically in randomly initialized deep linear networks and becomes
more severe with depth. We show that skip connections eliminate these singularities and provide evidence suggesting that they improve training partly by ameliorating the learning slow-down caused
by the singularities.
2
2.1
R ESULTS
S INGULARITIES IN FULLY- CONNECTED LAYERS AND HOW SKIP CONNECTIONS BREAK
THEM
In this work, we focus on three types of singularity that arise in fully-connected layers: elimination
and overlap singularities (Amari et al., 2006; Wei et al., 2008), and linear dependence singularities
1
Published as a conference paper at ICLR 2018
a
w
...
w =0
...
J =0
Elimination
singularities
J
w =0
w
...
...
J =0
identifiable? yes
no
skip
J
b wa
...
w b wa
...
Ja =Jb
wb
c
Ja =Jb
Overlap
singularity
wa wb wc
Ja
Jb
wa wb wc
Jc =αJa +βJb J
a
Jb
Jc
Linear dependence
singularity
Figure 1: Singularities in a fully connected layer and how skip connections break them. (a) In
elimination singularities, zero incoming weights, J = 0, eliminate units and make outgoing weights,
w, non-identifiable (red). Skip connections (blue) ensure units are active at least sometimes, so the
outgoing weights are identifiable (green). The reverse holds for zero outgoing weights, w = 0: skip
connections recover identifiability for J. (b) In overlap singularities, overlapping incoming weights,
Ja = Jb , make outgoing weights non-identifiable; skip connections again break the degeneracy. (c)
In linear dependence singularities, a subset of the hidden units become linearly dependent, making
their outgoing weights non-identifiable; skip connections break the linear dependence.
(Saxe et al., 2013). The linear dependence singularities can arise exactly only in linear networks,
whereas the elimination and overlap singularities can arise in non-linear networks as well. These
singularities are all related to the non-identifiability of the model. The Hessian of the loss function
becomes singular at these singularities (Supplementary Note 1), hence they are sometimes also
called degenerate or higher-order saddles (Anandkumar & Ge, 2016). Elimination singularities
arise when a hidden unit is effectively killed, e.g. when its incoming (or outgoing) weights become
zero (Figure 1a). This makes the outgoing (or incoming) connections of the unit non-identifiable.
Overlap singularities are caused by the permutation symmetry of the hidden units at a given layer
and they arise when two units become identical, e.g. when their incoming weights become identical
(Figure 1b). In this case, the outgoing connections of the units are no longer identifiable individually
(only their sum is identifiable). Linear dependence singularities arise when a subset of the hidden
units in a layer become linearly dependent (Figure 1c). Again, the outgoing connections of these
units are no longer identifiable individually (only a linear combination of them is identifiable).
How do skip connections eliminate these singularities? Skip connections between adjacent layers
break the elimination singularities by ensuring that the units are active at least for some inputs,
even when their adjustable incoming or outgoing connections become zero (Figure 1a; right). They
eliminate the overlap singularities by breaking the permutation symmetry of the hidden units at
a given layer (Figure 1b; right). Thus, even when the adjustable incoming weights of two units
become identical, the units do not collapse into each other, since their distinct skip connections
still disambiguate them. They also eliminate the linear dependence singularities by adding linearly
independent (in fact, orthogonal in most cases) inputs to the units (Figure 1c; right).
2.2
W HY ARE SINGULARITIES HARMFUL FOR LEARNING ?
The effect of elimination and overlap singularities on gradient-based learning has been analyzed
previously for shallow networks (Amari et al., 2006; Wei et al., 2008). Figure 2a shows the simplified two hidden unit model analyzed in Wei et al. (2008) and its reduction to a two-dimensional
system in terms of the overlap and elimination variables, h and z. Both types of singularity cause
degenerate manifolds in the loss landscape, represented by the lines h = 0 and z = ±1 in Figure 2b,
corresponding to the overlap and elimination singularities respectively. The elimination manifolds
divide the overlap manifolds into stable and unstable segments. According to the analysis presented
in Wei et al. (2008), these manifolds give rise to two types of plateaus in the learning dynamics:
on-singularity plateaus which are caused by the random walk behavior of stochastic gradient descent (SGD) along a stable segment of the overlap manifolds (thick segment on the h = 0 line in
Figure 2b) until it escapes the stable segment, and (more relevant in practical cases) near-singularity
plateaus which manifest themselves as a general slowing of the dynamics near the overlap manifolds, even when the initial location is not within the basin of attraction of the stable segment.
Although this analysis only holds for two hidden units, for higher dimensional cases, it suggests that
overlaps between hidden units significantly slow down learning along the overlap directions. These
overlap directions become more numerous as the number of hidden units increases, thus reducing
the effective dimensionality of the model. We provide empirical evidence for this claim below.
2
Published as a conference paper at ICLR 2018
Ja
wb
Jb
overlap
1
||J –J ||2 = h
2 a b
b
z
3
1
0
0
–1
–2
c
Gradient flow
2
0
h
1
–3
Near-singularity plateau
4
z
h
z,h
0
d
+
50
Time
100
1
0
–1
wa wb wc
Ja
0
e
linear
dependency
=
Jc
Jb
log speed
wa
elimination
parameter
wa–wb = z
wa+wb
log speed
a
–2
m
–3
||Ja+Jb –Jc || = m
–4
0
–5
0
||Jc ||
1
Figure 2: Why singularities are harmful for gradient-based learning. (a) Diagram of the analyzed
network and parameter reduction performed in the analysis in Wei et al. (2008). h = 0 corresponds
to the overlap singularity and z = ±1 correspond to the elimination singularities. (b) The gradient
flow field for the two-dimensional reduced system. The gradient norm is indicated by color. The
segment marked by the thick solid line is stable in this example; its basin of attraction is shaded
in gray. (c) Near-singularity plateau. Trajectory of learning dynamics starting from the black dot
indicated in b. Analysis and plots adapted from Wei et al. (2008). (d) Illustration of a linear dependence manifold in a toy model: the new coordinate m represents the distance to a particular linear
dependence manifold. (e) Gradient flow field for the toy model shown in d.
As mentioned earlier, linear dependence singularities arise exactly only in linear networks. However,
we expect them to hold approximately, and thus have consequences for learning, in the non-linear
case as well. Figure 2d-e shows an example in a toy single-layer nonlinear network: learning along
a linear dependence manifold, represented by m here, is much slower than learning along other
directions, e.g. the norm of the incoming weight vector Jc in the example shown here. Saxe et al.
(2013) demonstrated that this linear dependence problem arises generically, and becomes worse
with depth, in randomly initialized deep linear networks. Because learning is significantly slowed
down along linear dependence directions compared to other directions, these singularities effectively
reduce the dimensionality of the model, similarly to the overlap manifolds.
2.3 P LAIN NETWORKS ARE MORE DEGENERATE THAN NETWORKS WITH SKIP CONNECTIONS
To investigate the relationship between degeneracy, training difficulty and skip connections in deep
networks, we conducted several experiments with deep fully-connected networks. We compared
three different architectures. (i) The plain architecture is a fully-connected feedforward network
with no skip connections, described by the equation:
xl+1 = f (Wl xl + bl+1 )
l = 0, . . . , L − 1
where f is the ReLU nonlinearity and x0 denotes the input layer. (ii) The residual architecture introduces identity skip connections between adjacent layers (note that we do not allow skip connections
from the input layer):
x1 = f (W0 x0 + b1 ),
xl+1 = f (Wl xl + bl+1 ) + xl
l = 1, . . . , L − 1
(iii) The hyper-residual architecture adds skip connections between each layer and all layers above
it:
x1 = f (W0 x0 +b1 ), x2 = f (W1 x1 +b2 )+x1 , xl+1 = f (Wl xl +bl+1 )+xl +
l−1
1 X
Qk xk
l−1
l = 2, . . . , L−1
k=1
The skip connectivity from the immediately preceding layer is always the identity matrix, whereas
the remaining skip connections Qk are fixed, but allowed to be different from the identity (see
Supplementary Note 2 for further details). This architecture is inspired by the DenseNet architecture
(Huang et al., 2016). In both architectures, each layer projects skip connections to layers above it.
However, in the DenseNet architecture, the skip connectivity matrices are learned, whereas in the
hyper-residual architecture considered here, they are fixed.
In the experiments of this subsection, the networks all had L = 20 hidden layers (followed by
a softmax layer at the top) and n = 128 hidden units (ReLU) in each hidden layer. Hence, the
networks had the same total number of parameters. The biases were initialized to 0 and the weights
were initialized with the Glorot normal initialization scheme (Glorot & Bengio, 2010). The networks
were trained on the CIFAR-100 dataset (with coarse labels) using the Adam optimizer (Kingma &
3
CIFAR100 (20 layers)
50
0
Plain
Residual
Hyperres.
0
50
Epoch
100
b
c
100
99.9
0
50
100
Epoch
d
5e3
1e2
Ex. spectra after ep. 3
1e4
p(λ)
100
Negative eigs. (%)
Training acc. (%)
a
Degenerate eigs. (%)
Published as a conference paper at ICLR 2018
2.5e3
0
50
Epoch
100
1e6
1e8
50
0
50
λ
Figure 3: Model degeneracy increases training difficulty. (a) Training accuracy of different architectures. (b) Estimated fraction of degenerate eigenvalues during training. Error bars are standard
errors over 50 independent runs of the simulations. (c) Estimated fraction of negative eigenvalues
during training. (d) Example fitted spectra after 3 training epochs.
Ba, 2014) with learning rate 0.0005 and a batch size of 500. Because we are mainly interested
in understanding how singularities, and their removal, change the shape of the loss landscape and
consequently affect the optimization difficulty, we primarily monitor the training accuracy rather
than test accuracy in the results reported below.
To measure degeneracy, we estimated the eigenvalue density of the Hessian during training for
the three different network architectures. The probability of small eigenvalues in the eigenvalue
density reflects the dimensionality of the degenerate parameter space. To estimate this eigenvalue
density in our ∼ 1M-dimensional parameter spaces, we first estimated the first four moments of
the spectral density using the method of Skilling (Skilling, 1989) and fit the estimated moments
with a flexible mixture density model (see Supplementary Note 3 for details) consisting of a narrow
Gaussian component to capture the bulk of the spectral density, and a skew Gaussian density to
capture the tails (see Figure 3d for example fits). From the fitted mixture density, we estimated the
fraction of degenerate eigenvalues and the fraction of negative eigenvalues during training.
We validated our main results, as well as our mixture model for the spectral density, with smaller networks with ∼ 14K parameters where we could calculate all eigenvalues of the Hessian numerically
(Supplementary Note 4). For these smaller networks, the mixture model slightly underestimated the
fraction of degenerate eigenvalues and overestimated the fraction of negative eigenvalues; however,
there was a highly significant linear relationship between the actual and estimated fractions.
Figure 3b shows the evolution of the fraction of degenerate eigenvalues during training. A large
value at a particular point during optimization indicates a more degenerate model. By this measure,
the hyper-residual architecture is the least degenerate and the plain architecture is the most degenerate. We observe the opposite pattern for the fraction of negative eigenvalues (Figure 3c). The
differences between the architectures are more prominent early on in the training and there is an indication of a crossover later during training, with less degenerate models early on becoming slightly
more degenerate later on as the training performance starts to saturate (Figure 3b). Importantly, the
hyper-residual architecture has the highest training speed and the plain architecture has the lowest
training speed (Figure 3a), consistent with our hypothesis that the degeneracy of a model increases
the training difficulty and skip connections reduce the degeneracy.
2.4
T RAINING ACCURACY IS RELATED TO DISTANCE FROM DEGENERATE MANIFOLDS
To establish a more direct relationship between the elimination, overlap and linear dependence singularities discussed earlier on the one hand, and model degeneracy and training difficulty on the
other, we exploited the natural variability in training the same model caused by the stochasticity
of stochastic gradient descent (SGD) and random initialization. Specifically, we trained 100 plain
networks (30 hidden layers, 128 neurons per layer) on CIFAR-100 using different random initializations and random mini-batch selection. Training performance varied widely across runs. We
compared the best 10 and the worst 10 runs (measured by mean accuracy over 100 training epochs,
Figure 4a). The worst networks were more degenerate (Figure 4b); they were significantly closer to
elimination singularities, as measured by the average l2 -norm of the incoming weights of their hidden units (Figure 4c); they were significantly closer to overlap singularities (Figure 4d), as measured
by the mean correlation between the incoming weights of their hidden units; and their hidden units
4
Published as a conference paper at ICLR 2018
50
Epoch
100
0
50
Epoch
100
1
Mean overlap
99.99
e
d 0.02
0
50
Epoch
100
0
0
50
Epoch
100
Lin. dependence
0
c 1.4
Mean norm
Best 10
Worst 10
50
0
b 100
30 layers
Deg. eigs. (%)
Tr. accuracy (%)
a 100
1
0
0
50
100
Epoch
Figure 4: Training accuracy is correlated with distance from degenerate manifolds. (a) Training accuracies of the best 10 and the worst 10 plain networks trained on CIFAR-100. (b) Estimated fraction
of degenerate eigenvalues throughout training. Error bars are standard errors over 10 networks. (c)
Mean norm of the incoming weight vectors of hidden units. (d) Mean overlap of the weight vectors
of hidden units. (e) Linear dependence between hidden units in the same layer, measured by the
fraction of variance explained by the top three eigenmodes of their covariance matrix. These values
are averaged over the 30 layers of the network, yielding a single linear dependence score for each
network. For a replication of the results shown here for two other datasets, see Figure S7.
were significantly more linearly dependent (Figure 4e), as measured by the mean variance explained
by the top three eigenmodes of the covariance matrices of the hidden units in the same layer.
2.5
B ENEFITS OF SKIP CONNECTIONS AREN ’ T EXPLAINED BY GOOD INITIALIZATION ALONE
To investigate if the benefits of skip connections can be explained in terms of favorable initialization
of the parameters, we introduced a malicious initialization scheme for the residual network by subtracting the identity matrix from the initial weight matrices, Wl . If the benefits of skip connections
can be explained primarily by favorable initialization, this malicious initialization would be expected
to cancel the effects of skip connections at initialization and hence significantly deteriorate the performance. However, the malicious initialization only had a small adverse effect on the performance
of the residual network (Figure 5; ResMalInit), suggesting that the benefits of skip connections cannot be explained by favorable initialization alone. This result reveals a fundamental weakness in
previous explanations of the benefits of skip connections based purely on linear models (Hardt &
Ma, 2016; Li et al., 2016). In Supplementary Note 5 we show that skip connections do not eliminate
the singularities in deep linear networks, but only shift the landscape so that typical initializations
are farther from the singularities. Thus, in linear networks, any benefits of skip connections are due
entirely to better initialization. In contrast, skip connections genuinely eliminate the singularities
in nonlinear networks (Supplementary Note 1). The fact that malicious initialization of the residual
network does reduce its performance suggests that “ghosts” of these singularities still exist in the
loss landscape of nonlinear networks, but the performance reduction is only slight, suggesting that
skip connections alter the landscape around these ghosts to alleviate the learning slow-down that
would otherwise take place near them.
2.6 A LTERNATIVE WAYS OF ELIMINATING THE SINGULARITIES
If the success of skip connections can be attributed, at least partly, to eliminating singularities, then
alternative ways of eliminating them should also improve training. We tested this hypothesis by
introducing a particularly simple way of eliminating singularities: for each layer we drew random
target biases from a Gaussian distribution, N (µ, σ), and put an l2 -norm penalty on learned biases
deviating from those targets. This breaks the permutation symmetry between units and eliminates
the overlap singularities. In addition, positive µ values decrease the average threshold of the units
and make the elimination of units less likely (but not impossible), hence reducing the elimination
singularities. Decreased thresholds can also increase the dimensionality of the responses in a given
layer by reducing the fraction of times different units are identically zero, thereby making them less
linearly dependent. Note that setting µ = 0 and σ = 0 corresponds to the standard l2 -norm regularization of the biases, which does not eliminate any of the overlap or elimination singularities.
Hence, we expect the performance to be worse in this case than in cases with properly eliminated
singularities. On the other hand, although in general, larger values of µ and σ correspond to greater
elimination of singularities, the network also has to perform well in the classification task and very
large µ, σ values might be inconsistent with the latter requirement. Therefore, we expect the performance to be optimal for intermediate values of µ and σ. In the experiments reported below, we
optimized the hyperparameters µ, σ, and λ, i.e. the mean and the standard deviation of the target
5
Published as a conference paper at ICLR 2018
50
0
Residual
ResMalInit
BiasReg
BiasL2Reg (µ = 0, σ = 0)
Plain
0
50
Epoch
100
c
CIFAR100 (30 layers)
100
50
0
100
Degenerate eigs. (%)
Training accuracy (%)
b
CIFAR10 (30 layers)
100
Training accuracy (%)
a
0
50
Epoch
100
99.986
0
50
Epoch
100
Figure 5: Singularity elimination through bias regularization improves training. (a-b) Training accuracy of 30-layer networks on the CIFAR-10 and CIFAR-100 benchmarks. Error bars represent
±1 SEM over 50 independent runs. (c) Estimated fraction of degenerate eigenvalues in the plain,
residual and BiasReg networks.
bias distribution and the strength of the bias regularization term, through random search (Bergstra
& Bengio, 2012).
We trained 30-layer fully-connected feedforward networks on CIFAR-10 and CIFAR-100 datasets.
Figure 5a-b shows the training accuracy of different models on the two datasets. For both datasets,
among the models shown in Figure 5, the residual network performs the best and the plain network
the worst. Our simple singularity elimination through bias regularization scheme (BiasReg, cyan)
significantly improves performance over the plain network. Importantly, the standard l2 -norm regularization on the biases (BiasL2Reg (µ = 0, σ = 0), magenta) does not improve performance over
the plain network. These results are consistent with the singularity elimination hypothesis.
There is still a significant performance gap between our BiasReg network and the residual network
despite the fact that both break degeneracies. This can be partly attributed to the fact that the residual
network breaks the degeneracies more effectively than the BiasReg network (Figure 5c). Secondly,
even in models that completely eliminate the singularities, the learning speed would still depend on
the behavior of the gradient norms, and the residual network fares better than the BiasReg network
in this respect as well. At the beginning of training, the gradient norms with respect to the layer activities do not diminish in earlier layers of the residual network (Figure 6a, Epoch 0), demonstrating
that it effectively solves the vanishing gradients problem (Hochreiter, 1991; Bengio et al., 1994).
On the other hand, both in the plain network and in the BiasReg network, the gradient norms decay
quickly as one descends from the top of the network. Moreover, as training progresses (Figure 6a,
Epochs 1 and 2), the gradient norms are larger for the residual network than for the plain or the BiasReg network. Even for the maliciously initialized residual network, gradients do not decay quickly
at the beginning of training and the gradient norms behave similarly to those of the residual network
during training (Figure 6a; ResMalInit), suggesting that skip connections boost the gradient norms
near the ghosts of the singularities and reduce the learning slow-down that would otherwise take
place near them. Adding a single batch normalization layer (Ioffe & Szegedy, 2015) in the middle
of the BiasReg network alleviates the vanishing gradients problem for this network and brings its
performance closer to that of the residual network (Figure 6a-b; BiasReg+BN).
2.7
N ON - IDENTITY SKIP CONNECTIONS
If the singularity elimination hypothesis is correct, there should be nothing special about identity
skip connections. Skip connections other than identity should lead to training improvements if they
eliminate singularities. For the permutation symmetry breaking of the hidden units, ideally the skip
connection vector for each unit should disambiguate that unit maximally from all other units in that
layer. This is because as shown by the analysis in Wei et al. (2008) (Figure 2), even partial overlaps between hidden units significantly slow down learning (near-singularity plateaus). Mathematically, the maximal disambiguation requirement corresponds to an orthogonality condition on the
skip connectivity matrix (any full-rank matrix breaks the permutation symmetry, but only orthogonal matrices maximally disambiguate the units). Adding orthogonal vectors to different hidden units
is also useful for breaking potential (exact or approximate) linear dependencies between them. We
therefore tested random dense orthogonal matrices as skip connectivity matrices. Random dense
6
Published as a conference paper at ICLR 2018
Epoch 0
0.01
0
Epoch 1
b
Epoch 2
Tr. accuracy (%)
Mean grad. norm
a
Residual
ResMalInit
BiasReg+BN
BiasReg
Plain
1
15
30
Layer
1
15
30
1
15
30
100
50
0
0
50
Epoch
100
Figure 6: Skip connections effectively deal with the vanishing gradients problem. (a) Mean gradient
norms with respect to layer activities at the beginning of the first three epochs. Note that the mean
gradient norms of Plain and BiasReg networks are almost identical initially. (b) Training accuracy
of the networks on the CIFAR-100 dataset. The BiasReg network with a single batch normalization
layer inserted at layer 15 (BiasReg+BN) is shown in yellow. Its performance approaches the performance of the residual network. The results shown are averages over 50 independent runs. Standard
errors are small, hence are not shown for clarity.
1
2
4
8
16
32
64
128
identity
50
0
0
50
Epoch
100
CIFAR100 (30 layers)
100
b
Prob. of zero resp.
CIFAR10 (30 layers)
100
Training accuracy (%)
Training accuracy (%)
a
50
0
1
50
Epoch
100
CIFAR100 (30 layers)
0.07
0.03
identity
128 (dense ortho.)
1
50
Epoch
100
Figure 7: Random dense orthogonal skip connectivity matrices work slightly better than identity
skip connections. (a) Increasing the non-orthogonality of the skip connectivity matrix reduces the
performance (represented by lighter shades of gray). The results shown are averages over 10 independent runs of the simulations. (b) Probability of zero responses for residual networks with
identity skip connections (blue) and dense random orthogonal skip connections (black), averaged
over all hidden units and all training examples.
orthogonal matrices performed slightly better than identity skip connections in both CIFAR-10 and
CIFAR-100 datasets (Figure 7a, black vs. blue). This is because, even with skip connections, units
can be deactivated for some inputs because of the ReLU nonlinearity (recall that we do not allow
skip connections from the input layer). When this happens to a single unit at layer l, that unit is
effectively eliminated for that subset of inputs, hence eliminating the skip connection to the corresponding unit at layer l+1, if the skip connectivity is the identity. This causes a potential elimination
singularity for that particular unit. With dense skip connections, however, this possibility is reduced,
since all units in the previous layer are used. Moreover, when two distinct units at layer l are deactivated together, the identity skips cannot disambiguate the corresponding units at the next layer,
causing a potential overlap singularity. On the other hand, with dense orthogonal skips, because
all units at layer l are used, even if some of them are deactivated, the units at layer l + 1 can still
be disambiguated with the remaining active units. Figure 7b confirms for the CIFAR-100 dataset
that throughout most of the training, the hidden units of the network with dense orthogonal skip
connections have a lower probability of zero responses than those of the network with identity skip
connections.
Next, we gradually decreased the degree of “orthogonality” of the skip connectivity matrix to see
how the orthogonality of the matrix affects performance. Starting from a random dense orthogonal
matrix, we first divided the matrix into two halves and copied the first half to the second half. Starting
from n orthonormal vectors, this reduces the number of orthonormal vectors to n/2. We continued
on like this until the columns of the matrix were repeats of a single unit vector. We predict that as the
7
Published as a conference paper at ICLR 2018
0
1
64
128
Eigenvalue index
0
CIFAR100 (epoch 0)
c
= 0.025
= 0.02
= 0.015
= 0.01
= 0.005
=0
1
15
Layer
30
100
CIFAR100 (30 layers)
50
0
0
50
Epoch
100
d
Training accuracy (%)
0.5
0.05
Training accuracy (%)
b
1
Mean grad. norm
Eigvals of the cov. matrix
a
100
CIFAR10 (30 layers)
50
0
0
50
Epoch
100
Figure 8: Success of orthogonal skip connections cannot be explained by their ability to deal with
vanishing gradients only. (a) Eigenvalues of the covariance matrices with different τ values: τ = 0
corresponds to orthogonal skip connectivity matrices, larger τ values correspond to less orthogonal
matrices. Note that these eigenvalues are the eigenvalues of the covariance matrix of the skip connectivity vectors. The eigenvalue spectra of the skip connectivity matrices are always fixed to be
on the unit circle, hence equivalent to that of an orthogonal matrix. (b) Mean gradient norms with
respect to layer activations at the beginning of training. Gradients do not vanish in less orthogonal
skip connectivity matrices. If anything, gradient norms are typically larger with such matrices. (c-d)
Training accuracies on CIFAR-100 and CIFAR-10.
number of orthonormal vectors in the skip connectivity matrix is decreased, the performance should
deteriorate, because both the permutation symmetry-breaking capacity and the linear-dependencebreaking capacity of the skip connectivity matrix are reduced. Figure 7 shows the results for n = 128
hidden units. Darker colors correspond to “more orthogonal” matrices (e.g. “128” means all 128
skip vectors are orthonormal to each other, “1” means all 128 vectors are identical). The blue line is
the identity skip connectivity. More orthogonal skip connectivity matrices yield better performance,
consistent with our hypothesis.
The less orthogonal skip matrices also suffer from the vanishing gradients problem. So, their failure
could be partly attributed to the vanishing gradients problem. To control for this effect, we also designed skip connectivity matrices with eigenvalues on the unit circle (hence with eigenvalue spectra
equivalent to an orthogonal matrix), but with varying degrees of orthogonality (see Supplementary
Note 6 for details). More specifically, the columns (or rows) of an orthogonal matrix are orthonormal to each other, hence the covariance matrix of these vectors is the identity matrix. We designed
matrices where this covariance matrix was allowed to have non-zero off-diagonal values, reflecting
the fact that the vectors are not orthogonal any more. By controlling the magnitude of the correlations between the vectors, we manipulated the degree of orthogonality of the vectors. We achieved
this by setting the eigenvalue spectrum of the covariance matrix to be given by λi = exp(−τ (i − 1))
where λi denotes the i-th eigenvalue of the covariance matrix and τ is the parameter that controls the
degree of orthogonality: τ = 0 corresponds to the identity covariance matrix, hence to an orthonormal set of vectors, whereas larger values of τ correspond to gradually more correlated vectors. This
orthogonality manipulation was done while fixing the eigenvalue spectrum of the skip connectivity
matrix to be on the unit circle. Hence, the effects of this manipulation cannot be attributed to any
change in the eigenvalue spectrum, but only to the degree of orthogonality of the skip vectors. The
results of this experiment are shown in Figure 8. More orthogonal skip connectivity matrices still
perform better than less orthogonal ones (Figure 8c-d), even when their eigenvalue spectrum is fixed
and the vanishing gradients problem does not arise (Figure 8b), suggesting that the results of the
earlier experiment (Figure 7) cannot be explained solely by the vanishing gradients problem.
3 D ISCUSSION
In this paper, we proposed a novel explanation for the benefits of skip connections in terms of the
elimination of singularities. Our results suggest that elimination of singularities contributes at least
partly to the success of skip connections. However, we emphasize that singularity elimination is
not the only factor explaining the benefits of skip connections. Even in completely non-degenerate
models, other independent factors such as the behavior of gradient norms would affect training
performance. Indeed, we presented evidence suggesting that skip connections are also quite effective
at dealing with the problem of vanishing gradients and not every form of singularity elimination can
be expected to be equally good at dealing with such additional problems that beset the training of
deep networks.
8
Published as a conference paper at ICLR 2018
Alternative explanations: Several of our experiments rule out vanishing gradients as the sole explanation for training difficulties in deep networks and strongly suggest an independent role for the
singularities arising from the non-identifiability of the model. (i) In Figure 4, all nets have the exact
same plain architecture and similarly vanishing gradients at the beginning of training, yet they have
diverging performances correlated with measures of distance from singular manifolds. (ii) Vanishing gradients cannot explain the difference between identity skips and dense orthogonal skips in
Figure 7, because both eliminate vanishing gradients, yet dense orthogonal skips perform better. (iii)
In Figure 8, spectrum-equalized non-orthogonal skips often have larger gradient norms, yet worse
performance than orthogonal skips. (iv) Vanishing gradients cannot even explain the BiasReg results
in Figure 5. The BiasReg and the plain net have almost identical (and vanishing) gradients early on
in training (Figure 6a), yet the former has better performance as predicted by the symmetry-breaking
hypothesis. (v) Similar results hold for two-layer shallow networks where the problem of vanishing
gradients does not arise (Supplementary Note 7). In particular, shallow residual nets are less degenerate and have better accuracy than shallow plain nets; moreover, gradient norms and accuracy are
strongly correlated with distance from the overlap manifolds in these shallow nets.
Our malicious initialization experiment with residual nets (Figure 5) suggests that the benefits of
skip connections cannot be explained solely in terms of well-conditioning or improved initialization
either. This result reveals a fundamental weakness in purely linear explanations of the benefits of
skip connections (Hardt & Ma, 2016; Li et al., 2016). Unlike in nonlinear nets, improved initialization entirely explains the benefits of skip connections in linear nets (Supplementary Note 5).
A recent paper (Balduzzi et al., 2017) suggested that the loss of spatial structure in the covariance
of the gradients, a phenomenon called “shattered gradients”, could be partly responsible for training
difficulties in deep nonlinear networks. They argued that skip connections alleviate this problem by
essentially making the model “more linear”. It is easy to see that the shattered gradients problem
is distinct from both the vanishing/exploding gradients problem and the degeneracy problems considered in this paper, since shattered gradients arise only in sufficiently non-linear deep networks
(linear networks do not shatter gradients), whereas vanishing/exploding gradients, as well as the
degeneracies considered here, arise in linear networks too. The relative contribution of each of these
distinct problems to training difficulties in deep networks remains to be determined.
Symmetry-breaking in other architectures: We only reported results from experiments with fullyconnected networks, but we note that limited receptive field sizes and weight sharing between units
in a single feature channel in convolutional neural networks also reduce the permutation symmetry
in a given layer. The symmetry is not entirely eliminated since although individual units do not have
permutation symmetry in this case, feature channels do, but they are far fewer in number than the
number of individual units. Similarly, a recent extension of the residual architecture called ResNeXt
(Xie et al., 2016) uses parallel, segregated processing streams inside the “bottleneck” blocks, which
can again be seen as a way of reducing the permutation symmetry inside the block.
Our method of singularity reduction through bias regularization (BiasReg; Figure 5) can be thought
of as indirectly putting a prior over the unit activities. More complicated joint priors over hidden unit
responses that favor decorrelated (Cogswell et al., 2015) or clustered (Liao et al., 2016) responses
have been proposed before. Although the primary motivation for these regularization schemes was
to improve the generalizability or interpretability of the learned representations, they can potentially
be understood from a singularity elimination perspective as well. For example, a prior that favors
decorrelated responses can facilitate the breaking of permutation symmetries and linear dependencies between hidden units.
Our results lead to an apparent paradox: over-parametrization and redundancy in large neural network models have been argued to make optimization easier. Yet, our results seem to suggest the
opposite. However, there is no contradiction here. Any apparent contradiction is due to potential
ambiguities in the meanings of the terms “over-parametrization” and “redundancy”. The intuition
behind the benefits of over-parametrization for optimization is an increase in the effective capacity
of the model: over-parametrization in this sense leads to a large number of approximately equally
good ways of fitting the training data. On the other hand, the degeneracies considered in this paper
reduce the effective capacity of the model, leading to optimization difficulties. Our results suggest
that it could be useful for neural network researchers to pay closer attention to the degeneracies
inherent in their models. For better optimization, as a general design principle, we recommend reducing such degeneracies in a model as much as possible. Once the training performance starts to
9
Published as a conference paper at ICLR 2018
saturate, however, degeneracies may help the model achieve a better generalization performance.
Exploring this trade-off between the harmful and beneficial effects of degeneracies is an interesting
direction for future work.
Acknowledgments: AEO and XP were supported by the Intelligence Advanced Research Projects Activity
(IARPA) via Department of Interior/Interior Business Center (DoI/IBC) contract number D16PC00003. The
U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding
any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the
authors and should not be interpreted as necessarily representing the official policies or endorsements, either
expressed or implied, of IARPA, DoI/IBC, or the U.S. Government.
R EFERENCES
S. Amari, H. Park, and T. Ozeki. Singularities affect dynamics of learning in neuromanifolds. Neural
Comput., 18:1007–65, 2006.
A. Anandkumar and R. Ge. Efficient approaches for escaping higher order saddle points in nonconvex optimization. 2016. URL https://arxiv.org/abs/1602.05908.
D. Balduzzi, M. Frean, L. Leary, J.P. Lewis, K.W.-D. Ma, and B. McWilliams. The shattered
gradients problem: If resnets are the answer, then what is the question? ICML, 70:342–350,
2017.
Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is
difficult. IEEE Trans. Neural. Netw., 5:157–66, 1994.
J. Bergstra and Y. Bengio. Random search for hyper-parameter optimization. JMLR, 13:281–305,
2012.
M. Cogswell, F. Ahmed, R. Girshick, L. Zitnick, and D. Batra. Reducing overfitting in deep networks by decorrelating representations. 2015. URL https://arxiv.org/abs/1511.
06068.
X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks.
AISTAS, 9:249–256, 2010.
M. Hardt and T. Ma. Identity matters in deep learning. 2016. URL https://arxiv.org/abs/
1611.04231.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. 2015. URL
https://arxiv.org/abs/1512.03385.
K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks. 2016. URL
https://arxiv.org/abs/1603.05027.
S. Hochreiter. Untersuchungen zu dynamischen neuronalen Netzen. PhD thesis, Institut f. Informatik, Technische Univ. Munich, 1991.
G. Huang, Z. Liu, K.Q. Weinberger, and L. van der Maaten. Densely connected convolutional
networks. 2016. URL https://arxiv.org/abs/1608.06993.
S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. 2015. URL https://arxiv.org/abs/1502.03167.
D.P. Kingma and J.L. Ba. Adam: a method for stochastic optimization. 2014. URL https:
//arxiv.org/abs/1412.6980.
S. Li, J. Jiao, Y. Han, and T. Weissman. Demystifying resnet. 2016. URL https://arxiv.
org/abs/1611.01186.
R. Liao, A.G. Schwing, R.S. Zemel, and R. Urtasun. Learning deep parsimonious representations.
NIPS, pp. 5076–5084, 2016.
B.A. Pearlmutter. Fast exact multiplication by the hessian. Neural Comput., 6:147–60, 1994.
10
Published as a conference paper at ICLR 2018
D. Saad and S.A. Solla. On-line learning in soft committee machines. Phys. Rev. E, 52:4225, 1995.
A.M. Saxe, J.M. McClelland, and S. Ganguli. Exact solutions to the nonlinear dynamics of learning
in deep linear neural networks. 2013. URL https://arxiv.org/abs/1312.6120.
J. Skilling. The eigenvalues of mega-dimensional matrices. In Maximum Entropy and Bayesian
Methods, pp. 455–466. Kluwer Academic Publishers, 1989.
R.K. Srivastava, K. Greff, and J. Schmidhuber. Training very deep networks. NIPS, pp. 2377–2385,
2015.
H. Wei, J. Zhang, F. Cousseau, T. Ozeki, and S. Amari. Dynamics of learning near singularities in
layered networks. Neural Comput., 20:813–43, 2008.
S. Xie, R. Girshick, P. Dollar, Z. Tu, and K. He. Aggregated residual transformations for deep neural
networks. 2016. URL https://arxiv.org/abs/1611.05431.
S UPPLEMENTARY M ATERIALS
S UPPLEMENTARY
NETWORKS
N OTE 1: S INGULARITY OF THE H ESSIAN IN NON - LINEAR MULTILAYER
Because the cost function can be expressed as a sum over training examples, it is enough to consider
the cost for a single example: E = 12 ||y − xL ||2 ≡ 12 e> e, where xl are defined recursively as
xl = f (Wl−1 xl−1 ) for l = 1, . . . , L. We denote the inputs to units at layer l by the vector hl :
hl = Wl−1 xl−1 . We ignore the biases for simplicity. The derivative of the cost function with
respect to a single weight Wl,ij between layers l and l + 1 is given by:
>
0
..
.
∂E
0
>
0
>
0
>
= − f (hl+1,i )xl,j Wl+1
diag(fl+2
)Wl+2
diag(fl+3
) · · · WL−1
diag(fL0 )e
(1)
∂Wl,ij
.
..
0
Now, consider a different connection between the same output unit i at layer l + 1 and a different
input unit j 0 at layer l. The crucial thing to note is that if the units j and j 0 have the same set of
incoming weights, then the derivative of the cost function with respect to Wl,ij becomes identical to its derivative with respect to Wl,ij 0 : ∂E/∂Wl,ij = ∂E/∂Wl,ij 0 . This is because in this
condition xl,j 0 = xl,j for all possible inputs and all the remaining terms in Equation 1 are independent of the input index j. Thus, the columns (or rows) corresponding to the connections Wl,ij
and Wl,ij 0 in the Hessian become identical, making the Hessian degenerate. This is a re-statement
of the simple observation that when the units j and j 0 have the same set of incoming weights, the
parameters Wl,ij and Wl,ij 0 become non-identifiable (only their sum is identifiable). Thus, this corresponds to an overlap singularity. A similar argument shows that when a set of units at layer l, say
units indexed j, j 0 , j 00 become linearly dependent, the columns of the Hessian corresponding to the
weights Wl,ij , Wl,ij 0 and Wl,ij 00 become linearly dependent as well, thereby making the Hessian
singular. Again, this is just a re-statement of the fact that these weights are no longer individually
identifiable in this case (only a linear combination of them is identifiable). This corresponds to a
linear dependence singularity. In non-linear networks, except in certain degenerate cases where the
units saturate together, they may never be exactly linearly dependent, but they can be approximately
linearly dependent, which makes the Hessian close to singular.
Moreover, it is easy to see from Equation 1 that, when the presynaptic unit xl,j is always zero,
i.e. when that unit is effectively killed, the column (or row) of the Hessian corresponding to the
parameter Wl,ij becomes the zero vector for any i, and thus the Hessian becomes singular. This is
a re-statement of the simple observation that when the unit xl,j is always zero, its outgoing connections, Wl,ij , are no longer identifiable. This corresponds to an elimination singularity.
0
In the residual case, the only thing that changes in Equation 1 is that the factors Wk> diag(fk+1
) on
>
0
the right-hand side become Wk diag(fk+1 ) + I where I is an identity matrix of the appropriate size.
11
Published as a conference paper at ICLR 2018
The overlap singularities are eliminated, because xl,j 0 and xl,j cannot be the same for all possible
inputs in the residual case (even when the adjustable incoming weights of these units are identical).
Similarly, elimination singularities are also eliminated, because xl,j cannot be identically zero for all
possible inputs (even when the adjustable incoming weights of this unit are all zero), assuming that
the corresponding unit at the previous layer xl−1,j is not always zero, which, in turn, is guaranteed
with an identity skip connection if xl−2,j is not always zero etc., all the way down to the first hidden
layer. Any linear dependence between xl,j , xl,j 0 and xl,j 00 is also eliminated by adding linearly
independent inputs to them, assuming again that the corresponding units in the previous layer are
linearly independent.
S UPPLEMENTARY N OTE 2: S IMULATION DETAILS
In Figure 3, for the skip connections between non-adjacent layers in the hyper-residual networks,
i.e. Qk , we used matrices of the type labeled “32” in Figure 7, i.e. matrices consisting of four
copies of a set of 32 orthonormal vectors. We found that these matrices performed slightly better
than orthogonal matrices.
We augmented the training data in both CIFAR-10 and CIFAR-100 by adding reflected versions of
each training image, i.e. their mirror images. This yields a total of 100000 training images for both
datasets. The test data were not augmented, consisting of 10000 images in both cases. We used the
standard splits of the data into training and test sets.
For the BiasReg network of Figures 5-6, random hyperparameter search returned the following
values for the target bias distributions: µ = 0.51, σ = 0.96 for CIFAR-10 and µ = 0.91, σ = 0.03
for CIFAR-100.
The toy model shown in Figure 2b-c consists of the simulation of Equations 3.7 and 3.9 in Wei et al.
(2008). The toy model shown in Figure 2e is the simulation of learning dynamics in a network with
3 input, 3 hidden and 3 output units, parametrized in terms of the norms and unit-vector directions of
Ja − Jb − Jc , Ja + Jb − Jc , Jc , and the output weights. A teacher model with random parameters is
first chosen and a large set of “training data” is generated from the teacher model. Then the gradient
flow fields with respect to the two parameters m = ||Ja + Jb − Jc || and ||Jc || are plotted with
the assumption that the remaining parameters are already at their optima (a similar assumption was
made in the analysis of Wei et al. (2008)). We empirically confirmed that the flow field is generic.
S UPPLEMENTARY N OTE 3: E STIMATING THE EIGENVALUE SPECTRAL DENSITY OF THE
H ESSIAN IN DEEP NETWORKS
We use Skilling’s moment matching method (Skilling, 1989) to estimate the eigenvalue spectra
of the Hessian. We first estimate the first few non-central moments of the density by computing
mk = N1 r> Hk r where r is a random vector drawn from the standard multivariate Gaussian with
zero mean and identity covariance, H is the Hessian and N is the dimensionality of the parameter
space. Because the standard multivariate Gaussian is rotationally symmetric and the Hessian is a
symmetric matrix, it is easy to show that mk gives an unbiased estimate of the k-th moment of the
spectral density:
Z
N
1 > k
1 X 2 k
mk = r H r =
r̃ λ → p(λ)λk dλ as N → ∞
(2)
N
N i=1 i i
where λi are the eigenvalues of the Hessian, and p(λ) is the spectral density of the Hessian as
N → ∞. In Equation 2, we make use of the fact that r̃2i are random variables with expected value
1.
Despite appearances, the products in mk do not require the computation of the Hessian explicitly
and can instead be computed efficiently as follows:
v0 = r,
vk = Hvk−1 k = 1, . . . , K
(3)
where the Hessian times vector computation can be performed without computing the Hessian explicitly through Pearlmutter’s R-operator (Pearlmutter, 1994). In terms of the vectors vk , the estimates of the moments are given by the following:
1
1
m2k+1 = vk> vk+1
(4)
m2k = vk> vk ,
N
N
12
Published as a conference paper at ICLR 2018
For the results shown in Figure 3, we use 20-layer fully-connected feedforward networks and the
number of parameters is N = 709652. For the remaining simulations, we use 30-layer fullyconnected networks and the number of parameters is N = 874772.
We estimate the first four moments of the Hessian and fit the estimated moments with a parametric
density model. The parametric density model we use is a mixture of a narrow Gaussian distribution
(to capture the bulk of the density) and a skew-normal distribution (to capture the tails):
q(λ) = wSN (λ; ξ, ω, α) + (1 − w)N (λ; 0, σ = 0.001)
(5)
with 4 parameters in total: the mixture weight w, and the location ξ, scale ω and shape α parameters
of the skew-normal distribution. We fix the parameters of the Gaussian component to µ = 0 and
σ = 0.001. Since the densities are heavy-tailed, the moments are dominated by the tail behavior
of the model, hence the fits are not very sensitive to the precise choice of the parameters of the
Gaussian component. The moments of our model can be computed in closed-form. We had difficulty
fitting the parameters of the model with gradient-based methods, hence we used a simple grid search
method instead. The ranges searched over for each parameter was as follows. w: logarithmically
spaced between 10−9 and 10−3 ; α: linearly spaced between −50 and 50; ξ: linearly spaced between
−10 and 10; ω: logarithmically spaced between 10−1 and 103 . 100 parameters were evaluated along
each parameter dimension for a total of 108 parameter configurations evaluated.
The estimated moments ranged over several orders of magnitude. To make sure that the optimization
gave roughly equal weight to fitting each moment, we minimized a normalized objective function:
L(w, α, ξ, ω) =
4
X
|m̂k (w, α, ξ, ω) − mk |
|mk |
k=1
(6)
where m̂k (w, α, ξ, ω) is the model-derived estimate of the k-th moment.
S UPPLEMENTARY N OTE 4: VALIDATION OF THE RESULTS WITH SMALLER NETWORKS
Here, we validate our main results for smaller, numerically tractable networks. The networks in this
section are 10-layer fully-connected feedforward networks. The networks are trained on CIFAR100. The input dimensionality is reduced from 3072 to 128 through PCA. In what follows, we
calculate the fraction of degenerate eigenvalues by counting the number of eigenvalues inside a small
window of size 0.2 around 0, and the fraction of negative eigenvalues by the number of eigenvalues
to the left of this window.
We first compare residual networks with plain networks (Figure S1). The networks here have 16
hidden units in each layer yielding a total of 4852 parameters. This is small enough that we can
calculate all eigenvalues of the Hessian numerically. We observe that residual networks have better
training and test performance (Figure S1a-b); they are less degenerate (Figure S1d) and have more
negative eigenvalues than plain networks (Figure S1c). These results are consistent with the results
reported in Figure 3 for deeper and larger networks.
Next, we validate the results reported in Figure 4 by running 400 independent plain networks and
comparing the best-performing 40 with the worst-performing 40 among them (Figure S2). Again,
the networks here have 16 hidden units in each layer with a total of 4852 parameters. We observe
that the best networks are less degenerate (Figure S2d) and have more negative eigenvalues than
the worst networks (Figure S2c). Moreover, the hidden units of the best networks have less overlap
(Figure S2f), and, at least initially during training, have slightly larger weight norms than the worstperforming networks (Figure S2e). Again, these results are all consistent with those reported in
Figure 4 for deeper and larger networks.
Finally, using numerically tractable plain networks, we also tested whether we could reliably estimate the fractions of degenerate and negative eigenvalues with our mixture model. Just as we do
for the larger networks, we first fit the mixture model to the first four moments of the spectral density estimated with the method of Skilling (1989). We then estimate the fraction of degenerate and
negative eigenvalues from the fitted mixture model and compare these estimates with those obtained
from the numerically calculated actual eigenvalues. Because for the larger networks, the networks
were found to be highly degenerate, we restrict the analysis here to conditions where the fraction
of degenerate eigenvalues was at least 99.8%. We used 10-layer plain networks with 32 hidden
13
Published as a conference paper at ICLR 2018
40
b
Plain
Residual
0
0
150
Epoch
0
300
0
d 100
150
300
150
300
Epoch
Degenerate eigs. (%)
6
Negative eigs. (%)
c
40
Test accuracy (%)
Training accuracy (%)
a
0
0
150
Epoch
300
75
0
Epoch
Figure S1: Validation of the results with 10-layer plain and residual networks trained on CIFAR-100.
(a-b) Training and test accuracy. (c-d) Fraction of negative and degenerate eigenvalues throughout
training. The results are averages over 4 independent runs ±1 standard errors.
14
Published as a conference paper at ICLR 2018
a 35
0
c
0
15
0
30
0
e 1.2
15
Test accuracy (%)
0
d100
Negative eigs. (%)
5
0
b 35
Training accuracy (%)
Best 40
Worst 40
75
30
30
Degenerate eigs. (%)
0
f0.0025
Mean norm
15
15
30
Mean overlap
0
0.0025
0.005
1
0
15
Epoch
30
0.0075
0
15
Epoch
30
Figure S2: Validation of the results with 400 10-layer plain networks with 16 hidden units in each
layer (4852 parameters total) trained on CIFAR-100. We compare the best 40 networks with the
worst 40 networks, as in Figure 4. (a-b) Training and test accuracy. (c-d) Fraction of negative
and degenerate eigenvalues throughout training. Better performing networks are less degenerate and
have more negative eigenvalues. (e) Mean norms of the incoming weight vectors of the hidden units.
(f) Mean overlaps of the hidden units as measured by the mean correlation between their incoming
weight vectors. The results are averages over 40 best or worst runs ±1 standard errors.
15
Published as a conference paper at ICLR 2018
Degenerate eigs. (%)
99.95
0.04
99.90
0.03
99.85
99.80
99.75
99.75
Negative eigs. (%)
0.05
Estimated
Estimated
100
0.02
0.01
99.80
99.85
99.90
99.95
100
0
0
0.01
0.02
Actual
0.03
0.04
0.05
Actual
Figure S3: For 10-layer plain networks with 32 hidden units in each layer (14292 parameters total),
estimates obtained from the mixture model slightly underestimate the fraction of degenerate eigenvalues, and overestimate the fraction of negative eigenvalues; however, there is a highly significant
linear relationship between the actual values and the estimates. (a) Actual vs. estimated fraction of
degenerate eigenvalues. (b) Actual vs. estimated fraction of negative eigenvalues for the same networks. Dashed line shows the identity line. Dots and errorbars represent means and standard errors
of estimates in different bins; the solid lines and the shaded regions represent the linear regression
fits and the 95% confidence intervals.
units in each layer (with a total of 14292 parameters) for this analysis. We observe that, at least for
these small networks, the mixture model usually underestimates the fraction of degenerate eigenvalues and overestimates the fraction of negative eigenvalues. However, there is a highly significant
positive correlation between the actual and estimated fractions (Figure S3).
S UPPLEMENTARY
CONNECTIONS
N OTE 5: DYNAMICS OF LEARNING IN LINEAR NETWORKS WITH SKIP
To get a better analytic understanding of the effects of skip connections on the learning dynamics,
we turn to linear networks. In an L-layer linear plain network, the input-output mapping is given by
(again ignoring the biases for simplicity):
xL = WL−1 WL−2 . . . W1 x1
(7)
where x1 and xL are the input and output vectors, respectively. In linear residual networks with
identity skip connections between adjacent layers, the input-output mapping becomes:
xL = (WL−1 + I)(WL−2 + I) . . . (W1 + I)x1
(8)
Finally, in hyper-residual linear networks where all skip connection matrices are assumed to be the
identity, the input-output mapping is given by:
xL = WL−1 + (L − 1)I WL−2 + (L − 2)I . . . W1 + I x1
(9)
In the derivations to follow, we do not have to assume that the connectivity matrices are square
matrices. If they are rectangular matrices, the identity matrix I should be interpreted as a rectangular
identity matrix of the appropriate size. This corresponds to zero-padding the layers when they are
not the same size, as is usually done in practice.
Three-layer networks: Dynamics of learning in plain linear networks with no skip connections
was analyzed in Saxe et al. (2013). For a three-layer network (L = 3), the learning dynamics can be
expressed by the following differential equations (Saxe et al., 2013):
X
d
τ aα = (sα − aα · bα )bα −
(aα · bγ )bγ
(10)
dt
γ6=α
d
τ bα
dt
=
(sα − aα · bα )aα −
X
γ6=α
16
(aγ · bα )aγ
(11)
Published as a conference paper at ICLR 2018
Here aα and bα are n-dimensional column vectors (where n is the number of hidden units) connecting the hidden layer to the α-th input and output modes, respectively, of the input-output correlation
matrix and sα is the corresponding singular value (see Saxe et al. (2013) for further details). The
first term on the right-hand side of Equations 10-11 facilitates cooperation between aα and bα corresponding to the same input-output mode α, while the second term encourages competition between
vectors corresponding to different modes.
In the simplest scenario where there are only two input and output modes, the learning dynamics of
Equations 10, 11 reduces to:
d 1
a = (s1 − a1 · b1 )b1 − (a1 · b2 )b2
(12)
dt
d 2
a = (s2 − a2 · b2 )b2 − (a2 · b1 )b1
(13)
dt
d 1
b = (s1 − a1 · b1 )a1 − (a1 · b2 )a2
(14)
dt
d 2
b = (s2 − a2 · b2 )a2 − (a2 · b1 )a1
(15)
dt
How does adding skip connections between adjacent layers change the learning dynamics? Considering again a three-layer network (L = 3) with only two input and output modes, a straightforward
extension of Equations 12-15 shows that the learning dynamics changes as follows:
h
i
h
i
d 1
a = s1 − (a1 + v 1 ) · (b1 + u1 ) (b1 + u1 ) − (a1 + v 1 ) · (b2 + u2 ) (b2 + u2 ) (16)
dt
h
i
h
i
d 2
a = s2 − (a2 + v 2 ) · (b2 + u2 ) (b2 + u2 ) − (a2 + v 2 ) · (b1 + u1 ) (b1 + u1 ) (17)
dt
h
i
h
i
d 1
b = s1 − (a1 + v 1 ) · (b1 + u1 ) (a1 + v 1 ) − (a1 + v 1 ) · (b2 + u2 ) (a2 + v 2 ) (18)
dt
h
i
h
i
d 2
b = s2 − (a2 + v 2 ) · (b2 + u2 ) (a2 + v 2 ) − (a2 + v 2 ) · (b1 + u1 ) (a1 + v 1 ) (19)
dt
where u1 and u2 are orthonormal vectors (similarly for v 1 and v 2 ). The derivation proceeds essentially identically to the corresponding derivation for plain networks in Saxe et al. (2013). The
only differences are: (i) we substitute the plain weight matrices Wl with their residual counterparts
Wl + I and (ii) when changing the basis from the canonical basis for the weight matrices W1 , W2
to the input and output modes of the input-output correlation matrix, U and V, we note that:
W2 + I =
W1 + I =
UW2 + UU> = U(W2 + U> )
>
>
W1 V + VV = (W1 + V)V
α
α
α
(20)
>
(21)
α
where U and V are orthogonal matrices and the vectors a , b , u and v in Equations 16-19
>
correspond to the α-th columns of the matrices W1 , W2 , U and V, respectively.
Figure S4 shows, for two different initializations, the evolution of the variables a1 and a2 in plain
and residual networks with two input-output modes and two hidden units. When the variables are
initialized to small random values, the dynamics in the plain network initially evolves slowly (Figure S4a, blue); whereas it is much faster in the residual network (Figure S4a, red). This effect is
attributable to two factors. First, the added orthonormal vectors uα and v α increase the initial velocity of the variables in the residual network. Second, even when we equalize the initial norms of
the vectors, aα and aα + v α (and those of the vectors bα and bα + uα ) in the plain and the residual
networks, respectively, we still observe an advantage for the residual network (Figure S4b), because
the cooperative and competitive terms are orthogonal to each other in the residual network (or close
to orthogonal, depending on the initialization of aα and bα ; see right-hand side of Equations 16-19),
whereas in the plain network they are not necessarily orthogonal and hence can cancel each other
(Equations 12-15), thus slowing down convergence.
Singularity of the Hessian in linear three-layer networks: The dynamics in Equations 10, 11 can
be interpreted as gradient descent on the following energy function:
1 X
1 X α β 2
E=
(sα − aα · bα )2 +
(a · b )
(22)
2τ α
2τ
α6=β
17
1.5
0
-1.5
1.5
a2
a
a1
Published as a conference paper at ICLR 2018
plain
residual
0
200
0
-1.5
400
0
Time
2
1
1
0
400
Time
a2
a1
b
200
0
200
0
-0.5
400
Time
0
200
400
Time
Figure S4: Evolution of a1 and a2 in linear plain and residual networks (evolution of b1 and b2
proceeds similarly). The weights converge faster in residual networks. Simulation details are as
follows: the number of hidden units is 2 (the two solid lines for each color represent the weights
1
associated with the two hidden nodes, e.g. a11 and a√
2 on the
√ left), the singular values√are s1 =
√ 3.0,
s2 = 1.5. For the residual network, u1 = v1 = [1/ 2, 1/ 2]> and u2 = v2 = [1/ 2, −1/ 2]> .
In (a), the weights of both plain and residual networks are initialized to random values drawn from
a Gaussian with zero mean and standard deviation of 0.0001. The learning rate was set to 0.1. In
1
(b), the
the plain network are initialized as follows: the
and a2 are initialized
√ weights
√ of
√vectors√a >
>
1
2
to [1/ 2, 1/ 2] and the vectors b and b are initialized to [1/ 2, −1/ 2] ; the weights of the
residual network are all initialized to zero, thus equalizing the initial norms of the vectors aα and
aα + v α (and those of the vectors bα and bα + uα ) between the plain and residual networks. The
residual network still converges faster than the plain network. In (b), the learning rate was set to
0.01 to make the different convergence rates of the two networks more visible.
18
Published as a conference paper at ICLR 2018
This energy function is invariant to a (simultaneous) permutation of the elements of the vectors aα
and bα for all α. This causes degenerate manifolds in the landscape. Specifically, for the permutation
α
symmetry of hidden units, these manifolds are the hyperplanes aα
i = aj ∀α, for each pair of hidden
α
α
units i, j (similarly, the hyperplanes bi = bj ∀α) that make the model non-identifiable. Formally,
these correspond to the singularities of the Hessian or the Fisher information matrix. Indeed, we
α
shall quickly check below that when aα
i = aj ∀α for any pair of hidden units i, j, the Hessian
becomes singular (overlap singularities). The Hessian also has additional singularities at the hyperα
planes aα
i = 0 ∀α for any i and at bi = 0 ∀α for any i (elimination singularities).
Starting from the energy function in Equation 22 and taking the derivative with respect to a single
input-to-hidden layer weight, aα
i :
X
∂E
α
α α
=
−(s
−
a
·
b
)b
+
(aα · bβ )bβi
α
i
∂aα
i
(23)
β6=α
and the second derivatives are as follows:
∂2E
2
∂(aα
i )
=
∂2E
α
∂aα
i ∂aj
=
X
2
(bα
i ) +
(bβi )2 =
β6=α
α
bα
j bi +
X
X β
(bi )2
(24)
β
bβj bβi =
β6=α
X
bβi bβj
(25)
β
Note that the second derivatives are independent of mode index α, reflecting the fact that the energy
function is invariant to a permutation of the mode indices. Furthermore, when bβi = bβj for all β,
α
the columns in the Hessian corresponding to aα
i and aj become identical, causing an additional
α
degeneracy reflecting the non-identifiability of ai and aα
j . A similar derivation establishes that
aβi = aβj for all β also leads to a degeneracy in the Hessian, this time reflecting the non-identifiability
α
of bα
i and bj . These correspond to the overlap singularities.
In addition, it is easy to see from Equations 24, 25 that when bα
i = 0 ∀α, the right-hand sides of
both equations become identically zero, reflecting the non-identifiability of aα
i for all α. A similar
derivation shows that when aα
=
0
∀α,
the
columns
of
the
Hessian
corresponding
to bα
i
i become
α
identically zero for all α, this time reflecting the non-identifiability of bi for all α. These correspond
to the elimination singularities.
When we add skip connections between adjacent layers, i.e. in the residual architecture, the energy
function changes as follows:
1X α
1X
(sα − (aα + v α ) · (bα + uα ))2 +
((a + v α ) · (bβ + uβ ))2
(26)
E=
2 α
2
α6=β
and straightforward algebra yields the following second derivatives:
∂2E
2
∂(aα
i )
=
∂2E
α
∂aα
i ∂aj
=
X β
(bi + uβi )2
(27)
β
X β
(bi + uβi )(bβj + uβj )
(28)
β
Unlike in the plain network, setting bβi = bβj for all β, or setting bα
i = 0 ∀α, does not lead to a
degeneracy here, thanks to the orthogonal skip vectors uβ . However, this just shifts the locations of
the singularities. In particular, the residual network suffers from the same overlap and elimination
singularities as the plain network when we make the following change of variables: bβ → bβ − uβ
and aβ → aβ − v β .
Networks with more than three-layers: As shown in Saxe et al. (2013), in linear networks with
more than a single hidden layer, assuming that there are orthogonal matrices Rl and Rl+1 for each
layer l that diagonalize the initial weight matrix of the corresponding layer (i.e. R>
l+1 Wl (0)Rl =
Dl is a diagonal matrix), dynamics of different singular modes decouple from each other and each
19
b
a
2
Plain
0
-2
-2
Residual
2
Hyper−residual
2
0
0
a
2
0
-2
-2
0
-2
-2
2
0
b
2
u
Published as a conference paper at ICLR 2018
1
2
0
hyper-residual
residual
plain
0
200
400
Time
Figure S5: (a) Phase portraits for three-layer plain, residual and hyper-residual linear networks. (b)
QNl −1
Evolution of u = l=1
al for 10-layer plain, residual and hyper-residual linear networks. In the
plain network, u did not converge to its asymptotic value s within the simulated time window.
mode α evolves according to gradient descent dynamics in an energy landscape described by (Saxe
et al., 2013):
NY
l −1
2
1
Eplain =
sα −
aα
(29)
l
2τ
l=1
where aα
l can be interpreted as the strength of mode α at layer l and Nl is the total number of layers.
In residual networks, assuming further that the orthogonal matrices Rl satisfy R>
l+1 Rl = I, the
energy function changes to:
Eres
NY
l −1
2
1
=
sα −
(aα
l + 1)
2τ
(30)
l=1
and in hyper-residual networks, it is:
Ehyperres =
NY
l −1
2
1
sα −
(aα
l + l)
2τ
(31)
l=1
Figure S5a illustrates the effect of skip connections on the phase portrait of a three layer network.
The two axes, a and b, represent the mode strength variables for l = 1 and l = 2, respectively:
α
i.e. a ≡ aα
1 and b ≡ a2 . The plain network has a saddle point at (0, 0) (Figure S5a; left). The
dynamics around this point is slow, hence starting from small random values causes initially very
slow learning. The network funnels the dynamics through the unstable manifold a = b to the stable
hyperbolic solution corresponding to ab = s. Identity skip connections between adjacent layers in
the residual architecture move the saddle point to (−1, −1) (Figure S5a; middle). This speeds up the
dynamics around the origin, but not as much as in the hyper-residual architecture where the saddle
point is moved further away from the origin and the main diagonal to (−1, −2) (Figure S5a; right).
We found these effects to be more pronounced in deeper networks. Figure S5b shows the dynamics
of learning in 10-layer linear networks, demonstrating a clear advantage for the residual architecture
over the plain architecture and for the hyper-residual architecture over the residual architecture.
Singularity of the Hessian in reduced linear multilayer networks with skip connections: The
derivative of the cost function of a linear multilayer residual network (Equation 30) with respect
to the mode strength variable at layer i, ai , is given by (suppressing the mode index α and taking
τ = 1):
Y
∂E
= −(s − u) (al + 1)
(32)
∂ai
l6=i
and the second derivatives are:
∂2E
∂a2i
∂2E
∂ai ∂ak
=
hY
i2
(al + 1)
(33)
l6=i
=
i Y
h Y
2 (al + 1) − s
(al + 1)
l
l6=i,k
20
(34)
Published as a conference paper at ICLR 2018
It is easy to check that the columns (or rows) corresponding to ai and aj in the Hessian become
identical when ai = aj , making the Hessian degenerate. The hyper-residual architecture does not
eliminate these degeneracies but shifts them to different locations in the parameter space by adding
distinct constants to ai and aj (and to all other variables).
S UPPLEMENTARY N OTE 6: D ESIGNING SKIP CONNECTIVITY MATRICES WITH VARYING
DEGREES OF ORTHOGONALITY AND WITH EIGENVALUES ON THE UNIT CIRCLE
We generated the covariance matrix of the eigenvectors by S = QΛQ> , where Q is a random orthogonal matrix and Λ is the diagonal matrix of eigenvalues, Λii = exp(−τ (i − 1)), as explained in
the main text. We find the correlation matrix through R = D−1/2 SD−1/2 where D is the diagonal
matrix of the variances: i.e. Dii = Sii . We take the Cholesky decomposition of the correlation
matrix, R = TT> . Then the designed skip connectivity matrix is given by Σ = TULU−1 T−1 ,
where L and U are the matrices of eigenvalues and eigenvectors of another randomly generated orthogonal matrix, O: i.e. O = ULU> . With this construction, Σ has the same eigenvalue spectrum
as O, however the eigenvectors of Σ are linear combinations of the eigenvectors of O such that
their correlation matrix is given by R. Thus, the eigenvectors of Σ are not orthogonal to each other
unless τ = 0. Larger values of τ yield more correlated, hence less orthogonal, eigenvectors.
S UPPLEMENTARY N OTE 7: VALIDATION OF THE RESULTS WITH SHALLOW NETWORKS
To further demonstrate the generality of our results and the independence of the problem of singularities from the vanishing gradients problem in optimization, we performed an experiment with
shallow plain and residual networks with only two hidden layers and 16 units in each hidden layer.
Because we do not allow skip connections from the input layer, a network with two hidden layers is
the shallowest network we can use to compare the plain and residual architectures. Figure S6 shows
the results of this experiment. The residual network performs slightly better both on the training
and test data (Figure S6a-b); it is less degenerate (Figure S6d) and has more negative eigenvalues
(Figure S6c); it has larger gradients (Figure S6e) —note that the gradients in the plain network do
not vanish even at the beginning of training— and its hidden units have less overlap than the plain
network (Figure S6f). Moreover, the gradient norms closely track the mean overlap between the
hidden units and the degeneracy of the network (Figure S6d-f) throughout training. These results
suggest that the degeneracies caused by the overlaps of hidden units slow down learning, consistent
with our symmetry-breaking hypothesis and with the results from larger networks.
21
Published as a conference paper at ICLR 2018
40
b
Training acc. (%)
Plain
Residual
0
c
0
d
100
Negative eigs. (%)
Degenerate eigs. (%)
5
f
Mean grad. norm
Mean overlap
0
e 0.0015
0
40
Test acc. (%)
a
Layer 1
Layer 2
0
5
10
Epoch
15
20
75
0.02
0
0.02
25
0
5
10
Epoch
15
20
25
10
20
99.99
0
Epoch
100
30 layers (STL10)
50
0
Best 10
Worst 10
0
25
Epoch
10
20
1
0
Epoch
50
0
25
Epoch
20
0
0
Epoch
100
99.999
10
0.01
50
0
25
Epoch
20
1
0.5
0
Epoch
1.1
1
10
Lin. dependence
0
1.1
50
0.02
0
0
25
Epoch
10
20
Epoch
50
Lin. dependence
Best 10
Worst 10
1.2
Mean overlap
0
100
Mean overlap
50
Mean norm
30 layers (SVHN)
Mean norm
Tr. accuracy (%)
b
100
Deg. eigs. (%)
Tr. accuracy (%)
a
Deg. eigs. (%)
Figure S6: Main results hold for two-layer shallow nets trained on CIFAR-100. (a-b) Training and
test accuracies. Residual nets perform slightly better. (c-d) Fraction of negative and degenerate
eigenvalues. Residual nets are less degenerate. (e) Mean gradient norms with respect to the two
layer activations throughout training. (f) Mean overlap for the second hidden layer units, measured
as the mean correlation between the incoming weights of the hidden units. Results in (a-e) are
averages over 16 independent runs; error bars are small, hence not shown for clarity. In (f), error
bars represent standard errors.
1
0.5
0
25
50
Epoch
Figure S7: Replication of the results reported in Figure 4 for (a) the Street View House Numbers
(SVHN) dataset and (b) the STL-10 dataset.
22
| 9cs.NE
|
LabelFusion: A Pipeline for Generating Ground Truth Labels for Real
RGBD Data of Cluttered Scenes
arXiv:1707.04796v3 [cs.CV] 26 Sep 2017
Pat Marion*, Peter R. Florence*, Lucas Manuelli* and Russ Tedrake
Computer Science and Artificial Intelligence Laboratory
Massachusetts Institute of Technology
{pmarion,peteflo,manuelli,russt}@csail.mit.edu
∗ These authors contributed equally to this work.
Abstract— Deep neural network (DNN) architectures have
been shown to outperform traditional pipelines for object
segmentation and pose estimation using RGBD data, but the
performance of these DNN pipelines is directly tied to how
representative the training data is of the true data. Hence a
key requirement for employing these methods in practice is
to have a large set of labeled data for your specific robotic
manipulation task, a requirement that is not generally satisfied
by existing datasets. In this paper we develop a pipeline to
rapidly generate high quality RGBD data with pixelwise labels
and object poses. We use an RGBD camera to collect video
of a scene from multiple viewpoints and leverage existing
reconstruction techniques to produce a 3D dense reconstruction.
We label the 3D reconstruction using a human assisted ICPfitting of object meshes. By reprojecting the results of labeling
the 3D scene we can produce labels for each RGBD image of the
scene. This pipeline enabled us to collect over 1,000,000 labeled
object instances in just a few days. We use this dataset to answer
questions related to how much training data is required, and
of what quality the data must be, to achieve high performance
from a DNN architecture. Our dataset and annotation pipeline
are available at labelfusion.csail.mit.edu.
I. I NTRODUCTION
Advances in neural network architectures for deep learning
have made significant impacts on perception for robotic
manipulation tasks. State of the art networks are able to
produce high quality pixelwise segmentations of RGB images,
which can be used as a key component for 6DOF object pose
estimation in cluttered environments [1], [2]. However for
a network to be useful in practice it must be fine tuned
on labeled scenes of the specific objects targeted by the
manipulation task, and these networks can require tens to
hundreds of thousands of labeled training examples to achieve
adequate performance. To acquire sufficient data for each
specific robotics application using once-per-image human
labeling would be prohibitive, either in time or money.
While some work has investigated closing the gap with
simulated data [3], [4], [5], [6], our method can scale to
these magnitudes with real data.
In this paper we tackle this problem by developing an
open-source pipeline that vastly reduces the amount of human
annotation time needed to produce labeled RGBD datasets for
training image segmentation neural networks. The pipeline
produces ground truth segmentations and ground truth 6DOF
poses for multiple objects in scenes with clutter, occlusions,
and varied lighting conditions. The key components of the
pipeline are: leveraging dense RGBD reconstruction to fuse
together RGBD images taken from a variety of viewpoints,
labeling with ICP-assisted fitting of object meshes, and
automatically rendering labels using projected object meshes.
These techniques allow us to label once per scene, with each
scene containing thousands of images, rather than having to
annotate images individually. This reduces human annotation
time by several orders of magnitude over traditional techniques. We optimize our pipeline to both collect many views
of a scene and to collect many scenes with varied object
arrangements. Our goal is to enable manipulation researchers
and practitioners to generate customized datasets, which for
example can be used to train any of the available state-of-theart image segmentation neural network architectures. Using
this method we have collected over 1,000,000 labeled object
instances in multi-object scenes, with only a few days of data
collection and without using any crowd sourcing platforms
for human annotation.
Our primary contribution is the pipeline to rapidly generate
labeled data, which researchers can use to build their own
datasets, with the only hardware requirement being the RGBD
sensor itself. We also have made available our own dataset,
which is the largest available RGBD dataset with objectpose labels (352,000 labeled images, 1,000,000+ object
instances). Additionally, we contribute a number of empirical
results concerning the use of large datasets for practical
deep-learning-based pixelwise segmentation of manipulationrelevant scenes in clutter – specifically, we empirically
quantify the generalization value of varying aspects of the
training data: (i) multi-object vs single object scenes, (ii) the
number of background environments, and (iii) the number of
views per scene.
II. R ELATED W ORK
We review three areas of related work. First, we review
pipelines for generating labeled RGBD data. Second, we
review applications of this type of labeled data to 6DOF object
pose estimation in the context of robotic manipulation tasks.
Third, we review work related to our empirical evaluations,
concerning questions of scale and generalization for practical
learning in robotics-relevant contexts.
A. Methods for Generating Labeled RGBD Datasets
Rather than evaluate RGBD datasets based on the specific
dataset they provide, we evaluate the methods used to generate
them, and how well they scale. Firman [7] provides an
extensive overview of over 100 available RGBD datasets.
Only a few of the methods used are capable of generating
labels for 6DOF object poses, and none of these associated
datasets also provide per-pixel labeling of objects. One of
the most related methods to ours is that used to create the TLESS dataset [8], which contains approximately 49K RGBD
images of textureless objects labeled with the 6DOF pose of
each object. Compared to our approach, [8] requires highly
calibrated data collection equipment. They employ fiducials
for camera pose tracking which limits the ability of their
method to operate in arbitrary environments. Additionally
the alignment of the object models to the pointcloud is a
completely manual process with no algorithmic assistance.
Similarly, [1] describes a high-precision motion-capture-based
approach, which does have the benefit of generating highfidelity ground-truth pose, but its ability to scale to large
scale data generation is limited by: the confines of the motion
capture studio, motion capture markers on objects interfering
with the data collection, and time-intensive setup for each
object.
Although the approach is not capable of generating the
6 DOF poses of objects, a relevant method for per-pixel
labeling is described in [2]. They employ an automated data
collection pipeline in which the key idea is to use background
subtraction. Two images are taken with the camera at the
exact same location – in the first, no object is present, while
it is in the second. Background subtraction automatically
yields a pixelwise segmentation of the object. Using this
approach they generate 130,000 labeled images for their 39
objects. As a pixelwise labeling method, there are a few
drawbacks to this approach. The first is that in order to apply
the background subtraction method, they only have a single
object present in each scene. In particular there are no training
images with occlusions. They could in theory extend their
method to support multi-object scenes by adding objects to
the scene one-by-one, but this presents practical challenges.
Secondly the approach requires an accurately calibrated robot
arm to move the camera in a repeatable way. A benefit of the
method, however, is that it does enable pixelwise labeling of
even deformable objects.
The SceneNN [9] and ScanNet [10] data generation
pipelines share some features with our method. They both use
an RGBD sensor to produce a dense 3D reconstruction and
then perform annotations in 3D. However, since SceneNN and
ScanNet are focused on producing datasets for RGDB scene
understanding tasks, the type of annotation that is needed is
quite different. In particular their methods provide pixelwise
segmenation into generic object classes (floor, wall, couch
etc.). Neither SceneNN or ScaneNet have gometric models
for the specific objects in a scene and thus cannot provide
6DOF object poses. Whereas ScanNet and SceneNN focus
on producing datasets for benchmarking scene understanding
algorithms, we provide a pipeline to enable rapid generation
labeled data for your particular application and object set.
B. Object-Specific Pose Estimation in Clutter for Robotic
Manipulation
There have been a wide variety of methods to estimate
object poses for manipulation. A challenge is object specificity.
[1] and [2] are both state of the art pipelines for estimating
object poses from RGBD images in clutter – both approaches
use RGB pixelwise segmentation neural networks (trained
on their datasets described in the previous section) to crop
point clouds which are then fed into ICP-based algorithms
to estimate object poses by registering against prior known
meshes. Another approach is to directly learn pose estimation
[11]. The upcoming SIXD Challenge 2017 [12] will provide
a comparison of state of the art methods for 6DOF pose
estimation on a common dataset. The challenge dataset
contains RGBD images annotated with ground truth 6DOF
object poses. This is exactly the type of data produced by our
pipeline and we aim aim to submit our dataset to the 2018
challenge. There is also a trend in manipulation research
to bypass object pose estimation and work directly with
the raw sensor data [13], [14], [15]. Making these methods
object-specific in clutter could be aided by using the pipeline
presented here to train segmentation networks.
C. Empirical Evaluations of Data Requirements for Image
Segmentation Generalization
While the research community is more familiar with the
scale and variety of data needed for images in the style
of ImageNet [16], the type of visual data that robots have
available is much different than ImageNet-style images.
Additionally, higher object specificity may be desired. In
robotics contexts, there has been recent work in trying to
identify data requirements for achieving practical performance
for deep visual models trained on simulation data [3], [4], [5],
[6], and specifically augmenting small datasets of real data
with large datasets of simulation data [3], [4], [5], [6]. We do
not know of prior studies that have performed generalization
experiments with the scale of real data used here.
III. DATA G ENERATION P IPELINE
One of the main contributions of this paper is an efficient
pipeline for generating labeled RGBD training data. The
steps of the pipeline are described in the following sections:
RGBD data collection, dense 3D reconstruction, object mesh
generation, human assisted annotation, and rendering of
labeled images.
A. RGBD Data Collection
A feature of our approach is that the RGBD sensor can
either be mounted on an automated arm, as in Figure (1a),
or the the RGBD sensor can simply be hand-carried. The
benefit of the former option is a reduced human workload,
while the benefit of the latter option is that no sophisticated
equipment (i.e. motion capture, external markers, heavy robot
arm) is required, enabling data collection in a wide variety
(a)
(b)
(d)
(e)
(c)
(f)
Fig. 1: Overview of the data generation pipeline. (a) Xtion RGBD sensor mounted on Kuka IIWA arm for raw data collection.
(b) RGBD data processed by ElasticFusion into reconstructed pointcloud. (c) User annotation tool that allows for easy
alignment using 3 clicks. User clicks are shown as red and blue spheres. The transform mapping the red spheres to the
green spheres is then the user specified guess. (d) Cropped pointcloud coming from user specified pose estimate is shown in
green. The mesh model shown in grey is then finely aligned using ICP on the cropped pointcloud and starting from the user
provided guess. (e) All the aligned meshes shown in reconstructed pointcloud. (f) The aligned meshes are rendered as masks
in the RGB image, producing pixelwise labeled RGBD images for each view.
of environments. We captured 112 scenes using the handheld
approach. For the remaining 26 scenes we mounted the sensor
on a Kuka IIWA, as shown in Figure (1a). The IIWA was
programmed to perform a scanning pattern in both orientation
and azimuth. Note that the arm-automated method does not
require one to know the transform between the robot and the
camera; everything is done in camera frame. Our typical logs
averaged 120 seconds in duration with data captured at 30Hz
by the Asus Xtion Pro.
B. Dense 3D Reconstruction
The next step is to extract a dense 3D reconstruction
of the scene, shown in Figure (1b), from the raw RGBD
data. For this step we used the open source implementation
of ElasticFusion [17] with the default parameter settings,
which runs in realtime on our desktop with an NVIDIA GTX
1080 GPU. ElasticFusion also provides camera pose tracking
relative to the local reconstruction frame, a fact that we take
advantage of when rendering labeled images. Reconstruction
performance can be affected by the amount of geometric
features and RGB texture in the scene. Most natural indoor
scenes provide sufficient texture, but large, flat surfaces with
no RGB or depth texture can occasionally incur failure modes.
Our pipeline is designed in a modular fashion so that any
3D reconstruction method that provides camera pose tracking
can be used in place of ElasticFusion.
C. Object Mesh Generation
A pre-processing step for the pipeline is to obtain meshes
for each object. Once obtained, meshes speed annotation by
enabling alignment of the mesh model rather than manually
intensive pixelwise segmentation of the 3D reconstruction.
Using meshes necessitates rigid objects, but imposes no other
restrictions on the objects themselves. We tested several
different mesh construction techniques when building our
dataset. In total there are twelve objects. Four object meshes
were generated using an Artec Space Spider handheld scanner.
One object was scanned using Next Engine turntable scanner.
For the four objects which are part of the YCB dataset [18]
we used the provided meshes. One of our objects, a tissue
box, was modeled using primitive box geometry. In addition
our pipeline provides a volumetric meshing method using
the VTK implementation of [19] that operates directly on
the data already produced by ElasticFusion. Finally, there
exist several relatively low cost all-in-one solutions [20], [21],
[22] which use RGBD sensors such as the Asus Xtion, Intel
RealSense R300 and Occipital Structure Sensor, to generate
object meshes. The only requirement is that the mesh be
sufficiently high quality to enable the ICP based alignment
(see section III-D). RGB textures of meshes are not necessary.
D. Human Assisted Annotation
One of the key contributions of the paper is in reducing
the amount of human annotation time needed to generate
labeled per-pixel and pose data of objects in clutter. We
evaluated several global registration methods [23], [24], [25]
to try to automatically align our known objects to the 3D
reconstruction but none of them came close to providing
satisfactory results. This is due to a variety of reasons, but a
principle one is that many scene points didn’t belong to any
of the objects.
To circumvent this problem we developed a novel user
interface that utilizes human input to assist traditional
registration techniques. The user interface was developed
using Director [26], a robotics interface and visualization
framework. Typically the objects of interest are on a table
or another flat surface – if so, a single click from the user
segments out the table. The user identifies each object in
the scene and then selects the corresponding mesh from the
mesh library. They then perform a 3-click-based initialization
of the object pose. Our insight for the alignment stage
was that if the user provides a rough initial pose for the
object, then traditional ICP-based techniques can successfully
provide the fine alignment. The human provides the rough
initial alignment by clicking three points on the object in
the reconstructed pointcloud, and then clicking roughly the
same three points in the object mesh, see Figure (1c). The
transform that best aligns the 3 model points, shown in red,
with the three scene points, shown in blue, in a least squares
sense is found using the vtkLandmarkTransform function.
The resulting transform then specifies an initial alignment
of the object mesh to the scene, and a cropped pointcloud
is taken from the points within 1cm of the roughly aligned
model, as shown in green in Figure (1d). Finally, we perform
ICP to align this cropped pointcloud to the model, using the
rough aligment of the model as the initial seed. In practice
this results in very good alignments even for cluttered scenes
such as Figure (1e).
The entire human annotation process takes approximately
30 seconds per object. This is much faster than aligning
the full object meshes by hand without using the 3-click
technique which can take several minutes per object and
results in less accurate object poses. We also compared our
method with human labeling (polygon-drawing) each image,
and found intersection over union (IoU) above 80%, with
approximately four orders of magnitude less human effort
per image (supplementary figures on our website).
# objects
# distinct scenes
# unique object instances aligned
avg duration of single scene
# labeled RGBD frames
# labeled object instances
12
105 single/double object
33 with 6+ objects
339
120 seconds, 3600 RGBD frames
352,000
1,000,000+
TABLE I: Dataset Description
object models to the reconstructed pointcloud, we also have
object poses in each camera frame, for each image frame
in the log. Given object poses in camera frame it is easy to
get the pixelwise labels by projecting the object meshes into
the rendered images. An RGB image with projected object
meshes is displayed in Figure (1f).
F. Discussion
As compared to existing methods such as [8], [27], [1]
our method requires no sophisticated calibration, works for
arbitrary rigid objects in general environments, and requires
only 30 seconds of human annotation time per object per
scene. Since the human annotation is done on the full
3D reconstruction, one labeling effort automatically labels
thousands of RGBD images of the same scene from different
viewpoints.
IV. R ESULTS
We first analyze the effectiveness of the LabelFusion
data generation pipeline (Section IV-A). We then use data
generated from our pipeline to perform practical empirical
experiments to quantify the generalization value of different
aspects of training data (Section IV-B).
A. Evaluation of Data Generation Pipeline
LabelFusion has the capability to rapidly produce large
amounts of labeled data, with minimal human annotation time.
In total we generated over 352,000 labeled RGBD images,
of which over 200,000 were generated in approximately one
day by two people. Because many of our images are multiobject, this amounts to over 1,000,000 labeled object instances.
Detailed statistics are provided in Table I. The pipeline is opensource and intended for use. We were able to create training
data in a wide variety of scenarios; examples are provided
in Figure 2. In particular, we highlight the wide diversity of
environments enabled by hand-carried data collection, the
wide variety of lighting conditions, and the heavy clutter both
of backgrounds and of multi-labeled object scenes.
E. Rendering of Labeled Images and Object Poses
After the human annotation step of Section III-D, the rest of
the pipeline is automated. Given the previous steps it is easy
to generate per-pixel object labels by projecting the 3D object
poses back into the 2D RGB images. Since our reconstruction
method, ElasticFusion, provides camera poses relative to the
local reconstruction frame, and we have already aligned our
Fig. 3: Time required for each step of pipeline.
(a)
(b)
(c)
(d)
(e)
Fig. 2: Examples of labeled data generated by our pipeline: (a) heavily cluttered multi-object, (b) low light conditions, (c)
motion blur, (d) distance from object, (e) 25 different environments. All of these scenes were collected by hand-carrying the
RGBD sensor.
Fig. 4: Example segmentation performance (alpha-blended
with RGB image) of network (e) on a multi-object test scene.
For scaling to large scale data collection, the time required
to generate data is critical. Our pipeline is highly automated
and most components run at approximately real-time, as
shown in Figure 3. The amount of human time required
is approximately 30 seconds per object per scene, which
for a typical single-object scene is less than real-time. Postprocessing runtime is several times greater than real-time,
but is easily parallelizable – in practice, a small cluster of
2-4 modern desktop machines (quad-core Intel i7 and Nvidia
GTX 900 series or higher) can be made to post-process the
data from a single sensor at real-time rates. With a reasonable
amount of resources (one to two people and a handful of
computers), it would be possible to keep up with the real-time
rate of the sensor (generating labeled data at 30 Hz).
B. Empirical Evaluations: How Much Data Is Needed For
Practical Object-Specific Segmentation?
With the capability to rapidly generate a vast sum of labeled
real RGBD data, questions of “how much data is needed?” and
“which types of data are most valuable?” are accessible. We
explore practical generalization performance while varying
three axes of the training data: (i) whether the training set
includes multi-object scenes with occlusions or only singleobject scenes, (ii) the number of background environments,
and (iii) the number of views used per scene. For each, we
train a state-of-the-art ResNet segmentation network [28] with
different subsets of training data, and evaluate each network’s
generalization performance on common test sets. Further
experimental details are provided in our supplementary
material; due to space constraints we can only summarize
results here.
First, we investigate whether there is a benefit of using
training data with heavily occluded and cluttered multi-object
scenes, compared to training with only single-object scenes.
Although they encounter difficulties with heavy occlusions
in multi-object scenes, [2] uses purely single-object scenes
for training. We trained five different networks to enable
comparison of segmentation performance on novel scenes
(different placements of the objects) for a single background
environment. Results of segmentation performance on novel
scenes (measured using the mean IoU, intersection over union,
per object) show an advantage given multi-object occluded
scenes (b) compared to single-object scenes (a) (Figure 5,
right). In particular, the average IoU per object increases
190% given training set (b) instead of (a) in Figure 5, right,
even though (b) has strictly less labeled pixels than (a), due
to occlusions. This implies that the value of the multi-object
training data is more valuable per pixel than the single-object
Fig. 5: Comparisons of training on single-object vs. multi-object scenes and testing on single-object (left) and multi-object
(right) scenes.
training data. When the same amount of scenes for the singleobject scenes are used to train a network with multi-object
scenes (d), the increase in IoU performance averaged across
objects is 369%. Once the network has been trained on 18
multi-object scenes (d), an additional 18 single-object training
scenes have no noticeable effect on multi-object generalization
(e). For generalization performance on single-object scenes
(Figure 5, left), this effect is not observed; single-object
training scenes are sufficient for IoU performance above
60%.
Second, we ask: how does the performance curve grow
as more and more training data is added from different
background environments? To test this, we train different
networks respectively on 1, 2, 5, 10, 25, and 50 scenes
each labeled with a single drill object. The smaller datasets
are subsets of the larger datasets; this directly allows us to
measure the value of providing more data. The test set is
comprised of 11 background environments which none of
the networks have seen. We observe a steady increase in
segmentation performance that is approximately logarithmic
with the number of training scene backgrounds used (Figure
7, left). We also took our multi-object networks trained
on a single background and tested them on the 11 novel
environments with the drill. We observe an advantage of the
multi-object training data with occlusions over the singleobject training data in generalizing to novel background
environments (Figure 7, right).
Third, we investigate whether 30 Hz data is necessary, or
whether significantly less data suffices (Figure 9). We perform
experiments with downsampling the effective sensor rate
both for robot-arm-mounted multi-object single-background
training set (e), and the hand-carried many-environments
dataset with either 10 or 50 scenes. For each, we train four
different networks, where one has all data available and the
others have downsampled data at respectively 0.03, 0.3, and
3 Hz. We observe a monotonic increase in segmentation
performance as the effective sensor rate is increased, but
with heavily diminished returns after 0.3 Hz for the slower
robot-arm-mounted data (∼0.03 m/s camera motion velocity).
The hand-carried data (∼0.05 - 0.17 m/s) shows more gains
with higher rates.
V. C ONCLUSION
This paper introduces LabelFusion, our pipeline for efficiently generating RGBD data annotated with per-pixel labels
and ground truth object poses. Specifically only a few minutes
of human time are required for labeling a scene containing
thousands of RGBD images. LabelFusion is open source and
available for community use, and we also supply an example
dataset generated by our pipeline [29].
The capability to produce a large, labeled dataset enabled us
to answer several questions related to the type and quantity of
training data needed for practical deep learning segmentation
networks in a robotic manipulation context. Specifically we
found that networks trained on multi-object scenes performed
significantly better than those trained on single object scenes,
both on novel multi-object scenes with the same background,
and on single-object scenes with new backgrounds. Increasing
the variety of backgrounds in the training data for singleobject scenes also improved generalization performance for
new backgrounds, with approximately 50 different backgrounds breaking into above-50% IoU on entirely novel
scenes. Our recommendation is to focus on multi-object data
collection in a variety of backgrounds for the most gains in
generalization performance.
We hope that our pipeline lowers the barrier to entry for
using deep learning approaches for perception in support of
robotic manipulation tasks by reducing the amount of human
time needed to generate vast quantities of labeled data for
your specific environment and set of objects. It is also our
hope that our analysis of segmentation network performance
provides guidance on the type and quantity of data that needs
to be collected to achieve desired levels of generalization
performance.
ACKNOWLEDGEMENT
The authors thank Matthew O’Kelly for his guidance
with segmentation networks and manuscript feedback. We
also thank Allison Fastman and Sammy Creasey of Toyota
Research Institute for their help with hardware, including
object scanning and robot arm automation. David Johnson of
Draper Laboratory and Shuran Song of Princeton University
provided valuable input on training. We are grateful we were
(a)
18 singleobject scenes
(b)
3 multiobject scenes
(c)
(a) + (b)
(d)
18 multiobject scenes
(e)
(a) + (d)
RGB
image
Fig. 6: Comparison of segmentation performance on novel multi-object test scenes. Networks are either trained on (a) single
object scenes only, (b,d), multi-object test scenes only, or a mixture (c,e).
Fig. 7: (left) Generalization performance as a function of the number of environments provided at training time, for a set of
six networks trained on 50 different scenes or some subset ({1, 2, 5, 10, 25}) of those scenes. (right) Performance on the
same test set of unknown scenes, but measured for the 5 training configurations for the multi-object, single-environment-only
setup described previously.
# training environments
1
2
5
10
25
50
RGB image
Fig. 8: Comparison of segmentation performance on novel background environments. Networks were trained on {1, 2, 5, 10,
25, 50} background environments.
Robot-arm-mounted data (~0.03 m/s avg. velocity)
Tested on same background
Hand-carried data (~0.05 - 0.17 m/s avg. velocity)
Tested on novel backgrounds
(36 scenes)
Fig. 9: Pixelwise segmentation performance as a function of the number of views per scene, reduced by downsampling the
native 30 Hz sensor to {0.03, 0.3, 3.0.} Hz.
able to use the robot arm testing facility from Toyota Research
Insitute. This work was supported by the Air Force/Lincoln
Laboratory award no. 7000374874, by the Defense Advanced
Research Projects Agency via Air Force Research Laboratory
award FA8750-12-1-0321, and by NSF Contract IIS-1427050.
The views expressed are not endorsed by the sponsors.
R EFERENCES
[1] J. M. Wong, V. Kee, T. Le, S. Wagner, G.-L. Mariottini, A. Schneider,
L. Hamilton, R. Chipalkatty, M. Hebert, D. M. S. Johnson, J. Wu,
B. Zhou, and A. Torralba, “SegICP: Integrated Deep Semantic
Segmentation and Pose Estimation,” ArXiv e-prints, Mar. 2017.
[2] A. Zeng, K.-T. Yu, S. Song, D. Suo, E. Walker Jr, A. Rodriguez, and
J. Xiao, “Multi-view self-supervised deep learning for 6d pose estimation in the amazon picking challenge,” arXiv preprint arXiv:1609.09475,
2016.
[3] S. R. Richter, V. Vineet, S. Roth, and V. Koltun, “Playing for data:
Ground truth from computer games,” in European Conference on
Computer Vision. Springer, 2016, pp. 102–118.
[4] M. Johnson-Roberson, C. Barto, R. Mehta, S. N. Sridhar, K. Rosaen,
and R. Vasudevan, “Driving in the matrix: Can virtual worlds
replace human-generated annotations for real world tasks?” in IEEE
International Conference on Robotics and Automation, 2017, pp. 1–8.
[5] G. Ros, L. Sellart, J. Materzynska, D. Vazquez, and A. Lopez, “The
SYNTHIA Dataset: A large collection of synthetic images for semantic
segmentation of urban scenes,” 2016.
[6] H. Hattori, V. Naresh Boddeti, K. M. Kitani, and T. Kanade, “Learning
scene-specific pedestrian detectors without real data,” in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition,
2015, pp. 3819–3827.
[7] M. Firman, “RGBD Datasets: Past, Present and Future,” in CVPR
Workshop on Large Scale 3D Data: Acquisition, Modelling and Analysis,
2016.
[8] T. Hodan, P. Haluza, S. Obdrzalek, J. Matas, M. Lourakis, and
X. Zabulis, “T-less: An rgb-d dataset for 6d pose estimation of textureless objects,” arXiv preprint arXiv:1701.05498, 2017.
[9] B.-S. Hua, Q.-H. Pham, D. T. Nguyen, M.-K. Tran, L.-F. Yu, and S.-K.
Yeung, “Scenenn: A scene meshes dataset with annotations,” in 3D
Vision (3DV), 2016 Fourth International Conference on. IEEE, 2016,
pp. 92–101.
[10] A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and
M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor
scenes,” arXiv preprint arXiv:1702.04405, 2017.
[11] J. Yu, K. Weng, G. Liang, and G. Xie, “A vision-based robotic
grasping system using deep learning for 3d object recognition and
pose estimation,” in Robotics and Biomimetics (ROBIO), 2013 IEEE
International Conference on. IEEE, 2013, pp. 1175–1180.
[12] “SIXD Challenge.” [Online]. Available: http://cmp.felk.cvut.cz/sixd/
challenge_2017/
[13] S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end training
of deep visuomotor policies,” Journal of Machine Learning Research,
vol. 17, no. 39, pp. 1–40, 2016.
[14] M. Gualtieri, A. ten Pas, K. Saenko, and R. Platt, “High precision
grasp pose detection in dense clutter,” in Intelligent Robots and Systems
(IROS), 2016 IEEE/RSJ International Conference on. IEEE, 2016,
pp. 598–605.
[15] J. Mahler, J. Liang, S. Niyaz, M. Laskey, R. Doan, X. Liu, J. A. Ojea,
and K. Goldberg, “Dex-net 2.0: Deep learning to plan robust grasps
with synthetic point clouds and analytic grasp metrics,” arXiv preprint
arXiv:1703.09312, 2017.
[16] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large
scale visual recognition challenge,” International Journal of Computer
Vision, vol. 115, no. 3, pp. 211–252, 2015.
[17] T. Whelan, S. Leutenegger, R. Salas-Moreno, B. Glocker, and A. Davison, “Elasticfusion: Dense slam without a pose graph.” Robotics:
Science and Systems.
[18] B. Calli, A. Singh, J. Bruce, A. Walsman, K. Konolige, S. Srinivasa,
P. Abbeel, and A. M. Dollar, “Yale-cmu-berkeley dataset for
robotic manipulation research,” The International Journal of Robotics
Research, vol. 36, no. 3, pp. 261–268, 2017. [Online]. Available:
https://doi.org/10.1177/0278364917700714
[19] B. Curless and M. Levoy, “A volumetric method for building complex
models from range images,” in Proceedings of the 23rd annual
conference on Computer graphics and interactive techniques. ACM,
1996, pp. 303–312.
[20] “itSeez3D.” [Online]. Available: https://itseez3d.com
[21] “Sense for RealSense.” [Online]. Available: https://www.3dsystems.
com/shop/realsense/sense
[22] “Skanect.” [Online]. Available: http://skanect.occipital.com/
[23] Q.-Y. Zhou, J. Park, and V. Koltun, “Fast global registration,” in
European Conference on Computer Vision. Springer, 2016, pp. 766–
782.
[24] J. Yang, H. Li, D. Campbell, and Y. Jia, “Go-icp: A globally optimal
solution to 3d icp point-set registration,” IEEE Trans. Pattern Anal.
Mach. Intell., vol. 38, no. 11, pp. 2241–2254, Nov. 2016. [Online].
Available: https://doi.org/10.1109/TPAMI.2015.2513405
[25] N. Mellado, D. Aiger, and N. J. Mitra, “Super 4pcs fast global
pointcloud registration via smart indexing,” Computer Graphics
Forum, vol. 33, no. 5, pp. 205–215, 2014. [Online]. Available:
http://dx.doi.org/10.1111/cgf.12446
[26] P. Marion, “Director: A robotics interface and visualization framework,”
2015. [Online]. Available: http://github.com/RobotLocomotion/director
[27] C. Rennie, R. Shome, K. E. Bekris, and A. F. D. Souza, “A
dataset for improved rgbd-based object detection and pose estimation
for warehouse pick-and-place,” CoRR, vol. abs/1509.01277, 2015.
[Online]. Available: http://arxiv.org/abs/1509.01277
[28] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille,
“Deeplab: Semantic image segmentation with deep convolutional nets,
atrous convolution, and fully connected crfs,” arXiv:1606.00915, 2016.
[29] “LabelFusion,” http://labelfusion.csail.mit.edu.
| 1cs.CV
|
A Non-generative Framework and Convex Relaxations for
Unsupervised Learning
Elad Hazan
Tengyu Ma
arXiv:1610.01132v3 [cs.LG] 27 Dec 2016
December 28, 2016
Abstract
We give a novel formal theoretical framework for unsupervised learning with two distinctive
characteristics. First, it does not assume any generative model and based on a worst-case performance metric. Second, it is comparative, namely performance is measured with respect to a given
hypothesis class. This allows to avoid known computational hardness results and improper algorithms based on convex relaxations. We show how several families of unsupervised learning models,
which were previously only analyzed under probabilistic assumptions and are otherwise provably intractable, can be efficiently learned in our framework by convex optimization.
1 Introduction
Unsupervised learning is the task of learning structure from unlabelled examples. Informally, the main
goal of unsupervised learning is to extract structure from the data in a way that will enable efficient
learning from future labelled examples for potentially numerous independent tasks.
It is useful to recall the Probably Approximately Correct (PAC) learning theory for supervised learning [Val84], based on Vapnik’s statistical learning theory [Vap98]. In PAC learning, the learning can
access labelled examples from an unknown distribution. On the basis of these examples, the learner
constructs a hypothesis that generalizes to unseen data. A concept is said to be learnable with respect to
a hypothesis class if there exists an (efficient) algorithm that outputs a generalizing hypothesis with high
probability after observing polynomially many examples in terms of the input representation.
The great achievements of PAC learning that made it successful are its generality and algorithmic
applicability: PAC learning does not restrict the input domain in any way, and thus allows very general
learning, without generative or distributional assumptions on the world. Another important feature is
the restriction to specific hypothesis classes, without which there are simple impossibility results such
as the “no free lunch” theorem. This allows comparative and improper learning of computationally-hard
concepts.
The latter is a very important point which is often understated. Consider the example of sparse
regression, which is a canonical problem in high dimensional statistics. Fitting the best sparse vector
to linear prediction is an NP-hard problem [Nat95]. However, this does not prohibit improper learning,
since we can use a ℓ1 convex relaxation for the sparse vectors (famously known as LASSO [Tib96]).
Unsupervised learning, on the other hand, while extremely applicative and well-studied, has not
seen such an inclusive theory. The most common approaches, such as restricted Boltzmann machines,
topic models, dictionary learning, principal component analysis and metric clustering, are based almost
entirely on generative assumptions about the world. This is a strong restriction which makes it very hard
to analyze such approaches in scenarios for which the assumptions do not hold. A more discriminative
approach is based on compression, such as the Minimum Description Length criterion. This approach
gives rise to provably intractable problems and doesn’t allow improper learning.
1
Main results. We start by proposing a rigorous framework for unsupervised learning which allows
data-dependent, comparative learning without generative assumptions. It is general enough to encompass previous methods such as PCA, dictionary learning and topic models. Our main contribution are
optimization-based relaxations and efficient algorithms that are shown to improperly probably learn
previous models, specifically:
1. We consider the classes of hypothesis known as dictionary learning. We give a more general
hypothesis class which encompasses and generalizes it according to our definitions. We proceed
to give novel polynomial-time algorithms for learning the broader class. These algorithms are
based on new techniques in unsupervised learning, namely sum-of-squares convex relaxations.
As far as we know, this is the first result for efficient improper learning of dictionaries without
generative assumptions. Moreover, our result handles polynomially over-complete dictionaries,
while previous works [AGMM15, BKS15a] apply to at most constant factor over-completeness.
2. We give efficient algorithms for learning a new hypothesis class which we call spectral autoencoders. We show that this class generalizes, according to our definitions, the class of PCA (principal component analysis) and its kernel extensions.
Structure of this paper. In the following chapter we a non-generative, distribution-dependent definition for unsupervised learning which mirrors that of PAC learning for supervised learning. We then
proceed to an illustrative example and show how Principal Component Analysis can be formally learned
in this setting. The same section also gives a much more general class of hypothesis for unsupervised
learning which we call polynomial spectral decoding, and show how they can be efficient learned in our
framework using convex optimization. Finally, we get to our main contribution: a convex optimization
based methodology for improper learning a wide class of hypothesis, including dictionary learning.
1.1 Previous work
The vast majority of work on unsupervised learning, both theoretical as well as applicative, focuses on
generative models. These include topic models [BNJ03], dictionary learning [DH06], Deep Boltzmann
Machines and deep belief networks [Sal09] and many more. Many times these models entail non-convex
optimization problems that are provably NP-hard to solve in the worst-case.
A recent line of work in theoretical machine learning attempts to give efficient algorithms for these
models with provable guarantees. Such algorithms were given for topic models [AGM12], dictionary
learning [AGM13, AGMM15], mixtures of gaussians and hidden Markov models [HK13, AGH+ 14] and
more. However, these works retain, and at times even enhance, the probabilistic generative assumptions
of the underlying model. Perhaps the most widely used unsupervised learning methods are clustering
algorithms such as k-means, k-medians and principal component analysis (PCA), though these lack
generalization guarantees. An axiomatic approach to clustering was initiated by Kleinberg [Kle03]
and pursued further in [BDA09]. A discriminative generalization-based approach for clustering was
undertaken in [BBV08] within the model of similarity-based clustering.
Another approach from the information theory literature studies with online lossless compression.
The relationship between compression and machine learning goes back to the Minimum Description
Length criterion [Ris78]. More recent work in information theory gives online algorithms that attain
optimal compression, mostly for finite alphabets [ADJ+ 13, OSZ04]. For infinite alphabets, which are the
main object of study for unsupervised learning of signals such as images, there are known impossibility
results [JOS05]. This connection to compression was recently further advanced, mostly in the context
of textual data [PWMH13].
In terms of lossy compression, Rate Distortion Theory (RDT) [Ber71, CT06] is intimately related
to our definitions, as a framework for finding lossy compression with minimal distortion (which would
2
correspond to reconstruction error in our terminology). Our learnability definition can be seen of an
extension of RDT to allow improper learning and generalization error bounds. Another learning framework derived from lossy compression is the information bottleneck criterion [TPB00], and its learning
theoretic extensions [SST08]. The latter framework assumes an additional feedback signal, and thus is
not purely unsupervised.
The downside of the information-theoretic approaches is that worst-case competitive compression
is provably computationally hard under cryptographic assumptions. In contrast, our compression-based
approach is based on learning a restriction to a specific hypothesis class, much like PAC-learning. This
circumvents the impossibility results and allows for improper learning.
2 A formal framework for unsupervised learning
The basis constructs in an unsupervised learning setting are:
1. Instance domain X , such as images, text documents, etc. Target space, or range, Y. We usually
think of X = Rd , Y = Rk with d ≫ k. (Alternatively, Y can be all sparse vectors in a larger
space. )
2. An unknown, arbitrary distribution D on domain X .
3. A hypothesis class of decoding and encoding pairs,
H ⊆ {(h, g) ∈ {X 7→ Y} × {Y 7→ X }},
where h is the encoding hypothesis and g is the decoding hypothesis.
4. A loss function ℓ : H × X 7→ R>0 that measures the reconstruction error,
ℓ((g, h), x) .
For example, a natural choice is the ℓ2 -loss ℓ((g, h), x) = kg(h(x))− xk22 . The rationale here is to
learn structure without significantly compromising supervised learning for arbitrary future tasks.
Near-perfect reconstruction is sufficient as formally proved in Appendix A.1. Without generative
assumptions, it can be seen that near-perfect reconstruction is also necessary.
For convenience of notation, we use f as a shorthand for (h, g) ∈ H, a member of the hypothesis
class H. Denote the generalization ability of an unsupervised learning algorithm with respect to a
distribution D as
loss(f ) = E [ℓ(f, x)].
D
x∼D
We can now define the main object of study: unsupervised learning with respect to a given hypothesis
class. The definition is parameterized by real numbers: the first is the encoding length (measured in bits)
of the hypothesis class. The second is the bias, or additional error compared to the best hypothesis. Both
parameters are necessary to allow improper learning.
Definition 2.1. We say that instance D, X is (k, γ)-C -learnable with respect to hypothesis class H if
exists an algorithm that for every δ, ε > 0, after seeing m(ε, δ) = poly(1/ε, log(1/δ), d) examples,
returns an encoding and decoding pair (h, g) (not necessarily from H) such that:
1. with probability at least 1 − δ, lossD ((h, g)) 6 min(h,g)∈H lossD ((h, g)) + ε + γ.
2. h(x) has an explicit representation with length at most k bits.
3
For convenience we typically encode into real numbers instead of bits. Real encoding can often
(though not in the worst case) be trivially transformed to be binary with a loss of logarithmic factor.
Following PAC learning theory, we can use uniform convergence to bound the generalization error
of the empirical risk minimizer (ERM). Define the empirical loss for a given sample S ∼ D m as
loss(f ) =
S
1 X
·
ℓ(f, x)
m
x∈S
Define the ERM hypothesis for a given sample S ∼ D m as
fˆERM = arg min loss(fˆ) .
fˆ∈H
S
For a hypothesis class H, a loss function ℓ and a set of m samples S ∼ D m , define the empirical
Rademacher complexity of H with respect to ℓ and S as, 1
#
"
1 X
σi ℓ(f, x)
sup
RS,ℓ (H) =
E
σ∼{±1}m f ∈H m
x∈S
Let the Rademacher complexity of H with respect to distribution D and loss ℓ as Rm (H) =
ES∼Dm [RS,ℓ (H)]. When it’s clear from the context, we will omit the subscript ℓ.
We can now state and apply standard generalization error results. The proof of following theorem is
almost identical to [MRT12, Theorem 3.1]. For completeness we provide a proof in Appendix A.
Theorem 2.1. For any δ > 0, with probability 1 − δ, the generalization error of the ERM hypothesis is
bounded by:
s
loss(fˆERM ) 6 min loss(f ) + 6Rm (H) +
D
f ∈H D
4 log 1δ
2m
An immediate corollary of the theorem is that as long as the Rademacher complexity of a hypothesis
class approaches zero as the number of examples goes to infinity, it can be C learned by an inefficient
algorithm that optimizes over the hypothesis class by enumeration and outputs an best hypothesis with
encoding length k and bias γ = 0. Not surprisingly such optimization is often intractable and hences
the main challenge is to design efficient algorithms. As we will see in later sections, we often need to
trade the encoding length and bias slightly for computational efficiency.
Notation. For every vector z ∈ Rd1 ⊗ Rd2 , we can view it as a matrix of dimension d1 × d2 , which is
denoted as M(z). Therefore in this notation, M(u ⊗ v) = uv ⊤ .
Let vmax (·) : (Rd )⊗2 → Rd be the function that compute the top right-singular vector of some
vector in (Rd )⊗2 viewed as a matrix. That is, for z ∈ (Rd )⊗2 , then vmax (z) denotes the top rightsingular vector of M(z). We also overload the notation vmax for generalized eigenvectors of higher
order tensors. For T ∈ (Rd )⊗ℓ , let vmax (T ) = argmaxkxk61 T (x, x, . . . , x) where T (·) denotes the
multi-linear form defined by tensor T .
We use kAkℓp →ℓq to denote the induced
P operator norm of A from ℓp space to ℓq space. For simplicity,
we also define |A|1 = kAkℓ∞ →ℓ1 =
ij |Aij |, |A|∞ = kAkℓ1 →ℓ∞ = maxij |Aij |. We note that
kAkℓ1 →ℓ1 is the max column ℓ1 norm of A, and kAkℓ1 →ℓ2 is the largest column ℓ2 norm of A.
1
Technically, this is the Rademacher complexity of the class of functions ℓ ◦ H. However, since ℓ is usually fixed for certain
problem, we emphasize in the definition more the dependency on H.
4
3 Spectral autoencoders: unsupervised learning of algebraic manifolds
3.1 Algebraic manifolds
The goal of the spectral autoencoder hypothesis class we define henceforth is to learn the representation
of data that lies on a low-dimensional algebraic variety/manifolds. The linear variety, or linear manifold,
defined by the roots of linear equations, is simply a linear subspace. If the data resides in a linear
subspace, or close enough to it, then PCA is effective at learning its succinct representation.
One extension of the linear manifolds is the set of roots of low-degree polynomial equations. Fors
mally, let k, s be integers and let c1 , . . . , cds −k ∈ Rd be a set of vectors in ds dimension, and consider
the algebraic variety
n
o
M = x ∈ Rd : ∀i ∈ [ds − k], hci , x⊗s i = 0 .
Observe that here each constraint hci , x⊗s i is a degree-s polynomial over variables x, and when s = 1
s
the variety M becomes a liner subspace. Let a1 , . . . , ak ∈ Rd be a basis of the subspaces orthogonal
s
to all of c1 , . . . , cds −k , and let A ∈ Rk×d contains ai as rows. Then we have that given x ∈ M, the
encoding
y = Ax⊗s
pins down all the unknown information regarding x. In fact, for any x ∈ M, we have A⊤ Ax⊗s = x⊗s
and therefore x is decodable from y. The argument can also be extended to the situation when the data
point is close to M (according to a metric, as we discuss later). The goal of the rest of the subsections
is to learn the encoding matrix A given data points residing close to M.
3.2 Warm up: PCA and kernel PCA
In this section we illustrate our framework for agnostic unsupervised learning by showing how PCA and
kernel PCA can be efficiently learned within our model. The results of this sub-section are not new, and
given only for illustrative purposes.
The class of hypothesis corresponding to PCA operates on domain X = Rd and range Y = Rk
for some k < d via linear operators. In kernel PCA, the encoding linear operator applies to the s-th
tensor power x⊗s of the data. That is, the encoding and decoding are parameterized by a linear operator
s
A ∈ Rk×d ,
n
o
pca
Hk,s = (hA , gA ) : hA (x) = Ax⊗s , , gA (y) = A† y ,
where A† denotes the pseudo-inverse of A. The natural loss function here is the Euclidean norm,
ℓ((g, h), x) = kx⊗s − g(h(x))k2 = k(I − A† A)x⊗s k2 .
pca
Theorem 3.1. For a fixed constant s > 1, the class Hk,s is efficiently C -learnable with encoding length
k and bias γ = 0.
pca
The proof of the Theorem follows from two simple components: a) finding the ERM among Hk,s can
be efficiently solved by taking SVD of covariance matrix of the (lifted) data points. b) The Rademacher
complexity of the hypothesis class is bounded by O(ds /m) for m examples. Thus by Theorem 2.1 the
minimizer of ERM generalizes. The full proof is deferred to Appendix B.
5
3.3 Spectral Autoencoders
In this section we give a much broader set of hypothesis, encompassing PCA and kernel-PCA, and
show how to learn them efficiently. Throughout this section we assume that the data is normalized to
Euclidean norm 1, and consider the following class of hypothesis which naturally generalizes PCA:
sa as the following set of all hypothesis
Definition 3.1 (Spectral autoencoder). We define the class Hk,s
(g, h),
s
h(x) = Ax⊗s , A ∈ Rk×d
sa
s
Hk = (h, g) :
.
(3.1)
g(y) = vmax (By), B ∈ Rd ×k
pca
We note that this notion is more general than kernel PCA: suppose some (g, h) ∈ Hk,s has reconstruction error ε, namely, A† Ax⊗s is ε-close to x⊗s in Euclidean norm. Then by eigenvector perturbation
theorem, we have that vmax (A† Ax⊗s ) also reconstructs x with O(ε) error, and therefore there exists a
PSCA hypothesis with O(ε) error as well . Vice versa, it’s quite possible that for every A, the reconstruction A† Ax⊗s is far away from x⊗s so that kernel PCA doesn’t apply, but with spectral decoding we
can still reconstruct x from vmax (A† Ax⊗s ) since the top eigenvector of A† Ax⊗s is close x.
Here the key matter that distinguishes us from kernel PCA is in what metric x needs to be close
to the manifold so that it can be reconstructed. Using PCA, the requirement is that x is in Euclidean
distance close to M (which is a subspace), and using kernel PCA x⊗2 needs to be in Euclidean distance
close to the null space of ci ’s. However, Euclidean distances in the original space and lifted space
typically are meaningless for high-dimensional data since any two data points are far away with each
other in Euclidean distance. The advantage of using spectral autoencoders is that in the lifted space the
geometry is measured by spectral norm distance that is much smaller than Euclidean distance (with a
potential gap of d1/2 ). The key here is that though the dimension of lifted space is d2 , the objects of our
interests is the set of rank-1 tensors of the form x⊗2 . Therefore, spectral norm distance is a much more
effective measure of closeness since it exploits the underlying structure of the lifted data points.
We note that spectral autoencoders relate to vanishing component analysis [LLS+ 13]. When the
data is close to an algebraic manifold, spectral autoencoders aim to find the (small number of) essential
non-vanishing components in a noise robust manner.
3.4 Learnability of polynomial spectral decoding
For simplicity we focus on the case when s = 2. Ideally we would like to learn the best encodingdecoding scheme for any data distribution D. Though there are technical difficulties to achieve such a
general result. A natural attempt would be to optimize the loss function f (A, B) = kg(h(x)) − xk2 =
kx − vmax (BAx⊗2 )k2 . Not surprisingly, function f is not a convex function with respect to A, B, and
in fact it could be even non-continuous (if not ill-defined)!
Here we make a further realizability assumption that the data distribution D admits a reasonable
encoding and decoding pair with reasonable reconstruction error.
Definition 3.2. We say a data distribution D is (k, ε)-regularly spectral decodable if there exist A ∈
2
2
Rk×d and B ∈ Rd ×k with kBAkop 6 τ such that for x ∼ D, with probability 1, the encoding
y = Ax⊗2 satisfies that
M(By) = M(BAx⊗2 ) = xx⊤ + E ,
(3.2)
where kEkop 6 ε. Here τ > 1 is treated as a fixed constant globally.
To interpret the definition, we observe that if data distribution D is (k, ε)-regularly spectrally decodable, then by equation (3.2) and Wedin’s theorem (see e.g. [Vu11] ) on the robustness of eigenvector to
perturbation, M(By) has top eigenvector2 that is O(ε)-close to x itself. Therefore, definition 3.2 is a
2
Or right singular vector when M(By) is not symmetric
6
sufficient condition for the spectral decoding algorithm vmax (By) to return x approximately, though it
might be not necessary. Moreover, this condition partially addresses the non-continuity issue of using
objective f (A, B) = kx − vmax (BAx⊗2 )k2 , while f (A, B) remains (highly) non-convex. We resolve
this issue by using a convex surrogate.
Our main result concerning the learnability of the aforementioned hypothesis class is:
sa is C - learnable with encoding length O(τ 4 k 4 /δ 4 ) and bias δ
Theorem 3.2. The hypothesis class Hk,2
with respect to (k, ε)-regular distributions in polynomial time.
Our approach towards finding an encoding and decoding matrice A, B is to optimize the objective,
h
i
minimize f (R) = E Rx⊗2 − x⊗2 op
(3.3)
s.t. kRkS1 6 τ k
where k · kS1 denotes the Schatten 1-norm3 .
Suppose D is (k, ε)-regularly decodable, and suppose hA and gB are the corresponding encoding
and decoding function. Then we see that R = AB will satisfies that R has rank at most k and f (R) 6 ε.
On the other hand, suppose one obtains some R of rank k′ such that f (R) 6 δ, then we can produce hA
2
′
′
2
and gB with O(δ) reconstruction simply by choosing A ∈ Rk ×d B and B ∈ Rd ×k such that R = AB.
We use (non-smooth) Frank-Wolfe to solve objective (3.3), which in particular returns a low-rank
solution. We defer the proof of Theorem 3.2 to the Appendix B.1.
With a slightly stronger assumptions on the data distribution D, we can reduce the length of the code
to O(k2 /ε2 ) from O(k4 /ε4 ). See details in Appendix C.
4 A family of optimization encodings and efficient dictionary learning
In this section we give efficient algorithms for learning a family of unsupervised learning algorithms
commonly known as ”dictionary learning”. In contrast to previous approaches, we do not construct an
actual ”dictionary”, but rather improperly learn a comparable encoding via convex relaxations.
We consider a different family of codes which is motivated by matrix-based unsupervised learning
models such as topic-models, dictionary learning and PCA. This family is described by a matrix A ∈
Rd×r which has low complexity according to a certain norm k · kα , that is, kAkα 6 cα . We can
parametrize a family of hypothesis H according to these matrices, and define an encoding-decoding pair
according to
1
hA (x) = arg min |x − Ay|1 , gA (y) = Ay
d
kykβ 6k
We choose ℓ1 norm to measure the error mostly for convenience, though it can be quite flexible. The
different norms k·kα , k·kβ over A and y give rise to different learning models that have been considered
before. For example, if these are Euclidean norms, then we get PCA. If k · kα is the max column ℓ2 or
ℓ∞ norm and k · kb is the ℓ0 norm, then this corresponds to dictionary learning (more details in the next
section).
The optimal hypothesis in terms of reconstruction error is given by
1
1
⋆
|x − gA (hA (x))|1 = arg min E
|x − Ay|1 .
min
A = arg min E
r
kAkα 6cα x∼D y∈R :kykβ 6k d
kAkα 6cα x∼D d
The loss function can be generalized to other norms, e.g., squared ℓ2 loss, without any essential
change in the analysis. Notice that this optimization objective derived from reconstruction error is
3
Also known as nuclear norm or trace norm
7
identically the one used in the literature of dictionary learning. This can be seen as another justification
for the definition of unsupervised learning as minimizing reconstruction error subject to compression
constraints.
The optimization problem above is notoriously hard computationally, and significant algorithmic
and heuristic literature attempted to give efficient algorithms under various distributional assumptions(see [AGM13, AGMM15, AEB05] and the references therein). Our approach below circumvents
this computational hardness by convex relaxations that result in learning a different creature, albeit with
comparable compression and reconstruction objective.
4.1 Improper dictionary learning: overview
We assume the max column ℓ∞ norm of A is at most 1 and the ℓ1 norm of y is assumed to be at
most k. This is a more general setting than the random dictionaries (up to a re-scaling) that previous
works
[AGM13, AGMM15] studied. 4 In this case, the magnitude of each entry of x is on the order of
√
k if y has k random ±1 entries. We think of our target error per entry as much smaller than 15 . We
k that are parametrized by the dictionary matrix A = Rd×r ,
consider Hdict
o
n
Hkdict = (hA , gA ) : A ∈ Rd×r , kAkℓ1 →ℓ∞ 6 1 ,
where hA (x) = arg min |x − Ay|1 , gA (y) = Ay
kyk1 6k
Here we allow r to be larger than d, the case that is often called over-complete dictionary. The choice
of the loss can be replaced by ℓ2 loss (or other Lipschitz loss) without any additional efforts, though for
simplicity we stick to ℓ1 loss. Define A⋆ to be the the best dictionary under the model and ε⋆ to be the
optimal error,
A⋆ = arg minkAkℓ →ℓ∞ 61 Ex∼D miny∈Rr :kyk1 6k |x − Ay|1
(4.1)
1
ε⋆ = Ex∼D 1d · |x − gA⋆ (hA⋆ (x))|1 .
Algorithm 1 group encoding/decoding for improper dictionary learning
Inputs: N data points X ∈ Rd×N ∼ D N . Convex set Q. Sampling probability ρ.
1. Group encoding: Compute
Z = arg min |X − C|1 ,
(4.2)
C∈Q
and let
Y = h(X) = PΩ (Z) ,
where PΩ (B) is a random sampling of B where each entry is picked with probability ρ.
2. Group decoding: Compute
g(Y ) = arg min |PΩ (C) − Y |1 .
(4.3)
C∈Q
Theorem 4.1. For any δ > 0, p > 1, the hypothesis class Hkdict is C -learnable (by Algorithm 2) with
2
encoding length Õ(k2 r 1/p /δ2 ), bias δ + O(ε⋆ ) and sample complexity dO(p) in time nO(p )
√
The assumption can be relaxed to that A has ℓ∞ norm at most k and ℓ2 -norm√
at most d straightforwardly.
5
We are conservative in the scaling of the error here. Error much smaller than k is already meaningful.
4
8
We note that here r can be potentially much larger than d since by choosing
√ a large constant p the
overhead caused by r can be negligible. Since the average size of the entries is k, therefore we can get
the bias δ smaller than average size of the entries with code length roughly ≈ k.
The proof of Theorem 4.1 is deferred to Section 5.6. To demonstrate the key intuition and technique behind it, in the rest of the section we consider a simpler algorithm that achieves a weaker goal:
Algorithm 1 encodes multiple examples into some codes with the matching average encoding length
Õ(k2 r 1/p /δ2 ), and these examples can be decoded from the codes together with reconstruction error
ε⋆ + δ. Next, we outline the analysis of Algorithm 1, and we will show later that one can reduce the
problem of encoding a single examples to the problem of encoding multiple examples together.
Here we overload the notation gA⋆ (hA⋆ (·)) so that gA⋆ (hA⋆ (X)) denotes the collection of all the
gA⋆ (hA⋆ (xj )) where xj is the j-th column of X. Algorithm 1 assumes that there exists a convex set
Q ⊂ Rd×N such that
o
n
(4.4)
gA⋆ (hA⋆ (X)) : X ∈ Rd×N ⊂ {AY : kAkℓ1 →ℓ∞ 6 1, kY kℓ1 →ℓ1 6 k} ⊂ Q .
That is, Q is a convex relaxation of the group of reconstructions allowed in the class Hdict . Algorithm 1 first uses convex programming to denoise the data X into a clean version Z, which belongs to
the set Q. If the set Q has low complexity, then simple random sampling of Z ∈ Q serves as a good
encoding.
The following Lemma shows that if Q has low complexity in terms of sampling Rademacher width,
then Algorithm 1 will give a good group encoding and decoding scheme.
Lemma 4.2. Suppose convex Q ⊂ Rd×N satisfies condition (4.4). Then, Algorithm 1 gives a group
encoding andpdecoding pair such that
p with probability 1 − δ, the average reconstruction error is bounded
by ε⋆ + O( SRW m (Q) + O( log(1/δ)/m) where m = ρN d and SRW m (·) is the sampling
Rademacher width (defined in subsection 5.2), and the average encoding length is Õ(ρd).
The proofs here are technically standard: Lemma 4.2 simply follows from Lemma 5.1 and
Lemma 5.2 in Section 5. Lemma 5.1 shows that the difference between Z and X is comparable to
ε⋆ , which is a direct consequence of the optimization over a large set Q that contains optimal reconstruction. Lemma 5.2 shows that the sampling procedure doesn’t lose too much information given a
denoised version of the data is already observed, and therefore, one can reconstruct Z from Y .
The novelty here is to use these two steps together to denoise and achieve a short encoding. The
typical bottleneck of applying convex relaxation on matrix factorization based problem (or any other
problem) is the difficulty of rounding. Here instead of pursuing a rounding algorithm that output the
factor A and Y , we look for a convex relaxation that preserves the intrinsic complexity of the set which
enables the trivial sampling encoding. It turns out that controlling the width/complexity of the convex
relaxation boils down to proving concentration inequalities with sum-of-squares (SoS) proofs, which is
conceptually easier than rounding.
Therefore, the remaining challenge is to design convex set Q that simultaneously has the following
properties
(a) is a convex relaxation in the sense of satisfying condition (4.4)
(b) admits an efficient optimization algorithm
(c) has low complexity (that is, sampling rademacher width Õ(N poly(k)))
Most of the proofs need to be deferred to Section 5. We give a brief overview: In subsection 5.3 we will
design a convex set Q which satisfies condition (4.4) but not efficiently solvable, and in subsection 5.4
we verify that the sampling Rademacher width is O(N k log d). In subsection 5.5, we prove that a sumof-squares relaxation would give a set Qsos
p which satisfies (a), (b) and (c) approximately. Concretely,
we have the following theorem.
9
Theorem 4.3. For every p > 4, let N = dc0 p with a sufficiently large absolute constant c0 . Then, there
d×N (which is defined in subsection 5.5.2) such that (a) it satisfies condiexists a convex set Qsos
p ⊂ R
tion 4.4; (b) The optimization (4.2) and (4.3) are solvable by semidefinite
programming with run-time
p
2
2 2/p N/m).
is
bounded
by
SRW
nO(p ) ; (c) the sampling Rademacher width of Qsos
m (Q) 6 Õ(k r
p
We note that these three properties (with Lemma 4.2) imply that Algorithm 1 with Q = Qsos
p
and ρ = O(k2 r 2/p d−1 /δ2 · log d) gives a group encoding-decoding pair with average encoding length
O(k2 r 2/p /δ2 · log d) and bias δ.
Proof Overview of Theorem 4.3: At a very high level, the proof exploits the duality between sumof-squares relaxation and sum-of-squares proof system. Suppose w1 , . . . , wd are variables, then in SoS
relaxation an auxiliary variable WS is introduced for every subset S ⊂ [d] of size at most s, and valid
linear constraints and P
psd constraint for WS ’s are enforced. By convex duality, intuitively we P
have that if
2
a polynomial q(x) = |S|6s αS xS can be written
as
a
sum
of
squares
of
polynomial
q(x)
=
j rj (x) ,
P
then the corresponding linear form
P over XS , |S|6s αS XS is also nonnegative. Therefore, to certify
certain property of a linear form |S|6s αS XS over XS , it is sufficient (and also necessary by duality)
that the corresponding polynomial admit a sum-of-squares proof.
Here using the idea above, we first prove the Rademacher width of the convex hull of reconstructions Q0 = conv {Z = AY : kAkℓ1 →ℓ∞ 6 1, kY kℓ1 →ℓ1 6 k} using a SoS proof. Then the same proof
automatically applies to for the Rademacher width of the convex relaxation (which is essentially a set of
statements about linear combinations of the auxiliary variables). We lose a factor of r 2/p because SoS
proof is not strong enough for us to establish the optimal Rademacher width of Q0 .
5 Analysis of Improper Dictionary Learning
In this section we give the full proof of the Theorems and Lemmas in Section 4. We start by stating
general results on denoising, Rademacher width and factorizable norms, and proceed to give specialized
bounds for our setting in section 5.4.
5.1 Guarantees of denoising
In this subsection, we give the guarantees of the error caused by the denoising step. Recall that ε⋆ is the
optimal reconstruction error achievable by the optimal (proper) dictionary (equation (4.1)).
Lemma 5.1. Let Z be defined in equation (4.2). Then we have that
1
⋆
E [|Z − X|1 ] 6 ε
N d X∼DN
(5.1)
Proof. Let Y ⋆ = A⋆ hA⋆ (X) where hA⋆ (X) denote the collection of encoding of X using hA⋆ . Since
Y ⋆ ∈ {AY : kAk|ℓ1 →ℓ∞ 6 1, kY kℓ1 →ℓ1 6 k} ⊂ Q, we have that Y ⋆ is a feasible solution of optimization (4.2). Therefore, we have that N1d E [|Z − X|1 ] 6 N1d E [|X − Y ⋆ |1 ] = ε⋆ , where the equality is by
the definition of ε⋆ .
5.2 Sampling Rademacher Width of a Set
As long as the intrinsic complexity of the set Q is small then we can compress by random sampling.
The idea of viewing reconstruction error the test error of a supervised learning problem started with the
work of Srebro and Shraibman [SS05], and has been used for other completion problems, e.g., [BM15].
We use the terminology “Rademacher width” instead of “Rademacher complexity” to emphasize that
the notation defined below is a property of a set of vectors (instead of that of a hypothesis class).
10
For any set W ⊂ RD , and an integer m, we define its sampling Rademacher width (SRW) as,
1
sup hx, σiΩ ,
SRW m (W ) = E
(5.2)
σ,Ω m x∈W
where Ω is random subset of [D] of size m, ha, biΩ is defined as
P
i∈Ω ai bi
and σ ∼ {±1}D .
Lemma 5.2. ([BM15, Theorem 2.4]) With probability at least 1 − δ over the choice of Ω, for any z̃ ∈ RD
r
1
log(1/δ)
1
|z̃ − z|1 6 |PΩ (z̃) − PΩ (z)|1 + 2SRW m (W ) + M
.
D
m
m
where M = supz̃∈W,i∈[D] |zi − z̃i |.
5.3 Factorable norms
In this subsection, we define in general the factorable norms, from which we obtain a convex set Q
which satisfies condition (4.4) (see Lemma 5.3).
For any two norms k · kα , k · kβ that are defined on matrices of any dimension, we can define the
following quantity
Γα,β (Z) = inf kAkα kBkβ
Z=AB
(5.3)
For any p, q, s, t > 1, we use Γp,q,s,t(·) to denote the function Γℓp →ℓq ,ℓs →ℓt (·). When p = t,
Γp,q,s,p(Z) is the factorable norm [TJ89, Chapter 13] of matrix Z. In the case when p = t = 2,
q = ∞, s = 1, we have that Γ2,∞,1,2 (·) is the γ2 -norm [LMSS07] or max norm [SS05], which has been
used for matrix completion.
The following Lemma is the main purpose of this section which shows a construction of a convex
set Q that satisfies condition (4.4).
Lemma 5.3. For any q, t > 1 √
we have that Γ1,q,1,t (·) is a norm. As a consequence, letting Q1,∞,1,1 =
{C ∈ RN ×d : Γ1,∞,1,1 (C) 6 dk}, we have that Q1,∞,1,1 is a convex set and it satisfies condition (4.4).
Towards proving Lemma 5.3, we prove a stronger result that if p = s = 1, then Γp,q,s,t is also a
norm. This result is parallel to [TJ89, Chapter 13] where the case of p = t is considered.
Theorem 5.4. Suppose that k · kα and k · kβ are norms defined on matrices of any dimension such that
1. k[A, B]kα 6 max{kAkα , kBkα }
2. k · kβ is invariant with respect to appending a row with zero entries.
Then, Γα,β (·) is a norm.
Proof. Non-negativity: We have that Γα,β (Z) > 0 by definition as a product of two norms. Further,
Γα,β (Z) = 0 if and only if kAkα = 0 or kBkβ = 0, which is true if and only if A or B are zero, which
means that Z is zero.
Absolute scalability: For any positive t, we have that if Z = AB and Γα,β (Z) 6 kAkα kBkβ , then
tZ = (tA) · B and Γα,β (tZ) 6 tkAkα kBkβ . Therefore by definition of Γα,β (Z), we get Γα,β (tZ) 6
tΓα,β (Z).
If we replace Z by tZ and t by 1/t we obtain the other direction, namely Γα,β (Z) 6 1/t · Γα,β (tZ).
Therefore, Γα,β (tZ) = tΓα,β (Z) for any t > 0.
Triangle inequality: We next show that Γα,β (Z) satisfies triangle inequality, from which the result
follows. Let W and Z be two matrices of the same dimension. Suppose A, C satisfy that Z = AC and
11
Γα,β (Z) = kAkα kCkβ . Similarly,
−1 suppose W = BD, and Γα,β (W ) = kBkα kDkβ . Therefore, we
t C
have that W + Z = [tA, B]
, and that for any t > 0,
D
−1
t C
Γα,β (W + Z) 6 k[tA, B]kα
(by defintion of Γα,β )
D
β
!
C
0
(by triangle inquality)
+ t−1
6 k[tA, B]kα
0 β
D β
(by assumptions on k · kα and k · kβ )
6 max {t kAkα , kBkα } t−1 kCkβ + kDkβ
Pick t =
kBkα
kAkα ,
we obtain that,
Γα,β (W + Z) 6 kAkα kCkβ + kBkα kDkβ
= Γα,β (Z) + Γα,β (W ) .
Note that if k · kα is a ℓ1 → ℓq norm, then it’s also the max column-wise ℓq norm, and therefore it
satisfies the condition a) in Theorem 5.4. Moreover, for similar reason, k · kβ = k · kℓ1 →ℓt satisfies the
condition b) in Theorem 5.4. Hence, Lemma 5.3 is a direct consequence of Theorem 5.4. Lemma 5.3
gives a convex set Q that can be potentially used in Algorithm 1.
5.4 Sampling Rademacher width of level set of Γ1,∞,1,1
Here we give a Rademacher width bound for the specific set we’re interested in, namely the level sets of
Γ1,∞,1,1 , formally,
Q1,∞,1,1 = {C ∈ RN ×d : Γ1,∞,1,1 (C) 6 k}.
By the definition Q1,∞,1,1 satisfies condition (4.4). See section 5.2 for definition of Ramemacher
width.
Lemma 5.5. It holds that
e
SRW m (Q1,∞,1,1 ) 6 O
r
k2 N
m
!
.
Proof of Lemma 5.5. Recall the definition of the sample set Ω of coordinates from C, and their multiplication by i.i.d Rademacher variables in section 5.2. Reusing notation, let ξ = σ ⊙ Ω and we use Q as
a shorthand for Q1,∞,1,1 . Here ⊙ means the entry-wise Hadamard product (namely, each coordinate in
Ω is multiplied by an independent Rademacher random variable). We have that
#
"
1
sup hC, ξi
SRW m (Q) = E
ξ m C∈Q
"
#
1
=E
hAB, ξi
(by defintiion of Q)
sup
ξ m kAkℓ →ℓ∞ 61,kBkℓ →ℓ 6k
1
1
1
"
#
1
=E
hB, A⊤ ξi
sup
ξ m kAkℓ →ℓ∞ 61,kBkℓ →ℓ 6k
1
1
1
#
"
N
P
X
1
N
⊤
(By hU, V i 6
kA ξi k∞ .
kU
k
k
sup
6E
i ∞ kV kℓ1 →ℓ1 )
i=1
ξ m kAkℓ →ℓ∞ 61
1
i=1
12
m
be the probability of any entry belongs to Ω. Let ξi denote the i-th column of ξ, and Aj
Let ρ = dN
denotes the j-th column of A. Therefore, each entry of ξi has probability ρ/2 to be +1 and -1, and has
probability 1 − ρ of being 0. By concentration inequality we have that for ρ > logd r , and any fixed A
with kAkℓ1 →ℓ∞ = max kAj k∞ 6 1,
p
⊤
(5.4)
E[kA ξi k∞ ] 6 O( ρd log r log d) .
ξi
Moreover, we have that
p
Varξi [kA⊤ ξi k∞ ] 6 O( ρd log r log d) .
(5.5)
Moreover,
kA⊤ ξi k∞ has an sub-exponential P
tail. (Technically, its ψ1 -Orlicz norm is bounded by
√
⊤
O( ρd log r log d)). Note that the variance of N
i=1 kA ξi k∞ will decrease as N increases, and therefore for large enough N = (drρ)Ω(1) , we will have that with probability 1 − exp(−Ω(dr)Ω(1) ),
N
X
i=1
kA⊤ ξi k∞ 6 O(N
p
ρd log r log d)
Therefore, using the standard ε-net covering argument, we obtain that with high probability,
N
X
sup
√
kAkℓ1 →ℓ2 6 d i=1
kA⊤ ξi k∞ 6 O(N
p
ρd log r log d) .
Hence, altogether we have
!
r
2N
k
1
⊤
e
kkA ξkℓ1 →ℓ∞ 6 O
.
SRW m (Q) 6 E
sup
m kAkℓ →ℓ 6√d
m
ξ
1
2
5.5 Convex Relaxation for Γ1,∞,1,1 norm
5.5.1
Sum-of-squares relaxation
Here we will only briefly introduce the basic ideas of Sum-of-Squares (Lasserre) relaxation [Par00,
Las01] that will be used for this paper. We refer readers to the extensive study [Las15, Lau09, BS14] for
detailed discussions of sum of squares proofs and their applications to algorithm design. Recently, there
has been a popular line of research on applications of sum-of-squares algorithms to machine learning
problems [BKS15b, BKS14, BM16, MW15, GM15, HSS15, HSSS16, MSS16]. Here our technique in
the next subsection is most related to that of [BM16], with the main difference that we deal with ℓ1 norm
constraints that are not typically within the SoS framework.
Let R[x]d denote the set of all real polynomials of degree at most d with n variables x1 , . . . , xn . We
start by defining the notion of pseudo-expectation. The intuition is that the pseudo-expectation behave
like the actual expectation of a real probability distribution on squares of polynomials.
Definition 5.1 (pseudo-expectation). A degree-d pseudo-expectation Ẽ is a linear operator that maps
R[x]d to R and satisfies Ẽ(1) = 1 and Ẽ(p2 (x)) > 0 for all real polynomials p(x) of degree at most d/2.
Definition 5.2. Given a set of polynomial equations A = {q1 (x) = 0, . . . , qn (x) = 0}, we say degreed pseudo-expectation Ẽ satisfies constraints A if Ẽ [qi (x)r(x)] = 0 for every i and r(x) such that
deg(qi r) 6 d.
13
One can optimize over the set of pseudo-expectations that satisfy A in nO(d) time by the following
semidefinite program:
Variables
Subject to
Ẽ[xS ]
∀S : |S| 6 d
K
Ẽ qi (x)x = 0
∀i, K : |K| + deg(qi ) 6 d
i
h
Ẽ x⊗d/2 (x⊗d/2 )⊤ 0
Definition 5.3 (SoS proof of degree d). For a set of constraints A = {q1 (x) = 0, . . . , qn (x) = 0}, and
an integer d, we write
A ⊢d p(x) > q(x)
if there exists polynomials hi (x) for i = 0, 1, . . . , ℓ and gj (x) for j = 1, . . . , t such that deg(hi ) 6 d/2
and deg(gj rj ) 6 d that satisfy
p(x) − q(x) =
ℓ
X
2
hi (x) +
t
X
rj (x)gj (x) .
j=1
i=1
We will drop the subscript d when it is clear form the context.
The following fact follows from the definitions above but will be useful throughout the proof.
Proposition 5.6. Suppose A ⊢d p(x) > q(x). Then for any degree-d pseudo-expectation Ẽ that satisfies
A, we have Ẽ[p(x)] > Ẽ[q(x)].
5.5.2
Relaxation for Γ1,∞,1,1 norm
In this section, we introduce convex relaxations for the Γ1,∞,1,1 norm. For convenience, let p be a power
of 2. Let A and B be formal variables of dimension d × r and r × N in this section. We introduce
more formal variables for the relaxation. Let b be formal variables of dimension r × N . We consider the
following set of polynomial constraints over formal variables A, B, b:
)
(
r
X
p
p−1
p/(p−1)
2
bℓj 6 k
, ∀i, k, Aik 6 1 .
A = ∀i, j, Bij = bij ,
ℓ=1
For any real matrix C ∈ Rd×N , we define
A(C) = {C = AB} ∪ A .
We define our convex relaxation for Q1,∞,1,1 as follows,
d×N
Qsos
: ∃degree-O(p2 ) pseudo-expectation Ẽ that satisfies A(C)}
p = {C ∈ R
(5.6)
Lemma 5.7. For any p > 4, we have
and therefore Qsos
p satisfies condition (4.4).
Q1,∞,1,1 ⊂ Qsos
p
Proof. Suppose C ∈ Q1,∞,1,1 . Then by definition of the Γ1,∞,1,1 -norm (equation (5.3)), we have
that there exists matrices U, V such that C = U V and Uij2 6 1 and kV kℓ1 →ℓ1 6 k. Now construct
p/(p−1)
P
P
1/(p−1)
p−1
r
6 kp/(p−1) . Therefore, Then we have
vij = Vij
. We have rℓ=1 bpij 6
v
ℓ=1 ij
that A = U, B = V, b = v satisfies the constraint (5.6). Then the trivial pseudo-expectation operator
Ẽ[p(A, B, b)] = p(U, V, v) satisfies A(C) and therefore C ∈ Qsos
p by definition.
14
Theorem 5.8. Suppose N = dc0 p for large enough absolute constant c0 . Then the sampling Rademacher
complexity of Qsos
p is bounded by,
SRW m (Qsos
p )
6O
r
p2 N k 2 r 2/p log d
m
!
.
The proof of Theorem 5.8 finally boils down to prove certain empirical process statement with SoS
proofs. We start with the following two lemmas.
Lemma 5.9. Suppose N = dc0 p for larger enough constant c0 , and let ξ = σ ⊙ Ω where σ and Ω are
defined in Section 5.2. Then, we have
A ⊢ hAB, ξip 6 N p−1 kp
N
X
i=1
kA⊤ ξi kpp
Proof. Let ξi be the i-th column of ξ. We have that
!p
N
X
⊤
p
⊤
p
hA ξ, Bi i
A ⊢hAB, ξi = hA ξ, Bi =
i=1
N
X
6 N p−1
= N p−1
6 N p−1
i=1
N
X
(since ⊢
1
N
P
i∈N
αi
p
hA⊤ ξi , b⊙p−1
ip
i
i=1
N
X
i=1
6N
hA⊤ ξi , Bi ip
!
p−1 p
k
i=1
1
N
P
p
i∈[N ] αi )
(by Bij = bp−1
ij )
kA⊤ ξi kpp kbi kp(p−1)p (by ⊢
N
X
6
P
p−1
j a i bj
kA⊤ ξi kpp
p
P p−1
P
6 ( i api ) ( bpi ) ; see Lemma D.1)
(by constraint
Pr
p
ℓ=1 bℓj
6 kp/(p−1) )
Lemma 5.10. In the setting of Lemma 5.9, let ρ = m/(N d). Let x = (x1 , . . . , xd ) be indeterminates
and let B = {x21 6 1, . . . , x2d 6 1}. Suppose ρ > 1/d. With probability at least 1 − exp(−Ω(d)) over
the choice of ξ, there exists a SoS proof,
B ⊢ kξ ⊤ xkpp 6 N · O(ρdp2 )p/2 .
(5.7)
As an immediate consequence, let ξi be the i-th column of ξ, then, with probability at least 1 −
exp(−Ω(d)) over the choice of ξ there exists SoS proof,
A⊢
N
X
i=1
kA⊤ ξi kpp 6 N r · O(ρdp2 )p/2 .
(5.8)
Proof. Recall that p is an even number. We have that
B ⊢ kξ ⊤ xkpp = x⊗p/2
⊤
N
X
i=1
15
⊗p/2 ⊤
⊗p/2
ξi
ξi
!
x⊗p/2
PN ⊗p/2 ⊗p/2 ⊤
p/2
p/2
∈ Rd ×d .
ξi
Let T =
i=1 ξi
h
i
i
h
⊤
PN
p
E x⊗p/2 T x⊗p/2 = E
i=1 hξi , xi . It follows that
Then, by definition we have that
#
"N
⊤
X
hξi , xip
B ⊢ kξ ⊤ xkpp = x⊗p/2 (T − E[T ])x⊗p/2 + E
6 kxkp kT − E[T ]k + E
"
i=1
N
X
i=1
hξi , xip
#
(by ⊢ y ⊤ By 6 kyk2 kBk)
#
N
X
hξi , xip .
6 dp/2 kT − E[T ]k + E
"
i=1
Let ζ have the same distribution as ξi . Then we have that
X
p
t α xα .
E [hζ, xi ] =
α
where tα = 0 if xα contains an odd degree. Moreover, let |α| be the number xi with non-zero exponents
that xα has. Then we have tα 6 ρ|α| pp . Therefore we have that for ρ > 1/d,
B ⊢ E [hζ, xip ] 6 dp/2 ρp/2 pp .
It follows that
B⊢E
"
N
X
i=1
p
hξi , xi
#
6 N · O(ρdp2 )p/2 .
Next, we use Bernstein inequality to bound kT − E[T ]k. We control the variance by
#
"N
⊤
X
⊗p/2
⊗p/2
6 N d2p .
ξi
kξkp ξi
σ2 , E
i=1
⊗p/2
Moreover, each summand in the definition of T can be bounded by ξi
⊗p/2 ⊤
ξi
6 dp . Note that
these bounds are not necessarily tight since they already suffice for our purpose.
Therefore, we obtain
p
that with high probability over the chance of ξ, it holds that kT − E[T ]k 6 N dO(p) log N . Therefore
for N > dΩ(p) , we have that with high probability over the choice of ξ,
B ⊢ kξ ⊤ xkpp
6 dp/2 kT − E[T ]k + E
"
#
(5.9)
N
X
hξi , xip 6 N · O(ρdp2 )p/2 .
i=1
Finally
equation
implies equation (5.8). Let Ai be the i-th column of A. Note that we
Pwe show
P(5.7)
r
⊤ ξ kp =
⊤ A kp , since both left-hand side and right-hand side are equal to the
have N
kA
kξ
i p
i p
i=1
i=1
p-th power of all the entries of the matrix A⊤ ξ. Since A ⊢ {∀j, A2ij 6 1}, we can invoke the first part
of the Lemma and use equation (5.7) to complete the proof.
Combining Lemma 5.9 and Lemma 5.10 we obtain the following corollary,
Corollary 5.11. In the setting of Lemma 5.9, with probability 1 − exp(−Ω(d)), we have
A ⊢ hAB, ξip 6 N p kp r · O(ρd)p/2 .
16
(5.10)
Now we are ready to prove Theorem 5.8. Essentially the only thing that we need to do is take
pseudo-expectation on the both sides of equation (5.10).
Proof of Theorem 5.8. Recall that ξ = σ ⊙ Ω. We have that
#p
"
1
sos p
sup hC, ξi
SRW m (Qp ) = E
p
ξ m C∈Qsos
d
#
"
1
(by Jensen’s inequality)
6E
sup hC, ξip
p
ξ m C∈Qsos
d
"
#
1
=E
Ẽ [hAB, ξi]
(by definition of Qsos
sup
p )
p
ξ m Ẽ that satisfies A(C)
"
#
1
6E
Ẽ [hAB, ξi] | equation (5.8) holds
sup
mp Ẽ that satisfies A(C)
+ P [equation (5.8) doesn’t hold] · dO(1)
1
6 p N p kp r · O(ρdp2 )p/2 + exp(−Ω(d))dO(1)
m
(by Corollary 5.11 and Proposition 5.6)
1 p p
6 p N k r · O(ρdp2 )p/2 .
m
Therefore using the fact that ρ = m/(N d), taking p-th root on the equation above, we obtain,
!
r
p2 N r 2/p k2 log d
sos
.
SRW m (Qp ) 6 O
m
5.6 Proof of Theorem 4.1
In this subsection we prove Theorem 4.1 using the machinery developed in previous subsections. Essentially the idea here is to reduce the problem of encoding one example to the problem of encoding a
group of examples. To encode an example x, we draw N − 1 fresh examples x1 , . . . , xN −1 , and then
call the group encoding Algorithm 1.
We also note that the encoding procedure can be simplified by removing the denoising step, and this
still allows efficient decoding steps. Since in this case the encoding contains more noise from the data,
we prefer the current scheme.
Proof of Theorem 4.1. We will prove that Algorithm 2 with Q = Qpsos and ρ = O(k2 r 2/p d−1 /δ2 ·
log d) gives the desired encoding/decoding. Indeed with high probability, the encoding length is
Õ(ρd) = Õ(k2 r 2/p /δ2 ). Next we will show that we obtain good reconstruction in the sense that
1
1
⋆
d E [|x − g(h(x))|1 ] = d E [|x − z̃|1 ] 6 O(ε + δ). Here and throughout the proof, the expectation is over the randomness of x and the randomness of the algorithm (that is, the randomness of
x, x1 , . . . , xN −1 and the random sampling). First of all, by symmetry, we have that
E [|x − z̃1 |1 ] =
1
E [|[x1 , . . . , xN −1 , x] − [z̃1 , . . . , z̃N −1 , z̃]|1 ] .
Nd
(5.12)
Let X = [x1 , . . . , xN −1 , x], and Z̃ = [z̃1 , . . . , z̃N −1 , z̃] for simplicity of notations. Let [Z1,i , zi ] =
arg min[C,c]∈Q |[X1,i , xi ] − [C, c]|1 where X1,i be the samples drawn in the encoding process for xi .
17
Algorithm 2 Encoding/decoding for improper dictionary learning
Given: Integer N . Convex set Q ⊂ Rd×N . Sampling probability ρ.
1. Encoding: input: data point x ∈ Rd ∼ D, output: codeword y = h(x) ∈ Rk×N Draw N − 1
data points X1 ∈ Rd×N −1 ∼ D N −1 .
Denoising step:
[Z1 , z] = arg min |[X1 , x] − [C, c]|1 ,
(5.11)
[C,c]∈Q
Random sampling step:
h(x) = PΩ (z) ,
where PΩ (z) is a random sampling of z where each entry is picked with probability ρ.
2. Decoding: input: codeword y ∈ Rk×N output: reconstruction g(y) ∈ Rd×N
Take N more samples x1 , . . . , xN . Encode them to y1 = h(x1 ), . . . , yN = h(xN ). Then, compute
[z̃1 , . . . , z̃N −1 , z̃] = arg min |PΩ ([C, c]) − [y1 , . . . , yN , y]|1 .
C∈Q
Return g(y) = z̃.
Let Z = [z1 , . . . , zN −1 , z]. By Lemma 5.1 and the symmetry, we have that for any i,
1
1
⋆
⋆
dN E |[Z1,i , zi ] − [X1,i , xi ]|1 6 ε . Thus, we have dN E [|X − Z|1 ] 6 ε .
Let Z̃ ∈ Rd×N be a helper matrix in the analysis defined as
Ẑ = arg min |X − C|1 .
1
d
E [|zi − xi |1 ] 6
(5.13)
C∈Q
Then by Lemma 5.1 we have that
1
Nd
h
E Ẑ − X
1
i
6 ε⋆ . Then by triangle inequality we obtain that
i
h
1
6 2ε⋆ .
E Ẑ − Z
Nd
1
(5.14)
Note that by definition Ẑ ∈ Q. Therefore, by Lemma 5.2, we have that
i
i
h
1 h
1
6
+ 2SRW m (W )
(by Lemma 5.2)
Z̃
−
Z
P
(
Z̃)
−
P
(Z)
E
E
Ω
Ω
Nd
m
1
1
i
1 h
6
E PΩ (Ẑ) − PΩ (Z) + 2SRW m (W )
m
1
(since Z̃ is the minimizer, and Ẑ ∈ Q.)
i
h
1
=
+ 2SRW m (W )
E Ẑ − Z
Nd
1
6 ε⋆ + δ .
(using equation (5.14) and Theorem 4.3 (c))
Finally by triangle inequality again we have
h
h
i
i
1
1
1
⋆
6
+
E Z̃ − X
E Z̃ − Z
E [|X − Z|1 ] 6 ε .
Nd
Nd
Nd
1
1
Combining equation (5.12) and (5.15) we complete the proof.
18
(5.15)
6 Conclusions
We have defined a new framework for unsupervised learning which replaces generative assumptions by
notions of reconstruction error and encoding length. This framework is comparative, and allows learning
of particular hypothesis classes with respect to an unknown distribution by other hypothesis classes.
We demonstrate its usefulness by giving new polynomial time algorithms for two unsupervised
hypothesis classes. First, we give new polynomial time algorithms for dictionary models in significantly
broader range of parameters and assumptions. Another domain is the class of spectral encodings, for
which we consider a new class of models that is shown to strictly encompass PCA and kernel-PCA.
This new class is capable, in contrast to previous spectral models, learn algebraic manifolds. We give
efficient learning algorithms for this class based on convex relaxations.
Acknowledgements
We thank Sanjeev Arora for many illuminating discussions and crucial observations in earlier phases
of this work, amongst them that a representation which preserves information for all classifiers requires
lossless compression.
References
[ADJ+ 13]
Jayadev Acharya, Hirakendu Das, Ashkan Jafarpour, Alon Orlitsky, and Ananda Theertha
Suresh. Tight bounds for universal compression of large alphabets. In Proceedings of the
2013 IEEE International Symposium on Information Theory, Istanbul, Turkey, July 7-12,
2013, pages 2875–2879, 2013.
[AEB05]
Michal Aharon, Michael Elad, and Alfred Bruckstein. K-svd: Design of dictionaries for
sparse representation. In IN: PROCEEDINGS OF SPARS05, pages 9–12, 2005.
[AGH+ 14] Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M. Kakade, and Matus Telgarsky. Tensor decompositions for learning latent variable models. J. Mach. Learn. Res.,
15(1):2773–2832, January 2014.
[AGM12]
Sanjeev Arora, Rong Ge, and Ankur Moitra. Learning topic models–going beyond svd. In
Foundations of Computer Science (FOCS), 2012 IEEE 53rd Annual Symposium on, pages
1–10. IEEE, 2012.
[AGM13]
Sanjeev Arora, Rong Ge, and Ankur Moitra. New algorithms for learning incoherent and
overcomplete dictionaries. arXiv preprint arXiv:1308.6273, 2013.
[AGMM15] Sanjeev Arora, Rong Ge, Tengyu Ma, and Ankur Moitra. Simple, efficient, and neural
algorithms for sparse coding. In Proceedings of The 28th Conference on Learning Theory,
COLT 2015, Paris, France, July 3-6, 2015, pages 113–149, 2015.
[BBV08]
Maria-Florina Balcan, Avrim Blum, and Santosh Vempala. A discriminative framework for
clustering via similarity functions. In Proceedings of the Fortieth Annual ACM Symposium
on Theory of Computing, STOC ’08, pages 671–680, 2008.
[BDA09]
Shai Ben-David and Margareta Ackerman. Measures of clustering quality: A working
set of axioms for clustering. In D. Koller, D. Schuurmans, Y. Bengio, and L. Bottou,
editors, Advances in Neural Information Processing Systems 21, pages 121–128. Curran
Associates, Inc., 2009.
19
[Ber71]
Toby Berger. Rate distortion theory: A mathematical basis for data compression. 1971.
[BKS14]
Boaz Barak, Jonathan A. Kelner, and David Steurer. Rounding sum-of-squares relaxations.
In STOC, pages 31–40, 2014.
[BKS15a]
Boaz Barak, Jonathan A. Kelner, and David Steurer. Dictionary learning and tensor decomposition via the sum-of-squares method. In Proceedings of the Forty-Seventh Annual
ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA, June 14-17,
2015, pages 143–151, 2015.
[BKS15b]
Boaz Barak, Jonathan A. Kelner, and David Steurer. Dictionary learning and tensor decomposition via the sum-of-squares method. In Proceedings of the Forty-seventh Annual
ACM Symposium on Theory of Computing, STOC ’15, 2015.
[BM15]
Boaz Barak and Ankur Moitra. Tensor prediction, rademacher complexity and random
3-xor. CoRR, abs/1501.06521, 2015.
[BM16]
Boaz Barak and Ankur Moitra. Noisy tensor completion via the sum-of-squares hierarchy.
In Proceedings of the 29th Conference on Learning Theory, COLT 2016, New York, USA,
June 23-26, 2016, pages 417–445, 2016.
[BNJ03]
David M. Blei, Andrew Y. Ng, and Michael I. Jordan. Latent dirichlet allocation. J. Mach.
Learn. Res., 3:993–1022, March 2003.
[BS14]
Boaz Barak and David Steurer. Sum-of-squares proofs and the quest toward optimal algorithms. In Proceedings of International Congress of Mathematicians (ICM), 2014. To
appear.
[CT06]
Thomas M. Cover and Joy A. Thomas. Elements of Information Theory (Wiley Series in
Telecommunications and Signal Processing). Wiley-Interscience, 2006.
[DH06]
D. L. Donoho and X. Huo. Uncertainty principles and ideal atomic decomposition. IEEE
Trans. Inf. Theor., 47(7):2845–2862, September 2006.
[GM15]
Rong Ge and Tengyu Ma. Decomposing overcomplete 3rd order tensors using sum-ofsquares algorithms. In Approximation, Randomization, and Combinatorial Optimization.
Algorithms and Techniques, APPROX/RANDOM 2015, August 24-26, 2015, Princeton, NJ,
USA, pages 829–849, 2015.
[HK12]
Elad Hazan and Satyen Kale. Projection-free online learning. In Proceedings of the 29th
International Conference on Machine Learning, ICML 2012, Edinburgh, Scotland, UK,
June 26 - July 1, 2012, 2012.
[HK13]
Daniel Hsu and Sham M Kakade. Learning mixtures of spherical gaussians: moment
methods and spectral decompositions. In Proceedings of the 4th conference on Innovations
in Theoretical Computer Science, pages 11–20. ACM, 2013.
[HSS15]
Samuel B. Hopkins, Jonathan Shi, and David Steurer. Tensor principal component analysis
via sum-of-square proofs. In Proceedings of The 28th Conference on Learning Theory,
COLT 2015, Paris, France, July 3-6, 2015, pages 956–1006, 2015.
[HSSS16]
Samuel B. Hopkins, Tselil Schramm, Jonathan Shi, and David Steurer. Fast spectral algorithms from sum-of-squares proofs: tensor decomposition and planted sparse vectors. In
Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing, STOC
2016, Cambridge, MA, USA, June 18-21, 2016, pages 178–191, 2016.
20
[JOS05]
Nikola Jevtic, Alon Orlitsky, and Narayana P. Santhanam. A lower bound on compression
of unknown alphabets. Theor. Comput. Sci., 332(1-3):293–311, 2005.
[Kle03]
Jon M. Kleinberg. An impossibility theorem for clustering. In S. Becker, S. Thrun, and
K. Obermayer, editors, Advances in Neural Information Processing Systems 15, pages 463–
470. MIT Press, 2003.
[Las01]
Jean B. Lasserre. Global optimization with polynomials and the problem of moments.
SIAM Journal on Optimization, 11(3):796–817, 2001.
[Las15]
Jean Bernard Lasserre. An introduction to polynomial and semi-algebraic optimization.
Cambridge Texts in Applied Mathematics. Cambridge: Cambridge University Press. ,
2015.
[Lau09]
Monique Laurent. Sums of squares, moment matrices and optimization over polynomials.
In Mihai Putinar and Seth Sullivant, editors, Emerging Applications of Algebraic Geometry,
volume 149 of The IMA Volumes in Mathematics and its Applications, pages 157–270.
Springer New York, 2009.
[LLS+ 13]
Roi Livni, David Lehavi, Sagi Schein, Hila Nachlieli, Shai Shalev-Shwartz, and Amir
Globerson. Vanishing component analysis. In Proceedings of the 30th International Conference on Machine Learning, ICML 2013, Atlanta, GA, USA, 16-21 June 2013, pages
597–605, 2013.
[LMSS07] Nati Linial, Shahar Mendelson, Gideon Schechtman, and Adi Shraibman. Complexity
measures of sign matrices. Combinatorica, 27(4):439–463, 2007.
[MRT12]
Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of machine
learning. MIT press, 2012.
[MSS16]
Tengyu Ma, Jonathan Shi, and David Steurer. Polynomial-time tensor decompositions with
sum-of-squares. In FOCS 2016, to appear, 2016.
[MW15]
Tengyu Ma and Avi Wigderson. Sum-of-squares lower bounds for sparse PCA. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada,
pages 1612–1620, 2015.
[Nat95]
B. K. Natarajan. Sparse approximate solutions to linear systems. SIAM J. Comput.,
24(2):227–234, 1995.
[OSZ04]
Alon Orlitsky, Narayana P. Santhanam, and Junan Zhang. Universal compression of memoryless sources over unknown alphabets. IEEE Trans. Information Theory, 50(7):1469–
1481, 2004.
[Par00]
Pablo A. Parrilo. Structured Semidefinite Programs and Semialgebraic Geometry Methods
in Robustness and Optimization. PhD thesis, California Institute of Technology, 2000.
[PWMH13] Hristo S Paskov, Robert West, John C Mitchell, and Trevor Hastie. Compressive feature
learning. In Advances in Neural Information Processing Systems, pages 2931–2939, 2013.
[Ris78]
Jorma Rissanen. Modeling by shortest data description. Automatica, 14(5):465–471, 1978.
[Sal09]
Ruslan Salakhutdinov. Learning Deep Generative Models. PhD thesis, University of
Toronto, Toronto, Ont., Canada, Canada, 2009. AAINR61080.
21
[SS05]
Nathan Srebro and Adi Shraibman. Rank, trace-norm and max-norm. In Learning Theory,
18th Annual Conference on Learning Theory, COLT 2005, Bertinoro, Italy, June 27-30,
2005, Proceedings, pages 545–560, 2005.
[SST08]
Ohad Shamir, Sivan Sabato, and Naftali Tishby. Learning and Generalization with the
Information Bottleneck, pages 92–107. Springer Berlin Heidelberg, Berlin, Heidelberg,
2008.
[Tib96]
Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society. Series B (Methodological), 58(1):267–288, 1996.
[TJ89]
N. Tomczak-Jaegermann. Banach-Mazur distances and finite-dimensional operator ideals.
Pitman monographs and surveys in pure and applied mathematics. Longman Scientific &
Technical, 1989.
[TPB00]
Naftali Tishby, Fernando C. N. Pereira, and William Bialek. The information bottleneck
method. CoRR, physics/0004057, 2000.
[Val84]
Leslie G Valiant. A theory of the learnable. Communications of the ACM, 27(11):1134–
1142, 1984.
[Vap98]
Vladimir N. Vapnik. Statistical Learning Theory. Wiley-Interscience, 1998.
[Vu11]
Van Vu. Singular vectors under random perturbation. Random Structures and Algorithms,
39(4):526–538, 2011.
22
A
Proof of Theorem 2.1
Proof of Theorem 2.1. [MRT12, Theorem 3.1] asserts that with probability at least 1 − δ, we have that
for every hypothesis f ∈ H,
s
log 1δ
loss(f ) 6 loss(f ) + 2Rm (H) +
D
S
2m
by negating the loss function this gives
| loss(f ) − loss(f )| 6 2Rm (H) +
D
S
s
log 2δ
2m
and therefore, letting f ∗ = arg minf ∈H lossD (f ), we have
s
1
log δ
loss(fˆERM ) 6 loss(fˆERM ) + 2Rm (H) +
D
S
2m
s
log 1δ
6 loss(f ∗ ) + 2Rm (H) +
S
2m
s
4 log 1δ
6 loss(f ∗ ) + 6Rm (H) +
D
2m
(by [MRT12, Theorem 3.1])
( by definition of ERM)
( using [MRT12, Theorem 3.1] again)
A.1 Low reconstruction error is sufficient for supervised learning
This section observes that low reconstruction error is a sufficient condition for unsupervised learning to
allow supervised learning over any future task.
Lemma A.1. Consider any supervised learning problem with respect to distribution D over X × L that
is agnostically PAC-learnable with respect to L-Lipschitz loss function ℓ and with bias γ1 .
Suppose that unsupervised hypothesis class H is C -learnable with bias γ2 over distribution D and
domain X , by hypothesis f : X 7→ Y. Then the distribution D̃f over Y × L, which gives the pair
(f (x), y) the same measure as D gives to (x, y), is agnostically PAC-learnable with bias γ1 + Lγ2 .
Proof. Let h : X 7→ Y be a hypothesis that PAC-learns distribution D. Consider the hypothesis
h̃ : Y 7→ L , h̃(y) = (h ◦ g)(y)
Then by definition of reconstruction error and the Lipschitz property of ℓ we have
loss(h̃) =
D̃f
E
[ℓ(h̃(y), l)]
E
[ℓ((h ◦ g)(y), l)]
(y,l)∼D̃f
=
(y,l)∼D̃f
=
E
[ℓ(h(x̃), l)]
E
[ℓ(h(x), l)] +
(D(x) = D̃f (y))
(x,l)∼D
=
(x,l)∼D
= γ1 +
E
(x,l)∼D
E
(x,l)∼D
[ℓ(h(x̃), l) − ℓ(h(x), l)]
[ℓ(h(x̃), l) − ℓ(h(x), l)]
23
( PAC learnability)
6 γ1 + L E kx − x̃k
( Lipschitzness of ℓ ◦ h)
x∼D
= γ1 + L E kx − g ◦ f (x)k
x∼D
6 γ1 + Lγ2
( C -learnability)
B Proof of Theorem 3.1
Proof of Theorem 3.1. We assume without loss of generality s = 1. For s > 1 the proof will be identical
since one can assume x⊗s is the data points (and the dimension is raised to ds ).
Let x1 , . . . , xm be a set of examples ∼ D m . It can be shown that any minimizer of ERM
A∗ = arg min kxi − A† Axi k2
(B.1)
A∈Rd×k
P
⊤
satisfies that (A∗ )† A∗ is the the projection operator to the subspace of top k eigenvector of m
i=1 xi xi .
Therefore ERM (B.1) is efficiently solvable.
According to Theorem 2.1, the ERM hypothesis generalizes with rates governed by the Rademacher
complexity of the hypothesis class. Thus, it remains to compute the Rademacher complexity of the
hypothesis class for PCA. We assume for simplicity that all vectors in the domain have Euclidean norm
bounded by one.
#
"
X
1
σi ℓ((h, g), xi )
RS (Hkpca ) =
sup
E
σ∼{±1}m (h,g)∈Hpca m
i∈S
k
"
#
1 X
†
2
=
sup
σi kxi − A Axi k
E
σ∼{±1}m A∈Rd×k m
i∈S
!
#
"
m
X
X
1
(I − A† A)⊤ )
xi x⊤
σi Tr((I − A† A)
=
sup
E
i
σ∼{±1}m A∈Rd×k m
i=1
i∈S
"
!!#
m
1 X
†
⊤
sup Tr (I − A A)
=
.
σi xi xi
E
m
σ∼{±1}m A∈Rd×k
i=1
Then we apply Holder inequality, and effectively disentangle the part about σ and A:
"
!!#
m
X
1
sup Tr (I − A† A)
σi xi x⊤
E
i
m
σ∼{±1}m A∈Rd×k
i=1
"
#
m
1 X
†
⊤
6
sup kI − A AkF
σi xi xi
(Holder inequality)
E
m
σ∼{±1}m A∈Rd×k
i=1
F
#
"
m
X
√
1
( since A† A is a projection, kI − A† Ak 6 1.)
σi xi x⊤
6 d
E
i
m
m
σ∼{±1}
i=1
F
2 1/2
m
X
√
1
6 d
σi xi x⊤
(Cauchy-Schwarz inequality)
E
i
m
m
σ∼{±1}
i=1
F
v
u
m
√ u 1 X
⊤
t
hσi xi x⊤
(since E[σi σj ] = 0 for i 6= j)
6 d
i , σi xi xi i
m2
i=1
24
6
p
(by kxk 6 1)
d/m .
pca
Thus, from Theorem 2.1 we can conclude that the class Hk
Õ( εd2 )6 .
B.1
is learnable with sample complexity
Proof of Theorem 3.2
Theorem 3.2 follows from the following lemmas and the generalization theorem 2.1 straightforwardly.
Lemma B.1. Suppose distribution D is (k, ε)-regularly spectral decodable. Then for any δ > 0, solving
convex optimization (3.3) with non-smooth Frank-Wolfe algorithm [HK12, Theorem 4.4] with k′ =
O(τ 4 k4 /δ4 ) steps gives a solution R̂ of rank k′ such that f (R̂) 6 δ + ε.
Lemma B.2. The Rademacher ocomplexity of the class of function Φ
=
n
Rx⊗2 − x⊗2 op : R s.t. kRkS1 6 τ k with m examples is bounded from above by at most
p
Rm (Φ) 6 2τ k · 1/m
Here Lemma B.2 follows from the fact that Rx⊗2 − x⊗2 op is bounded above by 2τ k when kxk 6
1 and kRkS1 6 τ k. The rest of the section focuses on the proof of Lemma B.1.
Lemma B.1 basically follows from the fact that f is Lipschitz and guarantees of the Frank-Wolfe
algorithm.
Proposition B.3. The objective function f (R) is convex and 1-Lipschitz. Concretely, Let ℓx (R) =
kRx⊗2 − x⊗2 kop . Then
∂ℓx ∋ (u ⊗ v)(x⊗2 )⊤
where ∂ℓx is the set of sub-gradients of ℓx with respect to R, and u, v ∈ Rd are (one pair of) top left and
right singular vectors of M(Rx⊗2 − x⊗2 ).
Proof. This simply follows from calculating gradient with chain rule. Here we use the fact that A ∈
(Rd )⊗2 , the sub-gradient of kAkop contains the vector a ⊗ b where a, b are the top left and right singular
vectors of M(A). We can also verify by definition that (u ⊗ v)(x⊗2 )⊤ is a sub-gradient.
f (R′ ) − f (R) > (u ⊗ v)⊤ (R′ x⊗2 − Rx⊗2 )
= h(u ⊗ v)(x⊗2 )⊤ , R′ − Ri .
(by convexity of k · kop )
Now we are ready to prove Lemma B.1.
Proof of Lemma B.1. Since D is (k, ε)-regularly decodable, we know that there exists a rank-k solution
R∗ with f (R∗ ) 6 ε. Since kR∗ kop 6 τ , we have that kRkS1 6 rank(R∗ ) · kRkop 6 τ k. Therefore R∗
is feasible solution for the objective (3.3) with f (R∗ ) 6 ε.
By Proposition B.3, we have that f (R) is 1-Lipschitz. Moreover, for any R, S with kRkS1 6
τ k, kSkS1 6 τ k we have that kR − SkF 6 kRkF + kSkF 6 kRkS1 + kSkS1 6 2τ k. Therefore the
diameter of the constraint set is at most τ k.
By [HK12, Theorem 4.4], we have that Frank-Wolfe algorithm returns solution R with f (R) −
4
f (R∗ ) 6 ε + δ in τδk iteration.
6
For ℓ > 1 the sample complexity is Õ(dℓ /ε2 ).
25
C
Shorter codes with relaxed objective for Polynomial Spectral Components Analysis
Notations. For a matrix A, let σ1 (A) > σ2 (A) > .. be its singular values. Then the Schatten p-norm,
P
denoted by k · kSp , for p > 1 is defined as kAkSp = ( i σi (A)p )1/p . For even integer p, an equivalent
and simple definition is that kAkpSp , Tr((A⊤ A)p/2 ).
In this section we consider the following further relaxation of objective (3.3).
h
i
2
(C.1)
minimize f4 (R) := E Rx⊗2 − x⊗2 Sp
s.t. kRkS1 6 τ k
Since kAkF > kAkS4 > kAkS∞ = kAk, this is a relaxation of the objective (3.3), and it interpolates
between kernal PCA and spectral decoding. Our assumption is weaker than kernal PCA but stronger
than spectral decodable.
Definition C.1 (Extension of definition 3.2). We say a data distribution D is (k, ε)-regularly spectral
decodable with k · kSp norm if the error E in equation (5.10) is bounded by kEkSp 6 ε.
We can reduce the length of the code from O(k4 ) to O(k2 ) for any constant p.
Theorem C.1. Suppose data distribution is (k, ε)-spectral decodable with norm k · kSp for p = O(1),
then solving (C.1) using (usual) Frank-Wolfe algorithm gives a solution R̂ of k′ = O(k2 τ 2 /ε2 ) with
f (R) 6 ε + δ. As a direct consequence, we obtain encoding and decoding pair (gA , hB ) ∈ Hksa′ with
k′ = O(k2 τ 2 /ε2 ) and reconstruction error ε + δ.
The main advantage of using relaxed objective is its smoothness. This allows us to optimize over
the Schatten 1-norm constraint set much faster using usual Frank-Wolfe algorithm. Therefore the key
here is to establish the smoothness of the objective function. Theorem C.1 follows from the following
proposition straightforwardly.
Proposition C.2. Objective function fp (equation (C.1)) is convex and O(p)-smooth.
Proof. Since k·kSp is convex and composition of convex function with linear function gives convex function. Therefore, Rx⊗2 − x⊗2 Sp is a convex function. The square of an non-negative convex function
is also convex, and therefore we proved that fp is convex. We prove the smoothness by first showing
that kAk2Sp is a smooth function with respect to A. We use the definition kAkpSp = Tr((A⊤ A)p/2 ). Let
E be a small matrix that goes to 0, we have
kA + EkpSp = Tr((A⊤ A)p/2 ) + T1 + T2 + o(kEk2F )
(C.2)
where T1 , T2 denote the first order term and second order term respectively. Let U = A⊤ E +
E ⊤ A and V = A⊤ A. We note that T2 is a sum of the traces of matrices like U V U V U p/2−2 .
By Lieb-Thirring inequality, we have that all these term can be bounded by Tr(U p/2−2 V 2 ) =
2Tr((A⊤ A)p/2−2 A⊤ EE ⊤ A) + 2Tr((A⊤ A)p/2−2 A⊤ EA⊤ E ⊤ ). For the first term, we have that
2
Tr((A⊤ A)p/2−2 A⊤ EE ⊤ A) 6 k(AA⊤ )(p−2)/4 Ek2 6 k(AA⊤ )(p−2)/4 k2S∞ kEk2F = kAkp−2
S∞ kEkF
where in the first inequality we use Cauchy-Schwarz. Then for the second term we have
Tr((A⊤ A)p/2−2 A⊤ EA⊤ E ⊤ ) 6 k(A⊤ A)(p−2)/4 EkF kAE(A⊤ A)(p−4)/4 kF
6 k(A⊤ A)(p−2)/4 Ek2F
⊤ (p−2)/4 2
6 k(AA )
26
k
kEk2F
(by Lieb-Thirring inequality)
2
= kAkp−2
S∞ kEkF
(C.3)
Therefore, finally we got
2
T2 6 O(p2 ) · kAkp−2
Sp−2 kEkF
(C.4)
Therefore, we complete the proof by having,
kA + Ek2Sp 6 (kAkpSp + T1 + T2 + o(kEk2 ))2/p 6 kAkSp (1 + T1′ +
2
p/2
pkAkSp
T2 ) + o(kEk2 )
(by (1 + x)p/2 6 1 + 2x/p + o(kxk2 ))
6 kAk2Sp + T1′′ + O(p)kEk2 + o(kEk2 )
D
(by equation (C.3))
Toolbox
Lemma D.1. Let p > 2 be a power of 2 and u = [u1 , . . . , un ] and v = [v1 , . . . , vn ] be indeterminants.
Then there exists SoS proof,
p
!
X
X p X p p−1
p−1
(D.1)
vi
ui vj 6
ui
⊢
j
i
Proof Sketch. The inequality follows from repeated application of Cauchy-Schwarz. For example, for
p = 4 we have
⊢
X
i
u4i
!
X
vi4
3
>
X
u2i vi2
i
>
X
i
ui vi3
!2
!4
X
vi4
For p = 2s with s > 2, the statement can be proved inductively.
27
2
(by Cauchy-Schwarz)
(by Cauchy-Schwarz again)
| 8cs.DS
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.